I-ViT

I-ViT copied to clipboard

您好你可以提供一下以及训练好的模型，供下载吗，DeiT-T

HuPengsheet

Hi @zkkli, many thanks for your work, it is a quite nice contribution to the state-of-the-art. After training a model a model by running: `python quant_train.py --model deit_tiny --data --epochs...

fabriziojpiva

请问TVM的版本是否实现正确？

1

我使用的是deit_small模型，精度fine-tune至79.1%，  但是预测提供的cat.jpg的标签为  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？

zjnyly

您好，我这里所有图片的输出为什么都是这样的

5


husnxjsjsb

Issues on evaluating latency using TVM.

1

Hi I’m currently working on compiling I-ViT using TVM. On this project, The error appears. > Check failed: value < 1LL

rkdgmlqja

非常棒的工作，我比较好奇，纯int量化的优势在于速度，但是好像没有底层kernel的支持，还是以全精度（TVM）的方式去计算的，这样int量化的实际价值没有发挥出来，看论文中的数据实际latency没有较FasterTransformer提升太多。

XA23i

latency result slower than tensorrt fp16

3

Hi, I tried to replicate your speed experiment, I tested the deit_tiny, batch size=1, RTX3090 environment, after a few days of autotune, compared to tensorrt FP16, speed is still slower....

zhanglei1172

yanzhengmao

I thought I-ViT works based on PTQ when I read the paper but your code and readme shows that it's based on QAT. If I have the quantized int8 weight...

dedoogong

Hello, I am JIHO LEE. I thoroughly enjoyed reading your paper and was deeply impressed by it. I believe that the ultimate goal of this research is to improve efficiency...

jiho264