ValueError: num_channels must be divisible by num_groups with default config

[Carlos779988]

Running training with the provided config results in a GroupNorm error:

Traceback (most recent call last):
  File "/home/l/scratch/EndoIR/EndoIR/endoir/train.py", line 12, in <module>
    train_pipeline(root_path)
  File "/mnt/scratch/l/EndoIR/BasicSR-light/basicsr/train.py", line 124, in train_pipeline
    model = build_model(opt)
  File "/mnt/scratch/l/EndoIR/BasicSR-light/basicsr/models/__init__.py", line 27, in build_model
    model = MODEL_REGISTRY.get(opt['model_type'])(opt)
  File "/mnt/scratch/l/EndoIR/EndoIR/endoir/models/EndoIR_model.py", line 34, in __init__
    self.unet = build_network(opt['network_unet'])
  File "/mnt/scratch/l/EndoIR/BasicSR-light/basicsr/archs/__init__.py", line 22, in build_network
    net = ARCH_REGISTRY.get(network_type)(**opt)
  File "/mnt/scratch/l/EndoIR/EndoIR/endoir/archs/dualunet_arch.py", line 625, in __init__
    SingleResnetBlocWithAttn(  
  File "/mnt/scratch/l/EndoIR/EndoIR/endoir/archs/dualunet_arch.py", line 357, in __init__
    self.res_block = DecodeResnetBlock(
  File "/mnt/scratch/l/EndoIR/EndoIR/endoir/archs/dualunet_arch.py", line 187, in __init__
    self.block1 = Block(dim_out, dim_out, groups=norm_groups)
  File "/mnt/scratch/l/EndoIR/EndoIR/endoir/archs/dualunet_arch.py", line 151, in __init__
    nn.GroupNorm(groups, dim),
  File "/home/l/miniforge3/envs/EndoIR/lib/python3.10/site-packages/torch/nn/modules/normalization.py", line 258, in __init__
    raise ValueError('num_channels must be divisible by num_groups')
ValueError: num_channels must be divisible by num_groups

Questions:

1. What is the correct norm_groups value for this config?
2. Could you provide a working config for the SegSTRONG-C experiments?

[ff98li]

Yea I also got the same error. It's caused by the decoder reducing channel dimensions from 32 to 16, which is clearly not divisible by norm_groups of 32 set in their train.ymal file. It would be great if the authors could provide a clarification. I looked into the code and suspect that they were using norm_groups of 4 instead of 32:

EndoIR/EndoIR/endoir/archs/dualunet_arch.py

Lines 340 to 358 in ffef9d7

	class SingleResnetBlocWithAttn(nn.Module):
	def __init__(self, dim, dim_out, *, noise_level_emb_dim=None, norm_groups=4, dropout=0, ffn_factor=2.66,\
	with_attn=False, use_affine_level=False, attn_type=None, topk_ratio=None):
	super().__init__()
	self.with_attn = with_attn
	self.decoding = nn.Conv2d(dim, dim_out, kernel_size=1, bias=False)
	if noise_level_emb_dim is not None:
	self.noise_func = FeatureWiseAffine(
	noise_level_emb_dim, dim_out, use_affine_level)
	self.ffn = FeedForward(dim_out, ffn_factor, bias=False)
	self.norm = LayerNorm(dim_out, LayerNorm_type='WithBias')
	if topk_ratio is not None:
	self.use_topk = True
	self.routing = Feature_Router(dim_out, int(topk_ratio * dim_out))
	dim_out = int(topk_ratio * dim_out)
	self.res_block = DecodeResnetBlock(
	dim_out, dim_out, noise_level_emb_dim, norm_groups=norm_groups, dropout=dropout, use_affine_level=use_affine_level)