[BUG] module 'timm' has no attribute 'create_conv2d'

[sefibk]

Describe the bug Following the exact instructions to train EfficientDet I get the error: AttributeError: module 'timm' has no attribute 'create_conv2d'

To Reproduce Steps to reproduce the behavior:

1. Cloned efficientdet repo
2. Installed Timm package (tried both cloning repo and running python setup.py install and pip install timm)
3. Run ```

   !/bin/bash

   NUM_PROC=$1 shift python -m torch.distributed.launch --nproc_per_node=$NUM_PROC train.py "$@"

4. Process killed with error: AttributeError: module 'timm' has no attribute 'create_conv2d' comment: After installing Timm and starting python import timm succeeds including running timm.create_model() however timm.create_conv2d and other functions fails with the above message.

Expected behavior Expected training of EfficientDet to start

Desktop (please complete the following information):

• OS: Ubuntu 18.04
• This repository version [cloned Master no Feb 2nd - not sure how to check the version]
• Python 3.7.0
• pytorch 1.7.1 py3.7_cuda11.0.221_cudnn8.0.5_0 pytorch
• timm 0.3.4 pypi_0 pypi

Additional context I assume it is a problem with the Timm installation but I am not sure and do not know how to check this. Would appreciate any assistance

[rwightman]

@sefibk are you sure you don't have a different/old version of timm in your env (ie two timm) or a folder named timm in your working path?

[sefibk]

Thank you for replying however issue is still unsolved.

I tried removing the folder timm from the device => still same error Also tried removing the folder, uninstalling and reinstalling timm => still same error

BTW, when running pip install timm I get version 0.3.4 but when installing locally from the clone repo I get 0.4.1. Could this indicate something? Do you have any other ideas?

[sefibk]

I was able to solve those issues by editing __init__.py in timm but now I get the error:

Traceback (most recent call last):
  File "train_effdet.py", line 699, in <module>
    main()
  File "train_effdet.py", line 353, in main
    optimizer = timm.create_optimizer(args, model)
  File "/root/miniconda/lib/python3.7/site-packages/timm-0.4.1-py3.7.egg/timm/optim/optim_factory.py", line 53, in create_optimizer
    assert has_apex and torch.cuda.is_available(), 'APEX and CUDA required for fused optimizers'
AssertionError: APEX and CUDA required for fused optimizers

Do you use native APEX? must I install it? (I am running with torch==1.7.1)

[rwightman]

It's exactly as that assert says, you need it for the fused versions of optimizers. Remove the 'fused' if you don't have or don't want to install native APEX

[sefibk]

Thank you very