![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1965436%2F3153e16798fcb1ed291e678e648e9282%2FScreenshot%20from%202020-07-15%2010-48-10.png?generation=1594816395555302&alt=media\")
We have a licensing issue.\nAnd now it seems like we really need to share new information.
\n\nI'm going to start over with models that fit the license rules. \nLet's share some information. 😃
\n\n> \nModel: DETR\nEpoch: 50\nLB: 0.598
\n\n\n\n\nModel: EfficientDet-D6\n Epoch: 80\n LB: 0.717
\n
p.s. DETR is released under the Apache 2.0 license.
", "messages": [ { "id": "930094", "postDate": "07/15/2020 07:35:28", "content": "We have a licensing issue.\nAnd now it seems like we really need to share new information.
\n\nI'm going to start over with models that fit the license rules. \nLet's share some information. 😃
\n\n> \nModel: DETR\nEpoch: 50\nLB: 0.598
\n\n\n\n\nModel: EfficientDet-D6\n Epoch: 80\n LB: 0.717
\n
p.s. DETR is released under the Apache 2.0 license.
", "rawMarkdown": "We have a licensing issue.\nAnd now it seems like we really need to share new information.\n\nI'm going to start over with models that fit the license rules. \nLet's share some information. 😃 \n\n# Baseline\n> \nModel: DETR\nEpoch: 50\nLB: 0.598\n\n\n# Updated\n> \nModel: EfficientDet-D6\nEpoch: 80\nLB: 0.717\n\n\np.s. DETR is released under the Apache 2.0 license.", "votes": null }, { "id": "930151", "postDate": "07/15/2020 08:26:44", "content": "[update] Mine is at 0.764 without pseudo-labeling (custom pipeline). My current score (keeps updating) is with pseudo-labeling. I also have a DETR model at CV around 0.57, but I'm not planning to submit it before the end of the competition.
", "rawMarkdown": "[update] Mine is at 0.764 without pseudo-labeling (custom pipeline). My current score (keeps updating) is with pseudo-labeling. I also have a DETR model at CV around 0.57, but I'm not planning to submit it before the end of the competition.", "votes": null }, { "id": "930159", "postDate": "07/15/2020 08:29:51", "content": "Thanks for share it!\n0.7621 without pseudo-labeling is amazing. 👍
", "rawMarkdown": "Thanks for share it!\n0.7621 without pseudo-labeling is amazing. 👍", "votes": null }, { "id": "930397", "postDate": "07/15/2020 12:34:26", "content": "Hi @alexanderliao @piantic I've tried DETR for several days, but still hit no luck. My bbox_loss is oscillating drastically and periodically. Do you guys happen to come across such wired circumstances?
\n\n
I doubt that my input box format is wrong. I complied to the [xc, yc, w, h] / image_size format from the paper. How about yours? \nThanks you very much for your help!
I almost used the existing training kernel.\nhttps://www.kaggle.com/tanulsingh077/end-to-end-object-detection-with-transformers-detr
\n\nIt is a little difficult to confirm now. \nI'll check it as I continue experimenting.\n@stevenyin
", "rawMarkdown": "I almost used the existing training kernel.\nhttps://www.kaggle.com/tanulsingh077/end-to-end-object-detection-with-transformers-detr\n\nIt is a little difficult to confirm now. \nI'll check it as I continue experimenting.\n@stevenyin", "votes": null }, { "id": "930428", "postDate": "07/15/2020 13:02:17", "content": "would you mind to share you model info?
", "rawMarkdown": "would you mind to share you model info?", "votes": null }, { "id": "930432", "postDate": "07/15/2020 13:06:38", "content": "I don't think DETR works for this competition. A nice concept though.
", "rawMarkdown": "I don't think DETR works for this competition. A nice concept though.", "votes": null }, { "id": "930439", "postDate": "07/15/2020 13:12:31", "content": "@piantic Yes, I started from that kernel and do some customization to the pipeline, which does not hit the luck so far. Thanks for you help!
", "rawMarkdown": "piantic Yes, I started from that kernel and do some customization to the pipeline, which does not hit the luck so far. Thanks for you help!", "votes": null }, { "id": "930441", "postDate": "07/15/2020 13:14:01", "content": "Hi Liao, after reading the paper I begin to double about it as well. But can't get it started normally drives me a little annoyed...
", "rawMarkdown": "Hi Liao, after reading the paper I begin to double about it as well. But can't get it started normally drives me a little annoyed...", "votes": null }, { "id": "930870", "postDate": "07/15/2020 19:02:58", "content": "You can check out my existing threads for something along the same lines:\n- Best single model - No one has actually posted any YoloV5 scores here. So it is still relevant and I don't see the need to split the thread into a new one\n- Leaderboard of Compliant Submissions - Here I have compiled the top compliant scores into a leaderboard
", "rawMarkdown": "You can check out my existing threads for something along the same lines:\n- [Best single model](https://www.kaggle.com/c/global-wheat-detection/discussion/153084) - No one has actually posted any YoloV5 scores here. So it is still relevant and I don't see the need to split the thread into a new one\n- [Leaderboard of Compliant Submissions](https://www.kaggle.com/c/global-wheat-detection/discussion/166793) - Here I have compiled the top compliant scores into a leaderboard", "votes": null }, { "id": "930933", "postDate": "07/15/2020 20:25:31", "content": "How about FasterRCNN? Is EfficientDet better than FasterRCNN?
