Problems in run train.py

[JianchengZ]

Hello, I just got a problem.
When I run: [torchrun --nproc_per_node 4 train.py --scale small --data_dir ./Data --output_dir ./Results/ --exp_name clip_score_train_results],
I was told that: [from training.distributed import world_info_from_env
ModuleNotFoundError: No module named 'training'],
But I use pip or conda, I still can not have the module.

[JianchengZ]

Sorry, I have solved that problem above. But another problem still exits.
The problem is :
[W socket.cpp:464] [c10d] The server socket cannot be initialized on [::]:29500 (errno: 97 - Address family not supported by protocol).
[W socket.cpp:697] [c10d] The client socket cannot be initialized to connect to [localhost]:29500
Could you please help to answer it ?

[HariSeldon11988]

@JianchengZ

I have the same problem as you (from training.distributed ..... No module named 'training').
Can you tell me how you solved the problem? Would help me a lot :)

[JianchengZ]

@HariSeldon11988

Yes, here is the answer(commented by others):
The training module comes from open_clip, and you can find the module in the open_clip repository if you are interested in looking at the source code: https://github.com/mlfoundations/open_clip/tree/v2.16.1/src/training.