tf.keras works, tf.python.keras doesn't

[OverLordGoldDragon]

DOESN'T WORK:

from tensorflow.python.keras.layers import Input, Dense
from tensorflow.python.keras.models import Model
from tensorflow.python.keras.optimizers import Nadam

ipt = Input(shape=(4,))
out = Dense(1, activation='sigmoid')(ipt)

model = Model(ipt, out)
model.compile(optimizer=Nadam(lr=1e-4), loss='binary_crossentropy')

X = np.random.randn(32,4)
Y = np.random.randint(0,2,(32,1))
model.train_on_batch(X,Y)

WORKS: remove .python from above's imports. Above's error trace below.

Keras 2.3.0 and TensorFlow 2.0.0 freshly-installed via Anaconda, older versions uninstalled. Why the difference?

---

  File "<ipython-input-7-1e86d21d8fc4>", line 13, in <module>
    model.train_on_batch(X,Y)

  File "D:\Anaconda\envs\tf2_env\lib\site-packages\tensorflow_core\python\keras\engine\training.py", line 1017, in train_on_batch
    self._make_train_function()

  File "D:\Anaconda\envs\tf2_env\lib\site-packages\tensorflow_core\python\keras\engine\training.py", line 2116, in _make_train_function
    params=self._collected_trainable_weights, loss=self.total_loss)

  File "D:\Anaconda\envs\tf2_env\lib\site-packages\tensorflow_core\python\keras\optimizers.py", line 653, in get_updates
    grads = self.get_gradients(loss, params)

  File "D:\Anaconda\envs\tf2_env\lib\site-packages\tensorflow_core\python\keras\optimizers.py", line 92, in get_gradients
    if None in grads:

  File "D:\Anaconda\envs\tf2_env\lib\site-packages\tensorflow_core\python\ops\math_ops.py", line 1336, in tensor_equals
    return gen_math_ops.equal(self, other)

  File "D:\Anaconda\envs\tf2_env\lib\site-packages\tensorflow_core\python\ops\gen_math_ops.py", line 3627, in equal
    name=name)

  File "D:\Anaconda\envs\tf2_env\lib\site-packages\tensorflow_core\python\framework\op_def_library.py", line 545, in _apply_op_helper
    (input_name, err))

ValueError: Tried to convert 'y' to a tensor and failed. Error: None values not supported.

---

UPDATE: Debugging the two side-by-side, while both use the same files, execution diverges fairly quickly:

# .\tensorflow_core\python\keras\engine\training.py

### TF.KERAS
    if self._experimental_run_tf_function: #  TRUE
	
### TF.PYTHON.KERAS
    if self._experimental_run_tf_function: #  FALSE

Former proceeds to call training_v2_utils.train_on_batch(...) and returns thereafter, latter self._standardize_user_data(...) and others before ultimately failing.

One difference I noted between linked file and mine is, latter's short by ~100 lines - though I installed TF 2 via pip after the file's last update 12 days ago according to Github.

[OverLordGoldDragon]

The "one difference" was the key; pip failed - downloaded manually. Looking forward to Anaconda support.

[OverLordGoldDragon]

Nope, wasn't it - somehow made it temporarily work, then something else broke, then couldn't get it to work again with several fresh re-installs. Pip also fails to fetch Github's TensorFlow master branch. Likely a bug. Related SO

[pandrey-fr]

Keras 2.3.0 and TensorFlow 2.0.0 freshly-installed

This is a side note to your issue, but with tensorflow 2.0 you are not using the separate keras package (instead, you are using the keras implementation embarked within tensorflow).

That being said, the imports should work, both with and without specifying .python.. In general, if you place yourself as a user, I think you had better import from tensorflow.keras, which is the intended public display of tensorflow's submodules (if you simply import tensorflow, you have access to tensorflow.keras, but not to tensorflow.python). The python term in the imports refers to the source code organization, and is therefore preferably used when writing source code (which then uses utility functions to reformat its public naming at import time), although you may sometimes need to use it to access specific back-end functionalities which are normally not made public.

[gadagashwini-zz]

@OverLordGoldDragon,
Please use tf.keras instead of tf.python.keras. Please see the gist here. Thanks!

[OverLordGoldDragon]

@gadagashwini Your gist is already in my original post - my question concerns why tf.python.keras fails (it shouldn't)

[mihaimaruseac]

TL;DR: The API import is in the root of the package. Any other import is just Python allowing you to access privates with no consideration for good coding practices.

Long story short:

The only way that imports should be are

import tensorflow as tf

tf.keras....

We also provide support for

from tensorflow.keras import ...

though this is brittle and can break as we keep refactoring.

Importing from tensorflow.python or any other modules (including import tensorflow_core...) is not supported, and can break unannounced.

[OverLordGoldDragon]

@mihaimaruseac That's fair, except tf.keras itself is full of imports from tf.python.keras - and making custom optimizers often requires imports from either of the two, which still ultimately import from tf.python.keras. It occurred with me while building a custom optimizer - making it compatible with optimizer_v2 thus implies dealing with tf.python.keras and all its music. So even if it's dev-only, dev's still bug-prone - and I've opened a PR per this thread's bug.

Regardless, your answer may be the real one here; so to confirm, tf.python.keras is private, intended for development, rather than public use? Thanks

[mihaimaruseac]

Regardless, your answer may be the real one here; so to confirm, tf.python.keras is private, intended for development, rather than public use? Thanks

Yes, that's exactly the case. Anything under tf.python is private