The file output after training the model is a .ckpt file. Is this the final result of the model?

[niwenbo123]

Hello, I used the model you provided for training, and after completing the training, I noticed that the output file is a .ckpt file. I would like to confirm whether this .ckpt file is the final result of the model training? Do I need any further steps (such as converting it to another format) in order to use it for inference or deployment?

[henrytsui000]

Hi,

Yes, the .ckpt file can be considered the final result of the model training. To use it for inference or deployment, we currently recommend the following adjustments because we haven't fully merged the lightning training process and the original process.

In yolo/lazy.py, modify line 37:
https://github.com/WongKinYiu/YOLO/blob/51f507f61e733b2a1ef993a5bc544f4e8b749f1d/yolo/lazy.py#L37
to

model = ValidateModel.load_from_checkpoint('[your checkpoint path]', cfg=cfg)

This should allow you to load your checkpoint for inference or further deployment.

Best regards,
Henry Tsui

[ramonhollands]

@henrytsui000

I extended the code to do inference on my custom trained model:

""" if cfg.task.task == "inference":
if 'ckpt' in cfg.weight:
model = InferenceModel.load_from_checkpoint(cfg.weight, cfg=cfg)
else:
model = InferenceModel(cfg)
trainer.predict(model)
"""

This gives errors related to
⚠ Weight Not Found for key: 18.conv3.0.bottleneck.1.conv1.conv1 (because they are stored as model.model.
and
Unexpected key(s) in state_dict: "ema.model.0.conv.weight"

What do you recommend? I saw the related PR (152) but I think the 'load_from_checkpoint' is the way to go eventually.