Module 'gast' has no attribute 'Num'

[mgreenbe]

I got an error message including a request that I report a bug.

Here's the code:

import numpy as np
import tensorflow as tf
import platform

print()
print(f"PLATFORM:\n---------\n{platform.platform()}")
print("\nTENSORFLOW:\n----------")
for a in tf.version.__all__:
    print(f"{a}: {getattr(tf.version, a)}")

print(f"\nNUMPY:\n-----\n{np.version.version}")

print(f"\nPYTHON:\n-------\n{sys.version}\n")

np.random.seed(0)
tf.random.set_seed(0)


model = tf.keras.models.Sequential([
    tf.keras.layers.Dense(1, activation="linear")
])
model.compile(optimizer="sgd", loss="mse")

x = np.random.uniform(size=(1,1))
y = np.random.uniform(size=(1,))
model.fit(x, y, epochs=1)

And here's the output, including system info, module versions, etc.

PLATFORM:
---------
Darwin-18.7.0-x86_64-i386-64bit

TENSORFLOW:
----------
COMPILER_VERSION: 4.2.1 Compatible Apple LLVM 10.0.0 (clang-1000.11.45.5)
GIT_VERSION: v2.0.0-beta1-5101-gc75bb66a99
GRAPH_DEF_VERSION: 119
GRAPH_DEF_VERSION_MIN_CONSUMER: 0
GRAPH_DEF_VERSION_MIN_PRODUCER: 0
VERSION: 2.0.0-rc0

NUMPY:
-----
1.17.1

PYTHON:
-------
3.7.4 (default, Jul  9 2019, 18:13:23) 
[Clang 10.0.1 (clang-1001.0.46.4)]

Train on 1 samples
WARNING:tensorflow:Entity <function Function._initialize_uninitialized_variables.<locals>.initialize_variables at 0x1492748c0> could not be transformed and will be executed as-is. Please report this to the AutoGraph team. When filing the bug, set the verbosity to 10 (on Linux, `export AUTOGRAPH_VERBOSITY=10`) and attach the full output. Cause: module 'gast' has no attribute 'Num'
WARNING: Entity <function Function._initialize_uninitialized_variables.<locals>.initialize_variables at 0x1492748c0> could not be transformed and will be executed as-is. Please report this to the AutoGraph team. When filing the bug, set the verbosity to 10 (on Linux, `export AUTOGRAPH_VERBOSITY=10`) and attach the full output. Cause: module 'gast' has no attribute 'Num'
1/1 [==============================] - 0s 157ms/sample - loss: 1.2336

<tensorflow.python.keras.callbacks.History at 0x14927b5d0>

[tocosastalo]

Hello,

3 hours ago, gast library was updated from 0.2.2 to 0.3.0 and Num constant was removed (check commit serge-sans-paille/gast@231ee7a#diff-88b99bb28683bd5b7e3a204826ead112R99).

TF requires gast>=0.2.0, so newer version of gast module is automatically installed and that's probably why it fails with error AttributeError: module 'gast' has no attribute 'Num'.

This error may occur for example on the line 236 in tensorflow/python/autograph/pyct/qual_names.py (https://github.com/tensorflow/tensorflow/blob/r1.14/tensorflow/python/autograph/pyct/qual_names.py#L236):

if isinstance(s.value, gast.Num):

[jlebar]

@mgreenbe thank you for filing the bug, and @tocosastalo thanks for the info.

Workaround that worked for me: $ pip3 install 'gast==0.2.2'.

[ai-nikolai]

@jlebar & @tocosastalo & @mgreenbe thank you:

@mgreenbe thank you for filing the bug, and @tocosastalo thanks for the info.

Workaround that worked for me: $ pip3 install 'gast==0.2.2'.

This worked for me too: $ pip3 install gast==0.2.2.

[FeynmanDNA]

I encountered the same warning with tensorflow-gpu2.0beta1, and the above fix works for me too:

$ pip install gast==0.2.2
Collecting gast==0.2.2
Installing collected packages: gast
  Found existing installation: gast 0.3.1
    Uninstalling gast-0.3.1:
      Successfully uninstalled gast-0.3.1
Successfully installed gast-0.2.2

[mgreenbe]

Using the older version of gast worked for me too. I'll leave it to @jlebar or someone else on the tensorflow team to close this issue at their discretion, presumably once tensorflow and gast=0.3.x are interacting nicely.

[mihaimaruseac]

#32332 fixes this for 1.15, there's a similar PR for 2.0 and a commit on master. Adding @goldiegadde as she is fixing this

[wchargin]

I don’t understand what the status of this is. The corresponding commit
on master hasn’t fixed this in nightlies:

$ cat test.py
import tensorflow as tf

print(tf.__version__)

@tf.function
def foo():
  return 1

foo()
$ AUTOGRAPH_VERBOSITY=10 python test.py
2.0.0-dev20190910
WARNING:tensorflow:Entity <function foo at 0x7f71ac1990d0> could not be transformed and will be executed as-is. Please report this to the AutoGraph team. When filing the bug, set the verbosity to 10 (on Linux, `export AUTOGRAPH_VERBOSITY=10`) and attach the full output. Cause: Bad argument number for Name: 3, expecting 4

Running pip install gast==0.2.2 fixes the above issue.

I’m not sure what those changes in pip_new.sh do, but is there a plan
to update the Pip dependencies in setup.py so that people who
simply run pip install tf-nightly-2.0-preview get a working install?

[mihaimaruseac]

Classical error of having same dependency specified in two places. I'll send a fix, sorry for this

[wchargin]

thanks! and no worries :-)

[mihaimaruseac]

Should be fixed by c72125b, #32390 and #32391 (the last two need merging before fixing on corresponding branches)

[mihaimaruseac]

Should be fixed now, keeping issue open until verified

[arainboldt]

Thanks guys for the fix! Any chance we'll get a build for mac-python3.7?

[wchargin]

Confirmed fixed for me on tf-nightly-2.0-preview==2.0.0.dev20190911. Thank you!

Feel free to close at your discretion.

[mihaimaruseac]

In that case, closing.

Thank you for verifying.

@arainboldt the fix should get released in all pips we are releasing.

[mihaimaruseac]

New update:

gast has been pinned to 0.2.2 for 1.15 and 2.0 and later releases as well as nightly.
TF 1.14 or earlier don't have gast pinned and won't get it as we are not doing more releases there (except for security vulnerabilities).

If you encounter this issue on 1.14 or earlier, please install gast==0.2.2 using pip before installing tensorflow.

[bestwayyy]

I'm still getting the error even with gast 0.2.2.
I really need help to get this fixed.

[mihaimaruseac]

We are currently working on getting gast==0.3.2. Apologies for the delay.

[wookayin]

Starting from TensorFlow 2.2.0, it will work with gast 0.3.x but NOT gast 0.2.x any more. For TF <= 2.1, gast 0.2.2 is needed.

[ddrous]

Thank you @wookayin