[Error] Can not convert SaveModel to TFLite

[LaiThePhong]

Hello,
I copied all code from basic_training_notebook.ipynb and run on my collab 👈 I already shared it public, you can check the output in the last cells.

The steps like Generated positive data (the wakeword), Preparing data, Augmenting data, Training model, and Evaluation model are processed successfully. The trained model was saved as .h5 file as name: "last_weights.h5".
Then, to create TFlite file, I changed the flags: --train 0, --test_tf_nonstreaming 1 , the code executed 2 steps status below:

• Step 1: convert .h5 model to SaveModel: ✅ Success. (This time, I can see the .pb model in pathtrained_models/wakeword/non_stream/fingerprint.pb)
• Step 2: convert SaveModel to tflite: ❌ Error with display message: RuntimeError: failed to update filter tensor FULLY_CONNECTED node #21Node number 28 (TfLiteXNNPackDelegate) failed to prepare.

Same issue when I changed --train 0, --test_tflite_streaming 1: RuntimeError: failed to update filter tensor FULLY_CONNECTED node #48Node number 58 (TfLiteXNNPackDelegate) failed to prepare.

I am searching some solutions on the internet, but if you guys are familiar with this issue, please let me know to try it.
Thank you, so much.

[kahrendt]

Sorry, I don't think this one was fixed with #48, sorry for referencing this!

It doesn't look like you let the model train fully. It seems like it went through one training batch, but it didn't finish testing on the validation set. Its possible the weights weren't properly saved in that situation. Try letting the model train for more mini-batches.