def forward(self, x):
x = quantize_activations_gemm(x)
x1 = self.conv1(x)
conv1_weight, conv1_scale = quantize_weight_gemm(self.conv1.weight)
# conv1_weight = quantize_weights_bias_gemm(self.conv1.weight)
# conv1_scale = 1
conv1_bias = quantize_bias_gemm(self.conv1.bias) / conv1_scale
x = F.conv2d(x, conv1_weight, conv1_bias, stride=1,
padding=1) * conv1_scale
x = self.relu(x)
x = quantize_activations_gemm(x)
# x = self.maxpool(x)
x = self.layer1(x)
x = self.layer2(x)
x = self.layer3(x)
x = self.layer4(x)
x = F.avg_pool2d(x, 4)
x = x.view(x.size(0), -1)
x = self.linear(x)
# x = quantize_activations_gemm(x)
# x = self.scalar(x) # 修改
return x
def forward(self, x):
x = quantize_activations_gemm(x)
residual = x
out1 = self.conv1(x)
conv1_weight, conv1_scale = quantize_weight_gemm(self.conv1.weight)
# conv1_weight = quantize_weights_bias_gemm(self.conv1.weight)
# conv1_scale = 1
conv1_bias = quantize_bias_gemm(self.conv1.bias) / conv1_scale
out = F.conv2d(x, conv1_weight, conv1_bias) * conv1_scale
out = self.relu(out)
out = quantize_activations_gemm(out)
out1 = self.conv2(out)
conv2_weight, conv2_scale = quantize_weight_gemm(self.conv2.weight)
# conv2_weight = quantize_weights_bias_gemm(self.conv2.weight)
# conv2_scale = 1
conv2_bias = quantize_bias_gemm(self.conv2.bias) / conv2_scale
out = F.conv2d(
out, conv2_weight, conv2_bias, stride=self.stride,
padding=1) * conv2_scale
out = self.relu(out)
out = quantize_activations_gemm(out)
out1 = self.conv3(out)
conv3_weight, conv3_scale = quantize_weight_gemm(self.conv3.weight)
# conv3_weight = quantize_weights_bias_gemm(self.conv3.weight)
# conv3_scale = 1
conv3_bias = quantize_bias_gemm(self.conv3.bias) / conv3_scale
out = F.conv2d(out, conv3_weight, conv3_bias, padding=1) * conv3_scale
out = quantize_activations_gemm(out)
out1 += self.shortcut(residual)
if self.downsample:
short_weight, short_scale = quantize_weight_gemm(
self.shortcut[0].weight)
short_bias = quantize_bias_gemm(
self.shortcut[0].bias) / short_scale
residual = F.conv2d(
residual, short_weight, short_bias,
stride=self.stride) * short_scale
out += residual
out = self.relu(out)
return out
def forward(self, input):
qweight = quantize_weight_gemm_C(self.weight)
if self.bias is not None:
qbias = quantize_bias_gemm(self.bias)
else:
qbias = None
qinput = quantize_activations_gemm(input)
out = F.linear(qinput, qweight, qbias)
# out = quantize_activations_gemm(out)
return out
def forward(self, x):
out1 = self.conv(x)
conv_weight, conv_scale = quantize_weight_gemm(self.conv.weight)
conv_bias = quantize_bias_gemm(self.conv.bias / conv_scale)
out = F.conv2d(x, conv_weight, conv_bias, stride=1,
padding=1) * conv_scale
# s = nn.Parameter(torch.tensor(conv_scale))
out = self.relu(out)
out = quantize_activations_gemm(out)
return out
def forward(self, input):
qweight, scale = quantize_weight_gemm_S(self.weight)
if self.bias is not None:
qbias = quantize_bias_gemm(self.bias / scale)
else:
qbias = None
qinput = quantize_activations_gemm(input)
out = F.conv2d(qinput, qweight, qbias, self.stride, self.padding,
self.dilation, self.groups) * scale
return out
l_vgg = [
'.0.', '.2.', '.5.', '.7.', '.10.', '.12.', '.14.', '.17.', '.19.', '.21.',
'.24.', '.26.', '.28.'
]
scale = []
for index in range(params.shape[0]):
weight = state_dict[params[index][0]]
gamma = state_dict[params[index][1]]
beta = state_dict[params[index][2]]
running_mean = state_dict[params[index][3]]
running_var = state_dict[params[index][4]]
delta = gamma / (torch.sqrt(running_var + epsilon))
weight = weight * delta.view(-1, 1, 1, 1)
bias = (0 - running_mean) * delta + beta
qweight, s = quantize_weight_gemm(weight)
qbias = quantize_bias_gemm(bias)
scale.append(s)
merge_state_dict.update({
params[index][0]: qweight,
params[index][0][:-6] + "bias": qbias
})
# merge_state_dict.update({params[index][0][0:15] + l_vgg[index] + "weight": weight,
# params[index][0][0:15] + l_vgg[index] + "bias": bias})
merge_model.load_state_dict(merge_state_dict)
print(scale)
summary_writer = SummaryWriter(save_dir)
x = state_dict["module.conv1.weight"]
y = merge_state_dict["module.conv1.weight"]
z = y * scale[1]