{ "id": 220981, "title": "#6 Solution 🐼Tropic Thunder🐼 (journey and ensembling schemes)", "url": "/competitions/rfcx-species-audio-detection/discussion/220981", "author_name": "", "post_date": "2021-02-20T11:18:44.827479400Z", "votes": 19, "comment_count": 4, "views": 0, "content": "

Thanks to Kaggle and the Host for this competition. It's going to be special for me, I have got here my first Gold medal and my upgrade to Master tier in Competitions.

\n

Thanks a lot to my teammates: @jpison @pavelgonchar @antorsae. We worked together as a real team.

\n

After @antorsae post (https://www.kaggle.com/c/rfcx-species-audio-detection/discussion/220446) with his perspective, I'd like to give more information that could be useful to understand the influence of the differents experiments we did with our Tropic Model.

\n

First, find here our journey, under submissions perspective:
\n\"Submissions\"
\n@jpison and @amezet join in a team on January 19th
\n@pavelgonchar joined the team on February 1st
\n@antorsae joined the team on February 10th

\n

We worked in two different models (without hand labelling):

\n
    \n
  1. SED model. We worked with this base model, improving with mixup, intelligent cropping, better schedule: https://www.kaggle.com/gopidurgaprasad/rfcx-sed-model-stater.
    \nWe'd like to thank @gopidurgaprasad
    \nWe reached with this model 0.91760 in Public (0.92270 in Private)
    2. \n
  2. Tropic Model
    \n@antorsae explained in his post the details of the models.
    3. \n
\n

We learnt that the more diversity we have, the better score we got. So we did the experiments changing a lot the barebones, taking different from the complete list of TIMM models:
\nhttps://github.com/rwightman/pytorch-image-models/blob/master/results/results-imagenet.csv

\n

To select our two submissions, we prepare two scheme of ensembling:

\n

Submission selected 1:
\n\"Scheme

\n

Submission selected 2:
\n\"Scheme

\n

We could did experiments very fast because we have a lot of computing power. For this competition we used the following GPUs:
\n@antorsae
\n6x3090
\n@amezet
\n3x3090 + 2x2080Ti
\n@jpison
\n1x3090
\n@pavelgonchar
\n2x3090

", "messages": [ { "id": "1211617", "postDate": "02/20/2021 11:18:44", "content": "

Thanks to Kaggle and the Host for this competition. It's going to be special for me, I have got here my first Gold medal and my upgrade to Master tier in Competitions.

\n

Thanks a lot to my teammates: @jpison @pavelgonchar @antorsae. We worked together as a real team.

\n

After @antorsae post (https://www.kaggle.com/c/rfcx-species-audio-detection/discussion/220446) with his perspective, I'd like to give more information that could be useful to understand the influence of the differents experiments we did with our Tropic Model.

\n

First, find here our journey, under submissions perspective:
\n\"Submissions\"
\n@jpison and @amezet join in a team on January 19th
\n@pavelgonchar joined the team on February 1st
\n@antorsae joined the team on February 10th

\n

We worked in two different models (without hand labelling):

\n
    \n
  1. SED model. We worked with this base model, improving with mixup, intelligent cropping, better schedule: https://www.kaggle.com/gopidurgaprasad/rfcx-sed-model-stater.
    \nWe'd like to thank @gopidurgaprasad
    \nWe reached with this model 0.91760 in Public (0.92270 in Private)
    2. \n
  2. Tropic Model
    \n@antorsae explained in his post the details of the models.
    3. \n
\n

We learnt that the more diversity we have, the better score we got. So we did the experiments changing a lot the barebones, taking different from the complete list of TIMM models:
\nhttps://github.com/rwightman/pytorch-image-models/blob/master/results/results-imagenet.csv

\n

To select our two submissions, we prepare two scheme of ensembling:

\n

Submission selected 1:
\n\"Scheme

\n

Submission selected 2:
\n\"Scheme

\n

We could did experiments very fast because we have a lot of computing power. For this competition we used the following GPUs:
\n@antorsae
\n6x3090
\n@amezet
\n3x3090 + 2x2080Ti
\n@jpison
\n1x3090
\n@pavelgonchar
\n2x3090

