Problem adding AdaptiveAvgPool2d to finetune a pretrained densenet201

I want to add the following classifier to finetune a pretrained densenet201. The problem is it fails with a dimension error and I can’t figure out what’s up. Please help out

class ClassifierNew(nn.Module):
    def __init__(self, inp = 1920, h1=1024, out = 7, d=0.35):
        self.ap = nn.AdaptiveAvgPool2d((1,1)) = nn.AdaptiveMaxPool2d((1,1))
        self.fla = Flatten()
        self.bn0 = nn.BatchNorm1d(inp*2,eps=1e-05, momentum=0.1, affine=True)
        self.dropout0 = nn.Dropout(d)
        self.fc1 = nn.Linear(inp*2, h1)
        self.bn1 = nn.BatchNorm1d(h1,eps=1e-05, momentum=0.1, affine=True)
        self.dropout1 = nn.Dropout(d)
        self.fc2 = nn.Linear(h1, out)
    def forward(self, x):
        ap = self.ap(x)
        mp =
        x =,mp),dim=1)
        x = self.fla(x)
        x = self.bn0(x)
        x = self.dropout0(x)
        x = F.relu(self.fc1(x))
        x = self.bn1(x)
        x = self.dropout1(x)         
        x = self.fc2(x)
        return x

I get the following traceback

torch.Size([8, 1920])
torch.Size([8, 1920])
torch.Size([8, 1920])
torch.Size([8, 1920])
Traceback (most recent call last):
  File "", line 130, in <module>
    model = trainroutines.train_validate_model(model, optimizer, criterion, config.n_epochs,data_loader_train , data_loader_val, config.minibatch_print_freq)
  File "/dccstor/sgdermatology/scripts/sg_fair_Newton6/", line 55, in train_validate_model
    output = model(data)
  File "/u/nkinyanj/anaconda3/lib/python3.7/site-packages/torch/nn/modules/", line 493, in __call__
    result = self.forward(*input, **kwargs)
  File "/u/nkinyanj/anaconda3/lib/python3.7/site-packages/torch/nn/parallel/", line 152, in forward
    outputs = self.parallel_apply(replicas, inputs, kwargs)
  File "/u/nkinyanj/anaconda3/lib/python3.7/site-packages/torch/nn/parallel/", line 162, in parallel_apply
    return parallel_apply(replicas, inputs, kwargs, self.device_ids[:len(replicas)])
  File "/u/nkinyanj/anaconda3/lib/python3.7/site-packages/torch/nn/parallel/", line 83, in parallel_apply
    raise output
  File "/u/nkinyanj/anaconda3/lib/python3.7/site-packages/torch/nn/parallel/", line 59, in _worker
    output = module(*input, **kwargs)
  File "/u/nkinyanj/anaconda3/lib/python3.7/site-packages/torch/nn/modules/", line 493, in __call__
    result = self.forward(*input, **kwargs)
  File "/u/nkinyanj/anaconda3/lib/python3.7/site-packages/torchvision/models/", line 122, in forward
    out = self.classifier(out)
  File "/u/nkinyanj/anaconda3/lib/python3.7/site-packages/torch/nn/modules/", line 493, in __call__
    result = self.forward(*input, **kwargs)
  File "/dccstor/sgdermatology/scripts/sg_fair_Newton6/", line 570, in forward
    ap = self.ap(x)
  File "/u/nkinyanj/anaconda3/lib/python3.7/site-packages/torch/nn/modules/", line 493, in __call__
    result = self.forward(*input, **kwargs)
  File "/u/nkinyanj/anaconda3/lib/python3.7/site-packages/torch/nn/modules/", line 1060, in forward
    return F.adaptive_avg_pool2d(input, self.output_size)
  File "/u/nkinyanj/anaconda3/lib/python3.7/site-packages/torch/nn/", line 788, in adaptive_avg_pool2d
    _output_size = _list_with_default(output_size, input.size())
  File "/u/nkinyanj/anaconda3/lib/python3.7/site-packages/torch/nn/modules/", line 22, in _list_with_default
    raise ValueError('Input dimension should be at least {}'.format(len(out_size) + 1))
ValueError: Input dimension should be at least 3

Class Flatten is defined as

class Flatten(nn.Module):
    def forward(self, input):
        return input.view(input.size(0), -1)

Based on your model architecture, you should provide a 4-dimensional input as [batch_size, channels, height, width].
Both adaptive pooling layers will create an output of shape [batch_size, channels, 1, 1].
These outputs will be concatenated in their channel dimension and flattened, yielding an activation of [batch_size, channels*2].
The following nn.BatchNorm1d layer will normalize these channels and assume a sequence length of 1.
Given that its number of channels is defined as inp*2, your model should work with an input of shape [batch_size, 1920, h, w], where batch_size>=2 and h, w >= 1, e.g.:
[2, 1920, 2, 2].

Your output however seems to show that you are passing an input of [8, 1920].
Could you explain your use case and the architecture a bit, as I’m not sure to understand it completely?

1 Like

Thanks @ptrblck. I’m finetuning a pretrained densenet201 from PyTorch. The input image is RGB, image dimension is 600x450. I’m freezing the first layers and updating the weights of the other successive layers. When I replace the default pretrained densenet201 classifier with this class, I get the dimension error.

I have same problem, the problem seems to be that densenet is flattening the input before sending it to classifier. @PapRazzi Did you find any solution to this problem?

@gaurav8341 same issue here. Any solutions?