VQA MED add-ons

[tkornuta-ibm]

Added two flags to problem configuration:

• stream_images (DEFAULT: True) Flag indicating whether the problem will load and return images

• preload_images (DEFAULT: False) Flag indicating whether images will be preloaded (i.e. loaded once at start)

stream_images gives huge boost when working with question/answer pipelines.

WARNING: if preload_images option is active, the images will also be "preprocessed" at start. This means that preloading should not be used when one needs to use the random augmentations!

Using the preload_images option takes some time at the beginning, but during training/batch processing it is comparable (when it comes to time) to using DataLoader with 4 workers, with the advantage is lower utilization of CPU(s).

Some results on mimic_lstm_vgg16_ewm_is_cat_ffn_c1_loss (C1 problem, batch_size = 200, 2 GPUs):

• preload_images: True, workers: 0, (avg) single epoch time: 33(s)
• preload_images: False, workers: 0, (avg) single epoch time: 50(s)
• preload_images: False, workers: 1, (avg) single epoch time: 37(s)
• preload_images: False, workers: 4, (avg) single epoch time: 34(s)