zkkli / I-ViT

[ICCV 2023] I-ViT: Integer-only Quantization for Efficient Vision Transformer Inference
Apache License 2.0
155 stars 15 forks source link

请问TVM的版本是否实现正确? #6

Open zjnyly opened 10 months ago

zjnyly commented 10 months ago

我使用的是deit_small模型,精度fine-tune至79.1%, image

但是预测提供的cat.jpg的标签为 94e270515ebd33ea0715891c07f575b

我经过比对,发现这些标签都对不上cat这类,请问是为什么呢?

dedoogong commented 3 months ago

I also faced same problem. it's not TVM version issue. it's torch issue. if you run validation again with the checkpoint.pth.tar, you will get Top1 0 accuracy!!