Availability of the pretrained AutosegEdgeTPU-XS backbone

[J-Bracke]

Hello,

first thanks for providing this edgetpu optimized models for image segmentation, it helped me a lot!
I'm trying to transfer learning a model with the AutosegEdgeTPU-XS architecture at the moment, where a checkpoint of the pretrained AutosegEdgeTPU-XS backbone would be really helpful.
Could you please provide the missing checkpoints for the two AutoSegEdgeTPU backbones mentioned in your repo:
https://github.com/tensorflow/models/tree/v2.13.0/official/projects/edgetpu/vision

[laxmareddyp]

Hi @J-Bracke ,

We're glad you're using the official Model Garden for your research! We noticed you were looking for checkpoints for the two AutoSegEdgeTPU backbones,They're already available under the same path you mentioned above.

For confirmation, you can refer to the image attached to this message (assuming you're referring to the same image). Simply click on the model name and the checkpoint will automatically download to your local machine, ready for use.

We hope this helps! Let us know if you have any further questions.

[github-actions]

This issue has been marked stale because it has no recent activity since 7 days. It will be closed if no further activity occurs. Thank you.

[J-Bracke]

Thanks @laxmareddyp for the hint, although I had noticed the link-option by klicking on the autoseg_edgetpu_xs model name before already.
However I'm not sure, whether that provided checkpoint for the autoseg_edgetpu_xs backbone is the correct one.
If I run your default experiment "semantic_segmentation_searched_config.autoseg_edgetpu_xs()" with the downloaded checkpoint I get the error message:

AssertionError: Found 240 Python objects that were not bound to checkpointed values, likely due to changes in the Python program. Showing 10 of 240 unmatched objects: [SyncOnReadVariable:{ 0: <tf.Variable 'stack_16/block_0/fused_bn/moving_mean:0' shape=(576,) dtype=float32, numpy= ...
(shortened error message)

If I do the same with the downloaded mobilenet_edgetpuv2_xs backbone checkpoint and the experiment "semantic_segmentation_config.deeplabv3plus_mobilenet_edgetpuv2_xs_ade20k" it works.
What is different with the autoseg-models then? Could you help?