", "rawMarkdown": "How about FasterRCNN? Is EfficientDet better than FasterRCNN?", "votes": null }, { "id": "932016", "postDate": "07/16/2020 16:16:21", "content": "efficientdet-5
\nlb = 0.7445
\n[update]
\nlb = 0.7504
Efficientdet-D7\nLB: 0.7209
", "rawMarkdown": "Efficientdet-D7\nLB: 0.7209", "votes": null } ], "comments": [ { "id": 930151, "author_name": "alexanderliao", "author_url": "", "post_date": "07/15/2020 08:26:44", "content": "[update] Mine is at 0.764 without pseudo-labeling (custom pipeline). My current score (keeps updating) is with pseudo-labeling. I also have a DETR model at CV around 0.57, but I'm not planning to submit it before the end of the competition.
", "votes": null, "replies": [ { "id": 930159, "author_name": "piantic", "author_url": "", "post_date": "07/15/2020 08:29:51", "content": "Thanks for share it!\n0.7621 without pseudo-labeling is amazing. 👍
", "votes": null, "replies": [] }, { "id": 930428, "author_name": "thefatcat", "author_url": "", "post_date": "07/15/2020 13:02:17", "content": "would you mind to share you model info?
", "votes": null, "replies": [] } ] }, { "id": 930397, "author_name": "", "author_url": "", "post_date": "07/15/2020 12:34:26", "content": "Hi @alexanderliao @piantic I've tried DETR for several days, but still hit no luck. My bbox_loss is oscillating drastically and periodically. Do you guys happen to come across such wired circumstances?
\n\n
I doubt that my input box format is wrong. I complied to the [xc, yc, w, h] / image_size format from the paper. How about yours? \nThanks you very much for your help!
I almost used the existing training kernel.\nhttps://www.kaggle.com/tanulsingh077/end-to-end-object-detection-with-transformers-detr
\n\nIt is a little difficult to confirm now. \nI'll check it as I continue experimenting.\n@stevenyin
", "votes": null, "replies": [] }, { "id": 930432, "author_name": "alexanderliao", "author_url": "", "post_date": "07/15/2020 13:06:38", "content": "I don't think DETR works for this competition. A nice concept though.
", "votes": null, "replies": [] }, { "id": 930439, "author_name": "", "author_url": "", "post_date": "07/15/2020 13:12:31", "content": "@piantic Yes, I started from that kernel and do some customization to the pipeline, which does not hit the luck so far. Thanks for you help!
", "votes": null, "replies": [] }, { "id": 930441, "author_name": "", "author_url": "", "post_date": "07/15/2020 13:14:01", "content": "Hi Liao, after reading the paper I begin to double about it as well. But can't get it started normally drives me a little annoyed...
", "votes": null, "replies": [] } ] }, { "id": 930870, "author_name": "matthewmasters", "author_url": "", "post_date": "07/15/2020 19:02:58", "content": "You can check out my existing threads for something along the same lines:\n- Best single model - No one has actually posted any YoloV5 scores here. So it is still relevant and I don't see the need to split the thread into a new one\n- Leaderboard of Compliant Submissions - Here I have compiled the top compliant scores into a leaderboard
", "votes": null, "replies": [] }, { "id": 930933, "author_name": "haoanwang", "author_url": "", "post_date": "07/15/2020 20:25:31", "content": "How about FasterRCNN? Is EfficientDet better than FasterRCNN?
", "votes": null, "replies": [] }, { "id": 932016, "author_name": "ajax0564", "author_url": "", "post_date": "07/16/2020 16:16:21", "content": "efficientdet-5
\nlb = 0.7445
\n[update]
\nlb = 0.7504
Efficientdet-D7\nLB: 0.7209
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "930094": "We have a licensing issue.\nAnd now it seems like we really need to share new information.\n\nI'm going to start over with models that fit the license rules. \nLet's share some information. 😃 \n\n# Baseline\n> \nModel: DETR\nEpoch: 50\nLB: 0.598\n\n\n# Updated\n> \nModel: EfficientDet-D6\nEpoch: 80\nLB: 0.717\n\n\np.s. DETR is released under the Apache 2.0 license.", "930151": "[update] Mine is at 0.764 without pseudo-labeling (custom pipeline). My current score (keeps updating) is with pseudo-labeling. I also have a DETR model at CV around 0.57, but I'm not planning to submit it before the end of the competition.", "930159": "Thanks for share it!\n0.7621 without pseudo-labeling is amazing. 👍", "930397": "Hi @alexanderliao @piantic I've tried DETR for several days, but still hit no luck. My bbox_loss is oscillating drastically and periodically. Do you guys happen to come across such wired circumstances? \n\n\n\nI doubt that my input box format is wrong. I complied to the `[xc, yc, w, h] / image_size` format from the paper. How about yours? \nThanks you very much for your help!", "930427": "I almost used the existing training kernel.\nhttps://www.kaggle.com/tanulsingh077/end-to-end-object-detection-with-transformers-detr\n\nIt is a little difficult to confirm now. \nI'll check it as I continue experimenting.\n@stevenyin", "930428": "would you mind to share you model info?", "930432": "I don't think DETR works for this competition. A nice concept though.", "930439": "piantic Yes, I started from that kernel and do some customization to the pipeline, which does not hit the luck so far. Thanks for you help!", "930441": "Hi Liao, after reading the paper I begin to double about it as well. But can't get it started normally drives me a little annoyed...", "930870": "You can check out my existing threads for something along the same lines:\n- [Best single model](https://www.kaggle.com/c/global-wheat-detection/discussion/153084) - No one has actually posted any YoloV5 scores here. So it is still relevant and I don't see the need to split the thread into a new one\n- [Leaderboard of Compliant Submissions](https://www.kaggle.com/c/global-wheat-detection/discussion/166793) - Here I have compiled the top compliant scores into a leaderboard", "930933": "How about FasterRCNN? Is EfficientDet better than FasterRCNN?", "932016": "efficientdet-5\nlb = 0.7445\n[update]\nlb = 0.7504", "936756": "Efficientdet-D7\nLB: 0.7209" }, "source": "meta" }