{
  "id": 507258,
  "title": "What’s your best single model score?",
  "url": "/competitions/birdclef-2024/discussion/507258",
  "author_name": "Warrior",
  "post_date": "2024-05-25T05:17:00.688000",
  "votes": 15,
  "comment_count": 50,
  "views": 0,
  "content": "<p>Hi all,</p>\n<p>I have recently started to submit to the leaderboard. I need to get a sense of how my models are performing. Could you please share your best score with a single model so far?</p>",
  "messages": [
    {
      "id": 2834930,
      "postDate": "2024-05-25T05:17:00.690Z",
      "content": "<p>Hi all,</p>\n<p>I have recently started to submit to the leaderboard. I need to get a sense of how my models are performing. Could you please share your best score with a single model so far?</p>",
      "rawMarkdown": "Hi all,\n\nI have recently started to submit to the leaderboard. I need to get a sense of how my models are performing. Could you please share your best score with a single model so far?",
      "votes": 14
    },
    {
      "id": 2848205,
      "postDate": "2024-05-31T21:59:09.783Z",
      "content": "<p>1fold 0.68 just by chance. Another fold 0.62.</p>",
      "rawMarkdown": "1fold 0.68 just by chance. Another fold 0.62.",
      "votes": 9,
      "replies": [
        {
          "id": 2851488,
          "postDate": "2024-06-02T19:20:49.940Z",
          "content": "<p>This has a risk of going down in the private leaderboard. I think the reason for the sudden shift in scores is that the number of true labels in the public leaderboard is few so it's very sensitive.</p>",
          "rawMarkdown": "This has a risk of going down in the private leaderboard. I think the reason for the sudden shift in scores is that the number of true labels in the public leaderboard is few so it's very sensitive.\n",
          "votes": 1
        }
      ]
    },
    {
      "id": 2835425,
      "postDate": "2024-05-25T10:12:32.577Z",
      "content": "<p>ECA-NFNet-L0 5fold</p>\n<table>\n<thead>\n<tr>\n<th>fold0</th>\n<th>fold1</th>\n<th>fold2</th>\n<th>fold3</th>\n<th>fold4</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>0.68</td>\n<td>0.66</td>\n<td>0.67</td>\n<td>N/A</td>\n<td>N/A</td>\n</tr>\n</tbody>\n</table>",
      "rawMarkdown": "ECA-NFNet-L0 5fold\n\n|fold0|fold1|fold2|fold3|fold4|\n|---|---|---|---|---|\n|0.68|0.66|0.67|N/A|N/A|",
      "votes": 7,
      "replies": [
        {
          "id": 2835578,
          "postDate": "2024-05-25T11:57:51.107Z",
          "content": "<p>Thanks for sharing.<br>\nIn my case, eca_nfnet_b0 is out of time for inference in 5-fold ensemble with OpenVino.<br>\nWhat library did you use for faster inference?</p>\n<ul>\n<li>I know that inference time is also dependent on our trained weights.</li>\n</ul>",
          "rawMarkdown": "Thanks for sharing.\nIn my case, eca_nfnet_b0 is out of time for inference in 5-fold ensemble with OpenVino.\nWhat library did you use for faster inference?\n\n* I know that inference time is also dependent on our trained weights.",
          "votes": 2,
          "replies": [
            {
              "id": 2835641,
              "postDate": "2024-05-25T12:36:22.617Z",
              "content": "<p><a href=\"https://www.kaggle.com/johnlemon3\" target=\"_blank\">@johnlemon3</a> </p>\n<p>I haven't tried ensemble methods yet. The scores I mentioned are from individual models for each fold.</p>\n<p>I'm using OpenVINO for the inference engine. Since even a single eca_nfnet_b0 takes about 75 minutes, I'm considering an ensemble with lighter models like those in the EfficientNet series.</p>",
              "rawMarkdown": "@johnlemon3 \n\nI haven't tried ensemble methods yet. The scores I mentioned are from individual models for each fold.\n\nI'm using OpenVINO for the inference engine. Since even a single eca_nfnet_b0 takes about 75 minutes, I'm considering an ensemble with lighter models like those in the EfficientNet series.",
              "votes": 3
            },
            {
              "id": 2835682,
              "postDate": "2024-05-25T13:02:05.650Z",
              "content": "<p><a href=\"https://www.kaggle.com/akinosora\" target=\"_blank\">@akinosora</a> <br>\nOk, Thank you!</p>\n<p>And I agree with your statement about ensemble.<br>\nEca_nfnet_b0 will produce a good LB score, but this model takes a lot of time for inference.<br>\nConsidering the private test dataset, what we need to do is probably to make ensembles with lighter models at the cost of single heavy model's good score in public LB.</p>",
              "rawMarkdown": "@akinosora \nOk, Thank you!\n\nAnd I agree with your statement about ensemble.\nEca_nfnet_b0 will produce a good LB score, but this model takes a lot of time for inference.\nConsidering the private test dataset, what we need to do is probably to make ensembles with lighter models at the cost of single heavy model's good score in public LB.",
              "votes": 2
            },
            {
              "id": 2844442,
              "postDate": "2024-05-30T04:42:57.697Z",
              "rawMarkdown": "",
              "isDeleted": true
            },
            {
              "id": 2844488,
              "postDate": "2024-05-30T05:30:26.840Z",
              "content": "<p>Hey, <a href=\"https://www.kaggle.com/akinosora\" target=\"_blank\">@akinosora</a> is there any notebook you referred to for the OpenVINO inference engine submission?</p>",
              "rawMarkdown": "Hey, @akinosora is there any notebook you referred to for the OpenVINO inference engine submission?"
