Improve the AI-Polygon experience

[TDA-2030]

labelme with SAM integration is great, but my experience with it right now is a little annoying. Because I have to wait several seconds to compute image embedding for each image when I change to the next/prev image. Despite the caching mechanism, I still have to wait for calculations from time to time, which makes my computer very slow overall. I tried to change the provider of onnxruntime to GPU successfully, but it shows the following warning and doesn't seem to have much of a speedup.

[INFO   ] __init__:get_config:70 - Loading config file from: C:\Users\zhouli\.labelmerc
[DEBUG  ] canvas:initializeAiModel:140 - Initializing AI model: 'Segment-Anything (balanced)'
Computing MD5: C:\Users\zhouli\.cache/gdown\https-COLON--SLASH--SLASH-github.com-SLASH-wkentaro-SLASH-labelme-SLASH-releases-SLASH-download-SLASH-sam-20230416-SLASH-sam_vit_l_0b3195.quantized.encoder.onnx
MD5 matches: C:\Users\zhouli\.cache/gdown\https-COLON--SLASH--SLASH-github.com-SLASH-wkentaro-SLASH-labelme-SLASH-releases-SLASH-download-SLASH-sam-20230416-SLASH-sam_vit_l_0b3195.quantized.encoder.onnx
Computing MD5: C:\Users\zhouli\.cache/gdown\https-COLON--SLASH--SLASH-github.com-SLASH-wkentaro-SLASH-labelme-SLASH-releases-SLASH-download-SLASH-sam-20230416-SLASH-sam_vit_l_0b3195.quantized.decoder.onnx
MD5 matches: C:\Users\zhouli\.cache/gdown\https-COLON--SLASH--SLASH-github.com-SLASH-wkentaro-SLASH-labelme-SLASH-releases-SLASH-download-SLASH-sam-20230416-SLASH-sam_vit_l_0b3195.quantized.decoder.onnx
2023-05-17 15:03:20.0692669 [W:onnxruntime:, session_state.cc:1136 onnxruntime::VerifyEachNodeIsAssignedToAnEp] Some nodes were not assigned to the preferred execution providers which may or may not have an negative impact on performance. e.g. ORT explicitly assigns shape related ops to CPU to improve perf.
2023-05-17 15:03:20.0767142 [W:onnxruntime:, session_state.cc:1138 onnxruntime::VerifyEachNodeIsAssignedToAnEp] Rerunning with verbose output on a non-minimal build will show node assignments.
2023-05-17 15:03:21.4425535 [W:onnxruntime:, session_state.cc:1136 onnxruntime::VerifyEachNodeIsAssignedToAnEp] Some nodes were not assigned to the preferred execution providers which may or may not have an negative impact on performance. e.g. ORT explicitly assigns shape related ops to CPU to improve perf.
2023-05-17 15:03:21.4493971 [W:onnxruntime:, session_state.cc:1138 onnxruntime::VerifyEachNodeIsAssignedToAnEp] Rerunning with verbose output on a non-minimal build will show node assignments.
[DEBUG  ] segment_anything:_compute_and_cache_image_embedding:52 - Computing image embedding...
[DEBUG  ] segment_anything:_compute_and_cache_image_embedding:63 - Done computing image embedding.

As with the bad experience mentioned above, I think it would be a smoother experience if the user could import all the image embeddings computed on a powerful GPU and then do the AI labeling. The project https://github.com/anuragxel/salt did just that.

[wkentaro]

Despite the caching mechanism, I still have to wait for calculations from time to time

If this happens with the already computed images, we should fix that first.

I tried to change the provider of onnxruntime to GPU successfully, but it shows the following warning and doesn't seem to have much of a speedup.

This is also weird and should give better experience if we use GPU successfully.

[mohamedhasan-nazarbasha]

I think still there is an issue. Takes a lot of time to compute for new & computed images no matter what. Is there any possible way.

[left0ver]

metoo，I want to know if it can compute use gpu？