From 8b8b792802c34dd57f16353632fce65c164c360f Mon Sep 17 00:00:00 2001
From: Lornatang <liuchangyu1111@gmail.com>
Date: Tue, 16 Jun 2020 08:28:52 +0800
Subject: [PATCH 01/73] There is no need to download extra packages, official
 bring it with you

I submitted it once in your yolov3 project, you seem to accept it? I'm not sure. I'll submit PR again.
---
 utils/torch_utils.py | 19 ++++++++++++-------
 1 file changed, 12 insertions(+), 7 deletions(-)

diff --git a/utils/torch_utils.py b/utils/torch_utils.py
index 7eea4b4..e069792 100644
--- a/utils/torch_utils.py
+++ b/utils/torch_utils.py
@@ -7,6 +7,7 @@ import torch
 import torch.backends.cudnn as cudnn
 import torch.nn as nn
 import torch.nn.functional as F
+import torchvision.models as models
 
 
 def init_seeds(seed=0):
@@ -120,18 +121,22 @@ def model_info(model, verbose=False):
 
 def load_classifier(name='resnet101', n=2):
     # Loads a pretrained model reshaped to n-class output
-    import pretrainedmodels  # https://github.com/Cadene/pretrained-models.pytorch#torchvision
-    model = pretrainedmodels.__dict__[name](num_classes=1000, pretrained='imagenet')
+    model = models.__dict__[name](pretrained=True)
 
     # Display model properties
-    for x in ['model.input_size', 'model.input_space', 'model.input_range', 'model.mean', 'model.std']:
+    input_size = [3, 224, 224]
+    input_space = 'RGB'
+    input_range = [0, 1]
+    mean = [0.485, 0.456, 0.406]
+    std = [0.229, 0.224, 0.225]
+    for x in [input_size, input_space, input_range, mean, std]:
         print(x + ' =', eval(x))
 
     # Reshape output to n classes
-    filters = model.last_linear.weight.shape[1]
-    model.last_linear.bias = torch.nn.Parameter(torch.zeros(n))
-    model.last_linear.weight = torch.nn.Parameter(torch.zeros(n, filters))
-    model.last_linear.out_features = n
+    filters = model.fc.weight.shape[1]
+    model.fc.bias = torch.nn.Parameter(torch.zeros(n), requires_grad=True)
+    model.fc.weight = torch.nn.Parameter(torch.zeros(n, filters), requires_grad=True)
+    model.fc.out_features = n
     return model
 
 

From e1e33992865ddc9eb5a0d7f0ed1ccfb7d7a33e8c Mon Sep 17 00:00:00 2001
From: Glenn Jocher <glenn.jocher@ultralytics.com>
Date: Tue, 16 Jun 2020 09:59:42 -0700
Subject: [PATCH 02/73] ONNX export bug fix #93

---
 models/onnx_export.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/models/onnx_export.py b/models/onnx_export.py
index fd8abeb..59d13b1 100644
--- a/models/onnx_export.py
+++ b/models/onnx_export.py
@@ -25,7 +25,7 @@ if __name__ == '__main__':
 
     # Load pytorch model
     google_utils.attempt_download(opt.weights)
-    model = torch.load(opt.weights, map_location=torch.device('cpu'))['model']
+    model = torch.load(opt.weights, map_location=torch.device('cpu'))['model'].float()
     model.eval()
     model.fuse()
 

From ec81c7b5f2b83b8c74f1590f4bccdda94670bb0d Mon Sep 17 00:00:00 2001
From: Glenn Jocher <glenn.jocher@ultralytics.com>
Date: Tue, 16 Jun 2020 10:14:04 -0700
Subject: [PATCH 03/73] check_anchors() bug fix #90

---
 train.py       |  2 +-
 utils/utils.py | 12 ++++++------
 2 files changed, 7 insertions(+), 7 deletions(-)

diff --git a/train.py b/train.py
index 1e2d55a..4aff06c 100644
--- a/train.py
+++ b/train.py
@@ -199,7 +199,7 @@ def train(hyp):
     tb_writer.add_histogram('classes', c, 0)
 
     # Check anchors
-    check_best_possible_recall(dataset, anchors=model.model[-1].anchor_grid, thr=hyp['anchor_t'], imgsz=imgsz)
+    check_anchors(dataset, anchors=model.model[-1].anchor_grid, thr=hyp['anchor_t'], imgsz=imgsz)
 
     # Exponential moving average
     ema = torch_utils.ModelEMA(model)
diff --git a/utils/utils.py b/utils/utils.py
index 95d1198..adaadac 100755
--- a/utils/utils.py
+++ b/utils/utils.py
@@ -52,8 +52,9 @@ def check_img_size(img_size, s=32):
     return make_divisible(img_size, s)  # nearest gs-multiple
 
 
-def check_best_possible_recall(dataset, anchors, thr=4.0, imgsz=640):
+def check_anchors(dataset, model, thr=4.0, imgsz=640):
     # Check best possible recall of dataset with current anchors
+    anchors = model.module.model[-1].anchor_grid if hasattr(model, 'module') else model.model[-1].anchor_grid
     shapes = imgsz * dataset.shapes / dataset.shapes.max(1, keepdims=True)
     wh = torch.tensor(np.concatenate([l[:, 3:5] * s for s, l in zip(shapes, dataset.labels)])).float()  # wh
     ratio = wh[:, None] / anchors.view(-1, 2).cpu()[None]  # ratio
@@ -62,7 +63,6 @@ def check_best_possible_recall(dataset, anchors, thr=4.0, imgsz=640):
     mr = (m < thr).float().mean()  # match ratio
     print(('AutoAnchor labels:' + '%10s' * 6) % ('n', 'mean', 'min', 'max', 'matching', 'recall'))
     print(('                  ' + '%10.4g' * 6) % (wh.shape[0], wh.mean(), wh.min(), wh.max(), mr, bpr))
-
     assert bpr > 0.9, 'Best possible recall %.3g (BPR) below 0.9 threshold. Training cancelled. ' \
                       'Compute new anchors with utils.utils.kmeans_anchors() and update model before training.' % bpr
 
@@ -512,10 +512,10 @@ def build_targets(p, targets, model):
 
 
 def non_max_suppression(prediction, conf_thres=0.1, iou_thres=0.6, fast=False, classes=None, agnostic=False):
-    """
-    Performs  Non-Maximum Suppression on inference results
-    Returns detections with shape:
-        nx6 (x1, y1, x2, y2, conf, cls)
+    """Performs Non-Maximum Suppression (NMS) on inference results
+
+    Returns:
+         detections with shape: nx6 (x1, y1, x2, y2, conf, cls)
     """
     if prediction.dtype is torch.float16:
         prediction = prediction.float()  # to FP32

From 7c6fd4d6701d1e4aefc41af95be87b9430f171ca Mon Sep 17 00:00:00 2001
From: Glenn Jocher <glenn.jocher@ultralytics.com>
Date: Tue, 16 Jun 2020 10:22:22 -0700
Subject: [PATCH 04/73] --verbose bug fix #84

---
 test.py | 7 ++++---
 1 file changed, 4 insertions(+), 3 deletions(-)

diff --git a/test.py b/test.py
index 8d252ff..1f937a7 100644
--- a/test.py
+++ b/test.py
@@ -17,10 +17,10 @@ def test(data,
          save_json=False,
          single_cls=False,
          augment=False,
+         verbose=False,
          model=None,
          dataloader=None,
-         fast=False,
-         verbose=False):
+         fast=False):
     # Initialize/load model and set device
     if model is None:
         training = False
@@ -270,7 +270,8 @@ if __name__ == '__main__':
              opt.iou_thres,
              opt.save_json,
              opt.single_cls,
-             opt.augment)
+             opt.augment,
+             opt.verbose)
 
     elif opt.task == 'study':  # run over a range of settings and save/plot
         for weights in ['yolov5s.pt', 'yolov5m.pt', 'yolov5l.pt', 'yolov5x.pt']:

From 05b8ee5ca4a001eb7a4946d36788f4424f6749c1 Mon Sep 17 00:00:00 2001
From: Glenn Jocher <glenn.jocher@ultralytics.com>
Date: Tue, 16 Jun 2020 10:34:16 -0700
Subject: [PATCH 05/73] check_anchors() bug fix #102

---
 train.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/train.py b/train.py
index 4aff06c..db135a7 100644
--- a/train.py
+++ b/train.py
@@ -199,7 +199,7 @@ def train(hyp):
     tb_writer.add_histogram('classes', c, 0)
 
     # Check anchors
-    check_anchors(dataset, anchors=model.model[-1].anchor_grid, thr=hyp['anchor_t'], imgsz=imgsz)
+    check_anchors(dataset, model=model.model[-1].anchor_grid, thr=hyp['anchor_t'], imgsz=imgsz)
 
     # Exponential moving average
     ema = torch_utils.ModelEMA(model)

From 5a50491fa12a515f829407b2cfc4ec0acb05311d Mon Sep 17 00:00:00 2001
From: Glenn Jocher <glenn.jocher@ultralytics.com>
Date: Tue, 16 Jun 2020 10:36:35 -0700
Subject: [PATCH 06/73] check_anchors bug fix

---
 train.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/train.py b/train.py
index db135a7..484190a 100644
--- a/train.py
+++ b/train.py
@@ -199,7 +199,7 @@ def train(hyp):
     tb_writer.add_histogram('classes', c, 0)
 
     # Check anchors
-    check_anchors(dataset, model=model.model[-1].anchor_grid, thr=hyp['anchor_t'], imgsz=imgsz)
+    check_anchors(dataset, model=model, thr=hyp['anchor_t'], imgsz=imgsz)
 
     # Exponential moving average
     ema = torch_utils.ModelEMA(model)

From 57a0ae33501185882cf63cfd947f22aef59844c0 Mon Sep 17 00:00:00 2001
From: Glenn Jocher <glenn.jocher@ultralytics.com>
Date: Tue, 16 Jun 2020 13:12:22 -0700
Subject: [PATCH 07/73] AutoAnchor implementation

---
 utils/utils.py | 95 ++++++++++++++++++++++++--------------------------
 1 file changed, 45 insertions(+), 50 deletions(-)

diff --git a/utils/utils.py b/utils/utils.py
index adaadac..7069181 100755
--- a/utils/utils.py
+++ b/utils/utils.py
@@ -53,18 +53,23 @@ def check_img_size(img_size, s=32):
 
 
 def check_anchors(dataset, model, thr=4.0, imgsz=640):
-    # Check best possible recall of dataset with current anchors
+    # Check anchor fit to data, recompute if necessary
+    print('\nAnalyzing anchors... ', end='')
     anchors = model.module.model[-1].anchor_grid if hasattr(model, 'module') else model.model[-1].anchor_grid
     shapes = imgsz * dataset.shapes / dataset.shapes.max(1, keepdims=True)
     wh = torch.tensor(np.concatenate([l[:, 3:5] * s for s, l in zip(shapes, dataset.labels)])).float()  # wh
     ratio = wh[:, None] / anchors.view(-1, 2).cpu()[None]  # ratio
     m = torch.max(ratio, 1. / ratio).max(2)[0]  # max ratio
     bpr = (m.min(1)[0] < thr).float().mean()  # best possible recall
-    mr = (m < thr).float().mean()  # match ratio
-    print(('AutoAnchor labels:' + '%10s' * 6) % ('n', 'mean', 'min', 'max', 'matching', 'recall'))
-    print(('                  ' + '%10.4g' * 6) % (wh.shape[0], wh.mean(), wh.min(), wh.max(), mr, bpr))
-    assert bpr > 0.9, 'Best possible recall %.3g (BPR) below 0.9 threshold. Training cancelled. ' \
-                      'Compute new anchors with utils.utils.kmeans_anchors() and update model before training.' % bpr
+    # mr = (m < thr).float().mean()  # match ratio
+
+    print('Best Possible Recall (BPR) = %.3f' % bpr, end='')
+    if bpr < 0.99:  # threshold to recompute
+        print('. Generating new anchors for improved recall, please wait...' % bpr)
+        new_anchors = kmean_anchors(dataset, n=9, img_size=640, thr=4.0, gen=1000, verbose=False)
+        anchors[:] = torch.tensor(new_anchors).view_as(anchors).type_as(anchors)
+        print('New anchors saved to model. Update model *.yaml to use these anchors in the future.')
+    print('')  # newline
 
 
 def check_file(file):
@@ -689,14 +694,14 @@ def coco_single_class_labels(path='../coco/labels/train2014/', label_class=43):
             shutil.copyfile(src=img_file, dst='new/images/' + Path(file).name.replace('txt', 'jpg'))  # copy images
 
 
-def kmean_anchors(path='./data/coco128.yaml', n=9, img_size=(640, 640), thr=0.20, gen=1000):
+def kmean_anchors(path='./data/coco128.yaml', n=9, img_size=640, thr=4.0, gen=1000, verbose=True):
     """ Creates kmeans-evolved anchors from training dataset
 
         Arguments:
-            path: path to dataset *.yaml
+            path: path to dataset *.yaml, or a loaded dataset
             n: number of anchors
-            img_size: (min, max) image size used for multi-scale training (can be same values)
-            thr: IoU threshold hyperparameter used for training (0.0 - 1.0)
+            img_size: image size used for training
+            thr: anchor-label wh ratio threshold hyperparameter hyp['anchor_t'] used for training, default=4.0
             gen: generations to evolve anchors using genetic algorithm
 
         Return:
@@ -705,52 +710,41 @@ def kmean_anchors(path='./data/coco128.yaml', n=9, img_size=(640, 640), thr=0.20
         Usage:
             from utils.utils import *; _ = kmean_anchors()
     """
+    thr = 1. / thr
+
+    def metric(k):  # compute metrics
+        r = wh[:, None] / k[None]
+        x = torch.min(r, 1. / r).min(2)[0]  # ratio metric
+        # x = wh_iou(wh, torch.tensor(k))  # iou metric
+        return x, x.max(1)[0]  # x, best_x
 
-    from utils.datasets import LoadImagesAndLabels
+    def fitness(k):  # mutation fitness
+        _, best = metric(k)
+        return (best * (best > thr).float()).mean()  # fitness
 
     def print_results(k):
         k = k[np.argsort(k.prod(1))]  # sort small to large
-        iou = wh_iou(wh, torch.Tensor(k))
-        max_iou = iou.max(1)[0]
-        bpr, aat = (max_iou > thr).float().mean(), (iou > thr).float().mean() * n  # best possible recall, anch > thr
-
-        # thr = 5.0
-        # r = wh[:, None] / k[None]
-        # ar = torch.max(r, 1. / r).max(2)[0]
-        # max_ar = ar.min(1)[0]
-        # bpr, aat = (max_ar < thr).float().mean(), (ar < thr).float().mean() * n  # best possible recall, anch > thr
-
-        print('%.2f iou_thr: %.3f best possible recall, %.2f anchors > thr' % (thr, bpr, aat))
-        print('n=%g, img_size=%s, IoU_all=%.3f/%.3f-mean/best, IoU>thr=%.3f-mean: ' %
-              (n, img_size, iou.mean(), max_iou.mean(), iou[iou > thr].mean()), end='')
+        x, best = metric(k)
+        bpr, aat = (best > thr).float().mean(), (x > thr).float().mean() * n  # best possible recall, anch > thr
+        print('thr=%.2f: %.3f best possible recall, %.2f anchors past thr' % (thr, bpr, aat))
+        print('n=%g, img_size=%s, metric_all=%.3f/%.3f-mean/best, past_thr=%.3f-mean: ' %
+              (n, img_size, x.mean(), best.mean(), x[x > thr].mean()), end='')
         for i, x in enumerate(k):
             print('%i,%i' % (round(x[0]), round(x[1])), end=',  ' if i < len(k) - 1 else '\n')  # use in *.cfg
         return k
 
-    def fitness(k):  # mutation fitness
-        iou = wh_iou(wh, torch.Tensor(k))  # iou
-        max_iou = iou.max(1)[0]
-        return (max_iou * (max_iou > thr).float()).mean()  # product
-
-    # def fitness_ratio(k):  # mutation fitness
-    #     # wh(5316,2), k(9,2)
-    #     r = wh[:, None] / k[None]
-    #     x = torch.max(r, 1. / r).max(2)[0]
-    #     m = x.min(1)[0]
-    #     return 1. / (m * (m < 5).float()).mean()  # product
+    if isinstance(path, str):  # *.yaml file
+        with open(path) as f:
+            data_dict = yaml.load(f, Loader=yaml.FullLoader)  # model dict
+        from utils.datasets import LoadImagesAndLabels
+        dataset = LoadImagesAndLabels(data_dict['train'], augment=True, rect=True)
+    else:
+        dataset = path  # dataset
 
     # Get label wh
-    wh = []
-    with open(path) as f:
-        data_dict = yaml.load(f, Loader=yaml.FullLoader)  # model dict
-    dataset = LoadImagesAndLabels(data_dict['train'], augment=True, rect=True)
-    nr = 1 if img_size[0] == img_size[1] else 3  # number augmentation repetitions
-    for s, l in zip(dataset.shapes, dataset.labels):
-        # wh.append(l[:, 3:5] * (s / s.max()))  # image normalized to letterbox normalized wh
-        wh.append(l[:, 3:5] * s)  # image normalized to pixels
-    wh = np.concatenate(wh, 0).repeat(nr, axis=0)  # augment 3x
-    # wh *= np.random.uniform(img_size[0], img_size[1], size=(wh.shape[0], 1))  # normalized to pixels (multi-scale)
-    wh = wh[(wh > 2.0).all(1)]  # remove below threshold boxes (< 2 pixels wh)
+    shapes = img_size * dataset.shapes / dataset.shapes.max(1, keepdims=True)
+    wh = torch.tensor(np.concatenate([l[:, 3:5] * s for s, l in zip(shapes, dataset.labels)])).float()  # wh
+    wh = wh[(wh > 2.0).all(1)].numpy()  # filter > 2 pixels
 
     # Kmeans calculation
     from scipy.cluster.vq import kmeans
@@ -758,10 +752,10 @@ def kmean_anchors(path='./data/coco128.yaml', n=9, img_size=(640, 640), thr=0.20
     s = wh.std(0)  # sigmas for whitening
     k, dist = kmeans(wh / s, n, iter=30)  # points, mean distance
     k *= s
-    wh = torch.Tensor(wh)
+    wh = torch.tensor(wh)
     k = print_results(k)
 
-    # # Plot
+    # Plot
     # k, d = [None] * 20, [None] * 20
     # for i in tqdm(range(1, 21)):
     #     k[i-1], d[i-1] = kmeans(wh / s, i)  # points, mean distance
@@ -777,7 +771,7 @@ def kmean_anchors(path='./data/coco128.yaml', n=9, img_size=(640, 640), thr=0.20
     # Evolve
     npr = np.random
     f, sh, mp, s = fitness(k), k.shape, 0.9, 0.1  # fitness, generations, mutation prob, sigma
-    for _ in tqdm(range(gen), desc='Evolving anchors'):
+    for _ in tqdm(range(gen), desc='Evolving anchors with Genetic Algorithm:'):
         v = np.ones(sh)
         while (v == 1).all():  # mutate until a change occurs (prevent duplicates)
             v = ((npr.random(sh) < mp) * npr.random() * npr.randn(*sh) * s + 1).clip(0.3, 3.0)
@@ -785,7 +779,8 @@ def kmean_anchors(path='./data/coco128.yaml', n=9, img_size=(640, 640), thr=0.20
         fg = fitness(kg)
         if fg > f:
             f, k = fg, kg.copy()
-            print_results(k)
+            if verbose:
+                print_results(k)
     k = print_results(k)
     return k
 

From bafbc65ee352cd87c3f5f191995190a609a2a690 Mon Sep 17 00:00:00 2001
From: Glenn Jocher <glenn.jocher@ultralytics.com>
Date: Tue, 16 Jun 2020 14:26:02 -0700
Subject: [PATCH 08/73] AutoAnchor bug fix

---
 utils/utils.py | 10 +++++-----
 1 file changed, 5 insertions(+), 5 deletions(-)

diff --git a/utils/utils.py b/utils/utils.py
index 7069181..dd12201 100755
--- a/utils/utils.py
+++ b/utils/utils.py
@@ -719,7 +719,7 @@ def kmean_anchors(path='./data/coco128.yaml', n=9, img_size=640, thr=4.0, gen=10
         return x, x.max(1)[0]  # x, best_x
 
     def fitness(k):  # mutation fitness
-        _, best = metric(k)
+        _, best = metric(torch.tensor(k, dtype=torch.float32))
         return (best * (best > thr).float()).mean()  # fitness
 
     def print_results(k):
@@ -743,8 +743,8 @@ def kmean_anchors(path='./data/coco128.yaml', n=9, img_size=640, thr=4.0, gen=10
 
     # Get label wh
     shapes = img_size * dataset.shapes / dataset.shapes.max(1, keepdims=True)
-    wh = torch.tensor(np.concatenate([l[:, 3:5] * s for s, l in zip(shapes, dataset.labels)])).float()  # wh
-    wh = wh[(wh > 2.0).all(1)].numpy()  # filter > 2 pixels
+    wh = np.concatenate([l[:, 3:5] * s for s, l in zip(shapes, dataset.labels)])  # wh
+    wh = wh[(wh > 2.0).all(1)]  # filter > 2 pixels
 
     # Kmeans calculation
     from scipy.cluster.vq import kmeans
@@ -752,7 +752,7 @@ def kmean_anchors(path='./data/coco128.yaml', n=9, img_size=640, thr=4.0, gen=10
     s = wh.std(0)  # sigmas for whitening
     k, dist = kmeans(wh / s, n, iter=30)  # points, mean distance
     k *= s
-    wh = torch.tensor(wh)
+    wh = torch.tensor(wh, dtype=torch.float32)
     k = print_results(k)
 
