Keras 3.3.3 upgrade #165
Keras 3.3.3 upgrade #165
Conversation
Rolled over changes for Keras 3.3 for all tf.Model classes
bayesflow/coupling_networks.py
Outdated
| @@ -436,7 +439,7 @@ def _semantic_spline_parameters(self, parameters): | |||
| """ | |||
|
|
|||
| shape = tf.shape(parameters) | |||
| rank = len(shape) | |||
| rank = parameters.shape.rank | |||
There was a problem hiding this comment.
tensorflow-specific, len(shape) or parameters.ndim should be fine. Safest would probably be keras.ops.ndim
There was a problem hiding this comment.
Changed parameters.shape.rank to use keras.ops.ndim in coupling_networks.py (see 47fbbbd)
| # LearningRateSchedule instances need number of iterations | ||
| current_lr = optimizer.lr(optimizer.iterations).numpy() | ||
| elif hasattr(optimizer.lr, "numpy"): | ||
| current_lr = optimizer.learning_rate(optimizer.iterations).numpy() |
There was a problem hiding this comment.
not sure if this works under jax/numpy backend - can you test this? As an alternative, we could use keras.ops.convert_to_numpy(). We can then drop the hasattr check
There was a problem hiding this comment.
Changed learning_rate().numpy() to keras.ops.convert_to_numpy() in helper_functions.py (see 47fbbbd)
bayesflow/inference_networks.py
Outdated
| target_shape = tf.shape(targets) | ||
| condition_shape = tf.shape(condition) | ||
| target_shape = keras.backend.shape(targets) | ||
| condition_shape = keras.backend.shape(condition) |
There was a problem hiding this comment.
as discussed, please use keras.ops
There was a problem hiding this comment.
Changed requested keras.backend instances to keras.ops in inference_networks.py. (see 47fbbbd)
Will change remaining keras.backend to keras.ops calls with next update
There was a problem hiding this comment.
Remaining keras.backend calls switched to respective keras.ops functions (see b9d5f16)
Changed parameters.shape.rank to use keras.ops.ndim in coupling_networks.py, Changed learning_rate.numpy() to keras.ops.convert_to_numpy() in helper_functions.py, Changed requested keras.backend instances to keras.ops in inference_networks.py.
Changed all calls of keras.backend to the respective keras.ops function
|
We cannot merge this anymore due to the many changes on the branch, but we'll use this as a reference nonetheless. |
tf.model,tf.layer, andtf.optimizer@tf.functiondecorator tocoupling_networks._calculate_spline()to allow execution in non-eager environmentstf.Tensorin docstring toTensorw.r.t. the experimental Tensor typeThe included changes have worked successfully in all of the example / tutorial notebooks, however I'm sure there are a few introduced bugs laying around when TensorFlow is not the chosen backend.
Current environment dependencies: