PyTorch Diary

• Feature extractor:

• class FeatureExtractor(nn.Module):
      def __init__(self, submodule, extracted_layers):
          self.submodule = submodule
  
      def forward(self, x):
          outputs = []
          for name, module in self.submodule._modules.items():
              x = module(x)
              if name in self.extracted_layers:
                  outputs += [x]
          return outputs + [x]
  

• Feature Perceptual Loss:
• fast-neural-style
• pytorch-implementation-of-perceptual-losses-for-real-time-style-transfer
• Extracting features from pretrained model:
• class Vgg16_pretrained(nn.Module):
      def __init__(self):
          super(Vgg16_pretrained, self).__init__()
          features = list(vgg19(pretrained = True).features)[:36]
          # interested in following layers with corresponding indices
          # 1:  relu_1_1
          # 6:  relu_2_1
          # 11: relu_3_1
          # 20: relu_4_1
          # 29: relu_5_1
          # change it to eval() mode so that gradient won't be computed      
          self.features = nn.ModuleList(features).eval()
         
      def forward(self, x):
          results = []
          for ii,model in enumerate(self.features):           
              x = model(x)
              #pdb.set_trace()
              if ii in {1,6,11}:
                  print(ii, ':', model)
                  print(x.shape)
                  results.append(x)
• Loading pretrained model
• # init weights
  model_v1 =  EMNetv2(models)
  init_weights = []
  [init_weights.append(param.data) for param in model_v1.parameters()]
  
  
  
  # pretrained weights
  model_v2 =  EMNetv2(models)
  model_v2.load_state_dict(torch.load(model_ccd_pretrained))
  pretrained_weights = []
  [pretrained_weights.append(param.data) for param in model_v2.parameters()]

Funfact: artificial-intelligence-salaries-openai