Snapshot extension's autoloading raises AttributeError

[msakai]

trafficstars

If I run following script multiple times, it raises a AttributeError: 'ExtensionTrainerAdapter' object has no attribute 'torch_snapshot'.

import numpy as np
import torch
import torch.nn.functional as F
import ignite

from chainer.training import extensions
import chainer_pytorch_migration as cpm
import chainer_pytorch_migration.ignite

device = torch.device('cpu')

model = torch.nn.Linear(3,1).to(device)

X = torch.randn(10, 3)
y = torch.rand(10).to(torch.int64)
dataset = torch.utils.data.TensorDataset(X, y)
train_loader = torch.utils.data.DataLoader(dataset, shuffle=True, batch_size=2)

optimizer = torch.optim.Adam(model.parameters())
trainer = ignite.engine.create_supervised_trainer(
    model, optimizer, F.nll_loss, device=device)

optimizer.target = model
trainer.out = "result"
snapshot = extensions.snapshot(filename='snapshot_iteration-{.updater.iteration}', autoload=True)
cpm.ignite.add_trainer_extension(trainer, optimizer, snapshot, trigger=(1, 'iteration'))

trainer.run(train_loader, max_epochs=2)

$ python test.py 
$ python test.py 
Traceback (most recent call last):
  File "test.py", line 28, in <module>
    trainer.run(train_loader, max_epochs=2)
  File "/Users/sakai/.pyenv/versions/miniconda3-4.3.30/lib/python3.6/site-packages/chainer_pytorch_migration/ignite/extensions.py", line 293, in pre_run
    Engine.run(self.engine, data, max_epochs)
  File "/Users/sakai/.pyenv/versions/miniconda3-4.3.30/lib/python3.6/site-packages/ignite/engine/engine.py", line 446, in run
    self._handle_exception(e)
  File "/Users/sakai/.pyenv/versions/miniconda3-4.3.30/lib/python3.6/site-packages/ignite/engine/engine.py", line 410, in _handle_exception
    raise e
  File "/Users/sakai/.pyenv/versions/miniconda3-4.3.30/lib/python3.6/site-packages/ignite/engine/engine.py", line 429, in run
    self._fire_event(Events.STARTED)
  File "/Users/sakai/.pyenv/versions/miniconda3-4.3.30/lib/python3.6/site-packages/ignite/engine/engine.py", line 345, in _fire_event
    func(self, *(event_args + args), **kwargs)
  File "/Users/sakai/.pyenv/versions/miniconda3-4.3.30/lib/python3.6/site-packages/chainer_pytorch_migration/ignite/extensions.py", line 241, in set_training_started
    self.start_extensions()
  File "/Users/sakai/.pyenv/versions/miniconda3-4.3.30/lib/python3.6/site-packages/chainer_pytorch_migration/ignite/extensions.py", line 266, in start_extensions
    initializer(self)
  File "/Users/sakai/.pyenv/versions/miniconda3-4.3.30/lib/python3.6/site-packages/chainer/training/extensions/_snapshot.py", line 379, in initialize
    npz.load_npz(snapshot_file, target)
  File "/Users/sakai/.pyenv/versions/miniconda3-4.3.30/lib/python3.6/site-packages/chainer/serializers/npz.py", line 243, in load_npz
    d.load(obj)
  File "/Users/sakai/.pyenv/versions/miniconda3-4.3.30/lib/python3.6/site-packages/chainer/serializer.py", line 83, in load
    obj.serialize(self)
  File "/Users/sakai/.pyenv/versions/miniconda3-4.3.30/lib/python3.6/site-packages/chainer_pytorch_migration/ignite/extensions.py", line 307, in serialize
    state = torch.load(self.torch_snapshot)
AttributeError: 'ExtensionTrainerAdapter' object has no attribute 'torch_snapshot'

Maybe following description in Framework Migration GuideFramework Migration Guide implies that autoload is unsupported, but I create this issue just in case.

The correct way to restore a snapshot is by using cpm.ignite.load_chainer_snapshot(engine, optimizer, snapshot_path) with the Chainer snapshot path.

[emcastillo]

Exactly, autoload is currently not supported. We will update the migration guide to reflect this.

[msakai]

I see. Thanks!