search

AozhongZhang / COMQ

6 stars 0 forks source link

How to prepare dataset? #4

Open qifeng22 opened 4 months ago

qifeng22 commented 4 months ago

@AozhongZhang

qifeng22 commented 4 months ago

image I face this problem.

AozhongZhang commented 4 months ago

@qifeng22 Thanks for following our work. I recommend to use the dataset: ILSVRC2012. And make sure your data path is correct, which should be suited for the class Imagenet in datautils.py. I hope this screenshot will be help.

截屏2024-05-14 上午10 04 48
qifeng22 commented 4 months ago

okok，thanks

qifeng22 commented 4 months ago

image I test Full Float model, the accuracy is so low.

qifeng22 commented 4 months ago

@AozhongZhang

AozhongZhang commented 4 months ago

@qifeng22 I think this issue is related to Timm. Make sure that config, train_transforms and test_transforms in quantize_main.py match the model you use (vit_small_patch16_224). Data transform is very important in timm.

qifeng22 commented 4 months ago

ok, I will try check it.

qifeng22 commented 4 months ago

I run this command .

python quantize_main.py vit_small_patch16_224 --data_path --batchsize 1024 --wbits 4 --greedy

However, the result is always so bad, is there another way to find out problems where the model occurred? image

The following figure illustrates the content of the pretraining file. image

qifeng22 commented 4 months ago

@AozhongZhang

AozhongZhang commented 4 months ago

@qifeng22 I updated how to prepare the datasets, which will be helpful. Please see the prepare ImageNet datasets in README.md.

qifeng22 commented 4 months ago

OK， thank you