            },
            {
              "id": 2845222,
              "postDate": "2024-05-30T13:16:45.183Z",
              "content": "<ul>\n<li>InputSize==[4, 1, 128, 501]</li>\n<li>OpenVINO</li>\n<li>ThreadPoolExecutor(4)</li>\n</ul>",
              "rawMarkdown": "* InputSize==[4, 1, 128, 501]\n* OpenVINO\n* ThreadPoolExecutor(4)",
              "votes": 1
            },
            {
              "id": 2845246,
              "postDate": "2024-05-30T13:27:09.430Z",
              "content": "<p><a href=\"https://www.kaggle.com/rishavdash\" target=\"_blank\">@rishavdash</a> </p>\n<p>I didn’t use any notebooks. Instead, I referred to academic papers and source code:</p>\n<ul>\n<li><a href=\"https://arxiv.org/abs/2102.06171\" target=\"_blank\">https://arxiv.org/abs/2102.06171</a></li>\n<li><a href=\"https://github.com/huggingface/pytorch-image-models/blob/main/timm/models/nfnet.py#L270\" target=\"_blank\">https://github.com/huggingface/pytorch-image-models/blob/main/timm/models/nfnet.py#L270</a></li>\n<li><a href=\"https://github.com/huggingface/pytorch-image-models/blob/main/timm/layers/std_conv.py#L76\" target=\"_blank\">https://github.com/huggingface/pytorch-image-models/blob/main/timm/layers/std_conv.py#L76</a></li>\n</ul>\n<p>Note that ScaledStdConv2d includes BatchNorm (training=True), which can't be directly converted to OpenVINO. So, after training NFNet, if you properly replace ScaledStdConv2d with Conv2d, you can convert it to OpenVINO without losing accuracy.</p>",
              "rawMarkdown": "@rishavdash \n\nI didn’t use any notebooks. Instead, I referred to academic papers and source code:\n\n* https://arxiv.org/abs/2102.06171\n* https://github.com/huggingface/pytorch-image-models/blob/main/timm/models/nfnet.py#L270\n* https://github.com/huggingface/pytorch-image-models/blob/main/timm/layers/std_conv.py#L76\n\nNote that ScaledStdConv2d includes BatchNorm (training=True), which can't be directly converted to OpenVINO. So, after training NFNet, if you properly replace ScaledStdConv2d with Conv2d, you can convert it to OpenVINO without losing accuracy.",
              "votes": 1
            }
          ]
        },
        {
          "id": 2836370,
          "postDate": "2024-05-25T20:35:22.440Z",
          "content": "<p>Did you use SED or CNN architecture?</p>",
          "rawMarkdown": "Did you use SED or CNN architecture?",
          "replies": [
            {
              "id": 2837448,
              "postDate": "2024-05-26T13:44:13.023Z",
              "content": "<p><a href=\"https://www.kaggle.com/tjamali\" target=\"_blank\">@tjamali</a> </p>\n<p>It's a CNN model trained with 5-second audio clips. I think some tricks are needed to reach the top, but I'm not sure what they are yet.</p>",
              "rawMarkdown": "@tjamali \n\nIt's a CNN model trained with 5-second audio clips. I think some tricks are needed to reach the top, but I'm not sure what they are yet.",
              "votes": 4
            },
            {
              "id": 2843734,
              "postDate": "2024-05-29T17:26:53.793Z",
              "content": "<p>I am continuing with almost no hope in the short time remained :(</p>",
              "rawMarkdown": "I am continuing with almost no hope in the short time remained :(",
              "votes": 2
            },
            {
              "id": 2848473,
              "postDate": "2024-06-01T03:55:49.967Z",
              "content": "<p>Is your image_size = 256?</p>",
              "rawMarkdown": "Is your image_size = 256?"
            }
          ]
        }
      ]
    },
    {
      "id": 2836499,
      "postDate": "2024-05-26T00:58:55.543Z",
      "content": "<p>between 0.64 and 0.69 depending of the parameters and randomness.</p>",
      "rawMarkdown": "between 0.64 and 0.69 depending of the parameters and randomness.",
      "votes": 6,
      "replies": [
        {
          "id": 2836535,
          "postDate": "2024-05-26T01:59:43.170Z",
          "content": "<p>Thanks. Much score shift!</p>",
          "rawMarkdown": "Thanks. Much score shift!"
