AttributeError: module 'typing' has no attribute '_SpecialForm'

Amazon SageMaker
1

I’m getting the following error when trying to train wav2vec2 on SageMaker. I was not getting this error last week with the same code.

It appears to be an issue with datasets use of pyarrow.

Traceback (most recent call last):
  File "pretrain.py", line 48, in <module>
    import datasets
  File "/opt/conda/lib/python3.6/site-packages/datasets/__init__.py", line 22, in <module>
    import pyarrow
  File "/opt/conda/lib/python3.6/site-packages/pyarrow/__init__.py", line 63, in <module>
    import pyarrow.lib as _lib
  File "pyarrow/compat.pxi", line 49, in init pyarrow.lib
  File "/opt/conda/lib/python3.6/site-packages/cloudpickle/__init__.py", line 4, in <module>
    from cloudpickle.cloudpickle import *  # noqa
  File "/opt/conda/lib/python3.6/site-packages/cloudpickle/cloudpickle.py", line 63, in <module>
    import typing_extensions as _typing_extensions
  File "/opt/conda/lib/python3.6/site-packages/typing_extensions.py", line 159, in <module>
    class _FinalForm(typing._SpecialForm, _root=True):
AttributeError: module 'typing' has no attribute '_SpecialForm'

This code block is called and then it calls ‘pretrain.py’ which is just a copy of the tutorial here`

import sagemaker
import boto3
from pathlib import Path
from sagemaker.huggingface import HuggingFace
from sagemaker.s3 import S3Downloader


sess = sagemaker.Session()
role = sagemaker.get_execution_role()


# hyperparameters which are passed to the training job

hyperparameters=dict(
    dataset_name="patrickvonplaten/librispeech_asr_dummy",
    dataset_config_names="clean",
    dataset_split_names="validation",
    validation_split_percentage="3",
    model_name_or_path="patrickvonplaten/wav2vec2-base-v2",
    output_dir="/opt/ml/model/artefacts",
    max_train_steps="1",
    num_warmup_steps="1",
    gradient_accumulation_steps="1",
    learning_rate="0.005",
    weight_decay="0.01",
    max_duration_in_seconds="20.0",
    min_duration_in_seconds="2.0",
    logging_steps="1",
    saving_steps="1",
    per_device_train_batch_size="8",
    per_device_eval_batch_size="8",
    adam_beta1="0.9",
    adam_beta2="0.98",
    adam_epsilon="1e-06"
)

# create the Estimator
huggingface_estimator = HuggingFace(
        entry_point='pretrain.py',
        source_dir='./',
        instance_type='ml.p2.xlarge',
        instance_count=1,
        role=role,
        transformers_version='4.4',
        pytorch_version='1.6',
        py_version='py36',
        hyperparameters = hyperparameters
)

huggingface_estimator.fit()
2

Hey @mfox,

the issue seems that it is using some unsupported > py36 typings. So you can simply solve it by updating the transformers_version, pytorch_version , py_version to the latest one, which is

    transformers_version = '4.17',            # the transformers version used in the training job
    pytorch_version      = '1.10',            # the pytorch_version version used in the training job
    py_version           = 'py38',            # the python version used in the training job

Also we have a nice documented table of currently supported versions here: Reference