![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F4241674%2Fe153943075ee571c7b42182c0a3a41eb%2F2022-10-11%20214238.png?generation=1665492226981590&alt=media\")
Congrats to all the winners, and thanks to organizers and all participants. This is a joint writeup with my wonderful teammate @yamash73 and @ksork6s4 (thanks to awesome team work).
\nThings that worked
\nDataset
\nWe have noticed that Public LB scores vary greatly depending on the combination of datasets. Therefore, we are exploring the best combination of data sets in the early stages of the competition. Comparison of experimental results shows that \"Inshop\", \"Instre\" and \"Consumer-to-shop\" are effective. Based on these results, we have decided to use \"Products10k\", \"Landmark\", \"Instre\" and \"Consumer-to-shop\" as our basic dataset combination . Here \"Inshop\" and \"Consumer-to-shop\" are subsets of deepfashion, so only \"Consumer-to-shop\" is selected as our combination of datasets.
NOTE:
\nOur final combination of datasets is as follows:
\nModel
\nThe backbone of our final model is an ensemble of 3 OpenCLIP image encoders. OpenCLIP backbones contributed to the huge improvement of the LB score. We refered to the head/loss design from the past competitions (mainly Google Landmark Retrieval 2021).
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F4241674%2Fe9e01248f4d6655c00d75348c54aeda5%2F1.png?generation=1665490798502330&alt=media\")
Training
\nOnly head is trained like other teams. At first, we set the total epochs to 20, but after experiments showed that we could achieve the same level of performance with 5 epochs, we were able to increase the number of trials.
Things that didn't work
\n@koheiumeki Interesting thanks for sharing, we also follow a similar structure. Did you have problems when using 3 backbones considering the 9hrs limit?
", "votes": 5, "replies": [ { "id": 1983476, "author_name": "Koo", "author_url": "", "post_date": "2022-10-12T04:39:13.497000", "content": "@jesucristoe
\nThank you for your question and congratulations on 4th place!
We could fit 3 models into 9 hours limit without any special changes.
\nThe approximate submission times for each model are as follows:
In our training phase, fp16 is used for 3 backbones but it is not used for the submission model. This is our submission notebook.
\nhttps://www.kaggle.com/code/ksork6s4/guie-submission-openclip-3models-ensemble
Additional information:
\nViT-B is used in our final model, but it may not be necessary as similar scores can be reached with only ViT-H and ViT-L.
How did you fit OpenCLIP ViT-B + OpenCLIP ViT-L + OpenCLIP ViT-H into 9 hours of given submission time?
@manwithaflower
\nThank you for your question!
We could fit 3 models into 9 hours limit without any special changes.
\nThe approximate submission times for each model are as follows:
In our training phase, fp16 is used for 3 backbones but it is not used for the submission model. This is our submission notebook.
\nhttps://www.kaggle.com/code/ksork6s4/guie-submission-openclip-3models-ensemble
Additional information:
\nViT-B is used in our final model, but it may not be necessary as similar scores can be reached with only ViT-H and ViT-L.
Congratulations to Gold medal.
\nCongratulations to @ksork6s4 and @yamash73 for winning the master. I have to do my best too.
Congratulations! Can you share the training loss after 5 epochs?
", "votes": 2, "replies": [ { "id": 1983487, "author_name": "Koo", "author_url": "", "post_date": "2022-10-12T04:46:55.440000", "content": "@evilpsycho42
\nThank you for your comment and question!
The training loss after 5 epochs is 18.2422.
\nThis is the sum of ArcFace loss and focal loss.
BTW, what the LR do you use when unfreezing training?
", "votes": 1, "replies": [ { "id": 1983493, "author_name": "Koo", "author_url": "", "post_date": "2022-10-12T04:59:25.280000", "content": "@evilpsycho42
\nWe tested 1e-3 or 1e-4 after head-only training and scores did not improve.
\nBut we tried unfreezing training only in the beginning of the competition, so this may not be sufficient.
We also referred to this solution and parameters:
\nCongrats to all the winners, and thanks to organizers and all participants. This is a joint writeup with my wonderful teammate @yamash73 and @ksork6s4 (thanks to awesome team work).