     # Plot
@@ -771,7 +771,7 @@ def kmean_anchors(path='./data/coco128.yaml', n=9, img_size=640, thr=4.0, gen=10
     # Evolve
     npr = np.random
     f, sh, mp, s = fitness(k), k.shape, 0.9, 0.1  # fitness, generations, mutation prob, sigma
-    for _ in tqdm(range(gen), desc='Evolving anchors with Genetic Algorithm:'):
+    for _ in tqdm(range(gen), desc='Evolving anchors with Genetic Algorithm'):
         v = np.ones(sh)
         while (v == 1).all():  # mutate until a change occurs (prevent duplicates)
             v = ((npr.random(sh) < mp) * npr.random() * npr.randn(*sh) * s + 1).clip(0.3, 3.0)

From 72d06147b20edfc1b3227c5443ee6bb79ac8ef2a Mon Sep 17 00:00:00 2001
From: Glenn Jocher <glenn.jocher@ultralytics.com>
Date: Tue, 16 Jun 2020 14:42:30 -0700
Subject: [PATCH 09/73] Multi-GPU disabled in test.py

---
 test.py | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/test.py b/test.py
index 1f937a7..aa5cde3 100644
--- a/test.py
+++ b/test.py
@@ -1,7 +1,6 @@
 import argparse
 import json
 
-import yaml
 from torch.utils.data import DataLoader
 
 from utils.datasets import *
@@ -40,8 +39,9 @@ def test(data,
         if half:
             model.half()  # to FP16
 
-        if device.type != 'cpu' and torch.cuda.device_count() > 1:
-            model = nn.DataParallel(model)
+        # Multi-GPU disabled, incompatible with .half()
+        # if device.type != 'cpu' and torch.cuda.device_count() > 1:
+        #     model = nn.DataParallel(model)
 
     else:  # called by train.py
         training = True

From afe1df385b5ef37c3f75a49ceda94fdc448bf490 Mon Sep 17 00:00:00 2001
From: Glenn Jocher <glenn.jocher@ultralytics.com>
Date: Tue, 16 Jun 2020 15:08:14 -0700
Subject: [PATCH 10/73] dist.destroy_process_group() bug fix

---
 train.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/train.py b/train.py
index 484190a..e9be7e4 100644
--- a/train.py
+++ b/train.py
@@ -356,7 +356,7 @@ def train(hyp):
     if not opt.evolve:
         plot_results()  # save as results.png
     print('%g epochs completed in %.3f hours.\n' % (epoch - start_epoch + 1, (time.time() - t0) / 3600))
-    dist.destroy_process_group() if torch.cuda.device_count() > 1 else None
+    dist.destroy_process_group() if device.type != 'cpu' and torch.cuda.device_count() > 1 else None
     torch.cuda.empty_cache()
     return results
 

From 4052603e1fdb3e95b54cdec323b40ba22b2e359a Mon Sep 17 00:00:00 2001
From: Glenn Jocher <glenn.jocher@ultralytics.com>
Date: Tue, 16 Jun 2020 15:54:20 -0700
Subject: [PATCH 11/73] AutoAnchor update - improvement check

---
 utils/utils.py | 45 ++++++++++++++++++++++++++++++---------------
 1 file changed, 30 insertions(+), 15 deletions(-)

diff --git a/utils/utils.py b/utils/utils.py
index dd12201..cc4369a 100755
--- a/utils/utils.py
+++ b/utils/utils.py
@@ -58,17 +58,24 @@ def check_anchors(dataset, model, thr=4.0, imgsz=640):
     anchors = model.module.model[-1].anchor_grid if hasattr(model, 'module') else model.model[-1].anchor_grid
     shapes = imgsz * dataset.shapes / dataset.shapes.max(1, keepdims=True)
     wh = torch.tensor(np.concatenate([l[:, 3:5] * s for s, l in zip(shapes, dataset.labels)])).float()  # wh
-    ratio = wh[:, None] / anchors.view(-1, 2).cpu()[None]  # ratio
-    m = torch.max(ratio, 1. / ratio).max(2)[0]  # max ratio
-    bpr = (m.min(1)[0] < thr).float().mean()  # best possible recall
-    # mr = (m < thr).float().mean()  # match ratio
 
+    def metric(k):  # compute metric
+        r = wh[:, None] / k[None]
+        x = torch.min(r, 1. / r).min(2)[0]  # ratio metric
+        best = x.max(1)[0]  # best_x
+        return (best > 1. / thr).float().mean()  #  best possible recall
+
+    bpr = metric(anchors.clone().cpu().view(-1, 2))
     print('Best Possible Recall (BPR) = %.3f' % bpr, end='')
     if bpr < 0.99:  # threshold to recompute
-        print('. Generating new anchors for improved recall, please wait...' % bpr)
+        print('. Attempting to generate improved anchors, please wait...' % bpr)
         new_anchors = kmean_anchors(dataset, n=9, img_size=640, thr=4.0, gen=1000, verbose=False)
-        anchors[:] = torch.tensor(new_anchors).view_as(anchors).type_as(anchors)
-        print('New anchors saved to model. Update model *.yaml to use these anchors in the future.')
+        new_bpr = metric(new_anchors.reshape(-1, 2))
+        if new_bpr > bpr:
+            anchors[:] = torch.tensor(new_anchors).view_as(anchors).type_as(anchors)
+            print('New anchors saved to model. Update model *.yaml to use these anchors in the future.')
+        else:
+            print('Original anchors better than new anchors. Proceeding with original anchors.')
     print('')  # newline
 
 
@@ -712,19 +719,19 @@ def kmean_anchors(path='./data/coco128.yaml', n=9, img_size=640, thr=4.0, gen=10
     """
     thr = 1. / thr
 
-    def metric(k):  # compute metrics
+    def metric(k, wh):  # compute metrics
         r = wh[:, None] / k[None]
         x = torch.min(r, 1. / r).min(2)[0]  # ratio metric
         # x = wh_iou(wh, torch.tensor(k))  # iou metric
         return x, x.max(1)[0]  # x, best_x
 
     def fitness(k):  # mutation fitness
-        _, best = metric(torch.tensor(k, dtype=torch.float32))
+        _, best = metric(torch.tensor(k, dtype=torch.float32), wh)
         return (best * (best > thr).float()).mean()  # fitness
 
     def print_results(k):
         k = k[np.argsort(k.prod(1))]  # sort small to large
-        x, best = metric(k)
+        x, best = metric(k, wh0)
         bpr, aat = (best > thr).float().mean(), (x > thr).float().mean() * n  # best possible recall, anch > thr
         print('thr=%.2f: %.3f best possible recall, %.2f anchors past thr' % (thr, bpr, aat))
         print('n=%g, img_size=%s, metric_all=%.3f/%.3f-mean/best, past_thr=%.3f-mean: ' %
@@ -743,8 +750,14 @@ def kmean_anchors(path='./data/coco128.yaml', n=9, img_size=640, thr=4.0, gen=10
 
     # Get label wh
     shapes = img_size * dataset.shapes / dataset.shapes.max(1, keepdims=True)
-    wh = np.concatenate([l[:, 3:5] * s for s, l in zip(shapes, dataset.labels)])  # wh
-    wh = wh[(wh > 2.0).all(1)]  # filter > 2 pixels
+    wh0 = np.concatenate([l[:, 3:5] * s for s, l in zip(shapes, dataset.labels)])  # wh
+
+    # Filter
+    i = (wh0 < 4.0).any(1).sum()
+    if i:
+        print('WARNING: Extremely small objects found. '
+              '%g of %g labels are < 4 pixels in width or height.' % (i, len(wh0)))
+    wh = wh0[(wh0 >= 4.0).any(1)]  # filter > 2 pixels
 
     # Kmeans calculation
     from scipy.cluster.vq import kmeans
@@ -752,7 +765,8 @@ def kmean_anchors(path='./data/coco128.yaml', n=9, img_size=640, thr=4.0, gen=10
     s = wh.std(0)  # sigmas for whitening
     k, dist = kmeans(wh / s, n, iter=30)  # points, mean distance
     k *= s
-    wh = torch.tensor(wh, dtype=torch.float32)
+    wh = torch.tensor(wh, dtype=torch.float32)  # filtered
+    wh0 = torch.tensor(wh0, dtype=torch.float32)  # unflitered
     k = print_results(k)
 
     # Plot
@@ -781,8 +795,8 @@ def kmean_anchors(path='./data/coco128.yaml', n=9, img_size=640, thr=4.0, gen=10
             f, k = fg, kg.copy()
             if verbose:
                 print_results(k)
-    k = print_results(k)
-    return k
+
+    return print_results(k)
 
 
 def print_mutation(hyp, results, bucket=''):
@@ -1099,6 +1113,7 @@ def plot_labels(labels):
     ax[2].set_xlabel('width')
     ax[2].set_ylabel('height')
     plt.savefig('labels.png', dpi=200)
+    plt.close()
 
 
 def plot_evolution_results(hyp):  # from utils.utils import *; plot_evolution_results(hyp)

From 8db51c70025fd28ee7bd7ca60b41f183bab60c0e Mon Sep 17 00:00:00 2001
From: Glenn Jocher <glenn.jocher@ultralytics.com>
Date: Tue, 16 Jun 2020 16:05:28 -0700
Subject: [PATCH 12/73] tb_writer bug fix

---
 train.py | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)

diff --git a/train.py b/train.py
index e9be7e4..7df99c0 100644
--- a/train.py
+++ b/train.py
@@ -195,8 +195,9 @@ def train(hyp):
     c = torch.tensor(labels[:, 0])  # classes
     # cf = torch.bincount(c.long(), minlength=nc) + 1.
     # model._initialize_biases(cf.to(device))
-    plot_labels(labels)
-    tb_writer.add_histogram('classes', c, 0)
+    if tb_writer:
+        plot_labels(labels)
+        tb_writer.add_histogram('classes', c, 0)
 
     # Check anchors
     check_anchors(dataset, model=model, thr=hyp['anchor_t'], imgsz=imgsz)

From 1dfc28527ffa36e5f12e82e9956265761c75f480 Mon Sep 17 00:00:00 2001
From: Glenn Jocher <glenn.jocher@ultralytics.com>
Date: Tue, 16 Jun 2020 16:19:58 -0700
Subject: [PATCH 13/73] AutoAnchor BPR to 4 significant figures

---
 utils/utils.py | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/utils/utils.py b/utils/utils.py
index cc4369a..530cffd 100755
--- a/utils/utils.py
+++ b/utils/utils.py
@@ -66,7 +66,7 @@ def check_anchors(dataset, model, thr=4.0, imgsz=640):
         return (best > 1. / thr).float().mean()  #  best possible recall
 
     bpr = metric(anchors.clone().cpu().view(-1, 2))
-    print('Best Possible Recall (BPR) = %.3f' % bpr, end='')
+    print('Best Possible Recall (BPR) = %.4f' % bpr, end='')
     if bpr < 0.99:  # threshold to recompute
         print('. Attempting to generate improved anchors, please wait...' % bpr)
         new_anchors = kmean_anchors(dataset, n=9, img_size=640, thr=4.0, gen=1000, verbose=False)
@@ -733,7 +733,7 @@ def kmean_anchors(path='./data/coco128.yaml', n=9, img_size=640, thr=4.0, gen=10
         k = k[np.argsort(k.prod(1))]  # sort small to large
         x, best = metric(k, wh0)
         bpr, aat = (best > thr).float().mean(), (x > thr).float().mean() * n  # best possible recall, anch > thr
-        print('thr=%.2f: %.3f best possible recall, %.2f anchors past thr' % (thr, bpr, aat))
+        print('thr=%.2f: %.4f best possible recall, %.2f anchors past thr' % (thr, bpr, aat))
         print('n=%g, img_size=%s, metric_all=%.3f/%.3f-mean/best, past_thr=%.3f-mean: ' %
               (n, img_size, x.mean(), best.mean(), x[x > thr].mean()), end='')
         for i, x in enumerate(k):

From d1e57166109428f3462e848d18db613a681f429f Mon Sep 17 00:00:00 2001
From: Glenn Jocher <glenn.jocher@ultralytics.com>
Date: Tue, 16 Jun 2020 16:22:58 -0700
Subject: [PATCH 14/73] train with multi-gpu half test bug fix #99

---
 test.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/test.py b/test.py
index aa5cde3..df8ed67 100644
--- a/test.py
+++ b/test.py
@@ -46,7 +46,7 @@ def test(data,
     else:  # called by train.py
         training = True
         device = next(model.parameters()).device  # get model device
-        half = device.type != 'cpu'  # half precision only supported on CUDA
+        half = device.type != 'cpu' and torch.cuda.device_count() == 1  # half precision only supported on single-GPU
         if half:
             model.half()  # to FP16
 

From 2daa412575dd81615042b918d02964e1971a605f Mon Sep 17 00:00:00 2001
From: Glenn Jocher <glenn.jocher@ultralytics.com>
Date: Tue, 16 Jun 2020 16:28:25 -0700
Subject: [PATCH 15/73] train with multi-gpu half test bug fix #99

---
 test.py | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/test.py b/test.py
index df8ed67..da803b7 100644
--- a/test.py
+++ b/test.py
@@ -46,7 +46,8 @@ def test(data,
     else:  # called by train.py
         training = True
         device = next(model.parameters()).device  # get model device
-        half = device.type != 'cpu' and torch.cuda.device_count() == 1  # half precision only supported on single-GPU
+        # half disabled https://github.com/ultralytics/yolov5/issues/99
+        half = False  # device.type != 'cpu' and torch.cuda.device_count() == 1
         if half:
             model.half()  # to FP16
 

From ef75f7313d5ce72e248d7581f749c3566a5fcaf3 Mon Sep 17 00:00:00 2001
From: Glenn Jocher <glenn.jocher@ultralytics.com>
Date: Tue, 16 Jun 2020 16:48:01 -0700
Subject: [PATCH 16/73] AutoAnchor n, imgsz bug fix

---
 utils/utils.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/utils/utils.py b/utils/utils.py
index 530cffd..73d3d7f 100755
--- a/utils/utils.py
+++ b/utils/utils.py
@@ -69,7 +69,7 @@ def check_anchors(dataset, model, thr=4.0, imgsz=640):
     print('Best Possible Recall (BPR) = %.4f' % bpr, end='')
     if bpr < 0.99:  # threshold to recompute
         print('. Attempting to generate improved anchors, please wait...' % bpr)
-        new_anchors = kmean_anchors(dataset, n=9, img_size=640, thr=4.0, gen=1000, verbose=False)
+        new_anchors = kmean_anchors(dataset, n=anchors.numel() // 2, img_size=imgsz, thr=thr, gen=1000, verbose=False)
         new_bpr = metric(new_anchors.reshape(-1, 2))
         if new_bpr > bpr:
             anchors[:] = torch.tensor(new_anchors).view_as(anchors).type_as(anchors)

From 034609414e8b48f5dc324a5ec7729a78b8638cf4 Mon Sep 17 00:00:00 2001
From: Glenn Jocher <glenn.jocher@ultralytics.com>
Date: Tue, 16 Jun 2020 17:08:45 -0700
Subject: [PATCH 17/73] AutoAnchor fitness to screen

---
 utils/utils.py | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/utils/utils.py b/utils/utils.py
index 73d3d7f..40a9f81 100755
--- a/utils/utils.py
+++ b/utils/utils.py
@@ -785,7 +785,8 @@ def kmean_anchors(path='./data/coco128.yaml', n=9, img_size=640, thr=4.0, gen=10
     # Evolve
     npr = np.random
     f, sh, mp, s = fitness(k), k.shape, 0.9, 0.1  # fitness, generations, mutation prob, sigma
-    for _ in tqdm(range(gen), desc='Evolving anchors with Genetic Algorithm'):
+    pbar = tqdm(range(gen), desc='Evolving anchors with Genetic Algorithm')  # progress bar
+    for _ in pbar:
         v = np.ones(sh)
         while (v == 1).all():  # mutate until a change occurs (prevent duplicates)
             v = ((npr.random(sh) < mp) * npr.random() * npr.randn(*sh) * s + 1).clip(0.3, 3.0)
@@ -793,6 +794,7 @@ def kmean_anchors(path='./data/coco128.yaml', n=9, img_size=640, thr=4.0, gen=10
         fg = fitness(kg)
         if fg > f:
             f, k = fg, kg.copy()
+            pbar.desc = 'Evolving anchors with Genetic Algorithm: fitness = %.4f' % f
             if verbose:
                 print_results(k)
 

From a557b7d924c605d573f4f0e64c1bc3702276d9e8 Mon Sep 17 00:00:00 2001
From: Glenn Jocher <glenn.jocher@ultralytics.com>
Date: Tue, 16 Jun 2020 17:48:01 -0700
Subject: [PATCH 18/73] check_img_size() update

---
 utils/utils.py | 7 ++++---
 1 file changed, 4 insertions(+), 3 deletions(-)

diff --git a/utils/utils.py b/utils/utils.py
index 40a9f81..02aa6b6 100755
--- a/utils/utils.py
+++ b/utils/utils.py
@@ -47,9 +47,10 @@ def check_git_status():
 
 def check_img_size(img_size, s=32):
     # Verify img_size is a multiple of stride s
-    if img_size % s != 0:
-        print('WARNING: --img-size %g must be multiple of max stride %g' % (img_size, s))
-    return make_divisible(img_size, s)  # nearest gs-multiple
+    new_size = make_divisible(img_size, s)  # ceil gs-multiple
+    if new_size != img_size:
+        print('WARNING: --img-size %g must be multiple of max stride %g, updating to %g' % (img_size, s, new_size))
+    return new_size
 
 
 def check_anchors(dataset, model, thr=4.0, imgsz=640):

From 2368603484650e3a18e6c40ec7374beae7bd3391 Mon Sep 17 00:00:00 2001
From: Lornatang <liuchangyu1111@gmail.com>
Date: Wed, 17 Jun 2020 09:56:26 +0800
Subject: [PATCH 19/73] fix refrence bug

In torch==1.5, the import of the API has changed. Although it does not interrupt the operation of the program, it seems to me to be an implicit error and may throw an exception in later versions.
---
 detect.py | 4 +++-
 train.py  | 1 +
 2 files changed, 4 insertions(+), 1 deletion(-)

diff --git a/detect.py b/detect.py
index bf858e4..bc4ce62 100644
--- a/detect.py
+++ b/detect.py
@@ -1,5 +1,7 @@
 import argparse
 
+import torch.backends.cudnn as cudnn
+
 from utils.datasets import *
 from utils.utils import *
 
@@ -36,7 +38,7 @@ def detect(save_img=False):
     vid_path, vid_writer = None, None
     if webcam:
         view_img = True
-        torch.backends.cudnn.benchmark = True  # set True to speed up constant image size inference
+        cudnn.benchmark = True  # set True to speed up constant image size inference
         dataset = LoadStreams(source, img_size=imgsz)
     else:
         save_img = True
diff --git a/train.py b/train.py
index 7df99c0..6c559c1 100644
--- a/train.py
+++ b/train.py
@@ -4,6 +4,7 @@ import torch.distributed as dist
 import torch.nn.functional as F
 import torch.optim as optim
 import torch.optim.lr_scheduler as lr_scheduler
+import torch.utils.data
 from torch.utils.tensorboard import SummaryWriter
 
 import test  # import test.py to get mAP after each epoch

From 0071b699948874abd1032de290815aa931f61153 Mon Sep 17 00:00:00 2001
From: Prashant Rai <59931502+imprashr@users.noreply.github.com>
Date: Wed, 17 Jun 2020 14:15:27 +0530
Subject: [PATCH 20/73] onnx should be 1.6.0

for smoothly export model to onnx, onnx should be 1.6.0
---
 requirements.txt | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/requirements.txt b/requirements.txt
index 30660e9..a0130d2 100755
--- a/requirements.txt
+++ b/requirements.txt
@@ -21,4 +21,4 @@ git+https://github.com/cocodataset/cocoapi.git#subdirectory=PythonAPI
 # conda install -yc conda-forge scikit-image pycocotools tensorboard
 # conda install -yc spyder-ide spyder-line-profiler
 # conda install -yc pytorch pytorch torchvision
-# conda install -yc conda-forge protobuf numpy && pip install onnx  # https://github.com/onnx/onnx#linux-and-macos
+# conda install -yc conda-forge protobuf numpy && pip install onnx==1.6.0  # https://github.com/onnx/onnx#linux-and-macos

From b7ac446d62d275fa8f4377a4ba65988eb8e55a19 Mon Sep 17 00:00:00 2001
From: Glenn Jocher <glenn.jocher@ultralytics.com>
Date: Wed, 17 Jun 2020 12:56:47 -0700
Subject: [PATCH 21/73] update README.md

---
 README.md | 7 +++++--
 1 file changed, 5 insertions(+), 2 deletions(-)

diff --git a/README.md b/README.md
index 7e39727..35fbee6 100755
--- a/README.md
+++ b/README.md
@@ -74,9 +74,12 @@ Results saved to /content/yolov5/inference/output
 
 ## Reproduce Our Training
 
-Run command below. Training times for yolov5s/m/l/x are 2/4/6/8 days on a single V100 (multi-GPU times faster).
+Download [COCO](https://github.com/ultralytics/yolov5/blob/master/data/get_coco2017.sh), install [Apex](https://github.com/NVIDIA/apex) and run command below. Training times for yolov5s/m/l/x are 2/4/6/8 days on a single V100 (multi-GPU times faster). Use the largest `--batch-size` your GPU allows (batch sizes shown for 16 GB devices).
 ```bash
-$ python train.py --data coco.yaml --cfg yolov5s.yaml --weights '' --batch-size 16                                      
+$ python train.py --data coco.yaml --cfg yolov5s.yaml --weights '' --batch-size 64
+                                         yolov5m                                48
+                                         yolov5l                                32
+                                         yolov5x                                16
 ```
 <img src="https://user-images.githubusercontent.com/26833433/84186698-c4d54d00-aa45-11ea-9bde-c632c1230ccd.png" width="900">
 

From 8b6f5826bc9bfb16686fc693c9b3b4aeaedfc9d0 Mon Sep 17 00:00:00 2001
From: Glenn Jocher <glenn.jocher@ultralytics.com>
Date: Wed, 17 Jun 2020 12:58:01 -0700
Subject: [PATCH 22/73] update README.md

