Transfer Learning is not working

Hi there!

I`m trying to solve the kinship verification problem and for that I need to use Siamese type of networks, meaning there are two images on the input, and the output is binary classification probability.
Seems like model doing well on the train set (up to 80% acc, that is okay for such problem), but on the validation set model gives ~50% accuracy during whole training, with constantly increasing loss. On the test set model predict either all zeros or all ones, returning ~50% acc.

He is a whole network architecture:

  1. ResNet50 feature extractor, pretrained on the VGGFace2 dataset → embeddings
  2. Few combinations of the embeddings of two images, concatenating them
  3. Few FC layers, single output

For initialization pretrained ResNet50, I used code and weights from here (VGGFace2 section).

Here is the code:

class ResNet50(nn.Module):

    def __init__(self):
        super(ResNet50, self).__init__()
        self.meta = {'mean': [131.0912, 103.8827, 91.4953],
                     'std': [1, 1, 1],
                     'imageSize': [224, 224, 3]}
        self.conv1_7x7_s2 = nn.Conv2d(3, 64, kernel_size=[7, 7], stride=(2, 2), padding=(3, 3), bias=False)
        self.conv1_7x7_s2_bn = nn.BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        self.conv1_relu_7x7_s2 = nn.ReLU()
        self.pool1_3x3_s2 = nn.MaxPool2d(kernel_size=[3, 3], stride=[2, 2], padding=(0, 0), dilation=1, ceil_mode=True)
        self.conv2_1_1x1_reduce = nn.Conv2d(64, 64, kernel_size=[1, 1], stride=(1, 1), bias=False)
        self.conv2_1_1x1_reduce_bn = nn.BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        self.conv2_1_1x1_reduce_relu = nn.ReLU()
        self.conv2_1_3x3 = nn.Conv2d(64, 64, kernel_size=[3, 3], stride=(1, 1), padding=(1, 1), bias=False)
        self.conv2_1_3x3_bn = nn.BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        self.conv2_1_3x3_relu = nn.ReLU()
        self.conv2_1_1x1_increase = nn.Conv2d(64, 256, kernel_size=[1, 1], stride=(1, 1), bias=False)
        self.conv2_1_1x1_increase_bn = nn.BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        self.conv2_1_1x1_proj = nn.Conv2d(64, 256, kernel_size=[1, 1], stride=(1, 1), bias=False)
        self.conv2_1_1x1_proj_bn = nn.BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        self.conv2_1_relu = nn.ReLU()
        self.conv2_2_1x1_reduce = nn.Conv2d(256, 64, kernel_size=[1, 1], stride=(1, 1), bias=False)
        self.conv2_2_1x1_reduce_bn = nn.BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        self.conv2_2_1x1_reduce_relu = nn.ReLU()
        self.conv2_2_3x3 = nn.Conv2d(64, 64, kernel_size=[3, 3], stride=(1, 1), padding=(1, 1), bias=False)
        self.conv2_2_3x3_bn = nn.BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        self.conv2_2_3x3_relu = nn.ReLU()
        self.conv2_2_1x1_increase = nn.Conv2d(64, 256, kernel_size=[1, 1], stride=(1, 1), bias=False)
        self.conv2_2_1x1_increase_bn = nn.BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        self.conv2_2_relu = nn.ReLU()
        self.conv2_3_1x1_reduce = nn.Conv2d(256, 64, kernel_size=[1, 1], stride=(1, 1), bias=False)
        self.conv2_3_1x1_reduce_bn = nn.BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        self.conv2_3_1x1_reduce_relu = nn.ReLU()
        self.conv2_3_3x3 = nn.Conv2d(64, 64, kernel_size=[3, 3], stride=(1, 1), padding=(1, 1), bias=False)
        self.conv2_3_3x3_bn = nn.BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        self.conv2_3_3x3_relu = nn.ReLU()
        self.conv2_3_1x1_increase = nn.Conv2d(64, 256, kernel_size=[1, 1], stride=(1, 1), bias=False)
        self.conv2_3_1x1_increase_bn = nn.BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        self.conv2_3_relu = nn.ReLU()
        self.conv3_1_1x1_reduce = nn.Conv2d(256, 128, kernel_size=[1, 1], stride=(2, 2), bias=False)
        self.conv3_1_1x1_reduce_bn = nn.BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        self.conv3_1_1x1_reduce_relu = nn.ReLU()
        self.conv3_1_3x3 = nn.Conv2d(128, 128, kernel_size=[3, 3], stride=(1, 1), padding=(1, 1), bias=False)
        self.conv3_1_3x3_bn = nn.BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        self.conv3_1_3x3_relu = nn.ReLU()
        self.conv3_1_1x1_increase = nn.Conv2d(128, 512, kernel_size=[1, 1], stride=(1, 1), bias=False)
        self.conv3_1_1x1_increase_bn = nn.BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        self.conv3_1_1x1_proj = nn.Conv2d(256, 512, kernel_size=[1, 1], stride=(2, 2), bias=False)
        self.conv3_1_1x1_proj_bn = nn.BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        self.conv3_1_relu = nn.ReLU()
