TL;DR: \n- I'm using mmdetection framework and it's really convenient;\n- Best single model (cascade rcnn with imagenet pretrained resnext101) + TTA (horizontal flip, multi-scale testing (600, 900), (800, 1200), (1000, 1500), (1200, 1800), (1400,\n2100)) achieves 0.499 public;\n- Split datasets into 6 subsets by frequency and then finetune on them for 1-2 epochs. --> appr. 0.05+ increase;\n- Parent class expansion gives appr. 0.01+ increase;\n- Weighted ensemble (from ZFTurbo's solution last year) of all my high score submissions --> 0.608 final public score.
\n\n1. Single model\nAt the very beginning, I was going to attend the \"Visual Relationship Detection\" track. But then I realized that I didn't have a good object detection model for that one. So I started with faster rcnn+resnext101, it takes me about 20 days to train 24 epochs and results in 0.446 on public lb.\nSimilarly, I trained cascade rcnn+resnext101, cascade rcnn+senet154 for 12 and 8 epochs respectively.\nI just leave these models training for several weeks, do my daily work and give up \"visual relation detection\". \nThe best single model is cascade rcnn+resnext101, which was accidently trained for 19 epochs (6 epochs longer than planned). So maybe I should train longer for each model :).\nConclusion: my single models are weak. They should be trained longer.
\n\n2. Finetune\nSince the classes are very unbalanced, I split the dataset classes into 6 subsets simply according to frequency and finetune on them using faster rcnn+resnext101 model:\n- Classes 0-50, appr. 1411368images, 2 epochs, lr 0.001\n- Classes 51-100, appr. 308352 images, 2 epochs, lr 0.001\n- Classes 101-200: appr. 208096 images, 2 epochs, lr 0.001\n- Classes 201-300, appr. 93140 images, 2epochs, lr 0.001\n- Classes 301-400, appr. 45840 images, 1epoch, lr 0.001\n- Classes 401-500, appr. 19316 images, 1epoch, lr 0.001\nLast two weeks before final deadline, I found one huge bug in my code.\nAfter solving this bug, ensembling the predictions of finetuned models gave appr. 0.05+ increase on public LB.\nConclusion: solving class imbalance problem is the key to top silver or gold medal.
\n\n3. TTA and Final Ensemble\nSome teams merged and I may have the chance for solo gold.\nSo I did the followings:\n- TTA for each model: horizontal flip, multi-scale testing with (600, 900), (800, 1200), (1000, 1500), (1200, 1800), (1400, 2100) image size;\n- Expand parent class for each prediction after inference.\n- Weighted ensemble (from ZFTurbo's solution last year) of all models.
\n\nOther tricks:\n- Increase the number of limited boxes for each image, even though they are of low confidence. I choose 600 as upper limit in my final submission.\n- I wasn't able to finish 12 epochs for my cascade rcnn+senet154 model. Nevertheless, ensembling it gives slight improvement.
\n\nNot work for me:\nSoft-NMS: tried to use it for ensemble, wasn't going well.
\n\nPlanned but not implemented:\n- Use mask annotations from the segmentation track;\n- Multi-scale training and mixup augmentation.
\n\nFinal words:\nLet's play fair. \nPeace&love. 👍
", "messages": [ { "id": 640985, "postDate": "2019-10-04T11:14:07.517Z", "content": "TL;DR: \n- I'm using mmdetection framework and it's really convenient;\n- Best single model (cascade rcnn with imagenet pretrained resnext101) + TTA (horizontal flip, multi-scale testing (600, 900), (800, 1200), (1000, 1500), (1200, 1800), (1400,\n2100)) achieves 0.499 public;\n- Split datasets into 6 subsets by frequency and then finetune on them for 1-2 epochs. --> appr. 0.05+ increase;\n- Parent class expansion gives appr. 0.01+ increase;\n- Weighted ensemble (from ZFTurbo's solution last year) of all my high score submissions --> 0.608 final public score.
\n\n1. Single model\nAt the very beginning, I was going to attend the \"Visual Relationship Detection\" track. But then I realized that I didn't have a good object detection model for that one. So I started with faster rcnn+resnext101, it takes me about 20 days to train 24 epochs and results in 0.446 on public lb.\nSimilarly, I trained cascade rcnn+resnext101, cascade rcnn+senet154 for 12 and 8 epochs respectively.\nI just leave these models training for several weeks, do my daily work and give up \"visual relation detection\". \nThe best single model is cascade rcnn+resnext101, which was accidently trained for 19 epochs (6 epochs longer than planned). So maybe I should train longer for each model :).\nConclusion: my single models are weak. They should be trained longer.
