OP set 10 limitation ? #872
-
Hello, I am attempting to quantize https://github.com/gengshan-y/high-res-stereo . So far I have a working PyTorch model, ONNX model, and IR model (after patching ModelOptimizer), all FP32, and all of them have the same accuracy (avg error ~4.7), as expected. Since the tests with OpenVino POT gave me 2x gain in inference speed but my accuracy was severely affected ( from avg error 4.7 I got to 15 with AccuracyAwareQuantization and to 20-22 with DefaultQuantization, about 4-5 times worse), I wanted to give it a go with NNCF, since I tested this framework in the past, at version 1.6 and had good results on a different model. The reason I need OP set 11 is due to "index_put", at https://github.com/gengshan-y/high-res-stereo/blob/master/models/hsm.py , line 41: def feature_vol(self, refimg_fea, targetimg_fea,maxdisp, leftview=True) Thanks, |
Beta Was this translation helpful? Give feedback.
Replies: 2 comments 2 replies
-
Hi Cristi, Thanks for your interest in the NNCF. If you have troubles with export to opset 11 please report an issue for that. NNCF should support PyTorch 1.8 which supports opset 11 as I can see. |
Beta Was this translation helpful? Give feedback.
-
Sounds good to me. Let's proceed with an issue. |
Beta Was this translation helpful? Give feedback.
Sounds good to me. Let's proceed with an issue.