train

Could you please provide details about the computational resources used for training and for pre-extracting and caching the visual features? Specifically, how many GPUs were used, what model they were, and how long the training process took?