def understand_4d_im2col():
batch_size = 2
stride = 1
padding = 0
fh = 2
fw = 2
input_channel = 3
output_channel = 2
iw = 3
ih = 3
(output_height, output_width) = calculate_output_size(ih, iw, fh, fw, padding, stride)
wb = ConvWeightsBias(output_channel, input_channel, fh, fw, InitialMethod.MSRA, OptimizerName.SGD, 0.1)
wb.Initialize("test", "test", True)
wb.W = np.array(range(output_channel * input_channel * fh * fw)).reshape(output_channel, input_channel, fh, fw)
wb.B = np.array([0])
x = np.array(range(input_channel * iw * ih * batch_size)).reshape(batch_size, input_channel, ih, iw)
col = img2col(x, 2, 2, 1, 0)
w = wb.W.reshape(output_channel, -1).T
output = np.dot(col, w)
print("x=\n", x)
print("col_x=\n", col)
print("weights=\n", wb.W)
print("col_w=\n", w)
print("output=\n", output)
out2 = output.reshape(batch_size, output_height, output_width, -1)
print("out2=\n", out2)
out3 = np.transpose(out2, axes=(0, 3, 1, 2))
print("conv result=\n", out3)
def __init__(
self,
input_shape, # (InputChannelCount, H, W)
kernal_shape, # (OutputChannelCount, FH, FW)
conv_param, # (stride, padding)
activator,
param):
self.num_input_channel = input_shape[0]
self.input_height = input_shape[1]
self.input_width = input_shape[2]
self.num_output_channel = kernal_shape[0]
self.filter_height = kernal_shape[1]
self.filter_width = kernal_shape[2]
self.stride = conv_param[0]
self.padding = conv_param[1]
self.activator = activator
self.WeightsBias = ConvWeightsBias(self.num_output_channel,
self.num_input_channel,
self.filter_height,
self.filter_width,
param.init_method,
param.optimizer_name, param.eta)
(self.output_height, self.output_width) = calculate_output_size(
self.input_height, self.input_width, self.filter_height,
self.filter_width, self.padding, self.stride)
self.output_shape = (self.num_output_channel, self.output_height,
self.output_height)
def test_2d_conv():
batch_size = 1
stride = 1
padding = 0
fh = 2
fw = 2
input_channel = 1
output_channel = 1
iw = 3
ih = 3
(output_height, output_width) = calculate_output_size(ih, iw, fh, fw, padding, stride)
wb = ConvWeightsBias(output_channel, input_channel, fh, fw, InitialMethod.MSRA, OptimizerName.SGD, 0.1)
wb.Initialize("test", "test", True)
wb.W = np.array([3,2,1,0]).reshape(1,1,2,2)
wb.B = np.array([0])
x = np.array(range(9)).reshape(1,1,3,3)
output1 = jit_conv_4d(x, wb.W, wb.B, output_height, output_width, stride)
print("input=\n", x)
print("weights=\n", wb.W)
print("output=\n", output1)
col = img2col(x, 2, 2, 1, 0)
w = wb.W.reshape(4, 1)
output2 = np.dot(col, w)
print("input=\n", col)
print("weights=\n", w)
print("output2=\n", output2)
def initialize(self, folder, name, create_new=False):
self.WB = ConvWeightsBias(self.OutC, self.InC, self.FH, self.FW,
self.hp.init_method, self.hp.optimizer_name,
self.hp.eta)
self.WB.Initialize(folder, name, create_new)
(self.OutH,
self.OutW) = calculate_output_size(self.InH, self.InW, self.FH,
self.FW, self.padding, self.stride)
self.output_shape = (self.OutC, self.OutH, self.OutH)
def initialize(self, folder, name, create_new=False):
self.WB = ConvWeightsBias(
self.num_output_channel, self.num_input_channel, self.filter_height, self.filter_width,
