quantizable_resnet.py

class Bottleneck(nn.Module):
    ''' Standard bottleneck block
    input  = inplanes * H * W
    middle =   planes * H/stride * W/stride
    output = 4*planes * H/stride * W/stride
    '''
    expansion = 4

    def __init__(self, inplanes, planes, stride=1, dilation=1, downsample=None):
        super(Bottleneck, self).__init__()
        self.conv1 = nn.Conv2d(inplanes, planes, kernel_size=1, bias=False)
        self.bn1 = nn.BatchNorm2d(planes)
        self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, stride=stride,
                               padding=dilation, dilation=dilation, bias=False)
        self.bn2 = nn.BatchNorm2d(planes)
        self.conv3 = nn.Conv2d(planes, planes * 4, kernel_size=1, bias=False)
        self.bn3 = nn.BatchNorm2d(planes * 4)
        self.relu = nn.ReLU(inplace=True)
        self.downsample = downsample
        self.stride = stride

    def forward(self, x):
        residual = x

        out = self.conv1(x)
        out = self.bn1(out)
        out = self.relu(out)

        out = self.conv2(out)
        out = self.bn2(out)
        out = self.relu(out)

        out = self.conv3(out)
        out = self.bn3(out)

        if self.downsample is not None:
            residual = self.downsample(x)

        out += residual
        out = self.relu(out)

        return out

class QuantizableBottleneck(Bottleneck):
    expansion = 4
    def __init__(self, *args, **kwargs):
        super(QuantizableBottleneck, self).__init__(*args, **kwargs)
        self.skip_add_relu = nn.quantized.FloatFunctional()
        self.relu1 = nn.ReLU(inplace=False)
        self.relu2 = nn.ReLU(inplace=False)        

    def forward(self, x):

        identity = x
        out = self.conv1(x)
        out = self.bn1(out)
        out = self.relu1(out)
        out = self.conv2(out)
        out = self.bn2(out)
        out = self.relu2(out)

        out = self.conv3(out)
        out = self.bn3(out)

        if self.downsample is not None:
            identity = self.downsample(x)
        out = self.skip_add_relu.add_relu(out, identity)

        return out

    def fuse_model(self):
        fuse_modules(self, [['conv1', 'bn1', 'relu1'],
                            ['conv2', 'bn2', 'relu2'],
                            ['conv3', 'bn3']], inplace=True)
        if self.downsample:
            torch.quantization.fuse_modules(self.downsample, ['0', '1'], inplace=True)


class ResNet(nn.Module):
    """ A standard ResNet.
    """
    def __init__(self, block, layers, fc_out, model_name, self_similarity_radius=None, self_similarity_version=2):
        nn.Module.__init__(self)
        self.model_name = model_name

        # default values for a network pre-trained on imagenet
        self.rgb_means = [0.485, 0.456, 0.406]
        self.rgb_stds  = [0.229, 0.224, 0.225]
        self.input_size = (3, 224, 224)

        self.inplanes = 64
        self.conv1 = nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3,
                               bias=False)
        self.bn1 = nn.BatchNorm2d(64)
        self.relu = nn.ReLU(inplace=True)
        self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
        self.layer1 = self._make_layer(block, 64, layers[0], self_similarity_radius=self_similarity_radius, self_similarity_version=self_similarity_version)
        self.layer2 = self._make_layer(block, 128, layers[1], stride=2, self_similarity_radius=self_similarity_radius, self_similarity_version=self_similarity_version)
        self.layer3 = self._make_layer(block, 256, layers[2], stride=2, self_similarity_radius=self_similarity_radius, self_similarity_version=self_similarity_version)
        self.layer4 = self._make_layer(block, 512, layers[3], stride=2, self_similarity_radius=self_similarity_radius, self_similarity_version=self_similarity_version)

        reset_weights(self)

        self.fc = None
        self.fc_out = fc_out
        if self.fc_out > 0:
            self.avgpool = nn.AdaptiveAvgPool2d(1)
            self.fc = nn.Linear(512 * block.expansion, fc_out)
            self.fc_name = 'fc'

    def _make_layer(self, block, planes, blocks, stride=1, self_similarity_radius=None, self_similarity_version=1):
        downsample = None
        if stride != 1 or self.inplanes != planes * block.expansion:
            downsample = nn.Sequential(
                nn.Conv2d(self.inplanes, planes * block.expansion,
                          kernel_size=1, stride=stride, bias=False),
                nn.BatchNorm2d(planes * block.expansion),
            )

