Omniglot Dataset

[activatedgeek]

This is the loader for Omniglot Dataset.

One of the use cases of this dataset is for One-Shot Learning where we sample a pair from the dataset and train a Neural Network to learn the similarity metric between the pair.

P.S.: It is amazing how simple it is to write data loaders!

[activatedgeek]

I am particularly concerned as to how to implement the __len__ function for the randomized pair. Should I generate a preset number of pairs and just return them on every call to __getitem__?

[activatedgeek]

I precompute the random set of pairs with an added argument of pair_count to the new OmniglotRandomPair class constructor. I think this should solve the random pair generation problem and make it deterministic after the object is instantiated.

[alykhantejani]

Thanks for the PR - I've left some inline comments

[alykhantejani]

Thanks for the updates, I've left some more comments inline, mostly around the OmniglotRandomPair dataset

[activatedgeek]

Hey @alykhantejani, do you mind checking some updates to the randomized pair generation?

[activatedgeek]

@alykhantejani Writing the comments on the approach here because I couldn't find it anywhere when I had responded to your review earlier (weirdly)

If all you want are random pairs of images and whether they are the same character or not, then why not just sample on (img, idx) pair from self._flattened_character_images and then sample another continuously until they either match or not depending on what is_match is.

I believe this way, it is certainly harder to arrive at a collision when is_match=1. The way I have maintained the internal data helps me quickly arrive at a pair for both is_match = {0,1} because once I flatten the array, I lose information about the character class.

There's no need to precompute pairs, just make __len__ return then length of self._flattened_character_images and then when this is used with the pytorch DataLoader, it will stop sampling after __len__ samples.

The problem with this approach would be that I am not staying true to what I wanted to achieve with the class. The total number of pairs combinatorially possible are huge and that is why I introduced a pair_count defaulting to 10000 and then set that as the length. It gives the user enough flexibility to sample training pairs as well as setup a One-Shot task. If I limit it to the length of the flattened character list, it gives out the wrong perception and might confuse the user. I have also added random_seed to allow reproducibility.

[alykhantejani]

Hey @activatedgeek, sorry for the response.

Pinging @fmassa in case he has any opinions/thoughts on the random pair dataset

[fmassa]

Hi @activatedgeek , sorry for the delay in replying.

So, my first thought about the Pair dataset is the following: we should try not to add sampling logic into the dataset itself, which should only hold information on how to load individual elements, and let the sampling be performed by the Sampler/DataLoader.
It might look non-trivial in this case, but here is one possibility:

We introduce a new Dataset class, in a similar spirit of ConcatDataset, that performs n samples of the data and returns whatever you want?
Here is a draft implementation (untested):

class MultiDataset(object):
    def __init__(self, dataset, num_outputs=1, transforms=None):
        self.dataset = dataset
        self.num_outputs = num_outputs
        self.transforms = transforms

    def __getitem__(self, idx):
        # here comes the logic to convert a 1d index into a 
        # self.num_output indices, each of size len(self.dataset)
        individual_idx = []
        for i in range(self.num_outputs):
            individual_idx.append(idx % len(self.dataset))
            idx = idx // len(self.dataset)
        
        result = []
        for i in reversed(idx):
            result.append(self.dataset[i])

        if self.transforms is not None:
            result = self.transforms(result)

        return result    

    def __len__(self):
        return len(self.dataset) ** self.num_outputs

This way, you generate on-the-fly an arbitrarily large dataset, that can accommodate pairs/triplets/etc of elements of the same dataset. Plus, the logic of how to combine the different targets of each dataset becomes something that the user should do (via the transforms in the MultiDataset, each individual dataset has its own transforms as well).
And, if you want a biased sampler (say to have a roughly equal number of matches / non-matches) this could be handled by the sampler maybe?

This is just a rough idea, but let me know what you think.

[activatedgeek]

@fmassa That is a great idea. I was in fact wondering the same because I recently came across an obvious requirement of similar kind for the ImageNet/Mini-ImageNet datasets as well. It didn't feel right to create custom rules every now and then.

So here is what I will do - the Omniglot data loader will return a tuple (image, character_class_id) and I will wrap up this PR.

I will take up a MultiDataset like utility wrapper in another PR.

Does that sound good?

[fmassa]

Sounds good! Also, giving how generic this dataset is, might be worth considering sending it to pytorch/tnt, but we can discuss that later

[activatedgeek]

@fmassa @alykhantejani Can you please verify if everything is in order now?

[activatedgeek]

Any updates here?

cc @fmassa @alykhantejani

[activatedgeek]

Hi, I am wondering what the hold up is about? Is there something that I could further help with? It has been quite a while that this PR has been up and I would really appreciate if we could make some progress here. (Or closed if it doesn't align well).

cc @alykhantejani @fmassa

[activatedgeek]

@alykhantejani @fmassa Any possibility of this getting merged?

[fmassa]

@activatedgeek sorry for the delay, I'll have a look at it today.

[fmassa]

This looks very good, thanks! Once the merge conflicts are addressed, I think this can be merged.

[activatedgeek]

Hey @alykhantejani @fmassa , can we merge this? I can continue on a discussion on #338 after this.

[activatedgeek]

Pinging @fmassa and @alykhantejani here again. Can we please merge this?

[fmassa]

Awesome! Thanks a lot @activatedgeek !

[activatedgeek]

Hey @fmassa thanks! Do you mind taking a look at #338 as well?