tensorflow 2.3.1 incompatibility

[jakob-r]

Trying out the example I get the following error.

In [7]: from nasbench import api
---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
<ipython-input-7-5ba3f4b880ca> in <module>
----> 1 from nasbench import api

~/dump/nasbench/nasbench/api.py in <module>
     96 
     97 from nasbench.lib import config
---> 98 from nasbench.lib import evaluate
     99 from nasbench.lib import model_metrics_pb2
    100 from nasbench.lib import model_spec as _model_spec

~/dump/nasbench/nasbench/lib/evaluate.py in <module>
     22 
     23 from nasbench.lib import cifar
---> 24 from nasbench.lib import model_builder
     25 from nasbench.lib import training_time
     26 import numpy as np

~/dump/nasbench/nasbench/lib/model_builder.py in <module>
     29 
     30 from nasbench.lib import base_ops
---> 31 from nasbench.lib import training_time
     32 import numpy as np
     33 import tensorflow as tf

~/dump/nasbench/nasbench/lib/training_time.py in <module>
    128 
    129 
--> 130 class _TimingRunHook(tf.train.SessionRunHook):
    131   """Hook to stop the training after a certain amount of time."""
    132 

AttributeError: module 'tensorflow._api.v2.train' has no attribute 'SessionRunHook'

[ThomasBas]

I found a workaround by replacing 'train' with 'estimator' (tf.estimator.SessionRunHook) in the training_time.py file.

It will also raise a similar error for 'tf.train.CheckpointSaverListener' in training_time.py - line:174 and for 'tf.train.NanLossDuringTrainingError' in evaluate.py - line:30. Again replacing 'train' with 'estimator' should fix this for now.

Finally, in api.py replacing 'tf.python_io.tf_record_iterator' with 'tf.compat.v1.io.tf_record_iterator' should allow you to run example.py. Hope this helps!