Train using mmdetection and submit via Colab (Round 2)!

Hey everyone,

Since a lot of people seem to be finding it difficult to submit, we’ve converted the baseline code using mmdetection into a colab notebook which allows you to submit directly via colab. If you are not using mmdetection you can still check this notebook out and have a look at the submission steps. If you have any issues or need any help do feel free to post here.

MMdetection Starter(DIRECT SUBMIT!)

AICrowd Team

1 Like

Thanks for this, particularly the try/except block for the training script was useful.

I have one question about the baseline MMDetection submission (food-round2 repo by @nikhil_rayaprolu): if it’s not a secret, what settings were used to generate the baseline model? Is it correct to assume that it was trained with the provided training hyperparameters (type='SGD', lr=0.0025, momentum=0.9, weight_decay=0.0001), the default learning policy, no pre-training (load_from = None & resume_from = None) and then epoch 20 was taken for the final model?

The reason I’m asking is that I have been using similar settings to train the same architecture (except with pre-training) and I have gotten considerably worse results in 30 epochs with the evaluator, so I would like to troubleshoot if my settings aren’t optimal or there’s something else wrong.


I think the baseline(food-round2) was created using this config file. You can compare this with the config you are using right now and try find out where the issue might be.

1 Like

Awesome - that was my assumption as well, but I figured it was also a possibility that the actual config file wouldn’t be shared. Thanks for the quick answer - much appreciated!