Support PyTorch Dataset for oscar

[yuanchongbit]

What do these changes do?

I have done a demo of pytorch dataset API by inheriting torch.utils.data.Dataset. I also write some testcase in tests subfolder.

Related issue number#2224

I have try transfer a torch dataloader object to torch script but I have no idea how to do it. Now I am reading raySGD code to find a way to help mars to run torch or tf code more flexible.

[qinxuye]

Try to use mars dataset in run_pytorch_script. This requires passing Mars objects into script.

Before Mars 0.7, we have a concept called named_tensor and named_dataframe, that is, we can name a tensor or df via obj.execute(name='sth'), then in the script users can get it by providing a name like mt.named_tensor(name='sth').

But we don't like the concept because, the named tensor/df has no difference with the vanilla one except adding a name, but we have to add more APIs, and maybe for md.index or sth, we need named_index, that's a shame. Besides, named object bypassed the lifecycle mechanism in Mars, if an object executed with a name, we don't know when to delete it. However, for vanilla objects, if it's garbage collected, the reference count on the server side will minus 1, and deleted if reference count reaches 0.

In conclusion, we need a new way to pass Mars objects between driver and pytorch script.

I suggest to add the ability to run_script, which run_tensorflow_script and run_pytorch_script rely on, like,

import mars.dataframe as md
from mars.remote import run_script

df = md.read_xxx()
run_script('/to/path/script.py', inputs=[df])

In the script.

from mars import get_context

df = get_context().get_inputs()[0]

# other parts

The hard part is context can be got inside the script, and get_inputs() is absent for context.

Updated: see #2251.

[qinxuye]

#2251 has been merged to master branch, now could leverage it to support PyTorch dataset API.

[qinxuye]

I left some comments.

[qinxuye]

LGTM

[hekaisheng]

LGTM

[qinxuye]

Congrats that this PR got merged, look forward to seeing more contributions from you.