ValueError: sampler should be an instance of torch.utils.data.Sampler,

[bbrattoli]

Dear vadimkantorov,

thank you for your publishing this nice repo, very well written.
I'm running
"python train.py --dataset cub2011 --model margin --base resnet50"
with pytorch 1.0.1 and pythorn 3.6 but it crushes with the error

Traceback (most recent call last):
File "train.py", line 71, in
loader_train = torch.utils.data.DataLoader(dataset_train, sampler = adapt_sampler(opts.batch, dataset_train, opts.sampler), num_workers = opts.threads, batch_size = opts.batch, drop_last = True, pin_memory = True)
File "/export/home/bbrattol/anaconda2/envs/pytorch/lib/python3.6/site-packages/torch/utils/data/dataloader.py", line 805, in init
batch_sampler = BatchSampler(sampler, batch_size, drop_last)
File "/export/home/bbrattol/anaconda2/envs/pytorch/lib/python3.6/site-packages/torch/utils/data/sampler.py", line 146, in init
.format(sampler))
ValueError: sampler should be an instance of torch.utils.data.Sampler, but got sampler=<main. object at 0x7f199b08a9b0>

I guess it has something to do with the new pytorch version. Could you help me to make it run correctly?

Thanks

[vadimkantorov]

Could you try to modify:

adapt_sampler = lambda batch, dataset, sampler, **kwargs: type('', (), dict(__len__ = dataset.__len__, __iter__ = lambda _: itertools.chain.from_iterable(sampler(batch, dataset, **kwargs))))()

to read instead:

adapt_sampler = lambda batch, dataset, sampler, **kwargs: type('', (torch.utils.data.Sampler,), dict(__len__ = dataset.__len__, __iter__ = lambda _: itertools.chain.from_iterable(sampler(batch, dataset, **kwargs))))()

?

Please let me know if it works and don't hesitate to send a PR.

[bbrattoli]

I get this error now

Traceback (most recent call last):
File "train.py", line 77, in
loader_train = torch.utils.data.DataLoader(dataset_train, sampler = adapt_sampler(opts.batch, dataset_train, opts.sampler), num_workers = opts.threads, batch_size = opts.batch, drop_last = True, pin_memory = True)
File "train.py", line 71, in
dict(len = dataset.len, iter = lambda _: itertools.chain.from_iterable(
TypeError: type.new() argument 2 must be tuple, not type

I know it's just a missing parameter but I don't understand what's happening in this piece of code, so please help! :D

[vadimkantorov]

Just checking, are you using (torch.utils.data.Sampler,) and not (torch.utils.data.Sampler)? (the difference is the comma, but it's important)

[bbrattoli]

It gives me the same error in both cases

[vadimkantorov]

Sorry @bbrattoli, don't have time to look at this in detail these days. I'll update here if I check what's going on. Meanwhile, the way to go is to define yourself a Sampler subclass instead of my hacky adapt_sampler dynamic class creation.