Can not convert model lm_model4_opt.onnx to tflite

[tanphan07]

Thank you for your awesome project.
I have a problem when converting model lm_model4_opt.onnx to tflite.
I use this repo https://github.com/onnx/onnx-tensorflow to convert onnx to tflite and get this error BackendIsNotSupposedToImplementIt: FusedConv is not implemented. Can you support me to handle this case.
Thank you very much

[emilianavt]

You can try the unoptimized models found in #48. If those do not work for you, you may need to convert the original pytorch models in #1.

[tanphan07]

Thank for your reply, but i can not find the lm_model4 in the #48. Moreover, can you give me an advice training a model that has a good performance when eyes closed

[emilianavt]

Thank for your reply, but i can not find the lm_model4 in the #48.

In that case, you may have to reconvert it from pytorch weights, which can maybe, hopefully be found in #1. You can also try using model 3. It should have roughly the same tracking quality. 4 just has better support for winks (a single closed eye).

Moreover, can you give me an advice training a model that has a good performance when eyes closed

Make sure that you have a lot of correctly annotated images in various head poses with the eyes closed.

[tanphan07]

Thank for your reply, but i can not find the lm_model4 in the #48.

In that case, you may have to reconvert it from pytorch weights, which can maybe, hopefully be found in #1. You can also try using model 3. It should have roughly the same tracking quality. 4 just has better support for winks (a single closed eye).

Moreover, can you give me an advice training a model that has a good performance when eyes closed

Make sure that you have a lot of correctly annotated images in various head poses with the eyes closed.

In my use case, i want a model that can make high accuracy when the eyes closed, so i think only model 4 that can help me. Can you provide checkpoint of model 4. Thank you very much.

[emilianavt]

The accuracy of model 4 shouldn't be higher when both eyes are closed. It is only better at decoupling the closedness state of the eyes.

It may take some time before I can find the original checkpoint.

[tanphan07]

The accuracy of model 4 shouldn't be higher when both eyes are closed. It is only better at decoupling the closedness state of the eyes.

It may take some time before I can find the original checkpoint.

could you provide the dataset that has a lot of correctly annotated images in various head poses with the eyes closed? Thank you in advance

[emilianavt]

No, I cannot.