Tensorflowjs conversion error from tflite-model-maker generated SAVED_model

General Discussion
, ,
#1

Hi @khanhlvg

Thanks for your tutorials and colab notebook.
Since I was experimenting with using your android_figurine tflite model in a browser, I extracted it as SAVED_MODEL as shown below.

Because ExportFormat.TFJS is not allowed for object detection model yet…
I’ve tried to use this flow: Model maker → SAVED_MODEL → TFJS

Step1: Model maker → SAVED_MODEL (Success)

// model.export(export_dir='.', tflite_filename='android.tflite') // original code

model.export(export_dir="/js_export/", export_format=[ExportFormat.SAVED_MODEL])
!zip -r /js_export/ModelFiles.zip /js_export/

Step 2 SAVED_MODEL → TFJS (Failed)

스크린샷 2022-01-03 오후 5.33.03
스크린샷 2022-01-03 오후 5.33.031920×1093 212 KB

Error log

Traceback (most recent call last):
  File "/Users/judepark/Sites/TensorFlow/.direnv/python-3.6.8/lib/python3.6/site-packages/tensorflowjs/converters/wizard.py", line 619, in run
    converter.convert(arguments)
  File "/Users/judepark/Sites/TensorFlow/.direnv/python-3.6.8/lib/python3.6/site-packages/tensorflowjs/converters/converter.py", line 804, in convert
    weight_shard_size_bytes, metadata_map)
  File "/Users/judepark/Sites/TensorFlow/.direnv/python-3.6.8/lib/python3.6/site-packages/tensorflowjs/converters/converter.py", line 533, in _dispatch_converter
    metadata=metadata_map)
  File "/Users/judepark/Sites/TensorFlow/.direnv/python-3.6.8/lib/python3.6/site-packages/tensorflowjs/converters/tf_saved_model_conversion_v2.py", line 771, in convert_tf_saved_model
    metadata=metadata)
  File "/Users/judepark/Sites/TensorFlow/.direnv/python-3.6.8/lib/python3.6/site-packages/tensorflowjs/converters/tf_saved_model_conversion_v2.py", line 724, in _convert_tf_saved_model
    metadata=metadata)
  File "/Users/judepark/Sites/TensorFlow/.direnv/python-3.6.8/lib/python3.6/site-packages/tensorflowjs/converters/tf_saved_model_conversion_v2.py", line 157, in optimize_graph
    ', '.join(unsupported))
ValueError: Unsupported Ops in the model before optimization
TensorListConcatV2
#3

It seems some of the ops your TFLite model is using are not supported in TensorFlow.js. The only way to get around this error is to change the original model not to use those ops or to contribute the missing ops to the TensorFlow.js project.

1 Like
#5

This is my colab test notebook that uses @khanhlvg 's dataset and training jobs.
The only difference is that I extracted in SAVED_MODEL format, not TFLITE. And unfortunately even this saved model doesn’t work. Where exactly did I go wrong? please let me know.

#6

Can you try using TFLite Task Library for TF.js to run the TFLite model, instead of trying to convert the SavedModel to TF.js. There are ops used in the model that aren’t supported natively in TF.js.

Please follow the example below.

#8

@khanhlvg Thanks for your tip.

BTW, could you please let me know how i can load my custom model via TFLite pkg?

I trained my model and exported rico.tflite.
I’ve tried to load it on javascript but it didn’t work so far.