,
7 tweets,
2 min read
Read on Twitter
(1/7) My experience today to migrate my code from #Keras to Keras for @TensorFlow (tf.keras) has been a bit painful unfortunately. Some issue I encountered (@fchollet maybe this is of interest to you?):
(2/7) I found out that 1.13 doesn't have support for float16 in linear algebra operators, luckily this pull request now landed in 1.14-RC0 github.com/tensorflow/ten…
(3/7) It seems that Dropout on an input with a partially unknown shape, and a noise_shape of (None, 1, None), does not work in tf.keras. I could work around it but in the original Keras this works just fine.
(4/7) Existing multi_gpu_model utility doesn't seem te work, however I was unable to use MirroredStrategy because sample_weight_mode is not supported with tf.distribute.Strategy. Also see github.com/tensorflow/ten….
(5/7) I created my own Adam optimizer to deal with loss scaling when using float16, I need to refactor it because the interface of Optimizer changed.
(6/7) I was using weighted_masked_objective which is not available anymore in TF 1.14, likely I can use engine.training_utils.call_metric_function but need to refactor accordingly.
(7/7) My (conversation, sequence-to-sequence with extensions) model is still not training in tf.keras, so I'm expecting multiple other issues.
