def forward(self, x):
x = quantize_activations_gemm_A(x, self.act_scale[0])
residual = x
out1 = self.conv1(x)
conv1_weight, conv1_scale = quantize_weight_gemm_S(self.conv1.weight)
conv1_bias = quantize_bias_gemm(
self.conv1.bias) / (conv1_scale * self.bias_scale[0])
out = F.conv2d(x, conv1_weight, conv1_bias) * conv1_scale
out = self.relu1(out)
out = out * self.act_scale[0]
out = quantize_activations_gemm_A(out, self.act_scale[1])
out2 = self.conv2(out)
conv2_weight, conv2_scale = quantize_weight_gemm_S(self.conv2.weight)
conv2_bias = quantize_bias_gemm(
self.conv2.bias) / (conv2_scale * self.bias_scale[1])
out = F.conv2d(
out, conv2_weight, conv2_bias, stride=self.stride,
padding=1) * conv2_scale
out = self.relu2(out)
out = out * self.act_scale[1]
out = quantize_activations_gemm_A(out, self.act_scale[2])
out3 = self.conv3(out)
conv3_weight, conv3_scale = quantize_weight_gemm_S(self.conv3.weight)
conv3_bias = quantize_bias_gemm(
self.conv3.bias) / (conv2_scale * self.bias_scale[2])
out = F.conv2d(out, conv3_weight, conv3_bias, padding=1) * conv3_scale
out = out * self.act_scale[2]
out4 = self.shortcut(residual)
if self.downsample:
short_weight, short_scale = quantize_weight_gemm_S(
self.shortcut[0].weight)
short_bias = quantize_bias_gemm(
self.shortcut[0].bias) / (short_scale * self.act_scale[0])
residual = F.conv2d(
residual, short_weight, short_bias,
stride=self.stride) * short_scale
residual = residual * self.act_scale[0]
out += residual
out = self.relu(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.linear(qinput, qweight, qbias) * scale
# 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
def forward(self, x):
x1 = quantize_activations_gemm_A(x, self.act_scale)
out1 = self.conv(x1)
conv_weight, conv_scale = quantize_weight_gemm_S(self.conv.weight)
conv_bias = quantize_bias_gemm(self.conv.bias/(conv_scale*self.bias_scale))
h = x.size()[-1]
bias1 = conv_bias.repeat(h, h, 1)
bias = bias1.transpose(0, 2)
out = (F.conv2d(x1, conv_weight, stride=1, padding=1) + bias)*conv_scale
out = self.relu(out)
# out = quantize_activations_gemm_B(out)
out = out*self.act_scale
# out = quantize_activations_gemm(out)
return out
def forward(self, x):
x = quantize_activations_gemm_A(x, self.scale)
x1 = self.conv1(x)
conv1_weight, conv1_scale = quantize_weight_gemm_S(self.conv1.weight)
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.relu1(x)
x = x * self.scale
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)
return x