v1.6.3: cannot use checkpoint *.ckpt file to predict and convert to ONNX as we could do with *.pth files #924
Description
Search before asking
- I have searched the RF-DETR issues and found no similar feature requests.
Description
I try to load weights from "last.ckpt" into its model for prediction and ONNX conversion with specific resolution and device but meet error:
checkpoint_num_classes = checkpoint["model"]["class_embed.bias"].shape[0]
KeyError: 'model'
My code is:
model= RFDETRLarge(pretrain_weights= cfg.checkpoint_path, resolution= cfg.resolution, device= 'cuda')
model.export(output_dir= cfg.output_dir, verbose=False)
detections= model.predict(image, threshold=0.5)
I don't have this issue with other *.pth files. Please advise how to solve it. I wonder if we can back to *.pth file as v1.5.x
Use case
No response
Additional
No response
Are you willing to submit a PR?
- Yes I'd like to help by submitting a PR!