        },
        {
          "id": 2836564,
          "postDate": "2024-05-26T02:33:37.853Z",
          "content": "<p>Have you seen this much variation between folds of the same run?</p>",
          "rawMarkdown": "Have you seen this much variation between folds of the same run?",
          "replies": [
            {
              "id": 2837559,
              "postDate": "2024-05-26T15:08:11.623Z",
              "content": "<p>yes unfortunately</p>",
              "rawMarkdown": "yes unfortunately",
              "votes": 2
            },
            {
              "id": 2837877,
              "postDate": "2024-05-26T17:53:19.093Z",
              "content": "<p>Yeah I’ve definitely seen some variation between folds and I was curious if I should try to submit some additional folds from runs that had a high score.</p>",
              "rawMarkdown": "Yeah I’ve definitely seen some variation between folds and I was curious if I should try to submit some additional folds from runs that had a high score."
            },
            {
              "id": 2838103,
              "postDate": "2024-05-26T20:00:36.960Z",
              "content": "<p>Yeah you should try who knows</p>",
              "rawMarkdown": "Yeah you should try who knows",
              "votes": 1
            }
          ]
        }
      ]
    },
    {
      "id": 2836347,
      "postDate": "2024-05-25T20:11:56.473Z",
      "content": "<p>In my case, the best is single model with LB 0.69 and 40 minutes inference time. I don't use ensembling yet.</p>",
      "rawMarkdown": "In my case, the best is single model with LB 0.69 and 40 minutes inference time. I don't use ensembling yet.",
      "votes": 3,
      "replies": [
        {
          "id": 2836366,
          "postDate": "2024-05-25T20:28:23.950Z",
          "content": "<p>It's an amazing single model score! Is it only because of your base model or you did something beside it? I just want to know how much of this score comes from the model itself.</p>",
          "rawMarkdown": "It's an amazing single model score! Is it only because of your base model or you did something beside it? I just want to know how much of this score comes from the model itself.",
          "votes": 1,
          "replies": [
            {
              "id": 2836382,
              "postDate": "2024-05-25T20:45:27.077Z",
              "content": "<p>More impressive is the inference time imo, <a href=\"https://www.kaggle.com/kdmitrie\" target=\"_blank\">@kdmitrie</a> are you doing inference using openvino or onnx ?</p>",
              "rawMarkdown": "More impressive is the inference time imo, @kdmitrie are you doing inference using openvino or onnx ?",
              "votes": 2
            },
            {
              "id": 2836886,
              "postDate": "2024-05-26T06:40:50.343Z",
              "content": "<p>The model is an ordinary one. Without the tricks I applied during training it gets 0.61-0.63 LB. <br>\nAlso, as <a href=\"https://www.kaggle.com/ludovick\" target=\"_blank\">@ludovick</a> stated, the score depends on randomness.</p>\n<p>As for inference, I used both onnx and openvino. Openvino is several minutes faster.</p>",
              "rawMarkdown": "The model is an ordinary one. Without the tricks I applied during training it gets 0.61-0.63 LB. \nAlso, as @ludovick stated, the score depends on randomness.\n\nAs for inference, I used both onnx and openvino. Openvino is several minutes faster.",
              "votes": 2
            },
            {
              "id": 2837144,
              "postDate": "2024-05-26T09:52:59.620Z",
              "content": "<p>So far I have got two scores 0.61 and 0.63 exactly as you mentioned.</p>",
              "rawMarkdown": "So far I have got two scores 0.61 and 0.63 exactly as you mentioned.\n"
            },
            {
              "id": 2838632,
              "postDate": "2024-05-27T06:33:16.010Z",
              "content": "<p><a href=\"https://www.kaggle.com/kdmitrie\" target=\"_blank\">@kdmitrie</a> may i know whats the cv you got ? considering you are validating on first 5 seconds from you audio samples if i am not wrong.</p>",
              "rawMarkdown": "@kdmitrie may i know whats the cv you got ? considering you are validating on first 5 seconds from you audio samples if i am not wrong."
            }
          ]
        },
        {
          "id": 2843961,
          "postDate": "2024-05-29T19:39:26.357Z",
          "content": "<p>Just curious, did you use addition audios. I used <a href=\"https://www.kaggle.com/competitions/birdclef-2024/discussion/491687\" target=\"_blank\">additional audios </a>and my score didn't change.</p>",
          "rawMarkdown": "Just curious, did you use addition audios. I used [additional audios ](https://www.kaggle.com/competitions/birdclef-2024/discussion/491687)and my score didn't change.",
          "replies": [
            {
              "id": 2844046,
              "postDate": "2024-05-29T21:18:39.060Z",
              "content": "<p>I used both additional audios and unlabeled sounscapes</p>",
              "rawMarkdown": "I used both additional audios and unlabeled sounscapes",
              "votes": 3
            }
          ]
        }
      ]
    },
    {
      "id": 2835477,
      "postDate": "2024-05-25T10:40:06.630Z",
      "content": "<p>I believe there is a trick that can boost single model 0.66 score to more than 0.71, but I did not find out what it is yet.</p>",
      "rawMarkdown": "I believe there is a trick that can boost single model 0.66 score to more than 0.71, but I did not find out what it is yet.",
      "votes": 3,
      "replies": [
        {
          "id": 2835586,
          "postDate": "2024-05-25T12:00:36.087Z",
          "content": "<p>Are you ensembling for your current LB?</p>",
          "rawMarkdown": "Are you ensembling for your current LB?",
          "replies": [
            {
              "id": 2835612,
              "postDate": "2024-05-25T12:21:57.217Z",
              "content": "<p>yes I ensemble </p>",
              "rawMarkdown": "yes I ensemble ",
              "votes": 2
            }
          ]
        },
        {
          "id": 2836013,
          "postDate": "2024-05-25T16:50:16.107Z",
          "content": "<p>how many model do you ensemble?</p>",
          "rawMarkdown": "how many model do you ensemble?",
          "votes": 1
        },
        {
          "id": 2836367,
          "postDate": "2024-05-25T20:29:56.907Z",
          "content": "<p>Is your base model ECA-NFNet-L0 like other good single model score?</p>",
          "rawMarkdown": "Is your base model ECA-NFNet-L0 like other good single model score?"
