Support Apple Silicon

[tribustale]

Hi, is it possible that this library one day will be usable in the new generation of architecture Apple Silicon M1?

I tried to install it in a conda environment (through miniforge) after having installed also TensorFlow 2.4 benchmark on Mac M1, but it doesn't work! It gives an error related with bazel.

[alvarosg]

Thank your for your message. Is the problem related to this library specifically, or are you having troubles in general running any TensorFlow code on the hardware?

[tribustale]

I tried in three different ways;

• installing tensorflow_macos for the new arm64 architecture in a virtual environment: in this way tensorflow2.4 works but graph_nets cannot be imported due to problem installing bazel;
• installing tensorflow_macos for the old x86_64 architecture: in this way graph_nets can be installed but when I call tensorflow my jupyter kernel crashes;
• installing tensorflow from source, using terminal with Rosetta2, I can install graph_nets but tensorflow again crashes.

[alvarosg]

Could you share the full stack trace when importing graph_nets in the first case? Which line of the import fails?

I think it tensorflow works, then graph_nets should work, since graph_nets is pure python, so if it fails with a bazel related problem, this probably means tensorflow does not fully work.

[tribustale]

In the first case, using a jupyter-lab notebook I run

install_graph_nets_library = "Yes"  #@param ["Yes", "No"]

if install_graph_nets_library.lower() == "yes":
  print("Installing Graph Nets library and dependencies:")
  print("Output message from command:\n")
  !pip install "graph_nets>=1.1" "dm-sonnet>=2.0.0b0"
else:
  print("Skipping installation of Graph Nets library")

and everything goes fine. All the following packages seem installed:

• graph_nets (1.1.0)
• dm-sonnet (2.0.0)
• wrapt (1.12.1)
• dm-tree (0.1.5)
• tabulate (0.8.7)
• numpy (1.18.5)
• absl-py (0.11.0)
• six (1.15.0)
• networkx (2.5)
• setuptools (53.0.0)
• future (0.18.2)
• decorator (4.4.2)
• tensorflow_probability (0.12.1).

It appears that pip doesn't find tensorflow, since when running:

pip show tensorflow 

comes out

WARNING: Package(s) not found: tensorflow

even tough tensorflow is installed through tensorflow_macos. This is confirmed by importing tensorflow into the notebook:

import tensorflow as tf
tf.__version__

'2.4.0-rc0'

but when running

from graph_nets import utils_np
from graph_nets import utils_tf
from graph_nets import modules

the following error comes up :

---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
~/tdametta/lib/python3.8/site-packages/tensorflow/python/autograph/impl/api.py in to_graph(entity, recursive, experimental_optional_features)
    748             optional_features=experimental_optional_features))
--> 749     return autograph_artifact(_convert_actual(entity, program_ctx))
    750   except (ValueError, AttributeError, KeyError, NameError, AssertionError) as e:

~/tdametta/lib/python3.8/site-packages/tensorflow/python/autograph/impl/api.py in _convert_actual(entity, program_ctx)
    283 
--> 284   transformed, module, source_map = _TRANSPILER.transform(entity, program_ctx)
    285 

~/tdametta/lib/python3.8/site-packages/tensorflow/python/autograph/pyct/transpiler.py in transform(self, obj, user_context)
    285     if inspect.isfunction(obj) or inspect.ismethod(obj):
--> 286       return self.transform_function(obj, user_context)
    287 

~/tdametta/lib/python3.8/site-packages/tensorflow/python/autograph/pyct/transpiler.py in transform_function(self, fn, user_context)
    469           # TODO(mdan): Confusing overloading pattern. Fix.
--> 470           nodes, ctx = super(PyToPy, self).transform_function(fn, user_context)
    471 

~/tdametta/lib/python3.8/site-packages/tensorflow/python/autograph/pyct/transpiler.py in transform_function(self, fn, user_context)
    362     node = self._erase_arg_defaults(node)
--> 363     result = self.transform_ast(node, context)
    364 

~/tdametta/lib/python3.8/site-packages/tensorflow/python/autograph/impl/api.py in transform_ast(self, node, ctx)
    251     unsupported_features_checker.verify(node)
--> 252     node = self.initial_analysis(node, ctx)
    253 

~/tdametta/lib/python3.8/site-packages/tensorflow/python/autograph/impl/api.py in initial_analysis(self, node, ctx)
    238     graphs = cfg.build(node)
--> 239     node = qual_names.resolve(node)
    240     node = activity.resolve(node, ctx, None)

~/tdametta/lib/python3.8/site-packages/tensorflow/python/autograph/pyct/qual_names.py in resolve(node)
    251 def resolve(node):
--> 252   return QnResolver().visit(node)
    253 

/Applications/Xcode.app/Contents/Developer/Library/Frameworks/Python3.framework/Versions/3.8/lib/python3.8/ast.py in visit(self, node)
    359         visitor = getattr(self, method, self.generic_visit)
--> 360         return visitor(node)
    361 

/Applications/Xcode.app/Contents/Developer/Library/Frameworks/Python3.framework/Versions/3.8/lib/python3.8/ast.py in generic_visit(self, node)
    444             elif isinstance(old_value, AST):
--> 445                 new_node = self.visit(old_value)
    446                 if new_node is None:

/Applications/Xcode.app/Contents/Developer/Library/Frameworks/Python3.framework/Versions/3.8/lib/python3.8/ast.py in visit(self, node)
    359         visitor = getattr(self, method, self.generic_visit)
--> 360         return visitor(node)
    361 