self.hp.init_method, self.hp.optimizer_name, self.hp.eta)
self.WB.Initialize(folder, name, create_new)
(self.output_height, self.output_width) = ConvLayer.calculate_output_size(
self.input_height, self.input_width,
self.filter_height, self.filter_width,
self.padding, self.stride)
self.output_shape = (self.num_output_channel, self.output_height, self.output_height)
def understand_4d_col2img_complex():
batch_size = 2
stride = 1
padding = 0
fh = 2
fw = 2
input_channel = 3
output_channel = 2
iw = 3
ih = 3
(output_height,
output_width) = calculate_output_size(ih, iw, fh, fw, padding, stride)
wb = ConvWeightsBias(output_channel, input_channel, fh, fw,
InitialMethod.MSRA, OptimizerName.SGD, 0.1)
wb.Initialize("test", "test", True)
wb.W = np.array(range(output_channel * input_channel * fh * fw)).reshape(
output_channel, input_channel, fh, fw)
wb.B = np.array([0])
x = np.array(range(input_channel * iw * ih * batch_size)).reshape(
batch_size, input_channel, ih, iw)
print("x=\n", x)
col_x = img2col(x, fh, fw, stride, padding)
print("col_x=\n", col_x)
print("w=\n", wb.W)
col_w = wb.W.reshape(output_channel, -1).T
print("col_w=\n", col_w)
# backward
delta_in = np.array(
range(batch_size * output_channel * output_height *
output_width)).reshape(batch_size, output_channel, output_height,
output_width)
print("delta_in=\n", delta_in)
delta_in_2d = np.transpose(delta_in,
axes=(0, 2, 3, 1)).reshape(-1, output_channel)
print("delta_in_2d=\n", delta_in_2d)
dB = np.sum(delta_in_2d, axis=0, keepdims=True).T / batch_size
print("dB=\n", dB)
dW = np.dot(col_x.T, delta_in_2d) / batch_size
print("dW=\n", dW)
dW = np.transpose(dW, axes=(1, 0)).reshape(output_channel, input_channel,
fh, fw)
print("dW=\n", dW)
dcol = np.dot(delta_in_2d, col_w.T)
print("dcol=\n", dcol)
delta_out = col2img(dcol, x.shape, fh, fw, stride, padding, output_height,
output_width)
print("delta_out=\n", delta_out)
class ConvLayer(CLayer):
def __init__(self, num_input_channel, num_output_channel, num_filter_size, stride, padding, activator):
self.num_input_channel = num_input_channel
self.num_output_channel = num_output_channel
self.num_filter_size = num_filter_size
self.stride = stride
self.padding = padding
self.activator = activator
def Initialize(self):
self.weights = ConvWeightsBias(self.num_output_channel, self.num_input_channel, self.num_filter_size, self.num_filter_size)
self.weights.Initialize();
"""
输入数据
N：样本图片数量（比如一次计算10张图片）
C：图片通道数量（比如红绿蓝三通道）
H：图片高度（比如224）
W：图片宽度（比如224）
思维卷积操作
"""
def forward(self, x):
self.input_shape = x.shape
assert(x.ndim == 4)
self.x = x
# 把激活函数算做是当前层，上一层的误差传入后，先经过激活函数的导数，而得到本层的针对z值的误差
def backward(self, delta_in, flag):
def pre_update(self):
self.weights.pre_Update()
def update(self):
self.weights.Update()
def save_parameters(self, name):
self.weights.SaveResultValue(name)
def load_parameters(self, name):
self.weights.LoadResultValue(name)
class ConvLayer(CLayer):
# define the number of input and output channel, also the filter size
def __init__(
self,
input_shape, # (InputChannelCount, H, W)
kernal_shape, # (OutputChannelCount, FH, FW)
conv_param, # (stride, padding)
