Mohammad Khoshbin comments

Results 12 comments of


                                            Mohammad Khoshbin

How to improve the performance (mAP) of YOLOV7 on Custom Dataset?

It's kind of try and error. You can train with more data, change input size, add augmentation or change hyper parameters and train the model and see if **map** changes...

Inference time based on Hardware (GPU)

I had same experience with NVIDIA 1080 and 3070. On NVIDIA 3070, YOLOv7 is faster than YOLOv5 but on NVIDIA 1080 it's the opposite.

Small performance gain using TRT model

It is probably because of your GPU model. I tested yolov7 on 1080 and 3070. Inference time on 3070 is much less than 1080.

CUDA out of memory for yolo-pose

There is a line in `non_max_suppression_kpt` function that causes this error. I don't know why this happen. https://github.com/WongKinYiu/yolov7/blob/72a9c9628a22693a0c3a0eeb44c3be199d6cb1f1/utils/general.py#L742-L743 This line is used for computing conf for every object in multi-class...

“save_period” doesn't work?

There is no functionality to save checkpoint after every `save_period` epoch. The `save-period` option is only used for wandb log. I added PR #437 for this issue.

yolov7 inference slower than yolov5

I tested YOLOv7 on NVIDIA GeForce GTX 1080 Ti and NVIDIA GeForce RTX 3070. On 3070, YOLOv7 inference speed is approximately 50% less than 1080. Consider this in speed tests.

kpt, yolov7-w6-pose.pt, Can't call numpy() on Tensor that requires grad. Use tensor.detach().numpy() instead.

I added PR #333 to solve this problem.

kpt, yolov7-w6-pose.pt, Can't call numpy() on Tensor that requires grad. Use tensor.detach().numpy() instead.

use [export.py](https://github.com/WongKinYiu/yolov7/blob/main/export.py)

kpt, yolov7-w6-pose.pt, Can't call numpy() on Tensor that requires grad. Use tensor.detach().numpy() instead.

Do you use the latest YOLOv7 code? Take a look at #319

kpt, yolov7-w6-pose.pt, Can't call numpy() on Tensor that requires grad. Use tensor.detach().numpy() instead.

It's OK. Output of onnx model has shape of (number_of_detections, 7). Second part consists of (batch_id,x0,y0,x1,y1,cls_id,score)