{ "id": 320051, "title": "Hierarchical learning solution", "url": "/competitions/happy-whale-and-dolphin/discussion/320051", "author_name": "Vsst", "post_date": "2022-04-19T20:25:40.628000", "votes": 10, "comment_count": 0, "views": 0, "content": "

First of all, I would like to say congratulations to all the winners!šŸŽ‰šŸ”„ You did a good job with such a complex dataset!

\n

Although I didnā€™t do well in this competition (4 places to bronzešŸ˜©), I would like to share my solution. I think this information may be useful for further competitions.

\n

Architecture
\nAt the beginning of the competition I was interested in finding neural net architectures, which outperform the baseline effnet. I was sure that there are better architectures (Well, in the end I found that the differences are not so drastic). I stopped at the effnetv2_m/effnet_b6/convneXt_m in my final ensemble.
\nWhen I saw the data the first idea came to my mind was, that it is a hierarchical learning competition. After some search I found this interesting article: Your \"Flamingo\" is My \"Bird\": Fine-Grained, or Not. This article is applied to a classification problem, but I suggested it will work well in re-identification either (I modified an article approach a little making the individual embedding part larger than species). The final architecture of my solution you can see in the picture below.
\n\"arch1.png\"

\n

Augmentations
\n 2.1 Shear scale rotate
\nFrom the dataset you can see that there are a lot of different viewpoints for one whale. Using this augmentation I tried to simulate them.
\n 2.2 CLAHE at test and train time before all augmentation
\nThe purpose of using this augmentation is that some back fins are under water, some are darkened.
\n 2.3 Local grayscale transformation
\nThis transformation is based on this article, which gave a little boost to LB
\n 2.4 Color/brightness/hue/horizontal flip

\n

Dataset
\nFor all novice kagglers I would like to say that dataset is crucially important. I didnā€™t pay a lot of attention to it, but when I switched from detic to backfins dataset, the score was boosted a lot.

\n

Loss
\nA lot of people used ArcFace over CosFace, but in my experiments CosFace performed better. I used subcenter CosFace, because of the noise in the dataset. Also I followed dynamic margins as in the top solutions of Google Landmark competition

\n

Thing that didnā€™t work šŸ™
\nExperiments with triplet loss. I tried several approaches:

\n
    \n
  1. Simple hard triplet loss
    2. \n
  2. Triplet mining the following way. In the batch there was the same proportion of whales and dolphins.
    3. \n
  3. Centroid triplet loss
    \nAll this experiments performed worse than cosface
    4. \n
\n

šŸ‘€Summary of mentioned articles:
\n[1] Your \"Flamingo\" is My \"Bird\": Fine-Grained, or Not https://arxiv.org/pdf/2104.13643.pdf
\n[2] Sub-center ArcFace: Boosting Face Recognition by Large-scale Noisy Web Faces https://www.ecva.net/papers/eccv_2020/papers_ECCV/papers/123560715.pdf
\n[3] A general multi-modal data learning method for Person Re-identification
\nhttps://arxiv.org/pdf/2101.08533.pdf
\n[4] On the Unreasonable Effectiveness of Centroids in Image Retrieval
\nhttps://arxiv.org/pdf/2104.13643.pdf

", "messages": [ { "id": 1761359, "postDate": "2022-04-19T20:25:40.630Z", "content": "

First of all, I would like to say congratulations to all the winners!šŸŽ‰šŸ”„ You did a good job with such a complex dataset!

\n

Although I didnā€™t do well in this competition (4 places to bronzešŸ˜©), I would like to share my solution. I think this information may be useful for further competitions.

\n

Architecture
\nAt the beginning of the competition I was interested in finding neural net architectures, which outperform the baseline effnet. I was sure that there are better architectures (Well, in the end I found that the differences are not so drastic). I stopped at the effnetv2_m/effnet_b6/convneXt_m in my final ensemble.
\nWhen I saw the data the first idea came to my mind was, that it is a hierarchical learning competition. After some search I found this interesting article: Your \"Flamingo\" is My \"Bird\": Fine-Grained, or Not. This article is applied to a classification problem, but I suggested it will work well in re-identification either (I modified an article approach a little making the individual embedding part larger than species). The final architecture of my solution you can see in the picture below.
\n\"arch1.png\"

\n

Augmentations
\n 2.1 Shear scale rotate
\nFrom the dataset you can see that there are a lot of different viewpoints for one whale. Using this augmentation I tried to simulate them.
\n 2.2 CLAHE at test and train time before all augmentation
\nThe purpose of using this augmentation is that some back fins are under water, some are darkened.
\n 2.3 Local grayscale transformation
\nThis transformation is based on this article, which gave a little boost to LB
\n 2.4 Color/brightness/hue/horizontal flip

\n

Dataset
\nFor all novice kagglers I would like to say that dataset is crucially important. I didnā€™t pay a lot of attention to it, but when I switched from detic to backfins dataset, the score was boosted a lot.

