diff --git a/code/.idea/vcs.xml b/code/.idea/vcs.xml new file mode 100644 index 0000000..2e3f692 --- /dev/null +++ b/code/.idea/vcs.xml @@ -0,0 +1,6 @@ + + + + + + \ No newline at end of file diff --git a/code/1.jpg b/code/1.jpg index 71fe1bb..17d3631 100644 Binary files a/code/1.jpg and b/code/1.jpg differ diff --git a/code/__pycache__/mnist_cnn.cpython-37.pyc b/code/__pycache__/mnist_cnn.cpython-37.pyc index e81c16b..72beae9 100644 Binary files a/code/__pycache__/mnist_cnn.cpython-37.pyc and b/code/__pycache__/mnist_cnn.cpython-37.pyc differ diff --git a/code/__pycache__/mnist_cnn3.cpython-37.pyc b/code/__pycache__/mnist_cnn3.cpython-37.pyc index 7c3dc0d..62496a3 100644 Binary files a/code/__pycache__/mnist_cnn3.cpython-37.pyc and b/code/__pycache__/mnist_cnn3.cpython-37.pyc differ diff --git a/code/__pycache__/mnist_dense.cpython-37.pyc b/code/__pycache__/mnist_dense.cpython-37.pyc index 7c73fa5..9c9d7fb 100644 Binary files a/code/__pycache__/mnist_dense.cpython-37.pyc and b/code/__pycache__/mnist_dense.cpython-37.pyc differ diff --git a/code/__pycache__/read_image.cpython-37.pyc b/code/__pycache__/read_image.cpython-37.pyc index 9fa8fbc..6e562c2 100644 Binary files a/code/__pycache__/read_image.cpython-37.pyc and b/code/__pycache__/read_image.cpython-37.pyc differ diff --git a/code/checkpoint/checkpoint b/code/checkpoint/checkpoint index 9beb43e..65673c9 100644 --- a/code/checkpoint/checkpoint +++ b/code/checkpoint/checkpoint @@ -1,2 +1,2 @@ -model_checkpoint_path: "mnist.ckpt" -all_model_checkpoint_paths: "mnist.ckpt" +model_checkpoint_path: "mnist_cnn3.ckpt" +all_model_checkpoint_paths: "mnist_cnn3.ckpt" diff --git a/code/checkpoint/mnist.ckpt.data-00000-of-00001 b/code/checkpoint/mnist.ckpt.data-00000-of-00001 index 23d9b55..f87c7b0 100644 Binary files a/code/checkpoint/mnist.ckpt.data-00000-of-00001 and b/code/checkpoint/mnist.ckpt.data-00000-of-00001 differ diff --git a/code/checkpoint/mnist.ckpt.index b/code/checkpoint/mnist.ckpt.index index 2d65419..b19be27 100644 Binary files a/code/checkpoint/mnist.ckpt.index and b/code/checkpoint/mnist.ckpt.index differ diff --git a/code/checkpoint/mnist_cnn1.ckpt.data-00000-of-00001 b/code/checkpoint/mnist_cnn1.ckpt.data-00000-of-00001 index bdb4112..dd33a9d 100644 Binary files a/code/checkpoint/mnist_cnn1.ckpt.data-00000-of-00001 and b/code/checkpoint/mnist_cnn1.ckpt.data-00000-of-00001 differ diff --git a/code/checkpoint/mnist_cnn1.ckpt.index b/code/checkpoint/mnist_cnn1.ckpt.index index bbc3924..1004637 100644 Binary files a/code/checkpoint/mnist_cnn1.ckpt.index and b/code/checkpoint/mnist_cnn1.ckpt.index differ diff --git a/code/checkpoint/mnist_cnn3.ckpt.data-00000-of-00001 b/code/checkpoint/mnist_cnn3.ckpt.data-00000-of-00001 index 32d49d1..3ab49d3 100644 Binary files a/code/checkpoint/mnist_cnn3.ckpt.data-00000-of-00001 and b/code/checkpoint/mnist_cnn3.ckpt.data-00000-of-00001 differ diff --git a/code/checkpoint/mnist_cnn3.ckpt.index b/code/checkpoint/mnist_cnn3.ckpt.index index 6baf7dc..1bea2ad 100644 Binary files a/code/checkpoint/mnist_cnn3.ckpt.index and b/code/checkpoint/mnist_cnn3.ckpt.index differ diff --git a/code/main.py b/code/main.py index cd820fd..a854c03 100644 --- a/code/main.py +++ b/code/main.py @@ -51,7 +51,6 @@ def screenshot(*args): im.save('1.jpg') print("在使用%d模型"%model) result = read_image.predeiction('1.jpg', model) - text.set(str(result)) # 此函数用于清空画布 diff --git a/code/mnist_cnn.py b/code/mnist_cnn.py index 4a944ac..ef26d0a 100644 --- a/code/mnist_cnn.py +++ b/code/mnist_cnn.py @@ -1,52 +1,82 @@ import tensorflow as tf +import tensorflow.keras as keras import os from matplotlib import pyplot as plt from PIL import Image import numpy as np -mnist = tf.keras.datasets.mnist - +mnist = keras.datasets.mnist +# 