---
 README.md | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/README.md b/README.md
index 35fbee6..fbeaefa 100755
--- a/README.md
+++ b/README.md
@@ -74,7 +74,7 @@ Results saved to /content/yolov5/inference/output
 
 ## Reproduce Our Training
 
-Download [COCO](https://github.com/ultralytics/yolov5/blob/master/data/get_coco2017.sh), install [Apex](https://github.com/NVIDIA/apex) and run command below. Training times for yolov5s/m/l/x are 2/4/6/8 days on a single V100 (multi-GPU times faster). Use the largest `--batch-size` your GPU allows (batch sizes shown for 16 GB devices).
+Download [COCO](https://github.com/ultralytics/yolov5/blob/master/data/get_coco2017.sh), install [Apex](https://github.com/NVIDIA/apex) and run command below. Training times for YOLOv5s/m/l/x are 2/4/6/8 days on a single V100 (multi-GPU times faster). Use the largest `--batch-size` your GPU allows (batch sizes shown for 16 GB devices).
 ```bash
 $ python train.py --data coco.yaml --cfg yolov5s.yaml --weights '' --batch-size 64
                                          yolov5m                                48

From 9fdb0fbacf65ad6d8778da592ef9b87038a3f4c7 Mon Sep 17 00:00:00 2001
From: Glenn Jocher <glenn.jocher@ultralytics.com>
Date: Wed, 17 Jun 2020 19:51:15 -0700
Subject: [PATCH 23/73] AutoAnchor bug fix # 117

---
 train.py       |  4 +++-
 utils/utils.py | 13 ++++++++-----
 2 files changed, 11 insertions(+), 6 deletions(-)

diff --git a/train.py b/train.py
index 7df99c0..5163bb1 100644
--- a/train.py
+++ b/train.py
@@ -200,7 +200,8 @@ def train(hyp):
         tb_writer.add_histogram('classes', c, 0)
 
     # Check anchors
-    check_anchors(dataset, model=model, thr=hyp['anchor_t'], imgsz=imgsz)
+    if not opt.noautoanchor:
+        check_anchors(dataset, model=model, thr=hyp['anchor_t'], imgsz=imgsz)
 
     # Exponential moving average
     ema = torch_utils.ModelEMA(model)
@@ -374,6 +375,7 @@ if __name__ == '__main__':
     parser.add_argument('--resume', action='store_true', help='resume training from last.pt')
     parser.add_argument('--nosave', action='store_true', help='only save final checkpoint')
     parser.add_argument('--notest', action='store_true', help='only test final epoch')
+    parser.add_argument('--noautoanchor', action='store_true', help='disable autoanchor check')
     parser.add_argument('--evolve', action='store_true', help='evolve hyperparameters')
     parser.add_argument('--bucket', type=str, default='', help='gsutil bucket')
     parser.add_argument('--cache-images', action='store_true', help='cache images for faster training')
diff --git a/utils/utils.py b/utils/utils.py
index 02aa6b6..47f5219 100755
--- a/utils/utils.py
+++ b/utils/utils.py
@@ -56,7 +56,7 @@ def check_img_size(img_size, s=32):
 def check_anchors(dataset, model, thr=4.0, imgsz=640):
     # Check anchor fit to data, recompute if necessary
     print('\nAnalyzing anchors... ', end='')
-    anchors = model.module.model[-1].anchor_grid if hasattr(model, 'module') else model.model[-1].anchor_grid
+    m = model.module.model[-1] if hasattr(model, 'module') else model.model[-1]  # Detect()
     shapes = imgsz * dataset.shapes / dataset.shapes.max(1, keepdims=True)
     wh = torch.tensor(np.concatenate([l[:, 3:5] * s for s, l in zip(shapes, dataset.labels)])).float()  # wh
 
@@ -66,14 +66,17 @@ def check_anchors(dataset, model, thr=4.0, imgsz=640):
         best = x.max(1)[0]  # best_x
         return (best > 1. / thr).float().mean()  #  best possible recall
 
-    bpr = metric(anchors.clone().cpu().view(-1, 2))
+    bpr = metric(m.anchor_grid.clone().cpu().view(-1, 2))
     print('Best Possible Recall (BPR) = %.4f' % bpr, end='')
     if bpr < 0.99:  # threshold to recompute
         print('. Attempting to generate improved anchors, please wait...' % bpr)
-        new_anchors = kmean_anchors(dataset, n=anchors.numel() // 2, img_size=imgsz, thr=thr, gen=1000, verbose=False)
+        na = m.anchor_grid.numel() // 2  # number of anchors
+        new_anchors = kmean_anchors(dataset, n=na, img_size=imgsz, thr=thr, gen=1000, verbose=False)
         new_bpr = metric(new_anchors.reshape(-1, 2))
-        if new_bpr > bpr:
-            anchors[:] = torch.tensor(new_anchors).view_as(anchors).type_as(anchors)
+        if new_bpr > bpr:  # replace anchors
+            new_anchors = torch.tensor(new_anchors, device=m.anchors.device).type_as(m.anchors)
+            m.anchor_grid[:] = new_anchors.clone().view_as(m.anchor_grid)  # for inference
+            m.anchors[:] = new_anchors.clone().view_as(m.anchors) / m.stride.to(m.anchors.device).view(-1, 1, 1)  # loss
             print('New anchors saved to model. Update model *.yaml to use these anchors in the future.')
         else:
             print('Original anchors better than new anchors. Proceeding with original anchors.')

From d9b64c27c24db2001535bb480959aca015159510 Mon Sep 17 00:00:00 2001
From: Glenn Jocher <glenn.jocher@ultralytics.com>
Date: Wed, 17 Jun 2020 22:34:13 -0700
Subject: [PATCH 24/73] save ckpt in FP16 #119

---
 train.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/train.py b/train.py
index 5163bb1..ce46797 100644
--- a/train.py
+++ b/train.py
@@ -332,7 +332,7 @@ def train(hyp):
                 ckpt = {'epoch': epoch,
                         'best_fitness': best_fitness,
                         'training_results': f.read(),
-                        'model': ema.ema.module if hasattr(model, 'module') else ema.ema,
+                        'model': ema.ema.module.half() if hasattr(model, 'module') else ema.ema.half(),
                         'optimizer': None if final_epoch else optimizer.state_dict()}
 
             # Save last, best and delete

From cce95e744dad1d8cb487c34b1e641136063e77f3 Mon Sep 17 00:00:00 2001
From: Glenn Jocher <glenn.jocher@ultralytics.com>
Date: Thu, 18 Jun 2020 00:13:18 -0700
Subject: [PATCH 25/73] backbone as FP16, save default to FP32

---
 train.py       | 2 +-
 utils/utils.py | 3 +--
 2 files changed, 2 insertions(+), 3 deletions(-)

diff --git a/train.py b/train.py
index ce46797..5163bb1 100644
--- a/train.py
+++ b/train.py
@@ -332,7 +332,7 @@ def train(hyp):
                 ckpt = {'epoch': epoch,
                         'best_fitness': best_fitness,
                         'training_results': f.read(),
-                        'model': ema.ema.module.half() if hasattr(model, 'module') else ema.ema.half(),
+                        'model': ema.ema.module if hasattr(model, 'module') else ema.ema,
                         'optimizer': None if final_epoch else optimizer.state_dict()}
 
             # Save last, best and delete
diff --git a/utils/utils.py b/utils/utils.py
index 47f5219..9c3d7d2 100755
--- a/utils/utils.py
+++ b/utils/utils.py
@@ -627,13 +627,12 @@ def strip_optimizer(f='weights/best.pt'):  # from utils.utils import *; strip_op
 def create_backbone(f='weights/best.pt', s='weights/backbone.pt'):  # from utils.utils import *; create_backbone()
     # create backbone 's' from 'f'
     device = torch.device('cpu')
-    x = torch.load(f, map_location=device)
-    torch.save(x, s)  # update model if SourceChangeWarning
     x = torch.load(s, map_location=device)
 
     x['optimizer'] = None
     x['training_results'] = None
     x['epoch'] = -1
+    x['model'].half()  # to FP16
     for p in x['model'].parameters():
         p.requires_grad = True
     torch.save(x, s)

From 10145833c3a86add544f0291deaf860b61d43ea8 Mon Sep 17 00:00:00 2001
From: Glenn Jocher <glenn.jocher@ultralytics.com>
Date: Thu, 18 Jun 2020 00:16:47 -0700
Subject: [PATCH 26/73] strip_optimizer() default to save FP16 model

---
 utils/utils.py | 1 +
 1 file changed, 1 insertion(+)

diff --git a/utils/utils.py b/utils/utils.py
index 9c3d7d2..6bde407 100755
--- a/utils/utils.py
+++ b/utils/utils.py
@@ -620,6 +620,7 @@ def strip_optimizer(f='weights/best.pt'):  # from utils.utils import *; strip_op
     # Strip optimizer from *.pt files for lighter files (reduced by 1/2 size)
     x = torch.load(f, map_location=torch.device('cpu'))
     x['optimizer'] = None
+    x['model'].half()  # to FP16
     torch.save(x, f)
     print('Optimizer stripped from %s' % f)
 

From a3ee0a895498a597e3ebc0da8c14d48f66769485 Mon Sep 17 00:00:00 2001
From: Glenn Jocher <glenn.jocher@ultralytics.com>
Date: Thu, 18 Jun 2020 11:46:04 -0700
Subject: [PATCH 27/73] check_img_size() update

---
 .github/ISSUE_TEMPLATE/--bug-report.md | 34 ++++++++++++++++++--------
 1 file changed, 24 insertions(+), 10 deletions(-)

diff --git a/.github/ISSUE_TEMPLATE/--bug-report.md b/.github/ISSUE_TEMPLATE/--bug-report.md
index 500b606..44f25aa 100644
--- a/.github/ISSUE_TEMPLATE/--bug-report.md
+++ b/.github/ISSUE_TEMPLATE/--bug-report.md
@@ -7,29 +7,43 @@ assignees: ''
 
 ---
 
-Before submitting a bug report, please ensure that you are using the latest versions of:
- - Python
- - PyTorch
- - This repository (run `git fetch && git status -uno` to check and `git pull` to update)
+Before submitting a bug report, please be aware that your issue **must be reproducible** with all of the following, otherwise it is non-actionable, and we can not help you:
+ - **Current repository**: run `git fetch && git status -uno` to check and `git pull` to update your repo
+ - **Common dataset**: coco.yaml or coco128.yaml
+ - **Common environment**: Colab, Google Cloud, or Docker image. See https://github.com/ultralytics/yolov5#reproduce-our-environment
  
-**Your issue must be reproducible on a public dataset (i.e COCO) using the latest version of the repository, and you must supply code to reproduce, or we can not help you.**
-
-If this is a custom training question we suggest you include your `train*.jpg`, `test*.jpg` and `results.png` figures.
+If this is a custom dataset/training question you **must include** your `train*.jpg`, `test*.jpg` and `results.png` figures, or we can not help you. You can generate results.png with `utils.plot_results()`.
 
 
 ## 🐛 Bug
 A clear and concise description of what the bug is.
 
-## To Reproduce
-**REQUIRED**: Code to reproduce your issue below
+
+## To Reproduce (REQUIRED)
+
+Input:
+```
+import torch
+
+a = torch.tensor([5])
+c = a / 0
+```
+
+Output:
 ```
-python train.py ...
+Traceback (most recent call last):
+  File "/Users/glennjocher/opt/anaconda3/envs/env1/lib/python3.7/site-packages/IPython/core/interactiveshell.py", line 3331, in run_code
+    exec(code_obj, self.user_global_ns, self.user_ns)
+  File "<ipython-input-5-be04c762b799>", line 5, in <module>
+    c = a / 0
+RuntimeError: ZeroDivisionError
 ```
 
 
 ## Expected behavior
 A clear and concise description of what you expected to happen.
 
+
 ## Environment
 If applicable, add screenshots to help explain your problem.
 

From 245bd1d2424d141d810949e57b2c8cadbd7b5c96 Mon Sep 17 00:00:00 2001
From: Glenn Jocher <glenn.jocher@ultralytics.com>
Date: Thu, 18 Jun 2020 12:46:30 -0700
Subject: [PATCH 28/73] update README.md

---
 README.md | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/README.md b/README.md
index fbeaefa..12db50b 100755
--- a/README.md
+++ b/README.md
@@ -37,10 +37,10 @@ $ pip install -U -r requirements.txt
 
 ## Tutorials
 
-* <a href="https://colab.research.google.com/github/ultralytics/yolov5/blob/master/tutorial.ipynb"><img src="https://colab.research.google.com/assets/colab-badge.svg" alt="Open In Colab"></a>
+* [Notebook](https://github.com/ultralytics/yolov5/blob/master/tutorial.ipynb) <a href="https://colab.research.google.com/github/ultralytics/yolov5/blob/master/tutorial.ipynb"><img src="https://colab.research.google.com/assets/colab-badge.svg" alt="Open In Colab"></a>
 * [Train Custom Data](https://github.com/ultralytics/yolov5/wiki/Train-Custom-Data)
 * [Google Cloud Quickstart Guide](https://github.com/ultralytics/yolov5/wiki/GCP-Quickstart)
-* [Docker Quickstart Guide](https://github.com/ultralytics/yolov5/wiki/Docker-Quickstart) 
+* [Docker Quickstart Guide](https://github.com/ultralytics/yolov5/wiki/Docker-Quickstart) ![Docker Pulls](https://img.shields.io/docker/pulls/ultralytics/yolov5?logo=docker)
 
 
 ## Inference

From 80b82e8bdea882ead6e066564f9162b9935b9d9e Mon Sep 17 00:00:00 2001
From: Glenn Jocher <glenn.jocher@ultralytics.com>
Date: Thu, 18 Jun 2020 19:33:02 -0700
Subject: [PATCH 29/73] update yolo.py

---
 models/yolo.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/models/yolo.py b/models/yolo.py
index b433afb..9179c85 100644
--- a/models/yolo.py
+++ b/models/yolo.py
@@ -208,7 +208,7 @@ if __name__ == '__main__':
     parser.add_argument('--cfg', type=str, default='yolov5s.yaml', help='model.yaml')
     parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu')
     opt = parser.parse_args()
-    opt.cfg = glob.glob('./**/' + opt.cfg, recursive=True)[0]  # find file
+    opt.cfg = check_file(opt.cfg)  # check file
     device = torch_utils.select_device(opt.device)
 
     # Create model

From e249792c2f2e19e55f676185811693a29e805632 Mon Sep 17 00:00:00 2001
From: Glenn Jocher <glenn.jocher@ultralytics.com>
Date: Fri, 19 Jun 2020 10:15:08 -0700
Subject: [PATCH 30/73] update onnx_export.py usage

---
 models/onnx_export.py | 1 -
 1 file changed, 1 deletion(-)

diff --git a/models/onnx_export.py b/models/onnx_export.py
index 59d13b1..d17f8b2 100644
--- a/models/onnx_export.py
+++ b/models/onnx_export.py
@@ -1,7 +1,6 @@
 """Exports a pytorch *.pt model to *.onnx format
 
 Usage:
-    import torch
     $ export PYTHONPATH="$PWD" && python models/onnx_export.py --weights ./weights/yolov5s.pt --img 640 --batch 1
 """
 

From d4c6674c98e19df4c40e33a777610a18d1961145 Mon Sep 17 00:00:00 2001
From: Glenn Jocher <glenn.jocher@ultralytics.com>
Date: Fri, 19 Jun 2020 12:05:10 -0700
Subject: [PATCH 31/73] speeds updated from FP32 to FP16

---
 README.md | 16 ++++++++--------
 1 file changed, 8 insertions(+), 8 deletions(-)

diff --git a/README.md b/README.md
index 12db50b..abe7c11 100755
--- a/README.md
+++ b/README.md
@@ -4,7 +4,7 @@
 
 This repository represents Ultralytics open-source research into future object detection methods, and incorporates our lessons learned and best practices evolved over training thousands of models on custom client datasets with our previous YOLO repository https://github.com/ultralytics/yolov3. **All code and models are under active development, and are subject to modification or deletion without notice.** Use at your own risk.
 
-<img src="https://user-images.githubusercontent.com/26833433/84200349-729f2680-aa5b-11ea-8f9a-604c9e01a658.png" width="1000">** GPU Latency measures end-to-end latency per image averaged over 5000 COCO val2017 images using a V100 GPU with batch size 32, and includes image preprocessing, PyTorch FP32 inference, postprocessing and NMS.
+<img src="https://user-images.githubusercontent.com/26833433/84200349-729f2680-aa5b-11ea-8f9a-604c9e01a658.png" width="1000">** GPU Speed measures end-to-end time per image averaged over 5000 COCO val2017 images using a V100 GPU with batch size 32, and includes image preprocessing, PyTorch FP32 inference, postprocessing and NMS.
 
 - **June 9, 2020**: [CSP](https://github.com/WongKinYiu/CrossStagePartialNetworks) updates to all YOLOv5 models. New models are faster, smaller and more accurate. Credit to @WongKinYiu for his excellent work with CSP. 
 - **May 27, 2020**: Public release of repo. YOLOv5 models are SOTA among all known YOLO implementations, YOLOv5 family will be undergoing architecture research and development over Q2/Q3 2020 to increase performance. Updates may include [CSP](https://github.com/WongKinYiu/CrossStagePartialNetworks) bottlenecks, [YOLOv4](https://github.com/AlexeyAB/darknet) features, as well as PANet or BiFPN heads.
@@ -13,17 +13,17 @@ This repository represents Ultralytics open-source research into future object d
 
 ## Pretrained Checkpoints
 
-| Model | AP<sup>val</sup> | AP<sup>test</sup> | AP<sub>50</sub> | Latency<sub>GPU</sub> | FPS<sub>GPU</sub> || params | FLOPs |
+| Model | AP<sup>val</sup> | AP<sup>test</sup> | AP<sub>50</sub> | Speed<sub>GPU</sub> | FPS<sub>GPU</sub> || params | FLOPs |
 |---------- |------ |------ |------ | -------- | ------| ------ |------  |  :------: |
-| YOLOv5-s ([ckpt](https://drive.google.com/open?id=1Drs_Aiu7xx6S-ix95f9kNsA6ueKRpN2J))    | 35.5     | 35.5     | 55.0     | **2.5ms** | **400** || 7.1M   | 12.6B
-| YOLOv5-m ([ckpt](https://drive.google.com/open?id=1Drs_Aiu7xx6S-ix95f9kNsA6ueKRpN2J))    | 42.7     | 42.7     | 62.4     | 4.4ms     | 227     || 22.0M  | 39.0B
-| YOLOv5-l ([ckpt](https://drive.google.com/open?id=1Drs_Aiu7xx6S-ix95f9kNsA6ueKRpN2J))    | 45.7     | 45.9     | 65.1     | 6.8ms     | 147     || 50.3M  | 89.0B
-| YOLOv5-x ([ckpt](https://drive.google.com/open?id=1Drs_Aiu7xx6S-ix95f9kNsA6ueKRpN2J))    | **47.2** | **47.3** | **66.6** | 11.7ms    | 85      || 95.9M | 170.3B
-| YOLOv3-SPP ([ckpt](https://drive.google.com/open?id=1Drs_Aiu7xx6S-ix95f9kNsA6ueKRpN2J))  | 45.6     | 45.5     | 65.2     | 7.9ms     | 127     || 63.0M  | 118.0B
+| YOLOv5-s ([ckpt](https://drive.google.com/open?id=1Drs_Aiu7xx6S-ix95f9kNsA6ueKRpN2J))    | 35.5     | 35.5     | 55.0     | **2.1ms** | **476** || 7.1M   | 12.6B
+| YOLOv5-m ([ckpt](https://drive.google.com/open?id=1Drs_Aiu7xx6S-ix95f9kNsA6ueKRpN2J))    | 42.7     | 42.7     | 62.4     | 3.2ms     | 312     || 22.0M  | 39.0B
+| YOLOv5-l ([ckpt](https://drive.google.com/open?id=1Drs_Aiu7xx6S-ix95f9kNsA6ueKRpN2J))    | 45.7     | 45.9     | 65.1     | 4.1ms     | 243     || 50.3M  | 89.0B
+| YOLOv5-x ([ckpt](https://drive.google.com/open?id=1Drs_Aiu7xx6S-ix95f9kNsA6ueKRpN2J))    | **47.2** | **47.3** | **66.6** | 6.5ms     | 153     || 95.9M  | 170.3B
+| YOLOv3-SPP ([ckpt](https://drive.google.com/open?id=1Drs_Aiu7xx6S-ix95f9kNsA6ueKRpN2J))  | 45.6     | 45.5     | 65.2     | 4.8ms     | 208     || 63.0M  | 118.0B
 
 ** AP<sup>test</sup> denotes COCO [test-dev2017](http://cocodataset.org/#upload) server results, all other AP results in the table denote val2017 accuracy.  
 ** All AP numbers are for single-model single-scale without ensemble or test-time augmentation. Reproduce by `python test.py --img 736 --conf 0.001`  
-** Latency<sub>GPU</sub> measures end-to-end latency per image averaged over 5000 COCO val2017 images using a GCP [n1-standard-16](https://cloud.google.com/compute/docs/machine-types#n1_standard_machine_types) instance with one V100 GPU, and includes image preprocessing, PyTorch FP32 inference at batch size 32, postprocessing and NMS. Average NMS time included in this chart is 1-2ms/img.  Reproduce by `python test.py --img 640 --conf 0.1`  