\n\n2. Finetune\nSince the classes are very unbalanced, I split the dataset classes into 6 subsets simply according to frequency and finetune on them using faster rcnn+resnext101 model:\n- Classes 0-50, appr. 1411368images, 2 epochs, lr 0.001\n- Classes 51-100, appr. 308352 images, 2 epochs, lr 0.001\n- Classes 101-200: appr. 208096 images, 2 epochs, lr 0.001\n- Classes 201-300, appr. 93140 images, 2epochs, lr 0.001\n- Classes 301-400, appr. 45840 images, 1epoch, lr 0.001\n- Classes 401-500, appr. 19316 images, 1epoch, lr 0.001\nLast two weeks before final deadline, I found one huge bug in my code.\nAfter solving this bug, ensembling the predictions of finetuned models gave appr. 0.05+ increase on public LB.\nConclusion: solving class imbalance problem is the key to top silver or gold medal.
\n\n3. TTA and Final Ensemble\nSome teams merged and I may have the chance for solo gold.\nSo I did the followings:\n- TTA for each model: horizontal flip, multi-scale testing with (600, 900), (800, 1200), (1000, 1500), (1200, 1800), (1400, 2100) image size;\n- Expand parent class for each prediction after inference.\n- Weighted ensemble (from ZFTurbo's solution last year) of all models.
\n\nOther tricks:\n- Increase the number of limited boxes for each image, even though they are of low confidence. I choose 600 as upper limit in my final submission.\n- I wasn't able to finish 12 epochs for my cascade rcnn+senet154 model. Nevertheless, ensembling it gives slight improvement.
\n\nNot work for me:\nSoft-NMS: tried to use it for ensemble, wasn't going well.
\n\nPlanned but not implemented:\n- Use mask annotations from the segmentation track;\n- Multi-scale training and mixup augmentation.
\n\nFinal words:\nLet's play fair. \nPeace&love. 👍
", "rawMarkdown": "**TL;DR**: \n- I'm using mmdetection framework and it's really convenient;\n- Best single model (cascade rcnn with imagenet pretrained resnext101) + TTA (horizontal flip, multi-scale testing (600, 900), (800, 1200), (1000, 1500), (1200, 1800), (1400,\n2100)) achieves 0.499 public;\n- Split datasets into 6 subsets by frequency and then finetune on them for 1-2 epochs. --> appr. 0.05+ increase;\n- Parent class expansion gives appr. 0.01+ increase;\n- Weighted ensemble (from [ZFTurbo's solution last year](https://www.kaggle.com/c/google-ai-open-images-object-detection-track/discussion/64633#latest-590283)) of all my high score submissions --> 0.608 final public score.\n\n**1. Single model**\nAt the very beginning, I was going to attend the \"Visual Relationship Detection\" track. But then I realized that I didn't have a good object detection model for that one. So I started with faster rcnn+resnext101, it takes me about 20 days to train 24 epochs and results in 0.446 on public lb.\nSimilarly, I trained cascade rcnn+resnext101, cascade rcnn+senet154 for 12 and 8 epochs respectively.\nI just leave these models training for several weeks, do my daily work and give up \"visual relation detection\". \nThe best single model is cascade rcnn+resnext101, which was accidently trained for 19 epochs (6 epochs longer than planned). So maybe I should train longer for each model :).\n**Conclusion**: my single models are weak. They should be trained longer.\n\n**2. Finetune**\nSince the classes are very unbalanced, I split the dataset classes into 6 subsets simply according to frequency and finetune on them using faster rcnn+resnext101 model:\n- Classes 0-50, appr. 1411368images, 2 epochs, lr 0.001\n- Classes 51-100, appr. 308352 images, 2 epochs, lr 0.001\n- Classes 101-200: appr. 208096 images, 2 epochs, lr 0.001\n- Classes 201-300, appr. 93140 images, 2epochs, lr 0.001\n- Classes 301-400, appr. 45840 images, 1epoch, lr 0.001\n- Classes 401-500, appr. 19316 images, 1epoch, lr 0.001\nLast two weeks before final deadline, I found one huge bug in my code.\nAfter solving this bug, ensembling the predictions of finetuned models gave appr. 0.05+ increase on public LB.\n**Conclusion**: solving class imbalance problem is the key to top silver or gold medal.\n\n**3. TTA and Final Ensemble**\nSome teams merged and I may have the chance for solo gold.\nSo I did the followings:\n- TTA for each model: horizontal flip, multi-scale testing with (600, 900), (800, 1200), (1000, 1500), (1200, 1800), (1400, 2100) image size;\n- Expand parent class for each prediction after inference.\n- Weighted ensemble (from [ZFTurbo's solution last year](https://www.kaggle.com/c/google-ai-open-images-object-detection-track/discussion/64633#latest-590283)) of all models.\n\n**Other tricks:**\n- Increase the number of limited boxes for each image, even though they are of low confidence. I choose 600 as upper limit in my final submission.\n- I wasn't able to finish 12 epochs for my cascade rcnn+senet154 model. Nevertheless, ensembling it gives slight improvement.\n\n**Not work for me:**\nSoft-NMS: tried to use it for ensemble, wasn't going well. \n\n**Planned but not implemented:**\n- Use mask annotations from the segmentation track;\n- Multi-scale training and mixup augmentation.\n\n**Final words:**\nLet's play fair. \nPeace&love. 👍 ", "votes": 16 }, { "id": 644368, "postDate": "2019-10-08T17:34:49.070Z", "content": "Thanks for sharing your approach. What were the hardware specs you used and how much time did you train for?
