CODEWITHSUNDEEP
pythontensorflow
Fanto
Fanto

Reputation: 367

iterating over `tf.Tensor` is not allowed: AutoGraph did convert this function

I have a tensor with inside some integer value, i need to create a new tensor with all the value from the integer value*49 to value*49+49.

Example: tf.random.uniform() generates values [0,1] Then i need a new tensor with values: `[0,1,2,3,..48, 98, 99, ..., 147] But if i do something like this:

indices = tf.constant([], tf.int32)
for index in tf.random.uniform((51,), minval=0, maxval=62, dtype=tf.dtypes.int32):
    tensor1 = tf.constant([x for x in range(index * 49, index * 49 + 49)])
    indices = tf.concat([indices, tensor1], 0)
temp = tf.gather(huge_tensor, indices, axis=0)

I get this error: OperatorNotAllowedInGraphError: iterating over tf.Tensor is not allowed: AutoGraph did convert this function. This might indicate you are trying to use an unsupported feature.

I need to do something like this because i will later use tf.gather to gather from another huge tensor, all the rows using the indices tensor. If there is a way to tell tf.gather to grab all the rows directly instead of creating the new tensor indices is even better.

Upvotes: 0

Views: 652

Answers (1)

AloneTogether
AloneTogether

Reputation: 26718

You could try using tf.TensorArray and tf.range:

import tensorflow as tf

x = tf.random.uniform((51,), minval=0, maxval=62, dtype=tf.dtypes.int32)
indices = tf.TensorArray(tf.int32, size=0, dynamic_size=True)
huge_tensor = tf.random.normal((3136, 512))

for i in tf.range(tf.shape(x)[0]):
  indices = indices.write(indices.size(), tf.range(x[i] * 49, x[i] * 49 + 49))

tf.print(indices.stack().shape)
result = tf.gather(huge_tensor, indices.stack(), axis=0)
result_shape = tf.shape(result)
result = tf.reshape(result, (result_shape[0] * result_shape[1], result_shape[2]))
tf.print(result.shape)

TensorShape([51, 49])
TensorShape([2499, 512])

Using tf.range, I create 1D sequences based on each value in x. Then I apply stack() to the array indices, which returns the sequences as a stacked tensor.

Upvotes: 1

Related Questions