        self.conv3_2_1x1_reduce = nn.Conv2d(512, 128, kernel_size=[1, 1], stride=(1, 1), bias=False)
        self.conv3_2_1x1_reduce_bn = nn.BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        self.conv3_2_1x1_reduce_relu = nn.ReLU()
        self.conv3_2_3x3 = nn.Conv2d(128, 128, kernel_size=[3, 3], stride=(1, 1), padding=(1, 1), bias=False)
        self.conv3_2_3x3_bn = nn.BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        self.conv3_2_3x3_relu = nn.ReLU()
        self.conv3_2_1x1_increase = nn.Conv2d(128, 512, kernel_size=[1, 1], stride=(1, 1), bias=False)
        self.conv3_2_1x1_increase_bn = nn.BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        self.conv3_2_relu = nn.ReLU()
        self.conv3_3_1x1_reduce = nn.Conv2d(512, 128, kernel_size=[1, 1], stride=(1, 1), bias=False)
        self.conv3_3_1x1_reduce_bn = nn.BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        self.conv3_3_1x1_reduce_relu = nn.ReLU()
        self.conv3_3_3x3 = nn.Conv2d(128, 128, kernel_size=[3, 3], stride=(1, 1), padding=(1, 1), bias=False)
        self.conv3_3_3x3_bn = nn.BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        self.conv3_3_3x3_relu = nn.ReLU()
        self.conv3_3_1x1_increase = nn.Conv2d(128, 512, kernel_size=[1, 1], stride=(1, 1), bias=False)
        self.conv3_3_1x1_increase_bn = nn.BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        self.conv3_3_relu = nn.ReLU()
        self.conv3_4_1x1_reduce = nn.Conv2d(512, 128, kernel_size=[1, 1], stride=(1, 1), bias=False)
        self.conv3_4_1x1_reduce_bn = nn.BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        self.conv3_4_1x1_reduce_relu = nn.ReLU()
        self.conv3_4_3x3 = nn.Conv2d(128, 128, kernel_size=[3, 3], stride=(1, 1), padding=(1, 1), bias=False)
        self.conv3_4_3x3_bn = nn.BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        self.conv3_4_3x3_relu = nn.ReLU()
        self.conv3_4_1x1_increase = nn.Conv2d(128, 512, kernel_size=[1, 1], stride=(1, 1), bias=False)
        self.conv3_4_1x1_increase_bn = nn.BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        self.conv3_4_relu = nn.ReLU()
        self.conv4_1_1x1_reduce = nn.Conv2d(512, 256, kernel_size=[1, 1], stride=(2, 2), bias=False)
        self.conv4_1_1x1_reduce_bn = nn.BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        self.conv4_1_1x1_reduce_relu = nn.ReLU()
        self.conv4_1_3x3 = nn.Conv2d(256, 256, kernel_size=[3, 3], stride=(1, 1), padding=(1, 1), bias=False)
        self.conv4_1_3x3_bn = nn.BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        self.conv4_1_3x3_relu = nn.ReLU()
        self.conv4_1_1x1_increase = nn.Conv2d(256, 1024, kernel_size=[1, 1], stride=(1, 1), bias=False)
        self.conv4_1_1x1_increase_bn = nn.BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        self.conv4_1_1x1_proj = nn.Conv2d(512, 1024, kernel_size=[1, 1], stride=(2, 2), bias=False)
        self.conv4_1_1x1_proj_bn = nn.BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        self.conv4_1_relu = nn.ReLU()
        self.conv4_2_1x1_reduce = nn.Conv2d(1024, 256, kernel_size=[1, 1], stride=(1, 1), bias=False)
        self.conv4_2_1x1_reduce_bn = nn.BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        self.conv4_2_1x1_reduce_relu = nn.ReLU()
        self.conv4_2_3x3 = nn.Conv2d(256, 256, kernel_size=[3, 3], stride=(1, 1), padding=(1, 1), bias=False)
        self.conv4_2_3x3_bn = nn.BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        self.conv4_2_3x3_relu = nn.ReLU()
        self.conv4_2_1x1_increase = nn.Conv2d(256, 1024, kernel_size=[1, 1], stride=(1, 1), bias=False)
        self.conv4_2_1x1_increase_bn = nn.BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        self.conv4_2_relu = nn.ReLU()
        self.conv4_3_1x1_reduce = nn.Conv2d(1024, 256, kernel_size=[1, 1], stride=(1, 1), bias=False)
        self.conv4_3_1x1_reduce_bn = nn.BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        self.conv4_3_1x1_reduce_relu = nn.ReLU()
        self.conv4_3_3x3 = nn.Conv2d(256, 256, kernel_size=[3, 3], stride=(1, 1), padding=(1, 1), bias=False)
        self.conv4_3_3x3_bn = nn.BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        self.conv4_3_3x3_relu = nn.ReLU()
        self.conv4_3_1x1_increase = nn.Conv2d(256, 1024, kernel_size=[1, 1], stride=(1, 1), bias=False)
        self.conv4_3_1x1_increase_bn = nn.BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        self.conv4_3_relu = nn.ReLU()
        self.conv4_4_1x1_reduce = nn.Conv2d(1024, 256, kernel_size=[1, 1], stride=(1, 1), bias=False)