+** Speed<sub>GPU</sub> measures end-to-end time per image averaged over 5000 COCO val2017 images using a GCP [n1-standard-16](https://cloud.google.com/compute/docs/machine-types#n1_standard_machine_types) instance with one V100 GPU, and includes image preprocessing, PyTorch FP16 image inference at --batch-size 32 --img-size 640, postprocessing and NMS. Average NMS time included in this chart is 1-2ms/img.  Reproduce by `python test.py --img 640 --conf 0.1`  
 ** All checkpoints are trained to 300 epochs with default settings and hyperparameters (no autoaugmentation). 
 
 

From 34ec87ecf9336acf572fb996bbdd739d12f93b99 Mon Sep 17 00:00:00 2001
From: Glenn Jocher <glenn.jocher@ultralytics.com>
Date: Fri, 19 Jun 2020 12:13:15 -0700
Subject: [PATCH 32/73] Update README.md

---
 README.md | 1 +
 1 file changed, 1 insertion(+)

diff --git a/README.md b/README.md
index abe7c11..c9b8bd9 100755
--- a/README.md
+++ b/README.md
@@ -6,6 +6,7 @@ This repository represents Ultralytics open-source research into future object d
 
 <img src="https://user-images.githubusercontent.com/26833433/84200349-729f2680-aa5b-11ea-8f9a-604c9e01a658.png" width="1000">** GPU Speed measures end-to-end time per image averaged over 5000 COCO val2017 images using a V100 GPU with batch size 32, and includes image preprocessing, PyTorch FP32 inference, postprocessing and NMS.
 
+- **June 9, 2020**: [FP16](https://pytorch.org/docs/stable/nn.html#torch.nn.Module.half) as new default for smaller checkpoints and faster inference. Comparison in [d4c6674](https://github.com/ultralytics/yolov5/commit/d4c6674c98e19df4c40e33a777610a18d1961145).
 - **June 9, 2020**: [CSP](https://github.com/WongKinYiu/CrossStagePartialNetworks) updates to all YOLOv5 models. New models are faster, smaller and more accurate. Credit to @WongKinYiu for his excellent work with CSP. 
 - **May 27, 2020**: Public release of repo. YOLOv5 models are SOTA among all known YOLO implementations, YOLOv5 family will be undergoing architecture research and development over Q2/Q3 2020 to increase performance. Updates may include [CSP](https://github.com/WongKinYiu/CrossStagePartialNetworks) bottlenecks, [YOLOv4](https://github.com/AlexeyAB/darknet) features, as well as PANet or BiFPN heads.
 - **April 1, 2020**: Begin development of a 100% PyTorch, scaleable YOLOv3/4-based group of future models, in a range of compound-scaled sizes. Models will be defined by new user-friendly `*.yaml` files. New training methods will be simpler to start, faster to finish, and more robust to training a wider variety of custom dataset.

From 458c4d0e0b1925b2d49200bf4d00515ec7684c02 Mon Sep 17 00:00:00 2001
From: Glenn Jocher <glenn.jocher@ultralytics.com>
Date: Fri, 19 Jun 2020 12:14:06 -0700
Subject: [PATCH 33/73] Update README.md

---
 README.md | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/README.md b/README.md
index c9b8bd9..591a858 100755
--- a/README.md
+++ b/README.md
@@ -6,7 +6,7 @@ This repository represents Ultralytics open-source research into future object d
 
 <img src="https://user-images.githubusercontent.com/26833433/84200349-729f2680-aa5b-11ea-8f9a-604c9e01a658.png" width="1000">** GPU Speed measures end-to-end time per image averaged over 5000 COCO val2017 images using a V100 GPU with batch size 32, and includes image preprocessing, PyTorch FP32 inference, postprocessing and NMS.
 
-- **June 9, 2020**: [FP16](https://pytorch.org/docs/stable/nn.html#torch.nn.Module.half) as new default for smaller checkpoints and faster inference. Comparison in [d4c6674](https://github.com/ultralytics/yolov5/commit/d4c6674c98e19df4c40e33a777610a18d1961145).
+- **June 19, 2020**: [FP16](https://pytorch.org/docs/stable/nn.html#torch.nn.Module.half) as new default for smaller checkpoints and faster inference. Comparison in [d4c6674](https://github.com/ultralytics/yolov5/commit/d4c6674c98e19df4c40e33a777610a18d1961145).
 - **June 9, 2020**: [CSP](https://github.com/WongKinYiu/CrossStagePartialNetworks) updates to all YOLOv5 models. New models are faster, smaller and more accurate. Credit to @WongKinYiu for his excellent work with CSP. 
 - **May 27, 2020**: Public release of repo. YOLOv5 models are SOTA among all known YOLO implementations, YOLOv5 family will be undergoing architecture research and development over Q2/Q3 2020 to increase performance. Updates may include [CSP](https://github.com/WongKinYiu/CrossStagePartialNetworks) bottlenecks, [YOLOv4](https://github.com/AlexeyAB/darknet) features, as well as PANet or BiFPN heads.
 - **April 1, 2020**: Begin development of a 100% PyTorch, scaleable YOLOv3/4-based group of future models, in a range of compound-scaled sizes. Models will be defined by new user-friendly `*.yaml` files. New training methods will be simpler to start, faster to finish, and more robust to training a wider variety of custom dataset.

From 0a09882e4f51d29c1739c18d409d5d1d5e9559f8 Mon Sep 17 00:00:00 2001
From: Glenn Jocher <glenn.jocher@ultralytics.com>
Date: Fri, 19 Jun 2020 16:46:21 -0700
Subject: [PATCH 34/73] Update issue templates

---
 .github/ISSUE_TEMPLATE/--bug-report.md | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/.github/ISSUE_TEMPLATE/--bug-report.md b/.github/ISSUE_TEMPLATE/--bug-report.md
index 44f25aa..615a9ca 100644
--- a/.github/ISSUE_TEMPLATE/--bug-report.md
+++ b/.github/ISSUE_TEMPLATE/--bug-report.md
@@ -8,11 +8,11 @@ assignees: ''
 ---
 
 Before submitting a bug report, please be aware that your issue **must be reproducible** with all of the following, otherwise it is non-actionable, and we can not help you:
- - **Current repository**: run `git fetch && git status -uno` to check and `git pull` to update your repo
+ - **Current repo**: run `git fetch && git status -uno` to check and `git pull` to update repo
  - **Common dataset**: coco.yaml or coco128.yaml
  - **Common environment**: Colab, Google Cloud, or Docker image. See https://github.com/ultralytics/yolov5#reproduce-our-environment
  
-If this is a custom dataset/training question you **must include** your `train*.jpg`, `test*.jpg` and `results.png` figures, or we can not help you. You can generate results.png with `utils.plot_results()`.
+If this is a custom dataset/training question you **must include** your `train*.jpg`, `test*.jpg` and `results.png` figures, or we can not help you. You can generate these with `utils.plot_results()`.
 
 
 ## 🐛 Bug

From 899f1d4bde374b0d98fe56c767cf1985d21cb72f Mon Sep 17 00:00:00 2001
From: Lornatang <liuchangyu1111@gmail.com>
Date: Sat, 20 Jun 2020 13:00:03 +0800
Subject: [PATCH 35/73] Fix DDP bug in single process multiple device use cases

---
 requirements.txt | 2 +-
 train.py         | 1 +
 2 files changed, 2 insertions(+), 1 deletion(-)

diff --git a/requirements.txt b/requirements.txt
index a0130d2..52846b8 100755
--- a/requirements.txt
+++ b/requirements.txt
@@ -2,7 +2,7 @@
 Cython
 numpy==1.17
 opencv-python
-torch>=1.5
+torch==1.4
 matplotlib
 pillow
 tensorboard
diff --git a/train.py b/train.py
index f545ae1..94ccf57 100644
--- a/train.py
+++ b/train.py
@@ -151,6 +151,7 @@ def train(hyp):
                                 world_size=1,  # number of nodes
                                 rank=0)  # node rank
         model = torch.nn.parallel.DistributedDataParallel(model)
+        # pip install torch==1.4.0+cu100 torchvision==0.5.0+cu100 -f https://download.pytorch.org/whl/torch_stable.html
 
     # Dataset
     dataset = LoadImagesAndLabels(train_path, imgsz, batch_size,

From 5c2cd711df6099aa56b57bbb62381862cfcf9f18 Mon Sep 17 00:00:00 2001
From: Glenn Jocher <glenn.jocher@ultralytics.com>
Date: Sat, 20 Jun 2020 09:26:07 -0700
Subject: [PATCH 36/73] vid_formats bug fix #146

---
 utils/datasets.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/utils/datasets.py b/utils/datasets.py
index f3a5a4a..331092d 100755
--- a/utils/datasets.py
+++ b/utils/datasets.py
@@ -18,7 +18,7 @@ from utils.utils import xyxy2xywh, xywh2xyxy
 
 help_url = 'https://github.com/ultralytics/yolov5/wiki/Train-Custom-Data'
 img_formats = ['.bmp', '.jpg', '.jpeg', '.png', '.tif', '.dng']
-vid_formats = ['.mov', '.avi', '.mp4']
+vid_formats = ['.mov', '.avi', '.mp4', '.mpg']
 
 # Get orientation exif tag
 for orientation in ExifTags.TAGS.keys():

From 024a42f4fd7969a45b70e50e6ae5206844e94167 Mon Sep 17 00:00:00 2001
From: Glenn Jocher <glenn.jocher@ultralytics.com>
Date: Sat, 20 Jun 2020 09:53:37 -0700
Subject: [PATCH 37/73] datasets.py vid_formats updated

---
 utils/datasets.py | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)

diff --git a/utils/datasets.py b/utils/datasets.py
index 331092d..65e642d 100755
--- a/utils/datasets.py
+++ b/utils/datasets.py
@@ -18,7 +18,7 @@ from utils.utils import xyxy2xywh, xywh2xyxy
 
 help_url = 'https://github.com/ultralytics/yolov5/wiki/Train-Custom-Data'
 img_formats = ['.bmp', '.jpg', '.jpeg', '.png', '.tif', '.dng']
-vid_formats = ['.mov', '.avi', '.mp4', '.mpg']
+vid_formats = ['.mov', '.avi', '.mp4', '.mpg', '.mpeg', '.m4v', '.wmv', '.mkv']
 
 # Get orientation exif tag
 for orientation in ExifTags.TAGS.keys():
@@ -63,7 +63,8 @@ class LoadImages:  # for inference
             self.new_video(videos[0])  # new video
         else:
             self.cap = None
-        assert self.nF > 0, 'No images or videos found in ' + path
+        assert self.nF > 0, 'No images or videos found in %s. Supported formats are:\nimages: %s\nvideos: %s' % \
+                            (path, img_formats, vid_formats)
 
     def __iter__(self):
         self.count = 0

From 24dd150fbd0614f933c6a5041e005b8a7f712771 Mon Sep 17 00:00:00 2001
From: Glenn Jocher <glenn.jocher@ultralytics.com>
Date: Sat, 20 Jun 2020 11:55:39 -0700
Subject: [PATCH 38/73] update torch>=1.4

---
 requirements.txt | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/requirements.txt b/requirements.txt
index 52846b8..1100495 100755
--- a/requirements.txt
+++ b/requirements.txt
@@ -2,7 +2,7 @@
 Cython
 numpy==1.17
 opencv-python
-torch==1.4
+torch>=1.4
 matplotlib
 pillow
 tensorboard

From 1f1917ef56433acc51ddcd942b71d840d2d4e6ef Mon Sep 17 00:00:00 2001
From: Glenn Jocher <glenn.jocher@ultralytics.com>
Date: Sun, 21 Jun 2020 13:37:11 -0700
Subject: [PATCH 39/73] remove fast, add merge

---
 test.py        | 7 ++++---
 train.py       | 3 +--
 utils/utils.py | 7 +------
 3 files changed, 6 insertions(+), 11 deletions(-)

diff --git a/test.py b/test.py
index da803b7..2381ab4 100644
--- a/test.py
+++ b/test.py
@@ -19,7 +19,7 @@ def test(data,
          verbose=False,
          model=None,
          dataloader=None,
-         fast=False):
+         merge=False):
     # Initialize/load model and set device
     if model is None:
         training = False
@@ -65,7 +65,7 @@ def test(data,
         img = torch.zeros((1, 3, imgsz, imgsz), device=device)  # init img
         _ = model(img.half() if half else img) if device.type != 'cpu' else None  # run once
 
-        fast |= conf_thres > 0.001  # enable fast mode
+        merge = opt.merge  # use Merge NMS
         path = data['test'] if opt.task == 'test' else data['val']  # path to val/test images
         dataset = LoadImagesAndLabels(path,
                                       imgsz,
@@ -109,7 +109,7 @@ def test(data,
 
             # Run NMS
             t = torch_utils.time_synchronized()
-            output = non_max_suppression(inf_out, conf_thres=conf_thres, iou_thres=iou_thres, fast=fast)
+            output = non_max_suppression(inf_out, conf_thres=conf_thres, iou_thres=iou_thres, merge=merge)
             t1 += torch_utils.time_synchronized() - t
 
         # Statistics per image
@@ -254,6 +254,7 @@ if __name__ == '__main__':
     parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu')
     parser.add_argument('--single-cls', action='store_true', help='treat as single-class dataset')
     parser.add_argument('--augment', action='store_true', help='augmented inference')
+    parser.add_argument('--merge', action='store_true', help='use Merge NMS')
     parser.add_argument('--verbose', action='store_true', help='report mAP by class')
     opt = parser.parse_args()
     opt.img_size = check_img_size(opt.img_size)
diff --git a/train.py b/train.py
index 94ccf57..921f0e3 100644
--- a/train.py
+++ b/train.py
@@ -305,8 +305,7 @@ def train(hyp):
                                              save_json=final_epoch and opt.data.endswith(os.sep + 'coco.yaml'),
                                              model=ema.ema,
                                              single_cls=opt.single_cls,
-                                             dataloader=testloader,
-                                             fast=epoch < epochs / 2)
+                                             dataloader=testloader)
 
         # Write
         with open(results_file, 'a') as f:
diff --git a/utils/utils.py b/utils/utils.py
index 6bde407..8457643 100755
--- a/utils/utils.py
+++ b/utils/utils.py
@@ -527,7 +527,7 @@ def build_targets(p, targets, model):
     return tcls, tbox, indices, anch
 
 
-def non_max_suppression(prediction, conf_thres=0.1, iou_thres=0.6, fast=False, classes=None, agnostic=False):
+def non_max_suppression(prediction, conf_thres=0.1, iou_thres=0.6, merge=False, classes=None, agnostic=False):
     """Performs Non-Maximum Suppression (NMS) on inference results
 
     Returns:
@@ -544,12 +544,7 @@ def non_max_suppression(prediction, conf_thres=0.1, iou_thres=0.6, fast=False, c
     max_det = 300  # maximum number of detections per image
     time_limit = 10.0  # seconds to quit after
     redundant = True  # require redundant detections
-    fast |= conf_thres > 0.001  # fast mode
     multi_label = nc > 1  # multiple labels per box (adds 0.5ms/img)
-    if fast:
-        merge = False
-    else:
-        merge = True  # merge for best mAP (adds 0.5ms/img)
 
     t = time.time()
     output = [None] * prediction.shape[0]

From fc171e2678ce303527eb2cffa000f6cd72ce5319 Mon Sep 17 00:00:00 2001
From: Glenn Jocher <glenn.jocher@ultralytics.com>
Date: Sun, 21 Jun 2020 14:08:59 -0700
Subject: [PATCH 40/73] check_anchor_order() update

---
 models/yolo.py |  3 ++-
 utils/utils.py | 14 +++++++++++++-
 2 files changed, 15 insertions(+), 2 deletions(-)

diff --git a/models/yolo.py b/models/yolo.py
index 9179c85..e491e64 100644
--- a/models/yolo.py
+++ b/models/yolo.py
@@ -61,8 +61,9 @@ class Model(nn.Module):
 
         # Build strides, anchors
         m = self.model[-1]  # Detect()
-        m.stride = torch.tensor([64 / x.shape[-2] for x in self.forward(torch.zeros(1, ch, 64, 64))])  # forward
+        m.stride = torch.tensor([128 / x.shape[-2] for x in self.forward(torch.zeros(1, ch, 128, 128))])  # forward
         m.anchors /= m.stride.view(-1, 1, 1)
+        check_anchor_order(m)
         self.stride = m.stride
 
         # Init weights, biases
diff --git a/utils/utils.py b/utils/utils.py
index 8457643..f1f5db5 100755
--- a/utils/utils.py
+++ b/utils/utils.py
@@ -58,7 +58,8 @@ def check_anchors(dataset, model, thr=4.0, imgsz=640):
     print('\nAnalyzing anchors... ', end='')
     m = model.module.model[-1] if hasattr(model, 'module') else model.model[-1]  # Detect()
     shapes = imgsz * dataset.shapes / dataset.shapes.max(1, keepdims=True)
-    wh = torch.tensor(np.concatenate([l[:, 3:5] * s for s, l in zip(shapes, dataset.labels)])).float()  # wh
+    scale = np.random.uniform(0.9, 1.1, size=(shapes.shape[0], 1))  # augment scale
+    wh = torch.tensor(np.concatenate([l[:, 3:5] * s for s, l in zip(shapes * scale, dataset.labels)])).float()  # wh
 
     def metric(k):  # compute metric
         r = wh[:, None] / k[None]
@@ -77,12 +78,23 @@ def check_anchors(dataset, model, thr=4.0, imgsz=640):
             new_anchors = torch.tensor(new_anchors, device=m.anchors.device).type_as(m.anchors)
             m.anchor_grid[:] = new_anchors.clone().view_as(m.anchor_grid)  # for inference
             m.anchors[:] = new_anchors.clone().view_as(m.anchors) / m.stride.to(m.anchors.device).view(-1, 1, 1)  # loss
+            check_anchor_order(m)
             print('New anchors saved to model. Update model *.yaml to use these anchors in the future.')
         else:
             print('Original anchors better than new anchors. Proceeding with original anchors.')
     print('')  # newline
 
 
+def check_anchor_order(m):
+    # Check anchor order against stride order for YOLOv5 Detect() module m, and correct if necessary
+    a = m.anchor_grid.prod(-1).view(-1)  # anchor area
+    da = a[-1] - a[0]  # delta a
+    ds = m.stride[-1] - m.stride[0]  # delta s
+    if da.sign() != ds.sign():  # same order
+        m.anchors[:] = m.anchors.flip(0)
+        m.anchor_grid[:] = m.anchor_grid.flip(0)
+
+
 def check_file(file):
     # Searches for file if not found locally
     if os.path.isfile(file):

From bbd12c7cfef914299db7ced156c6781f999af2b8 Mon Sep 17 00:00:00 2001
From: Glenn Jocher <glenn.jocher@ultralytics.com>
Date: Mon, 22 Jun 2020 10:14:58 -0700
Subject: [PATCH 41/73] detect.py fast bug fix

---
 detect.py | 3 +--
 1 file changed, 1 insertion(+), 2 deletions(-)

diff --git a/detect.py b/detect.py
index bc4ce62..567c9e6 100644
--- a/detect.py
+++ b/detect.py
@@ -64,8 +64,7 @@ def detect(save_img=False):
         pred = model(img, augment=opt.augment)[0]
 
         # Apply NMS
-        pred = non_max_suppression(pred, opt.conf_thres, opt.iou_thres,
-                                   fast=True, classes=opt.classes, agnostic=opt.agnostic_nms)
+        pred = non_max_suppression(pred, opt.conf_thres, opt.iou_thres, classes=opt.classes, agnostic=opt.agnostic_nms)
         t2 = torch_utils.time_synchronized()
 
         # Apply Classifier

From ef58dac9631da217abf72d0e39aff59909c1f1b9 Mon Sep 17 00:00:00 2001
From: Glenn Jocher <glenn.jocher@ultralytics.com>
Date: Mon, 22 Jun 2020 10:24:56 -0700
Subject: [PATCH 42/73] update Dockerfile

---
 Dockerfile | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/Dockerfile b/Dockerfile
index 1ce9671..d73affc 100644
--- a/Dockerfile
+++ b/Dockerfile
@@ -1,10 +1,6 @@
 # Start FROM Nvidia PyTorch image https://ngc.nvidia.com/catalog/containers/nvidia:pytorch
 FROM nvcr.io/nvidia/pytorch:20.03-py3
 
-# Install dependencies (pip or conda)
-RUN pip install -U gsutil
-# RUN pip install -U -r requirements.txt
-
 # Create working directory
 RUN mkdir -p /usr/src/app
