{ "id": 171965, "title": "what have u learnt from this competition ?( any experience)", "url": "/competitions/global-wheat-detection/discussion/171965", "author_name": "", "post_date": "2020-08-03T07:17:34.885250500Z", "votes": 13, "comment_count": 24, "views": 0, "content": "", "messages": [ { "id": "956033", "postDate": "08/03/2020 07:17:34", "content": "", "rawMarkdown": "", "votes": null }, { "id": "956133", "postDate": "08/03/2020 09:14:30", "content": "

My first object detection experience. Learned a lot)

", "rawMarkdown": "My first object detection experience. Learned a lot)", "votes": null }, { "id": "956482", "postDate": "08/03/2020 14:34:43", "content": "

If a competition has no valid leaderboard, if organizer's statement about how they deal with compliance issue is not clear enough, stay away from it.

", "rawMarkdown": "If a competition has no valid leaderboard, if organizer's statement about how they deal with compliance issue is not clear enough, stay away from it.", "votes": null }, { "id": "956748", "postDate": "08/03/2020 18:41:56", "content": "

I think price is not all. I learn more about this competition and got first team up working.

", "rawMarkdown": "I think price is not all. I learn more about this competition and got first team up working.", "votes": null }, { "id": "956763", "postDate": "08/03/2020 18:50:38", "content": "

In addition to many technical aspects of this competition, here's some other general takeaways:\n1. Always make sure to read the rules from the start! This has been an obvious pain in this competition due to the licensing issue. Determining from the start what is or isn't allowed will hopefully avoid these issues in the future.\n2. Make everything reproducible! One of my highest scores was attained without the inference notebook being 100% reproducible so I couldn't get the same score every time. Not only is it bad design but it may lead to not getting the prize you deserve in the end.\n3. Constantly experiment. I ended up waiting until towards the end of the competition to try out some of my experiments and it has put me on a real time crunch. Always be experimenting with your ideas throughout the competition.\n4. Only change one thing at a time! I'm at fault of this myself. It is much better to change only one variable at a time and see it's impact on CV/LB score than to make multiple changes and not sure how each of them contributes to the new score. If you add two great changes but one really bad one, your results might make you think they are all bad.

\n\n

I'm sure there's much more I could say but I will leave it at these main points

", "rawMarkdown": "In addition to many technical aspects of this competition, here's some other general takeaways:\n1. Always make sure to read the rules from the start! This has been an obvious pain in this competition due to the licensing issue. Determining from the start what is or isn't allowed will hopefully avoid these issues in the future.\n2. Make everything reproducible! One of my highest scores was attained without the inference notebook being 100% reproducible so I couldn't get the same score every time. Not only is it bad design but it may lead to not getting the prize you deserve in the end.\n3. Constantly experiment. I ended up waiting until towards the end of the competition to try out some of my experiments and it has put me on a real time crunch. Always be experimenting with your ideas throughout the competition.\n4. Only change one thing at a time! I'm at fault of this myself. It is much better to change only one variable at a time and see it's impact on CV/LB score than to make multiple changes and not sure how each of them contributes to the new score. If you add two great changes but one really bad one, your results might make you think they are all bad.\n\nI'm sure there's much more I could say but I will leave it at these main points", "votes": null }, { "id": "956803", "postDate": "08/03/2020 19:49:07", "content": "

Some great points there. I'm sure a lot of people agree with a lot of these.

", "rawMarkdown": "Some great points there. I'm sure a lot of people agree with a lot of these.", "votes": null }, { "id": "956925", "postDate": "08/03/2020 23:32:11", "content": "

I form a habit of waking up early :))

", "rawMarkdown": "I form a habit of waking up early :))", "votes": null }, { "id": "956967", "postDate": "08/04/2020 01:32:12", "content": "

Pseudo labels on new data > never seeing the data before even if the labels are not \"perfect\" the feature extractor will probably learning enough to improve overall result specially if test is from a different distribution ... self supervised learning is coming 🔥

", "rawMarkdown": "Pseudo labels on new data > never seeing the data before even if the labels are not \"perfect\" the feature extractor will probably learning enough to improve overall result specially if test is from a different distribution ... self supervised learning is coming 🔥", "votes": null }, { "id": "956980", "postDate": "08/04/2020 01:39:42", "content": "

PseudoLabeling | TTA | WBF | Gradient Accumulation | Mixed Precision | Mixup and cutmix | Trying different architectures.\n2 months full of learning and investigating different paths.

