Sampling from a PyTorch dataset

1

I would like to randomly sample from a PyTorch dataset and create a PyTorch dataset from the samples.
Currently, I am using data.random_split but I’m not sure if this is the best approach.
This is the code:

def generate_random_coreset(dataset, num_samples):
    # dataset is a pytorch dataset
    return data.random_split(dataset, [len(dataset) - num_samples, num_samples])

Does PyTorch offer a possible function to solve this?
I would appreciate any guidance and hits given on this question.

2

Could you please describe your use case in more detail? Also please checkout torch.utils.data.DataLoader which allows automatic batched loading of data possibly using parallel worker processes.

3

Thank you for the reply.
My use case is creating a coreset (or memory) from the task dataset in a continual learning setting. I would like to do the sampling before using the dataloaders.

4

You can use torch.utils.data.RandomSampler to create a Random Sampler for your Dataloader.
E.g.

random_sampler = data.RandomSampler(dataset, num_samples=num_samples)
dataloader = data.DataLoader(dataset, batch_size=k, sampler=random_sampler)

Note that use can pass a generator object to RandomSampler to get the same subset of random samples every time.

1 Like