 WORKDIR /usr/src/app
@@ -12,6 +8,10 @@ WORKDIR /usr/src/app
 # Copy contents
 COPY . /usr/src/app
 
+# Install dependencies (pip or conda)
+#RUN pip install -r requirements.txt
+RUN pip install -U gsutil
+
 # Copy weights
 #RUN python3 -c "from models import *; \
 #attempt_download('weights/yolov5s.pt'); \

From 364fcfd7dba53f46edd4f04c037a039c0a287972 Mon Sep 17 00:00:00 2001
From: Glenn Jocher <glenn.jocher@ultralytics.com>
Date: Mon, 22 Jun 2020 14:27:17 -0700
Subject: [PATCH 43/73] PANet update

---
 README.md              | 15 +++++++------
 models/yolov3-spp.yaml |  3 +--
 models/yolov5l.yaml    | 49 ++++++++++++++++++++++++------------------
 models/yolov5m.yaml    | 49 ++++++++++++++++++++++++------------------
 models/yolov5s.yaml    | 49 ++++++++++++++++++++++++------------------
 models/yolov5x.yaml    | 49 ++++++++++++++++++++++++------------------
 utils/utils.py         |  8 ++++---
 7 files changed, 126 insertions(+), 96 deletions(-)

diff --git a/README.md b/README.md
index 591a858..df49f67 100755
--- a/README.md
+++ b/README.md
@@ -4,7 +4,7 @@
 
 This repository represents Ultralytics open-source research into future object detection methods, and incorporates our lessons learned and best practices evolved over training thousands of models on custom client datasets with our previous YOLO repository https://github.com/ultralytics/yolov3. **All code and models are under active development, and are subject to modification or deletion without notice.** Use at your own risk.
 
-<img src="https://user-images.githubusercontent.com/26833433/84200349-729f2680-aa5b-11ea-8f9a-604c9e01a658.png" width="1000">** GPU Speed measures end-to-end time per image averaged over 5000 COCO val2017 images using a V100 GPU with batch size 32, and includes image preprocessing, PyTorch FP32 inference, postprocessing and NMS.
+<img src="https://user-images.githubusercontent.com/26833433/85336627-c6663280-b493-11ea-9b0a-289b0f182b84.png" width="1000">** GPU Speed measures end-to-end time per image averaged over 5000 COCO val2017 images using a V100 GPU with batch size 8, and includes image preprocessing, PyTorch FP16 inference, postprocessing and NMS.
 
 - **June 19, 2020**: [FP16](https://pytorch.org/docs/stable/nn.html#torch.nn.Module.half) as new default for smaller checkpoints and faster inference. Comparison in [d4c6674](https://github.com/ultralytics/yolov5/commit/d4c6674c98e19df4c40e33a777610a18d1961145).
 - **June 9, 2020**: [CSP](https://github.com/WongKinYiu/CrossStagePartialNetworks) updates to all YOLOv5 models. New models are faster, smaller and more accurate. Credit to @WongKinYiu for his excellent work with CSP. 
@@ -14,13 +14,14 @@ This repository represents Ultralytics open-source research into future object d
 
 ## Pretrained Checkpoints
 
-| Model | AP<sup>val</sup> | AP<sup>test</sup> | AP<sub>50</sub> | Speed<sub>GPU</sub> | FPS<sub>GPU</sub> || params | FLOPs |
+| Model | AP<sup>val</sup> | AP<sup>test</sup> | AP<sub>50</sub> | Speed<sub>GPU</sub> | FPS<sub>GPU</sub> || params | FLOPS |
 |---------- |------ |------ |------ | -------- | ------| ------ |------  |  :------: |
-| YOLOv5-s ([ckpt](https://drive.google.com/open?id=1Drs_Aiu7xx6S-ix95f9kNsA6ueKRpN2J))    | 35.5     | 35.5     | 55.0     | **2.1ms** | **476** || 7.1M   | 12.6B
-| YOLOv5-m ([ckpt](https://drive.google.com/open?id=1Drs_Aiu7xx6S-ix95f9kNsA6ueKRpN2J))    | 42.7     | 42.7     | 62.4     | 3.2ms     | 312     || 22.0M  | 39.0B
-| YOLOv5-l ([ckpt](https://drive.google.com/open?id=1Drs_Aiu7xx6S-ix95f9kNsA6ueKRpN2J))    | 45.7     | 45.9     | 65.1     | 4.1ms     | 243     || 50.3M  | 89.0B
-| YOLOv5-x ([ckpt](https://drive.google.com/open?id=1Drs_Aiu7xx6S-ix95f9kNsA6ueKRpN2J))    | **47.2** | **47.3** | **66.6** | 6.5ms     | 153     || 95.9M  | 170.3B
-| YOLOv3-SPP ([ckpt](https://drive.google.com/open?id=1Drs_Aiu7xx6S-ix95f9kNsA6ueKRpN2J))  | 45.6     | 45.5     | 65.2     | 4.8ms     | 208     || 63.0M  | 118.0B
+| [YOLOv5s](https://drive.google.com/open?id=1Drs_Aiu7xx6S-ix95f9kNsA6ueKRpN2J)    | 36.5     | 36.5     | 55.6     | **2.2ms** | **455** || 7.5M   | 13.2B
+| [YOLOv5m](https://drive.google.com/open?id=1Drs_Aiu7xx6S-ix95f9kNsA6ueKRpN2J)    | 43.4     | 43.4     | 62.4     | 3.0ms     | 333     || 21.8M  | 39.4B
+| [YOLOv5l](https://drive.google.com/open?id=1Drs_Aiu7xx6S-ix95f9kNsA6ueKRpN2J)    | 46.6     | 46.7     | 65.4     | 3.9ms     | 256     || 47.8M  | 88.1B
+| [YOLOv5x](https://drive.google.com/open?id=1Drs_Aiu7xx6S-ix95f9kNsA6ueKRpN2J)    | **48.2** | **48.3** | **66.9** | 6.1ms     | 164     || 89.0M  | 166.4B
+| [YOLOv3-SPP](https://drive.google.com/open?id=1Drs_Aiu7xx6S-ix95f9kNsA6ueKRpN2J)  | 45.6     | 45.5     | 65.2     | 4.5ms     | 222     || 63.0M  | 118.0B
+
 
 ** AP<sup>test</sup> denotes COCO [test-dev2017](http://cocodataset.org/#upload) server results, all other AP results in the table denote val2017 accuracy.  
 ** All AP numbers are for single-model single-scale without ensemble or test-time augmentation. Reproduce by `python test.py --img 736 --conf 0.001`  
diff --git a/models/yolov3-spp.yaml b/models/yolov3-spp.yaml
index 3dad009..6508dc4 100644
--- a/models/yolov3-spp.yaml
+++ b/models/yolov3-spp.yaml
@@ -25,8 +25,7 @@ backbone:
    [-1, 4, Bottleneck, [1024]],  # 10
   ]
 
-# yolov3-spp head
-# na = len(anchors[0])
+# YOLOv3-SPP head
 head:
   [[-1, 1, Bottleneck, [1024, False]],  # 11
    [-1, 1, SPP, [512, [5, 9, 13]]],
diff --git a/models/yolov5l.yaml b/models/yolov5l.yaml
index f270fdc..959d4bd 100644
--- a/models/yolov5l.yaml
+++ b/models/yolov5l.yaml
@@ -5,41 +5,48 @@ width_multiple: 1.0  # layer channel multiple
 
 # anchors
 anchors:
-  - [10,13, 16,30, 33,23]  # P3/8
-  - [30,61, 62,45, 59,119]  # P4/16
   - [116,90, 156,198, 373,326]  # P5/32
+  - [30,61, 62,45, 59,119]  # P4/16
+  - [10,13, 16,30, 33,23]  # P3/8
 
-# yolov5 backbone
+# YOLOv5 backbone
 backbone:
   # [from, number, module, args]
-  [[-1, 1, Focus, [64, 3]],  # 1-P1/2
-   [-1, 1, Conv, [128, 3, 2]],  # 2-P2/4
-   [-1, 3, Bottleneck, [128]],
-   [-1, 1, Conv, [256, 3, 2]],  # 4-P3/8
+  [[-1, 1, Focus, [64, 3]],  # 0-P1/2
+   [-1, 1, Conv, [128, 3, 2]],  # 1-P2/4
+   [-1, 3, BottleneckCSP, [128]],
+   [-1, 1, Conv, [256, 3, 2]],  # 3-P3/8
    [-1, 9, BottleneckCSP, [256]],
-   [-1, 1, Conv, [512, 3, 2]],  # 6-P4/16
+   [-1, 1, Conv, [512, 3, 2]],  # 5-P4/16
    [-1, 9, BottleneckCSP, [512]],
-   [-1, 1, Conv, [1024, 3, 2]], # 8-P5/32
+   [-1, 1, Conv, [1024, 3, 2]], # 7-P5/32
    [-1, 1, SPP, [1024, [5, 9, 13]]],
-   [-1, 6, BottleneckCSP, [1024]],  # 10
   ]
 
-# yolov5 head
+# YOLOv5 head
 head:
-  [[-1, 3, BottleneckCSP, [1024, False]],  # 11
-   [-1, 1, nn.Conv2d, [na * (nc + 5), 1, 1]],  # 12 (P5/32-large)
+  [[-1, 3, BottleneckCSP, [1024, False]],  # 9
 
-   [-2, 1, nn.Upsample, [None, 2, 'nearest']],
-   [[-1, 6], 1, Concat, [1]],  # cat backbone P4
    [-1, 1, Conv, [512, 1, 1]],
-   [-1, 3, BottleneckCSP, [512, False]],
-   [-1, 1, nn.Conv2d, [na * (nc + 5), 1, 1]],  # 17 (P4/16-medium)
+   [-1, 1, nn.Upsample, [None, 2, 'nearest']],
+   [[-1, 6], 1, Concat, [1]],  # cat backbone P4
+   [-1, 3, BottleneckCSP, [512, False]],  # 13
 
-   [-2, 1, nn.Upsample, [None, 2, 'nearest']],
-   [[-1, 4], 1, Concat, [1]],  # cat backbone P3
    [-1, 1, Conv, [256, 1, 1]],
+   [-1, 1, nn.Upsample, [None, 2, 'nearest']],
+   [[-1, 4], 1, Concat, [1]],  # cat backbone P3
    [-1, 3, BottleneckCSP, [256, False]],
-   [-1, 1, nn.Conv2d, [na * (nc + 5), 1, 1]],  # 22 (P3/8-small)
+   [-1, 1, nn.Conv2d, [na * (nc + 5), 1, 1]],  # 18 (P3/8-small)
+
+   [-2, 1, Conv, [256, 3, 2]],
+   [[-1, 14], 1, Concat, [1]],  # cat head P4
+   [-1, 3, BottleneckCSP, [512, False]],
+   [-1, 1, nn.Conv2d, [na * (nc + 5), 1, 1]],  # 22 (P4/16-medium)
+
+   [-2, 1, Conv, [512, 3, 2]],
+   [[-1, 10], 1, Concat, [1]],  # cat head P5
+   [-1, 3, BottleneckCSP, [1024, False]],
+   [-1, 1, nn.Conv2d, [na * (nc + 5), 1, 1]],  # 26 (P5/32-large)
 
-   [[], 1, Detect, [nc, anchors]],  # Detect(P3, P4, P5)
+   [[], 1, Detect, [nc, anchors]],  # Detect(P5, P4, P3)
   ]
diff --git a/models/yolov5m.yaml b/models/yolov5m.yaml
index 8498004..60037c2 100644
--- a/models/yolov5m.yaml
+++ b/models/yolov5m.yaml
@@ -5,41 +5,48 @@ width_multiple: 0.75  # layer channel multiple
 
 # anchors
 anchors:
-  - [10,13, 16,30, 33,23]  # P3/8
-  - [30,61, 62,45, 59,119]  # P4/16
   - [116,90, 156,198, 373,326]  # P5/32
+  - [30,61, 62,45, 59,119]  # P4/16
+  - [10,13, 16,30, 33,23]  # P3/8
 
-# yolov5 backbone
+# YOLOv5 backbone
 backbone:
   # [from, number, module, args]
-  [[-1, 1, Focus, [64, 3]],  # 1-P1/2
-   [-1, 1, Conv, [128, 3, 2]],  # 2-P2/4
-   [-1, 3, Bottleneck, [128]],
-   [-1, 1, Conv, [256, 3, 2]],  # 4-P3/8
+  [[-1, 1, Focus, [64, 3]],  # 0-P1/2
+   [-1, 1, Conv, [128, 3, 2]],  # 1-P2/4
+   [-1, 3, BottleneckCSP, [128]],
+   [-1, 1, Conv, [256, 3, 2]],  # 3-P3/8
    [-1, 9, BottleneckCSP, [256]],
-   [-1, 1, Conv, [512, 3, 2]],  # 6-P4/16
+   [-1, 1, Conv, [512, 3, 2]],  # 5-P4/16
    [-1, 9, BottleneckCSP, [512]],
-   [-1, 1, Conv, [1024, 3, 2]], # 8-P5/32
+   [-1, 1, Conv, [1024, 3, 2]], # 7-P5/32
    [-1, 1, SPP, [1024, [5, 9, 13]]],
-   [-1, 6, BottleneckCSP, [1024]],  # 10
   ]
 
-# yolov5 head
+# YOLOv5 head
 head:
-  [[-1, 3, BottleneckCSP, [1024, False]],  # 11
-   [-1, 1, nn.Conv2d, [na * (nc + 5), 1, 1]],  # 12 (P5/32-large)
+  [[-1, 3, BottleneckCSP, [1024, False]],  # 9
 
-   [-2, 1, nn.Upsample, [None, 2, 'nearest']],
-   [[-1, 6], 1, Concat, [1]],  # cat backbone P4
    [-1, 1, Conv, [512, 1, 1]],
-   [-1, 3, BottleneckCSP, [512, False]],
-   [-1, 1, nn.Conv2d, [na * (nc + 5), 1, 1]],  # 17 (P4/16-medium)
+   [-1, 1, nn.Upsample, [None, 2, 'nearest']],
+   [[-1, 6], 1, Concat, [1]],  # cat backbone P4
+   [-1, 3, BottleneckCSP, [512, False]],  # 13
 
-   [-2, 1, nn.Upsample, [None, 2, 'nearest']],
-   [[-1, 4], 1, Concat, [1]],  # cat backbone P3
    [-1, 1, Conv, [256, 1, 1]],
+   [-1, 1, nn.Upsample, [None, 2, 'nearest']],
+   [[-1, 4], 1, Concat, [1]],  # cat backbone P3
    [-1, 3, BottleneckCSP, [256, False]],
-   [-1, 1, nn.Conv2d, [na * (nc + 5), 1, 1]],  # 22 (P3/8-small)
+   [-1, 1, nn.Conv2d, [na * (nc + 5), 1, 1]],  # 18 (P3/8-small)
+
+   [-2, 1, Conv, [256, 3, 2]],
+   [[-1, 14], 1, Concat, [1]],  # cat head P4
+   [-1, 3, BottleneckCSP, [512, False]],
+   [-1, 1, nn.Conv2d, [na * (nc + 5), 1, 1]],  # 22 (P4/16-medium)
+
+   [-2, 1, Conv, [512, 3, 2]],
+   [[-1, 10], 1, Concat, [1]],  # cat head P5
+   [-1, 3, BottleneckCSP, [1024, False]],
+   [-1, 1, nn.Conv2d, [na * (nc + 5), 1, 1]],  # 26 (P5/32-large)
 
-   [[], 1, Detect, [nc, anchors]],  # Detect(P3, P4, P5)
+   [[], 1, Detect, [nc, anchors]],  # Detect(P5, P4, P3)
   ]
diff --git a/models/yolov5s.yaml b/models/yolov5s.yaml
index 482d1dd..1eaef97 100644
--- a/models/yolov5s.yaml
+++ b/models/yolov5s.yaml
@@ -5,41 +5,48 @@ width_multiple: 0.50  # layer channel multiple
 
 # anchors
 anchors:
-  - [10,13, 16,30, 33,23]  # P3/8
-  - [30,61, 62,45, 59,119]  # P4/16
   - [116,90, 156,198, 373,326]  # P5/32
+  - [30,61, 62,45, 59,119]  # P4/16
+  - [10,13, 16,30, 33,23]  # P3/8
 
-# yolov5 backbone
+# YOLOv5 backbone
 backbone:
   # [from, number, module, args]
-  [[-1, 1, Focus, [64, 3]],  # 1-P1/2
-   [-1, 1, Conv, [128, 3, 2]],  # 2-P2/4
-   [-1, 3, Bottleneck, [128]],
-   [-1, 1, Conv, [256, 3, 2]],  # 4-P3/8
+  [[-1, 1, Focus, [64, 3]],  # 0-P1/2
+   [-1, 1, Conv, [128, 3, 2]],  # 1-P2/4
+   [-1, 3, BottleneckCSP, [128]],
+   [-1, 1, Conv, [256, 3, 2]],  # 3-P3/8
    [-1, 9, BottleneckCSP, [256]],
-   [-1, 1, Conv, [512, 3, 2]],  # 6-P4/16
+   [-1, 1, Conv, [512, 3, 2]],  # 5-P4/16
    [-1, 9, BottleneckCSP, [512]],
-   [-1, 1, Conv, [1024, 3, 2]], # 8-P5/32
+   [-1, 1, Conv, [1024, 3, 2]], # 7-P5/32
    [-1, 1, SPP, [1024, [5, 9, 13]]],
-   [-1, 6, BottleneckCSP, [1024]],  # 10
   ]
 
-# yolov5 head
+# YOLOv5 head
 head:
-  [[-1, 3, BottleneckCSP, [1024, False]],  # 11
-   [-1, 1, nn.Conv2d, [na * (nc + 5), 1, 1]],  # 12 (P5/32-large)
+  [[-1, 3, BottleneckCSP, [1024, False]],  # 9
 
-   [-2, 1, nn.Upsample, [None, 2, 'nearest']],
-   [[-1, 6], 1, Concat, [1]],  # cat backbone P4
    [-1, 1, Conv, [512, 1, 1]],
-   [-1, 3, BottleneckCSP, [512, False]],
-   [-1, 1, nn.Conv2d, [na * (nc + 5), 1, 1]],  # 17 (P4/16-medium)
+   [-1, 1, nn.Upsample, [None, 2, 'nearest']],
+   [[-1, 6], 1, Concat, [1]],  # cat backbone P4
+   [-1, 3, BottleneckCSP, [512, False]],  # 13
 
-   [-2, 1, nn.Upsample, [None, 2, 'nearest']],
-   [[-1, 4], 1, Concat, [1]],  # cat backbone P3
    [-1, 1, Conv, [256, 1, 1]],
+   [-1, 1, nn.Upsample, [None, 2, 'nearest']],
+   [[-1, 4], 1, Concat, [1]],  # cat backbone P3
    [-1, 3, BottleneckCSP, [256, False]],
-   [-1, 1, nn.Conv2d, [na * (nc + 5), 1, 1]],  # 22 (P3/8-small)
+   [-1, 1, nn.Conv2d, [na * (nc + 5), 1, 1]],  # 18 (P3/8-small)
+
+   [-2, 1, Conv, [256, 3, 2]],
+   [[-1, 14], 1, Concat, [1]],  # cat head P4
+   [-1, 3, BottleneckCSP, [512, False]],
+   [-1, 1, nn.Conv2d, [na * (nc + 5), 1, 1]],  # 22 (P4/16-medium)
+
+   [-2, 1, Conv, [512, 3, 2]],
+   [[-1, 10], 1, Concat, [1]],  # cat head P5
+   [-1, 3, BottleneckCSP, [1024, False]],
+   [-1, 1, nn.Conv2d, [na * (nc + 5), 1, 1]],  # 26 (P5/32-large)
 
-   [[], 1, Detect, [nc, anchors]],  # Detect(P3, P4, P5)
+   [[], 1, Detect, [nc, anchors]],  # Detect(P5, P4, P3)
   ]
diff --git a/models/yolov5x.yaml b/models/yolov5x.yaml
index 47658e8..dcd6fbc 100644
--- a/models/yolov5x.yaml
+++ b/models/yolov5x.yaml
@@ -5,41 +5,48 @@ width_multiple: 1.25  # layer channel multiple
 
 # anchors
 anchors:
-  - [10,13, 16,30, 33,23]  # P3/8
-  - [30,61, 62,45, 59,119]  # P4/16
   - [116,90, 156,198, 373,326]  # P5/32
+  - [30,61, 62,45, 59,119]  # P4/16
+  - [10,13, 16,30, 33,23]  # P3/8
 
-# yolov5 backbone
+# YOLOv5 backbone
 backbone:
   # [from, number, module, args]
-  [[-1, 1, Focus, [64, 3]],  # 1-P1/2
-   [-1, 1, Conv, [128, 3, 2]],  # 2-P2/4
-   [-1, 3, Bottleneck, [128]],
-   [-1, 1, Conv, [256, 3, 2]],  # 4-P3/8
+  [[-1, 1, Focus, [64, 3]],  # 0-P1/2
+   [-1, 1, Conv, [128, 3, 2]],  # 1-P2/4
+   [-1, 3, BottleneckCSP, [128]],
+   [-1, 1, Conv, [256, 3, 2]],  # 3-P3/8
    [-1, 9, BottleneckCSP, [256]],
-   [-1, 1, Conv, [512, 3, 2]],  # 6-P4/16
+   [-1, 1, Conv, [512, 3, 2]],  # 5-P4/16
    [-1, 9, BottleneckCSP, [512]],
-   [-1, 1, Conv, [1024, 3, 2]], # 8-P5/32
+   [-1, 1, Conv, [1024, 3, 2]], # 7-P5/32
    [-1, 1, SPP, [1024, [5, 9, 13]]],
-   [-1, 6, BottleneckCSP, [1024]],  # 10
   ]
 
-# yolov5 head
+# YOLOv5 head
 head:
-  [[-1, 3, BottleneckCSP, [1024, False]],  # 11
-   [-1, 1, nn.Conv2d, [na * (nc + 5), 1, 1]],  # 12 (P5/32-large)
+  [[-1, 3, BottleneckCSP, [1024, False]],  # 9
 
-   [-2, 1, nn.Upsample, [None, 2, 'nearest']],
-   [[-1, 6], 1, Concat, [1]],  # cat backbone P4
    [-1, 1, Conv, [512, 1, 1]],
-   [-1, 3, BottleneckCSP, [512, False]],
-   [-1, 1, nn.Conv2d, [na * (nc + 5), 1, 1]],  # 17 (P4/16-medium)
+   [-1, 1, nn.Upsample, [None, 2, 'nearest']],
+   [[-1, 6], 1, Concat, [1]],  # cat backbone P4
+   [-1, 3, BottleneckCSP, [512, False]],  # 13
 
-   [-2, 1, nn.Upsample, [None, 2, 'nearest']],
-   [[-1, 4], 1, Concat, [1]],  # cat backbone P3
    [-1, 1, Conv, [256, 1, 1]],
+   [-1, 1, nn.Upsample, [None, 2, 'nearest']],
+   [[-1, 4], 1, Concat, [1]],  # cat backbone P3
    [-1, 3, BottleneckCSP, [256, False]],
-   [-1, 1, nn.Conv2d, [na * (nc + 5), 1, 1]],  # 22 (P3/8-small)
+   [-1, 1, nn.Conv2d, [na * (nc + 5), 1, 1]],  # 18 (P3/8-small)
+
+   [-2, 1, Conv, [256, 3, 2]],
+   [[-1, 14], 1, Concat, [1]],  # cat head P4
+   [-1, 3, BottleneckCSP, [512, False]],
+   [-1, 1, nn.Conv2d, [na * (nc + 5), 1, 1]],  # 22 (P4/16-medium)
+
+   [-2, 1, Conv, [512, 3, 2]],
+   [[-1, 10], 1, Concat, [1]],  # cat head P5
+   [-1, 3, BottleneckCSP, [1024, False]],
+   [-1, 1, nn.Conv2d, [na * (nc + 5), 1, 1]],  # 26 (P5/32-large)
 
-   [[], 1, Detect, [nc, anchors]],  # Detect(P3, P4, P5)
+   [[], 1, Detect, [nc, anchors]],  # Detect(P5, P4, P3)
   ]
diff --git a/utils/utils.py b/utils/utils.py
index f1f5db5..9dd1d65 100755
--- a/utils/utils.py
+++ b/utils/utils.py
@@ -1094,12 +1094,14 @@ def plot_study_txt(f='study.txt', x=None):  # from utils.utils import *; plot_st
 
     ax2.plot(1E3 / np.array([209, 140, 97, 58, 35, 18]), [33.5, 39.1, 42.5, 45.9, 49., 50.5],
              'k.-', linewidth=2, markersize=8, alpha=.25, label='EfficientDet')
+
+    ax2.grid()
     ax2.set_xlim(0, 30)
-    ax2.set_ylim(25, 50)
-    ax2.set_xlabel('GPU Latency (ms)')
+    ax2.set_ylim(28, 50)
+    ax2.set_yticks(np.arange(30, 55, 5))
+    ax2.set_xlabel('GPU Speed (ms/img)')
     ax2.set_ylabel('COCO AP val')
     ax2.legend(loc='lower right')
-    ax2.grid()
     plt.savefig('study_mAP_latency.png', dpi=300)
     plt.savefig(f.replace('.txt', '.png'), dpi=200)
 

From 69ff48669300dec69806d54aecd93925b88191e5 Mon Sep 17 00:00:00 2001
From: Glenn Jocher <glenn.jocher@ultralytics.com>
Date: Mon, 22 Jun 2020 14:32:05 -0700
Subject: [PATCH 44/73] Update README.md

---
 README.md | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/README.md b/README.md
index df49f67..59c5682 100755
--- a/README.md
+++ b/README.md
@@ -6,7 +6,8 @@ This repository represents Ultralytics open-source research into future object d
 
 <img src="https://user-images.githubusercontent.com/26833433/85336627-c6663280-b493-11ea-9b0a-289b0f182b84.png" width="1000">** GPU Speed measures end-to-end time per image averaged over 5000 COCO val2017 images using a V100 GPU with batch size 8, and includes image preprocessing, PyTorch FP16 inference, postprocessing and NMS.
 
-- **June 19, 2020**: [FP16](https://pytorch.org/docs/stable/nn.html#torch.nn.Module.half) as new default for smaller checkpoints and faster inference. Comparison in [d4c6674](https://github.com/ultralytics/yolov5/commit/d4c6674c98e19df4c40e33a777610a18d1961145).
+- **June 22, 2020**: [PANet](https://arxiv.org/abs/1803.01534) updates: increased layers, reduced parameters, faster inference and improved mAP [364fcfd](https://github.com/ultralytics/yolov5/commit/364fcfd7dba53f46edd4f04c037a039c0a287972).
+- **June 19, 2020**: [FP16](https://pytorch.org/docs/stable/nn.html#torch.nn.Module.half) as new default for smaller checkpoints and faster inference [d4c6674](https://github.com/ultralytics/yolov5/commit/d4c6674c98e19df4c40e33a777610a18d1961145).
