Turns out private board is not so friendly for EfficientDet.... But sharing my solution anyways
\n\nOverview
\n\narvalis_3 and u_saskTurns out private board is not so friendly for EfficientDet.... But sharing my solution anyways
\n\nOverview
\n\narvalis_3 and u_saskGreat work, Thanks for sharing!\nDo you have any reason not use arvalis_3 and u_sask in pseudo-labeling stage?
The boxes quality looked really bad. Also, they have low validation mAP, while the images in the public LB seemed a lot \"easier.\"
", "rawMarkdown": "The boxes quality looked really bad. Also, they have low validation mAP, while the images in the public LB seemed a lot \"easier.\"", "votes": null }, { "id": "958591", "postDate": "08/05/2020 04:26:04", "content": "Was lowkey hoping you to win. If only there were medals for the most effort, you'd get gold.\nAnyway, adding to your solution with observations from our own (single fold model) and top solutions;\n- Single fold model: Not so friendly\n- Multi fold/model ensemble: Very stable high private and public score
\n\nYours is a single model, right? Tell me your thoughts on the same.
\n\nThe above conclusions are only for Effdet, YoloV5 single fold models (No Pseudo Labelling, just base model with simple TTA) we submitted when it wasn't banned are high. (Public: 0.76 - Private:0.69)
", "rawMarkdown": "Was lowkey hoping you to win. If only there were medals for the most effort, you'd get gold.\nAnyway, adding to your solution with observations from our own (single fold model) and top solutions;\n- Single fold model: Not so friendly\n- Multi fold/model ensemble: Very stable high private and public score\n\nYours is a single model, right? Tell me your thoughts on the same. \n\nThe above conclusions are only for Effdet, YoloV5 single fold models (No Pseudo Labelling, just base model with simple TTA) we submitted when it wasn't banned are high. (Public: 0.76 - Private:0.69)", "votes": null }, { "id": "958602", "postDate": "08/05/2020 04:33:42", "content": "Yes, mine is a single fold model. Moving on to the next challenge!
", "rawMarkdown": "Yes, mine is a single fold model. Moving on to the next challenge!", "votes": null }, { "id": "959176", "postDate": "08/05/2020 11:58:08", "content": "Thanks for sharing. \nYou've tried a lot, and it's great to share the experience.
\n\n", "rawMarkdown": "Thanks for sharing. \nYou've tried a lot, and it's great to share the experience.\n\n@alexanderliao", "votes": null } ], "comments": [ { "id": 958565, "author_name": "yummycocopalm", "author_url": "", "post_date": "08/05/2020 04:05:17", "content": "Great work, Thanks for sharing!\nDo you have any reason not use arvalis_3 and u_sask in pseudo-labeling stage?
The boxes quality looked really bad. Also, they have low validation mAP, while the images in the public LB seemed a lot \"easier.\"
", "votes": null, "replies": [] } ] }, { "id": 958591, "author_name": "ii5m0k3ii", "author_url": "", "post_date": "08/05/2020 04:26:04", "content": "Was lowkey hoping you to win. If only there were medals for the most effort, you'd get gold.\nAnyway, adding to your solution with observations from our own (single fold model) and top solutions;\n- Single fold model: Not so friendly\n- Multi fold/model ensemble: Very stable high private and public score
\n\nYours is a single model, right? Tell me your thoughts on the same.
\n\nThe above conclusions are only for Effdet, YoloV5 single fold models (No Pseudo Labelling, just base model with simple TTA) we submitted when it wasn't banned are high. (Public: 0.76 - Private:0.69)
", "votes": null, "replies": [ { "id": 958602, "author_name": "alexanderliao", "author_url": "", "post_date": "08/05/2020 04:33:42", "content": "Yes, mine is a single fold model. Moving on to the next challenge!
", "votes": null, "replies": [] } ] }, { "id": 959176, "author_name": "piantic", "author_url": "", "post_date": "08/05/2020 11:58:08", "content": "Thanks for sharing. \nYou've tried a lot, and it's great to share the experience.
\n\n", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "958535": "Turns out private board is not so friendly for EfficientDet.... But sharing my solution anyways\n\nOverview\n\n- EfficientDet-D6 with input size = 1024\n- Heavy augmentation with 1:1 Mixup (full beta for me didn't work), and 4-CutMix\n- Trained on original data, finetuned on external data\n- External data: full-sized image by jigsaw kernel + automatic box fusion on the edges. Available here: https://www.kaggle.com/alexanderliao/wheatfullsize\n- Pseudo-labeling: Freezing backbone, fine-tuning on all sources except `arvalis_3` and `u_sask`", "958565": "Great work, Thanks for sharing!\nDo you have any reason not use `arvalis_3` and `u_sask` in pseudo-labeling stage?", "958586": "The boxes quality looked really bad. Also, they have low validation mAP, while the images in the public LB seemed a lot \"easier.\"", "958591": "Was lowkey hoping you to win. If only there were medals for the most effort, you'd get gold.\nAnyway, adding to your solution with observations from our own (single fold model) and top solutions;\n- Single fold model: Not so friendly\n- Multi fold/model ensemble: Very stable high private and public score\n\nYours is a single model, right? Tell me your thoughts on the same. \n\nThe above conclusions are only for Effdet, YoloV5 single fold models (No Pseudo Labelling, just base model with simple TTA) we submitted when it wasn't banned are high. (Public: 0.76 - Private:0.69)", "958602": "Yes, mine is a single fold model. Moving on to the next challenge!", "959176": "Thanks for sharing. \nYou've tried a lot, and it's great to share the experience.\n\n@alexanderliao" }, "source": "meta" }