{ "id": 303839, "title": "What is your best ensembling score?", "url": "/competitions/tensorflow-great-barrier-reef/discussion/303839", "author_name": "", "post_date": "2022-01-29T15:20:09.091557400Z", "votes": null, "comment_count": 20, "views": 0, "content": "

What is your best ensembling score?Let's have a talk!

", "messages": [ { "id": "1668388", "postDate": "01/29/2022 15:20:09", "content": "

What is your best ensembling score?Let's have a talk!

", "rawMarkdown": "What is your best ensembling score?Let's have a talk!", "votes": null }, { "id": "1668392", "postDate": "01/29/2022 15:21:35", "content": "

Oh My friends！！！ This is a nice comment

", "rawMarkdown": "Oh My friends！！！ This is a nice comment", "votes": null }, { "id": "1668393", "postDate": "01/29/2022 15:22:17", "content": "

The fusion effect of the two models will be better? Can Yolo series get higher scores by integrating Yolo series?

", "rawMarkdown": "The fusion effect of the two models will be better? Can Yolo series get higher scores by integrating Yolo series?", "votes": null }, { "id": "1669002", "postDate": "01/30/2022 08:37:37", "content": "

My ensembling score << a single yolov5 score😂

", "rawMarkdown": "My ensembling score << a single yolov5 score😂", "votes": null }, { "id": "1670966", "postDate": "02/01/2022 03:33:54", "content": "

my best ensembling score yet is 0.630 with inference size is 3600. In my case, it outperforms all my models in the same resolution inference.

", "rawMarkdown": "my best ensembling score yet is 0.630 with inference size is 3600. In my case, it outperforms all my models in the same resolution inference.", "votes": null }, { "id": "1670975", "postDate": "02/01/2022 03:45:56", "content": "

single model 0.71, ensembling 0.70 …

", "rawMarkdown": "single model 0.71, ensembling 0.70 ...", "votes": null }, { "id": "1670978", "postDate": "02/01/2022 03:50:57", "content": "

Wow, can i ask your score come from hyperparameters, data or post processing sir? i tried lots of experiments but seems nothing works yet. 👀

", "rawMarkdown": "Wow, can i ask your score come from hyperparameters, data or post processing sir? i tried lots of experiments but seems nothing works yet. 👀", "votes": null }, { "id": "1670982", "postDate": "02/01/2022 03:58:00", "content": "

all of them i think

", "rawMarkdown": "all of them i think", "votes": null }, { "id": "1671180", "postDate": "02/01/2022 08:17:29", "content": "

Wow!!!Could you share sth.?Which model do you use?Yolov5 or something else?

", "rawMarkdown": "Wow!!!Could you share sth.?Which model do you use?Yolov5 or something else?", "votes": null }, { "id": "1671673", "postDate": "02/01/2022 16:39:01", "content": "

ensemble create more bounding box. it will increase recall.

\n

unless each of your model has low FP, it is likely that FP will increase tool.

\n

i suggest ensemble + post verification to remove FP. it is like cascade rcnn, there are many stages for propsal generation and roi classification.

", "rawMarkdown": "ensemble create more bounding box. it will increase recall.\n\nunless each of your model has low FP, it is likely that FP will increase tool.\n\ni suggest ensemble + post verification to remove FP. it is like cascade rcnn, there are many stages for propsal generation and roi classification.", "votes": null }, { "id": "1671728", "postDate": "02/01/2022 17:23:02", "content": "

Ensembling + increasing conf works best for me. But I'm moving towards doing exactly that!

", "rawMarkdown": "Ensembling + increasing conf works best for me. But I'm moving towards doing exactly that!", "votes": null }, { "id": "1672594", "postDate": "02/02/2022 07:22:33", "content": "

YoloV5's TTA is essentially an ensemble of the model at different inference sizes

", "rawMarkdown": "YoloV5's TTA is essentially an ensemble of the model at different inference sizes", "votes": null }, { "id": "1672741", "postDate": "02/02/2022 09:11:48", "content": "

just yolov5

", "rawMarkdown": "just yolov5", "votes": null }, { "id": "1673916", "postDate": "02/03/2022 05:05:48", "content": "

I got the similar results.Tracker can not improve my best model（lb0.661），but I am try to find some other ways to get better score and figure out it.