        layers = []
        layers.append(block(self.inplanes, planes=planes, stride=stride, downsample=downsample))
        self.inplanes = planes * block.expansion
        for i in range(1, blocks):
            layers.append(block(self.inplanes, planes))
        if self_similarity_radius:
            if self_similarity_version == 1:
                from . self_sim import SelfSimilarity1
                layers.append(SelfSimilarity1(self_similarity_radius, self.inplanes))
            else:
                from . self_sim import SelfSimilarity2
                layers.append(SelfSimilarity2(self_similarity_radius, self.inplanes))
        return nn.Sequential(*layers)

    def forward(self, x, out_layer=0):
        x = self.conv1(x)
        x = self.bn1(x)
        x = self.relu(x)
        x = self.maxpool(x)

        x = self.layer1(x)
        x = self.layer2(x)
        x = self.layer3(x)
        if out_layer==-1:
            return x, self.layer4(x)
        x = self.layer4(x)

        if self.fc_out > 0:
            x = self.avgpool(x)
            x = x.view(x.size(0), -1)
            x = self.fc(x)
        return x
      
    def fuse_model(self):
        r"""Fuse conv/bn/relu modules in resnet models
        Fuse conv+bn+relu/ Conv+relu/conv+Bn modules to prepare for quantization.
        Model is modified in place.  Note that this operation does not change numerics
        and the model after modification is in floating point
        """

        fuse_modules(self, ['conv1', 'bn1', 'relu'], inplace=True)
        for m in self.modules():
            if type(m) == QuantizableBottleneck:
                m.fuse_model()

class ResNet_RMAC(ResNet):
    """ ResNet for RMAC (without ROI pooling)
    """
    def __init__(self, block, layers, model_name, out_dim=2048, norm_features=False,
                       pooling='gem', gemp=3, center_bias=0,
                       dropout_p=None, without_fc=False, **kwargs):
        ResNet.__init__(self, block, layers, 0, model_name, **kwargs)
        self.norm_features = norm_features
        self.without_fc = without_fc
        self.pooling = pooling
        self.center_bias = center_bias
        if pooling == 'max':
            self.adpool = nn.AdaptiveMaxPool2d(output_size=1)
        elif pooling == 'avg':
            self.adpool = nn.AdaptiveAvgPool2d(output_size=1)
        elif pooling.startswith('gem'):
            self.adpool = GeneralizedMeanPoolingP(norm=gemp)
        else:
            raise ValueError(pooling)

        self.dropout = nn.Dropout(dropout_p) if dropout_p is not None else None
        self.fc = nn.Linear(512 * block.expansion, out_dim)
        self.fc_name = 'fc'
        self.feat_dim = out_dim
        self.detach = False

    def _forward(self, x):
        bs, _, H, W = x.shape

        x = ResNet.forward(self, x)

        if self.dropout is not None:
            x = self.dropout(x)

        if self.detach:
            # stop the back-propagation here, if needed
            x = Variable(x.detach())
            x = self.id(x)  # fake transformation

        if self.center_bias > 0:
            b = self.center_bias
            bias = 1 + torch.FloatTensor([[[[0,0,0,0],[0,b,b,0],[0,b,b,0],[0,0,0,0]]]]).to(x.device)
            bias = torch.nn.functional.interpolate(bias, size=x.shape[-2:], mode='bilinear', align_corners=True)
            x = x*bias

        # global pooling
        x = self.adpool(x)

        if self.norm_features:
            x = l2_normalize(x, axis=1)

        x.squeeze_()
        if not self.without_fc:
            x = self.fc(x)

        x = l2_normalize(x, axis=-1)
        return x

    forward = _forward


class QuantizableRMAC(ResNet_RMAC):

    def __init__(self, *args, **kwargs):
        super(QuantizableRMAC, self).__init__(*args, **kwargs)
        self.quant = torch.quantization.QuantStub()
        self.dequant = torch.quantization.DeQuantStub()
    
    def forward(self, x):
        x = self.quant(x)
        x = self._forward(x)
        x = self.dequant(x)
        return x
    
    def fuse_model(self):
        super().fuse_model()