\nThings that worked
\nDataset
\nWe have noticed that Public LB scores vary greatly depending on the combination of datasets. Therefore, we are exploring the best combination of data sets in the early stages of the competition. Comparison of experimental results shows that \"Inshop\", \"Instre\" and \"Consumer-to-shop\" are effective. Based on these results, we have decided to use \"Products10k\", \"Landmark\", \"Instre\" and \"Consumer-to-shop\" as our basic dataset combination . Here \"Inshop\" and \"Consumer-to-shop\" are subsets of deepfashion, so only \"Consumer-to-shop\" is selected as our combination of datasets.
NOTE:
\nOur final combination of datasets is as follows:
\nModel
\nThe backbone of our final model is an ensemble of 3 OpenCLIP image encoders. OpenCLIP backbones contributed to the huge improvement of the LB score. We refered to the head/loss design from the past competitions (mainly Google Landmark Retrieval 2021).
Training
\nOnly head is trained like other teams. At first, we set the total epochs to 20, but after experiments showed that we could achieve the same level of performance with 5 epochs, we were able to increase the number of trials.
Things that didn't work
\n@koheiumeki Interesting thanks for sharing, we also follow a similar structure. Did you have problems when using 3 backbones considering the 9hrs limit?
", "rawMarkdown": "@koheiumeki Interesting thanks for sharing, we also follow a similar structure. Did you have problems when using 3 backbones considering the 9hrs limit?", "votes": 5, "replies": [ { "id": 1983476, "postDate": "2022-10-12T04:39:13.497Z", "content": "@jesucristoe
\nThank you for your question and congratulations on 4th place!
We could fit 3 models into 9 hours limit without any special changes.
\nThe approximate submission times for each model are as follows:
In our training phase, fp16 is used for 3 backbones but it is not used for the submission model. This is our submission notebook.
\nhttps://www.kaggle.com/code/ksork6s4/guie-submission-openclip-3models-ensemble
Additional information:
\nViT-B is used in our final model, but it may not be necessary as similar scores can be reached with only ViT-H and ViT-L.
How did you fit OpenCLIP ViT-B + OpenCLIP ViT-L + OpenCLIP ViT-H into 9 hours of given submission time?
@manwithaflower
\nThank you for your question!
We could fit 3 models into 9 hours limit without any special changes.
\nThe approximate submission times for each model are as follows:
In our training phase, fp16 is used for 3 backbones but it is not used for the submission model. This is our submission notebook.
\nhttps://www.kaggle.com/code/ksork6s4/guie-submission-openclip-3models-ensemble
Additional information:
\nViT-B is used in our final model, but it may not be necessary as similar scores can be reached with only ViT-H and ViT-L.
Congratulations to Gold medal.
\nCongratulations to @ksork6s4 and @yamash73 for winning the master. I have to do my best too.
Congratulations! Can you share the training loss after 5 epochs?
", "rawMarkdown": "Congratulations! Can you share the training loss after 5 epochs?", "votes": 2, "replies": [ { "id": 1983487, "postDate": "2022-10-12T04:46:55.440Z", "content": "@evilpsycho42
\nThank you for your comment and question!
The training loss after 5 epochs is 18.2422.
\nThis is the sum of ArcFace loss and focal loss.
BTW, what the LR do you use when unfreezing training?
", "rawMarkdown": "BTW, what the LR do you use when unfreezing training?", "votes": 1, "replies": [ { "id": 1983493, "postDate": "2022-10-12T04:59:25.280Z", "content": "@evilpsycho42
\nWe tested 1e-3 or 1e-4 after head-only training and scores did not improve.
\nBut we tried unfreezing training only in the beginning of the competition, so this may not be sufficient.
We also referred to this solution and parameters:
\n", "rawMarkdown": "@evilpsycho42 \nWe tested 1e-3 or 1e-4 after head-only training and scores did not improve.\nBut we tried unfreezing training only in the beginning of the competition, so this may not be sufficient. \n\nWe also referred to this solution and parameters:\n- solution: https://www.kaggle.com/competitions/landmark-retrieval-2021/discussion/277099\n- code: https://github.com/ChristofHenkel/kaggle-landmark-2021-1st-place/tree/main/configs" } ] } ], "complete": true, "source": "live", "fetched_at": "2026-10-24T13:23:33Z", "competition_id": 36414, "competition_slug": "google-universal-image-embedding", "team_id": 8884347, "solution_url": "/competitions/google-universal-image-embedding/writeups/icd-10th-place-solution" }