This model seems to be very popular among birdclef competition. I understand that it is used in the No 4 solution in last year's competition, which @lihaoweicvch used to achieve 0.68 lb.
\nBut why this model? On huggingface, it is rarely used. It's top1 and top5 accuracys are worse than efficientnet of the same size (b4). For my case, it tends to have worse performance compared to models with other backbones. I sometimes even have a hard time training it as it produce NAN loss in the middle of an epoch (could be my problem).
\nWhat are your thoughts?
", "messages": [ { "id": "2804378", "postDate": "05/10/2024 03:14:59", "content": "This model seems to be very popular among birdclef competition. I understand that it is used in the No 4 solution in last year's competition, which @lihaoweicvch used to achieve 0.68 lb.
\nBut why this model? On huggingface, it is rarely used. It's top1 and top5 accuracys are worse than efficientnet of the same size (b4). For my case, it tends to have worse performance compared to models with other backbones. I sometimes even have a hard time training it as it produce NAN loss in the middle of an epoch (could be my problem).
\nWhat are your thoughts?
", "rawMarkdown": "This model seems to be very popular among birdclef competition. I understand that it is used in the No 4 solution in last year's competition, which @lihaoweicvch used to achieve 0.68 lb. \n\nBut why this model? On huggingface, it is rarely used. It's top1 and top5 accuracys are worse than efficientnet of the same size (b4). For my case, it tends to have worse performance compared to models with other backbones. I sometimes even have a hard time training it as it produce NAN loss in the middle of an epoch (could be my problem).\n\nWhat are your thoughts?", "votes": null }, { "id": "2804663", "postDate": "05/10/2024 06:30:57", "content": "We are confirming that the choice of backbone maximizing CV varies depending on how the mel spectrogram is constructed. When adopting specific preprocessing and employing GeM in the model, we confirm that eca_nfnet_l0 still achieves the best scores within the observed range.
", "rawMarkdown": "We are confirming that the choice of backbone maximizing CV varies depending on how the mel spectrogram is constructed. When adopting specific preprocessing and employing GeM in the model, we confirm that eca_nfnet_l0 still achieves the best scores within the observed range.", "votes": null }, { "id": "2804758", "postDate": "05/10/2024 07:32:03", "content": "\n\nIt's top1 and top5 accuracys are worse than efficientnet of the same size (b4).
\n
We can't really use this figure because it's probably for ImageNet classification leaderboard. Here we are doing mel spectogram.
", "rawMarkdown": ">It's top1 and top5 accuracys are worse than efficientnet of the same size (b4).\n\nWe can't really use this figure because it's probably for ImageNet classification leaderboard. Here we are doing mel spectogram.", "votes": null } ], "comments": [ { "id": 2804663, "author_name": "miyamotodaiya", "author_url": "", "post_date": "05/10/2024 06:30:57", "content": "We are confirming that the choice of backbone maximizing CV varies depending on how the mel spectrogram is constructed. When adopting specific preprocessing and employing GeM in the model, we confirm that eca_nfnet_l0 still achieves the best scores within the observed range.
", "votes": null, "replies": [] }, { "id": 2804758, "author_name": "nyleve", "author_url": "", "post_date": "05/10/2024 07:32:03", "content": "\n\nIt's top1 and top5 accuracys are worse than efficientnet of the same size (b4).
\n
We can't really use this figure because it's probably for ImageNet classification leaderboard. Here we are doing mel spectogram.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "2804378": "This model seems to be very popular among birdclef competition. I understand that it is used in the No 4 solution in last year's competition, which @lihaoweicvch used to achieve 0.68 lb. \n\nBut why this model? On huggingface, it is rarely used. It's top1 and top5 accuracys are worse than efficientnet of the same size (b4). For my case, it tends to have worse performance compared to models with other backbones. I sometimes even have a hard time training it as it produce NAN loss in the middle of an epoch (could be my problem).\n\nWhat are your thoughts?", "2804663": "We are confirming that the choice of backbone maximizing CV varies depending on how the mel spectrogram is constructed. When adopting specific preprocessing and employing GeM in the model, we confirm that eca_nfnet_l0 still achieves the best scores within the observed range.", "2804758": ">It's top1 and top5 accuracys are worse than efficientnet of the same size (b4).\n\nWe can't really use this figure because it's probably for ImageNet classification leaderboard. Here we are doing mel spectogram." }, "source": "meta" }