Closed GraphGrailAi closed 5 years ago
Unfortunately, we're not able to release pretrained models for this purpose. Hmm - what kind of time-to-response are you concerned with? Yeah, pretraining took us a few days and fine-tuning took around an hour.
time-to-response - is time to process one sentence to get result with angry or happy class. Also, what is the final size of model?
For the 4-layer retrieval Transformer model, I measure typical forward-pass times of 0.1 sec (for a batch size of 512) and a model size of ~350 MB.
Can we download pretrained models to use them for inference? (Also what is time-to-response for this?) As i see in paper, it takes 2-3 days to train myself