import numpy as np import lqmtest_x3_5_pool import lqmtest_x3_2_load_data import lqmtest_x3_4_conv_proc import lqmtest_x3_6_fullconn import lqmtest_x3_7_nonlinear import lqmtest_x3_8_classify import lqmtest_x3_9_label import lqmtest_x3_10_error class ModelObj: # 网络对象 def __init__(self, ObjID, ObjType, ObjLable, ParaString, ObjX, ObjY): self.ObjID = ObjID # 图元号 self.ObjType = ObjType # 图元类别 self.ObjLable = ObjLable # 对象标签 self.ParaString = ParaString # 参数字符串 self.ObjX = ObjX # 对象位置x坐标 self.ObjY = ObjY # 对象位置y坐标 class AjConv_Class(ModelObj): # 卷积调整对象 def __init__(self, ObjID, ObjType, ObjLable, ParaString, ObjX, ObjY): super().__init__(ObjID, ObjType, ObjLable, ParaString, ObjX, ObjY) self.AjConvProc = self.ajconv_proc # 基本操作函数 self.SetAjConvPara = self.setajconv_para # 参数设置函数 def ajconv_proc(self, images, AjConvPara): kernel_grad_list = [] bias_grad = 0 # 初始化偏置项梯度为零 for c in images: # 计算卷积核和偏置项的梯度 # 初始化卷积核梯度为零矩阵 kernel_grad = np.zeros_like(AjConvPara['kernel_info']) for i in range(AjConvPara['loss'].shape[0]): for j in range(AjConvPara['loss'].shape[1]): # 将输入数据数组中对应的子矩阵旋转180度， # 与误差值相乘，累加到卷积核梯度上 kernel_grad+=np.rot90(c[i:i+kernel_grad.shape[0], j:j + kernel_grad.shape[0]], 2) * AjConvPara['loss'][i, j] # 将误差值累加到偏置项梯度上 bias_grad += AjConvPara['loss'][i, j] kernel_grad_list.append(kernel_grad) # 使用stack函数沿着第0个轴把一百个a数组堆叠起来 result = np.stack(kernel_grad_list, axis=0) # 沿着第0个维度求和 kernel_grad = np.sum(result, axis=0) / len(images) # 更新卷积核和偏置项参数 kernel = AjConvPara['kernel_info']-AjConvPara['learning_rate' ] * kernel_grad # 卷积核参数减去学习率乘以卷积核梯度 return kernel # 返回更新后的卷积核 def setajconv_para(self, loss, ConvPara): kernel = ConvPara['kernel'] # 卷积核信息 learning_rate = float(input("请输入卷积调整的学习率: ")) # 0.01 学习率 loss = np.array([[loss]]) AjConvPara = {'kernel_info': kernel, 'learning_rate': learning_rate, 'loss': loss} return AjConvPara if __name__ == '__main__': DataSet = lqmtest_x3_2_load_data.Data_Class("DataSet1", 1, "数据集1", [], 120, 330) # setload_data()函数，获取加载数据集的参数 DataPara = DataSet.SetDataPara() train_images, test_images = DataSet.LoadData(DataPara) Conv = lqmtest_x3_4_conv_proc.Conv_Class("Conv1", 2, "卷积1", [], 250, 330) ConvPara = Conv.SetConvPara() for i in range(len(train_images) // 32): images = train_images[i * 32:(i + 1) * 32] conv_images = [] # 存储卷积处理后的图片的列表 for image in images: # 获取训练集的图片数据 dim = len(image.shape) # 获取矩阵的维度 if dim == 2: # 如果是二维矩阵,则转化为三维矩阵 image_h, image_w = image.shape image = np.reshape(image, (1, image_h, image_w)) # 调用ConvProc()函数，根据ConvPara参数完成卷积计算 output = Conv.ConvProc(image, ConvPara) conv_images.append(output) # 将卷积结果存储到列表 elif dim == 3: # 若为三维矩阵，则保持不变直接卷积处理 output = Conv.ConvProc(image, ConvPara) conv_images.append(output) # 将卷积处理后的图片列表转换为数组形式，方便后续处理 conv_images = np.array(conv_images) Pool = lqmtest_x3_5_pool.Pool_Class("Pool1", 3, "最大池化1", [], 380, 330) PoolPara = Pool.SetPollPara() pool_images = [] # 存储池化处理后的图片的列表 for image in conv_images: # 获取卷积后的图片数据 output = Pool.MaxPoolProc(image, PoolPara) pool_images.append(output) # 将池化结果存储到列表 # 将池化处理后的图片列表转换为数组形式，方便后续处理 pool_images = np.array(pool_images) _, _, poolH, poolW = pool_images.shape FullConn = lqmtest_x3_6_fullconn.FullConn_Class("FullConn1", 4, "全连接1", [], 510, 330) FullConnPara = FullConn.SetFullConnPara(poolH, poolW) fullconn_images = [] # 存储全连接处理后的图片的列表 for image in pool_images: # 获取池化后的图片数据 output = FullConn.FullConnProc(image, FullConnPara) fullconn_images.append(output) # 将全连接处理后的结果存储到列表 # 将全连接处理后的图片列表转换为数组形式，方便后续处理 fullconn_images = np.array(fullconn_images) Nonline = lqmtest_x3_7_nonlinear.Nonline_Class("Nonline1", 5, "非线性函数1", [], 640, 330) NonLPara = Nonline.SetNonLPara() # 存储非线性处理后的图片的列表 nonlinear_images = [] for image in fullconn_images: # 获取全连接处理后的图片数据 output = Nonline.NonlinearProc(image, NonLPara) # 将非线性处理后的结果存储到列表 nonlinear_images.append(output) # 将非线性处理后的图片列表转换为数组形式，方便后续处理 nonlinear_images = np.array(nonlinear_images) Classifier=lqmtest_x3_8_classify.Classifier_Class("Classifier1", 6, "分类1", [], 780, 330) ClassifyPara = Classifier.SetClassifyPara() classifier_images = [] # 存储分类处理后的图片的列表 prob_images = [] # 存储分类处理后的概率向量 for image in nonlinear_images: # 获取非线性处理后的图片数据 # 调用softmax函数，得到概率分布向量 prob = Classifier.softmax(image) prob_images.append(prob) # 将概率向量结果存储到列表 output = Classifier.ClassifierProc(image, ClassifyPara) classifier_images.append(output) # 将分类结果存储到列表 # 将分类的结果列表转换为数组形式，方便后续处理 classifier_images = np.array(classifier_images) LabelPara = lqmtest_x3_9_label.Label() label_dict = LabelPara.setlabel_para() right_label = LabelPara.label_array(i) labeled_samples = LabelPara.label_proc(images, right_label, label_dict) Error = lqmtest_x3_10_error.Error_Class("Error1", 7, "误差计算1", [], 710, 124) ErrorPara = Error.SetErrorPara() # 输入值是一个概率矩阵，真实标签值是一个类别列表 prob_images = np.squeeze(prob_images) loss = Error.ErrorProc(prob_images, right_label, ErrorPara) AjConv = AjConv_Class("AjConv1", 8, "卷积调整1", [], 250, 70) AjConvPara = AjConv.SetAjConvPara(loss, ConvPara) ConvPara['kernel'] = AjConv.AjConvProc(images, AjConvPara)