activator,
param):
self.num_input_channel = input_shape[0]
self.input_height = input_shape[1]
self.input_width = input_shape[2]
self.num_output_channel = kernal_shape[0]
self.filter_height = kernal_shape[1]
self.filter_width = kernal_shape[2]
self.stride = conv_param[0]
self.padding = conv_param[1]
self.activator = activator
self.WeightsBias = ConvWeightsBias(self.num_output_channel,
self.num_input_channel,
self.filter_height,
self.filter_width,
param.init_method,
param.optimizer_name, param.eta)
(self.output_height, self.output_width) = calculate_output_size(
self.input_height, self.input_width, self.filter_height,
self.filter_width, self.padding, self.stride)
self.output_shape = (self.num_output_channel, self.output_height,
self.output_height)
"""
输入数据
N：样本图片数量（比如一次计算10张图片）
C：图片通道数量（比如红绿蓝三通道）
H：图片高度（比如224）
W：图片宽度（比如224）
思维卷积操作
"""
def forward(self, x):
assert (x.ndim == 4)
self.x = x
assert (self.x.shape[1] == self.num_input_channel)
assert (self.x.shape[2] == self.input_height)
assert (self.x.shape[3] == self.input_width)
self.batch_size = self.x.shape[0]
if self.padding > 0:
self.padded = np.pad(self.x,
((0, 0), (0, 0), (self.padding, self.padding),
(self.padding, self.padding)), 'constant')
else:
self.padded = self.x
#end if
self.z = jit_conv_4d(self.padded, self.WeightsBias.W,
self.WeightsBias.B, self.output_height,
self.output_width, self.stride)
self.a = self.activator.forward(self.z)
return self.a
def forward_fast(self, x):
FN, C, FH, FW = self.WeightsBias.W.shape
N, C, H, W = x.shape
out_h = 1 + int((H + 2 * self.padding - FH) / self.stride)
out_w = 1 + int((W + 2 * self.padding - FW) / self.stride)
col_x = im2col(x, FH, FW, self.stride, self.padding)
col_W = self.WeightsBias.W.reshape(FN, -1).T
out = np.dot(col_x, col_W) + self.WeightsBias.B.reshape(-1, FN)
self.z = out.reshape(N, out_h, out_w, -1).transpose(0, 3, 1, 2)
self.x = x
self.col_x = col_x
self.col_W = col_W
self.a = self.activator.forward(self.z)
return self.z, self.a
# 把激活函数算做是当前层，上一层的误差传入后，先经过激活函数的导数，而得到本层的针对z值的误差
def backward(self, delta_in, flag):
assert (delta_in.ndim == 4)
assert (delta_in.shape == self.a.shape)
# 计算激活函数的导数
dz, _ = self.activator.backward(self.z, self.a, delta_in)
# 转换误差矩阵尺寸
dz_stride_1 = expand_delta_map(dz, self.batch_size,
self.num_output_channel,
self.input_height, self.input_width,
self.output_height, self.output_width,
self.filter_height, self.filter_width,
self.padding, self.stride)
# 求本层的输出误差矩阵时，应该用本层的输入误差矩阵互相关计算本层的卷积核的旋转
# 由于输出误差矩阵的尺寸必须与本层的输入数据的尺寸一致，所以必须根据卷积核的尺寸，调整本层的输入误差矩阵的尺寸
(pad_h,
pad_w) = calculate_padding_size(dz_stride_1.shape[2],
dz_stride_1.shape[3],
self.filter_height, self.filter_width,
self.input_height, self.input_width)
dz_padded = np.pad(dz_stride_1,
((0, 0), (0, 0), (pad_h, pad_h), (pad_w, pad_w)),
'constant')
# 计算本层的权重矩阵的梯度
self._calculate_weightsbias_grad(dz_stride_1)
# 计算本层输出到下一层的误差矩阵
delta_out = self._calculate_delta_out(dz_padded, flag)
return delta_out
# 用输入数据乘以回传入的误差矩阵,得到卷积核的梯度矩阵
def _calculate_weightsbias_grad(self, dz):
self.WeightsBias.ClearGrads()
(pad_h, pad_w) = calculate_padding_size(self.input_height,
self.input_width, dz.shape[2],