        self.conv4_4_1x1_reduce_bn = nn.BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        self.conv4_4_1x1_reduce_relu = nn.ReLU()
        self.conv4_4_3x3 = nn.Conv2d(256, 256, kernel_size=[3, 3], stride=(1, 1), padding=(1, 1), bias=False)
        self.conv4_4_3x3_bn = nn.BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        self.conv4_4_3x3_relu = nn.ReLU()
        self.conv4_4_1x1_increase = nn.Conv2d(256, 1024, kernel_size=[1, 1], stride=(1, 1), bias=False)
        self.conv4_4_1x1_increase_bn = nn.BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        self.conv4_4_relu = nn.ReLU()
        self.conv4_5_1x1_reduce = nn.Conv2d(1024, 256, kernel_size=[1, 1], stride=(1, 1), bias=False)
        self.conv4_5_1x1_reduce_bn = nn.BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        self.conv4_5_1x1_reduce_relu = nn.ReLU()
        self.conv4_5_3x3 = nn.Conv2d(256, 256, kernel_size=[3, 3], stride=(1, 1), padding=(1, 1), bias=False)
        self.conv4_5_3x3_bn = nn.BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        self.conv4_5_3x3_relu = nn.ReLU()
        self.conv4_5_1x1_increase = nn.Conv2d(256, 1024, kernel_size=[1, 1], stride=(1, 1), bias=False)
        self.conv4_5_1x1_increase_bn = nn.BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        self.conv4_5_relu = nn.ReLU()
        self.conv4_6_1x1_reduce = nn.Conv2d(1024, 256, kernel_size=[1, 1], stride=(1, 1), bias=False)
        self.conv4_6_1x1_reduce_bn = nn.BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        self.conv4_6_1x1_reduce_relu = nn.ReLU()
        self.conv4_6_3x3 = nn.Conv2d(256, 256, kernel_size=[3, 3], stride=(1, 1), padding=(1, 1), bias=False)
        self.conv4_6_3x3_bn = nn.BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        self.conv4_6_3x3_relu = nn.ReLU()
        self.conv4_6_1x1_increase = nn.Conv2d(256, 1024, kernel_size=[1, 1], stride=(1, 1), bias=False)
        self.conv4_6_1x1_increase_bn = nn.BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        self.conv4_6_relu = nn.ReLU()
        self.conv5_1_1x1_reduce = nn.Conv2d(1024, 512, kernel_size=[1, 1], stride=(2, 2), bias=False)
        self.conv5_1_1x1_reduce_bn = nn.BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        self.conv5_1_1x1_reduce_relu = nn.ReLU()
        self.conv5_1_3x3 = nn.Conv2d(512, 512, kernel_size=[3, 3], stride=(1, 1), padding=(1, 1), bias=False)
        self.conv5_1_3x3_bn = nn.BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        self.conv5_1_3x3_relu = nn.ReLU()
        self.conv5_1_1x1_increase = nn.Conv2d(512, 2048, kernel_size=[1, 1], stride=(1, 1), bias=False)
        self.conv5_1_1x1_increase_bn = nn.BatchNorm2d(2048, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        self.conv5_1_1x1_proj = nn.Conv2d(1024, 2048, kernel_size=[1, 1], stride=(2, 2), bias=False)
        self.conv5_1_1x1_proj_bn = nn.BatchNorm2d(2048, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        self.conv5_1_relu = nn.ReLU()
        self.conv5_2_1x1_reduce = nn.Conv2d(2048, 512, kernel_size=[1, 1], stride=(1, 1), bias=False)
        self.conv5_2_1x1_reduce_bn = nn.BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        self.conv5_2_1x1_reduce_relu = nn.ReLU()
        self.conv5_2_3x3 = nn.Conv2d(512, 512, kernel_size=[3, 3], stride=(1, 1), padding=(1, 1), bias=False)
        self.conv5_2_3x3_bn = nn.BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        self.conv5_2_3x3_relu = nn.ReLU()
        self.conv5_2_1x1_increase = nn.Conv2d(512, 2048, kernel_size=[1, 1], stride=(1, 1), bias=False)
        self.conv5_2_1x1_increase_bn = nn.BatchNorm2d(2048, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        self.conv5_2_relu = nn.ReLU()
        self.conv5_3_1x1_reduce = nn.Conv2d(2048, 512, kernel_size=[1, 1], stride=(1, 1), bias=False)
        self.conv5_3_1x1_reduce_bn = nn.BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        self.conv5_3_1x1_reduce_relu = nn.ReLU()
        self.conv5_3_3x3 = nn.Conv2d(512, 512, kernel_size=[3, 3], stride=(1, 1), padding=(1, 1), bias=False)
        self.conv5_3_3x3_bn = nn.BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        self.conv5_3_3x3_relu = nn.ReLU()
        self.conv5_3_1x1_increase = nn.Conv2d(512, 2048, kernel_size=[1, 1], stride=(1, 1), bias=False)
        self.conv5_3_1x1_increase_bn = nn.BatchNorm2d(2048, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        self.conv5_3_relu = nn.ReLU()
        self.pool5_7x7_s1 = nn.AvgPool2d(kernel_size=[7, 7], stride=[1, 1], padding=0)
        self.classifier = nn.Conv2d(2048, 8631, kernel_size=[1, 1], stride=(1, 1))