", "rawMarkdown": "Thanks for sharing your approach. What were the hardware specs you used and how much time did you train for? " }, { "id": 641896, "postDate": "2019-10-05T09:21:34.773Z", "content": "Thanks and congrats! So where was the huge bug? :)
", "rawMarkdown": "Thanks and congrats! So where was the huge bug? :)" }, { "id": 641027, "postDate": "2019-10-04T11:52:59.340Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 644368, "author_name": "Rohit Midha", "author_url": "", "post_date": "2019-10-08T17:34:49.070000", "content": "Thanks for sharing your approach. What were the hardware specs you used and how much time did you train for?
", "votes": 0, "replies": [] }, { "id": 641896, "author_name": "Artyom Palvelev", "author_url": "", "post_date": "2019-10-05T09:21:34.773000", "content": "Thanks and congrats! So where was the huge bug? :)
", "votes": 0, "replies": [] }, { "id": 641027, "author_name": "", "author_url": "", "post_date": "2019-10-04T11:52:59.340000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "640985": "**TL;DR**: \n- I'm using mmdetection framework and it's really convenient;\n- Best single model (cascade rcnn with imagenet pretrained resnext101) + TTA (horizontal flip, multi-scale testing (600, 900), (800, 1200), (1000, 1500), (1200, 1800), (1400,\n2100)) achieves 0.499 public;\n- Split datasets into 6 subsets by frequency and then finetune on them for 1-2 epochs. --> appr. 0.05+ increase;\n- Parent class expansion gives appr. 0.01+ increase;\n- Weighted ensemble (from [ZFTurbo's solution last year](https://www.kaggle.com/c/google-ai-open-images-object-detection-track/discussion/64633#latest-590283)) of all my high score submissions --> 0.608 final public score.\n\n**1. Single model**\nAt the very beginning, I was going to attend the \"Visual Relationship Detection\" track. But then I realized that I didn't have a good object detection model for that one. So I started with faster rcnn+resnext101, it takes me about 20 days to train 24 epochs and results in 0.446 on public lb.\nSimilarly, I trained cascade rcnn+resnext101, cascade rcnn+senet154 for 12 and 8 epochs respectively.\nI just leave these models training for several weeks, do my daily work and give up \"visual relation detection\". \nThe best single model is cascade rcnn+resnext101, which was accidently trained for 19 epochs (6 epochs longer than planned). So maybe I should train longer for each model :).\n**Conclusion**: my single models are weak. They should be trained longer.\n\n**2. Finetune**\nSince the classes are very unbalanced, I split the dataset classes into 6 subsets simply according to frequency and finetune on them using faster rcnn+resnext101 model:\n- Classes 0-50, appr. 1411368images, 2 epochs, lr 0.001\n- Classes 51-100, appr. 308352 images, 2 epochs, lr 0.001\n- Classes 101-200: appr. 208096 images, 2 epochs, lr 0.001\n- Classes 201-300, appr. 93140 images, 2epochs, lr 0.001\n- Classes 301-400, appr. 45840 images, 1epoch, lr 0.001\n- Classes 401-500, appr. 19316 images, 1epoch, lr 0.001\nLast two weeks before final deadline, I found one huge bug in my code.\nAfter solving this bug, ensembling the predictions of finetuned models gave appr. 0.05+ increase on public LB.\n**Conclusion**: solving class imbalance problem is the key to top silver or gold medal.\n\n**3. TTA and Final Ensemble**\nSome teams merged and I may have the chance for solo gold.\nSo I did the followings:\n- TTA for each model: horizontal flip, multi-scale testing with (600, 900), (800, 1200), (1000, 1500), (1200, 1800), (1400, 2100) image size;\n- Expand parent class for each prediction after inference.\n- Weighted ensemble (from [ZFTurbo's solution last year](https://www.kaggle.com/c/google-ai-open-images-object-detection-track/discussion/64633#latest-590283)) of all models.\n\n**Other tricks:**\n- Increase the number of limited boxes for each image, even though they are of low confidence. I choose 600 as upper limit in my final submission.\n- I wasn't able to finish 12 epochs for my cascade rcnn+senet154 model. Nevertheless, ensembling it gives slight improvement.\n\n**Not work for me:**\nSoft-NMS: tried to use it for ensemble, wasn't going well. \n\n**Planned but not implemented:**\n- Use mask annotations from the segmentation track;\n- Multi-scale training and mixup augmentation.\n\n**Final words:**\nLet's play fair. \nPeace&love. 👍 ", "644368": "Thanks for sharing your approach. What were the hardware specs you used and how much time did you train for? ", "641896": "Thanks and congrats! So where was the huge bug? :)", "641027": "" } }