#Accelerating the DeViSE Objective in CUDA Sameer Lal, Prateek Goel
To compile:
make
For specific types (single image, mini-batch, or mini-batch chunks):
make <type>
To run:
./<type> <image_vectors> <held_out_image_vectors> <image_labels> <held_out_image_labels> <word_vectors>
The single image method computes the gradient with respect to the objective a single image at a time. Mini-batch moves multiple images at a time to the GPU for computation. Chunks moves multiple mini-batches onto the GPU for each kernel invocation.