How to define dynamic-shape variable when building computational graph with Tensorflow 1.15

Question

System information

1. Have I written custom code (as opposed to using a stock example script provided in TensorFlow): No

2. OS Platform and Distribution (eg, Linux Ubuntu 16.04): Linux Ubuntu 18.04

3. Mobile device (eg iPhone 8, Pixel 2, Samsung Galaxy) if the issue happens on mobile device:

4. TensorFlow installed from (source or binary): Conda repo

5. TensorFlow version (use command below): 1.15

6. Python version: 3.7.7

7. Bazel version (if compiling from source):

8. GCC/Compiler version (if compiling from source):

9. CUDA/cuDNN version: 10.1

10. GPU model and memory: Tesla V100-SMX3-32GB

11. Describe the current behavior

    tensorflow.python.framework.errors_impl.InvalidArgumentError: Assign requires shapes of both tensors to match. lhs shape= [] rhs shape= [1,1] [[{{node Variable/Assign}}]]

Describe the expected behavior

No error

Standalone code to reproduce the issue

import tensorflow as tf
import numpy as np
import os
os.environ["CUDA_VISIBLE_DEVICES"]='0'

with tf.Session() as sess:
    v = tf.Variable(np.zeros(shape=[1,1]),shape=tf.TensorShape(None))
    sess.run(tf.global_variables_initializer())

Obseration: The error did not appear when I use eager_execution_mode()

Code:

tf.enable_eager_execution()
v = tf.Variable(np.zeros([1,1]),shape=tf.TensorShape(None))
tf.print(v)
v.assign(np.ones([2,2]))
tf.print(v)    

Output:

[[0]]
[[1 1]
 [1 1]]

Link to a MWE: https://colab.research.google.com/gist/amahendrakar/3fe8345db4092d520246205be4b97948/41620.ipynb

Answer 1

Just have to enable resource variable as dynamic shape behavior is only available for this 'updated' Variable class.

import tensorflow as tf
import numpy as np
import os
os.environ["CUDA_VISIBLE_DEVICES"]='0'
tf.compat.v1.enable_resource_variables()

with tf.Session() as sess:
    v = tf.Variable(np.zeros(shape=[1,1]),shape=tf.TensorShape(None))
    sess.run(tf.global_variables_initializer())