Using fixed set of pos/neg pairs?

[zabzug-pfpt]

I see the sampler to generate a fixed set of triplets and the miner that uses that fixed set. Is there an analogous sampler/miner duo for generating and using a fixed set of positive and negative pairs?

Also, is there any relevant example code?

[KevinMusgrave]

That doesn't exist currently. But pull requests are welcome! 😄

[zabzug-pfpt]

@KevinMusgrave it looks like the EmbeddingsAlreadyPackagedAsTriplets miner will trivially fail for batch sizes that aren't multiples of 3:

        batch_size = embeddings.size(0)
        a = torch.arange(0, batch_size, 3)
        p = torch.arange(1, batch_size, 3)
        n = torch.arange(2, batch_size, 3)

From the BasePostGradientMiner's output assertion method:

            self.num_triplets = len(output[0])
            assert self.num_triplets == len(output[1]) == len(output[2])

For example, with a batch size of 32, we get len(a) == len(p) == 11 but len(n) == 10. Is that expected? Am I missing something?

[KevinMusgrave]

Yeah the batch size has to be a multiple of 3. Using the FixedSetOfTriplets sampler should make the dataloader return triplets arranged as [anchor1, positive1, negative1, anchor2, positive2, negative2, ...]. If the batch size is not a multiple of 3, then you'll be missing a part of a triplet.