        }
      ]
    },
    {
      "id": 2835233,
      "postDate": "2024-05-25T08:07:40.440Z",
      "content": "<p>0.65 single model score</p>",
      "rawMarkdown": "0.65 single model score",
      "votes": 3,
      "replies": [
        {
          "id": 2836373,
          "postDate": "2024-05-25T20:37:21.610Z",
          "content": "<p>Could you let me know which base model you used?</p>",
          "rawMarkdown": "Could you let me know which base model you used?",
          "replies": [
            {
              "id": 2836384,
              "postDate": "2024-05-25T20:45:54.457Z",
              "content": "<p>Efficientnet b0</p>",
              "rawMarkdown": "Efficientnet b0",
              "votes": 4
            },
            {
              "id": 2843956,
              "postDate": "2024-05-29T19:35:36.427Z",
              "content": "<p>Just curious, did you use addition audio or just the original one. I used <a href=\"https://www.kaggle.com/competitions/birdclef-2024/discussion/491687\" target=\"_blank\">additional audios </a>and my score didn't change at all.</p>",
              "rawMarkdown": "Just curious, did you use addition audio or just the original one. I used [additional audios ](https://www.kaggle.com/competitions/birdclef-2024/discussion/491687)and my score didn't change at all."
            }
          ]
        }
      ]
    },
    {
      "id": 2852784,
      "postDate": "2024-06-03T13:10:18.337Z",
      "content": "<p>I got lots of 0.62 from EfficientnetB0/eca_nf_l1. The best one I got is 0.64 (amost 0.65 I guess because I'm the top 1 in 0.64 range) using eca_nf_l1 with strat-k-fold, xy_mask, and random h-flip. I have no idea about how to improve. 😭</p>\n<p>Update 1: EfficientnetB0 0.65(2 I guess) </p>",
      "rawMarkdown": "I got lots of 0.62 from EfficientnetB0/eca_nf_l1. The best one I got is 0.64 (amost 0.65 I guess because I'm the top 1 in 0.64 range) using eca_nf_l1 with strat-k-fold, xy_mask, and random h-flip. I have no idea about how to improve. 😭\n\nUpdate 1: EfficientnetB0 0.65(2 I guess) ",
      "votes": 1,
      "replies": [
        {
          "id": 2853218,
          "postDate": "2024-06-03T16:58:37.127Z",
          "content": "<p>Hope you can get a good score in the private leaderboard.</p>",
          "rawMarkdown": "Hope you can get a good score in the private leaderboard."
        }
      ]
    },
    {
      "id": 2845266,
      "postDate": "2024-05-30T13:42:19.460Z",
      "content": "<p>so many 0.66</p>",
      "rawMarkdown": "so many 0.66",
      "votes": 2
    },
    {
      "id": 2838970,
      "postDate": "2024-05-27T10:00:45.400Z",
      "content": "<p>0.66 LB with single model. </p>",
      "rawMarkdown": "0.66 LB with single model. ",
      "votes": 2
    },
    {
      "id": 2837096,
      "postDate": "2024-05-26T09:15:09.530Z",
      "content": "<p>My best single model is 0.68 PB.</p>",
      "rawMarkdown": "My best single model is 0.68 PB.",
      "votes": 2,
      "replies": [
        {
          "id": 2838165,
          "postDate": "2024-05-26T21:19:17.103Z",
          "content": "<p>This gives me hope</p>",
          "rawMarkdown": "This gives me hope",
          "votes": 1,
          "replies": [
            {
              "id": 2838263,
              "postDate": "2024-05-27T01:08:21.450Z",
              "content": "<p>I try add nocall to background noise and pretrain on 2020, 2021, 2022, 2023 data but not improve.</p>",
              "rawMarkdown": "I try add nocall to background noise and pretrain on 2020, 2021, 2022, 2023 data but not improve.",
              "votes": 4
            }
          ]
        }
      ]
    },
    {
      "id": 2835582,
      "postDate": "2024-05-25T11:59:47.613Z",
      "content": "<p>0.66 single model score</p>",
      "rawMarkdown": "0.66 single model score",
      "votes": 2
    },
    {
      "id": 2837599,
      "postDate": "2024-05-26T15:41:55.683Z",
      "rawMarkdown": "",
      "votes": -29,
      "replies": [
        {
          "id": 2838088,
          "postDate": "2024-05-26T19:29:50.680Z",
          "content": "<p>How do you get a LB score without submitting?</p>\n<p>I guess this is why you got down votes.</p>",
          "rawMarkdown": "How do you get a LB score without submitting?\n\nI guess this is why you got down votes.",
          "votes": 4,
          "replies": [
            {
              "id": 2838164,
              "postDate": "2024-05-26T21:17:14.510Z",
              "content": "<p>You keep it in the theoretical realm. This guy knows marketing 😉</p>",
              "rawMarkdown": "You keep it in the theoretical realm. This guy knows marketing 😉",
              "votes": 2
            }
          ]
        }
      ]
    },
    {
      "id": 2840974,
      "postDate": "2024-05-28T10:37:37.230Z",
      "content": "<p>Thank you for your support. </p>",
      "rawMarkdown": "Thank you for your support. ",
      "votes": 1
    },
    {
      "id": 2840740,
      "postDate": "2024-05-28T08:42:36.520Z",
