将 tensorflow 1.xx model 装入 tensorflow 2.xx

Question

I have a SavedModel created with TF1 being loaded with TF2.

I am getting a warning for each variable in the graph it seems, which is:

WARNING:tensorflow:Unable to create a python object for variable <tf.Variable 'Encoder_en/hidden_layers/tanh_layer_0/bias:0' shape=(512,) dtype=float32_ref> because it is a reference variable. It may not be visible to training APIs. If this is a problem, consider rebuilding the SavedModel after running tf.compat.v1.enable_resource_variables().

I would like to preferably fix this warning, or just suppress it!

So far I have tried:

# In my python app
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'

# In my Dockerfile
ENV TF_CPP_MIN_LOG_LEVEL 2

EDIT: this model is from tensorflow hub, therefore I have not built it.

Answer 1

Logging in TensorFlow changed in more recent versions, and TF_CPP_MIN_LOG_LEVEL is not used anymore (see issues #26348 and #31870 ). Try with tf.get_logger().setLevel('ERROR') .

import tensorflow as tf
tf.get_logger().warning('test')
# WARNING:tensorflow:test
tf.get_logger().setLevel('ERROR')
tf.get_logger().warning('test')
# (silence)