about quantize

[shenshen0318]

In quantize.py, I find QConv2d, QLinear and QConvTranspose2d ( I guess they are quantized Conv2d, Linear and ConvTranspose2d). But in model_infe.py, I just find quantize's value is True, but i don't know how to construct a model using them. Furthermore, i want to build a Dis contains QConv2d(model is a list,like[nn.ReflectionPad2d(padding), QConv2d, nn.LeakyReLU(inplace=True)]), my forward function just write return self.model(x) is enough?

[tilmto]

If you set C.quantize = True in config_train.py, the model will be trained with default 8 bit. You can change the default bit-width of the class QConv2d in quantize.py. I think the self.model(x)is enough for your case.

[shenshen0318]

Thank you for your reply！！！Another one question, the file of operations.py is your custom class(like Conv, ConvTranspose2dNorm, ConvNorm), they are quantized and used to build NAS_GAN?

[tilmto]

Yes, they are the basic searchable blocks to build the supernet and will be used to construct the final derived architecture. But they will be quantized only if you set C.quantize = True in config_train.py, otherwise they will be full precision ones.

[shenshen0318]

Thanks！I have no question at present. You could close this issue.