Shuffling in ds.pytorch()
Understanding data shuffling in Deep Lake's pytorch dataloader
The Deep Lake shuffling algorithm is based upon a shuffle buffer that preloads a specified amount of data (in MB) determined by the
ds.pytorch(buffer_size = 2048). First, the dataloader randomly selects chunks of from the applicable tensors until the shuffle buffer is full. Next, the indices that are available in shuffle buffer are randomly sampled to construct the batches that are returned by the dataloader. As the data in the shuffle buffer is consumed, new chunks are randomly selected and added to the buffer.
As a consequence, if many chunks in the buffer contain data from the same class, which may occur if data was uploaded in non-random order, the shuffle buffer may contain fewer unique classes than if the samples were chosen fully randomly based on index. The most extreme case of reduced randomness occurs when datasets are much larger than the shuffle buffer, when they have many classes, and when those classes occur in sequence within the dataset indices.
One example dataset is Imagenet, which has 1000 classes, 1.2M images, 140GB of data, and approximately 140 images per 16MB chunk. For the case when the images are uploaded in sequence, the plot below shows how many unique classes are returned by the loader vs the number of images that have been returned in total. It is evident that randomly sampling by index returns many more unique values than the Deep Lake's dataloader.