search

tensorflow / tfjs

A WebGL accelerated JavaScript library for training and deploying ML models.
https://js.tensorflow.org
Apache License 2.0
18.3k stars 1.91k forks source link

Error in loading a Keras model in TensorFlow.js #3458

Closed Gheith-Abandah closed 4 years ago

Gheith-Abandah commented 4 years ago

TensorFlow.js version

2.0.0

Browser version

Chrome Version 83.0.4103.97 (Official Build) (64-bit)

Summary

I have a problem in loading a trained model that was developed using Python and TensorFlow Keras into TensorFlow.js.

Procedure

The model was developed using Python, TensorFlow, and Keras using: model.fit(...) model.save('./model_python_tf_keras.h5) This model is attached.

I then converted this Keras model into a TensorFlow.js model using: $ tensorflowjs_converter --input_format=keras ./model_python_tf_keras.h5 ./model_js The resulting folder is attached.

Problem

I get a problem when I try to load the new model in an HTML document (index_min.html is attached) using the following Javascript code: model = await tf.loadLayersModel('http://localhost:63342/powem_meters/model_js/model.json'); The error message is: Uncaught (in promise) Error: Provided weight data has no target variable: bidirectional/forward_lstm/lstm_cell_1/kernel

As a result, I am not able to use my model in a Javascript program.

Notes

I have tried to save the model directly from Python using: tfjs.converters.save_keras_model(model, "./model_js") And I tried running the tensorflowjs_converted using tf version 1.4.0 and tfjs version 1.2.6. But I keep getting the same error.

I have also noticed that the above procedure works for a simple fully-connected model. However, my model is not a custom one. The model definition code is: model = Sequential() model.add(Embedding(num_tokens+1, 32, input_length=max_seq_length)) model.add(Bidirectional(LSTM(latent_dim, input_shape=(None,num_tokens), return_sequences=True, dropout=0.1, recurrent_dropout=0.3), merge_mode='concat')) model.add(Bidirectional(LSTM(latent_dim, dropout=0.1, recurrent_dropout=0.3), merge_mode='concat')) model.add(Dense(output_data.shape[1], activation='softmax')) model.compile(optimizer='adam', loss='categorical_crossentropy', metrics=['accuracy'])

Details

My environment is as follows: OS: Ubuntu 20.04 Python version 3.8.2 tensorflow 2.2.0rc4
tensorflowjs 2.0.0
files.zip

rthadur commented 4 years ago

@Gheith-Abandah thanks for reporting , can you please check here for similar issue.

Gheith-Abandah commented 4 years ago

Thank you for this hint. I have checked Issue #755 before. I am now checking it again and I will come back to you when I reach a conclusion.

rthadur commented 4 years ago

@Gheith-Abandah thank you , please check this comment as well.

Gheith-Abandah commented 4 years ago

Using tfjs version 2.0.1 solves this problem thank you.