Recursive copying of attribute dictionaries for TensorImage subclass

[restlessronin]

I ran into a problem with batch_to_sample when I was trying to apply it to multi-spectral images. I simplified the problematic code to the following

import torch
import fastai
from fastai.vision.all import *

class _T(TensorImage):
    pass

t = _T([[1.,2.],[3.,4.],[5.,6.]], kw=1)

test_eq(t.__dict__,{'kw': 1})

l = torch.unbind(t)

Basically I was expecting this to be true

test_eq(l[0].__dict__,{'kw': 1})

Instead I was getting

test_eq(l[0].__dict__,{})

If what I was expecting is the desired behaviour, then this branch contains my proposed fix. I'm a relative newbie to both python and fastai, so apologies if I'm missing something obvious.

I was also unable to get nbdev_test running properly on the fastai repo on my mac (lots of seemingly spurious test failures). Presumably the CI will run as part of this PR.

I would like to add my appreciation and admiration for Jeremy and the other incredible people who have helped create this community and movement. Well done!!!

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[restlessronin]

Modified the code to cover the edge case

t1 = _T([[1.,2.]], kw=1)
test_eq(t1.__dict__,{'kw': 1})
l1 = torch.unbind(t1)
test_eq(l1[0].__dict__,{'kw': 1})

[jph00]

Nice PR - thanks!