 - **June 9, 2020**: [CSP](https://github.com/WongKinYiu/CrossStagePartialNetworks) updates to all YOLOv5 models. New models are faster, smaller and more accurate. Credit to @WongKinYiu for his excellent work with CSP. 
 - **May 27, 2020**: Public release of repo. YOLOv5 models are SOTA among all known YOLO implementations, YOLOv5 family will be undergoing architecture research and development over Q2/Q3 2020 to increase performance. Updates may include [CSP](https://github.com/WongKinYiu/CrossStagePartialNetworks) bottlenecks, [YOLOv4](https://github.com/AlexeyAB/darknet) features, as well as PANet or BiFPN heads.
 - **April 1, 2020**: Begin development of a 100% PyTorch, scaleable YOLOv3/4-based group of future models, in a range of compound-scaled sizes. Models will be defined by new user-friendly `*.yaml` files. New training methods will be simpler to start, faster to finish, and more robust to training a wider variety of custom dataset.

From 3804578b78d7199361544620bdfc5ab826bc882b Mon Sep 17 00:00:00 2001
From: Glenn Jocher <glenn.jocher@ultralytics.com>
Date: Mon, 22 Jun 2020 14:33:47 -0700
Subject: [PATCH 45/73] Update README.md

---
 README.md | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/README.md b/README.md
index 59c5682..874141f 100755
--- a/README.md
+++ b/README.md
@@ -8,7 +8,7 @@ This repository represents Ultralytics open-source research into future object d
 
 - **June 22, 2020**: [PANet](https://arxiv.org/abs/1803.01534) updates: increased layers, reduced parameters, faster inference and improved mAP [364fcfd](https://github.com/ultralytics/yolov5/commit/364fcfd7dba53f46edd4f04c037a039c0a287972).
 - **June 19, 2020**: [FP16](https://pytorch.org/docs/stable/nn.html#torch.nn.Module.half) as new default for smaller checkpoints and faster inference [d4c6674](https://github.com/ultralytics/yolov5/commit/d4c6674c98e19df4c40e33a777610a18d1961145).
-- **June 9, 2020**: [CSP](https://github.com/WongKinYiu/CrossStagePartialNetworks) updates to all YOLOv5 models. New models are faster, smaller and more accurate. Credit to @WongKinYiu for his excellent work with CSP. 
+- **June 9, 2020**: [CSP](https://github.com/WongKinYiu/CrossStagePartialNetworks) updates: improved speed, size, and accuracy. Credit to @WongKinYiu for excellent CSP work. 
 - **May 27, 2020**: Public release of repo. YOLOv5 models are SOTA among all known YOLO implementations, YOLOv5 family will be undergoing architecture research and development over Q2/Q3 2020 to increase performance. Updates may include [CSP](https://github.com/WongKinYiu/CrossStagePartialNetworks) bottlenecks, [YOLOv4](https://github.com/AlexeyAB/darknet) features, as well as PANet or BiFPN heads.
 - **April 1, 2020**: Begin development of a 100% PyTorch, scaleable YOLOv3/4-based group of future models, in a range of compound-scaled sizes. Models will be defined by new user-friendly `*.yaml` files. New training methods will be simpler to start, faster to finish, and more robust to training a wider variety of custom dataset.
 

From 7306dadf78a7937c466a256be032be99b4f0297b Mon Sep 17 00:00:00 2001
From: Glenn Jocher <glenn.jocher@ultralytics.com>
Date: Mon, 22 Jun 2020 15:16:10 -0700
Subject: [PATCH 46/73] Update README.md

---
 README.md | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/README.md b/README.md
index 874141f..d74e355 100755
--- a/README.md
+++ b/README.md
@@ -4,7 +4,7 @@
 
 This repository represents Ultralytics open-source research into future object detection methods, and incorporates our lessons learned and best practices evolved over training thousands of models on custom client datasets with our previous YOLO repository https://github.com/ultralytics/yolov3. **All code and models are under active development, and are subject to modification or deletion without notice.** Use at your own risk.
 
-<img src="https://user-images.githubusercontent.com/26833433/85336627-c6663280-b493-11ea-9b0a-289b0f182b84.png" width="1000">** GPU Speed measures end-to-end time per image averaged over 5000 COCO val2017 images using a V100 GPU with batch size 8, and includes image preprocessing, PyTorch FP16 inference, postprocessing and NMS.
+<img src="https://user-images.githubusercontent.com/26833433/85340570-30360a80-b49b-11ea-87cf-bdf33d53ae15.png" width="1000">** GPU Speed measures end-to-end time per image averaged over 5000 COCO val2017 images using a V100 GPU with batch size 8, and includes image preprocessing, PyTorch FP16 inference, postprocessing and NMS.
 
 - **June 22, 2020**: [PANet](https://arxiv.org/abs/1803.01534) updates: increased layers, reduced parameters, faster inference and improved mAP [364fcfd](https://github.com/ultralytics/yolov5/commit/364fcfd7dba53f46edd4f04c037a039c0a287972).
 - **June 19, 2020**: [FP16](https://pytorch.org/docs/stable/nn.html#torch.nn.Module.half) as new default for smaller checkpoints and faster inference [d4c6674](https://github.com/ultralytics/yolov5/commit/d4c6674c98e19df4c40e33a777610a18d1961145).

From 6c1b87a42ee5cdbc0c996e046acf77c26f85e15e Mon Sep 17 00:00:00 2001
From: Glenn Jocher <glenn.jocher@ultralytics.com>
Date: Mon, 22 Jun 2020 23:00:23 -0700
Subject: [PATCH 47/73] update google_utils import

---
 detect.py      | 1 +
 test.py        | 1 +
 train.py       | 1 +
 utils/utils.py | 2 +-
 4 files changed, 4 insertions(+), 1 deletion(-)

diff --git a/detect.py b/detect.py
index 567c9e6..e4a7ddb 100644
--- a/detect.py
+++ b/detect.py
@@ -2,6 +2,7 @@ import argparse
 
 import torch.backends.cudnn as cudnn
 
+from utils import google_utils
 from utils.datasets import *
 from utils.utils import *
 
diff --git a/test.py b/test.py
index 2381ab4..10de7d8 100644
--- a/test.py
+++ b/test.py
@@ -3,6 +3,7 @@ import json
 
 from torch.utils.data import DataLoader
 
+from utils import google_utils
 from utils.datasets import *
 from utils.utils import *
 
diff --git a/train.py b/train.py
index 921f0e3..2dfe781 100644
--- a/train.py
+++ b/train.py
@@ -9,6 +9,7 @@ from torch.utils.tensorboard import SummaryWriter
 
 import test  # import test.py to get mAP after each epoch
 from models.yolo import Model
+from utils import google_utils
 from utils.datasets import *
 from utils.utils import *
 
diff --git a/utils/utils.py b/utils/utils.py
index 9dd1d65..cf819c5 100755
--- a/utils/utils.py
+++ b/utils/utils.py
@@ -20,7 +20,7 @@ import yaml
 from scipy.signal import butter, filtfilt
 from tqdm import tqdm
 
-from . import torch_utils, google_utils  #  torch_utils, google_utils
+from . import torch_utils  #  torch_utils, google_utils
 
 # Set printoptions
 torch.set_printoptions(linewidth=320, precision=5, profile='long')

From db898131f965ecf5193c7304a5d7992e072fa917 Mon Sep 17 00:00:00 2001
From: Glenn Jocher <glenn.jocher@ultralytics.com>
Date: Mon, 22 Jun 2020 23:07:22 -0700
Subject: [PATCH 48/73] study update

---
 test.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/test.py b/test.py
index 10de7d8..227c557 100644
--- a/test.py
+++ b/test.py
@@ -279,7 +279,7 @@ if __name__ == '__main__':
     elif opt.task == 'study':  # run over a range of settings and save/plot
         for weights in ['yolov5s.pt', 'yolov5m.pt', 'yolov5l.pt', 'yolov5x.pt']:
             f = 'study_%s_%s.txt' % (Path(opt.data).stem, Path(weights).stem)  # filename to save to
-            x = list(range(288, 896, 64))  # x axis
+            x = list(range(352, 832, 64))  # x axis
             y = []  # y axis
             for i in x:  # img-size
                 print('\nRunning %s point %s...' % (f, i))

From 2a550333c0c8b39a76d4638dab530a84da247bcb Mon Sep 17 00:00:00 2001
From: Glenn Jocher <glenn.jocher@ultralytics.com>
Date: Mon, 22 Jun 2020 23:20:37 -0700
Subject: [PATCH 49/73] update README.md

---
 README.md | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/README.md b/README.md
index d74e355..cf1a0e8 100755
--- a/README.md
+++ b/README.md
@@ -20,7 +20,7 @@ This repository represents Ultralytics open-source research into future object d
 | [YOLOv5s](https://drive.google.com/open?id=1Drs_Aiu7xx6S-ix95f9kNsA6ueKRpN2J)    | 36.5     | 36.5     | 55.6     | **2.2ms** | **455** || 7.5M   | 13.2B
 | [YOLOv5m](https://drive.google.com/open?id=1Drs_Aiu7xx6S-ix95f9kNsA6ueKRpN2J)    | 43.4     | 43.4     | 62.4     | 3.0ms     | 333     || 21.8M  | 39.4B
 | [YOLOv5l](https://drive.google.com/open?id=1Drs_Aiu7xx6S-ix95f9kNsA6ueKRpN2J)    | 46.6     | 46.7     | 65.4     | 3.9ms     | 256     || 47.8M  | 88.1B
-| [YOLOv5x](https://drive.google.com/open?id=1Drs_Aiu7xx6S-ix95f9kNsA6ueKRpN2J)    | **48.2** | **48.3** | **66.9** | 6.1ms     | 164     || 89.0M  | 166.4B
+| [YOLOv5x](https://drive.google.com/open?id=1Drs_Aiu7xx6S-ix95f9kNsA6ueKRpN2J)    | **48.4** | **48.4** | **66.9** | 6.1ms     | 164     || 89.0M  | 166.4B
 | [YOLOv3-SPP](https://drive.google.com/open?id=1Drs_Aiu7xx6S-ix95f9kNsA6ueKRpN2J)  | 45.6     | 45.5     | 65.2     | 4.5ms     | 222     || 63.0M  | 118.0B
 
 

From a63e1c9b3ca4cef623e2f8b83787e1580926b91b Mon Sep 17 00:00:00 2001
From: Glenn Jocher <glenn.jocher@ultralytics.com>
Date: Mon, 22 Jun 2020 23:23:50 -0700
Subject: [PATCH 50/73] study update

---
 test.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/test.py b/test.py
index 227c557..b3ff968 100644
--- a/test.py
+++ b/test.py
@@ -277,7 +277,7 @@ if __name__ == '__main__':
              opt.verbose)
 
     elif opt.task == 'study':  # run over a range of settings and save/plot
-        for weights in ['yolov5s.pt', 'yolov5m.pt', 'yolov5l.pt', 'yolov5x.pt']:
+        for weights in ['yolov5s.pt', 'yolov5m.pt', 'yolov5l.pt', 'yolov5x.pt', 'yolov3-spp.pt']:
             f = 'study_%s_%s.txt' % (Path(opt.data).stem, Path(weights).stem)  # filename to save to
             x = list(range(352, 832, 64))  # x axis
             y = []  # y axis

From 6cea1bff60452ffe5d9a65877d4e9e8d9366d88f Mon Sep 17 00:00:00 2001
From: Glenn Jocher <glenn.jocher@ultralytics.com>
Date: Tue, 23 Jun 2020 10:11:34 -0700
Subject: [PATCH 51/73] Update README.md

---
 README.md | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/README.md b/README.md
index cf1a0e8..4268583 100755
--- a/README.md
+++ b/README.md
@@ -17,7 +17,7 @@ This repository represents Ultralytics open-source research into future object d
 
 | Model | AP<sup>val</sup> | AP<sup>test</sup> | AP<sub>50</sub> | Speed<sub>GPU</sub> | FPS<sub>GPU</sub> || params | FLOPS |
 |---------- |------ |------ |------ | -------- | ------| ------ |------  |  :------: |
-| [YOLOv5s](https://drive.google.com/open?id=1Drs_Aiu7xx6S-ix95f9kNsA6ueKRpN2J)    | 36.5     | 36.5     | 55.6     | **2.2ms** | **455** || 7.5M   | 13.2B
+| [YOLOv5s](https://drive.google.com/open?id=1Drs_Aiu7xx6S-ix95f9kNsA6ueKRpN2J)    | 36.5     | 36.5     | 55.6     | **2.1ms** | **476** || 7.5M   | 13.2B
 | [YOLOv5m](https://drive.google.com/open?id=1Drs_Aiu7xx6S-ix95f9kNsA6ueKRpN2J)    | 43.4     | 43.4     | 62.4     | 3.0ms     | 333     || 21.8M  | 39.4B
 | [YOLOv5l](https://drive.google.com/open?id=1Drs_Aiu7xx6S-ix95f9kNsA6ueKRpN2J)    | 46.6     | 46.7     | 65.4     | 3.9ms     | 256     || 47.8M  | 88.1B
 | [YOLOv5x](https://drive.google.com/open?id=1Drs_Aiu7xx6S-ix95f9kNsA6ueKRpN2J)    | **48.4** | **48.4** | **66.9** | 6.1ms     | 164     || 89.0M  | 166.4B

From 53294f28b4b23cec757b8fb9c704f7ae436b854c Mon Sep 17 00:00:00 2001
From: Glenn Jocher <glenn.jocher@ultralytics.com>
Date: Tue, 23 Jun 2020 10:53:42 -0700
Subject: [PATCH 52/73] update yolo.py

---
 models/yolo.py | 4 +---
 1 file changed, 1 insertion(+), 3 deletions(-)

diff --git a/models/yolo.py b/models/yolo.py
index e491e64..158a241 100644
--- a/models/yolo.py
+++ b/models/yolo.py
@@ -1,7 +1,5 @@
 import argparse
 
-import yaml
-
 from models.experimental import *
 
 
@@ -206,7 +204,7 @@ def parse_model(md, ch):  # model_dict, input_channels(3)
 
 if __name__ == '__main__':
     parser = argparse.ArgumentParser()
-    parser.add_argument('--cfg', type=str, default='yolov5s.yaml', help='model.yaml')
+    parser.add_argument('--cfg', type=str, default='yolov5m-p6.yaml', help='model.yaml')
     parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu')
     opt = parser.parse_args()
     opt.cfg = check_file(opt.cfg)  # check file

From f9ae460eeccd30bdc43a89a37f74b9cc7b93d52f Mon Sep 17 00:00:00 2001
From: Glenn Jocher <glenn.jocher@ultralytics.com>
Date: Tue, 23 Jun 2020 20:12:23 -0700
Subject: [PATCH 53/73] update README.md

---
 README.md | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/README.md b/README.md
index 4268583..8596da3 100755
--- a/README.md
+++ b/README.md
@@ -17,7 +17,7 @@ This repository represents Ultralytics open-source research into future object d
 
 | Model | AP<sup>val</sup> | AP<sup>test</sup> | AP<sub>50</sub> | Speed<sub>GPU</sub> | FPS<sub>GPU</sub> || params | FLOPS |
 |---------- |------ |------ |------ | -------- | ------| ------ |------  |  :------: |
-| [YOLOv5s](https://drive.google.com/open?id=1Drs_Aiu7xx6S-ix95f9kNsA6ueKRpN2J)    | 36.5     | 36.5     | 55.6     | **2.1ms** | **476** || 7.5M   | 13.2B
+| [YOLOv5s](https://drive.google.com/open?id=1Drs_Aiu7xx6S-ix95f9kNsA6ueKRpN2J)    | 36.6     | 36.6     | 55.8     | **2.1ms** | **476** || 7.5M   | 13.2B
 | [YOLOv5m](https://drive.google.com/open?id=1Drs_Aiu7xx6S-ix95f9kNsA6ueKRpN2J)    | 43.4     | 43.4     | 62.4     | 3.0ms     | 333     || 21.8M  | 39.4B
 | [YOLOv5l](https://drive.google.com/open?id=1Drs_Aiu7xx6S-ix95f9kNsA6ueKRpN2J)    | 46.6     | 46.7     | 65.4     | 3.9ms     | 256     || 47.8M  | 88.1B
 | [YOLOv5x](https://drive.google.com/open?id=1Drs_Aiu7xx6S-ix95f9kNsA6ueKRpN2J)    | **48.4** | **48.4** | **66.9** | 6.1ms     | 164     || 89.0M  | 166.4B

From 8699c319ca48253eb9983ee3be4c7ab0c3159d60 Mon Sep 17 00:00:00 2001
From: Glenn Jocher <glenn.jocher@ultralytics.com>
Date: Wed, 24 Jun 2020 10:42:01 -0700
Subject: [PATCH 54/73] update yolo.py

---
 models/yolo.py | 10 ++++++----
 1 file changed, 6 insertions(+), 4 deletions(-)

diff --git a/models/yolo.py b/models/yolo.py
index 158a241..2d853cb 100644
--- a/models/yolo.py
+++ b/models/yolo.py
@@ -96,8 +96,11 @@ class Model(nn.Module):
                 x = y[m.f] if isinstance(m.f, int) else [x if j == -1 else y[j] for j in m.f]  # from earlier layers
 
             if profile:
-                import thop
-                o = thop.profile(m, inputs=(x,), verbose=False)[0] / 1E9 * 2  # FLOPS
+                try:
+                    import thop
+                    o = thop.profile(m, inputs=(x,), verbose=False)[0] / 1E9 * 2  # FLOPS
+                except:
+                    o = 0
                 t = torch_utils.time_synchronized()
                 for _ in range(10):
                     _ = m(x)
@@ -217,11 +220,10 @@ if __name__ == '__main__':
     # Profile
     # img = torch.rand(8 if torch.cuda.is_available() else 1, 3, 640, 640).to(device)
     # y = model(img, profile=True)
-    # print([y[0].shape] + [x.shape for x in y[1]])
 
     # ONNX export
     # model.model[-1].export = True
-    # torch.onnx.export(model, img, f.replace('.yaml', '.onnx'), verbose=True, opset_version=11)
+    # torch.onnx.export(model, img, opt.cfg.replace('.yaml', '.onnx'), verbose=True, opset_version=11)
 
     # Tensorboard
     # from torch.utils.tensorboard import SummaryWriter

From b1b3634145aa206cd177a322109fce1c776e4779 Mon Sep 17 00:00:00 2001
From: Glenn Jocher <glenn.jocher@ultralytics.com>
Date: Wed, 24 Jun 2020 11:00:03 -0700
Subject: [PATCH 55/73] update onnx_export.py #134

---
 models/onnx_export.py | 1 +
 1 file changed, 1 insertion(+)

diff --git a/models/onnx_export.py b/models/onnx_export.py
index d17f8b2..fe0287e 100644
--- a/models/onnx_export.py
+++ b/models/onnx_export.py
@@ -9,6 +9,7 @@ import argparse
 import onnx
 
 from models.common import *
+from utils import google_utils
 
 if __name__ == '__main__':
     parser = argparse.ArgumentParser()

From 0825cb7fd891563a5e85e378d896970fc4b03604 Mon Sep 17 00:00:00 2001
From: Glenn Jocher <glenn.jocher@ultralytics.com>
Date: Wed, 24 Jun 2020 11:58:23 -0700
Subject: [PATCH 56/73] create models/hub

---
 models/{ => hub}/yolov3-spp.yaml |  0
 models/hub/yolov5-fpn.yaml       | 45 +++++++++++++++++++++++++++
 models/hub/yolov5-panet.yaml     | 52 ++++++++++++++++++++++++++++++++
 3 files changed, 97 insertions(+)
 rename models/{ => hub}/yolov3-spp.yaml (100%)
 create mode 100644 models/hub/yolov5-fpn.yaml
 create mode 100644 models/hub/yolov5-panet.yaml

diff --git a/models/yolov3-spp.yaml b/models/hub/yolov3-spp.yaml
similarity index 100%
rename from models/yolov3-spp.yaml
rename to models/hub/yolov3-spp.yaml
diff --git a/models/hub/yolov5-fpn.yaml b/models/hub/yolov5-fpn.yaml
new file mode 100644
index 0000000..52cec54
--- /dev/null
+++ b/models/hub/yolov5-fpn.yaml
@@ -0,0 +1,45 @@
+# parameters
+nc: 80  # number of classes
+depth_multiple: 1.0  # model depth multiple
+width_multiple: 1.0  # layer channel multiple
+
+# anchors
+anchors:
+  - [10,13, 16,30, 33,23]  # P3/8
+  - [30,61, 62,45, 59,119]  # P4/16
+  - [116,90, 156,198, 373,326]  # P5/32
+
+# YOLOv5 backbone
+backbone:
+  # [from, number, module, args]
+  [[-1, 1, Focus, [64, 3]],  # 0-P1/2
+   [-1, 1, Conv, [128, 3, 2]],  # 1-P2/4
+   [-1, 3, Bottleneck, [128]],
+   [-1, 1, Conv, [256, 3, 2]],  # 3-P3/8
+   [-1, 9, BottleneckCSP, [256]],
+   [-1, 1, Conv, [512, 3, 2]],  # 5-P4/16
+   [-1, 9, BottleneckCSP, [512]],
+   [-1, 1, Conv, [1024, 3, 2]], # 7-P5/32
+   [-1, 1, SPP, [1024, [5, 9, 13]]],
+   [-1, 6, BottleneckCSP, [1024]],  # 9
+  ]
+
+# YOLOv5 FPN head
+head:
+  [[-1, 3, BottleneckCSP, [1024, False]],
+   [-1, 1, nn.Conv2d, [na * (nc + 5), 1, 1]],  # 11 (P5/32-large)
+
+   [-2, 1, nn.Upsample, [None, 2, 'nearest']],
+   [[-1, 6], 1, Concat, [1]],  # cat backbone P4
+   [-1, 1, Conv, [512, 1, 1]],
+   [-1, 3, BottleneckCSP, [512, False]],
+   [-1, 1, nn.Conv2d, [na * (nc + 5), 1, 1]],  # 16 (P4/16-medium)
+
+   [-2, 1, nn.Upsample, [None, 2, 'nearest']],
+   [[-1, 4], 1, Concat, [1]],  # cat backbone P3
+   [-1, 1, Conv, [256, 1, 1]],
+   [-1, 3, BottleneckCSP, [256, False]],
+   [-1, 1, nn.Conv2d, [na * (nc + 5), 1, 1]],  # 21 (P3/8-small)
+
+   [[], 1, Detect, [nc, anchors]],  # Detect(P3, P4, P5)
+  ]
diff --git a/models/hub/yolov5-panet.yaml b/models/hub/yolov5-panet.yaml
new file mode 100644
index 0000000..aed1e23
--- /dev/null
+++ b/models/hub/yolov5-panet.yaml
@@ -0,0 +1,52 @@
+# parameters
+nc: 80  # number of classes
+depth_multiple: 1.0  # model depth multiple
+width_multiple: 1.0  # layer channel multiple
+
+# anchors
