Yolov5: Why setting model[-1].export = True during export

[govindamagrawal]

❔Question

Why are we setting model[-1].export = True in export.py?

model.model[-1].export = True  # set Detect() layer export=True

Why I am asking this as the prediction using the model, with and without setting the export is different. So,

model = torch.load(opt.weights, map_location=torch.device('cpu'))['model'].float()
model.eval()
model.model[-1].export = True  # set Detect() layer export=True
img = torch.zeros((1,3,224,224)
pred = model(image)

Here, pred[0].shape = torch.Size([1,3,7,7,85])

Now, when I do,

model = torch.load(opt.weights, map_location=torch.device('cpu'))['model'].float()
model.eval()
img = torch.zeros((1,3,224,224)
pred = model(image)

Here, pred[0].shape = torch.Size([1,3087,85])

Can anyone please explain this. I guess it is to generalize for any input shape.

Also, if setting the export=True is the correct way to export, how to get the correct output of shape ([1,3087,85]), as this is further used for non-max compression and matrix of ([1,3,7,7,85]) cannot be used.

[love2learn-7179]

I think w/ the export flag it returns training output, while w/o the flag it returns inference output. So you shouldn't set this flag for inferencing. plz see line 34 @https://github.com/ultralytics/yolov5/blob/master/models/yolo.py

[github-actions[bot]]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.