Yolov5s number of parameters

[andreysher]

Hello, and thanks for your work! I have a question about FLOPs for yolov5 models.
I use yolov5s from Ultralitics repo
on 640*640 images, and according to fvcore FLOPs analizer it have 8.2653184 GFLOPs, but in your table YOLOv5s have 13.7 GFLOPs. Which resolution, or version of model do you use in this table?

[NetsPresso-admin]

The NetsPresso Model Compressor measures the FLOPs using thop (Link) to measure the FLOPS.

In addition, according to the official YOLOv5 git repository, the FLOPs of the YOLOv5s seem to contain 16.5FLOPs (Link).

The difference between 13.7GFLOPs and the 16.5GFLOPs may be because of the activation function.
As thop cannot find the rule for the torch.nn.modules.activations.SiLU the FLOPs from the NetsPresso Model Compressor may smaller than the Actual FLOPs as YOLOv5s.

Thank you very much for your report.

[andreysher]

Hello again! I found that Ultralitics also use thop for their table and multiply it's result for some reason (link).
In this case your FLOPs seems strange, because of the same profiler and different results.

[NetsPresso-admin]

Sorry for the late reply.
Values from the official YOLOv5 git repository are for COCO pre-trained YOLOv5 models.
NetsPresso uses VOC pre-trained YOLOv5 models so the FLOPs can be different from the official repo.

[andreysher]

The official repo has config for Pascal VOC. I create a model for Pascal VOC training by the command:

python train.py --weights='' --cfg=models/yolov5s.yaml --imgsz=640 --data=data/VOC.yaml

I check the shape of this model's output. It is [1, 25200, 25]. There is 25 values in the last component: object confidence, bbox coordinates and 20 class confidences, so my model have 20 classes. The number of flops for this model, measured by thop calculator according to ultralitics FLOPS calculation procedure is 16.1 GFLOPS.

The code for FLOPS measurement:

from thop import profile
import torch
import argparse

if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument('--model-path', type=str, help='Path to model.pt')
    args = parser.parse_args()

    data = torch.ones((1, 3, 640, 640))
    model = torch.load(args.model_path, map_location='cpu')['model'].to(torch.float32)
    thop_macs, thop_params = profile(model, inputs=(data,))
    print(f"Number of FLOPS: {thop_macs * 2 * 1e-6}") # more information here: https://github.com/ultralytics/yolov5/issues/10042

But your model still have 13.7 GFLOPS. What is the difference between ultralitics yolov5s and your baseline? Can you please try to reproduce my latency calculations on your baseline model and provide the results to me?

[NetsPresso-admin]

Thank you for the clarification.
We recently found that the columns were written incorrectly.
Model Size and GFLOPS in the table is changed right now.
Please see the documentation, and thank you for your comments!