", "rawMarkdown": "I got the similar results.Tracker can not improve my best model（lb0.661），but I am try to find some other ways to get better score and figure out it.", "votes": null }, { "id": "1675084", "postDate": "02/04/2022 03:02:58", "content": "

0.659 with higher conf settings of 3 different public models + tracking. my best score right now is a single model that I trained

", "rawMarkdown": "0.659 with higher conf settings of 3 different public models + tracking. my best score right now is a single model that I trained", "votes": null }, { "id": "1675093", "postDate": "02/04/2022 03:12:38", "content": "

which one doea you use?Yolov5s6 or m6?

", "rawMarkdown": "which one doea you use?Yolov5s6 or m6?", "votes": null }, { "id": "1675096", "postDate": "02/04/2022 03:16:43", "content": "

i think all of them were yolov5s6. my current model is a modified yolov5s6 that I pretrained on another dataset

", "rawMarkdown": "i think all of them were yolov5s6. my current model is a modified yolov5s6 that I pretrained on another dataset", "votes": null }, { "id": "1675104", "postDate": "02/04/2022 03:27:28", "content": "

how about talking together?

", "rawMarkdown": "how about talking together?", "votes": null }, { "id": "1675106", "postDate": "02/04/2022 03:30:39", "content": "

what do you mean? I am a bit confused. I don't want to reveal too much since it is towards the end of the competition.

", "rawMarkdown": "what do you mean? I am a bit confused. I don't want to reveal too much since it is towards the end of the competition.", "votes": null }, { "id": "1675128", "postDate": "02/04/2022 04:09:34", "content": "

I think it is not a steady score to own silver

", "rawMarkdown": "I think it is not a steady score to own silver", "votes": null }, { "id": "1675385", "postDate": "02/04/2022 08:29:52", "content": "

You made the topic for \"talking together\"? Interesting, your team is full of five people.

", "rawMarkdown": "You made the topic for \"talking together\"? Interesting, your team is full of five people.", "votes": null } ], "comments": [ { "id": 1668392, "author_name": "catchbat", "author_url": "", "post_date": "01/29/2022 15:21:35", "content": "

Oh My friends！！！ This is a nice comment

", "votes": null, "replies": [] }, { "id": 1668393, "author_name": "yuntianzhang", "author_url": "", "post_date": "01/29/2022 15:22:17", "content": "

The fusion effect of the two models will be better? Can Yolo series get higher scores by integrating Yolo series?

", "votes": null, "replies": [] }, { "id": 1669002, "author_name": "rainfalllove", "author_url": "", "post_date": "01/30/2022 08:37:37", "content": "

My ensembling score << a single yolov5 score😂

", "votes": null, "replies": [] }, { "id": 1670966, "author_name": "locbaop", "author_url": "", "post_date": "02/01/2022 03:33:54", "content": "

my best ensembling score yet is 0.630 with inference size is 3600. In my case, it outperforms all my models in the same resolution inference.

", "votes": null, "replies": [] }, { "id": 1670975, "author_name": "snaker", "author_url": "", "post_date": "02/01/2022 03:45:56", "content": "

single model 0.71, ensembling 0.70 …

", "votes": null, "replies": [ { "id": 1670978, "author_name": "locbaop", "author_url": "", "post_date": "02/01/2022 03:50:57", "content": "

Wow, can i ask your score come from hyperparameters, data or post processing sir? i tried lots of experiments but seems nothing works yet. 👀

", "votes": null, "replies": [] }, { "id": 1670982, "author_name": "snaker", "author_url": "", "post_date": "02/01/2022 03:58:00", "content": "

all of them i think

", "votes": null, "replies": [] }, { "id": 1671180, "author_name": "rishengyang", "author_url": "", "post_date": "02/01/2022 08:17:29", "content": "

Wow!!!Could you share sth.?Which model do you use?Yolov5 or something else?

", "votes": null, "replies": [] }, { "id": 1672741, "author_name": "snaker", "author_url": "", "post_date": "02/02/2022 09:11:48", "content": "

just yolov5

", "votes": null, "replies": [] }, { "id": 1673916, "author_name": "freshair1996", "author_url": "", "post_date": "02/03/2022 05:05:48", "content": "

I got the similar results.Tracker can not improve my best model（lb0.661），but I am try to find some other ways to get better score and figure out it.

", "votes": null, "replies": [] } ] }, { "id": 1671673, "author_name": "hengck23", "author_url": "", "post_date": "02/01/2022 16:39:01", "content": "

ensemble create more bounding box. it will increase recall.

