深度学习现在大火,虽然自己上过深度学习课程、用过keras做过一些实验,始终觉得理解不透彻。最近仔细学习前辈和学者的著作,感谢他们的无私奉献,整理得到本文,共勉。
(1)神经网络的缺陷
在神经网络一文中简单介绍了其原理,可以发现不同层之间是全连接的,当神经网络的深度、节点数变大,会导致过拟合、参数过多等问题。
(2)计算机视觉(图像)背景
如果你觉得这篇文章看起来稍微还有些吃力,或者想要更系统地学习人工智能,那么推荐你去看床长人工智能教程。非常棒的大神之作,我也被圈粉了。教程不仅通俗易懂,而且很风趣幽默。点击这里可以查看教程。
根据前言中的两方面,这里介绍卷积神经网络的两个特性。
(1)局部感知
图1:全连接网络。如果L1层有1000×1000像素的图像,L2层有1000,000个隐层神经元,每个隐层神经元都连接L1层图像的每一个像素点,就有1000x1000x1000,000=10^12个连接,也就是10^12个权值参数。
图2:局部连接网络。L2层每一个节点与L1层节点同位置附近10×10的窗口相连接,则1百万个隐层神经元就只有100w乘以100,即10^8个参数。其权值连接个数比原来减少了四个数量级。
(2)权值共享
就图2来说,权值共享,不是说,所有的红色线标注的连接权值相同。而是说,每一个颜色的线都有一个红色线的权值与之相等,所以第二层的每个节点,其从上一层进行卷积的参数都是相同的。
图2中隐层的每一个神经元都连接10×10个图像区域,也就是说每一个神经元存在10×10=100个连接权值参数。如果我们每个神经元这100个参数是相同的?也就是说每个神经元用的是同一个卷积核去卷积图像。这样L1层我们就只有100个参数。但是这样,只提取了图像一种特征?如果需要提取不同的特征,就加多几种卷积核。所以假设我们加到100种卷积核,也就是1万个参数。
每种卷积核的参数不一样,表示它提出输入图像的不同特征(不同的边缘)。这样每种卷积核去卷积图像就得到对图像的不同特征的放映,我们称之为Feature Map,也就是特征图。
以LeCun的LeNet-5为例,不包含输入,LeNet-5共有7层,每层都包含连接权值(可训练参数)。输入图像为32*32大小。我们先要明确一点:每个层有多个特征图,每个特征图通过一种卷积滤波器提取输入的一种特征,然后每个特征图有多个神经元。
C1、C3、C5是卷积层,S2、S4、S6是下采样层。利用图像局部相关性的原理,对图像进行下抽样,可以减少数据处理量同时保留有用信息。
在神经网络一文中已经详细介绍过全连接和激励层的前向传播过程,这里主要介绍卷积层、下采样(池化)层。
(1)卷积层
如图4所示,输入图片是一个5×5的图片,用一个3×3的卷积核对该图片进行卷积操作。本质上是一个点积操作。举例:1×1+0×1+1×1+0×0+1×1+0×1+1×0+0×0+1×1=4
def conv2(X, k): x_row, x_col = X.shape k_row, k_col = k.shape ret_row, ret_col = x_row - k_row + 1, x_col - k_col + 1 ret = np.empty((ret_row, ret_col)) for y in range(ret_row): for x in range(ret_col): sub = X[y : y + k_row, x : x + k_col] ret[y,x] = np.sum(sub * k) return retclass ConvLayer: def __init__(self, in_channel, out_channel, kernel_size): self.w = np.random.randn(in_channel, out_channel, kernel_size, kernel_size) self.b = np.zeros((out_channel)) def _relu(self, x): x[x < 0] = 0 def forward(self, in_data): # assume the first index is channel index in_channel, in_row, in_col = in_data.shape out_channel, kernel_row, kernel_col = self.w.shape[1], self.w.shape[2], self.w.shape[3] self.top_val = np.zeros((out_channel, in_row - kernel_row + 1, in_col - kernel_col + 1)) for j in range(out_channel): for i in range(in_channel): self.top_val[j] += conv2(in_data[i], self.w[i, j]) self.top_val[j] += self.b[j] self.top_val[j] = self._relu(self.topval[j]) return self.top_val
(2)下采样(池化)层
下采样,即池化,目的是减小特征图,池化规模一般为2×2。常用的池化方法有:
class MaxPoolingLayer: def __init__(self, kernel_size, name='MaxPool'): self.kernel_size = kernel_size def forward(self, in_data): in_batch, in_channel, in_row, in_col = in_data.shape k = self.kernel_size out_row = in_row / k + (1 if in_row % k != 0 else 0) out_col = in_col / k + (1 if in_col % k != 0 else 0) self.flag = np.zeros_like(in_data) ret = np.empty((in_batch, in_channel, out_row, out_col)) for b_id in range(in_batch): for c in range(in_channel): for oy in range(out_row): for ox in range(out_col): height = k if (oy + 1) * k <= in_row else in_row - oy * k width = k if (ox + 1) * k <= in_col else in_col - ox * k idx = np.argmax(in_data[b_id, c, oy * k: oy * k + height, ox * k: ox * k + width]) offset_r = idx / width offset_c = idx % width self.flag[b_id, c, oy * k + offset_r, ox * k + offset_c] = 1 ret[b_id, c, oy, ox] = in_data[b_id, c, oy * k + offset_r, ox * k + offset_c] return ret
