Image multi label recognition model does not converge

I built a image multi label recognition model based on resnet-50, the network is as follows (it’s exactly the same as in the torchvision.models.resnet50 , with the only difference in the final softmax layer in the forward function).

My dataset contain 6000 images with 15 classes, each image may have 1 to 5 labels. But after training, the model seems not converge, with always predicting zeros for all labels.

class ResNetMultiTag(nn.Module):

    def __init__(self, block, layers, num_classes=1000, zero_init_residual=False,
                groups=1, width_per_group=64, replace_stride_with_dilation=None,
        super(ResNetMultiTag, self).__init__()
        if norm_layer is None:
            norm_layer = nn.BatchNorm2d
        self._norm_layer = norm_layer

        self.inplanes = 64
        self.dilation = 1
        if replace_stride_with_dilation is None:
            # each element in the tuple indicates if we should replace
            # the 2x2 stride with a dilated convolution instead
            replace_stride_with_dilation = [False, False, False]
        if len(replace_stride_with_dilation) != 3:
            raise ValueError("replace_stride_with_dilation should be None "
                            "or a 3-element tuple, got {}".format(replace_stride_with_dilation))
        self.groups = groups
        self.base_width = width_per_group
        self.conv1 = nn.Conv2d(3, self.inplanes, kernel_size=7, stride=2, padding=3,
        self.bn1 = norm_layer(self.inplanes)
        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.layer2 = self._make_layer(block, 128, layers[1], stride=2,
        self.layer3 = self._make_layer(block, 256, layers[2], stride=2,
        self.layer4 = self._make_layer(block, 512, layers[3], stride=2,
        self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
        self.fc = nn.Linear(512 * block.expansion, num_classes * 2) # special
        self.softmax = nn.Softmax(dim=-1) # special

        for m in self.modules():
            if isinstance(m, nn.Conv2d):
                nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
            elif isinstance(m, (nn.BatchNorm2d, nn.GroupNorm)):
                nn.init.constant_(m.weight, 1)
                nn.init.constant_(m.bias, 0)

        # Zero-initialize the last BN in each residual branch,
        # so that the residual branch starts with zeros, and each residual block behaves like an identity.
        # This improves the model by 0.2~0.3% according to
        if zero_init_residual:
            for m in self.modules():
                if isinstance(m, Bottleneck):
                    nn.init.constant_(m.bn3.weight, 0)
                elif isinstance(m, BasicBlock):
                    nn.init.constant_(m.bn2.weight, 0)

    def _make_layer(self, block, planes, blocks, stride=1, dilate=False):
        norm_layer = self._norm_layer
        downsample = None
        previous_dilation = self.dilation
        if dilate:
            self.dilation *= stride
            stride = 1
        if stride != 1 or self.inplanes != planes * block.expansion:
            downsample = nn.Sequential(
                conv1x1(self.inplanes, planes * block.expansion, stride),
                norm_layer(planes * block.expansion),

        layers = []
        layers.append(block(self.inplanes, planes, stride, downsample, self.groups,
                            self.base_width, previous_dilation, norm_layer))
        self.inplanes = planes * block.expansion
        for _ in range(1, blocks):
            layers.append(block(self.inplanes, planes, groups=self.groups,
                                base_width=self.base_width, dilation=self.dilation,

        return nn.Sequential(*layers)

    def forward(self, x):
        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)
        x = self.layer4(x)

        x = self.avgpool(x)
        x = x.reshape(x.size(0), -1)
        x = self.fc(x)
        x = x.reshape(x.size(0), -1, 2) # special
        x = self.softmax(x) # special

        return x

This is the loss function I use:

def get_tag_loss(self, pred, label):
    pred: [batchsize, num_classes, 2]
    label: [batchsize, num_classes, 2]
    pred = pred.contiguous().view(-1,2)
    label = label.contiguous().view(-1,2)

    label_scalar = label[:, 1]

    logp = torch.nn.functional.log_softmax(pred, dim=-1)
    ce = -logp[range(label_scalar.shape[0]), label_scalar.long()]
    loss = torch.sum( ce )
    return loss

I use SGD with lr ranges from 1e-2 to 1e-4, I have 6000 images, 15 classes. Each image may have 1 to 5 labels. The number of labels is un-balanced but I have done the resampling to keep the balance.

However, the model does not converge.

I tried to train the model on only one image, this case the model converges to the correct labels. However, even if I use only two images, just add one more image, the model failed to converge.

What’s wrong with my implementation?

Did I use the wrong loss function, wrong model, or I make other mistakes?

Thank you all for helping me!!

Update. The reason may be that the dataset is too small. 6000 images, 15 classes, multi-label, may be too difficult.
I’ve tried to build a smaller dataset, with 3000 images, 5 classes, multi-label. The training result, though not good enough, is better. The model can identify some easy labels.
But I’m new to multi label recognition. I hope there is other hint on this problem.