CODEWITHSUNDEEP
pythontensorflowmachine-learningkerastensorflow2.0
sebtheiler
sebtheiler

Reputation: 2557

Simple Keras Network in GradientTape: LookupError: No gradient defined for operation 'IteratorGetNext' (op type: IteratorGetNext)

I've built a very simple TensorFlow Keras model with a single dense layer. It works perfectly fine outside a GradientTape block, but inside a GradientTape block it raises LookupError: No gradient defined for operation 'IteratorGetNext' (op type: IteratorGetNext)

Code to reproduce:

from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Dense
import tensorflow as tf
import numpy as np

print(tf.__version__)

model = Sequential()
model.add(Dense(1, input_shape=(16,)))

fake_data = np.random.random((1, 16))

print(model.predict(fake_data).shape) # works
with tf.GradientTape() as tape:
    print(model.predict(fake_data).shape) # LookupError: No gradient defined for operation 'IteratorGetNext' (op type: IteratorGetNext)

This appears to work in TensorFlow 2.0.0, however it fails in TensorFlow 2.1.0 and 2.2.0

Here is a notebook that replicates the issue.

Upvotes: 12

Views: 6448

Answers (1)

Marco Cerliani
Marco Cerliani

Reputation: 22031

try to redefine the predict operation in GradientTape in this way

with tf.GradientTape() as tape:
    print(model(fake_data).shape)

Upvotes: 15

Related Questions