Ignite: DCGAN Example pickle exception

[Daiver]

First of all i'm sorry for my poor English
I just tried to run dcgan example and faced with following exception:

Traceback (most recent call last):
  File "C:/Users/Daiver/Documents/Depth/ignite_examples/dcgan_orig.py", line 438, in <module>
    alpha=0.98, output_dir=output_dir)
  File "C:/Users/Daiver/Documents/Depth/ignite_examples/dcgan_orig.py", line 417, in main
    trainer.run(loader, epochs)
  File "C:\Users\Daiver\Anaconda3\lib\site-packages\ignite\engine\engine.py", line 326, in run
    self._handle_exception(e)
  File "C:\Users\Daiver\Anaconda3\lib\site-packages\ignite\engine\engine.py", line 289, in _handle_exception
    self._fire_event(Events.EXCEPTION_RAISED, e)
  File "C:\Users\Daiver\Anaconda3\lib\site-packages\ignite\engine\engine.py", line 226, in _fire_event
    func(self, *(event_args + args), **kwargs)
  File "C:/Users/Daiver/Documents/Depth/ignite_examples/dcgan_orig.py", line 414, in handle_exception
    raise e
  File "C:\Users\Daiver\Anaconda3\lib\site-packages\ignite\engine\engine.py", line 313, in run
    hours, mins, secs = self._run_once_on_dataset()
  File "C:\Users\Daiver\Anaconda3\lib\site-packages\ignite\engine\engine.py", line 280, in _run_once_on_dataset
    self._handle_exception(e)
  File "C:\Users\Daiver\Anaconda3\lib\site-packages\ignite\engine\engine.py", line 289, in _handle_exception
    self._fire_event(Events.EXCEPTION_RAISED, e)
  File "C:\Users\Daiver\Anaconda3\lib\site-packages\ignite\engine\engine.py", line 226, in _fire_event
    func(self, *(event_args + args), **kwargs)
  File "C:/Users/Daiver/Documents/Depth/ignite_examples/dcgan_orig.py", line 414, in handle_exception
    raise e
  File "C:\Users\Daiver\Anaconda3\lib\site-packages\ignite\engine\engine.py", line 268, in _run_once_on_dataset
    for batch in self.state.dataloader:
  File "C:\Users\Daiver\Anaconda3\lib\site-packages\torch\utils\data\dataloader.py", line 819, in __iter__
    return _DataLoaderIter(self)
  File "C:\Users\Daiver\Anaconda3\lib\site-packages\torch\utils\data\dataloader.py", line 560, in __init__
    w.start()
  File "C:\Users\Daiver\Anaconda3\lib\multiprocessing\process.py", line 105, in start
    self._popen = self._Popen(self)
  File "C:\Users\Daiver\Anaconda3\lib\multiprocessing\context.py", line 223, in _Popen
    return _default_context.get_context().Process._Popen(process_obj)
  File "C:\Users\Daiver\Anaconda3\lib\multiprocessing\context.py", line 322, in _Popen
    return Popen(process_obj)
  File "C:\Users\Daiver\Anaconda3\lib\multiprocessing\popen_spawn_win32.py", line 65, in __init__
    reduction.dump(process_obj, to_child)
  File "C:\Users\Daiver\Anaconda3\lib\multiprocessing\reduction.py", line 60, in dump
    ForkingPickler(file, protocol).dump(obj)
AttributeError: Can't pickle local object 'check_dataset.<locals>.<lambda>'

Process finished with exit code 1

I'm pretty sure this is due this line to_rgb = transforms.Lambda(lambda img: img.convert('RGB')). (i commented this and example failed with channels mismatch error)
Unfortunately i'm not very good with Python/Torch/Ignite so i have no idea why this example attempts to pickle lambda.

This is exactly what i run: https://gist.github.com/Daiver/cfaa55410d3d1f9082197e4113bf7328 (just dcgan example with hardcoded options)

My Python/packages versions

torch version: 1.0.0
torchvision version: 0.2.1
ignite version:  0.1.2
Python version: 3.6.5 |Anaconda, Inc.| (default, Mar 29 2018, 13:32:41) [MSC v.1900 64 bit (AMD64)] (Windows)

Please, tell me if i can provide more data.

[vfdev-5]

@Daiver thanks for the feedback! Maybe the problem you have is related to your OS (Windows).
You can try to replace the lambda by defining a function:

def to_rgb(img):
    return img.convert('RGB')

def check_dataset(dataset, dataroot):
    # to_rgb = transforms.Lambda(lambda img: img.convert('RGB'))
    # Probably transforms.Lambda is not necessary
    ...
    dataset = dset.MNIST(root=dataroot, download=True, transform=transforms.Compose([to_rgb,
                                                                                         resize,
                                                                                         to_tensor, normalize])

Feel free to ask more if need more support on this.

[Daiver]

@vfdev-5 thank you for answer!
Replacing lambda by function helps me, at least training starts. Should i close the issue?

[vfdev-5]

@Daiver glad that it helps !

Let's leave it open. I think we need to modify DCGAN example such that we can run it on any platform: Win, LInux, OSX.

And if you would like to contribute, PRs are welcome :)
See contributing for more info.

[anmolsjoshi]

@vfdev-5 looking further into this issue, there might be three solutions below.

• num_workers=0 seems to solve the issue, but will result in a slower process
• align dcgan example closer to the pytorch example, not forcing Discriminator input_channels as 3, but dependent on the dataset. In the official pytorch example
• setup to_rgb as you have shown above.

thoughts?

[vfdev-5]

@anmolsjoshi the 2nd choice looks good. However I would like to ensure that this really solves the problem...

[anmolsjoshi]

@vfdev-5 just wanted to clarify, aligning it closer to the Pytorch example would also remove to_rgb that is used in MNIST. I'll try tonight and report back.

[vfdev-5]

@anmolsjoshi yes, you can introduce a parameter as nc:

https://github.com/pytorch/examples/blob/0d3fe14a1c5a00795e3671ea3473caef6f0da72d/dcgan/main.py#L82-L89

I'll try tonight and report back.

Sounds good