    def forward(self, data):
        conv1_7x7_s2 = self.conv1_7x7_s2(data)
        conv1_7x7_s2_bn = self.conv1_7x7_s2_bn(conv1_7x7_s2)
        conv1_7x7_s2_bnxx = self.conv1_relu_7x7_s2(conv1_7x7_s2_bn)
        pool1_3x3_s2 = self.pool1_3x3_s2(conv1_7x7_s2_bnxx)
        conv2_1_1x1_reduce = self.conv2_1_1x1_reduce(pool1_3x3_s2)
        conv2_1_1x1_reduce_bn = self.conv2_1_1x1_reduce_bn(conv2_1_1x1_reduce)
        conv2_1_1x1_reduce_bnxx = self.conv2_1_1x1_reduce_relu(conv2_1_1x1_reduce_bn)
        conv2_1_3x3 = self.conv2_1_3x3(conv2_1_1x1_reduce_bnxx)
        conv2_1_3x3_bn = self.conv2_1_3x3_bn(conv2_1_3x3)
        conv2_1_3x3_bnxx = self.conv2_1_3x3_relu(conv2_1_3x3_bn)
        conv2_1_1x1_increase = self.conv2_1_1x1_increase(conv2_1_3x3_bnxx)
        conv2_1_1x1_increase_bn = self.conv2_1_1x1_increase_bn(conv2_1_1x1_increase)
        conv2_1_1x1_proj = self.conv2_1_1x1_proj(pool1_3x3_s2)
        conv2_1_1x1_proj_bn = self.conv2_1_1x1_proj_bn(conv2_1_1x1_proj)
        conv2_1 = torch.add(conv2_1_1x1_proj_bn, 1, conv2_1_1x1_increase_bn)
        conv2_1x = self.conv2_1_relu(conv2_1)
        conv2_2_1x1_reduce = self.conv2_2_1x1_reduce(conv2_1x)
        conv2_2_1x1_reduce_bn = self.conv2_2_1x1_reduce_bn(conv2_2_1x1_reduce)
        conv2_2_1x1_reduce_bnxx = self.conv2_2_1x1_reduce_relu(conv2_2_1x1_reduce_bn)
        conv2_2_3x3 = self.conv2_2_3x3(conv2_2_1x1_reduce_bnxx)
        conv2_2_3x3_bn = self.conv2_2_3x3_bn(conv2_2_3x3)
        conv2_2_3x3_bnxx = self.conv2_2_3x3_relu(conv2_2_3x3_bn)
        conv2_2_1x1_increase = self.conv2_2_1x1_increase(conv2_2_3x3_bnxx)
        conv2_2_1x1_increase_bn = self.conv2_2_1x1_increase_bn(conv2_2_1x1_increase)
        conv2_2 = torch.add(conv2_1x, 1, conv2_2_1x1_increase_bn)
        conv2_2x = self.conv2_2_relu(conv2_2)
        conv2_3_1x1_reduce = self.conv2_3_1x1_reduce(conv2_2x)
        conv2_3_1x1_reduce_bn = self.conv2_3_1x1_reduce_bn(conv2_3_1x1_reduce)
        conv2_3_1x1_reduce_bnxx = self.conv2_3_1x1_reduce_relu(conv2_3_1x1_reduce_bn)
        conv2_3_3x3 = self.conv2_3_3x3(conv2_3_1x1_reduce_bnxx)
        conv2_3_3x3_bn = self.conv2_3_3x3_bn(conv2_3_3x3)
        conv2_3_3x3_bnxx = self.conv2_3_3x3_relu(conv2_3_3x3_bn)
        conv2_3_1x1_increase = self.conv2_3_1x1_increase(conv2_3_3x3_bnxx)
        conv2_3_1x1_increase_bn = self.conv2_3_1x1_increase_bn(conv2_3_1x1_increase)
        conv2_3 = torch.add(conv2_2x, 1, conv2_3_1x1_increase_bn)
        conv2_3x = self.conv2_3_relu(conv2_3)
        conv3_1_1x1_reduce = self.conv3_1_1x1_reduce(conv2_3x)
        conv3_1_1x1_reduce_bn = self.conv3_1_1x1_reduce_bn(conv3_1_1x1_reduce)
        conv3_1_1x1_reduce_bnxx = self.conv3_1_1x1_reduce_relu(conv3_1_1x1_reduce_bn)
        conv3_1_3x3 = self.conv3_1_3x3(conv3_1_1x1_reduce_bnxx)
        conv3_1_3x3_bn = self.conv3_1_3x3_bn(conv3_1_3x3)
        conv3_1_3x3_bnxx = self.conv3_1_3x3_relu(conv3_1_3x3_bn)
        conv3_1_1x1_increase = self.conv3_1_1x1_increase(conv3_1_3x3_bnxx)
        conv3_1_1x1_increase_bn = self.conv3_1_1x1_increase_bn(conv3_1_1x1_increase)
