Compatibility with tensorflow = 2.5.0 and Keras = 2.4

[zeydabadi]

I tried to run the example code below and receive the following error:
model = tf.keras.Sequential( layers=tf.keras.layers.Dense(units=2), name="Linear") model.compile(loss=tf.keras.losses.MeanSquaredError(), optimizer="adam")
ImportanceTraining(model).fit( x_train, y_train, batch_size=32, epochs=10, verbose=1,validation_data=(x_val, y_val) )

The error is below:

`---------------------------------------------------------------------------
RuntimeError Traceback (most recent call last)
in
----> 1 ImportanceTraining(model).fit(
2 x_train, y_train,
3 batch_size=32,
4 epochs=10,
5 verbose=1,validation_data=(x_val, y_val)

~/myvenv/mykears3.9/lib/python3.9/site-packages/importance_sampling/training.py in init(self, model, presample, tau_th, forward_batch_size, score, layer)
368
369 # Call the parent to wrap the model
--> 370 super(ImportanceTraining, self).init(model, score, layer)
371
372 def sampler(self, dataset, batch_size, steps_per_epoch, epochs):

~/myvenv/mykears3.9/lib/python3.9/site-packages/importance_sampling/training.py in init(self, model, score, layer)
338 self._reweighting = BiasedReweightingPolicy(1.0) # no bias
339
--> 340 super(_UnbiasedImportanceTraining, self).init(model, score, layer)
341
342 @Property

~/myvenv/mykears3.9/lib/python3.9/site-packages/importance_sampling/training.py in init(self, model, score, layer)
33 self._check_model(model)
34 self.original_model = model
---> 35 self.model = OracleWrapper(
36 model,
37 self.reweighting,

~/myvenv/mykears3.9/lib/python3.9/site-packages/importance_sampling/model_wrappers.py in init(self, model, reweighting, score, layer)
167 # Augment the model with reweighting, scoring etc
168 # Save the new model and the training functions in member variables
--> 169 self._augment_model(model, score, reweighting)
170
171 def _gnorm_layer(self, model, layer):

~/myvenv/mykears3.9/lib/python3.9/site-packages/importance_sampling/model_wrappers.py in _augment_model(self, model, score, reweighting)
203 loss = model.loss
204 optimizer = model.optimizer.class(**model.optimizer.get_config())
--> 205 output_shape = model.get_output_shape_at(0)[1:]
206 if isinstance(loss, str) and loss.startswith("sparse"):
207 output_shape = output_shape[:-1] + (1,)

~/myvenv/mykears3.9/lib/python3.9/site-packages/tensorflow/python/keras/engine/base_layer.py in get_output_shape_at(self, node_index)
1986 RuntimeError: If called in Eager mode.
1987 """
-> 1988 return self._get_node_attribute_at_index(node_index, 'output_shapes',
1989 'output shape')
1990

~/myvenv/mykears3.9/lib/python3.9/site-packages/tensorflow/python/keras/engine/base_layer.py in _get_node_attribute_at_index(self, node_index, attr, attr_name)
2582 """
2583 if not self._inbound_nodes:
-> 2584 raise RuntimeError('The layer has never been called '
2585 'and thus has no defined ' + attr_name + '.')
2586 if not len(self._inbound_nodes) > node_index:

RuntimeError: The layer has never been called and thus has no defined output shape.
`

[angeloskath]

Hi,

I will close this issue since it is a known incompatibility. Check #25 for more information.

Cheers,
Angelos

[QuetzalcoatlRosso]

Here is a working example (as of earlier this year):
https://github.com/ibarrien/Adam-with-Bandit-Sampling/blob/main/banditutils/setup.py

See in general that repo:
https://github.com/ibarrien/Adam-with-Bandit-Sampling/

[angeloskath]

@QuetzalcoatlRosso wow thanks for the help :-). Admittedly, I should put this in the README.

[QuetzalcoatlRosso]

@angeloskath : cool, most welcome! README + setup.py that explicate the required versions would be awesome