train my own model

[wmkai]

I want to train my own model from scratch, but the log shows that

Training with a single process on 1 GPUs.
Traceback (most recent call last):
  File "/home/tiger/code/pytorch-image-models/train.py", line 866, in <module>
    main()
  File "/home/tiger/code/pytorch-image-models/train.py", line 397, in main
    model = create_model(
  File "/home/tiger/code/pytorch-image-models/timm/models/factory.py", line 71, in create_model
    model = create_fn(pretrained=pretrained, pretrained_cfg=pretrained_cfg, **kwargs)
  File "/home/tiger/code/pytorch-image-models/timm/models/topformer.py", line 662, in topformer_base
    return _create_topformer('topformer_base', pretrained, **model_args)
  File "/home/tiger/code/pytorch-image-models/timm/models/topformer.py", line 575, in _create_topformer
    return build_model_with_cfg(Topformer, variant, pretrained, **kwargs)
  File "/home/tiger/code/pytorch-image-models/timm/models/helpers.py", line 511, in build_model_with_cfg
    pretrained_cfg = resolve_pretrained_cfg(variant, pretrained_cfg=pretrained_cfg)
  File "/home/tiger/code/pytorch-image-models/timm/models/helpers.py", line 469, in resolve_pretrained_cfg
    assert pretrained_cfg
AssertionError
(torch-env) 

[rwightman]

@wmkai questions (ie non-issues, non-features) are for discussions... moving

[rwightman]

@wmkai you need to follow template of other models if you want to use the factory, so follow other models and make sure there is a default_cfgs {} dict in the model module w/ an entry matching each of your model name fns, OR the simplest approach if you just want to get something training but use the build_model.. fn, just pass a dict for pretrained_cfg arg that has defaults for your model (the resolution, interpolation, crop size, etc are used to helpset the preprocessing pipeline in train script)

[wmkai]

@wmkai you need to follow template of other models if you want to use the factory, so follow other models and make sure there is a default_cfgs {} dict in the model module w/ an entry matching each of your model name fns, OR the simplest approach if you just want to get something training but use the build_model.. fn, just pass a dict for pretrained_cfg arg that has defaults for your model (the resolution, interpolation, crop size, etc are used to helpset the preprocessing pipeline in train script)

Thanks for the reply! I did try to follow template of other models but I saw there are url in the default_cfgs of other models. I guess it's a url of a pretrained model. However, I want to train my own models from scratch, so I did not use the url in the default_cfgs. I don't know whether or not it is wrong. And I tried to look for document for training my own model but didn't find it. If there are some relevant documents, can you send it to me?

[rwightman]

@wmkai I've updated the function that checks pretrained cfg to be more tolerant of one not being passed in (warn instead of assert), it will use some defaults if no config exists, but should allow one to start training without needing to register default cfgs, etc...