Skip to content
This repository has been archived by the owner on Oct 1, 2020. It is now read-only.

Support for nn.ConvTranspose2d #72

Open
Amitdedhia6 opened this issue Aug 19, 2020 · 0 comments
Open

Support for nn.ConvTranspose2d #72

Amitdedhia6 opened this issue Aug 19, 2020 · 0 comments

Comments

@Amitdedhia6
Copy link

Hi - is there a plan to support nn.ConvTranspose2d for both QNNPACK and FBGEMM? I am using this in my model. After applying QNNPACK quantization techniques (both post training static and QAT), I am getting huge drop in accuracy, as well as not great improvement in speed. I suspect it could be due to ConvTranspose2d.

I am using Unet model for semantic segmentation. In the upsampling side, there are series of layers each having ConvTranspose2d operation. In the quantized model, it has to keep de-quantizing weights/activations to float for ConvTranspose2d layer and then back to quantized form for subsequent convolution/BN/Relu layers.

Sign up for free to subscribe to this conversation on GitHub. Already have an account? Sign in.
Labels
None yet
Projects
None yet
Development

No branches or pull requests

1 participant