Add_test passes valid transforms to DeviceDataLoader

fastai fastai dev
1

Hi!

I was wondering why my test images where getting resized while I was passing no transforms to add_test and I noticed that it was implemented like this:

    def add_test(self, items:Iterator, label:Any=None, tfms=None, tfm_y=None)->None:
        "Add the `items` as a test set. Pass along `label` otherwise label them with `EmptyLabel`."
        self.label_list.add_test(items, label=label, tfms=tfms, tfm_y=tfm_y)
        vdl = self.valid_dl
        dl = DataLoader(self.label_list.test, vdl.batch_size, shuffle=False, drop_last=False, num_workers=vdl.num_workers)
        self.test_dl = DeviceDataLoader(dl, vdl.device, vdl.tfms, vdl.collate_fn)

The DeviceDataLoader(dl, vdl.device, vdl.tfms, vdl.collate_fn) seems to explain my problem, therefore I was wondering if it was intentional that it passes vdl.tfms instead of tfms. If not, I can make a PR to correct this (that will require some very hard work as you can see).

2

vdl.tfms are the validation transforms on the validation dataloader (typically normalization). There is no resize there as you are already on the GPU and batched by that point.

3

Oh right, that makes sense! I’ll keep investigating as to why my test images are getting resized then.

4

Found it!

def new(self, x, y, tfms=None, tfm_y=None, **kwargs)->'LabelList':
    tfms,tfm_y = ifnone(tfms, self.tfms),ifnone(tfm_y, self.tfm_y)
    if isinstance(x, ItemList):
        return self.__class__(x, y, tfms=tfms, tfm_y=tfm_y, **self.tfmargs)
    else:
        return self.new(self.x.new(x, **kwargs), self.y.new(y, **kwargs), tfms=tfms, tfm_y=tfm_y).process()

As self.tfmargs is passed to the constructor, the argument size that I gave when creating the ItemList gets passed to the test list. After that, even though the list of transforms is empty, apply_tfms includes a resize part, meaning my test images get resized even when I don’t want to.
I’m not sure it would be beneficial to also pass custom tfmargs through add_test, as it is not very hard to patch when I want to (learner.data.label_list.test.tfmargs = None).

1 Like