RuntimeError: The size of tensor a (12828) must match the size of tensor b (3000) at non-singleton dimension 1

[SJavad]

Hi 🖐 I Trained an SSD model with --resolution=640 and when I want to convert the .pth model to .onnx, it gives me an error:

    locations[..., :2] * center_variance * priors[..., 2:] + priors[..., :2],
RuntimeError: The size of tensor a (12828) must match the size of tensor b (3000) at non-singleton dimension 1

when I want to convert the model to 300*300 size, it is OK and converts successfully... but when I set the --width=640 and --height=640, this error happened.

exported command:

Namespace(net='ssd-mobilenet', input='', output='', labels='labels.txt', width=640, height=640, batch_size=1, model_dir='.\\models\\GED\\')
=> running on device cuda:0
=> found best checkpoint with loss 2.568583369255066 (.\models\GED\mb1-ssd-Epoch-6-Loss-2.568583369255066.pth)
creating network:  ssd-mobilenet
num classes:       3
=> loading checkpoint:  .\models\GED\mb1-ssd-Epoch-6-Loss-2.568583369255066.pth
=> exporting model to ONNX...

[SJavad]

A LOT of open issues and... 🤯 no answer... Hmmm 🤔 Interesting... 🧐

[Zalways]

hey！i met the same problem with you! have you solved this problem?looking forward to your reply!

[SJavad]

hey！i met the same problem with you! have you solved this problem?looking forward to your reply!

@Zalways Hi I Talk with dusty-nv and get say, pull again the repository (in dusty-nv) repo to solve this problem seems that is the bug and solved