2

I understand that Eager mode is a new alpha feature on the nightly builds and that it is not perfect yet, but I do not know if there are any tf.keras workarounds for this problem.

The error Layer.input not supported in Eager mode. triggers on the block

model = tf.keras.models.Sequential()
model.add(tf.layers.Dense(2, input_shape = (None, 1)))
model.add(tf.layers.Dense(units = 1))
model.compile(optimizer = "sgd", loss = "mean_squared_error")

I do not know anything about keras or the keras tensorflow API and I was wondering if there was a way to avoid Layer.input with keras techniques so as to stay within Eager mode. Following a tutorial in the tf.Eager docs I have confirmed that model = tf.layers.Dense(1) works but I don't know how to add another layer.

Any help is very much appreciated.

EDIT As of tensorflow v1.10, keras is supported in eager mode.

abeoliver
  • 152
  • 1
  • 10
  • Did you activate tensorflow eager mode by calling `tfe.enable_eager_execution()` before running this code block? Eager mode is disabled by default. – Shanqing Cai Nov 09 '17 at 19:40
  • @scai Yes I did, I believe that whatever interface keras wants to use simply isn't supported yet. I am wondering what keras methods there are to avoid `Layer.input` – abeoliver Nov 13 '17 at 19:36
  • I believe this should be resolved now, as Eager now supports Keras Models. – Max Ghenis Sep 09 '18 at 19:07

1 Answers1

1

Keras Models are not yet supported with eager execution, but Keras layers are. Which means that while you can't use tf.keras.models.Sequential yet, you could combine layers yourself. See the user guide.

Hope that helps.

ash
  • 6,681
  • 3
  • 18
  • 30