Flair: RuntimeError: storage has wrong size when loading en-sentiment

[rookie32]

When trying to load the en-sentiment model, I get the error:

RuntimeError: storage has wrong size: expected -1862414276 got 22700

I am able to load other models such as de-offensive-language and the ner SequenceTagger.

Code:

from flair.models import TextClassifier
from flair.data import Sentence
classifier = TextClassifier.load('en-sentiment')

Possibly related to issue https://github.com/zalandoresearch/flair/issues/172 ?
But the fix for that was merged back in November? https://github.com/zalandoresearch/flair/pull/194

Have tried deleting and redownloading the model.

Platform: Windows 10, Torch 1.0 (cpu), Python 3.6.4 (Anaconda)

[alanakbik]

Hi @rookie32 - yes it is the same bug as reported in #194, namely that very large models (over 2GB) cannot be de-serialized in certain windows and mac setups. This is a known issue in the pickle library which is used by pytorch to serialize.

We currently have a workaround in place in which we make sure that models do not go over 2GB in size, but overlooked this with the en-sentiment model. We will fix this for the next release!

[alanakbik]

Here also the link to the corresponding pickle issue.

[pdsing]

I am facing the same issue on my windows machine (windows 10, 8 GB RAM). Do we have a fix for this issue?

[alanakbik]

Yes, we are working on this at the moment. We will push an updated version with the upcoming 0.4.1 release, within the next weeks.

[highway11git]

It's pretty weird to me that this hasn't been solved, as the discussion about this bug in pickle was reported in 2015.

Anyways, I created a temporary workaround.

Disclaimer: this will use a big amount of RAM. If there's a more memory efficient solution that doesn't change the way Torch serializes the models (i.e. doesn't change serialization.py), let us know

Step 1:
Create a text_classifier.py file, which should be a copy of flair/models/text_classification_model.py

Step 2:
Replace the _load_state function with the following two functions:
https://gist.github.com/highway11git/1fa64621ea1b04e6494bf6672a750b24

Step 3:
Instead of using from flair.models import TextClassifier in your project, import the TextClassifier class from your newly created file. Like so: from text_classifier import TextClassifier

[alanakbik]

Hi @highway11git thanks very much for this! We'll test it and see if we can directly integrate it. Perhaps we could add a special flag to the model loader to allow users to specify which method they want to use to load a model, with the default being the current one.

[highway11git]

I've updated the gist:

1. load_big_file now uses mmap instead of BytesIO, which makes it _much more memory efficient_
2. _load_state now also uses the load_big_file function if the user doesn't use cuda

Thanks for testing it, and keep us up to date on any progress.

[alanakbik]

Cool, thanks! We'll test the solution and let you know!

[alanakbik]

@highway11git @pdsing @rookie32 I've added the fix proposed by @highway11git into a pull request.

Since we could not reproduce the error on our setups, could you check if this fixed it for you?

[dcavar]

Using the newest Flair and torch on Win 10 I still get this error for (other models worked):
tagger = SequenceTagger.load('frame-fast')
or
tagger = SequenceTagger.load('frame')

---

NotImplementedError Traceback (most recent call last)
in
----> 1 tagger = SequenceTagger.load('frame-fast')
2
3 sentence_1 = Sentence('George returned to Berlin to return his hat .')
4 sentence_2 = Sentence('He had a look at different hats .')
5

c:\program files\python37\lib\site-packages\flair\nn.py in load(cls, model)
84 # see https://github.com/zalandoresearch/flair/issues/351
85 f = file_utils.load_big_file(str(model_file))
---> 86 state = torch.load(f, map_location=flair.device)
87
88 model = cls._init_model_with_state_dict(state)

c:\program files\python37\lib\site-packages\torchserialization.py in load(f, map_location, pickle_module, *pickle_load_args)
527 with _open_zipfile_reader(f) as opened_zipfile:
528 return _load(opened_zipfile, map_location, pickle_module, *pickle_load_args)
--> 529 return _legacy_load(opened_file, map_location, pickle_module, **pickle_load_args)
530
531

c:\program files\python37\lib\site-packages\torchserialization.py in _legacy_load(f, map_location, pickle_module, *pickle_load_args)
700 unpickler = pickle_module.Unpickler(f, *pickle_load_args)
701 unpickler.persistent_load = persistent_load
--> 702 result = unpickler.load()
703
704 deserialized_storage_keys = pickle_module.load(f, **pickle_load_args)

c:\program files\python37\lib\pathlib.py in __new__(cls, args, *kwargs)
1023 if not self._flavour.is_supported:
1024 raise NotImplementedError("cannot instantiate %r on your system"
-> 1025 % (cls.__name__,))
1026 self._init()
1027 return self

NotImplementedError: cannot instantiate 'PosixPath' on your system