Keras-retinanet: BBox do not appear

[rbothra3]

Hi all,

I did the following to train my dataset:

1. Installed Retinanet
2. Created a csv with my data in the recommended format
3. Ran: retinanet-train csv /home/rohit/Downloads/data_csv4.csv /home/rohit/Downloads/catcsv.csv
4. The training ran for about 1.5 days, 50 echos and created 50 snapshots in the snapshots folder
5. Edited the example file to change the path to:

model_path = "/home/rohit/Downloads/keras-retinanet/snapshots-old/resnet50_csv_10.h5"
When I run the example file, I do not get a bounding box. I get one warning which states:

/home/rohit/.local/lib/python3.6/site-packages/keras/models.py:255: UserWarning: No training configuration found in save file: the model was *not* compiled. Compile it manually.
  warnings.warn('No training configuration found in save file: '

Did I miss something? Again, thank you for your help!

[yhenon]

There was a significant bug(https://github.com/fizyr/keras-retinanet/pull/240) until recently that affected training. Could you check if the code you used was affected by the bug? If so, you'll have to retrain unfortunately (although it should take less than 1 day before you can check to see if you are starting to get predictions at low confidences).

[rbothra3]

I used the master from three days ago. Running it again to verify the performance.

[hoangminhtoan]

Dear,

I updated the #240 bug and re-run with pascal 2007 dataset. I still got the warning "No training configuration found in save file:". And one more thing, after I trained with default configs (50 epochs, 10000 steps/ epoch, weights = 'resnet50') I did not get the good result as your example. It only detected a guy on the screen and recognized him as a bird.

[hgaiser]

It's unfortunate, but it seems https://github.com/fizyr/keras-retinanet/pull/244 could have also been negatively affecting your results. Could you try retraining after that patch?

[hoangminhtoan]

Thanks, I am re-running with different weight ('imagenet' as default) to see the performance. It'll take me 1.5 days for training. then I'll update #244 and see.

[hgaiser]

@hoangminhtoan if your current training does not include #244 , but does include any form of data augmentation (even flips, which are enabled by default) then it is pointless to continue training as the result will be suboptimal. My advice is to restart the procedure on the latest master version.

[hoangminhtoan]

@hgaiser. thanks for your support, I pulled the newest version and have been running it since yesterday.

[hgaiser]

Alright, if it solved your issue, would you close this?

[rbothra3]

244 Fixed the issues. Thanks!