reorganize training.py, code is clearer and handles more corner cases…

[ahundt]

This is the final item in the PR chain necessary to support input tensors. Here is an overview of what will be supported:

# tf yield ops such as those created by the tf `RecordInput()` API
# which supply dataset images and labels
x_train_batch, y_train_batch = read_and_decode_recordinput(...)

# create a basic cnn
x_train_input = Input(tensor=x_train_batch)
x_train_out = cnn_layers(x_train_input)

# y label batch is input & output
# Perhaps this aspect of API usage can be improved?
y_train_in = Input(tensor=y_train_batch)

# ==bugfix==
# This call causes a crash without this patch because
# an invalid call is made that is equivalent to:
# K.placeholder(dtype=x_train_input)
train_model = Model(inputs=[x_train_in], outputs=[x_train_out])

# ==bugfix==
# This call will crash without this patch because
# it is assumed the parameters `x` and `y` are
# provided here and not via the ops
# x_train_batch and y_train_batch 
train_model.compile(optimizer='rmsprop',
                    loss='categorical_crossentropy',
                    metrics=['accuracy'])

# ==bugfix== + ==new param==
# This call will crash without this patch because
# the changes in tensor order caused by the
# constructor update, which accepts yield ops,
# were not previously accounted for.
#
# A new param steps_per_epoch is added
# which works just like in fit_generator()
train_model.fit(None, y_train_in, 
                batch_size=batch_size,
                epochs=epochs,
                steps_per_epoch=10000)

The combined version of these changes (with more detail) is in #6928.

Changes to class Model(Container):

• relocate _prepare_sample_weights() to be a distinct function
• define _make_ins() so backend Function() params are created consistently
• add additional checks to None in many places
• raise errors in additional exceptional situations
• add _make_function() so all utilities that create an instance of backend.Function() behave consistently (modulo deliberate / correct differences)

Depends on:
#7067, #7066

Additional fixes in:
#7064, #7067, #7068, #7069, #7071, #7072

Tested by:
#7061

Also see the comments by @Dref360 on the equivalent changes. #7046 (review)

Note: travis failure will be resolved once dependencies above are merged

[ahundt]

@fchollet can we take the next steps on this PR and its dependencies? They are valid error detection and reporting improvements independent of the API Design Review #7102 since they leave the API unchanged.

Here is a good order to look at them that accounts for deps:

1. training.py _standardize_input_data() cleanup input tensor corner cases #7067 _standardize_input_data() support non-numpy tensor inputs, clear error messages for unsupported inputs, and unit test
2. training.py _check_loss_and_target_compatibility() fix crash if y is None #7071 training.py _check_loss_and_target_compatibility()
   • bugfix: don't access y param of model.fit() if it is None
3. Tensorflow backend Function class __init__() and __call__() error handling + fetches + feed_dict #7064 TensorFlow backend Function class init() and call() error handling
4. is_placeholder() implemented #7066 K.is_placeholder(x) implemented
   • Returns whether x contains a K.placeholder()
5. model.fit(steps_per_epoch), mnist_tfrecord.py, progbar np.mean #7113 model.fit(steps_per_epoch) added
6. reorganize training.py, code is clearer and handles more corner cases… #7072 reorganize training.py, code is clearer and handles more corner cases
   • Depends on training.py _standardize_input_data() cleanup input tensor corner cases #7067, training.py _check_loss_and_target_compatibility() fix crash if y is None #7071, is_placeholder() implemented #7066 (passes once those are merged)
7. Yield op tests #7061 yield op unit tests
   • might need a small tweak once above are merged to pass
   • unit test for reorganize training.py, code is clearer and handles more corner cases… #7072

[stale]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed after 30 days if no further activity occurs, but feel free to re-open a closed issue if needed.

[ahundt]

This would need to be redone due to changes on master and I've run out of cycles to keep it up to date, closing for now. If others are interested in incorporating similar changes please let me know and I'll be happy to help.