2 ms·
Very expensive. AFAIK the model can’t be quantized during backprop, so right there you’d need a ton of RAM. Backprop is faster bc it can be parallelized, but
by dartos 3y ago
Very expensive.
AFAIK the model can’t be quantized during backprop, so right there you’d need a ton of RAM.
Backprop is faster bc it can be parallelized, but IIRC you need to hold an entire copy of the model for each backprop process.
- scribu 3y agoActually, there have been attempts to do quantized backprop, but not sure how successfully.