在神经网络一文中已经详细介绍过全连接和激励层的后向传播过程,这里主要介绍卷积层、下采样(池化)层。
(1)卷积层
当一个卷积层L的下一层(L+1)为采样层,并假设我们已经计算得到了采样层的残差,现在计算该卷积层的残差。从最上面的网络结构图我们知道,采样层(L+1)的map大小是卷积层L的1/(scale*scale),以scale=2为例,但这两层的map个数是一样的,卷积层L的某个map中的4个单元与L+1层对应map的一个单元关联,可以对采样层的残差与一个scale*scale的全1矩阵进行克罗内克积 进行扩充,使得采样层的残差的维度与上一层的输出map的维度一致。
扩展过程:
利用卷积计算卷积层的残差:
def backward(self, residual): in_channel, out_channel, kernel_size = self.w.shape in_batch = residual.shape[0] # gradient_b self.gradient_b = residual.sum(axis=3).sum(axis=2).sum(axis=0) / self.batch_size # gradient_w self.gradient_w = np.zeros_like(self.w) for b_id in range(in_batch): for i in range(in_channel): for o in range(out_channel): self.gradient_w[i, o] += conv2(self.bottom_val[b_id], residual[o]) self.gradient_w /= self.batch_size # gradient_x gradient_x = np.zeros_like(self.bottom_val) for b_id in range(in_batch): for i in range(in_channel): for o in range(out_channel): gradient_x[b_id, i] += conv2(padding(residual, kernel_size - 1), rot180(self.w[i, o])) gradient_x /= self.batch_size # update self.prev_gradient_w = self.prev_gradient_w * self.momentum - self.gradient_w self.w += self.lr * self.prev_gradient_w self.prev_gradient_b = self.prev_gradient_b * self.momentum - self.gradient_b self.b += self.lr * self.prev_gradient_b return gradient_x
(2)下采样(池化)层
当某个采样层L的下一层是卷积层(L+1),并假设我们已经计算出L+1层的残差,现在计算L层的残差。采样层到卷积层直接的连接是有权重和偏置参数的,因此不像卷积层到采样层那样简单。现再假设L层第j个map Mj与L+1层的M2j关联,按照BP的原理,L层的残差Dj是L+1层残差D2j的加权和,但是这里的困难在于,我们很难理清M2j的那些单元通过哪些权重与Mj的哪些单元关联,这里需要两个小的变换(rot180°和padding):
rot180°:旋转:表示对矩阵进行180度旋转(可通过行对称交换和列对称交换完成)
def rot180(in_data): ret = in_data.copy() yEnd = ret.shape[0] - 1 xEnd = ret.shape[1] - 1 for y in range(ret.shape[0] / 2): for x in range(ret.shape[1]): ret[yEnd - y][x] = ret[y][x] for y in range(ret.shape[0]): for x in range(ret.shape[1] / 2): ret[y][xEnd - x] = ret[y][x] return ret
padding:扩充
def padding(in_data, size): cur_r, cur_w = in_data.shape[0], in_data.shape[1] new_r = cur_r + size * 2 new_w = cur_w + size * 2 ret = np.zeros((new_r, new_w)) ret[size:cur_r + size, size:cur_w+size] = in_data return ret
import numpy as npimport sysdef conv2(X, k): # as a demo code, here we ignore the shape check x_row, x_col = X.shape k_row, k_col = k.shape ret_row, ret_col = x_row - k_row + 1, x_col - k_col + 1 ret = np.empty((ret_row, ret_col)) for y in range(ret_row): for x in range(ret_col): sub = X[y : y + k_row, x : x + k_col] ret[y,x] = np.sum(sub * k) return retdef rot180(in_data): ret = in_data.copy() yEnd = ret.shape[0] - 1 xEnd = ret.shape[1] - 1 for y in range(ret.shape[0] / 2): for x in range(ret.shape[1]): ret[yEnd - y][x] = ret[y][x] for y in range(ret.shape[0]): for x in range(ret.shape[1] / 2): ret[y][xEnd - x] = ret[y][x] return retdef padding(in_data, size): cur_r, cur_w = in_data.shape[0], in_data.shape[1] new_r = cur_r + size * 2 new_w = cur_w + size * 2 ret = np.zeros((new_r, new_w)) ret[size:cur_r + size, size:cur_w+size] = in_data return retdef discreterize(in_data, size): num = in_data.shape[0] ret = np.zeros((num, size)) for i, idx in enumerate(in_data): ret[i, idx] = 1 return retclass ConvLayer: def __init__(self, in_channel, out_channel, kernel_size, lr=0.01, momentum=0.9, name='Conv'): self.w = np.random.randn(in_channel, out_channel, kernel_size, kernel_size) self.b = np.zeros((out_channel)) self.layer_name = name self.lr = lr self.momentum = momentum self.prev_gradient_w = np.zeros_like(self.w) self.prev_gradient_b = np.zeros_like(self.b) # def _relu(self, x): # x[x < 0] = 0 # return x def forward(self, in_data): # assume the first index is channel index print 'conv forward:' + str(in_data.shape) in_batch, in_channel, in_row, in_col = in_data.shape out_channel, kernel_size = self.w.shape[1], self.w.shape[2] self.top_val = np.zeros((in_batch, out_channel, in_row - kernel_size + 1, in_col - kernel_size + 1)) self.bottom_val = in_data for b_id in range(in_batch): for o in range(out_channel): for i in range(in_channel): self.top_val[b_id, o] += conv2(in_data[b_id, i], self.w[i, o]) self.top_val[b_id, o] += self.b[o] return self.top_val def backward(self, residual): in_channel, out_channel, kernel_size = self.w.shape in_batch = residual.shape[0] # gradient_b self.gradient_b = residual.sum(axis=3).sum(axis=2).sum(axis=0) / self.batch_size # gradient_w self.gradient_w = np.zeros_like(self.w) for b_id in range(in_batch): for i in range(in_channel): for o in range(out_channel): self.gradient_w[i, o] += conv2(self.bottom_val[b_id], residual[o]) self.gradient_w /= self.batch_size # gradient_x gradient_x = np.zeros_like(self.bottom_val) for b_id in range(in_batch): for i in range(in_channel): for o in range(out_channel): gradient_x[b_id, i] += conv2(padding(residual, kernel_size - 1), rot180(self.w[i, o])) gradient_x /= self.batch_size # update self.prev_gradient_w = self.prev_gradient_w * self.momentum - self.gradient_w self.w += self.lr * self.prev_gradient_w self.prev_gradient_b = self.prev_gradient_b * self.momentum - self.gradient_b self.b += self.lr * self.prev_gradient_b return gradient_xclass FCLayer: def __init__(self, in_num, out_num, lr = 0.01, momentum=0.9): self._in_num = in_num self._out_num = out_num self.w = np.random.randn(in_num, out_num) self.b = np.zeros((out_num, 1)) self.lr = lr self.momentum = momentum self.prev_grad_w = np.zeros_like(self.w) self.prev_grad_b = np.zeros_like(self.b) # def _sigmoid(self, in_data): # return 1 / (1 + np.exp(-in_data)) def forward(self, in_data): print 'fc forward=' + str(in_data.shape) self.topVal = np.dot(self.w.T, in_data) + self.b self.bottomVal = in_data return self.topVal def backward(self, loss): batch_size = loss.shape[0] # residual_z = loss * self.topVal * (1 - self.topVal) grad_w = np.dot(self.bottomVal, loss.T) / batch_size grad_b = np.sum(loss) / batch_size residual_x = np.dot(self.w, loss) self.prev_grad_w = self.prev_grad_w * momentum - grad_w self.prev_grad_b = self.prev_grad_b * momentum - grad_b self.w -= self.lr * self.prev_grad_w self.b -= self.lr * self.prev_grad_b return residual_xclass ReLULayer: def __init__(self, name='ReLU'): pass def forward(self, in_data): self.top_val = in_data ret = in_data.copy() ret[ret < 0] = 0 return ret def backward(self, residual): gradient_x = residual.copy() gradient_x[self.top_val < 0] = 0 return gradient_xclass MaxPoolingLayer: def __init__(self, kernel_size, name='MaxPool'): self.kernel_size = kernel_size def forward(self, in_data): in_batch, in_channel, in_row, in_col = in_data.shape k = self.kernel_size out_row = in_row / k + (1 if in_row % k != 0 else 0) out_col = in_col / k + (1 if in_col % k != 0 else 0) self.flag = np.zeros_like(in_data) ret = np.empty((in_batch, in_channel, out_row, out_col)) for b_id in range(in_batch): for c in range(in_channel): for oy in range(out_row): for ox in range(out_col): height = k if (oy + 1) * k <= in_row else in_row - oy * k width = k if (ox + 1) * k <= in_col else in_col - ox * k idx = np.argmax(in_data[b_id, c, oy * k: oy * k + height, ox * k: ox * k + width]) offset_r = idx / width offset_c = idx % width self.flag[b_id, c, oy * k + offset_r, ox * k + offset_c] = 1 ret[b_id, c, oy, ox] = in_data[b_id, c, oy * k + offset_r, ox * k + offset_c] return ret def backward(self, residual): in_batch, in_channel, in_row, in_col = self.flag k = self.kernel_size out_row, out_col = residual.shape[2], residual.shape[3] gradient_x = np.zeros_like(self.flag) for b_id in range(in_batch): for c in range(in_channel): for oy in range(out_row): for ox in range(out_col): height = k if (oy + 1) * k <= in_row else in_row - oy * k width = k if (ox + 1) * k <= in_col else in_col - ox * k gradient_x[b_id, c, oy * k + offset_r, ox * k + offset_c] = residual[b_id, c, oy, ox] gradient_x[self.flag == 0] = 0 return gradient_xclass FlattenLayer: def __init__(self, name='Flatten'): pass def forward(self, in_data): self.in_batch, self.in_channel, self.r, self.c = in_data.shape return in_data.reshape(self.in_batch, self.in_channel * self.r * self.c) def backward(self, residual): return residual.reshape(self.in_batch, self.in_channel, self.r, self.c)class SoftmaxLayer: def __init__(self, name='Softmax'): pass def forward(self, in_data): exp_out = np.exp(in_data) self.top_val = exp_out / np.sum(exp_out, axis=1) return self.top_val def backward(self, residual): return self.top_val - residualclass Net: def __init__(self): self.layers = [] def addLayer(self, layer): self.layers.append(layer) def train(self, trainData, trainLabel, validData, validLabel, batch_size, iteration): train_num = trainData.shape[0] for iter in range(iteration): print 'iter=' + str(iter) for batch_iter in range(0, train_num, batch_size): if batch_iter + batch_size < train_num: self.train_inner(trainData[batch_iter: batch_iter + batch_size], trainLabel[batch_iter: batch_iter + batch_size]) else: self.train_inner(trainData[batch_iter: train_num], trainLabel[batch_iter: train_num]) print "eval=" + str(self.eval(validData, validLabel)) def train_inner(self, data, label): lay_num = len(self.layers) in_data = data for i in range(lay_num): out_data = self.layers[i].forward(in_data) in_data = out_data residual_in = label for i in range(0, lay_num, -1): residual_out = self.layers[i].backward(residual_in) residual_in = residual_out def eval(self, data, label): lay_num = len(self.layers) in_data = data for i in range(lay_num): out_data = self.layers[i].forward(in_data) in_data = out_data out_idx = np.argmax(in_data, axis=1) label_idx = np.argmax(label, axis=1) return np.sum(out_idx == label_idx) / float(out_idx.shape[0])