zkkli / repq-vit Goto Github PK

Hello! In your four detection experiments, I noticed a significant difference between the reported result and the test result for mask_rcnn_swin_small. The reported result is 44.2 mAP, but the tested result is 42.8 mAP. However, the results for the other three detection experiments match the reported results. I was wondering if there are any specific settings for mask_rcnn_swin_small, such as only quantizing the backbone?

Thank you for your contributions in the field of model quantification and generously sharing your code.

I encountered a problem where I ran a quantitative example and obtained the corresponding top 1 and top 5 data using the same W and A, which were the same as the author's results, but the q I saved_ Model. pth, open using Netron and see that it is still in float32 format. How can I save the quantization model correctly.

The code I use to save the results is:

Save. pth format:

torch. save (q_model, 'my/save/path')

Save onnx format:

dummy_ Input=torch. randn ((1, 3, 224, 224)). to (device)
torch. onnx. export (q_model, dummy input, onnx_path, verbose=False, inputnames=['input '], outputnames=['output'], opset_version=11)

The pytorch version I am using is:
torch==1.11.0+cu113
torchvision==0.12.0+cu113
timm==0.4.12

It seems that, after the log sqrt2 quantization, the activations have different scale parameters at different positions. So, in actual hardware, would it be impossible to perform subsequent matrix multiplications using integer form?

Hi, I notice in Table 1. there is "No REC" in column 3. And I wonder what it means and does it affects inference speed a lot. Thank you.

When I try to run the code, there is a problem in line 117 of test_quant.py "next_module = father_module.reduction". my shell is "python test_quant.py --model deit_small --dataset /root/data/Imagenet". Is there any problem with my shell?

I am trying to run your code, but I cannot install the packages based on the instructions provided. Can you give the latest instructions for running your code?

Hi~ the concept of channel-wise in Conv is quite intuitive, but how to understand channel in tensor with shape [batch_size, num_seq, num_hidden] ? weather the channel refers to the num_seq dim or not?