MobileNetEdgeTPU-V2 final model to be used

[mohitmundhragithub]

The links that were shared earlier has different files:

Link1: https://tfhub.dev/google/edgetpu/vision/mobilenet-edgetpu-v2/l/1
Link2: https://drive.google.com/file/d/1--A9EI8Ntr_Q5CH0WblFvUctVI_6VkcR/view?usp=sharing

The TF model shared in Link2 has the dimensions fixed to 1x224x224x3.
The TF model shared in Link1 has the dimensions fixed to ?x224x224x3.

As per the description in the slide Link 2 must have been extracted from Link1. But need confirmation on the final model. I think the one with dynamic batch would be be better.

[freedomtan]

Agree that we should support dynamic batch size for offline.

Actually, I think the official one is from https://github.com/tensorflow/models/tree/master/official/projects/edgetpu/vision#edgetpu-optimized-vision-models

[mohitmundhragithub]

Agree that we should support dynamic batch size for offline.

Actually, I think the official one is from https://github.com/tensorflow/models/tree/master/official/projects/edgetpu/vision#edgetpu-optimized-vision-models

Hi,

I checked this link. The TF version of the MobileNetEdgeTPUv2-L shared here is the checkpoint. And the tflite version shared in link is a fixed 1 batch model only.

I think the tfhub model shared in the slide (link 1 in my comment in this issue) should be good to use.

[freedomtan]

Agree that we should support dynamic batch size for offline.
Actually, I think the official one is from https://github.com/tensorflow/models/tree/master/official/projects/edgetpu/vision#edgetpu-optimized-vision-models

Hi,

I checked this link. The TF version of the MobileNetEdgeTPUv2-L shared here is the checkpoint. And the tflite version shared in link is a fixed 1 batch model only.

I think the tfhub model shared in the slide (link 1 in my comment in this issue) should be good to use.

Yes. the TF version is the checkpoint. But I thinks it's the original source. To get frozen pb or saved_model, you have to clone the repo and export the model to use the checkpoints / weights.

Nope, the TFLite model does support dynamic batch size. When I run it with benchmark_model --graph=... --use_nnapi=1 --input_layer=input --input_layer_shape=n,224,224,3, where n > 1, I can get expected results on Pixel and MTK devices.

[mohitmundhragithub]

Agree that we should support dynamic batch size for offline.
Actually, I think the official one is from https://github.com/tensorflow/models/tree/master/official/projects/edgetpu/vision#edgetpu-optimized-vision-models

Hi,
I checked this link. The TF version of the MobileNetEdgeTPUv2-L shared here is the checkpoint. And the tflite version shared in link is a fixed 1 batch model only.
I think the tfhub model shared in the slide (link 1 in my comment in this issue) should be good to use.

Yes. the TF version is the checkpoint. But I thinks it's the original source. To get frozen pb or saved_model, you have to clone the repo and export the model to use the checkpoints / weights.

Nope, the TFLite model does support dynamic batch size. When I run it with benchmark_model --graph=... --use_nnapi=1 --input_layer=input --input_layer_shape=n,224,224,3, where n > 1, I can get expected results on Pixel and MTK devices.

when i use this tflite model, somehow it doesn't work for us.

Is it ok to use the saved model from link 1? The accuracy reported by Parham is also based on this model only.

[mohitmundhragithub]

It was discussed in the meeting today that it's ok to use Link 1 as well.
We may need to save a copy of this model in github.com/mlcommons/mobile_open as well.

[freedomtan]

It was discussed in the meeting today that it's ok to use Link 1 as well. We may need to save a copy of this model in github.com/mlcommons/mobile_open as well.

Agree. Let's send PR to mobile_open or mobile_models

[mohitmundhragithub]

The tfhub link (link 1) now redirects to kaggle... but the downloadable model is the same.

[mohitmundhragithub]

I can't create a PR for the model.
@anhappdev can you please help for it?

[anhappdev]

I can't create a PR for the model. @anhappdev can you please help for it?

Can you show a screenshot of the error?

[mohitmundhragithub]

No technical issues as such... Its more of a licensing issue.
For me to download the model and then upload to mlcommons repo will require more approvals.

[freedomtan]

We chose another model. Close this for now.