Huffman code and Bidmach
Opened this issue · 1 comments
arita37 commented
Hello,
For the DNN and large memory ML algo, wondering if you will intend to use Huffman code to reduce the memory footprint of the parameters (weights for DNN).
Also, wondering if you will use techniques like pruning and weight sharing by quantization
to reduce the memory usage ?
Thanks
jcanny commented
Probably not. We'd have to keep uncompressing them with each operation, which would undercut the value of using the GPU.