      "content": "<p>Thank Youuu for idea.</p>",
      "rawMarkdown": "Thank Youuu for idea.",
      "votes": 1
    }
  ],
  "comments": [
    {
      "id": 2848205,
      "author_name": "gezi",
      "author_url": "",
      "post_date": "2024-05-31T21:59:09.783000",
      "content": "<p>1fold 0.68 just by chance. Another fold 0.62.</p>",
      "votes": 9,
      "replies": [
        {
          "id": 2851488,
          "author_name": "Warrior",
          "author_url": "",
          "post_date": "2024-06-02T19:20:49.940000",
          "content": "<p>This has a risk of going down in the private leaderboard. I think the reason for the sudden shift in scores is that the number of true labels in the public leaderboard is few so it's very sensitive.</p>",
          "votes": 1,
          "replies": []
        }
      ]
    },
    {
      "id": 2835425,
      "author_name": "Moyashii",
      "author_url": "",
      "post_date": "2024-05-25T10:12:32.577000",
      "content": "<p>ECA-NFNet-L0 5fold</p>\n<table>\n<thead>\n<tr>\n<th>fold0</th>\n<th>fold1</th>\n<th>fold2</th>\n<th>fold3</th>\n<th>fold4</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>0.68</td>\n<td>0.66</td>\n<td>0.67</td>\n<td>N/A</td>\n<td>N/A</td>\n</tr>\n</tbody>\n</table>",
      "votes": 7,
      "replies": [
        {
          "id": 2835578,
          "author_name": "MBOOK",
          "author_url": "",
          "post_date": "2024-05-25T11:57:51.107000",
          "content": "<p>Thanks for sharing.<br>\nIn my case, eca_nfnet_b0 is out of time for inference in 5-fold ensemble with OpenVino.<br>\nWhat library did you use for faster inference?</p>\n<ul>\n<li>I know that inference time is also dependent on our trained weights.</li>\n</ul>",
          "votes": 2,
          "replies": [
            {
              "id": 2835641,
              "author_name": "Moyashii",
              "author_url": "",
              "post_date": "2024-05-25T12:36:22.617000",
              "content": "<p><a href=\"https://www.kaggle.com/johnlemon3\" target=\"_blank\">@johnlemon3</a> </p>\n<p>I haven't tried ensemble methods yet. The scores I mentioned are from individual models for each fold.</p>\n<p>I'm using OpenVINO for the inference engine. Since even a single eca_nfnet_b0 takes about 75 minutes, I'm considering an ensemble with lighter models like those in the EfficientNet series.</p>",
              "votes": 3,
              "replies": []
            },
            {
              "id": 2835682,
              "author_name": "MBOOK",
              "author_url": "",
              "post_date": "2024-05-25T13:02:05.650000",
              "content": "<p><a href=\"https://www.kaggle.com/akinosora\" target=\"_blank\">@akinosora</a> <br>\nOk, Thank you!</p>\n<p>And I agree with your statement about ensemble.<br>\nEca_nfnet_b0 will produce a good LB score, but this model takes a lot of time for inference.<br>\nConsidering the private test dataset, what we need to do is probably to make ensembles with lighter models at the cost of single heavy model's good score in public LB.</p>",
              "votes": 2,
              "replies": []
            },
            {
              "id": 2844442,
              "author_name": "",
              "author_url": "",
              "post_date": "2024-05-30T04:42:57.697000",
              "content": "",
              "votes": 0,
              "replies": []
            },
            {
              "id": 2844488,
              "author_name": "Rishav dash",
              "author_url": "",
              "post_date": "2024-05-30T05:30:26.840000",
              "content": "<p>Hey, <a href=\"https://www.kaggle.com/akinosora\" target=\"_blank\">@akinosora</a> is there any notebook you referred to for the OpenVINO inference engine submission?</p>",
              "votes": 0,
              "replies": []
            },
            {
              "id": 2845222,
              "author_name": "Moyashii",
              "author_url": "",
              "post_date": "2024-05-30T13:16:45.183000",
              "content": "<ul>\n<li>InputSize==[4, 1, 128, 501]</li>\n<li>OpenVINO</li>\n<li>ThreadPoolExecutor(4)</li>\n</ul>",
              "votes": 1,
              "replies": []
            },
            {
              "id": 2845246,
              "author_name": "Moyashii",
              "author_url": "",
              "post_date": "2024-05-30T13:27:09.430000",
              "content": "<p><a href=\"https://www.kaggle.com/rishavdash\" target=\"_blank\">@rishavdash</a> </p>\n<p>I didn’t use any notebooks. Instead, I referred to academic papers and source code:</p>\n<ul>\n<li><a href=\"https://arxiv.org/abs/2102.06171\" target=\"_blank\">https://arxiv.org/abs/2102.06171</a></li>\n<li><a href=\"https://github.com/huggingface/pytorch-image-models/blob/main/timm/models/nfnet.py#L270\" target=\"_blank\">https://github.com/huggingface/pytorch-image-models/blob/main/timm/models/nfnet.py#L270</a></li>\n<li><a href=\"https://github.com/huggingface/pytorch-image-models/blob/main/timm/layers/std_conv.py#L76\" target=\"_blank\">https://github.com/huggingface/pytorch-image-models/blob/main/timm/layers/std_conv.py#L76</a></li>\n</ul>\n<p>Note that ScaledStdConv2d includes BatchNorm (training=True), which can't be directly converted to OpenVINO. So, after training NFNet, if you properly replace ScaledStdConv2d with Conv2d, you can convert it to OpenVINO without losing accuracy.</p>",
