Exp 1: First 5 seconds Training (Efficient B0) (2024 Data) [LB 0.66]
\nExp 2: (5 seconds from first 7 seconds for classes count > 160 samples, 5 seconds from first 20 seconds for classes with count < 160) [LB: 0.60]
\nI guess, still first 5 seconds work best, but I refuse to accept that there is no way to use remaining duration of recordings.
", "messages": [ { "id": "2810561", "postDate": "05/13/2024 10:15:35", "content": "Exp 1: First 5 seconds Training (Efficient B0) (2024 Data) [LB 0.66]
\nExp 2: (5 seconds from first 7 seconds for classes count > 160 samples, 5 seconds from first 20 seconds for classes with count < 160) [LB: 0.60]
\nI guess, still first 5 seconds work best, but I refuse to accept that there is no way to use remaining duration of recordings.
", "rawMarkdown": "Exp 1: First 5 seconds Training (Efficient B0) (2024 Data) [LB 0.66]\n\nExp 2: (5 seconds from first 7 seconds for classes count > 160 samples, 5 seconds from first 20 seconds for classes with count < 160) [LB: 0.60]\n\n\nI guess, still first 5 seconds work best, but I refuse to accept that there is no way to use remaining duration of recordings.", "votes": null }, { "id": "2810662", "postDate": "05/13/2024 11:42:38", "content": "Have you tried to use google bird-vocalization-analyzer to pick the \"best\" segment during training?
How does this work?
", "rawMarkdown": "How does this work?", "votes": null }, { "id": "2810987", "postDate": "05/13/2024 14:51:52", "content": "I tried. No help. It only even decreased lb.
", "rawMarkdown": "I tried. No help. It only even decreased lb.", "votes": null }, { "id": "2810991", "postDate": "05/13/2024 14:55:25", "content": "My training code uses all duration. I split each audio to 5s snippets and for each epoch I iterate over all snippets. It produces a 0.66 lb
", "rawMarkdown": "My training code uses all duration. I split each audio to 5s snippets and for each epoch I iterate over all snippets. It produces a 0.66 lb", "votes": null }, { "id": "2811010", "postDate": "05/13/2024 15:15:20", "content": "That's a great idea! Thanks for sharing this with everyone. I have no idea how I didn't think of that! How did you implement this?
", "rawMarkdown": "That's a great idea! Thanks for sharing this with everyone. I have no idea how I didn't think of that! How did you implement this?", "votes": null }, { "id": "2811012", "postDate": "05/13/2024 15:17:01", "content": "It's consistent with my results - I tried pseudolabeling unlabeled soundscapes with it, but got LB decrease.
", "rawMarkdown": "It's consistent with my results - I tried pseudolabeling unlabeled soundscapes with it, but got LB decrease.", "votes": null }, { "id": "2811205", "postDate": "05/13/2024 16:39:20", "content": "As a fellow refuser, I'm totally in alignment there's got to be a way to use it. It's throwing away precious, beautiful data. I don't know if any has mentioned this before, but I wonder if you do some sort of decreasing labeling smoothing on 5s clips if that works. First 5s 1.0, next 5s 0.98, next 5s 0.96, etc or something? Unfortunately, I haven't had time to try it myself.
", "rawMarkdown": "As a fellow refuser, I'm totally in alignment there's got to be a way to use it. It's throwing away precious, beautiful data. I don't know if any has mentioned this before, but I wonder if you do some sort of decreasing labeling smoothing on 5s clips if that works. First 5s 1.0, next 5s 0.98, next 5s 0.96, etc or something? Unfortunately, I haven't had time to try it myself.", "votes": null }, { "id": "2811316", "postDate": "05/13/2024 17:35:13", "content": "What LR do you use for your efficientnet? I cant get mine to give al cv above 0.82
", "rawMarkdown": "What LR do you use for your efficientnet? I cant get mine to give al cv above 0.82", "votes": null }, { "id": "2811739", "postDate": "05/13/2024 21:18:46", "content": "May I ask if your cv goes up with your Exp 2 setting?
", "rawMarkdown": "May I ask if your cv goes up with your Exp 2 setting?", "votes": null }, { "id": "2811825", "postDate": "05/13/2024 23:32:25", "content": "Yes, my CV with exp 1 in 0.972 and with EXP 2 its 0.976
", "rawMarkdown": "Yes, my CV with exp 1 in 0.972 and with EXP 2 its 0.976", "votes": null }, { "id": "2811826", "postDate": "05/13/2024 23:33:18", "content": "Here is my recent notebook you can use a starting point. CV goes more than 0.97.
\nhttps://www.kaggle.com/code/salmanahmedtamu/training-0-65-0-66/notebook
What did you do with the last segment < 5 sec? Did you dropped it, or padded it up to 5 secs?