dz.shape[3],
self.filter_height,
self.filter_width, 1)
input_padded = np.pad(self.x,
((0, 0), (0, 0), (pad_h, pad_h), (pad_w, pad_w)),
'constant')
for bs in range(self.batch_size):
for oc in range(self.num_output_channel): # == kernal count
for ic in range(self.num_input_channel): # == filter count
w_grad = np.zeros((self.filter_height, self.filter_width))
conv2d(input_padded[bs, ic], dz[bs, oc], 0, w_grad)
self.WeightsBias.W_grad[oc, ic] += w_grad
#end ic
self.WeightsBias.B_grad[oc] += dz[bs, oc].sum()
#end oc
#end bs
self.WeightsBias.MeanGrads(self.batch_size)
# 用输入误差矩阵乘以（旋转180度后的）卷积核
def _calculate_delta_out(self, dz, flag):
delta_out = np.zeros(self.x.shape)
if flag != LayerIndexFlags.FirstLayer:
rot_weights = self.WeightsBias.Rotate180()
for bs in range(batch_size):
for oc in range(self.num_output_channel): # == kernal count
delta_per_input = np.zeros(
(self.input_height, self.input_width))
for ic in range(self.num_input_channel): # == filter count
conv2d(dz[bs, oc], rot_weights[oc, ic], 0,
delta_per_input)
delta_out[bs, ic] += delta_per_input
#END IC
#end oc
#end bs
# end if
return delta_out
def pre_update(self):
self.weights.pre_Update()
def update(self):
self.WeightsBias.Update()
def save_parameters(self, name):
self.WeightsBias.Save(name)
def load_parameters(self, name):
self.WeightsBias.Load(name)
class ConvLayer(CLayer):
# define the number of input and output channel, also the filter size
def __init__(self,
input_shape, # (InputChannelCount, H, W)
kernal_shape, # (OutputChannelCount, FH, FW)
conv_param, # (stride, padding)
hp):
self.InC = input_shape[0] # input channel count
self.InH = input_shape[1] # input image height
self.InW = input_shape[2] # input image width
self.OutC = kernal_shape[0] # output channel count
self.FH = kernal_shape[1] # kernal/filter height
self.FW = kernal_shape[2] # kernal/filter width
self.stride = conv_param[0]
self.padding = conv_param[1]
self.hp = hp
def initialize(self, folder, name, create_new=False):
self.WB = ConvWeightsBias(
self.OutC, self.InC, self.FH, self.FW,
self.hp.init_method, self.hp.optimizer_name, self.hp.eta)
self.WB.Initialize(folder, name, create_new)
(self.OutH, self.OutW) = calculate_output_size(
self.InH, self.InW,
self.FH, self.FW,
self.padding, self.stride)
self.output_shape = (self.OutC, self.OutH, self.OutH)
def set_filter(self, w, b):
if w is not None:
self.WB.W = w
if b is not None:
self.WB.B = b
def forward(self, x, train=True):
return self.forward_img2col(x, train)
def backward(self, delta_in, layer_idx):
delta_out, dw, db = self.backward_col2img(delta_in, layer_idx)
return delta_out
def forward_img2col(self, x, train=True):
self.x = x
self.batch_size = self.x.shape[0]
assert(self.x.shape == (self.batch_size, self.InC, self.InH, self.InW))
self.col_x = img2col(x, self.FH, self.FW, self.stride, self.padding)
self.col_w = self.WB.W.reshape(self.OutC, -1).T
self.col_b = self.WB.B.reshape(-1, self.OutC)
out1 = np.dot(self.col_x, self.col_w) + self.col_b
out2 = out1.reshape(self.batch_size, self.OutH, self.OutW, -1)
self.z = np.transpose(out2, axes=(0, 3, 1, 2))
return self.z
def backward_col2img(self, delta_in, layer_idx):