载入mnist数据集 (x_train, y_train), (x_test, y_test) = mnist.load_data() -x_train, x_test = x_train/255.0, x_test/255.0 + +# 归一化处理样本，把每个像素点的数据范围由[0-255]转为[0-1] +x_train = x_train/255.0 +x_test = x_test/255.0 + +# 增加一个维度 +x_train = tf.expand_dims(x_train, -1) +x_test = tf.expand_dims(x_test, -1) def creat_model(): - model = tf.keras.models.Sequential([ - tf.keras.layers.Flatten(), - tf.keras.layers.Dense(128, activation='relu'), - tf.keras.layers.Dense(10, activation='softmax') + + model = keras.Sequential([ + # 卷积层定义 + keras.layers.Conv2D(32, 3, padding="SAME"), + keras.layers.BatchNormalization(), + keras.layers.Activation("relu"), + keras.layers.MaxPool2D(pool_size=(2, 2), strides=2), + keras.layers.Dropout(0.2), + # 分类器定义 + keras.layers.Flatten(), + keras.layers.Dense(128, activation="relu"), + keras.layers.Dense(10, activation="softmax") ]) - model.compile(optimizer='adam', - loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=False), + model.compile(optimizer='adam', # 优化方法选用adam + loss=keras.losses.SparseCategoricalCrossentropy(from_logits=False), + # 指定每个批次训练误差的减小方法 metrics=["sparse_categorical_accuracy"]) + # 评价函数 return model -def model_fit(model,check_save_path): + +def eva_acc(str, model): + # 加载测试集判断准确率 + model.load_weights(str) + final_loss, final_acc = model.evaluate(x_test, y_test, verbose=2) + print("Model accuracy: ", final_acc, ", model loss: ", final_loss) + + + +def model_fit(model, check_save_path): if os.path.exists(check_save_path+'.index'): print("load modals...") model.load_weights(check_save_path) - cp_callback = tf.keras.callbacks.ModelCheckpoint(filepath=check_save_path, save_weights_only=True, save_best_only=True) + ES_callback = keras.callbacks.EarlyStopping( + monitor="val_sparse_categorical_accuracy", # 数据的监视入口 + min_delta=0.001, # 增大或减小的阈值，只有大于这个部分才算作improvement。 + patience=1, # 能够容忍多少个epo + # ch内都没有improvement。 + mode='max' # ’auto’, ‘min’, ‘,max’三个可能。如果知道是要上升还是下降，建议设置一下。 + ) - history = model.fit(x_train, y_train, batch_size=32, epochs=5, validation_split=0.15, validation_freq=1, callbacks=[cp_callback]) - model.summary() - - final_loss, final_acc = model.evaluate(x_test, y_test, verbose=2) - print("Model accuracy: ", final_acc, ", model loss: ", final_loss) + history = model.fit(x_train, y_train, batch_size=32, epochs=5, validation_split=0.15, validation_freq=1, callbacks=[cp_callback, ES_callback]) + # 输出参数计算 + model.summary() acc = history.history['sparse_categorical_accuracy'] val_acc = history.history['val_sparse_categorical_accuracy'] loss = history.history['loss'] val_loss = history.history['val_loss'] - + # 图形化 plt.subplot(1, 2, 1) plt.plot(acc, label='Training Accuracy') plt.plot(val_acc, label='Validation Accuracy') plt.title('Training and Validation Accuracy') plt.legend() - plt.subplot(1, 2, 2) plt.plot(loss, label='Training Loss') plt.plot(val_loss, label='Validation Loss') @@ -54,13 +84,9 @@ def model_fit(model,check_save_path): plt.legend() plt.show() -def eva_acc(str, model): - model.load_weights(str) - final_loss, final_acc = model.evaluate(x_test, y_test, verbose=2) - print("Model accuracy: ", final_acc, ", model loss: ", final_loss) if __name__=="__main__": - check_save_path = "./checkpoint/mnist.ckpt" + check_save_path = "./checkpoint/mnist_cnn1.ckpt" model = creat_model() model_fit(model, check_save_path) eva_acc(check_save_path, model) diff --git a/code/mnist_cnn3.py b/code/mnist_cnn3.py index cc10dfc..178e14d 100644 --- a/code/mnist_cnn3.py +++ b/code/mnist_cnn3.py @@ -5,14 +5,16 @@ import tensorflow as tf import matplotlib.pyplot as plt mnist = keras.datasets.mnist (x_train, y_train), (x_test, y_test) = mnist.load_data() +# 数据归一化 x_train = x_train/255.0 x_test = x_test/255.0 +# 为了卷积层输入格式，增加一个维度 x_train = tf.expand_dims(x_train, -1) x_test = tf.expand_dims(x_test, -1) -print("train shape:", x_train.shape) -print("test shape:", x_test.shape) +# print("train shape:", x_train.shape) +# print("test shape:", x_test.shape) # 使用此类进行图形增强 datagen = keras.preprocessing.image.ImageDataGenerator( @@ -24,6 +26,7 @@ datagen = keras.preprocessing.image.ImageDataGenerator( horizontal_flip=False # 布尔值,随机水平翻转。 ) + train_datagen = datagen.flow( x_train, y_train, @@ -38,20 +41,31 @@ validation_genetor = datagen.flow( subset="validation" ) def creat_model(): + """ + 此函数用于构建模型。 + 调用keras.Sequential()建立模型的框架。 + 使用compile()配置神经网络训练方法 + :return: + """ model = keras.Sequential([ keras.layers.Reshape((28, 28, 1)), + # 卷积层定义 keras.layers.Conv2D(filters=32, kernel_size=(5, 5), activation="relu", padding="same", input_shape=(28, 28, 1)), + # 池化层定义 keras.layers.MaxPool2D((2, 2)), + # 卷积层、池化层 keras.layers.Conv2D(filters=64, kernel_size=(3, 3), activation="relu", padding="same"), keras.layers.Conv2D(filters=64, kernel_size=(3, 3), activation="relu", padding="same"), keras.layers.MaxPool2D((2, 2)), + # 卷积层、池化层 keras.layers.Conv2D(filters=128, kernel_size=(3, 3), activation="relu", padding="same"), keras.layers.Conv2D(filters=128, kernel_size=(3, 3), activation="relu", padding="same"), keras.layers.MaxPool2D((2, 2)), + # 分类器：多个全连接层 keras.layers.Flatten(), keras.layers.Dense(512, activation="sigmoid"), keras.layers.Dropout(0.25), @@ -65,29 +79,53 @@ def creat_model(): keras.layers.Dense(10, activation="sigmoid") ]) + # 使用compile()配置神经网络训练方法 + # model.compile(optimizer=优化器, + # loss=损失函数, + # metrics=["评测指标"]) model.compile(optimizer='adam', loss=keras.losses.SparseCategoricalCrossentropy(from_logits=False), metrics=["sparse_categorical_accuracy"]) return model def model_fit(model, check_save_path): + +# 判断是否有保存的模型，有就加载之 if os.path.exists(check_save_path+'.index'): print("load modals...") model.load_weights(check_save_path) + # 使用callbacks.ModelCheckpoint()保存模型 + # tf.keras.callbacks.ModelCheckpoint(filepath=保存模型的路径, + # save_weights_only=若设置为True则仅保存模型权重,设置为False则保存整个模型 + # save_best_only=若设置为True将只保存在验证集上性能最好的模型 + # ) cp_callback = tf.keras.callbacks.ModelCheckpoint(filepath=check_save_path, save_weights_only=True, save_best_only=True) + # 用于动态调整学习率，参数如下： + # monitor：监测的值，可以是accuracy，val_loss,val_accuracy + # factor：缩放学习率的值，学习率将以lr = lr*factor的形式被减少 + # patience：当patience个epoch过去而模型性能不提升时，学习率减少的动作会被触发 + # min_lr：学习率最小值，能缩小到的下限 reduce_lr = tf.keras.callbacks.ReduceLROnPlateau(monitor='val_loss', factor=0.1, - patience=5, + patience=2, min_lr=0.000001, verbose=1) - - history = model.fit(train_datagen, epochs=1, validation_data=validation_genetor, callbacks=[reduce_lr,cp_callback],verbose=1) + # 早停机制 + # monitor：要监测的数量 + # min_delta：在被监测的数据中被认为是提升的最小变化，即绝对变化小于min_delta，将被视为没有提升。 + # patience：没有进步的训练轮数，在这之后训练就会被停止。 + earlystop_callback = keras.callbacks.EarlyStopping( + monitor='val_sparse_categorical_accuracy', + min_delta=0.0001, + patience=2) + history = model.fit(train_datagen, epochs=10, validation_data=validation_genetor, callbacks=[reduce_lr, cp_callback, earlystop_callback], verbose=1) model.summary() + # 用于可视化 acc = history.history['sparse_categorical_accuracy'] val_acc = history.history['val_sparse_categorical_accuracy'] loss = history.history['loss'] @@ -106,7 +144,7 @@ def model_fit(model, check_save_path): plt.legend() plt.show() - +# 此函数通过测试集检测准确率 def model_valtest(model, check_save_path): model.load_weights(check_save_path) final_loss, final_acc = model.evaluate(x_test, y_test, verbose=2) @@ -116,5 +154,5 @@ if __name__ == "__main__": check_save_path = "./checkpoint/mnist_cnn3.ckpt" model = creat_model() - # model_fit(model, check_save_path) + #model_fit(model, check_save_path) model_valtest(model, check_save_path) \ No newline at end of file