Batch size = 1 leads to error

[matteosodano]

I tried to train with default parameters and sunrgbd dataset with batch_size set to 1. The following error arises:

  File "train.py", line 577, in <module>
    train_main()
  File "train.py", line 206, in train_main
    label_downsampling_rates, args.epochs, args.results_dir, monitoring_images, debug_mode=args.debug)
  File "train.py", line 310, in train_one_epoch
    pred_scales = model(image, depth)
  File "/home/matteo/anaconda3/envs/rgbd_segmentation/lib/python3.7/site-packages/torch/nn/modules/module.py", line 541, in __call__
    result = self.forward(*input, **kwargs)
  File "/home/matteo/Code/Semantic_Segmentation/ESANet/src/models/model.py", line 238, in forward
    out = self.context_module(fuse)
  File "/home/matteo/anaconda3/envs/rgbd_segmentation/lib/python3.7/site-packages/torch/nn/modules/module.py", line 541, in __call__
    result = self.forward(*input, **kwargs)
  File "/home/matteo/Code/Semantic_Segmentation/ESANet/src/models/context_modules.py", line 72, in forward
    y = f(x)
  File "/home/matteo/anaconda3/envs/rgbd_segmentation/lib/python3.7/site-packages/torch/nn/modules/module.py", line 541, in __call__
    result = self.forward(*input, **kwargs)
  File "/home/matteo/anaconda3/envs/rgbd_segmentation/lib/python3.7/site-packages/torch/nn/modules/container.py", line 92, in forward
    input = module(input)
  File "/home/matteo/anaconda3/envs/rgbd_segmentation/lib/python3.7/site-packages/torch/nn/modules/module.py", line 541, in __call__
    result = self.forward(*input, **kwargs)
  File "/home/matteo/anaconda3/envs/rgbd_segmentation/lib/python3.7/site-packages/torch/nn/modules/container.py", line 92, in forward
    input = module(input)
  File "/home/matteo/anaconda3/envs/rgbd_segmentation/lib/python3.7/site-packages/torch/nn/modules/module.py", line 541, in __call__
    result = self.forward(*input, **kwargs)
  File "/home/matteo/anaconda3/envs/rgbd_segmentation/lib/python3.7/site-packages/torch/nn/modules/batchnorm.py", line 81, in forward
    exponential_average_factor, self.eps)
  File "/home/matteo/anaconda3/envs/rgbd_segmentation/lib/python3.7/site-packages/torch/nn/functional.py", line 1666, 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, 256, 1, 1])

Any hint on how to fix it?

[danielS91]

The error comes from the global average pooling in the context module. The output of the global branch has no spatial resolution any more (see your error). If you are using batchsize 1, there is only one element left for determining mean and std in the subsequent batch normalization, which is just not possible.
So you need a batchsize > 1 or you have to remove the global branch or the entire context module.

Note that this is requirement is only for training. Inference with batchsize 1 works well as the inference mode for batch normalization is different.