Add NPU backend support for val and inference

[MengqingCao]

I am a user of NPU. When I used TIMM recently, I found that it does not support NPU natively. It's pleasure to see that someone has made some contributions on leveraging NPU to TIMM #2102. But it currently only offers the feature of using NPU during training. This PR extends NPU support to the validate and inference entries, thus addressing this limitation.

Specify the device as "npu", then you can use NPU as accelerator during inferencing and validating.

It is tested on:

• model: tiny_vit_21m_512
• dataset: the val subset of ImageNet-1K

Validate Scripts

python validate.py ../open_clip/data/ImageNet-1000/val/ --device npu --model ./model_ckpts/tiny_vit_21m_512 --batch-size 64 --pretrained

ScreenShot

It shows the validation results on val subset of ImageNet-1K are as following:

top-1 acc	top-5 acc
86.040%	97.750%

Inference Scripts

python inference.py ./data/ --device npu --batch-size 64 --model ./model_ckpts/tiny_vit_21m_512 --label-type detail --topk 5

ScreenShot

results

Here offers some results of predicting the top-5 classification results by inferencing on tiny_vit_21m_512. Everything goes well on npu.

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[MengqingCao]

cc @rwightman

[rwightman]

@MengqingCao see #2138 ... I need a better design to centralize device specific acccelerator module loading, etc instead of spreading it out across many files, it's not a sustainable approach.

Also, another challenge here is I don't have easy access to many potential accelerators so definitely need help testing as I can't realistically run my normal CI or tests with them as I do across my github and local CI right now...

[MengqingCao]

@MengqingCao see #2138 ... I need a better design to centralize device specific acccelerator module loading, etc instead of spreading it out across many files, it's not a sustainable approach.

Also, another challenge here is I don't have easy access to many potential accelerators so definitely need help testing as I can't realistically run my normal CI or tests with them as I do across my github and local CI right now...

Good day! @rwightman, thanks for your reply.

for your first concern, I agree that importing the device specific modules in many files is not a smart way to enable the devices. I was inspired by the way of centralizing device related modules loading in train.py#L415 that we could do a autoloading when the whole lib is initing. Because the way in train.py#L415 also needs to do redundant processing in many files.

My initial idea was to load the device accelerator module via a specific environment variable (e.g. TIMM_DEVICE_EXT). This variable is set in timm/init.py by reading the configuration infos in a specific file (e.g., a json file), and then the module is preloaded according to this variable, so that device-related modules import can be activated within the entire TIMM library, instead of having to import them separately everywhere. But I think the device-specific hardcoding has to be modified.

For your second concern, making a mechine with Ascend NPU available to community is on my to-do list, so that we could ensure that the correctness of the code could be verified and maintained.

Let me know if you have any ideas or confusion!

[MengqingCao]

Hi, @rwightman. I have just committed the code implementation of the above solution, please review it, thx!