              "votes": 1,
              "replies": []
            }
          ]
        },
        {
          "id": 2836370,
          "author_name": "Warrior",
          "author_url": "",
          "post_date": "2024-05-25T20:35:22.440000",
          "content": "<p>Did you use SED or CNN architecture?</p>",
          "votes": 0,
          "replies": [
            {
              "id": 2837448,
              "author_name": "Moyashii",
              "author_url": "",
              "post_date": "2024-05-26T13:44:13.023000",
              "content": "<p><a href=\"https://www.kaggle.com/tjamali\" target=\"_blank\">@tjamali</a> </p>\n<p>It's a CNN model trained with 5-second audio clips. I think some tricks are needed to reach the top, but I'm not sure what they are yet.</p>",
              "votes": 4,
              "replies": []
            },
            {
              "id": 2843734,
              "author_name": "Warrior",
              "author_url": "",
              "post_date": "2024-05-29T17:26:53.793000",
              "content": "<p>I am continuing with almost no hope in the short time remained :(</p>",
              "votes": 2,
              "replies": []
            },
            {
              "id": 2848473,
              "author_name": "Warrior",
              "author_url": "",
              "post_date": "2024-06-01T03:55:49.967000",
              "content": "<p>Is your image_size = 256?</p>",
              "votes": 0,
              "replies": []
            }
          ]
        }
      ]
    },
    {
      "id": 2836499,
      "author_name": "Shiro",
      "author_url": "",
      "post_date": "2024-05-26T00:58:55.543000",
      "content": "<p>between 0.64 and 0.69 depending of the parameters and randomness.</p>",
      "votes": 6,
      "replies": [
        {
          "id": 2836535,
          "author_name": "Warrior",
          "author_url": "",
          "post_date": "2024-05-26T01:59:43.170000",
          "content": "<p>Thanks. Much score shift!</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 2836564,
          "author_name": "Will Rice",
          "author_url": "",
          "post_date": "2024-05-26T02:33:37.853000",
          "content": "<p>Have you seen this much variation between folds of the same run?</p>",
          "votes": 0,
          "replies": [
            {
              "id": 2837559,
              "author_name": "Shiro",
              "author_url": "",
              "post_date": "2024-05-26T15:08:11.623000",
              "content": "<p>yes unfortunately</p>",
              "votes": 2,
              "replies": []
            },
            {
              "id": 2837877,
              "author_name": "Will Rice",
              "author_url": "",
              "post_date": "2024-05-26T17:53:19.093000",
              "content": "<p>Yeah I’ve definitely seen some variation between folds and I was curious if I should try to submit some additional folds from runs that had a high score.</p>",
              "votes": 0,
              "replies": []
            },
            {
              "id": 2838103,
              "author_name": "Shiro",
              "author_url": "",
              "post_date": "2024-05-26T20:00:36.960000",
              "content": "<p>Yeah you should try who knows</p>",
              "votes": 1,
              "replies": []
            }
          ]
        }
      ]
    },
    {
      "id": 2836347,
      "author_name": "Konstantin Dmitriev",
      "author_url": "",
      "post_date": "2024-05-25T20:11:56.473000",
      "content": "<p>In my case, the best is single model with LB 0.69 and 40 minutes inference time. I don't use ensembling yet.</p>",
      "votes": 3,
      "replies": [
        {
          "id": 2836366,
          "author_name": "Warrior",
          "author_url": "",
          "post_date": "2024-05-25T20:28:23.950000",
          "content": "<p>It's an amazing single model score! Is it only because of your base model or you did something beside it? I just want to know how much of this score comes from the model itself.</p>",
          "votes": 1,
          "replies": [
            {
              "id": 2836382,
              "author_name": "Shubham Thapa",
              "author_url": "",
              "post_date": "2024-05-25T20:45:27.077000",
              "content": "<p>More impressive is the inference time imo, <a href=\"https://www.kaggle.com/kdmitrie\" target=\"_blank\">@kdmitrie</a> are you doing inference using openvino or onnx ?</p>",
              "votes": 2,
              "replies": []
            },
            {
              "id": 2836886,
