Open peterdudfield opened 1 year ago
Each worker thread is the multiprocessing though, and each example is emitted one by one by the datapipes before it, so not sure multiprocessing that code would speed it up at all. Adding more workers seems like probably the best solution to it, I would think.
Unless we want to break that each datapipe is independent of other ones, in which case we could maybe make a function or datapipe that loads multiple examples per step at once, and then gives them all to the next datapipe, still not sure that would be faster though
Instead of looping on the batch, could we use multi processing?
code is here