        conv3_1_1x1_proj = self.conv3_1_1x1_proj(conv2_3x)
        conv3_1_1x1_proj_bn = self.conv3_1_1x1_proj_bn(conv3_1_1x1_proj)
        conv3_1 = torch.add(conv3_1_1x1_proj_bn, 1, conv3_1_1x1_increase_bn)
        conv3_1x = self.conv3_1_relu(conv3_1)
        conv3_2_1x1_reduce = self.conv3_2_1x1_reduce(conv3_1x)
        conv3_2_1x1_reduce_bn = self.conv3_2_1x1_reduce_bn(conv3_2_1x1_reduce)
        conv3_2_1x1_reduce_bnxx = self.conv3_2_1x1_reduce_relu(conv3_2_1x1_reduce_bn)
        conv3_2_3x3 = self.conv3_2_3x3(conv3_2_1x1_reduce_bnxx)
        conv3_2_3x3_bn = self.conv3_2_3x3_bn(conv3_2_3x3)
        conv3_2_3x3_bnxx = self.conv3_2_3x3_relu(conv3_2_3x3_bn)
        conv3_2_1x1_increase = self.conv3_2_1x1_increase(conv3_2_3x3_bnxx)
        conv3_2_1x1_increase_bn = self.conv3_2_1x1_increase_bn(conv3_2_1x1_increase)
        conv3_2 = torch.add(conv3_1x, 1, conv3_2_1x1_increase_bn)
        conv3_2x = self.conv3_2_relu(conv3_2)
        conv3_3_1x1_reduce = self.conv3_3_1x1_reduce(conv3_2x)
        conv3_3_1x1_reduce_bn = self.conv3_3_1x1_reduce_bn(conv3_3_1x1_reduce)
        conv3_3_1x1_reduce_bnxx = self.conv3_3_1x1_reduce_relu(conv3_3_1x1_reduce_bn)
        conv3_3_3x3 = self.conv3_3_3x3(conv3_3_1x1_reduce_bnxx)
        conv3_3_3x3_bn = self.conv3_3_3x3_bn(conv3_3_3x3)
        conv3_3_3x3_bnxx = self.conv3_3_3x3_relu(conv3_3_3x3_bn)
        conv3_3_1x1_increase = self.conv3_3_1x1_increase(conv3_3_3x3_bnxx)
        conv3_3_1x1_increase_bn = self.conv3_3_1x1_increase_bn(conv3_3_1x1_increase)
        conv3_3 = torch.add(conv3_2x, 1, conv3_3_1x1_increase_bn)
        conv3_3x = self.conv3_3_relu(conv3_3)
        conv3_4_1x1_reduce = self.conv3_4_1x1_reduce(conv3_3x)
        conv3_4_1x1_reduce_bn = self.conv3_4_1x1_reduce_bn(conv3_4_1x1_reduce)
        conv3_4_1x1_reduce_bnxx = self.conv3_4_1x1_reduce_relu(conv3_4_1x1_reduce_bn)
        conv3_4_3x3 = self.conv3_4_3x3(conv3_4_1x1_reduce_bnxx)
        conv3_4_3x3_bn = self.conv3_4_3x3_bn(conv3_4_3x3)
        conv3_4_3x3_bnxx = self.conv3_4_3x3_relu(conv3_4_3x3_bn)
        conv3_4_1x1_increase = self.conv3_4_1x1_increase(conv3_4_3x3_bnxx)
        conv3_4_1x1_increase_bn = self.conv3_4_1x1_increase_bn(conv3_4_1x1_increase)
        conv3_4 = torch.add(conv3_3x, 1, conv3_4_1x1_increase_bn)
        conv3_4x = self.conv3_4_relu(conv3_4)
        conv4_1_1x1_reduce = self.conv4_1_1x1_reduce(conv3_4x)
        conv4_1_1x1_reduce_bn = self.conv4_1_1x1_reduce_bn(conv4_1_1x1_reduce)
        conv4_1_1x1_reduce_bnxx = self.conv4_1_1x1_reduce_relu(conv4_1_1x1_reduce_bn)
        conv4_1_3x3 = self.conv4_1_3x3(conv4_1_1x1_reduce_bnxx)
        conv4_1_3x3_bn = self.conv4_1_3x3_bn(conv4_1_3x3)
        conv4_1_3x3_bnxx = self.conv4_1_3x3_relu(conv4_1_3x3_bn)
        conv4_1_1x1_increase = self.conv4_1_1x1_increase(conv4_1_3x3_bnxx)
        conv4_1_1x1_increase_bn = self.conv4_1_1x1_increase_bn(conv4_1_1x1_increase)
        conv4_1_1x1_proj = self.conv4_1_1x1_proj(conv3_4x)
        conv4_1_1x1_proj_bn = self.conv4_1_1x1_proj_bn(conv4_1_1x1_proj)