\n

unless each of your model has low FP, it is likely that FP will increase tool.

\n

i suggest ensemble + post verification to remove FP. it is like cascade rcnn, there are many stages for propsal generation and roi classification.

", "votes": null, "replies": [ { "id": 1671728, "author_name": "outwrest", "author_url": "", "post_date": "02/01/2022 17:23:02", "content": "

Ensembling + increasing conf works best for me. But I'm moving towards doing exactly that!

", "votes": null, "replies": [] } ] }, { "id": 1672594, "author_name": "alexchwong", "author_url": "", "post_date": "02/02/2022 07:22:33", "content": "

YoloV5's TTA is essentially an ensemble of the model at different inference sizes

", "votes": null, "replies": [] }, { "id": 1675084, "author_name": "outwrest", "author_url": "", "post_date": "02/04/2022 03:02:58", "content": "

0.659 with higher conf settings of 3 different public models + tracking. my best score right now is a single model that I trained

", "votes": null, "replies": [ { "id": 1675093, "author_name": "rishengyang", "author_url": "", "post_date": "02/04/2022 03:12:38", "content": "

which one doea you use?Yolov5s6 or m6?

", "votes": null, "replies": [] }, { "id": 1675096, "author_name": "outwrest", "author_url": "", "post_date": "02/04/2022 03:16:43", "content": "

i think all of them were yolov5s6. my current model is a modified yolov5s6 that I pretrained on another dataset

", "votes": null, "replies": [] }, { "id": 1675104, "author_name": "rishengyang", "author_url": "", "post_date": "02/04/2022 03:27:28", "content": "

how about talking together?

", "votes": null, "replies": [] }, { "id": 1675106, "author_name": "outwrest", "author_url": "", "post_date": "02/04/2022 03:30:39", "content": "

what do you mean? I am a bit confused. I don't want to reveal too much since it is towards the end of the competition.

", "votes": null, "replies": [] }, { "id": 1675128, "author_name": "rishengyang", "author_url": "", "post_date": "02/04/2022 04:09:34", "content": "

I think it is not a steady score to own silver

", "votes": null, "replies": [] }, { "id": 1675385, "author_name": "freshair1996", "author_url": "", "post_date": "02/04/2022 08:29:52", "content": "

You made the topic for \"talking together\"? Interesting, your team is full of five people.

", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1668388": "What is your best ensembling score?Let's have a talk!", "1668392": "Oh My friends！！！ This is a nice comment", "1668393": "The fusion effect of the two models will be better? Can Yolo series get higher scores by integrating Yolo series?", "1669002": "My ensembling score << a single yolov5 score😂", "1670966": "my best ensembling score yet is 0.630 with inference size is 3600. In my case, it outperforms all my models in the same resolution inference.", "1670975": "single model 0.71, ensembling 0.70 ...", "1670978": "Wow, can i ask your score come from hyperparameters, data or post processing sir? i tried lots of experiments but seems nothing works yet. 👀", "1670982": "all of them i think", "1671180": "Wow!!!Could you share sth.?Which model do you use?Yolov5 or something else?", "1671673": "ensemble create more bounding box. it will increase recall.\n\nunless each of your model has low FP, it is likely that FP will increase tool.\n\ni suggest ensemble + post verification to remove FP. it is like cascade rcnn, there are many stages for propsal generation and roi classification.", "1671728": "Ensembling + increasing conf works best for me. But I'm moving towards doing exactly that!", "1672594": "YoloV5's TTA is essentially an ensemble of the model at different inference sizes", "1672741": "just yolov5", "1673916": "I got the similar results.Tracker can not improve my best model（lb0.661），but I am try to find some other ways to get better score and figure out it.", "1675084": "0.659 with higher conf settings of 3 different public models + tracking. my best score right now is a single model that I trained", "1675093": "which one doea you use?Yolov5s6 or m6?", "1675096": "i think all of them were yolov5s6. my current model is a modified yolov5s6 that I pretrained on another dataset", "1675104": "how about talking together?", "1675106": "what do you mean? I am a bit confused. I don't want to reveal too much since it is towards the end of the competition.", "1675128": "I think it is not a steady score to own silver", "1675385": "You made the topic for \"talking together\"? Interesting, your team is full of five people." }, "source": "meta" }