TQDMProgressBar not working in TF-2.2.0rc1

[bthorsted]

System information

• OS Platform and Distribution: Linux Ubuntu 18.04
• TensorFlow version and how it was installed (source or binary): TF-2.2.0rc1 (wheel compiled from source)
• TensorFlow-Addons version and how it was installed (source or binary): 0.8.3 installed via pip
• Python version: 3.7.6
• Is GPU used? (yes/no): Yes

Describe the bug

Executing model.fit() with the TQDMProgressBar() callback results in KeyError: 'metrics' because of a change in TF-2.2 that moves initialization of model.metrics (and model.metrics_names) from compile stage to train stage.

Code to reproduce the issue

import numpy as np
import tensorflow as tf
import tensorflow_addons as tfa

x = np.random.random((5,1,5))
y = np.random.random((5,1,5))

inputs = tf.keras.layers.Input(shape=(3,))
outputs = tf.keras.layers.Dense(2, name="out_1")(inputs)
model = tf.keras.models.Model(inputs=inputs, outputs=outputs)
model.compile(optimizer="Adam", loss="mse", metrics=["acc"])

pg = tfa.callbacks.TQDMProgressBar()
model_callbacks = [pg, ]
VERBOSE=0
history = model.fit(
    x,
    y,
    epochs=100,
    verbose=VERBOSE,
    callbacks=model_callbacks
)

Other info / logs

---------------------------------------------------------------------------
KeyError                                  Traceback (most recent call last)
<ipython-input-23-fdbb03f574a1> in <module>
     48 #   class_weight=class_weights,
     49     verbose=VERBOSE,
---> 50     callbacks=model_callbacks,
     51 )

~/.pyenv/versions/3.7.6/lib/python3.7/site-packages/tensorflow/python/keras/engine/training.py in _method_wrapper(self, *args, **kwargs)
     63   def _method_wrapper(self, *args, **kwargs):
     64     if not self._in_multi_worker_mode():  # pylint: disable=protected-access
---> 65       return method(self, *args, **kwargs)
     66 
     67     # Running inside `run_distribute_coordinator` already.

~/.pyenv/versions/3.7.6/lib/python3.7/site-packages/tensorflow/python/keras/engine/training.py in fit(self, x, y, batch_size, epochs, verbose, callbacks, validation_split, validation_data, shuffle, class_weight, sample_weight, initial_epoch, steps_per_epoch, validation_steps, validation_batch_size, validation_freq, max_queue_size, workers, use_multiprocessing, **kwargs)
    763       self.stop_training = False
    764       train_function = self.make_train_function()
--> 765       callbacks.on_train_begin()
    766       # Handle fault-tolerance for multi-worker.
    767       # TODO(omalleyt): Fix the ordering issues that mean this has to

~/.pyenv/versions/3.7.6/lib/python3.7/site-packages/tensorflow/python/keras/callbacks.py in on_train_begin(self, logs)
    445     logs = self._process_logs(logs)
    446     for callback in self.callbacks:
--> 447       callback.on_train_begin(logs)
    448 
    449   def on_train_end(self, logs=None):

~/.pyenv/versions/3.7.6/lib/python3.7/site-packages/tensorflow_addons/callbacks/tqdm_progress_bar.py in on_train_begin(self, logs)
    100     def on_train_begin(self, logs=None):
    101         self.num_epochs = self.params["epochs"]
--> 102         self.metrics = self.params["metrics"]
    103 
    104         if self.show_overall_progress:

KeyError: 'metrics'

[gabrieldemarmiesse]

I believe that it's going to be fixed with #1365 . We lack tests for those, which explains why they break unexpectedly.

[bthorsted]

@gabrieldemarmiesse I can confirm that the patch you mentioned fixed my issue 👍 It does introduce a new issue, however. Validation metrics are not printed at the end of an epoch anymore.

[gabrieldemarmiesse]

Yeah we definitly lack some proper testing for the callbacks in Addons.