search

dlunion / DBFace

DBFace is a real-time, single-stage detector for face detection, with faster speed and higher accuracy
1.35k stars 303 forks source link

Onnx input dummy question #26

Closed ckcraig01 closed 4 years ago

ckcraig01 commented 4 years ago

Dear Author:

Thank you very much.

The dummy is set to dummy = torch.zeros((1, 3, 32, 32)).cuda(): https://github.com/dlunion/DBFace/blob/1b408dbc37ed5b235c619b52dc3bec9276fd38f9/train/small/onnx.py#L34

However, the input size of model seems to be, 800x800:

https://github.com/dlunion/DBFace/blob/1b408dbc37ed5b235c619b52dc3bec9276fd38f9/train/small/train-small-H-keep12-noext-ignoresmall2.py#L19

Not sure if you mean that we shall adjust the model input size (for example 512x512) and re-train the model, then use the very size (1,3,512,512) to generated the final onnx. Thank you very much.

dlunion commented 4 years ago

DBFace is a fully convolutional network, so when exporting onnx, the resolution is 32x32, and the input resolution will be modified as needed during inference. So 32x32 has nothing to do with the trained 800x800

Refer here: https://github.com/dlunion/tensorRTIntegrate/blob/7543614a4ff5f6e022526a4e137d47f4e63f5f96/src/examples/dbface.cpp#L148

ckcraig01 commented 4 years ago

Dear Author:

Thank you for the help!