", "rawMarkdown": "PseudoLabeling | TTA | WBF | Gradient Accumulation | Mixed Precision | Mixup and cutmix | Trying different architectures.\n2 months full of learning and investigating different paths.", "votes": null }, { "id": "957223", "postDate": "08/04/2020 06:43:09", "content": "

I was exploring retinanet a few weeks before the competition started. So wanted to test it on this competition as the dataset was comparatively smaller than Coco. Some of the things I learned are

\n\n
    \n
  1. A strong data pipeline is important.
    2. \n
  2. Try out ideas every day, I left the competition in the middle as I could not find ways to improve the score.
    3. \n
  3. I was trying to stick with smaller models like resnet34, resnext50 till the last week of the competition. I should have started using larger models earlier.
    4. \n
  4. Build a strong CV as earlier as possible.
    5. \n
", "rawMarkdown": "I was exploring retinanet a few weeks before the competition started. So wanted to test it on this competition as the dataset was comparatively smaller than Coco. Some of the things I learned are\n\n1. A strong data pipeline is important.\n2. Try out ideas every day, I left the competition in the middle as I could not find ways to improve the score.\n3. I was trying to stick with smaller models like resnet34, resnext50 till the last week of the competition. I should have started using larger models earlier.\n4. Build a strong CV as earlier as possible.", "votes": null }, { "id": "957274", "postDate": "08/04/2020 07:29:32", "content": "

indeed. We should always fix seed while training experiments)

", "rawMarkdown": "indeed. We should always fix seed while training experiments)", "votes": null }, { "id": "957276", "postDate": "08/04/2020 07:31:11", "content": "

I`m afraid that models can overfit heavily to LB due to big public part of final dataset

", "rawMarkdown": "I`m afraid that models can overfit heavily to LB due to big public part of final dataset", "votes": null }, { "id": "957334", "postDate": "08/04/2020 08:21:42", "content": "

Learnt quite a lot actually like

\n\n
    \n
  1. How to handle overlapping bounding boxes using techniques like Non Maximum Suppression and Weighted Boxes Fusion (although slower than NMS for inference but overall an interesting approach)
    2. \n
  2. First time playing with albumentation package for incorporating data augmentation techniques using PyTorch. (Have only used ImageDataGenerator of Keras previously)
    3. \n
  3. Implementation of RetinaNet, EfficientNet and EfficientDet and their architectural differences
    4. \n
  4. Importance of using licensed packages or models
    5. \n
  5. Pseudo Labelling technique
    6. \n
", "rawMarkdown": "Learnt quite a lot actually like\n\n1. How to handle overlapping bounding boxes using techniques like Non Maximum Suppression and Weighted Boxes Fusion (although slower than NMS for inference but overall an interesting approach)\n2. First time playing with albumentation package for incorporating data augmentation techniques using PyTorch. (Have only used ImageDataGenerator of Keras previously)\n3. Implementation of RetinaNet, EfficientNet and EfficientDet and their architectural differences\n4. Importance of using licensed packages or models\n5. Pseudo Labelling technique", "votes": null }, { "id": "957536", "postDate": "08/04/2020 11:36:29", "content": "

Learnt some tricks and techniques which no course or a book could have taught.\nApart from this, it was great to connect with new team mates from across the globe who are experienced and got to learn a lot from their thought process, decision-making strategies etc...

