search

google-coral / tutorials

Colab/Jupyter tutorials about training TensorFlow models for Edge TPU, and other tutorials
Apache License 2.0
177 stars 39 forks source link

TF2 retrain_classification_ptq_tf2 notebook always gives 0 accuracy #3

Closed kpham123 closed 2 years ago

kpham123 commented 4 years ago

Run all and running inference for the output of the TF2 model, on a Coral TPU board always gives 0.00 accuracy, even for images from the training dataset.

Using the same input images on the TF1 compiled model works fine.

kpham123 commented 4 years ago

image

Namburger commented 4 years ago

@kpham123 what is your model's IO tensor type?

kpham123 commented 4 years ago

This is actually the model in google's tutorials here: https://github.com/google-coral/tutorials/blob/master/retrain_classification_ptq_tf2.ipynb

Seems like it was recently updated with a new commit. I should be able to load and check this by the end of week

Namburger commented 4 years ago

@kpham123 hey, yeah, the reason why I asked is because the model produced on an older version of that will have float IO tensors. Which means that you'll have to process the outputs differently than normal.

The newest revision (2 days ago) brought us back to uint8 IO tensors so you can use our normal demo on this model.

hjonnala commented 2 years ago

Closing this as this issue no longer exists.