How to Get a pretrained model to do PTQ?

[youdutaidi]

Thank you for sharing your PTQ code but I test Deit-Small with min-max quant technic, top@1 accuracy in validation dataset is only 0.11, can not match the papers 75.05%，is there anything wrong?

[linyang-zhh]

Hi! Our models are loaded with the official open source checkpoints, and it can be found in link1 and link2. You can check whether the pretrained models are loaded correctly.

[youdutaidi]

Hi! Our models are loaded with the official open source checkpoints, and it can be found in link1 and link2. You can check whether the pretrained models are loaded correctly.

Yes, I loaded the pre-trained model in link and test it ,but the validation result performs poorly, I print the pred and target lable (where the original code is :

accuracy(output, target, topk=(1,)):
    maxk = max(topk)
    batch_size = target.size(0)
    _, pred = output.topk(maxk, 1, True, True)
    pred = pred.t()
    print("pred: ",pred)
    print("target: ",target.reshape(1, -1).expand_as(pred))
    correct = pred.eq(target.reshape(1, -1).expand_as(pred))
    res = []
    for k in topk:
        correct_k = correct[:k].reshape(-1).float().sum(0)
        res.append(correct_k.mul_(100.0 / batch_size))
    return res

And the printed results are like this ：

pred: tensor([[171, 819, 672, 30, 688, 549, 549, 591, 591, 145, 591, 622, 591, 745, 688, 557, 688, 591, 978, 591, 492, 591, 492, 622, 549, 549, 688, 591, 492, 492, 622, 688, 688, 688, 591, 492, 152, 171, 557, 688, 145, 616, 978, 978, 591, 591, 591, 591, 492, 622, 591, 622, 978, 745, 492, 591, 591, 492, 591, 336, 622, 846, 492, 740, 336, 492, 745, 492, 336, 688, 688, 30, 928, 978, 30, 624, 152, 591, 727, 622, 688, 549, 692, 492, 928, 622, 624, 492, 549, 756, 492, 622, 853, 688, 492, 591, 156, 727, 745, 688], [622, 150, 591, 688, 907, 941, 745, 152, 740, 557, 740, 171, 740, 492, 513, 745, 591, 941, 427, 549, 169, 152, 745, 171, 745, 492, 672, 688, 622, 622, 492, 740, 740, 740, 740, 622, 591, 622, 571, 591, 622, 614, 853, 744, 672, 798, 152, 783, 622, 492, 549, 145, 728, 370, 145, 152, 152, 745, 744, 624, 492, 549, 549, 591, 624, 745, 549, 622, 624, 30, 907, 688, 688, 819, 688, 169, 331, 30, 624, 492, 30, 591, 744, 923, 911, 492, 336, 557, 745, 273, 622, 171, 978, 692, 273, 740, 461, 336, 61, 740], [196, 336, 30, 591, 928, 798, 657, 549, 672, 745, 692, 492, 672, 663, 672, 492, 740, 978, 672, 152, 387, 928, 549, 196, 657, 745, 30, 740, 923, 727, 145, 309, 591, 30, 783, 145, 692, 492, 745, 740, 492, 269, 941, 819, 740, 152, 549, 346, 923, 145, 740, 196, 672, 61, 336, 740, 187, 557, 672, 387, 557, 609, 387, 688, 169, 557, 464, 923, 907, 740, 882, 513, 911, 624, 513, 196, 675, 740, 614, 196, 740, 798, 549, 387, 291, 557, 907, 571, 387, 111, 557, 530, 728, 907, 418, 152, 170, 492, 370, 455], [492, 978, 874, 513, 740, 591, 492, 798, 337, 492, 152, 196, 688, 464, 874, 418, 30, 187, 707, 740, 846, 549, 387, 492, 387, 387, 874, 152, 145, 196, 196, 672, 30, 591, 152, 196, 662, 145, 145, 672, 557, 203, 591, 218, 337, 928, 928, 672, 846, 557, 152, 171, 688, 273, 557, 928, 692, 61, 218, 853, 196, 387, 27, 30, 907, 528, 663, 452, 513, 591, 797, 907, 246, 150, 591, 259, 591, 688, 702, 145, 874, 152, 381, 622, 688, 145, 513, 622, 492, 928, 418, 203, 688, 152, 448, 549, 530, 923, 273, 337], [978, 387, 978, 874, 591, 745, 692, 740, 30, 571, 898, 978, 152, 549, 728, 622, 289, 629, 153, 978, 549, 740, 923, 145, 635, 635, 337, 30, 387, 495, 495, 106, 672, 672, 346, 171, 675, 196, 492, 30, 875, 336, 549, 672, 30, 549, 798, 619, 387, 196, 783, 492, 513, 657, 622, 688, 740, 273, 819, 367, 145, 306, 325, 672, 622, 61, 61, 196, 853, 672, 600, 728, 907, 387, 129, 609, 372, 672, 907, 557, 513, 692, 306, 336, 692, 196, 387, 145, 635, 370, 923, 978, 30, 591, 370, 692, 962, 387, 464, 106]], device='cuda:0') target: tensor([[14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15], [14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15], [14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15], [14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15], [14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15]], device='cuda:0')

[linyang-zhh]

@youdutaidi
Hi, sorry for that, and can u provide your command and the code (if the code were modified). I will check it.

[youdutaidi]

@youdutaidi Hi, sorry for that, and can u provide your command and the code (if the code were modified). I will check it.

Thank you very much for your friendly help.
I didn't modify the original code ,and the command is :

CUDA_VISIBLE_DEVICES=1 python test_quant.py deit_small /home/dataset/cailingling/ImageNet/ --quant --ptf --lis --quant-method minmax

result is ：

[PeiqinSun]

Please first verify your dataset and float model is right?

1. You can use command:
   CUDA_VISIBLE_DEVICES=1 python test_quant.py deit_small /home/dataset/cailingling/ImageNet/