", "rawMarkdown": "What did you do with the last segment < 5 sec? Did you dropped it, or padded it up to 5 secs?", "votes": null }, { "id": "2812989", "postDate": "05/14/2024 14:15:29", "content": "Does exp 2 increase LB or not?
", "rawMarkdown": "Does exp 2 increase LB or not?", "votes": null }, { "id": "2812991", "postDate": "05/14/2024 14:16:39", "content": "I pad it to 5 sec. However I dont think I have a strong effect on the result
", "rawMarkdown": "I pad it to 5 sec. However I dont think I have a strong effect on the result", "votes": null }, { "id": "2812993", "postDate": "05/14/2024 14:17:31", "content": "I simply split and save the 5 sec spectrograms in a place and create a new df based on the files
", "rawMarkdown": "I simply split and save the 5 sec spectrograms in a place and create a new df based on the files", "votes": null }, { "id": "2814074", "postDate": "05/15/2024 06:31:58", "content": "FWIW so far I’ve exclusively trained b0 using random crops and my submissions have ranged from 0.59-0.66. I think the idea to use the first 5 sec is a bit overblown (although I can see the intuition, it’s unclear to me if it really works)
", "rawMarkdown": "FWIW so far I’ve exclusively trained b0 using random crops and my submissions have ranged from 0.59-0.66. I think the idea to use the first 5 sec is a bit overblown (although I can see the intuition, it’s unclear to me if it really works)", "votes": null }, { "id": "2818147", "postDate": "05/17/2024 09:51:51", "content": "I’ve tried it too, but for me the CV remains the same. I think it’s enough to use 5 seconds per audio
", "rawMarkdown": "I’ve tried it too, but for me the CV remains the same. I think it’s enough to use 5 seconds per audio", "votes": null }, { "id": "2821079", "postDate": "05/17/2024 22:38:46", "content": "Do you use prev competitions data?
", "rawMarkdown": "Do you use prev competitions data?", "votes": null }, { "id": "2822334", "postDate": "05/18/2024 14:37:51", "content": "no i did not
", "rawMarkdown": "no i did not", "votes": null } ], "comments": [ { "id": 2810662, "author_name": "atamazian", "author_url": "", "post_date": "05/13/2024 11:42:38", "content": "Have you tried to use google bird-vocalization-analyzer to pick the \"best\" segment during training?
How does this work?
", "votes": null, "replies": [] }, { "id": 2810987, "author_name": "llleeeoooh", "author_url": "", "post_date": "05/13/2024 14:51:52", "content": "I tried. No help. It only even decreased lb.
", "votes": null, "replies": [ { "id": 2811012, "author_name": "atamazian", "author_url": "", "post_date": "05/13/2024 15:17:01", "content": "It's consistent with my results - I tried pseudolabeling unlabeled soundscapes with it, but got LB decrease.
", "votes": null, "replies": [] } ] } ] }, { "id": 2810991, "author_name": "llleeeoooh", "author_url": "", "post_date": "05/13/2024 14:55:25", "content": "My training code uses all duration. I split each audio to 5s snippets and for each epoch I iterate over all snippets. It produces a 0.66 lb
", "votes": null, "replies": [ { "id": 2811010, "author_name": "max1mum", "author_url": "", "post_date": "05/13/2024 15:15:20", "content": "That's a great idea! Thanks for sharing this with everyone. I have no idea how I didn't think of that! How did you implement this?
", "votes": null, "replies": [ { "id": 2812993, "author_name": "llleeeoooh", "author_url": "", "post_date": "05/14/2024 14:17:31", "content": "I simply split and save the 5 sec spectrograms in a place and create a new df based on the files
", "votes": null, "replies": [] } ] }, { "id": 2812819, "author_name": "atamazian", "author_url": "", "post_date": "05/14/2024 12:39:10", "content": "What did you do with the last segment < 5 sec? Did you dropped it, or padded it up to 5 secs?
", "votes": null, "replies": [ { "id": 2812991, "author_name": "llleeeoooh", "author_url": "", "post_date": "05/14/2024 14:16:39", "content": "I pad it to 5 sec. However I dont think I have a strong effect on the result
", "votes": null, "replies": [] } ] }, { "id": 2818147, "author_name": "demko1", "author_url": "", "post_date": "05/17/2024 09:51:51", "content": "I’ve tried it too, but for me the CV remains the same. I think it’s enough to use 5 seconds per audio
", "votes": null, "replies": [] }, { "id": 2821079, "author_name": "darkhann", "author_url": "", "post_date": "05/17/2024 22:38:46", "content": "Do you use prev competitions data?