", "rawMarkdown": "Learnt some tricks and techniques which no course or a book could have taught.\nApart from this, it was great to connect with new team mates from across the globe who are experienced and got to learn a lot from their thought process, decision-making strategies etc...", "votes": null }, { "id": "957788", "postDate": "08/04/2020 15:03:16", "content": "

First competition for me! I only started this competition about a week and a half ago, so I lost out on a ton of time. I learned so much, and what a great community Kaggle is. Firstly, data is king. I spent way too much time tuning hyperparameters without a robust validation set, so many models were a shot in the dark. I also learned the importance of data augmentation and TTA, especially when pseudolabelling. Ultimately, I ran out of ideas for how to improve my yolov4 pseudolabelling kernel, and not having enough time to make a robust val set, resorted to pseudolabelling. I got a score of 0.7322, which I think is alright. Learned a ton about teamwork also. I wish I had time to ensemble with my team (who were using effdet), but it is what it is. I'm sure I'll learn even more looking at the solutions of top performers, especially Tereka. No clue how Tereka is so far ahead of other compliant submissions, and even yolov5 submissions!

", "rawMarkdown": "First competition for me! I only started this competition about a week and a half ago, so I lost out on a ton of time. I learned so much, and what a great community Kaggle is. Firstly, data is king. I spent way too much time tuning hyperparameters without a robust validation set, so many models were a shot in the dark. I also learned the importance of data augmentation and TTA, especially when pseudolabelling. Ultimately, I ran out of ideas for how to improve my yolov4 pseudolabelling kernel, and not having enough time to make a robust val set, resorted to pseudolabelling. I got a score of 0.7322, which I think is alright. Learned a ton about teamwork also. I wish I had time to ensemble with my team (who were using effdet), but it is what it is. I'm sure I'll learn even more looking at the solutions of top performers, especially Tereka. No clue how Tereka is so far ahead of other compliant submissions, and even yolov5 submissions!", "votes": null }, { "id": "957944", "postDate": "08/04/2020 16:36:08", "content": "
    \n
  1. Good understanding of Object Detection in General and models like Yolo & EffDet
    2. \n
  2. Power of Augmentations during Training and Test Time(TTA). Albumentations is awesome
    3. \n
  3. Handling very large image sizes using Gradient Accumulation & Mixed Precision.
    4. \n
  4. A better intuition of how training time and model size affects eventual results.
    5. \n
", "rawMarkdown": "1. Good understanding of Object Detection in General and models like Yolo & EffDet \n2. Power of Augmentations during Training and Test Time(TTA). Albumentations is awesome\n3. Handling very large image sizes using Gradient Accumulation & Mixed Precision.\n4. A better intuition of how training time and model size affects eventual results.", "votes": null }, { "id": "957996", "postDate": "08/04/2020 17:20:37", "content": "

I've learnt a lot about object detection. It was a very valuable experience for me. It was my first experience with Pseudo-labeling and data cleaning as well. Thanks a lot to the authors of the public notebooks. By the way, it was my first Kaggle competition but not the first experience in ML. I hope to get a good result but in any case, I have a good time.

", "rawMarkdown": "I've learnt a lot about object detection. It was a very valuable experience for me. It was my first experience with Pseudo-labeling and data cleaning as well. Thanks a lot to the authors of the public notebooks. By the way, it was my first Kaggle competition but not the first experience in ML. I hope to get a good result but in any case, I have a good time.", "votes": null }, { "id": "958251", "postDate": "08/04/2020 21:53:05", "content": "

Don't participate in a poorly labeled competition in the first place....

\n\n

But, in all seriousness, I will probably post something in the upcoming days.

", "rawMarkdown": "Don't participate in a poorly labeled competition in the first place....\n\nBut, in all seriousness, I will probably post something in the upcoming days.", "votes": null }, { "id": "958260", "postDate": "08/04/2020 22:04:20", "content": "

Good luck with however the shakeup / YoloV5 removal plays out. What you have shared has definitely added to the experience of this competition. It would be great to read about your solution and what you learnt.

", "rawMarkdown": "Good luck with however the shakeup / YoloV5 removal plays out. What you have shared has definitely added to the experience of this competition. It would be great to read about your solution and what you learnt.", "votes": null }, { "id": "958295", "postDate": "08/04/2020 23:08:46", "content": "

peace & love

", "rawMarkdown": "peace & love", "votes": null }, { "id": "958417", "postDate": "08/05/2020 01:19:39", "content": "

This is my first experience in Object Detection.

\n\n", "rawMarkdown": "This is my first experience in Object Detection.\n\n- Learned the evolution and ideas of different algorithms, including Faster RCNN, YOLO, and latest state-of-arts\n- Learned the nice code shared by awesome kagglers. (Although YOLOv5 is not a compliant solution but the code is really well written.)\n- Looking forward to seeing the top solutions!", "votes": null }, { "id": "958466", "postDate": "08/05/2020 02:25:33", "content": "

LOL, thanks nvnn!!

", "rawMarkdown": "LOL, thanks nvnn!!", "votes": null }, { "id": "958898", "postDate": "08/05/2020 07:52:50", "content": "

The differences between different open source licenses. XD

", "rawMarkdown": "The differences between different open source licenses. XD", "votes": null }, { "id": "958907", "postDate": "08/05/2020 07:59:08", "content": "

😂 Absolutely, \nI never knew there were so many licenses before this comp

", "rawMarkdown": "😂 Absolutely, \nI never knew there were so many licenses before this comp", "votes": null }, { "id": "959398", "postDate": "08/05/2020 14:53:02", "content": "

That object detection tasks are not as sophisticated as I thought\nTools: pytorch, Colab\nModels: resnext, efficientDet, yolo\nAlso, advanced augmentations like cutmix and others\nSome training notebooks were very useful, thanks Kaggle community

", "rawMarkdown": "That object detection tasks are not as sophisticated as I thought\nTools: pytorch, Colab\nModels: resnext, efficientDet, yolo\nAlso, advanced augmentations like cutmix and others\nSome training notebooks were very useful, thanks Kaggle community", "votes": null } ], "comments": [ { "id": 956133, "author_name": "kupchanski", "author_url": "", "post_date": "08/03/2020 09:14:30", "content": "

My first object detection experience. Learned a lot)

", "votes": null, "replies": [] }, { "id": 956482, "author_name": "jielu0728", "author_url": "", "post_date": "08/03/2020 14:34:43", "content": "

If a competition has no valid leaderboard, if organizer's statement about how they deal with compliance issue is not clear enough, stay away from it.

", "votes": null, "replies": [ { "id": 956748, "author_name": "doanquanvietnamca", "author_url": "", "post_date": "08/03/2020 18:41:56", "content": "

I think price is not all. I learn more about this competition and got first team up working.

", "votes": null, "replies": [] } ] }, { "id": 956763, "author_name": "matthewmasters", "author_url": "", "post_date": "08/03/2020 18:50:38", "content": "

In addition to many technical aspects of this competition, here's some other general takeaways:\n1. Always make sure to read the rules from the start! This has been an obvious pain in this competition due to the licensing issue. Determining from the start what is or isn't allowed will hopefully avoid these issues in the future.\n2. Make everything reproducible! One of my highest scores was attained without the inference notebook being 100% reproducible so I couldn't get the same score every time. Not only is it bad design but it may lead to not getting the prize you deserve in the end.\n3. Constantly experiment. I ended up waiting until towards the end of the competition to try out some of my experiments and it has put me on a real time crunch. Always be experimenting with your ideas throughout the competition.\n4. Only change one thing at a time! I'm at fault of this myself. It is much better to change only one variable at a time and see it's impact on CV/LB score than to make multiple changes and not sure how each of them contributes to the new score. If you add two great changes but one really bad one, your results might make you think they are all bad.

\n\n

I'm sure there's much more I could say but I will leave it at these main points

", "votes": null, "replies": [ { "id": 956803, "author_name": "datahobbit", "author_url": "", "post_date": "08/03/2020 19:49:07", "content": "

Some great points there. I'm sure a lot of people agree with a lot of these.

", "votes": null, "replies": [] }, { "id": 957274, "author_name": "kupchanski", "author_url": "", "post_date": "08/04/2020 07:29:32", "content": "

indeed. We should always fix seed while training experiments)

", "votes": null, "replies": [] } ] }, { "id": 956925, "author_name": "nvnnghia", "author_url": "", "post_date": "08/03/2020 23:32:11", "content": "

I form a habit of waking up early :))

", "votes": null, "replies": [ { "id": 958466, "author_name": "superkevingit", "author_url": "", "post_date": "08/05/2020 02:25:33", "content": "

LOL, thanks nvnn!!

