Add support for multi sample item in optimize and yielding from the _getitem_ of the StreamingDataset

[tchaton]

🚀 Feature

Motivation

It would be great to be able to create a batch of sub sample from a given sample. Right now, you can't do that.

However a user could support this.

def optimize(...):

    sample = 
	return MultiSample(sample, num_samples=X)

Under the hood, we know this sample could be used to generate multiple random samples.

class MyStreamingDataset(StreamingDataset):

	def __getitem__(self, index, sample_id):
			sample  = super().__getitem__(index)
            
            # do some transformation
            return  data

A use case would be image detection where each image can be used to generate multiple sub boxes and we might want to have them as different training samples.

Pitch

Alternatives

Additional context