              "author_name": "Konstantin Dmitriev",
              "author_url": "",
              "post_date": "2024-05-26T06:40:50.343000",
              "content": "<p>The model is an ordinary one. Without the tricks I applied during training it gets 0.61-0.63 LB. <br>\nAlso, as <a href=\"https://www.kaggle.com/ludovick\" target=\"_blank\">@ludovick</a> stated, the score depends on randomness.</p>\n<p>As for inference, I used both onnx and openvino. Openvino is several minutes faster.</p>",
              "votes": 2,
              "replies": []
            },
            {
              "id": 2837144,
              "author_name": "Warrior",
              "author_url": "",
              "post_date": "2024-05-26T09:52:59.620000",
              "content": "<p>So far I have got two scores 0.61 and 0.63 exactly as you mentioned.</p>",
              "votes": 0,
              "replies": []
            },
            {
              "id": 2838632,
              "author_name": "Shubham Thapa",
              "author_url": "",
              "post_date": "2024-05-27T06:33:16.010000",
              "content": "<p><a href=\"https://www.kaggle.com/kdmitrie\" target=\"_blank\">@kdmitrie</a> may i know whats the cv you got ? considering you are validating on first 5 seconds from you audio samples if i am not wrong.</p>",
              "votes": 0,
              "replies": []
            }
          ]
        },
        {
          "id": 2843961,
          "author_name": "Warrior",
          "author_url": "",
          "post_date": "2024-05-29T19:39:26.357000",
          "content": "<p>Just curious, did you use addition audios. I used <a href=\"https://www.kaggle.com/competitions/birdclef-2024/discussion/491687\" target=\"_blank\">additional audios </a>and my score didn't change.</p>",
          "votes": 0,
          "replies": [
            {
              "id": 2844046,
              "author_name": "Konstantin Dmitriev",
              "author_url": "",
              "post_date": "2024-05-29T21:18:39.060000",
              "content": "<p>I used both additional audios and unlabeled sounscapes</p>",
              "votes": 3,
              "replies": []
            }
          ]
        }
      ]
    },
    {
      "id": 2835477,
      "author_name": "yuanzhe zhou",
      "author_url": "",
      "post_date": "2024-05-25T10:40:06.630000",
      "content": "<p>I believe there is a trick that can boost single model 0.66 score to more than 0.71, but I did not find out what it is yet.</p>",
      "votes": 3,
      "replies": [
        {
          "id": 2835586,
          "author_name": "Will Rice",
          "author_url": "",
          "post_date": "2024-05-25T12:00:36.087000",
          "content": "<p>Are you ensembling for your current LB?</p>",
          "votes": 0,
          "replies": [
            {
              "id": 2835612,
              "author_name": "yuanzhe zhou",
              "author_url": "",
              "post_date": "2024-05-25T12:21:57.217000",
              "content": "<p>yes I ensemble </p>",
              "votes": 2,
              "replies": []
            }
          ]
        },
        {
          "id": 2836013,
          "author_name": "Arunodhayan",
          "author_url": "",
          "post_date": "2024-05-25T16:50:16.107000",
          "content": "<p>how many model do you ensemble?</p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 2836367,
          "author_name": "Warrior",
          "author_url": "",
          "post_date": "2024-05-25T20:29:56.907000",
          "content": "<p>Is your base model ECA-NFNet-L0 like other good single model score?</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 2835233,
      "author_name": "Arunodhayan",
      "author_url": "",
      "post_date": "2024-05-25T08:07:40.440000",
      "content": "<p>0.65 single model score</p>",
      "votes": 3,
      "replies": [
        {
          "id": 2836373,
          "author_name": "Warrior",
          "author_url": "",
          "post_date": "2024-05-25T20:37:21.610000",
          "content": "<p>Could you let me know which base model you used?</p>",
          "votes": 0,
          "replies": [
            {
              "id": 2836384,
              "author_name": "Arunodhayan",
              "author_url": "",
              "post_date": "2024-05-25T20:45:54.457000",
              "content": "<p>Efficientnet b0</p>",
              "votes": 4,
              "replies": []
            },
            {
              "id": 2843956,
              "author_name": "Warrior",
              "author_url": "",
              "post_date": "2024-05-29T19:35:36.427000",
              "content": "<p>Just curious, did you use addition audio or just the original one. I used <a href=\"https://www.kaggle.com/competitions/birdclef-2024/discussion/491687\" target=\"_blank\">additional audios </a>and my score didn't change at all.</p>",
              "votes": 0,
              "replies": []
            }
          ]
        }
      ]
    },
    {
      "id": 2852784,