        conv4_1 = torch.add(conv4_1_1x1_proj_bn, 1, conv4_1_1x1_increase_bn)
        conv4_1x = self.conv4_1_relu(conv4_1)
        conv4_2_1x1_reduce = self.conv4_2_1x1_reduce(conv4_1x)
        conv4_2_1x1_reduce_bn = self.conv4_2_1x1_reduce_bn(conv4_2_1x1_reduce)
        conv4_2_1x1_reduce_bnxx = self.conv4_2_1x1_reduce_relu(conv4_2_1x1_reduce_bn)
        conv4_2_3x3 = self.conv4_2_3x3(conv4_2_1x1_reduce_bnxx)
        conv4_2_3x3_bn = self.conv4_2_3x3_bn(conv4_2_3x3)
        conv4_2_3x3_bnxx = self.conv4_2_3x3_relu(conv4_2_3x3_bn)
        conv4_2_1x1_increase = self.conv4_2_1x1_increase(conv4_2_3x3_bnxx)
        conv4_2_1x1_increase_bn = self.conv4_2_1x1_increase_bn(conv4_2_1x1_increase)
        conv4_2 = torch.add(conv4_1x, 1, conv4_2_1x1_increase_bn)
        conv4_2x = self.conv4_2_relu(conv4_2)
        conv4_3_1x1_reduce = self.conv4_3_1x1_reduce(conv4_2x)
        conv4_3_1x1_reduce_bn = self.conv4_3_1x1_reduce_bn(conv4_3_1x1_reduce)
        conv4_3_1x1_reduce_bnxx = self.conv4_3_1x1_reduce_relu(conv4_3_1x1_reduce_bn)
        conv4_3_3x3 = self.conv4_3_3x3(conv4_3_1x1_reduce_bnxx)
        conv4_3_3x3_bn = self.conv4_3_3x3_bn(conv4_3_3x3)
        conv4_3_3x3_bnxx = self.conv4_3_3x3_relu(conv4_3_3x3_bn)
        conv4_3_1x1_increase = self.conv4_3_1x1_increase(conv4_3_3x3_bnxx)
        conv4_3_1x1_increase_bn = self.conv4_3_1x1_increase_bn(conv4_3_1x1_increase)
        conv4_3 = torch.add(conv4_2x, 1, conv4_3_1x1_increase_bn)
        conv4_3x = self.conv4_3_relu(conv4_3)
        conv4_4_1x1_reduce = self.conv4_4_1x1_reduce(conv4_3x)
        conv4_4_1x1_reduce_bn = self.conv4_4_1x1_reduce_bn(conv4_4_1x1_reduce)
        conv4_4_1x1_reduce_bnxx = self.conv4_4_1x1_reduce_relu(conv4_4_1x1_reduce_bn)
        conv4_4_3x3 = self.conv4_4_3x3(conv4_4_1x1_reduce_bnxx)
        conv4_4_3x3_bn = self.conv4_4_3x3_bn(conv4_4_3x3)
        conv4_4_3x3_bnxx = self.conv4_4_3x3_relu(conv4_4_3x3_bn)
        conv4_4_1x1_increase = self.conv4_4_1x1_increase(conv4_4_3x3_bnxx)
        conv4_4_1x1_increase_bn = self.conv4_4_1x1_increase_bn(conv4_4_1x1_increase)
        conv4_4 = torch.add(conv4_3x, 1, conv4_4_1x1_increase_bn)
        conv4_4x = self.conv4_4_relu(conv4_4)
        conv4_5_1x1_reduce = self.conv4_5_1x1_reduce(conv4_4x)
        conv4_5_1x1_reduce_bn = self.conv4_5_1x1_reduce_bn(conv4_5_1x1_reduce)
        conv4_5_1x1_reduce_bnxx = self.conv4_5_1x1_reduce_relu(conv4_5_1x1_reduce_bn)
        conv4_5_3x3 = self.conv4_5_3x3(conv4_5_1x1_reduce_bnxx)
        conv4_5_3x3_bn = self.conv4_5_3x3_bn(conv4_5_3x3)
        conv4_5_3x3_bnxx = self.conv4_5_3x3_relu(conv4_5_3x3_bn)
        conv4_5_1x1_increase = self.conv4_5_1x1_increase(conv4_5_3x3_bnxx)
        conv4_5_1x1_increase_bn = self.conv4_5_1x1_increase_bn(conv4_5_1x1_increase)
        conv4_5 = torch.add(conv4_4x, 1, conv4_5_1x1_increase_bn)
        conv4_5x = self.conv4_5_relu(conv4_5)
        conv4_6_1x1_reduce = self.conv4_6_1x1_reduce(conv4_5x)
        conv4_6_1x1_reduce_bn = self.conv4_6_1x1_reduce_bn(conv4_6_1x1_reduce)
