Updates on model weights and inference timings

[veronica1428]

Hi,

I tried to use the pretrained weights on the Lisa Daysequence1 dataset, but somehow I am not getting the mAP as 0.91 (the one mentioned in ReadMe).

Are there any update on the pretrained model weights ? Also, just to confirm, are these weights are for 67th epoch only ?

Also, if possible, can you let us know the inference timings of the current pretrained model.

Thanks and Regards
Veronica

[sovit-123]

Did you try training again. That mAP is the validation mAP at 67th epoch.
And that is the final best weight. No other models were trained after that. For the inference timings, I will have search for the files a bit more. Might take some time.

[veronica1428]

Hi,

Yes, I tried training again and for more epochs. But mAP and Recall are decreasing for more epochs.

Just curious to know why after 67th epoch specifically for YoloV3 all metrics are going down ?
You also faced the same issue ?

Also, it written in README,
best_model_12.pt: Trained for 67 epochs on all the traffic signs. Current mAP is 0.919
Is it traffic lights, right ?

Thanks and Regards
Veronica

[sovit-123]

So, after 67 epochs, the mode is starting to overfit. That's why I chose that model.

And I will correct the traffic signs to traffic lights. Thanks for pointing that out.

[veronica1428]

Hi, yup, so even I got recall and precision quite high, but thats on validation dataset.
Specifically, for test dataset (lisa), getting really very low recall and precision compared to val.

Just wanted to ask, you mAP was 0.9 on test dataset or validation ?

[sovit-123]

Yes, on the validation dataset.