基于python实现的CNN卷积神经网络手写数字识别项目源码（高分项目）.zip

from abc import ABC, abstractmethod import numpy as np from method import Xavier, He from tools import * class LayerBase(ABC): """ @抽象基类，所有层需要重载覆盖父类函数才可以定义 @需要覆盖的函数: 前向传播 forward(self, x) 后向传播 backward(self, dout) 参数更新，传进参数更新的对象，例如SGD,AdaGrad update(self, learning_object) 参数保存 saveParams(self) 参数加载 loadParams(self) """ def __init__(self) -> None: pass @abstractmethod def forward(self, x): pass @abstractmethod def backward(self, dout): pass def __call__(self, x): return self.forward(x) @abstractmethod def update(self, learning_object): pass @abstractmethod def saveParams(self): pass @abstractmethod def loadParams(self, layer_params): pass class AffineLayer(LayerBase): """ 全连接层 @params output_size: 输出size method: 参数初始化方法 """ def __init__(self, output_size, method) -> None: self.method = method self.output_size = output_size self.is_init_params = False # 参数是否已经初始化 # 梯度 self.db = None self.dW = None # 参数的形状会根据x以及输出形状动态生成 def ininParams(self, x, output_size, method): self.init_method = eval(method)() self.w, self.b = self.init_method.getAffineParams( x.shape, output_size) self.is_init_params = True def forward(self, x): # 第一次调用需要初始化参数 if not self.is_init_params: self.ininParams(x, self.output_size, self.method) # 保存x形状，方便反向传播 self.x_shape = x.shape # 将高维度的数据转为二维 x = x.reshape(x.shape[0], -1) self.x = x # 计算点乘 return np.dot(self.x, self.w)+self.b def backward(self, dout): self.dW = np.dot(self.x.T, dout) self.db = np.sum(dout, axis=0) # 转为传进前向传播参数的形状 return np.dot(dout, self.w.T).reshape(self.x_shape) # dx # 根据更新方法进行动态调用函数 def update(self, learning_object: object): self.w = learning_object(self.w, self.dW, self.__str__()+"w") self.b = learning_object(self.b, self.db, self.__str__()+"b") # 参数保存 def saveParams(self): params_dict = {} layer_params = {} layer_params['w'] = self.w layer_params['b'] = self.b layer_params['method'] = self.method layer_params['output_size'] = self.output_size params_dict[self.__class__.__name__] = layer_params return params_dict # 参数加载 def loadParams(self, layer_params): self.w = layer_params['w'] self.b = layer_params['b'] self.is_init_params = True class ReluLayer(LayerBase): """ Relu层 """ def __init__(self) -> None: self.mask = None def forward(self, x): self.mask = (x <= 0) x[self.mask] = 0 return x def backward(self, dout): dout[self.mask] = 0 return dout def update(self, learning_object: object): pass def saveParams(self): params_dict = {} params_dict[self.__class__.__name__] = {} return params_dict def loadParams(self, layer_params): pass class SoftMaxLayer(LayerBase): """ SoftMax层 """ # 要处理溢出问题 def forward(self, x): if x.ndim == 2: x = x.T x = x - np.max(x, axis=0) y = np.exp(x) / np.sum(np.exp(x), axis=0) return y.T x = x - np.max(x) # 溢出对策 return np.exp(x) / np.sum(np.exp(x)) def backward(self): pass def update(self, learning_object: object): pass def saveParams(self): pass def loadParams(self, layer_params): pass class SoftMaxLossLayer(LayerBase): """ SoftMax + 交叉熵 层 """ def __init__(self) -> None: self.train_result = None self.target = None self.softmax_layer = SoftMaxLayer() def cross_entropy_error(self, y, t): if y.ndim == 1: t = t.reshape(1, t.size) y = y.reshape(1, y.size) # 监督数据是one-hot-vector的情况下，转换为正确解标签的索引 if t.size == y.size: t = t.argmax(axis=1) batch_size = y.shape[0] return -np.sum(np.log(y[np.arange(batch_size), t] + 1e-7)) / batch_size def forward(self, x): self.train_result = self.softmax_layer(x) self.loss = self.cross_entropy_error(self.train_result, self.target) return self.loss def backward(self, dout): batch_size = self.target.shape[0] if self.target.size == self.train_result.size: # 监督数据是one-hot-vector的情况 dx = (self.train_result - self.target) / batch_size else: dx = self.train_result.copy() dx[np.arange(batch_size), self.target] -= 1 dx = dx / batch_size return dx def __call__(self, x, t): self.target = t return self.forward(x) def update(self, learning_object: object): pass def saveParams(self): params_dict = {} layer_params = {} layer_params['is_last_layer'] = 1 params_dict[self.__class__.__name__] = layer_params return params_dict def loadParams(self, layer_params): pass class ConvolutionLayer(LayerBase): """ 卷积层 @params: stride 步幅 pad 填充 method 参数初始化方法 filter_size 滤波器数目 channels 传进图片的通道数 """ def __init__(self, stride, pad, method, filter_size, filter_num, channels): self.stride = stride self.pad = pad self.method = method self.filter_size = filter_size self.filter_num = filter_num self.channels = channels self.initParams(method, filter_size, filter_num, channels) # 中间数据（backward时使用） self.x = None self.col = None self.col_W = None # 权重和偏置参数的梯度 self.dW = None self.db = None def initParams(self, method, fiter_size, filter_num, channels): self.init_method = eval(method)() self.W, self.b = self.init_method.getParams( fiter_size, filter_num, channels) def forward(self, x): FN, C, FH, FW = self.W.shape N, C, H, W = x.shape out_h = 1 + int((H + 2*self.pad - FH) / self.stride) out_w = 1 + int((W + 2*self.pad - FW) / self.stride) col = im2col(x, FH, FW, self.stride, self.pad) col_W = self.W.reshape(FN, -1).T out = np.dot(col, col_W) + self.b out = out.reshape(N, out_h, out_w, -1).transpose(0, 3, 1, 2) self.x = x self.col = col self.col_W = col_W return out def backward(self, dout): FN, C, FH, FW = self.W.shape dout = dout.transpose(0, 2, 3, 1).reshape(-1, FN) self.db = np.sum(dout, axis=0) self.dW = np.dot(self.col.T, dout) self.dW = self.dW.transpose(1, 0).reshape(FN, C, FH, FW) dcol = np.dot(dout, self.col_W.T) dx = col2im(dcol, self.x.shape, FH, FW, self.stride, self.pad) return dx def update(self, learning_object: object): self.W = learning_object(self.W, self.dW, self.__str__()+"w") self.b = learning_object(self.b, self.db, self.__str__()+"b") def savePar