+anchors:
+  - [116,90, 156,198, 373,326]  # P5/32
+  - [30,61, 62,45, 59,119]  # P4/16
+  - [10,13, 16,30, 33,23]  # P3/8
+
+# YOLOv5 backbone
+backbone:
+  # [from, number, module, args]
+  [[-1, 1, Focus, [64, 3]],  # 0-P1/2
+   [-1, 1, Conv, [128, 3, 2]],  # 1-P2/4
+   [-1, 3, BottleneckCSP, [128]],
+   [-1, 1, Conv, [256, 3, 2]],  # 3-P3/8
+   [-1, 9, BottleneckCSP, [256]],
+   [-1, 1, Conv, [512, 3, 2]],  # 5-P4/16
+   [-1, 9, BottleneckCSP, [512]],
+   [-1, 1, Conv, [1024, 3, 2]], # 7-P5/32
+   [-1, 1, SPP, [1024, [5, 9, 13]]],
+  ]
+
+# YOLOv5 PANet head
+head:
+  [[-1, 3, BottleneckCSP, [1024, False]],
+   [-1, 1, Conv, [512, 1, 1]],  # 10
+
+   [-1, 1, nn.Upsample, [None, 2, 'nearest']],
+   [[-1, 6], 1, Concat, [1]],  # cat backbone P4
+   [-1, 3, BottleneckCSP, [512, False]],
+   [-1, 1, Conv, [256, 1, 1]],  # 14
+
+   [-1, 1, nn.Upsample, [None, 2, 'nearest']],
+   [[-1, 4], 1, Concat, [1]],  # cat backbone P3
+   [-1, 3, BottleneckCSP, [256, False]],
+   [-1, 1, nn.Conv2d, [na * (nc + 5), 1, 1]],  # 18 (P3/8-small)
+
+   [-2, 1, Conv, [256, 3, 2]],
+   [[-1, 14], 1, Concat, [1]],  # cat head P4
+   [-1, 3, BottleneckCSP, [512, False]],
+   [-1, 1, nn.Conv2d, [na * (nc + 5), 1, 1]],  # 22 (P4/16-medium)
+
+   [-2, 1, Conv, [512, 3, 2]],
+   [[-1, 10], 1, Concat, [1]],  # cat head P5
+   [-1, 3, BottleneckCSP, [1024, False]],
+   [-1, 1, nn.Conv2d, [na * (nc + 5), 1, 1]],  # 26 (P5/32-large)
+
+   [[], 1, Detect, [nc, anchors]],  # Detect(P5, P4, P3)
+  ]

From 3b062254a6cb22bc7b992d01267888e5795ccec4 Mon Sep 17 00:00:00 2001
From: Glenn Jocher <glenn.jocher@ultralytics.com>
Date: Wed, 24 Jun 2020 12:53:43 -0700
Subject: [PATCH 57/73] update utils.create_pretrained()

---
 Dockerfile     | 4 ++--
 utils/utils.py | 6 +++---
 2 files changed, 5 insertions(+), 5 deletions(-)

diff --git a/Dockerfile b/Dockerfile
index d73affc..bda78c0 100644
--- a/Dockerfile
+++ b/Dockerfile
@@ -1,5 +1,6 @@
 # Start FROM Nvidia PyTorch image https://ngc.nvidia.com/catalog/containers/nvidia:pytorch
 FROM nvcr.io/nvidia/pytorch:20.03-py3
+RUN pip install -U gsutil
 
 # Create working directory
 RUN mkdir -p /usr/src/app
@@ -10,7 +11,6 @@ COPY . /usr/src/app
 
 # Install dependencies (pip or conda)
 #RUN pip install -r requirements.txt
-RUN pip install -U gsutil
 
 # Copy weights
 #RUN python3 -c "from models import *; \
@@ -41,7 +41,7 @@ RUN pip install -U gsutil
 
 # Bash into running container
 # sudo docker container exec -it ba65811811ab bash
-# python -c "from utils.utils import *; create_backbone('weights/last.pt')" && gsutil cp weights/backbone.pt gs://*
+# python -c "from utils.utils import *; create_pretrained('weights/last.pt')" && gsutil cp weights/pretrained.pt gs://*
 
 # Bash into stopped container
 # sudo docker commit 6d525e299258 user/test_image && sudo docker run -it --gpus all --ipc=host -v "$(pwd)"/coco:/usr/src/coco --entrypoint=sh user/test_image
diff --git a/utils/utils.py b/utils/utils.py
index cf819c5..c33f41f 100755
--- a/utils/utils.py
+++ b/utils/utils.py
@@ -632,8 +632,8 @@ def strip_optimizer(f='weights/best.pt'):  # from utils.utils import *; strip_op
     print('Optimizer stripped from %s' % f)
 
 
-def create_backbone(f='weights/best.pt', s='weights/backbone.pt'):  # from utils.utils import *; create_backbone()
-    # create backbone 's' from 'f'
+def create_pretrained(f='weights/best.pt', s='weights/pretrained.pt'):  # from utils.utils import *; create_pretrained()
+    # create pretrained checkpoint 's' from 'f' (create_pretrained(x, x) for x in glob.glob('./*.pt'))
     device = torch.device('cpu')
     x = torch.load(s, map_location=device)
 
@@ -644,7 +644,7 @@ def create_backbone(f='weights/best.pt', s='weights/backbone.pt'):  # from utils
     for p in x['model'].parameters():
         p.requires_grad = True
     torch.save(x, s)
-    print('%s modified for backbone use and saved as %s' % (f, s))
+    print('%s saved as pretrained checkpoint %s' % (f, s))
 
 
 def coco_class_count(path='../coco/labels/train2014/'):

From b8557f87e3e983ef541fa34a245a16fc30e98455 Mon Sep 17 00:00:00 2001
From: Glenn Jocher <glenn.jocher@ultralytics.com>
Date: Wed, 24 Jun 2020 13:02:27 -0700
Subject: [PATCH 58/73] add stride to datasets.py

---
 test.py           | 1 +
 train.py          | 6 ++++--
 utils/datasets.py | 4 ++--
 3 files changed, 7 insertions(+), 4 deletions(-)

diff --git a/test.py b/test.py
index b3ff968..ae6451f 100644
--- a/test.py
+++ b/test.py
@@ -73,6 +73,7 @@ def test(data,
                                       batch_size,
                                       rect=True,  # rectangular inference
                                       single_cls=opt.single_cls,  # single class mode
+                                      stride=int(max(model.stride)),  # model stride
                                       pad=0.5)  # padding
         batch_size = min(batch_size, len(dataset))
         nw = min([os.cpu_count(), batch_size if batch_size > 1 else 0, 8])  # number of workers
diff --git a/train.py b/train.py
index 2dfe781..d9d2884 100644
--- a/train.py
+++ b/train.py
@@ -160,7 +160,8 @@ def train(hyp):
                                   hyp=hyp,  # augmentation hyperparameters
                                   rect=opt.rect,  # rectangular training
                                   cache_images=opt.cache_images,
-                                  single_cls=opt.single_cls)
+                                  single_cls=opt.single_cls,
+                                  stride=gs)
     mlc = np.concatenate(dataset.labels, 0)[:, 0].max()  # max label class
     assert mlc < nc, 'Label class %g exceeds nc=%g in %s. Correct your labels or your model.' % (mlc, nc, opt.cfg)
 
@@ -179,7 +180,8 @@ def train(hyp):
                                                                  hyp=hyp,
                                                                  rect=True,
                                                                  cache_images=opt.cache_images,
-                                                                 single_cls=opt.single_cls),
+                                                                 single_cls=opt.single_cls,
+                                                                 stride=gs),
                                              batch_size=batch_size,
                                              num_workers=nw,
                                              pin_memory=True,
diff --git a/utils/datasets.py b/utils/datasets.py
index 65e642d..37d773e 100755
--- a/utils/datasets.py
+++ b/utils/datasets.py
@@ -258,7 +258,7 @@ class LoadStreams:  # multiple IP or RTSP cameras
 
 class LoadImagesAndLabels(Dataset):  # for training/testing
     def __init__(self, path, img_size=416, batch_size=16, augment=False, hyp=None, rect=False, image_weights=False,
-                 cache_images=False, single_cls=False, pad=0.0):
+                 cache_images=False, single_cls=False, stride=32, pad=0.0):
         try:
             path = str(Path(path))  # os-agnostic
             parent = str(Path(path).parent) + os.sep
@@ -325,7 +325,7 @@ class LoadImagesAndLabels(Dataset):  # for training/testing
                 elif mini > 1:
                     shapes[i] = [1, 1 / mini]
 
-            self.batch_shapes = np.ceil(np.array(shapes) * img_size / 32. + pad).astype(np.int) * 32
+            self.batch_shapes = np.ceil(np.array(shapes) * img_size / stride + pad).astype(np.int) * stride
 
         # Cache labels
         self.imgs = [None] * n

From e071b8dd3a68b5ebc25361f0aac290b295f99796 Mon Sep 17 00:00:00 2001
From: Glenn Jocher <glenn.jocher@ultralytics.com>
Date: Wed, 24 Jun 2020 13:12:23 -0700
Subject: [PATCH 59/73] update models/common.py for Conv() flexible padding

---
 models/common.py | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/models/common.py b/models/common.py
index 10f6fd6..3c4a0d7 100644
--- a/models/common.py
+++ b/models/common.py
@@ -13,7 +13,8 @@ class Conv(nn.Module):
     # Standard convolution
     def __init__(self, c1, c2, k=1, s=1, g=1, act=True):  # ch_in, ch_out, kernel, stride, groups
         super(Conv, self).__init__()
-        self.conv = nn.Conv2d(c1, c2, k, s, k // 2, groups=g, bias=False)
+        p = k // 2 if isinstance(k, int) else [x // 2 for x in k]  # padding
+        self.conv = nn.Conv2d(c1, c2, k, s, p, groups=g, bias=False)
         self.bn = nn.BatchNorm2d(c2)
         self.act = nn.LeakyReLU(0.1, inplace=True) if act else nn.Identity()
 

From 227aa735dccea28f44bc069b24a5a6356e9a707a Mon Sep 17 00:00:00 2001
From: Glenn Jocher <glenn.jocher@ultralytics.com>
Date: Wed, 24 Jun 2020 13:20:14 -0700
Subject: [PATCH 60/73] updates

---
 detect.py | 5 +++++
 test.py   | 2 +-
 2 files changed, 6 insertions(+), 1 deletion(-)

diff --git a/detect.py b/detect.py
index e4a7ddb..2fb96ea 100644
--- a/detect.py
+++ b/detect.py
@@ -156,3 +156,8 @@ if __name__ == '__main__':
 
     with torch.no_grad():
         detect()
+
+        # Update all models
+        # for opt.weights in ['yolov5s.pt', 'yolov5m.pt', 'yolov5l.pt', 'yolov5x.pt', 'yolov3-spp.pt']:
+        #    detect()
+        #    create_pretrained(opt.weights, opt.weights)
diff --git a/test.py b/test.py
index ae6451f..04e1983 100644
--- a/test.py
+++ b/test.py
@@ -246,7 +246,7 @@ def test(data,
 if __name__ == '__main__':
     parser = argparse.ArgumentParser(prog='test.py')
     parser.add_argument('--weights', type=str, default='weights/yolov5s.pt', help='model.pt path')
-    parser.add_argument('--data', type=str, default='data/coco.yaml', help='*.data path')
+    parser.add_argument('--data', type=str, default='data/coco128.yaml', help='*.data path')
     parser.add_argument('--batch-size', type=int, default=32, help='size of each image batch')
     parser.add_argument('--img-size', type=int, default=640, help='inference size (pixels)')
     parser.add_argument('--conf-thres', type=float, default=0.001, help='object confidence threshold')

From 9a9c4f1259c91ea6495095ce17d71f4ffe8a4304 Mon Sep 17 00:00:00 2001
From: Glenn Jocher <glenn.jocher@ultralytics.com>
Date: Wed, 24 Jun 2020 18:00:43 -0700
Subject: [PATCH 61/73] update yolo.py default --cfg

---
 models/yolo.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/models/yolo.py b/models/yolo.py
index 2d853cb..c9e6c49 100644
--- a/models/yolo.py
+++ b/models/yolo.py
@@ -207,7 +207,7 @@ def parse_model(md, ch):  # model_dict, input_channels(3)
 
 if __name__ == '__main__':
     parser = argparse.ArgumentParser()
-    parser.add_argument('--cfg', type=str, default='yolov5m-p6.yaml', help='model.yaml')
+    parser.add_argument('--cfg', type=str, default='yolov5s.yaml', help='model.yaml')
     parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu')
     opt = parser.parse_args()
     opt.cfg = check_file(opt.cfg)  # check file

From b50fdf16aff8311a8dba3ab8efdf2fd1b15e0d6e Mon Sep 17 00:00:00 2001
From: Glenn Jocher <glenn.jocher@ultralytics.com>
Date: Wed, 24 Jun 2020 22:22:13 -0700
Subject: [PATCH 62/73] model.names multi-GPU bug fix #94

---
 detect.py | 2 +-
 train.py  | 2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)

diff --git a/detect.py b/detect.py
index 2fb96ea..9d922ae 100644
--- a/detect.py
+++ b/detect.py
@@ -46,7 +46,7 @@ def detect(save_img=False):
         dataset = LoadImages(source, img_size=imgsz)
 
     # Get names and colors
-    names = model.names if hasattr(model, 'names') else model.modules.names
+    names = model.module.names if hasattr(model, 'module') else model.names
     colors = [[random.randint(0, 255) for _ in range(3)] for _ in range(len(names))]
 
     # Run inference
diff --git a/train.py b/train.py
index d9d2884..ea72f15 100644
--- a/train.py
+++ b/train.py
@@ -79,6 +79,7 @@ def train(hyp):
     # Create model
     model = Model(opt.cfg).to(device)
     assert model.md['nc'] == nc, '%s nc=%g classes but %s nc=%g classes' % (opt.data, nc, opt.cfg, model.md['nc'])
+    model.names = data_dict['names']
 
     # Image sizes
     gs = int(max(model.stride))  # grid size (max stride)
@@ -193,7 +194,6 @@ def train(hyp):
     model.hyp = hyp  # attach hyperparameters to model
     model.gr = 1.0  # giou loss ratio (obj_loss = 1.0 or giou)
     model.class_weights = labels_to_class_weights(dataset.labels, nc).to(device)  # attach class weights
-    model.names = data_dict['names']
 
     # Class frequency
     labels = np.concatenate(dataset.labels, 0)

From b2d43078489cc329d90489ae29e6867713fba82d Mon Sep 17 00:00:00 2001
From: Glenn Jocher <glenn.jocher@ultralytics.com>
Date: Thu, 25 Jun 2020 10:59:42 -0700
Subject: [PATCH 63/73] Update README.md

---
 README.md | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/README.md b/README.md
index 8596da3..12873f8 100755
--- a/README.md
+++ b/README.md
@@ -91,7 +91,7 @@ $ python train.py --data coco.yaml --cfg yolov5s.yaml --weights '' --batch-size
 
 To access an up-to-date working environment (with all dependencies including CUDA/CUDNN, Python and PyTorch preinstalled), consider a:
 
-- **GCP** Deep Learning VM with $300 free credit offer: See our [GCP Quickstart Guide](https://github.com/ultralytics/yolov5/wiki/GCP-Quickstart) 
+- **Google Cloud** Deep Learning VM with $300 free credit offer: See our [GCP Quickstart Guide](https://github.com/ultralytics/yolov5/wiki/GCP-Quickstart) 
 - **Google Colab Notebook** with 12 hours of free GPU time. <a href="https://colab.research.google.com/github/ultralytics/yolov5/blob/master/tutorial.ipynb"><img src="https://colab.research.google.com/assets/colab-badge.svg" alt="Open In Colab"></a>
 - **Docker Image** https://hub.docker.com/r/ultralytics/yolov5. See [Docker Quickstart Guide](https://github.com/ultralytics/yolov5/wiki/Docker-Quickstart) ![Docker Pulls](https://img.shields.io/docker/pulls/ultralytics/yolov5?logo=docker)
 
@@ -104,7 +104,7 @@ To access an up-to-date working environment (with all dependencies including CUD
 ## About Us
 
 Ultralytics is a U.S.-based particle physics and AI startup with over 6 years of expertise supporting government, academic and business clients. We offer a wide range of vision AI services, spanning from simple expert advice up to delivery of fully customized, end-to-end production solutions, including:
-- **Cloud-based AI** surveillance systems operating on **hundreds of HD video streams in realtime.**
+- **Cloud-based AI** systems operating on **hundreds of HD video streams in realtime.**
 - **Edge AI** integrated into custom iOS and Android apps for realtime **30 FPS video inference.**
 - **Custom data training**, hyperparameter evolution, and model exportation to any destination.
 

From 8f7e11dcef21ab39330a4fbc0540f1ada3aacee4 Mon Sep 17 00:00:00 2001
From: Glenn Jocher <glenn.jocher@ultralytics.com>
Date: Thu, 25 Jun 2020 11:53:13 -0700
Subject: [PATCH 64/73] Update README.md

---
 README.md | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/README.md b/README.md
index 12873f8..7eda5f5 100755
--- a/README.md
+++ b/README.md
@@ -9,8 +9,8 @@ This repository represents Ultralytics open-source research into future object d
 - **June 22, 2020**: [PANet](https://arxiv.org/abs/1803.01534) updates: increased layers, reduced parameters, faster inference and improved mAP [364fcfd](https://github.com/ultralytics/yolov5/commit/364fcfd7dba53f46edd4f04c037a039c0a287972).
 - **June 19, 2020**: [FP16](https://pytorch.org/docs/stable/nn.html#torch.nn.Module.half) as new default for smaller checkpoints and faster inference [d4c6674](https://github.com/ultralytics/yolov5/commit/d4c6674c98e19df4c40e33a777610a18d1961145).
 - **June 9, 2020**: [CSP](https://github.com/WongKinYiu/CrossStagePartialNetworks) updates: improved speed, size, and accuracy. Credit to @WongKinYiu for excellent CSP work. 
-- **May 27, 2020**: Public release of repo. YOLOv5 models are SOTA among all known YOLO implementations, YOLOv5 family will be undergoing architecture research and development over Q2/Q3 2020 to increase performance. Updates may include [CSP](https://github.com/WongKinYiu/CrossStagePartialNetworks) bottlenecks, [YOLOv4](https://github.com/AlexeyAB/darknet) features, as well as PANet or BiFPN heads.
-- **April 1, 2020**: Begin development of a 100% PyTorch, scaleable YOLOv3/4-based group of future models, in a range of compound-scaled sizes. Models will be defined by new user-friendly `*.yaml` files. New training methods will be simpler to start, faster to finish, and more robust to training a wider variety of custom dataset.
+- **May 27, 2020**: Public release of repo. YOLOv5 models are SOTA among all known YOLO implementations.
+- **April 1, 2020**: Start development of future [YOLOv3](https://github.com/ultralytics/yolov3)/[YOLOv4](https://github.com/AlexeyAB/darknet)-based PyTorch models in a range of compound-scaled sizes.
 
 
 ## Pretrained Checkpoints

From 8669f4561c14313d38661c03c66798356d1aabff Mon Sep 17 00:00:00 2001
From: Glenn Jocher <glenn.jocher@ultralytics.com>
Date: Thu, 25 Jun 2020 12:26:30 -0700
Subject: [PATCH 65/73] Update README.md

---
 README.md | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/README.md b/README.md
index 7eda5f5..6f0d1fa 100755
--- a/README.md
+++ b/README.md
@@ -98,7 +98,7 @@ To access an up-to-date working environment (with all dependencies including CUD
 
 ## Citation
 
-[![DOI](https://zenodo.org/badge/146165888.svg)](https://zenodo.org/badge/latestdoi/146165888)
+[![DOI](https://zenodo.org/badge/264818686.svg)](https://zenodo.org/badge/latestdoi/264818686)
 
 
 ## About Us

From c910ed397cad0b4ffbf9b736ab605746ee680a4f Mon Sep 17 00:00:00 2001
From: Glenn Jocher <glenn.jocher@ultralytics.com>
Date: Thu, 25 Jun 2020 14:27:45 -0700
Subject: [PATCH 66/73] Update README.md

---
 README.md | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/README.md b/README.md
index 6f0d1fa..1e29d18 100755
--- a/README.md
+++ b/README.md
@@ -6,9 +6,9 @@ This repository represents Ultralytics open-source research into future object d
 
 <img src="https://user-images.githubusercontent.com/26833433/85340570-30360a80-b49b-11ea-87cf-bdf33d53ae15.png" width="1000">** GPU Speed measures end-to-end time per image averaged over 5000 COCO val2017 images using a V100 GPU with batch size 8, and includes image preprocessing, PyTorch FP16 inference, postprocessing and NMS.
 
-- **June 22, 2020**: [PANet](https://arxiv.org/abs/1803.01534) updates: increased layers, reduced parameters, faster inference and improved mAP [364fcfd](https://github.com/ultralytics/yolov5/commit/364fcfd7dba53f46edd4f04c037a039c0a287972).
