Have anyone tried Efficient net v2 series?
\nFor me, I use an efficient v2 m, but it drops my rate (0.82+ -> 0.768).
\nIt seems strange as the efficient v looks more modern than b. (personal opinion)
\nBTW, what's the single best model for you?
\nmy efficient net b4 got 0.821 (still improving)
Have anyone tried Efficient net v2 series?
\nFor me, I use an efficient v2 m, but it drops my rate (0.82+ -> 0.768).
\nIt seems strange as the efficient v looks more modern than b. (personal opinion)
\nBTW, what's the single best model for you?
\nmy efficient net b4 got 0.821 (still improving)
I've only used efficientnet v1 b0 for now, but I'm curious what resolution spectrogram did you use to train? It might have an impact, since the EfficientNetV2 paper used a fairly involved \"progressive learning\" training routine, slowly increasing image size from 128 -> 300 while increasing regularization over 350+ epochs. The also reported V2 has somewhat worse accuracy than V1 for small models when not using progressive learning (table 11).
", "rawMarkdown": "I've only used efficientnet v1 b0 for now, but I'm curious what resolution spectrogram did you use to train? It might have an impact, since the [EfficientNetV2 paper](https://arxiv.org/abs/2104.00298) used a fairly involved \"progressive learning\" training routine, slowly increasing image size from 128 -> 300 while increasing regularization over 350+ epochs. The also reported V2 has somewhat worse accuracy than V1 for small models when not using progressive learning (table 11).", "votes": null }, { "id": "3188954", "postDate": "04/28/2025 14:32:05", "content": "thx for your reply. I also notice that the V2 is trained by a \"progressive learning\". However, I think progressive learning represents a large workload and I have not figure out how to combine progressive learning with K fold, since they are a little bit contrary. Also, some info show that efficient v2 can be defined as a backbone for object detection model, which may indicates that it can be used noramlly.
", "rawMarkdown": "thx for your reply. I also notice that the V2 is trained by a \"progressive learning\". However, I think progressive learning represents a large workload and I have not figure out how to combine progressive learning with K fold, since they are a little bit contrary. Also, some info show that efficient v2 can be defined as a backbone for object detection model, which may indicates that it can be used noramlly.", "votes": null }, { "id": "3188957", "postDate": "04/28/2025 14:36:23", "content": "I only use the 380 size as input. It may be a little bit high i think
", "rawMarkdown": "I only use the 380 size as input. It may be a little bit high i think", "votes": null } ], "comments": [ { "id": 3188929, "author_name": "robbynevels", "author_url": "", "post_date": "04/28/2025 13:19:21", "content": "I've only used efficientnet v1 b0 for now, but I'm curious what resolution spectrogram did you use to train? It might have an impact, since the EfficientNetV2 paper used a fairly involved \"progressive learning\" training routine, slowly increasing image size from 128 -> 300 while increasing regularization over 350+ epochs. The also reported V2 has somewhat worse accuracy than V1 for small models when not using progressive learning (table 11).
", "votes": null, "replies": [ { "id": 3188954, "author_name": "xinaocheng", "author_url": "", "post_date": "04/28/2025 14:32:05", "content": "thx for your reply. I also notice that the V2 is trained by a \"progressive learning\". However, I think progressive learning represents a large workload and I have not figure out how to combine progressive learning with K fold, since they are a little bit contrary. Also, some info show that efficient v2 can be defined as a backbone for object detection model, which may indicates that it can be used noramlly.
", "votes": null, "replies": [] }, { "id": 3188957, "author_name": "xinaocheng", "author_url": "", "post_date": "04/28/2025 14:36:23", "content": "I only use the 380 size as input. It may be a little bit high i think
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "3188888": "Have anyone tried Efficient net v2 series?\nFor me, I use an efficient v2 m, but it drops my rate (0.82+ -> 0.768). \nIt seems strange as the efficient v looks more modern than b. (personal opinion)\nBTW, what's the single best model for you?\nmy efficient net b4 got 0.821 (still improving)", "3188929": "I've only used efficientnet v1 b0 for now, but I'm curious what resolution spectrogram did you use to train? It might have an impact, since the [EfficientNetV2 paper](https://arxiv.org/abs/2104.00298) used a fairly involved \"progressive learning\" training routine, slowly increasing image size from 128 -> 300 while increasing regularization over 350+ epochs. The also reported V2 has somewhat worse accuracy than V1 for small models when not using progressive learning (table 11).", "3188954": "thx for your reply. I also notice that the V2 is trained by a \"progressive learning\". However, I think progressive learning represents a large workload and I have not figure out how to combine progressive learning with K fold, since they are a little bit contrary. Also, some info show that efficient v2 can be defined as a backbone for object detection model, which may indicates that it can be used noramlly.", "3188957": "I only use the 380 size as input. It may be a little bit high i think" }, "source": "meta" }