", "votes": null, "replies": [] } ] }, { "id": 956967, "author_name": "ma7moud", "author_url": "", "post_date": "08/04/2020 01:32:12", "content": "

Pseudo labels on new data > never seeing the data before even if the labels are not \"perfect\" the feature extractor will probably learning enough to improve overall result specially if test is from a different distribution ... self supervised learning is coming 🔥

", "votes": null, "replies": [ { "id": 957276, "author_name": "kupchanski", "author_url": "", "post_date": "08/04/2020 07:31:11", "content": "

I`m afraid that models can overfit heavily to LB due to big public part of final dataset

", "votes": null, "replies": [] } ] }, { "id": 956980, "author_name": "nacirbouazizi", "author_url": "", "post_date": "08/04/2020 01:39:42", "content": "

PseudoLabeling | TTA | WBF | Gradient Accumulation | Mixed Precision | Mixup and cutmix | Trying different architectures.\n2 months full of learning and investigating different paths.

", "votes": null, "replies": [] }, { "id": 957223, "author_name": "vishnus", "author_url": "", "post_date": "08/04/2020 06:43:09", "content": "

I was exploring retinanet a few weeks before the competition started. So wanted to test it on this competition as the dataset was comparatively smaller than Coco. Some of the things I learned are

\n\n
    \n
  1. A strong data pipeline is important.
    2. \n
  2. Try out ideas every day, I left the competition in the middle as I could not find ways to improve the score.
    3. \n
  3. I was trying to stick with smaller models like resnet34, resnext50 till the last week of the competition. I should have started using larger models earlier.
    4. \n
  4. Build a strong CV as earlier as possible.
    5. \n
", "votes": null, "replies": [] }, { "id": 957334, "author_name": "aravindram11", "author_url": "", "post_date": "08/04/2020 08:21:42", "content": "

Learnt quite a lot actually like

\n\n
    \n
  1. How to handle overlapping bounding boxes using techniques like Non Maximum Suppression and Weighted Boxes Fusion (although slower than NMS for inference but overall an interesting approach)
    2. \n
  2. First time playing with albumentation package for incorporating data augmentation techniques using PyTorch. (Have only used ImageDataGenerator of Keras previously)
    3. \n
  3. Implementation of RetinaNet, EfficientNet and EfficientDet and their architectural differences
    4. \n
  4. Importance of using licensed packages or models
    5. \n
  5. Pseudo Labelling technique
    6. \n
", "votes": null, "replies": [] }, { "id": 957536, "author_name": "nitindatta", "author_url": "", "post_date": "08/04/2020 11:36:29", "content": "

Learnt some tricks and techniques which no course or a book could have taught.\nApart from this, it was great to connect with new team mates from across the globe who are experienced and got to learn a lot from their thought process, decision-making strategies etc...

", "votes": null, "replies": [] }, { "id": 957788, "author_name": "stanleyjzheng", "author_url": "", "post_date": "08/04/2020 15:03:16", "content": "

First competition for me! I only started this competition about a week and a half ago, so I lost out on a ton of time. I learned so much, and what a great community Kaggle is. Firstly, data is king. I spent way too much time tuning hyperparameters without a robust validation set, so many models were a shot in the dark. I also learned the importance of data augmentation and TTA, especially when pseudolabelling. Ultimately, I ran out of ideas for how to improve my yolov4 pseudolabelling kernel, and not having enough time to make a robust val set, resorted to pseudolabelling. I got a score of 0.7322, which I think is alright. Learned a ton about teamwork also. I wish I had time to ensemble with my team (who were using effdet), but it is what it is. I'm sure I'll learn even more looking at the solutions of top performers, especially Tereka. No clue how Tereka is so far ahead of other compliant submissions, and even yolov5 submissions!

