Not a priority, but batch transforms (parallelized on GPU) will make training marginally faster. Currently, I let the DataLoader in Torch take care of it example by example, using its own optimizations like loading the batch while updating weights. Relevant issue here: https://github.com/pytorch/vision/issues/157
Not a priority, but batch transforms (parallelized on GPU) will make training marginally faster. Currently, I let the DataLoader in Torch take care of it example by example, using its own optimizations like loading the batch while updating weights. Relevant issue here: https://github.com/pytorch/vision/issues/157