      "author_name": "SakuraYuyuko",
      "author_url": "",
      "post_date": "2024-06-03T13:10:18.337000",
      "content": "<p>I got lots of 0.62 from EfficientnetB0/eca_nf_l1. The best one I got is 0.64 (amost 0.65 I guess because I'm the top 1 in 0.64 range) using eca_nf_l1 with strat-k-fold, xy_mask, and random h-flip. I have no idea about how to improve. 😭</p>\n<p>Update 1: EfficientnetB0 0.65(2 I guess) </p>",
      "votes": 1,
      "replies": [
        {
          "id": 2853218,
          "author_name": "Warrior",
          "author_url": "",
          "post_date": "2024-06-03T16:58:37.127000",
          "content": "<p>Hope you can get a good score in the private leaderboard.</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 2845266,
      "author_name": "Ksana",
      "author_url": "",
      "post_date": "2024-05-30T13:42:19.460000",
      "content": "<p>so many 0.66</p>",
      "votes": 2,
      "replies": []
    },
    {
      "id": 2838970,
      "author_name": "Phaedrus",
      "author_url": "",
      "post_date": "2024-05-27T10:00:45.400000",
      "content": "<p>0.66 LB with single model. </p>",
      "votes": 2,
      "replies": []
    },
    {
      "id": 2837096,
      "author_name": "Quan Vu",
      "author_url": "",
      "post_date": "2024-05-26T09:15:09.530000",
      "content": "<p>My best single model is 0.68 PB.</p>",
      "votes": 2,
      "replies": [
        {
          "id": 2838165,
          "author_name": "Will Rice",
          "author_url": "",
          "post_date": "2024-05-26T21:19:17.103000",
          "content": "<p>This gives me hope</p>",
          "votes": 1,
          "replies": [
            {
              "id": 2838263,
              "author_name": "Quan Vu",
              "author_url": "",
              "post_date": "2024-05-27T01:08:21.450000",
              "content": "<p>I try add nocall to background noise and pretrain on 2020, 2021, 2022, 2023 data but not improve.</p>",
              "votes": 4,
              "replies": []
            }
          ]
        }
      ]
    },
    {
      "id": 2835582,
      "author_name": "Will Rice",
      "author_url": "",
      "post_date": "2024-05-25T11:59:47.613000",
      "content": "<p>0.66 single model score</p>",
      "votes": 2,
      "replies": []
    },
    {
      "id": 2837599,
      "author_name": "",
      "author_url": "",
      "post_date": "2024-05-26T15:41:55.683000",
      "content": "",
      "votes": -29,
      "replies": [
        {
          "id": 2838088,
          "author_name": "CPMP",
          "author_url": "",
          "post_date": "2024-05-26T19:29:50.680000",
          "content": "<p>How do you get a LB score without submitting?</p>\n<p>I guess this is why you got down votes.</p>",
          "votes": 4,
          "replies": [
            {
              "id": 2838164,
              "author_name": "Will Rice",
              "author_url": "",
              "post_date": "2024-05-26T21:17:14.510000",
              "content": "<p>You keep it in the theoretical realm. This guy knows marketing 😉</p>",
              "votes": 2,
              "replies": []
            }
          ]
        }
      ]
    },
    {
      "id": 2840974,
      "author_name": "RAJDIP PARMAR",
      "author_url": "",
      "post_date": "2024-05-28T10:37:37.230000",
      "content": "<p>Thank you for your support. </p>",
      "votes": 1,
      "replies": []
    },
    {
      "id": 2840740,
      "author_name": "Janki P Sodha",
      "author_url": "",
      "post_date": "2024-05-28T08:42:36.520000",
      "content": "<p>Thank Youuu for idea.</p>",
      "votes": 1,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "2834930": "Hi all,\n\nI have recently started to submit to the leaderboard. I need to get a sense of how my models are performing. Could you please share your best score with a single model so far?",
    "2848205": "1fold 0.68 just by chance. Another fold 0.62.",
    "2835425": "ECA-NFNet-L0 5fold\n\n|fold0|fold1|fold2|fold3|fold4|\n|---|---|---|---|---|\n|0.68|0.66|0.67|N/A|N/A|",
    "2836499": "between 0.64 and 0.69 depending of the parameters and randomness.",
    "2836347": "In my case, the best is single model with LB 0.69 and 40 minutes inference time. I don't use ensembling yet.",
    "2835477": "I believe there is a trick that can boost single model 0.66 score to more than 0.71, but I did not find out what it is yet.",
    "2835233": "0.65 single model score",
    "2852784": "I got lots of 0.62 from EfficientnetB0/eca_nf_l1. The best one I got is 0.64 (amost 0.65 I guess because I'm the top 1 in 0.64 range) using eca_nf_l1 with strat-k-fold, xy_mask, and random h-flip. I have no idea about how to improve. 😭\n\nUpdate 1: EfficientnetB0 0.65(2 I guess) ",
    "2845266": "so many 0.66",
    "2838970": "0.66 LB with single model. ",
    "2837096": "My best single model is 0.68 PB.",
    "2835582": "0.66 single model score",
    "2837599": "",
    "2840974": "Thank you for your support. ",
    "2840740": "Thank Youuu for idea."
  }
}