", "votes": null, "replies": [ { "id": 2822334, "author_name": "llleeeoooh", "author_url": "", "post_date": "05/18/2024 14:37:51", "content": "no i did not
", "votes": null, "replies": [] } ] } ] }, { "id": 2811205, "author_name": "msthil", "author_url": "", "post_date": "05/13/2024 16:39:20", "content": "As a fellow refuser, I'm totally in alignment there's got to be a way to use it. It's throwing away precious, beautiful data. I don't know if any has mentioned this before, but I wonder if you do some sort of decreasing labeling smoothing on 5s clips if that works. First 5s 1.0, next 5s 0.98, next 5s 0.96, etc or something? Unfortunately, I haven't had time to try it myself.
", "votes": null, "replies": [] }, { "id": 2811316, "author_name": "dylanc21", "author_url": "", "post_date": "05/13/2024 17:35:13", "content": "What LR do you use for your efficientnet? I cant get mine to give al cv above 0.82
", "votes": null, "replies": [ { "id": 2811826, "author_name": "salmanahmedtamu", "author_url": "", "post_date": "05/13/2024 23:33:18", "content": "Here is my recent notebook you can use a starting point. CV goes more than 0.97.
\nhttps://www.kaggle.com/code/salmanahmedtamu/training-0-65-0-66/notebook
May I ask if your cv goes up with your Exp 2 setting?
", "votes": null, "replies": [ { "id": 2811825, "author_name": "salmanahmedtamu", "author_url": "", "post_date": "05/13/2024 23:32:25", "content": "Yes, my CV with exp 1 in 0.972 and with EXP 2 its 0.976
", "votes": null, "replies": [ { "id": 2812989, "author_name": "max1mum", "author_url": "", "post_date": "05/14/2024 14:15:29", "content": "Does exp 2 increase LB or not?
", "votes": null, "replies": [] } ] } ] }, { "id": 2814074, "author_name": "robbynevels", "author_url": "", "post_date": "05/15/2024 06:31:58", "content": "FWIW so far I’ve exclusively trained b0 using random crops and my submissions have ranged from 0.59-0.66. I think the idea to use the first 5 sec is a bit overblown (although I can see the intuition, it’s unclear to me if it really works)
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "2810561": "Exp 1: First 5 seconds Training (Efficient B0) (2024 Data) [LB 0.66]\n\nExp 2: (5 seconds from first 7 seconds for classes count > 160 samples, 5 seconds from first 20 seconds for classes with count < 160) [LB: 0.60]\n\n\nI guess, still first 5 seconds work best, but I refuse to accept that there is no way to use remaining duration of recordings.", "2810662": "Have you tried to use google `bird-vocalization-analyzer` to pick the \"best\" segment during training?", "2810846": "How does this work?", "2810987": "I tried. No help. It only even decreased lb.", "2810991": "My training code uses all duration. I split each audio to 5s snippets and for each epoch I iterate over all snippets. It produces a 0.66 lb", "2811010": "That's a great idea! Thanks for sharing this with everyone. I have no idea how I didn't think of that! How did you implement this?", "2811012": "It's consistent with my results - I tried pseudolabeling unlabeled soundscapes with it, but got LB decrease.", "2811205": "As a fellow refuser, I'm totally in alignment there's got to be a way to use it. It's throwing away precious, beautiful data. I don't know if any has mentioned this before, but I wonder if you do some sort of decreasing labeling smoothing on 5s clips if that works. First 5s 1.0, next 5s 0.98, next 5s 0.96, etc or something? Unfortunately, I haven't had time to try it myself.", "2811316": "What LR do you use for your efficientnet? I cant get mine to give al cv above 0.82", "2811739": "May I ask if your cv goes up with your Exp 2 setting?", "2811825": "Yes, my CV with exp 1 in 0.972 and with EXP 2 its 0.976", "2811826": "Here is my recent notebook you can use a starting point. CV goes more than 0.97. \nhttps://www.kaggle.com/code/salmanahmedtamu/training-0-65-0-66/notebook", "2812819": "What did you do with the last segment < 5 sec? Did you dropped it, or padded it up to 5 secs?", "2812989": "Does exp 2 increase LB or not?", "2812991": "I pad it to 5 sec. However I dont think I have a strong effect on the result", "2812993": "I simply split and save the 5 sec spectrograms in a place and create a new df based on the files", "2814074": "FWIW so far I’ve exclusively trained b0 using random crops and my submissions have ranged from 0.59-0.66. I think the idea to use the first 5 sec is a bit overblown (although I can see the intuition, it’s unclear to me if it really works)", "2818147": "I’ve tried it too, but for me the CV remains the same. I think it’s enough to use 5 seconds per audio", "2821079": "Do you use prev competitions data?", "2822334": "no i did not" }, "source": "meta" }