        conv4_6_1x1_reduce_bnxx = self.conv4_6_1x1_reduce_relu(conv4_6_1x1_reduce_bn)
        conv4_6_3x3 = self.conv4_6_3x3(conv4_6_1x1_reduce_bnxx)
        conv4_6_3x3_bn = self.conv4_6_3x3_bn(conv4_6_3x3)
        conv4_6_3x3_bnxx = self.conv4_6_3x3_relu(conv4_6_3x3_bn)
        conv4_6_1x1_increase = self.conv4_6_1x1_increase(conv4_6_3x3_bnxx)
        conv4_6_1x1_increase_bn = self.conv4_6_1x1_increase_bn(conv4_6_1x1_increase)
        conv4_6 = torch.add(conv4_5x, 1, conv4_6_1x1_increase_bn)
        conv4_6x = self.conv4_6_relu(conv4_6)
        conv5_1_1x1_reduce = self.conv5_1_1x1_reduce(conv4_6x)
        conv5_1_1x1_reduce_bn = self.conv5_1_1x1_reduce_bn(conv5_1_1x1_reduce)
        conv5_1_1x1_reduce_bnxx = self.conv5_1_1x1_reduce_relu(conv5_1_1x1_reduce_bn)
        conv5_1_3x3 = self.conv5_1_3x3(conv5_1_1x1_reduce_bnxx)
        conv5_1_3x3_bn = self.conv5_1_3x3_bn(conv5_1_3x3)
        conv5_1_3x3_bnxx = self.conv5_1_3x3_relu(conv5_1_3x3_bn)
        conv5_1_1x1_increase = self.conv5_1_1x1_increase(conv5_1_3x3_bnxx)
        conv5_1_1x1_increase_bn = self.conv5_1_1x1_increase_bn(conv5_1_1x1_increase)
        conv5_1_1x1_proj = self.conv5_1_1x1_proj(conv4_6x)
        conv5_1_1x1_proj_bn = self.conv5_1_1x1_proj_bn(conv5_1_1x1_proj)
        conv5_1 = torch.add(conv5_1_1x1_proj_bn, 1, conv5_1_1x1_increase_bn)
        conv5_1x = self.conv5_1_relu(conv5_1)
        conv5_2_1x1_reduce = self.conv5_2_1x1_reduce(conv5_1x)
        conv5_2_1x1_reduce_bn = self.conv5_2_1x1_reduce_bn(conv5_2_1x1_reduce)
        conv5_2_1x1_reduce_bnxx = self.conv5_2_1x1_reduce_relu(conv5_2_1x1_reduce_bn)
        conv5_2_3x3 = self.conv5_2_3x3(conv5_2_1x1_reduce_bnxx)
        conv5_2_3x3_bn = self.conv5_2_3x3_bn(conv5_2_3x3)
        conv5_2_3x3_bnxx = self.conv5_2_3x3_relu(conv5_2_3x3_bn)
        conv5_2_1x1_increase = self.conv5_2_1x1_increase(conv5_2_3x3_bnxx)
        conv5_2_1x1_increase_bn = self.conv5_2_1x1_increase_bn(conv5_2_1x1_increase)
        conv5_2 = torch.add(conv5_1x, 1, conv5_2_1x1_increase_bn)
        conv5_2x = self.conv5_2_relu(conv5_2)
        conv5_3_1x1_reduce = self.conv5_3_1x1_reduce(conv5_2x)
        conv5_3_1x1_reduce_bn = self.conv5_3_1x1_reduce_bn(conv5_3_1x1_reduce)
        conv5_3_1x1_reduce_bnxx = self.conv5_3_1x1_reduce_relu(conv5_3_1x1_reduce_bn)
        conv5_3_3x3 = self.conv5_3_3x3(conv5_3_1x1_reduce_bnxx)
        conv5_3_3x3_bn = self.conv5_3_3x3_bn(conv5_3_3x3)
        conv5_3_3x3_bnxx = self.conv5_3_3x3_relu(conv5_3_3x3_bn)
        conv5_3_1x1_increase = self.conv5_3_1x1_increase(conv5_3_3x3_bnxx)
        conv5_3_1x1_increase_bn = self.conv5_3_1x1_increase_bn(conv5_3_1x1_increase)
        conv5_3 = torch.add(conv5_2x, 1, conv5_3_1x1_increase_bn)
        conv5_3x = self.conv5_3_relu(conv5_3)
        pool5_7x7_s1 = self.pool5_7x7_s1(conv5_3x)
        pool5_7x7_s1 = pool5_7x7_s1.view(pool5_7x7_s1.shape[0], pool5_7x7_s1.shape[1])
        classifier = self.classifier(pool5_7x7_s1)
        # classifier = classifier_preflatten.view(classifier_preflatten.size(0), -1)
        return classifier, pool5_7x7_s1

