assert len(set(num_channels_list)) == 1

[jxncyym]

Describe the issue:
I compress the model: https://github.com/chenjun2hao/DDRNet.pytorch

when add this code "pruner._unwrap_model()", meet the error

Traceback (most recent call last):
File "/algdata02/yiming.yu/DDRNet.pytorch_pruner/tools/pruner_nni_L1filter.py", line 142, in
main()
File "/algdata02/yiming.yu/DDRNet.pytorch_pruner/tools/pruner_nni_L1filter.py", line 130, in main
m_speedup.speedup_model()
File "/algdata02/yiming.yu/DDRNet.pytorch_pruner/envp_3/lib/python3.6/site-packages/nni/compression/pytorch/speedup/compressor.py", line 500, in speedup_model
fix_mask_conflict(self.masks, self.bound_model, self.dummy_input)
File "/algdata02/yiming.yu/DDRNet.pytorch_pruner/envp_3/lib/python3.6/site-packages/nni/compression/pytorch/utils/mask_conflict.py", line 50, in fix_mask_conflict
masks = fix_channel_mask.fix_mask()
File "/algdata02/yiming.yu/DDRNet.pytorch_pruner/envp_3/lib/python3.6/site-packages/nni/compression/pytorch/utils/mask_conflict.py", line 259, in fix_mask
assert len(set(num_channels_list)) == 1
AssertionError

Environment:

• NNI version:2.4
• Training service (local|remote|pai|aml|etc):
• Client OS: window
• Server OS (for remote mode only):unbuntu 18
• Python version:3.6
• PyTorch/TensorFlow version:pytorch 1.7.1
• Is conda/virtualenv/venv used?:virtualenv
• Is running in Docker?:no

Configuration:

• Experiment config (remember to remove secrets!): L1FilterPruner config_list = [{ 'sparsity': 0.7, 'op_types': ['Conv2d'] }]
• Search space:

Log message:

• nnimanager.log:
• dispatcher.log:
• nnictl stdout and stderr:

How to reproduce it?:

[flyingmrwang]

the same for speedup pruned ghostnet, and the num channels are 16 and 1

[J-shang]

Hi @jxncyym @flyingmrwang ,
If possible, could you provide the model and pruning code for our reproduction?

[jxncyym]

@J-shang the latest code refer to :https://pan.baidu.com/s/19xoJOC9yaD3E6-RkZkOf9Q，g4og

[J-shang]

This dependency set leads to error, but from the model structure, it seems that they should not be in the same group.

{'spp.scale2.3': 128, 'spp.scale4.3': 128, 'spp.scale0.2': 128, 'layer5.0.conv3': 512, 'spp.process1.2': 128, 'layer5.0.downsample.0': 512, 'spp.process2.2': 128, 'spp.scale1.3': 128, 'spp.scale3.3': 128, 'spp.process3.2': 128}

line 174 in ddrnet_23_slim.py
out = self.compression(torch.cat(x_list, 1)) + self.shortcut(x)

Seems the dependency is related to this line.

@zheng-ningxin please help to look into this issue.

[jxncyym]

Is there any progress @zheng-ningxin @J-shang

[zheng-ningxin]

The size operation on the graph misleads the model topology analysis.

[jxncyym]

then how to solve this problem？@zheng-ningxin

[zheng-ningxin]

I'll submit a PR to fix this ASAP. I'll let you know when this pr merged. Thanks!

[vinhtq115]

I'll submit a PR to fix this ASAP. I'll let you know when this pr merged. Thanks!

Have you been able to fix this? I have the same issue with YOLOv5-Face yolov5n-0.5 model.

[HappyPeanuts]

I have meet the same problem:'assert len(set(num_channels_list)) == 1' when i try to prune my model DLASeg, i want to know how to resove it~ thanks a lot

[HappyPeanuts]

I'll submit a PR to fix this ASAP. I'll let you know when this pr merged. Thanks!

Hi, have you resolved this problem? I have meet the same error and I guess it was caused by the use of DCNv2. I want to now how to resolve this problem. Thanks ~

[zheng-ningxin]

@HappyPeanuts Sincerely sorry for the late PR, already working on this pr and need some time. Thanks~

[HappyPeanuts]

thank you for your reply，I will try other methods，if convenient ，please notice me when you resolved this problem  发自我的iPhone

…

------------------ Original ------------------ From: Ningxin Zheng ***@***.***> Date: Mon,Feb 28,2022 4:49 PM To: microsoft/nni ***@***.***> Cc: 杨志伟 ***@***.***>, Mention ***@***.***> Subject: Re: [microsoft/nni] assert len(set(num_channels_list)) == 1 (#4160) @HappyPeanuts Sincerely sorry for the late PR, already working on this pr and need some time. Thanks~ — Reply to this email directly, view it on GitHub, or unsubscribe. Triage notifications on the go with GitHub Mobile for iOS or Android. You are receiving this because you were mentioned.Message ID: ***@***.***>

[HK017]

@zheng-ningxin,hello, What does this code assert len(set(num_channels_list)) == 1 mean for me to examine my own model structure? thanks

[dianxin556]

I also meet the same problem , so how to solve it?

[dreamlychina]

I also meet the same problem , so how to solve it?

[J-shang]

hello @dianxin556 @dreamlychina , could you show your model for us to reproduce your issue? This problem is highly related to the model structure.