+- **June 22, 2020**: [PANet](https://arxiv.org/abs/1803.01534) updates: new heads, reduced parameters, faster inference and improved mAP [364fcfd](https://github.com/ultralytics/yolov5/commit/364fcfd7dba53f46edd4f04c037a039c0a287972).
 - **June 19, 2020**: [FP16](https://pytorch.org/docs/stable/nn.html#torch.nn.Module.half) as new default for smaller checkpoints and faster inference [d4c6674](https://github.com/ultralytics/yolov5/commit/d4c6674c98e19df4c40e33a777610a18d1961145).
-- **June 9, 2020**: [CSP](https://github.com/WongKinYiu/CrossStagePartialNetworks) updates: improved speed, size, and accuracy. Credit to @WongKinYiu for excellent CSP work. 
+- **June 9, 2020**: [CSP](https://github.com/WongKinYiu/CrossStagePartialNetworks) updates: improved speed, size, and accuracy (credit to @WongKinYiu for CSP).
 - **May 27, 2020**: Public release of repo. YOLOv5 models are SOTA among all known YOLO implementations.
 - **April 1, 2020**: Start development of future [YOLOv3](https://github.com/ultralytics/yolov3)/[YOLOv4](https://github.com/AlexeyAB/darknet)-based PyTorch models in a range of compound-scaled sizes.
 

From e670a3353b80a238b73157998eeead60c1acdade Mon Sep 17 00:00:00 2001
From: Glenn Jocher <glenn.jocher@ultralytics.com>
Date: Thu, 25 Jun 2020 15:09:04 -0700
Subject: [PATCH 67/73] update test.py for FP16 testing during training

---
 test.py | 15 +++++++--------
 1 file changed, 7 insertions(+), 8 deletions(-)

diff --git a/test.py b/test.py
index 04e1983..8b94f54 100644
--- a/test.py
+++ b/test.py
@@ -25,7 +25,6 @@ def test(data,
     if model is None:
         training = False
         device = torch_utils.select_device(opt.device, batch_size=batch_size)
-        half = device.type != 'cpu'  # half precision only supported on CUDA
 
         # Remove previous
         for f in glob.glob('test_batch*.jpg'):
@@ -37,20 +36,19 @@ def test(data,
         torch_utils.model_info(model)
         model.fuse()
         model.to(device)
-        if half:
-            model.half()  # to FP16
 
-        # Multi-GPU disabled, incompatible with .half()
+        # Multi-GPU disabled, incompatible with .half() https://github.com/ultralytics/yolov5/issues/99
         # if device.type != 'cpu' and torch.cuda.device_count() > 1:
         #     model = nn.DataParallel(model)
 
     else:  # called by train.py
         training = True
         device = next(model.parameters()).device  # get model device
-        # half disabled https://github.com/ultralytics/yolov5/issues/99
-        half = False  # device.type != 'cpu' and torch.cuda.device_count() == 1
-        if half:
-            model.half()  # to FP16
+
+    # Half
+    half = device.type != 'cpu' and torch.cuda.device_count() == 1  # half precision only supported on single-GPU
+    if half:
+        model.half()  # to FP16
 
     # Configure
     model.eval()
@@ -237,6 +235,7 @@ def test(data,
                   'See https://github.com/cocodataset/cocoapi/issues/356')
 
     # Return results
+    model.float()  # for training
     maps = np.zeros(nc) + map
     for i, c in enumerate(ap_class):
         maps[c] = ap[i]

From 256a3e89d2d78ab90eff44a4771de2586f62b15c Mon Sep 17 00:00:00 2001
From: Glenn Jocher <glenn.jocher@ultralytics.com>
Date: Thu, 25 Jun 2020 17:52:56 -0700
Subject: [PATCH 68/73] small dataset bug fix #140

---
 train.py | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/train.py b/train.py
index ea72f15..94139a0 100644
--- a/train.py
+++ b/train.py
@@ -287,10 +287,10 @@ def train(hyp):
 
             # Plot
             if ni < 3:
-                f = 'train_batch%g.jpg' % i  # filename
-                res = plot_images(images=imgs, targets=targets, paths=paths, fname=f)
-                if tb_writer:
-                    tb_writer.add_image(f, res, dataformats='HWC', global_step=epoch)
+                f = 'train_batch%g.jpg' % ni  # filename
+                result = plot_images(images=imgs, targets=targets, paths=paths, fname=f)
+                if tb_writer and result is not None:
+                    tb_writer.add_image(f, result, dataformats='HWC', global_step=epoch)
                     # tb_writer.add_graph(model, imgs)  # add model to tensorboard
 
             # end batch ------------------------------------------------------------------------------------------------

From 68f63616b3e53c1c4156900f239eccba6c8279b7 Mon Sep 17 00:00:00 2001
From: Anish Hiranandani <anish.hira@gmail.com>
Date: Thu, 25 Jun 2020 18:23:36 -0700
Subject: [PATCH 69/73] Fix after suggestion

---
 detect.py | 8 ++++++--
 1 file changed, 6 insertions(+), 2 deletions(-)

diff --git a/detect.py b/detect.py
index e4a7ddb..9c4990e 100644
--- a/detect.py
+++ b/detect.py
@@ -95,8 +95,12 @@ def detect(save_img=False):
                 for *xyxy, conf, cls in det:
                     if save_txt:  # Write to file
                         xywh = (xyxy2xywh(torch.tensor(xyxy).view(1, 4)) / gn).view(-1).tolist()  # normalized xywh
-                        with open(save_path[:save_path.rfind('.')] + '.txt', 'a') as file:
-                            file.write(('%g ' * 5 + '\n') % (cls, *xywh))  # label format
+                        if dataset.frame == 0:
+                            with open(save_path[:save_path.rfind('.')] + '.txt', 'a') as f:
+                                f.write(('%g ' * 5 + '\n') % (cls, *xywh))  # label format
+                        else:
+                            with open(save_path[:save_path.rfind('.')] + '_' + str(dataset.frame) + '.txt', 'a') as f:
+                                f.write(('%g ' * 5 + '\n') % (cls, *xywh))  # label format
 
                     if save_img or view_img:  # Add bbox to image
                         label = '%s %.2f' % (names[int(cls)], conf)

From 496ec33a33dd18de0b8dc14de1e571955acf5135 Mon Sep 17 00:00:00 2001
From: Glenn Jocher <glenn.jocher@ultralytics.com>
Date: Thu, 25 Jun 2020 19:19:15 -0700
Subject: [PATCH 70/73] Update detect.py

Added some recent updates that were missing, and updated the filename with an if else.
---
 detect.py | 16 +++++++++-------
 1 file changed, 9 insertions(+), 7 deletions(-)

diff --git a/detect.py b/detect.py
index 9c4990e..07631dd 100644
--- a/detect.py
+++ b/detect.py
@@ -46,7 +46,7 @@ def detect(save_img=False):
         dataset = LoadImages(source, img_size=imgsz)
 
     # Get names and colors
-    names = model.names if hasattr(model, 'names') else model.modules.names
+    names = model.module.names if hasattr(model, 'module') else model.names
     colors = [[random.randint(0, 255) for _ in range(3)] for _ in range(len(names))]
 
     # Run inference
@@ -80,6 +80,7 @@ def detect(save_img=False):
                 p, s, im0 = path, '', im0s
 
             save_path = str(Path(out) / Path(p).name)
+            txt_path = save_path[:save_path.rfind('.')] + ('_%g' % dataset.frame if dataset.mode == 'video' else '')
             s += '%gx%g ' % img.shape[2:]  # print string
             gn = torch.tensor(im0.shape)[[1, 0, 1, 0]]  #  normalization gain whwh
             if det is not None and len(det):
@@ -95,12 +96,8 @@ def detect(save_img=False):
                 for *xyxy, conf, cls in det:
                     if save_txt:  # Write to file
                         xywh = (xyxy2xywh(torch.tensor(xyxy).view(1, 4)) / gn).view(-1).tolist()  # normalized xywh
-                        if dataset.frame == 0:
-                            with open(save_path[:save_path.rfind('.')] + '.txt', 'a') as f:
-                                f.write(('%g ' * 5 + '\n') % (cls, *xywh))  # label format
-                        else:
-                            with open(save_path[:save_path.rfind('.')] + '_' + str(dataset.frame) + '.txt', 'a') as f:
-                                f.write(('%g ' * 5 + '\n') % (cls, *xywh))  # label format
+                        with open(txt_path + '.txt', 'a') as f:
+                            f.write(('%g ' * 5 + '\n') % (cls, *xywh))  # label format
 
                     if save_img or view_img:  # Add bbox to image
                         label = '%s %.2f' % (names[int(cls)], conf)
@@ -160,3 +157,8 @@ if __name__ == '__main__':
 
     with torch.no_grad():
         detect()
+
+        # Update all models
+        # for opt.weights in ['yolov5s.pt', 'yolov5m.pt', 'yolov5l.pt', 'yolov5x.pt', 'yolov3-spp.pt']:
+        #    detect()
+        #    create_pretrained(opt.weights, opt.weights)

From 8c5b6220f5576b2296e6ca18b041598b3edcad68 Mon Sep 17 00:00:00 2001
From: Glenn Jocher <glenn.jocher@ultralytics.com>
Date: Thu, 25 Jun 2020 19:26:23 -0700
Subject: [PATCH 71/73] Update detect.py

---
 detect.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/detect.py b/detect.py
index 07631dd..69a37de 100644
--- a/detect.py
+++ b/detect.py
@@ -80,7 +80,7 @@ def detect(save_img=False):
                 p, s, im0 = path, '', im0s
 
             save_path = str(Path(out) / Path(p).name)
-            txt_path = save_path[:save_path.rfind('.')] + ('_%g' % dataset.frame if dataset.mode == 'video' else '')
+            txt_path = str(Path(out) / Path(p).stem) + ('_%g' % dataset.frame if dataset.mode == 'video' else '')
             s += '%gx%g ' % img.shape[2:]  # print string
             gn = torch.tensor(im0.shape)[[1, 0, 1, 0]]  #  normalization gain whwh
             if det is not None and len(det):

From 22fb2b0c25e6f60f95e6ef5d8588e586bdbc06f7 Mon Sep 17 00:00:00 2001
From: Glenn Jocher <glenn.jocher@ultralytics.com>
Date: Fri, 26 Jun 2020 18:56:13 -0700
Subject: [PATCH 72/73] refactor dataloader

---
 test.py           | 22 +++-------------------
 train.py          | 35 ++++++-----------------------------
 utils/datasets.py | 22 +++++++++++++++++++++-
 3 files changed, 30 insertions(+), 49 deletions(-)

diff --git a/test.py b/test.py
index 8b94f54..c0bda5f 100644
--- a/test.py
+++ b/test.py
@@ -1,8 +1,6 @@
 import argparse
 import json
 
-from torch.utils.data import DataLoader
-
 from utils import google_utils
 from utils.datasets import *
 from utils.utils import *
@@ -56,30 +54,16 @@ def test(data,
         data = yaml.load(f, Loader=yaml.FullLoader)  # model dict
     nc = 1 if single_cls else int(data['nc'])  # number of classes
     iouv = torch.linspace(0.5, 0.95, 10).to(device)  # iou vector for mAP@0.5:0.95
-    # iouv = iouv[0].view(1)  # comment for mAP@0.5:0.95
     niou = iouv.numel()
 
     # Dataloader
     if dataloader is None:  # not training
+        merge = opt.merge  # use Merge NMS
         img = torch.zeros((1, 3, imgsz, imgsz), device=device)  # init img
         _ = model(img.half() if half else img) if device.type != 'cpu' else None  # run once
-
-        merge = opt.merge  # use Merge NMS
         path = data['test'] if opt.task == 'test' else data['val']  # path to val/test images
-        dataset = LoadImagesAndLabels(path,
-                                      imgsz,
-                                      batch_size,
-                                      rect=True,  # rectangular inference
-                                      single_cls=opt.single_cls,  # single class mode
-                                      stride=int(max(model.stride)),  # model stride
-                                      pad=0.5)  # padding
-        batch_size = min(batch_size, len(dataset))
-        nw = min([os.cpu_count(), batch_size if batch_size > 1 else 0, 8])  # number of workers
-        dataloader = DataLoader(dataset,
-                                batch_size=batch_size,
-                                num_workers=nw,
-                                pin_memory=True,
-                                collate_fn=dataset.collate_fn)
+        dataloader = create_dataloader(path, imgsz, batch_size, int(max(model.stride)), opt,
+                                       hyp=None, augment=False, cache=False, pad=0.5, rect=True)[0]
 
     seen = 0
     names = model.names if hasattr(model, 'names') else model.module.names
diff --git a/train.py b/train.py
index 94139a0..4238713 100644
--- a/train.py
+++ b/train.py
@@ -155,38 +155,15 @@ def train(hyp):
         model = torch.nn.parallel.DistributedDataParallel(model)
         # pip install torch==1.4.0+cu100 torchvision==0.5.0+cu100 -f https://download.pytorch.org/whl/torch_stable.html
 
-    # Dataset
-    dataset = LoadImagesAndLabels(train_path, imgsz, batch_size,
-                                  augment=True,
-                                  hyp=hyp,  # augmentation hyperparameters
-                                  rect=opt.rect,  # rectangular training
-                                  cache_images=opt.cache_images,
-                                  single_cls=opt.single_cls,
-                                  stride=gs)
+    # Trainloader
+    dataloader, dataset = create_dataloader(train_path, imgsz, batch_size, gs, opt,
+                                            hyp=hyp, augment=True, cache=opt.cache_images, rect=opt.rect)
     mlc = np.concatenate(dataset.labels, 0)[:, 0].max()  # max label class
     assert mlc < nc, 'Label class %g exceeds nc=%g in %s. Correct your labels or your model.' % (mlc, nc, opt.cfg)
 
-    # Dataloader
-    batch_size = min(batch_size, len(dataset))
-    nw = min([os.cpu_count(), batch_size if batch_size > 1 else 0, 8])  # number of workers
-    dataloader = torch.utils.data.DataLoader(dataset,
-                                             batch_size=batch_size,
-                                             num_workers=nw,
-                                             shuffle=not opt.rect,  # Shuffle=True unless rectangular training is used
-                                             pin_memory=True,
-                                             collate_fn=dataset.collate_fn)
-
     # Testloader
-    testloader = torch.utils.data.DataLoader(LoadImagesAndLabels(test_path, imgsz_test, batch_size,
-                                                                 hyp=hyp,
-                                                                 rect=True,
-                                                                 cache_images=opt.cache_images,
-                                                                 single_cls=opt.single_cls,
-                                                                 stride=gs),
-                                             batch_size=batch_size,
-                                             num_workers=nw,
-                                             pin_memory=True,
-                                             collate_fn=dataset.collate_fn)
+    testloader = create_dataloader(test_path, imgsz_test, batch_size, gs, opt,
+                                            hyp=hyp, augment=False, cache=opt.cache_images, rect=True)[0]
 
     # Model parameters
     hyp['cls'] *= nc / 80.  # scale coco-tuned hyp['cls'] to current dataset
@@ -218,7 +195,7 @@ def train(hyp):
     maps = np.zeros(nc)  # mAP per class
     results = (0, 0, 0, 0, 0, 0, 0)  # 'P', 'R', 'mAP', 'F1', 'val GIoU', 'val Objectness', 'val Classification'
     print('Image sizes %g train, %g test' % (imgsz, imgsz_test))
-    print('Using %g dataloader workers' % nw)
+    print('Using %g dataloader workers' % dataloader.num_workers)
     print('Starting training for %g epochs...' % epochs)
     # torch.autograd.set_detect_anomaly(True)
     for epoch in range(start_epoch, epochs):  # epoch ------------------------------------------------------------------
diff --git a/utils/datasets.py b/utils/datasets.py
index 37d773e..00f2338 100755
--- a/utils/datasets.py
+++ b/utils/datasets.py
@@ -41,6 +41,26 @@ def exif_size(img):
     return s
 
 
+def create_dataloader(path, imgsz, batch_size, stride, opt, hyp=None, augment=False, cache=False, pad=0.0, rect=False):
+    dataset = LoadImagesAndLabels(path, imgsz, batch_size,
+                                  augment=augment,  # augment images
+                                  hyp=hyp,  # augmentation hyperparameters
+                                  rect=rect,  # rectangular training
+                                  cache_images=cache,
+                                  single_cls=opt.single_cls,
+                                  stride=stride,
+                                  pad=pad)
+
+    batch_size = min(batch_size, len(dataset))
+    nw = min([os.cpu_count(), batch_size if batch_size > 1 else 0, 0])  # number of workers
+    dataloader = torch.utils.data.DataLoader(dataset,
+                                             batch_size=batch_size,
+                                             num_workers=nw,
+                                             pin_memory=True,
+                                             collate_fn=LoadImagesAndLabels.collate_fn)
+    return dataloader, dataset
+
+
 class LoadImages:  # for inference
     def __init__(self, path, img_size=416):
         path = str(Path(path))  # os-agnostic
@@ -712,7 +732,7 @@ def random_affine(img, targets=(), degrees=10, translate=.1, scale=.1, shear=10,
         area = w * h
         area0 = (targets[:, 3] - targets[:, 1]) * (targets[:, 4] - targets[:, 2])
         ar = np.maximum(w / (h + 1e-16), h / (w + 1e-16))  # aspect ratio
-        i = (w > 4) & (h > 4) & (area / (area0 * s + 1e-16) > 0.2) & (ar < 10)
+        i = (w > 2) & (h > 2) & (area / (area0 * s + 1e-16) > 0.2) & (ar < 20)
 
         targets = targets[i]
         targets[:, 1:5] = xy[i]

From 77fb8ee0823e7a5b99ea7b9a07a439d1e108e30d Mon Sep 17 00:00:00 2001
From: Glenn Jocher <glenn.jocher@ultralytics.com>
Date: Fri, 26 Jun 2020 19:30:09 -0700
Subject: [PATCH 73/73] refactor dataloader

---
 utils/datasets.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/utils/datasets.py b/utils/datasets.py
index 00f2338..aee891c 100755
--- a/utils/datasets.py
+++ b/utils/datasets.py
@@ -52,7 +52,7 @@ def create_dataloader(path, imgsz, batch_size, stride, opt, hyp=None, augment=Fa
                                   pad=pad)
 
     batch_size = min(batch_size, len(dataset))
-    nw = min([os.cpu_count(), batch_size if batch_size > 1 else 0, 0])  # number of workers
+    nw = min([os.cpu_count(), batch_size if batch_size > 1 else 0, 8])  # number of workers
     dataloader = torch.utils.data.DataLoader(dataset,
                                              batch_size=batch_size,
                                              num_workers=nw,