And here is the code for whole architecture:

class VGGFace(nn.Module):

    def __init__(self, desc='resnet50):
        super().__init__()

        # init descriptor
        if desc=="resnet50":
            self.descriptor = ResNet50()
            assert desc_out_shape == 512

            weights_path = "weights/resnet50_ft_dag.pth"

            # load weights, pretrained on VGGFace2 dataset
            state_dict = torch.load(weights_path)
            self.descriptor.load_state_dict(state_dict)

            # freeze pretrained descriptor, as it should not be updated during further training
            for param in self.descriptor.parameters():
                param.requires_grad = False

            # change last layer of the model, so it will return not 8631d embiddings, but 512d
            num_ftrs = self.descriptor.classifier.in_channels
            self.descriptor.classifier = nn.Linear(num_ftrs, 512)


        # create two FC layers and dropout
        self.fc1 = nn.Linear(desc_out_shape * 3, 128)
        self.fc2 = nn.Linear(128, 1)
        self.dropout = nn.Dropout(0.5)
        self.relu = nn.ReLU()


    def forward(self, x1, x2):
        # get embedding, by passing images through descriptor
        x1_emb = self.descriptor(x1)[0]
        x2_emb = self.descriptor(x2)[0]

        # different combinations of embeddings
        # x^2 - y^2
        comb_1 = torch.sub(x1_emb.pow(2), x2_emb.pow(2))

        # (x - y)^2
        comb_2 = torch.sub(x1_emb, x2_emb).pow(2)

        # x * y
        comb_3 = x1_emb * x2_emb

        # concatenate all combinations into 1 vector
        x = torch.cat([comb_1, comb_2, comb_3], dim=1)


        # pass through FC layers resulting combination
        x = self.relu(self.fc1(x))
        x = self.dropout(x)
        x = self.fc2(x)

        return x

After loading pretrained weights into the feature extractor, I freeze all parameters in in, and change last layer, so that will return not 8631 length embedding, but 512.

And here is a part of my training code:

model = VGGFace(desc='resnet50')

device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
model.to(device)

criterion = nn.BCEWithLogitsLoss()
optimizer = optim.Adam(model.parameters(), lr=0.0003)

for epoch in range(NUM_EPOCHS):
        model.train()
        running_loss = []
        for batch in tqdm(train_dataloader):
            x1, x2, labels = batch

            # move data to the corresponding hardware (CPU or GPU)
            x1 = x1.to(device)
            x2 = x2.to(device)
            labels = labels.to(device)

            # zero the parameter gradients
            optimizer.zero_grad()

            # forward + backward + optimize
            outputs = model(x1, x2)

            loss = criterion(outputs, labels)
            loss.backward()
            optimizer.step()

            # multiply by the size of batch
            running_loss += loss.item() * x1.size(0)

        train_loss = running_loss / len(train_dataset)
        running_loss = 0

        model.eval()

        for batch in tqdm(val_dataloader):
            x1, x2, labels = batch

            # move data to the corresponding hardware (CPU or GPU)
            x1 = x1.to(device)
            x2 = x2.to(device)
            labels = labels.to(device)

            # forward
            with torch.no_grad():
                outputs = model(x1, x2)

            loss = criterion(outputs.float(), labels.float())

            # multiply by the size of batch
            running_loss += loss.item() * x1.size(0)

        val_loss = running_loss / len(val_dataset)

Data: 2 images, 224x224, label 0 or 1

I’ve read tons of discussion, articles, and tutorials, but it doesn’t help me at all.

I’ve tried:

  • decrease number of layer (remove a different combination of the embeddings)
  • decrease number of neurons (originally, there are ~1.3 Mil parameters to train, I’ve tried reduce to 300k - didn’t help)
  • training on different dataset sizes, from few hundreds up to 200k
  • adding Dropout for regularization
  • adding relu between FC
  • trained for different number of epochs, from 20 to 150

I assume that can be caused by one of 3 issues:

  • model has to small capacity, and can not learn such a hard problem
  • model overfits, but then I don’t know how to solve that
  • there is some stupid bug in the code, maybe connected with parameters updating

I will be very-very grateful for any help.