\n

Loss
\nA lot of people used ArcFace over CosFace, but in my experiments CosFace performed better. I used subcenter CosFace, because of the noise in the dataset. Also I followed dynamic margins as in the top solutions of Google Landmark competition

\n

Thing that didnā€™t work šŸ™
\nExperiments with triplet loss. I tried several approaches:

\n
    \n
  1. Simple hard triplet loss
    2. \n
  2. Triplet mining the following way. In the batch there was the same proportion of whales and dolphins.
    3. \n
  3. Centroid triplet loss
    \nAll this experiments performed worse than cosface
    4. \n
\n

šŸ‘€Summary of mentioned articles:
\n[1] Your \"Flamingo\" is My \"Bird\": Fine-Grained, or Not https://arxiv.org/pdf/2104.13643.pdf
\n[2] Sub-center ArcFace: Boosting Face Recognition by Large-scale Noisy Web Faces https://www.ecva.net/papers/eccv_2020/papers_ECCV/papers/123560715.pdf
\n[3] A general multi-modal data learning method for Person Re-identification
\nhttps://arxiv.org/pdf/2101.08533.pdf
\n[4] On the Unreasonable Effectiveness of Centroids in Image Retrieval
\nhttps://arxiv.org/pdf/2104.13643.pdf

", "rawMarkdown": "First of all, I would like to say congratulations to all the winners!šŸŽ‰šŸ”„ You did a good job with such a complex dataset!\n\nAlthough I didnā€™t do well in this competition (4 places to bronzešŸ˜©), I would like to share my solution. I think this information may be useful for further competitions.\n\n**Architecture**\nAt the beginning of the competition I was interested in finding neural net architectures, which outperform the baseline effnet. I was sure that there are better architectures (Well, in the end I found that the differences are not so drastic). I stopped at the effnetv2_m/effnet_b6/convneXt_m in my final ensemble.\nWhen I saw the data the first idea came to my mind was, that it is a hierarchical learning competition. After some search I found this interesting article: **[Your \"Flamingo\" is My \"Bird\": Fine-Grained, or Not]( https://arxiv.org/pdf/2104.13643.pdf)**. This article is applied to a classification problem, but I suggested it will work well in re-identification either (I modified an article approach a little making the individual embedding part larger than species). The final architecture of my solution you can see in the picture below.\n[![arch1.png](https://i.postimg.cc/bvcXWFCw/arch1.png)](https://postimg.cc/4Y5FhwRD)\n\n\n**Augmentations**\n 2.1 **Shear scale rotate**\nFrom the dataset you can see that there are a lot of different viewpoints for one whale. Using this augmentation I tried to simulate them.\n 2.2 **CLAHE** at test and train time before all augmentation\nThe purpose of using this augmentation is that some back fins are under water, some are darkened. \n 2.3 **Local grayscale transformation**\nThis transformation is based on this [article](https://arxiv.org/pdf/2101.08533.pdf), which gave a little boost to LB\n 2.4 **Color/brightness/hue/horizontal flip**\n\n**Dataset**\nFor all novice kagglers I would like to say that **dataset is crucially important**. I didnā€™t pay a lot of attention to it, but when I switched from detic to backfins dataset, the score was boosted a lot. \n\n**Loss**\nA lot of people used ArcFace over CosFace, but in my experiments CosFace performed better. I used [subcenter CosFace](https://www.ecva.net/papers/eccv_2020/papers_ECCV/papers/123560715.pdf), because of the noise in the dataset. Also I followed dynamic margins as in the top solutions of Google Landmark competition\n\n**Thing that didnā€™t work šŸ™**\nExperiments with triplet loss. I tried several approaches:\n1. Simple hard triplet loss\n2. Triplet mining the following way. In the batch there was the same proportion of whales and dolphins. \n3. [ Centroid triplet loss ](https://arxiv.org/pdf/2104.13643.pdf)\nAll this experiments performed