Model Quantization

[pepsbn (Peyman)]

Hi all, is there any official implementation/model for supporting model quantization in Pytorch? I mean sth similar to that of TF (https://www.tensorflow.org/performance/quantization)

Cheers
-P

[Felix_Lessange (Damien MENIGAUX)]

Shouldn’t this be possible with 1.0 ? I mean, it seems it was advertised like this. Can’t find it though.

[justusschock]

This should be possible with glow (haven’t tested it yet)

[dingyongchao]

@Felix_Lessange

though

I want training quantification such as TF as well! Waiting!

[Felix_Lessange (Damien MENIGAUX)]

I’m still very interested in this. Current state isn’t workable (and doesn’t support all operators).

Has pytorch dropped the idea of allowing speed optimization inside the framework ?