Tensor value error with inception_v3 model

vision
1

I’m fixing a inception v3 model for image captioning.
But after 1 epoch of training, i got value error.
How can I fix this problem?

Traceback (most recent call last):
  File "train.py", line 107, in <module>
    main(args)
  File "train.py", line 61, in main
    features = encoder(images)
  File "/home/mango/anaconda3/envs/img/lib/python3.7/site-packages/torch/nn/modules/module.py", line 727, in _call_impl
    result = self.forward(*input, **kwargs)
  File "/home/mango/바탕화면/lab/pytorch-tutorial/tutorials/03-advanced/image_captioning/model.py", line 20, in forward
    embed = self.inception(images)
  File "/home/mango/anaconda3/envs/img/lib/python3.7/site-packages/torch/nn/modules/module.py", line 727, in _call_impl
    result = self.forward(*input, **kwargs)
  File "/home/mango/anaconda3/envs/img/lib/python3.7/site-packages/torchvision/models/inception.py", line 199, in forward
    x, aux = self._forward(x)
  File "/home/mango/anaconda3/envs/img/lib/python3.7/site-packages/torchvision/models/inception.py", line 169, in _forward
    aux = self.AuxLogits(x)
  File "/home/mango/anaconda3/envs/img/lib/python3.7/site-packages/torch/nn/modules/module.py", line 727, in _call_impl
    result = self.forward(*input, **kwargs)
  File "/home/mango/anaconda3/envs/img/lib/python3.7/site-packages/torchvision/models/inception.py", line 419, in forward
    x = self.conv1(x)
  File "/home/mango/anaconda3/envs/img/lib/python3.7/site-packages/torch/nn/modules/module.py", line 727, in _call_impl
    result = self.forward(*input, **kwargs)
  File "/home/mango/anaconda3/envs/img/lib/python3.7/site-packages/torchvision/models/inception.py", line 440, in forward
    x = self.bn(x)
  File "/home/mango/anaconda3/envs/img/lib/python3.7/site-packages/torch/nn/modules/module.py", line 727, in _call_impl
    result = self.forward(*input, **kwargs)
  File "/home/mango/anaconda3/envs/img/lib/python3.7/site-packages/torch/nn/modules/batchnorm.py", line 136, in forward
    self.weight, self.bias, bn_training, exponential_average_factor, self.eps)
  File "/home/mango/anaconda3/envs/img/lib/python3.7/site-packages/torch/nn/functional.py", line 2054, in batch_norm
    _verify_batch_size(input.size())
  File "/home/mango/anaconda3/envs/img/lib/python3.7/site-packages/torch/nn/functional.py", line 2037, in _verify_batch_size
    raise ValueError('Expected more than 1 value per channel when training, got input size {}'.format(size))
ValueError: Expected more than 1 value per channel when training, got input size torch.Size([1, 768, 1, 1])


class Inception(nn.Module):
    def __init__(self, embed_size):
        super(Inception, self).__init__()
        self.inception = models.inception_v3(pretrained=True)
        #in_features = self.inception.fc.in_features
        #self.linear = nn.Linear(in_features, embed_size)
        self.linear = nn.Linear(self.inception.fc.in_features, embed_size)
        self.bn = nn.BatchNorm1d(embed_size, momentum=0.01)
        #self.inception.fc = self.linear
    
    def forward(self, images):
        embed = self.inception(images)
        return embed

class DecoderRNN(nn.Module):
    def __init__(self, embed_size, hidden_size, vocab_size, num_layers, max_seq_length=20):
        """Set the hyper-parameters and build the layers."""
        super(DecoderRNN, self).__init__()
        self.embed = nn.Embedding(vocab_size, embed_size)
        self.lstm = nn.LSTM(embed_size, hidden_size, num_layers, batch_first=True)
        self.linear = nn.Linear(hidden_size, vocab_size)
        self.max_seg_length = max_seq_length
        
    def forward(self, features, captions, lengths):
        """Decode image feature vectors and generates captions."""
        embeddings = self.embed(captions)
        #print(embeddings.size())
        #print(features.logits.unsqueeze(1).size())
        embeddings = torch.cat((features.logits.unsqueeze(1), embeddings), 1)

        packed = pack_padded_sequence(embeddings, lengths, batch_first=True) 
        hiddens, _ = self.lstm(packed)
        outputs = self.linear(hiddens[0])
        return outputs
 
    def sample(self, features, states=None):
        """Generate captions for given image features using greedy search."""
        sampled_ids = []
        inputs = features.unsqueeze(1)
        for i in range(self.max_seg_length):
            hiddens, states = self.lstm(inputs, states)          # hiddens: (batch_size, 1, hidden_size)
            outputs = self.linear(hiddens.squeeze(1))            # outputs:  (batch_size, vocab_size)
            _, predicted = outputs.max(1)                        # predicted: (batch_size)
            sampled_ids.append(predicted)
            inputs = self.embed(predicted)                       # inputs: (batch_size, embed_size)
            inputs = inputs.unsqueeze(1)                         # inputs: (batch_size, 1, embed_size)
        sampled_ids = torch.stack(sampled_ids, 1)                # sampled_ids: (batch_size, max_seq_length)
        return sampled_ids
2

The error is raised by some batchnorm layers, which cannot calculate the stats from a single sample.
If you leave the model in training mode, you would have to provide at least 2 samples in the batch.
Assuming this error is only raised at the end of the epoch, I guess the last batch might be smaller than the rest and you could drop it via setting drop_last=True in the DataLoader.

2 Likes
3

Thank you. I solved this problem with drop_last=True.