worse than cosface\n\nšŸ‘€**Summary of mentioned articles:**\n[1] Your \"Flamingo\" is My \"Bird\": Fine-Grained, or Not https://arxiv.org/pdf/2104.13643.pdf\n[2] Sub-center ArcFace: Boosting Face Recognition by Large-scale Noisy Web Faces https://www.ecva.net/papers/eccv_2020/papers_ECCV/papers/123560715.pdf\n[3] A general multi-modal data learning method for Person Re-identification\nhttps://arxiv.org/pdf/2101.08533.pdf\n[4] On the Unreasonable Effectiveness of Centroids in Image Retrieval\nhttps://arxiv.org/pdf/2104.13643.pdf\n", "votes": 10 } ], "comments": [], "raw_markdown_by_id": { "1761359": "First of all, I would like to say congratulations to all the winners!šŸŽ‰šŸ”„ You did a good job with such a complex dataset!\n\nAlthough I didnā€™t do well in this competition (4 places to bronzešŸ˜©), I would like to share my solution. I think this information may be useful for further competitions.\n\n**Architecture**\nAt the beginning of the competition I was interested in finding neural net architectures, which outperform the baseline effnet. I was sure that there are better architectures (Well, in the end I found that the differences are not so drastic). I stopped at the effnetv2_m/effnet_b6/convneXt_m in my final ensemble.\nWhen I saw the data the first idea came to my mind was, that it is a hierarchical learning competition. After some search I found this interesting article: **[Your \"Flamingo\" is My \"Bird\": Fine-Grained, or Not]( https://arxiv.org/pdf/2104.13643.pdf)**. This article is applied to a classification problem, but I suggested it will work well in re-identification either (I modified an article approach a little making the individual embedding part larger than species). The final architecture of my solution you can see in the picture below.\n[![arch1.png](https://i.postimg.cc/bvcXWFCw/arch1.png)](https://postimg.cc/4Y5FhwRD)\n\n\n**Augmentations**\n 2.1 **Shear scale rotate**\nFrom the dataset you can see that there are a lot of different viewpoints for one whale. Using this augmentation I tried to simulate them.\n 2.2 **CLAHE** at test and train time before all augmentation\nThe purpose of using this augmentation is that some back fins are under water, some are darkened. \n 2.3 **Local grayscale transformation**\nThis transformation is based on this [article](https://arxiv.org/pdf/2101.08533.pdf), which gave a little boost to LB\n 2.4 **Color/brightness/hue/horizontal flip**\n\n**Dataset**\nFor all novice kagglers I would like to say that **dataset is crucially important**. I didnā€™t pay a lot of attention to it, but when I switched from detic to backfins dataset, the score was boosted a lot. \n\n**Loss**\nA lot of people used ArcFace over CosFace, but in my experiments CosFace performed better. I used [subcenter CosFace](https://www.ecva.net/papers/eccv_2020/papers_ECCV/papers/123560715.pdf), because of the noise in the dataset. Also I followed dynamic margins as in the top solutions of Google Landmark competition\n\n**Thing that didnā€™t work šŸ™**\nExperiments with triplet loss. I tried several approaches:\n1. Simple hard triplet loss\n2. Triplet mining the following way. In the batch there was the same proportion of whales and dolphins. \n3. [ Centroid triplet loss ](https://arxiv.org/pdf/2104.13643.pdf)\nAll this experiments performed worse than cosface\n\nšŸ‘€**Summary of mentioned articles:**\n[1] Your \"Flamingo\" is My \"Bird\": Fine-Grained, or Not https://arxiv.org/pdf/2104.13643.pdf\n[2] Sub-center ArcFace: Boosting Face Recognition by Large-scale Noisy Web Faces https://www.ecva.net/papers/eccv_2020/papers_ECCV/papers/123560715.pdf\n[3] A general multi-modal data learning method for Person Re-identification\nhttps://arxiv.org/pdf/2101.08533.pdf\n[4] On the Unreasonable Effectiveness of Centroids in Image Retrieval\nhttps://arxiv.org/pdf/2104.13643.pdf\n" } }