", "votes": null, "replies": [] }, { "id": 957944, "author_name": "watzisname", "author_url": "", "post_date": "08/04/2020 16:36:08", "content": "
    \n
  1. Good understanding of Object Detection in General and models like Yolo & EffDet
    2. \n
  2. Power of Augmentations during Training and Test Time(TTA). Albumentations is awesome
    3. \n
  3. Handling very large image sizes using Gradient Accumulation & Mixed Precision.
    4. \n
  4. A better intuition of how training time and model size affects eventual results.
    5. \n
", "votes": null, "replies": [] }, { "id": 957996, "author_name": "vadimtimakin", "author_url": "", "post_date": "08/04/2020 17:20:37", "content": "

I've learnt a lot about object detection. It was a very valuable experience for me. It was my first experience with Pseudo-labeling and data cleaning as well. Thanks a lot to the authors of the public notebooks. By the way, it was my first Kaggle competition but not the first experience in ML. I hope to get a good result but in any case, I have a good time.

", "votes": null, "replies": [] }, { "id": 958251, "author_name": "alexanderliao", "author_url": "", "post_date": "08/04/2020 21:53:05", "content": "

Don't participate in a poorly labeled competition in the first place....

\n\n

But, in all seriousness, I will probably post something in the upcoming days.

", "votes": null, "replies": [ { "id": 958260, "author_name": "datahobbit", "author_url": "", "post_date": "08/04/2020 22:04:20", "content": "

Good luck with however the shakeup / YoloV5 removal plays out. What you have shared has definitely added to the experience of this competition. It would be great to read about your solution and what you learnt.

", "votes": null, "replies": [] } ] }, { "id": 958295, "author_name": "creatrol", "author_url": "", "post_date": "08/04/2020 23:08:46", "content": "

peace & love

", "votes": null, "replies": [] }, { "id": 958417, "author_name": "gxygomes", "author_url": "", "post_date": "08/05/2020 01:19:39", "content": "

This is my first experience in Object Detection.

\n\n", "votes": null, "replies": [] }, { "id": 958898, "author_name": "qingyuanwang", "author_url": "", "post_date": "08/05/2020 07:52:50", "content": "

The differences between different open source licenses. XD

", "votes": null, "replies": [ { "id": 958907, "author_name": "nitindatta", "author_url": "", "post_date": "08/05/2020 07:59:08", "content": "

😂 Absolutely, \nI never knew there were so many licenses before this comp

", "votes": null, "replies": [] } ] }, { "id": 959398, "author_name": "alturutin", "author_url": "", "post_date": "08/05/2020 14:53:02", "content": "

That object detection tasks are not as sophisticated as I thought\nTools: pytorch, Colab\nModels: resnext, efficientDet, yolo\nAlso, advanced augmentations like cutmix and others\nSome training notebooks were very useful, thanks Kaggle community

