Btukfyl

I saved my model automatically by callbacks.ModelCheckpoint() with a HDF5 file.

# Checkpoint In the /output folder

filepath = "./model/mnist-cnn-best.hd5"



# Keep only a single checkpoint, the best over test accuracy.

checkpoint = keras.callbacks.ModelCheckpoint(filepath, monitor='val_acc', 

                                             verbose=1, save_best_only=True,

                                             mode='max')



# Train

model.fit(x_train, y_train,

          batch_size=batch_size,

          epochs=epochs,

          verbose=1,

          validation_data=(x_test, y_test),

          callbacks=[checkpoint])

When I load a model, an error occured.

  model = keras.models.load_model("./mnist-cnn-best.hd5")



  File "D:Program FilesAnaconda3libsite-packagestensorflowpythonkerasenginesaving.py", line 251, in load_model

    training_config['weighted_metrics'])

KeyError: 'weighted_metrics'

If I load model with param 'compile=False', it works correctly.

I know the normal way to save model in keras is:

from keras.models import load_model



model.save('my_model.h5')  # creates a HDF5 file 'my_model.h5'

del model  # deletes the existing model



# returns a compiled model

# identical to the previous one

model = load_model('my_model.h5')

By the way, this error also happened when me convert this model by Tensorflow Lite.
But I don't know what's wrong with my model.
Does anyone has an idea?

I hit a similar problem that yields the same error message, but the cause might be different than yours:

Code: (Tensorflow 1.11 and tf.keras.version: 2.1.6-tf)

 if load_model_path.endswith('.h5'):

        model = tf.keras.models.load_model(load_model_path)

Error message:

  File "...../lib/python3.6/site-packages/tensorflow/python/keras/engine/saving.py", line 251, in load_model

    training_config['weighted_metrics'])

KeyError: 'weighted_metrics'

And I found out it's because the model was saved in an older Keras version.
I had to comment out the code related to weighted_metrics to be able to load the model. However, it's just a workaround before I can find a sustainable solution to the mismatching problem. Interestingly, @fchollet just added weighted_metrics to the latest Keras version recently (Oct 2018).
https://github.com/keras-team/keras/blob/master/keras/engine/saving.py#L136
I hope this will help the people who hit the same problem as I did.

If you haven't figured out the answer to this yet, I think I've got it.

I haven't dug into the code to exactly figure out why, but basically the model checkpoint callback can only be loaded with the load_weights() function, which is then used for evaluation.

If you want to save a model that you can load to train again later you need to use model.save and model.load_model. Hopefully helpful to someone who wanders upon this.