col_delta_in = np.transpose(delta_in, axes=(0,2,3,1)).reshape(-1, self.OutC)
self.WB.dB = np.sum(col_delta_in, axis=0, keepdims=True).T / self.batch_size
col_dW = np.dot(self.col_x.T, col_delta_in) / self.batch_size
self.WB.dW = np.transpose(col_dW, axes=(1, 0)).reshape(self.OutC, self.InC, self.FH, self.FW)
col_delta_out = np.dot(col_delta_in, self.col_w.T)
delta_out = col2img(col_delta_out, self.x.shape, self.FH, self.FW, self.stride, self.padding, self.OutH, self.OutW)
return delta_out, self.WB.dW, self.WB.dB
def forward_numba(self, x, train=True):
assert(x.ndim == 4)
self.x = x
assert(self.x.shape[1] == self.InC)
assert(self.x.shape[2] == self.InH)
assert(self.x.shape[3] == self.InW)
self.batch_size = self.x.shape[0]
if self.padding > 0:
self.padded = np.pad(self.x, ((0,0), (0,0), (self.padding,self.padding), (self.padding,self.padding)), 'constant')
#self.padded = np.pad(self.x, mode="constant", constant_value=0, pad_width=(0,0,0,0,self.padding,self.padding,self.padding,self.padding))
else:
self.padded = self.x
#end if
self.z = jit_conv_4d(self.padded, self.WB.W, self.WB.B, self.OutH, self.OutW, self.stride)
return self.z
def backward_numba(self, delta_in, flag):
assert(delta_in.ndim == 4)
assert(delta_in.shape == self.z.shape)
# 如果正向计算中的stride不是1，转换成是1的等价误差数组
dz_stride_1 = expand_delta_map(delta_in, self.batch_size, self.OutC, self.InH, self.InW, self.OutH, self.OutW, self.FH, self.FW, self.padding, self.stride)
# 计算本层的权重矩阵的梯度
self._calculate_weightsbias_grad(dz_stride_1)
# 求本层的输出误差矩阵时，应该用本层的输入误差矩阵互相关计算本层的卷积核的旋转
# 由于输出误差矩阵的尺寸必须与本层的输入数据的尺寸一致，所以必须根据卷积核的尺寸，调整本层的输入误差矩阵的尺寸
(pad_h, pad_w) = calculate_padding_size(
dz_stride_1.shape[2], dz_stride_1.shape[3],
self.FH, self.FW,
self.InH, self.InW)
dz_padded = np.pad(dz_stride_1, ((0,0),(0,0),(pad_h, pad_h),(pad_w, pad_w)), 'constant')
# 计算本层输出到下一层的误差矩阵
delta_out = self._calculate_delta_out(dz_padded, flag)
#return delta_out
return delta_out, self.WB.dW, self.WB.dB
# 用输入数据乘以回传入的误差矩阵,得到卷积核的梯度矩阵
def _calculate_weightsbias_grad(self, dz):
self.WB.ClearGrads()
# 先把输入矩阵扩大，周边加0
(pad_h, pad_w) = calculate_padding_size(
self.InH, self.InW,
dz.shape[2], dz.shape[3],
self.FH, self.FW, 1)
input_padded = np.pad(self.x, ((0,0),(0,0),(pad_h, pad_h),(pad_w,pad_w)), 'constant')
# 输入矩阵与误差矩阵卷积得到权重梯度矩阵
(self.WB.dW, self.WB.dB) = calcalate_weights_grad(
input_padded, dz, self.batch_size,
self.OutC, self.InC,
self.FH, self.FW,
self.WB.dW, self.WB.dB)
self.WB.MeanGrads(self.batch_size)
# 用输入误差矩阵乘以（旋转180度后的）卷积核
def _calculate_delta_out(self, dz, layer_idx):
if layer_idx == 0:
return None
# 旋转卷积核180度
rot_weights = self.WB.Rotate180()
# 定义输出矩阵形状
delta_out = np.zeros(self.x.shape).astype(np.float32)
# 输入梯度矩阵卷积旋转后的卷积核，得到输出梯度矩阵
delta_out = calculate_delta_out(dz, rot_weights, self.batch_size,
self.InC, self.OutC,
self.InH, self.InW, delta_out)
return delta_out
def pre_update(self):
pass
def update(self):
self.WB.Update()
def save_parameters(self):
self.WB.SaveResultValue()
def load_parameters(self):
self.WB.LoadResultValue()
return rs
if __name__ == '__main__':
stride = 1
padding = 0
fh = 3
fw = 3
input_channel = 3
output_channel = 4
