search

afzalsayed96 / intel-scene-classification

1st place solution for AV Intel scene classification challenge
https://datahack.analyticsvidhya.com/contest/practice-problem-intel-scene-classification-challe/
39 stars 16 forks source link

Should imagenet stats for place dataset be used? #1

Closed gloriamacia closed 4 years ago

gloriamacia commented 4 years ago

Hey! I was going through your amazing work. Thanks for sharing In 2-intel_fastai_res50_places.ipynb you are loading places weights but using image_net stats. Not sure if this is correct. Just to make sure.

afzalsayed96 commented 4 years ago

Hey,

You have to pass stats to normalize dataset and you have to calculate stats based on the dataset. But imagenet_stats are decent enough and it is safe to assume that it should work with most image datasets which it did in our case.

You can read more about it here: https://docs.fast.ai/vision.data.html#ImageDataBunch.normalize

If you are running the experiments yourself then feel free to try without imagenet_stats and see how it varies.

On Tue, 23 Jul, 2019, 9:49 PM Glòria Macià Muñoz, notifications@github.com wrote:

Hey! I was going through your amazing work. Thanks for sharing In 2-intel_fastai_res50_places.ipynb you are loading places weights but using image_net stats. Not sure if this is correct. Just to make sure.

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/afzalsayed96/intel_scene_classification/issues/1?email_source=notifications&email_token=ADLBEO63UJPNMG2F7OT6WGTQA4VRNA5CNFSM4IGGI6HKYY3PNVWWK3TUL52HS4DFUVEXG43VMWVGG33NNVSW45C7NFSM4HA67J3A, or mute the thread https://github.com/notifications/unsubscribe-auth/ADLBEO2TGT7NLEOOUIDP2ALQA4VRNANCNFSM4IGGI6HA .

gloriamacia commented 4 years ago

Ah, thanks a lot! I was really wondering. Awesome code, Afzal. Quick questions (just curiosity): you talk about different ensembles. How many models you end up using at the end?

afzalsayed96 commented 4 years ago

Hey,

There were not many different models (most of them were resnet50 with places365) but rather different techniques such as mixup, TTA, etc were used in ensembling. If I remember correctly I ended up ensembling around 50 submissions with various combination of techniques used in each of them to get the final solution.

You can read more about it in my write-up here:

“1st Place Solution for Intel Scene Classification Challenge” by Afzal Sayed https://link.medium.com/E1XlHajGAY

On Wed, 24 Jul, 2019, 1:03 PM Glòria Macià Muñoz, notifications@github.com wrote:

Ah, thanks a lot! I was really wondering. Awesome code, Afzal. Quick questions (just curiosity): you talk about different ensembles. How many models you end up using at the end?

— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/afzalsayed96/intel_scene_classification/issues/1?email_source=notifications&email_token=ADLBEO4Q6SNN7463UQNKEA3QBAAVBA5CNFSM4IGGI6HKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOD2VOJAY#issuecomment-514516099, or mute the thread https://github.com/notifications/unsubscribe-auth/ADLBEO5JBWHKZ4TAGZK5ZKTQBAAVBANCNFSM4IGGI6HA .

gloriamacia commented 4 years ago

Awesome! Congrats again :) Feel free to add me on linkedin. Truly great work!