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "956033": "", "956133": "My first object detection experience. Learned a lot)", "956482": "If a competition has no valid leaderboard, if organizer's statement about how they deal with compliance issue is not clear enough, stay away from it.", "956748": "I think price is not all. I learn more about this competition and got first team up working.", "956763": "In addition to many technical aspects of this competition, here's some other general takeaways:\n1. Always make sure to read the rules from the start! This has been an obvious pain in this competition due to the licensing issue. Determining from the start what is or isn't allowed will hopefully avoid these issues in the future.\n2. Make everything reproducible! One of my highest scores was attained without the inference notebook being 100% reproducible so I couldn't get the same score every time. Not only is it bad design but it may lead to not getting the prize you deserve in the end.\n3. Constantly experiment. I ended up waiting until towards the end of the competition to try out some of my experiments and it has put me on a real time crunch. Always be experimenting with your ideas throughout the competition.\n4. Only change one thing at a time! I'm at fault of this myself. It is much better to change only one variable at a time and see it's impact on CV/LB score than to make multiple changes and not sure how each of them contributes to the new score. If you add two great changes but one really bad one, your results might make you think they are all bad.\n\nI'm sure there's much more I could say but I will leave it at these main points", "956803": "Some great points there. I'm sure a lot of people agree with a lot of these.", "956925": "I form a habit of waking up early :))", "956967": "Pseudo labels on new data > never seeing the data before even if the labels are not \"perfect\" the feature extractor will probably learning enough to improve overall result specially if test is from a different distribution ... self supervised learning is coming 🔥", "956980": "PseudoLabeling | TTA | WBF | Gradient Accumulation | Mixed Precision | Mixup and cutmix | Trying different architectures.\n2 months full of learning and investigating different paths.", "957223": "I was exploring retinanet a few weeks before the competition started. So wanted to test it on this competition as the dataset was comparatively smaller than Coco. Some of the things I learned are\n\n1. A strong data pipeline is important.\n2. Try out ideas every day, I left the competition in the middle as I could not find ways to improve the score.\n3. I was trying to stick with smaller models like resnet34, resnext50 till the last week of the competition. I should have started using larger models earlier.\n4. Build a strong CV as earlier as possible.", "957274": "indeed. We should always fix seed while training experiments)", "957276": "I`m afraid that models can overfit heavily to LB due to big public part of final dataset", "957334": "Learnt quite a lot actually like\n\n1. How to handle overlapping bounding boxes using techniques like Non Maximum Suppression and Weighted Boxes Fusion (although slower than NMS for inference but overall an interesting approach)\n2. First time playing with albumentation package for incorporating data augmentation techniques using PyTorch. (Have only used ImageDataGenerator of Keras previously)\n3. Implementation of RetinaNet, EfficientNet and EfficientDet and their architectural differences\n4. Importance of using licensed packages or models\n5. Pseudo Labelling technique", "957536": "Learnt some tricks and techniques which no course or a book could have taught.\nApart from this, it was great to connect with new team mates from across the globe who are experienced and got to learn a lot from their thought process, decision-making strategies etc...", "957788": "First competition for me! I only started this competition about a week and a half ago, so I lost out on a ton of time. I learned so much, and what a great community Kaggle is. Firstly, data is king. I spent way too much time tuning hyperparameters without a robust validation set, so many models were a shot in the dark. I also learned the importance of data augmentation and TTA, especially when pseudolabelling. Ultimately, I ran out of ideas for how to improve my yolov4 pseudolabelling kernel, and not having enough time to make a robust val set, resorted to pseudolabelling. I got a score of 0.7322, which I think is alright. Learned a ton about teamwork also. I wish I had time to ensemble with my team (who were using effdet), but it is what it is. I'm sure I'll learn even more looking at the solutions of top performers, especially Tereka. No clue how Tereka is so far ahead of other compliant submissions, and even yolov5 submissions!", "957944": "1. Good understanding of Object Detection in General and models like Yolo & EffDet \n2. Power of Augmentations during Training and Test Time(TTA). Albumentations is awesome\n3. Handling very large image sizes using Gradient Accumulation & Mixed Precision.\n4. A better intuition of how training time and model size affects eventual results.", "957996": "I've learnt a lot about object detection. It was a very valuable experience for me. It was my first experience with Pseudo-labeling and data cleaning as well. Thanks a lot to the authors of the public notebooks. By the way, it was my first Kaggle competition but not the first experience in ML. I hope to get a good result but in any case, I have a good time.", "958251": "Don't participate in a poorly labeled competition in the first place....\n\nBut, in all seriousness, I will probably post something in the upcoming days.", "958260": "Good luck with however the shakeup / YoloV5 removal plays out. What you have shared has definitely added to the experience of this competition. It would be great to read about your solution and what you learnt.", "958295": "peace & love", "958417": "This is my first experience in Object Detection.\n\n- Learned the evolution and ideas of different algorithms, including Faster RCNN, YOLO, and latest state-of-arts\n- Learned the nice code shared by awesome kagglers. (Although YOLOv5 is not a compliant solution but the code is really well written.)\n- Looking forward to seeing the top solutions!", "958466": "LOL, thanks nvnn!!", "958898": "The differences between different open source licenses. XD", "958907": "😂 Absolutely, \nI never knew there were so many licenses before this comp", "959398": "That object detection tasks are not as sophisticated as I thought\nTools: pytorch, Colab\nModels: resnext, efficientDet, yolo\nAlso, advanced augmentations like cutmix and others\nSome training notebooks were very useful, thanks Kaggle community" }, "source": "meta" }