ValueError: Expected more than 1 value per channel when training, got input size torch.Size([1, 32, 1, 1])

Question

ValueError: Expected more than 1 value per channel when training, got input size torch.Size([1, 32, 1, 1])

571502680 opened this issue 5 years ago · comments

LiQi commented 5 years ago

How does it solve？

Tramac · Answer 1 · Thu May 16 2019 22:21:29 GMT+0800 (China Standard Time)

Hi, What's the size of the input image?

LiQi · Answer 2 · Thu May 16 2019 22:59:39 GMT+0800 (China Standard Time)

Hi, What's the size of the input image?
4806403，Does segmentation have relation with the size of input image？

Tramac · Answer 3 · Thu May 16 2019 23:17:28 GMT+0800 (China Standard Time)

Segmentation task don't have relation with the size of input image, but the spatial resolution of feature map will be 1/8. Can the size be written as weight * width?
Please give a detailed error log.

Tramac · Answer 4 · Thu May 16 2019 23:59:51 GMT+0800 (China Standard Time)

Please ensure batch_size > 1.

LiQi · Answer 5 · Fri May 17 2019 07:43:45 GMT+0800 (China Standard Time)

Segmentation task don't have relation with the size of input image, but the spatial resolution of feature map will be 1/8. Can the size be written as weight * width?
Please give a detailed error log.
When I use the cityspace dataset, it occurs to this error.and batch_size =32

Tramac · Answer 6 · Fri May 17 2019 09:29:03 GMT+0800 (China Standard Time)

Hi, is there a more detailed error log?

LiQi · Answer 7 · Fri May 17 2019 16:52:57 GMT+0800 (China Standard Time)

Hi, is there a more detailed error log?

Nothing，i delete BatchNorm and it is ok.

Tramac · Answer 8 · Fri May 17 2019 16:55:35 GMT+0800 (China Standard Time)

BatchNorm is important , maybe this will affect performance.

LiQi · Answer 9 · Fri May 17 2019 16:57:57 GMT+0800 (China Standard Time)

-_-! Does BatchNorm have connection with Batchsize?it doesn't work if not delete BN

LiQi · Answer 10 · Fri May 17 2019 17:05:30 GMT+0800 (China Standard Time)

When classifiying,is the background a class?

Tramac · Answer 11 · Fri May 17 2019 17:13:30 GMT+0800 (China Standard Time)

Namespace(aux=False, aux_weight=0.4, base_size=520, batch_size=1, crop_size=224, dataset='citys', device=device(type='cpu'), epochs=160, eval=False, lr=0.01, model='fast_scnn', momentum=0.9, no_val=True, resume=None, save_folder='./weights', start_epoch=0, train_split='train', weight_decay=0.0001)
Found 2975 images in the folder ./datasets/citys/leftImg8bit/train
Found 500 images in the folder ./datasets/citys/leftImg8bit/val
w/ class balance
Starting Epoch: 0, Total Epochs: 160
Traceback (most recent call last):
  File "/home/Fast-SCNN-pytorch/train.py", line 200, in <module>
    trainer.train()
  File "/home/Fast-SCNN-pytorch/train.py", line 135, in train
    outputs = self.model(images)
  File "/home/.pyenv/versions/anaconda3-4.4.0/lib/python3.6/site-packages/torch/nn/modules/module.py", line 493, in __call__
    result = self.forward(*input, **kwargs)
  File "/home/Fast-SCNN-pytorch/models/fast_scnn.py", line 36, in forward
    x = self.global_feature_extractor(higher_res_features)
  File "/home/.pyenv/versions/anaconda3-4.4.0/lib/python3.6/site-packages/torch/nn/modules/module.py", line 493, in __call__
    result = self.forward(*input, **kwargs)
  File "/home/Fast-SCNN-pytorch/models/fast_scnn.py", line 186, in forward
    x = self.ppm(x)
  File "/home/.pyenv/versions/anaconda3-4.4.0/lib/python3.6/site-packages/torch/nn/modules/module.py", line 493, in __call__
    result = self.forward(*input, **kwargs)
  File "/home/Fast-SCNN-pytorch/models/fast_scnn.py", line 139, in forward
    feat1 = self.upsample(self.conv1(self.pool(x, 1)), size)
  File "/home/.pyenv/versions/anaconda3-4.4.0/lib/python3.6/site-packages/torch/nn/modules/module.py", line 493, in __call__
    result = self.forward(*input, **kwargs)
  File "/home/Fast-SCNN-pytorch/models/fast_scnn.py", line 61, in forward
    return self.conv(x)
  File "/home/.pyenv/versions/anaconda3-4.4.0/lib/python3.6/site-packages/torch/nn/modules/module.py", line 493, in __call__
    result = self.forward(*input, **kwargs)
  File "/home/.pyenv/versions/anaconda3-4.4.0/lib/python3.6/site-packages/torch/nn/modules/container.py", line 92, in forward
    input = module(input)
  File "/home/.pyenv/versions/anaconda3-4.4.0/lib/python3.6/site-packages/torch/nn/modules/module.py", line 493, in __call__
    result = self.forward(*input, **kwargs)
  File "/home/.pyenv/versions/anaconda3-4.4.0/lib/python3.6/site-packages/torch/nn/modules/batchnorm.py", line 83, in forward
    exponential_average_factor, self.eps)
  File "/home/.pyenv/versions/anaconda3-4.4.0/lib/python3.6/site-packages/torch/nn/functional.py", line 1693, in batch_norm
    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, 32, 1, 1])

background is class 0.
BatchNorm has connection with batch size, when the model in train() mode, the batch_size cann't be set 1.
If I set batch_size=1, it will appear bug above, if batch_size = 4 or others (8, 16,...), there is no bug.

LiQi · Answer 12 · Fri May 17 2019 17:18:03 GMT+0800 (China Standard Time)

However I set batch_size = 4, it is amazing!

Tramac · Answer 13 · Fri May 17 2019 17:20:36 GMT+0800 (China Standard Time)

I will fix bug when batch_size=32 as soon as possible.

LiQi · Answer 14 · Fri May 17 2019 17:22:37 GMT+0800 (China Standard Time)

OK，Thank you very much!

Sanchayan Sarkar · Answer 15 · Mon Oct 11 2021 10:53:28 GMT+0800 (China Standard Time)

There is one solution . If you encounter a batch_size = 1 , then simply add a dummy tensor of Zeros . And before computing loss simply eliminate that tensor so its not affecting gradients. But this would require a computational overhead.