Whether to output all intermediates from functional control flow ops.
tf.experimental.output_all_intermediates(
state
)
The "default" behavior to is to output all intermediates when using v2 control flow inside Keras models in graph mode (possibly inside Estimators). This is needed to support taking gradients of v2 control flow. In graph mode, Keras can sometimes freeze the forward graph before the gradient computation which does not work for v2 control flow since it requires updating the forward ops to output the needed intermediates. We work around this by proactively outputting the needed intermediates when building the forward pass itself. Ideally any such extra tensors should be pruned out at runtime. However, if for any reason this doesn't work for you or if you have an infernce-only model you can turn this behavior off using tf.compat.v1.experimental.output_all_intermediates(False).
If with the default behavior you are still seeing errors of the form "Connecting to invalid output X of source node Y which has Z outputs" try setting tf.compat.v1.experimental.output_all_intermediates(True) and please file an issue at https://github.com/tensorflow/tensorflow/issues.
| Args | |
|---|---|
state | True, False or None. None restores the default behavior. |
© 2020 The TensorFlow Authors. All rights reserved.
Licensed under the Creative Commons Attribution License 3.0.
Code samples licensed under the Apache 2.0 License.
https://www.tensorflow.org/versions/r1.15/api_docs/python/tf/experimental/output_all_intermediates