iw = 28
ih = 28
(output_height,
output_width) = ConvLayer.calculate_output_size(ih, iw, fh, fw, padding,
stride)
wb = ConvWeightsBias(output_channel, input_channel, fh, fw,
InitialMethod.MSRA, OptimizerName.SGD, 0.1)
wb.Initialize("test", "test", True)
batch_size = 64
x = np.random.randn(batch_size, input_channel, iw, ih)
# dry run
output1 = conv_4d(x, wb.W, wb.B, output_height, output_width, stride)
s1 = time.time()
for i in range(10):
output1 = conv_4d(x, wb.W, wb.B, output_height, output_width, stride)
e1 = time.time()
print("Time used for Python:", e1 - s1)
# dry run
output2 = jit_conv_4d(x, wb.W, wb.B, output_height, output_width, stride)
s2 = time.time()
for i in range(10):
class ConvLayer(CLayer):
# define the number of input and output channel, also the filter size
def __init__(self,
input_shape, # (InputChannelCount, H, W)
kernal_shape, # (OutputChannelCount, FH, FW)
conv_param, # (stride, padding)
hp):
self.num_input_channel = input_shape[0]
self.input_height = input_shape[1]
self.input_width = input_shape[2]
self.num_output_channel = kernal_shape[0]
self.filter_height = kernal_shape[1]
self.filter_width = kernal_shape[2]
self.stride = conv_param[0]
self.padding = conv_param[1]
self.hp = hp
def initialize(self, folder, name, create_new=False):
self.WB = ConvWeightsBias(
self.num_output_channel, self.num_input_channel, self.filter_height, self.filter_width,
self.hp.init_method, self.hp.optimizer_name, self.hp.eta)
self.WB.Initialize(folder, name, create_new)
(self.output_height, self.output_width) = ConvLayer.calculate_output_size(
self.input_height, self.input_width,
self.filter_height, self.filter_width,
self.padding, self.stride)
self.output_shape = (self.num_output_channel, self.output_height, self.output_height)
def forward(self, x, train=True):
return self.forward_img2col(x, train)
def backward(self, delta_in, layer_idx):
delta_out, dw, db = self.backward_col2img(delta_in, layer_idx)
return delta_out
def forward_img2col(self, x, train=True):
self.x = x
assert(self.x.shape[1] == self.num_input_channel)
assert(self.x.shape[2] == self.input_height)
assert(self.x.shape[3] == self.input_width)
self.batch_size = self.x.shape[0]
FN, C, FH, FW = self.WB.W.shape
N, C, H, W = x.shape
out_h = 1 + int((H + 2 * self.padding - FH) / self.stride)
out_w = 1 + int((W + 2 * self.padding - FW) / self.stride)
self.col_x = img2col(x, FH, FW, self.stride, self.padding)
self.col_w = self.WB.W.reshape(FN, -1).T
out1 = np.dot(self.col_x, self.col_w) + self.WB.B.reshape(-1,FN)
out2 = out1.reshape(N, out_h, out_w, -1)
self.z = np.transpose(out2, axes=(0, 3, 1, 2))
return self.z
def backward_col2img(self, delta_in, layer_idx):
FN, C, FH, FW = self.WB.W.shape
dout = np.transpose(delta_in, axes=(0,2,3,1)).reshape(-1, FN)
self.WB.dB = np.sum(dout, axis=0, keepdims=True).T / self.batch_size
dW = np.dot(self.col_x.T, dout)
self.WB.dW = np.transpose(dW, axes=(1, 0)).reshape(FN, C, FH, FW) / self.batch_size
dcol = np.dot(dout, self.col_w.T)
delta_out = col2img(dcol, self.x.shape, FH, FW, self.stride, self.padding)
return delta_out, self.WB.dW, self.WB.dB
def forward_numba(self, x, train=True):
assert(x.ndim == 4)
self.x = x
assert(self.x.shape[1] == self.num_input_channel)
assert(self.x.shape[2] == self.input_height)
assert(self.x.shape[3] == self.input_width)
self.batch_size = self.x.shape[0]
if self.padding > 0:
self.padded = np.pad(self.x, ((0,0), (0,0), (self.padding,self.padding), (self.padding,self.padding)), 'constant')
#self.padded = np.pad(self.x, mode="constant", constant_value=0, pad_width=(0,0,0,0,self.padding,self.padding,self.padding,self.padding))
else:
self.padded = self.x
#end if
self.z = jit_conv_4d(self.padded, self.WB.W, self.WB.B, self.output_height, self.output_width, self.stride)
return self.z
def backward_numba(self, delta_in, flag):
assert(delta_in.ndim == 4)
assert(delta_in.shape == self.z.shape)
# 转换误差矩阵尺寸
dz_stride_1 = expand_delta_map(delta_in, self.batch_size, self.num_output_channel, self.input_height, self.input_width, self.output_height, self.output_width, self.filter_height, self.filter_width, self.padding, self.stride)
# 求本层的输出误差矩阵时，应该用本层的输入误差矩阵互相关计算本层的卷积核的旋转
# 由于输出误差矩阵的尺寸必须与本层的输入数据的尺寸一致，所以必须根据卷积核的尺寸，调整本层的输入误差矩阵的尺寸
(pad_h, pad_w) = calculate_padding_size(
dz_stride_1.shape[2], dz_stride_1.shape[3],
self.filter_height, self.filter_width,
self.input_height, self.input_width)
dz_padded = np.pad(dz_stride_1, ((0,0),(0,0),(pad_h, pad_h),(pad_w, pad_w)), 'constant')
# 计算本层的权重矩阵的梯度
self._calculate_weightsbias_grad(dz_stride_1)
# 计算本层输出到下一层的误差矩阵
delta_out = self._calculate_delta_out(dz_padded, flag)
#return delta_out
return delta_out, self.WB.dW, self.WB.dB
# 用输入数据乘以回传入的误差矩阵,得到卷积核的梯度矩阵
def _calculate_weightsbias_grad(self, dz):
self.WB.ClearGrads()
# 先把输入矩阵扩大，周边加0
(pad_h, pad_w) = calculate_padding_size(
self.input_height, self.input_width,
dz.shape[2], dz.shape[3],
self.filter_height, self.filter_width, 1)
input_padded = np.pad(self.x, ((0,0),(0,0),(pad_h, pad_h),(pad_w,pad_w)), 'constant')
# 输入矩阵与误差矩阵卷积得到权重梯度矩阵
(self.WB.dW, self.WB.dB) = calcalate_weights_grad(
input_padded, dz, self.batch_size,
self.num_output_channel, self.num_input_channel,
self.filter_height, self.filter_width,
self.WB.dW, self.WB.dB)
self.WB.MeanGrads(self.batch_size)
# 用输入误差矩阵乘以（旋转180度后的）卷积核
def _calculate_delta_out(self, dz, layer_idx):
if layer_idx == 0:
return None
# 旋转卷积核180度
rot_weights = self.WB.Rotate180()
delta_out = np.zeros(self.x.shape).astype(np.float32)
# 输入梯度矩阵卷积旋转后的卷积核，得到输出梯度矩阵
delta_out = calculate_delta_out(dz, rot_weights, self.batch_size,
self.num_input_channel, self.num_output_channel,
self.input_height, self.input_width, delta_out)
return delta_out
def pre_update(self):
self.weights.pre_Update()
def update(self):
self.WB.Update()
def save_parameters(self):
self.WB.SaveResultValue()
def load_parameters(self):
self.WB.LoadResultValue()
@staticmethod
def calculate_output_size(input_h, input_w, filter_h, filter_w, padding, stride=1):
output_h = (input_h - filter_h + 2 * padding) // stride + 1
output_w = (input_w - filter_w + 2 * padding) // stride + 1
return (output_h, output_w)
def Initialize(self):
self.weights = ConvWeightsBias(self.num_output_channel, self.num_input_channel, self.num_filter_size, self.num_filter_size)
self.weights.Initialize();