How to use intermediate blocks in densenet?

In the torch vision

for i, num_layers in enumerate(block_config):    
      block = _DenseBlock(num_layers=num_layers, num_input_features=num_features, bn_size=bn_size, growth_rate=growth_rate, drop_rate=drop_rate) 
      self.features.add_module('denseblock%d' % (i + 1), block) 

I want to apply a classifier for the block ( likes deep supervision) . How can I implement it? Thanks

You can load the pretrained model and then take apart the network and add your classifiers and have these outputs returned in your forward pass. You can take apart the network by iterating through named_children or children.

Thanks. But it is not clear to me. I means I want to obtain the graph likes below

I haven’t tested this, but this is the general idea.

import torch.nn as nn
import torchvision.models as model_zoo

class Net(nn.Module):
  def __init__(self):
    super(Net, self).__init__()
    self.dnet = model_zoo.densenet121(pretrained=True)
  def forward(self, input):
    dblock_outs = []
    x = input
    for name, module in self.dnet["features"].named_children():
      if "denseblock" in name:
        x = module(x)
        x = module(x)
    x = self.dnet["classifier"](x)
    return x, dblock_outs

model = Net()
out, dblock_outs_for_classifiers = model(minibatch)
1 Like