/Applications/Xcode.app/Contents/Developer/Library/Frameworks/Python3.framework/Versions/3.8/lib/python3.8/ast.py in generic_visit(self, node)
    435                     if isinstance(value, AST):
--> 436                         value = self.visit(value)
    437                         if value is None:

/Applications/Xcode.app/Contents/Developer/Library/Frameworks/Python3.framework/Versions/3.8/lib/python3.8/ast.py in visit(self, node)
    359         visitor = getattr(self, method, self.generic_visit)
--> 360         return visitor(node)
    361 

~/tdametta/lib/python3.8/site-packages/tensorflow/python/autograph/pyct/qual_names.py in visit_Name(self, node)
    216   def visit_Name(self, node):
--> 217     node = self.generic_visit(node)
    218     anno.setanno(node, anno.Basic.QN, QN(node.id))

/Applications/Xcode.app/Contents/Developer/Library/Frameworks/Python3.framework/Versions/3.8/lib/python3.8/ast.py in generic_visit(self, node)
    444             elif isinstance(old_value, AST):
--> 445                 new_node = self.visit(old_value)
    446                 if new_node is None:

/Applications/Xcode.app/Contents/Developer/Library/Frameworks/Python3.framework/Versions/3.8/lib/python3.8/ast.py in visit(self, node)
    359         visitor = getattr(self, method, self.generic_visit)
--> 360         return visitor(node)
    361 

~/tdametta/lib/python3.8/site-packages/tensorflow/python/autograph/pyct/qual_names.py in visit_Subscript(self, node)
    231     s = node.slice
--> 232     if not isinstance(s, gast.Index):
    233       # TODO(mdan): Support range and multi-dimensional indices.

AttributeError: module 'gast' has no attribute 'Index'

During handling of the above exception, another exception occurred:

ConversionError                           Traceback (most recent call last)
<ipython-input-39-29055414a904> in <module>
----> 1 from graph_nets import utils_np
      2 from graph_nets import utils_tf
      3 from graph_nets import modules

~/tdametta/lib/python3.8/site-packages/graph_nets/__init__.py in <module>
     19 from __future__ import print_function
     20 
---> 21 from graph_nets import blocks
     22 from graph_nets import graphs
     23 from graph_nets import modules

~/tdametta/lib/python3.8/site-packages/graph_nets/blocks.py in <module>
     33 from __future__ import print_function
     34 
---> 35 from graph_nets import _base
     36 from graph_nets import graphs
     37 from graph_nets import utils_tf

~/tdametta/lib/python3.8/site-packages/graph_nets/_base.py in <module>
     51 
     52 import six
---> 53 import sonnet as snt
     54 
     55 _sonnet_version = snt.__version__

~/tdametta/lib/python3.8/site-packages/sonnet/__init__.py in <module>
     19 from __future__ import print_function
     20 
---> 21 from sonnet import distribute
     22 from sonnet import initializers
     23 from sonnet import mixed_precision

~/tdametta/lib/python3.8/site-packages/sonnet/distribute.py in <module>
     19 from __future__ import print_function
     20 
---> 21 from sonnet.src.distribute.batch_norm import CrossReplicaBatchNorm
     22 from sonnet.src.distribute.replicator import create_variables_eagerly
     23 from sonnet.src.distribute.replicator import Replicator

~/tdametta/lib/python3.8/site-packages/sonnet/src/distribute/batch_norm.py in <module>
     20 from __future__ import print_function
     21 
---> 22 from sonnet.src import batch_norm
     23 from sonnet.src import initializers
     24 from sonnet.src import metrics

~/tdametta/lib/python3.8/site-packages/sonnet/src/batch_norm.py in <module>
     32 
     33 
---> 34 class BaseBatchNorm(base.Module):
     35   r"""Batch normalization module.
     36 

~/tdametta/lib/python3.8/site-packages/sonnet/src/batch_norm.py in BaseBatchNorm()
    130 
    131   @utils.smart_autograph
--> 132   def __call__(self,
    133                inputs: tf.Tensor,
    134                is_training: types.BoolLike,

~/tdametta/lib/python3.8/site-packages/sonnet/src/utils.py in smart_autograph(f)
    178     A wrapper for `f` that dispatches to the original or autograph version of f.
    179   """
--> 180   f_autograph = tf.autograph.to_graph(f)
    181 
    182   @functools.wraps(f)

~/tdametta/lib/python3.8/site-packages/tensorflow/python/autograph/impl/api.py in to_graph(entity, recursive, experimental_optional_features)
    750   except (ValueError, AttributeError, KeyError, NameError, AssertionError) as e:
    751     logging.error(1, 'Error converting %s', entity, exc_info=True)
--> 752     raise ConversionError('converting {}: {}: {}'.format(
    753         entity, e.__class__.__name__, str(e)))
    754 

ConversionError: converting <function BaseBatchNorm.__call__ at 0x124539ca0>: AttributeError: module 'gast' has no attribute 'Index'

What should I try? It appears it has a problem with gast and the new version 0.4 (that doesn't have .Index()), I tried with downgrade to 0.3.3 but it doesn't change.

[alvarosg]

The problem seems to happen when importing sonnet, so I would recommend to try to reproduce the error just by importing and installing sonnet, and if it still happens, raise an issue against the sonnet repo.

[alvarosg]

(closing due to inactivity)