", "rawMarkdown": "Thanks to Kaggle and the Host for this competition. It's going to be special for me, I have got here my first Gold medal and my upgrade to Master tier in Competitions.\n\nThanks a lot to my teammates: @jpison @pavelgonchar @antorsae. We worked together as a real team. \n\nAfter @antorsae post (https://www.kaggle.com/c/rfcx-species-audio-detection/discussion/220446) with his perspective, I'd like to give more information that could be useful to understand the influence of the differents experiments we did with our **Tropic Model**.\n\nFirst, find here **our journey, under submissions perspective**:\n![Submissions](https://raw.githubusercontent.com/amezet/kaggle_rfcx/main/Submissions%20-%20historical.jpg)\n@jpison and @amezet join in a team on January 19th\n@pavelgonchar joined the team on February 1st\n@antorsae joined the team on February 10th\n\nWe worked in two different models (**without hand labelling**):\n1. **SED model**. We worked with this base model, improving with mixup, intelligent cropping, better schedule: https://www.kaggle.com/gopidurgaprasad/rfcx-sed-model-stater.\nWe'd like to thank @gopidurgaprasad\nWe reached with this model 0.91760 in Public (0.92270 in Private)\n2. **Tropic Model**\n@antorsae explained in his post the details of the models.\n\nWe learnt that the more diversity we have, the better score we got. So we did the experiments changing a lot the barebones, taking different from the complete list of TIMM models:\nhttps://github.com/rwightman/pytorch-image-models/blob/master/results/results-imagenet.csv\n\nTo select our two submissions, we prepare two scheme of ensembling:\n\n**Submission selected 1:**\n![Scheme 1](https://raw.githubusercontent.com/amezet/kaggle_rfcx/main/SCHEME%20submission%20selected%201.jpg)\n\n**Submission selected 2:**\n![Scheme 2](https://raw.githubusercontent.com/amezet/kaggle_rfcx/main/SCHEME%20submission%20selected%202.jpg)\n\nWe could did experiments very fast because we have a lot of computing power. For this competition we used the following GPUs:\n@antorsae\n6x3090\n@amezet\n3x3090 + 2x2080Ti\n@jpison \n1x3090 \n@pavelgonchar\n2x3090", "votes": null }, { "id": "1215732", "postDate": "02/23/2021 22:36:38", "content": "

Congratulations, nice visualisation.
\nThanks for sharing!

", "rawMarkdown": "Congratulations, nice visualisation.\nThanks for sharing!", "votes": null }, { "id": "1216076", "postDate": "02/24/2021 05:24:55", "content": "

i like your progress chart. good work!

", "rawMarkdown": "i like your progress chart. good work!", "votes": null }, { "id": "1216096", "postDate": "02/24/2021 05:44:50", "content": "

I like the hardware flex. Good work.

", "rawMarkdown": "I like the hardware flex. Good work.", "votes": null }, { "id": "1217551", "postDate": "02/25/2021 06:17:31", "content": "

masters of GPUssss boss!

", "rawMarkdown": "masters of GPUssss boss!", "votes": null } ], "comments": [ { "id": 1215732, "author_name": "aliabdin1", "author_url": "", "post_date": "02/23/2021 22:36:38", "content": "

Congratulations, nice visualisation.
\nThanks for sharing!

", "votes": null, "replies": [] }, { "id": 1216076, "author_name": "hengck23", "author_url": "", "post_date": "02/24/2021 05:24:55", "content": "

i like your progress chart. good work!

", "votes": null, "replies": [] }, { "id": 1216096, "author_name": "underwearfitting", "author_url": "", "post_date": "02/24/2021 05:44:50", "content": "

I like the hardware flex. Good work.

", "votes": null, "replies": [] }, { "id": 1217551, "author_name": "wubinbai", "author_url": "", "post_date": "02/25/2021 06:17:31", "content": "

masters of GPUssss boss!

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1211617": "Thanks to Kaggle and the Host for this competition. It's going to be special for me, I have got here my first Gold medal and my upgrade to Master tier in Competitions.\n\nThanks a lot to my teammates: @jpison @pavelgonchar @antorsae. We worked together as a real team. \n\nAfter @antorsae post (https://www.kaggle.com/c/rfcx-species-audio-detection/discussion/220446) with his perspective, I'd like to give more information that could be useful to understand the influence of the differents experiments we did with our **Tropic Model**.\n\nFirst, find here **our journey, under submissions perspective**:\n![Submissions](https://raw.githubusercontent.com/amezet/kaggle_rfcx/main/Submissions%20-%20historical.jpg)\n@jpison and @amezet join in a team on January 19th\n@pavelgonchar joined the team on February 1st\n@antorsae joined the team on February 10th\n\nWe worked in two different models (**without hand labelling**):\n1. **SED model**. We worked with this base model, improving with mixup, intelligent cropping, better schedule: https://www.kaggle.com/gopidurgaprasad/rfcx-sed-model-stater.\nWe'd like to thank @gopidurgaprasad\nWe reached with this model 0.91760 in Public (0.92270 in Private)\n2. **Tropic Model**\n@antorsae explained in his post the details of the models.\n\nWe learnt that the more diversity we have, the better score we got. So we did the experiments changing a lot the barebones, taking different from the complete list of TIMM models:\nhttps://github.com/rwightman/pytorch-image-models/blob/master/results/results-imagenet.csv\n\nTo select our two submissions, we prepare two scheme of ensembling:\n\n**Submission selected 1:**\n![Scheme 1](https://raw.githubusercontent.com/amezet/kaggle_rfcx/main/SCHEME%20submission%20selected%201.jpg)\n\n**Submission selected 2:**\n![Scheme 2](https://raw.githubusercontent.com/amezet/kaggle_rfcx/main/SCHEME%20submission%20selected%202.jpg)\n\nWe could did experiments very fast because we have a lot of computing power. For this competition we used the following GPUs:\n@antorsae\n6x3090\n@amezet\n3x3090 + 2x2080Ti\n@jpison \n1x3090 \n@pavelgonchar\n2x3090", "1215732": "Congratulations, nice visualisation.\nThanks for sharing!", "1216076": "i like your progress chart. good work!", "1216096": "I like the hardware flex. Good work.", "1217551": "masters of GPUssss boss!" }, "source": "meta" }