Spaces:
Running
Running
wd-v1-4-swinv2-tagger-v3 Torch Model is either Incorrect or Model is vastly different
#9
by
djmcflush
- opened
Problem:
It would seem that predictions made from the torch model and the onnx model are vastly different. So much so that I believe that the uploaded torch model may be from a different checkpoint.
Description:
Tested this out with the following images:
Please check out the reference implementations mentioned in the model page: https://huggingface.co/SmilingWolf/wd-swinv2-tagger-v3
In particular, make sure the image has got the correct channel order (BGR), correct scaling (0..255 -> -1..1) and that the logits of the pytorch models go through the sigmoid activation function for a 1:1 comparison.
No further news from the interested party. Closing.
SmilingWolf
changed discussion status to
closed