As seen from various winning solutions of Cornell Birdcall Identification competition , the top performers used external data and also aggressive augmentations to do well . I had earlier used a baseline solution without augmentations and it was able to score 0.78X , I applied Audio Data Augmentations and simplified model architecture(removed some dense layers) and for the same backbone architecture I was able to get a score of 0.83X , that too for only after training for single fold , I added some dense layers with augmentations and the score dropped to 0.78X again , so not only using augmentations but also using simple architecture is the key to place well in this comp.
\nPS: I am yet to try more ideas for this competition😄
", "messages": [ { "id": "1169425", "postDate": "01/25/2021 13:47:23", "content": "As seen from various winning solutions of Cornell Birdcall Identification competition , the top performers used external data and also aggressive augmentations to do well . I had earlier used a baseline solution without augmentations and it was able to score 0.78X , I applied Audio Data Augmentations and simplified model architecture(removed some dense layers) and for the same backbone architecture I was able to get a score of 0.83X , that too for only after training for single fold , I added some dense layers with augmentations and the score dropped to 0.78X again , so not only using augmentations but also using simple architecture is the key to place well in this comp.
\nPS: I am yet to try more ideas for this competition😄
", "rawMarkdown": "As seen from various winning solutions of [Cornell Birdcall Identification](https://www.kaggle.com/c/birdsong-recognition/overview) competition , the top performers used external data and also aggressive augmentations to do well . I had earlier used a baseline solution without augmentations and it was able to score **0.78X** , I applied Audio Data Augmentations and simplified model architecture(removed some dense layers) and for the same backbone architecture I was able to get a score of **0.83X** , that too for only after training for single fold , I added some dense layers with augmentations and the score dropped to **0.78X** again , so not only using augmentations but also using simple architecture is the key to place well in this comp.\n\n**PS: I am yet to try more ideas for this competition😄**", "votes": null }, { "id": "1169709", "postDate": "01/25/2021 17:39:13", "content": "Thanks for sharing @sayedathar11 , what augmentations are using? I tried GaussNoise, TimeMask and FrequencyMask, sometimes they looks good on CV but all failed on LB 😂
", "rawMarkdown": "Thanks for sharing @sayedathar11 , what augmentations are using? I tried GaussNoise, TimeMask and FrequencyMask, sometimes they looks good on CV but all failed on LB 😂", "votes": null } ], "comments": [ { "id": 1169709, "author_name": "superchenhao", "author_url": "", "post_date": "01/25/2021 17:39:13", "content": "Thanks for sharing @sayedathar11 , what augmentations are using? I tried GaussNoise, TimeMask and FrequencyMask, sometimes they looks good on CV but all failed on LB 😂
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1169425": "As seen from various winning solutions of [Cornell Birdcall Identification](https://www.kaggle.com/c/birdsong-recognition/overview) competition , the top performers used external data and also aggressive augmentations to do well . I had earlier used a baseline solution without augmentations and it was able to score **0.78X** , I applied Audio Data Augmentations and simplified model architecture(removed some dense layers) and for the same backbone architecture I was able to get a score of **0.83X** , that too for only after training for single fold , I added some dense layers with augmentations and the score dropped to **0.78X** again , so not only using augmentations but also using simple architecture is the key to place well in this comp.\n\n**PS: I am yet to try more ideas for this competition😄**", "1169709": "Thanks for sharing @sayedathar11 , what augmentations are using? I tried GaussNoise, TimeMask and FrequencyMask, sometimes they looks good on CV but all failed on LB 😂" }, "source": "meta" }