Convert output of retrain.py to tensorflow.js

Question

The script retrain.py described in How to Retrain an Image Classifier for New Categories was run as

python retrain.py --tfhub_module https://tfhub.dev/google/imagenet/mobilenet_v2_100_224/feature_vector/2 --image_dir /tmp/test

and produced the output file /tmp/output_graph.pb. Converting this with

tensorflowjs_converter --input_format=tf_saved_model --output_format=tfjs_graph_model /tmp/output_graph.pb /tmp/model

failed with

IOError: SavedModel file does not exist at: /tmp/output_graph.pb/{saved_model.pbtxt|saved_model.pb}

If the file output_graph.pb is renamed to saved_model.pb (by @edkeveked), the error changes to

RuntimeError: MetaGraphDef associated with tags 'serve' could not be found in SavedModel. To inspect available tag-sets in the SavedModel, please use the SavedModel CLI: saved_model_cli

saved_model_cli show --dir . reports an empty tag set.

How can this be fixed?

Croix

The error says it all. The file is not found – Cornelius

@edkeveked: the file /tmp/output_graph.pb exists along with output_labels.txt. Is this a different format? – Croix

Have you solved it? I ran into the same problem using the tfjs sample mobilenet V2, – Spandau

@季谢尔: No. The alternative failed as well. Even a bounty on #55849809 yielded no result. See the issue at github.com/tensorflow/tensorflow/issues/28365 – Croix

@季谢尔: see the answer – Croix

Answer 1

As hinted by @Ping Yu in Retrain image detection with MobileNet, you can use

python retrain.py --tfhub_module https://tfhub.dev/google/imagenet/mobilenet_v2_100_224/feature_vector/2 \
    --image_dir /tmp/flower_photos --saved_model_dir /tmp/saved_retrained_model
tensorflowjs_converter --input_format=tf_saved_model \
    --output_format=tfjs_graph_model \
    --saved_model_tags=serve \
    /tmp/saved_retrained_model/ /tmp/converted_model/

This saves the model using the saved model format.

Croix

Answer 2

The input path is the path of the folder and not of the file. Consider the following:

tensorflowjs_converter --input_format=tf_saved_model --output_format=tfjs_graph_model /tmp /tmp/model

Cornelius

After changing to the directory and renaming the .pb file, a different error occurs. See the edit. Still a step in the right direction. Thanks – Croix

Though I can't tell exactly why you're having the current error, I could suggest to add other flags --output_node_names="final_result", --saved_model_tags=serve to tfjs-converter command. – Cornelius

both together yielded "TensorFlow.js model converters.: error: unrecognized arguments: --output_node_names=final_result", skipping that resulted in the same "RuntimeError: MetaGraphDef associated with tags 'serve' could not be found in SavedModel. To inspect available tag-sets in the SavedModel, please use the SavedModel CLI: saved_model_cli" – Croix