python tfjs.converter.load_keras_model fails if Variable names already used

To get help from the community, check out our Google group.

TensorFlow.js version

tensorflowjs version 0.4.0 keras version 2.1.4

Browser version

None ( this is a python issue)

Describe the problem or feature request

If you create a model, save it to it to tfjs.json format and try to load it within the same python session, it will fail because of a weight renaming issue.

Code to reproduce the bug / link to feature request

import os
import tensorflowjs as tfjs
import keras
OUTPUT_DIR = "/tmp/tfjs-output-dir"

dense_a = keras.layers.Dense(2, input_dim = 3)
model = keras.models.Sequential([dense_a])
tfjs.converters.save_keras_model(model os.path.join(OUTPUT_DIR, "seq_a_tfjs"))
model_prime = tfjs.converters.load_keras_model(os.path.join(
    OUTPUT_DIR, "seq_a_tfjs", "model.json"))

outputs

Using TensorFlow backend.
Traceback (most recent call last):
  File "tiny_nested_model_a_min.py", line 26, in <module>
    model_prime = tfjs.converters.load_keras_model(os.path.join(OUTPUT_DIR, "seq_a_tfjs", "model.json"))
  File "/usr/local/google/home/bileschi/env/a/local/lib/python2.7/site-packages/tensorflowjs/converters/keras_tfjs_loader.py", line 208, in load_keras_model
    use_unique_name_scope=use_unique_name_scope)
  File "/usr/local/google/home/bileschi/env/a/local/lib/python2.7/site-packages/tensorflowjs/converters/keras_tfjs_loader.py", line 74, in _deserialize_keras_model
    weights_list.append(weights_dict[weight_name])
KeyError: u'dense_1_1/kernel'

[bileschi]

If you load the model three times you get three versions o the same model, no weight sharing. It modifies the eight names with the number plus underscore method.

On Sat, Jun 2, 2018 at 1:49 PM Daniel Smilkov [e-mail] wrote:

What does Keras do if you save and load the model in the same session. That is:

model.save(‘my_model.h5’) # creates a HDF5 file ‘my_model.h5’# Will the next line error? model = load_model(‘my_model.h5’)

I’m assuming it will error as well since in their example snippet https://keras.io/getting-started/faq/#how-can-i-save-a-keras-model they do call del model in between.

So it might actually be better to explicitly error than to silently create a whole new graph and double the memory usage.

— You are receiving this because you authored the thread.

Reply to this email directly, view it on GitHub https://github.com/tensorflow/tfjs/issues/373#issuecomment-394104744, or mute the thread https://github.com/notifications/unsubscribe-auth/AAhZTtAhuKlaLgCttzt8v5rHp7eVHEb5ks5t4tAagaJpZM4UXM3l .

– Stan Bileschi Ph.D. | SWE | [e-mail] | 617-230-8081

[caisq]

Reassigning to @davidsoergel for follow-up investigation.