Can you provide some details about these lines in yolov5_tflite about onnx_tflite_yolov3 HOT 2 CLOSED

These lines of code is basically building output from previous convolutions, which is part of the yolov5 Detect module.

a = torch.tensor(anchors).float().view(nl, -1, 2).to(device)
anchor_grid = a.clone().view(nl, 1, -1, 1, 2)  # shape(nl,1,na,1,2)

builds anchors, which read from yaml files (I will change it to read from *.pt weights ultralytics/yolov5#1127 (comment)).

    _, _, ny_nx, _ = x[i].shape
    r = imgsz[0] / imgsz[1]
    nx = int(np.sqrt(ny_nx / r))
    ny = int(r * nx)

gets the output shape ny (vertical) and nx (horizontal) of the i-th yolo head.

    grid[i] = _make_grid(nx, ny).to(x[i].device)

creates meshgrid from the shape ny and nx.

    y[..., 0:2] = (y[..., 0:2] * 2. - 0.5 + grid[i].to(x[i].device)) * stride  # xy
    y[..., 2:4] = (y[..., 2:4] * 2) ** 2 * anchor_grid[i]  # wh

are the equations to output center coordinates xy and, width and height of the output bounding boxes.

Or you could post your question on the PRs:
ultralytics/yolov5#959
ultralytics/yolov5#1127

from onnx_tflite_yolov3.

Close this issue due to there're other suitable places for this issue (ultralytics/yolov5#959, ultralytics/yolov5#1127).

from onnx_tflite_yolov3.