{
  "id": 294205,
  "title": "Journey from LB 0.038 ---> 0.502 ---> 0.515 🖤 🔥",
  "url": "/competitions/tensorflow-great-barrier-reef/discussion/294205",
  "author_name": "Mahipal Singh",
  "post_date": "2021-12-09T06:40:30.494000",
  "votes": 55,
  "comment_count": 44,
  "views": 0,
  "content": "<p>Hello, kagglers<br>\nMy last submission score 0.515 🔥. This is my first competition, Thanks to Kaggle, I'm learning a lot here. <br>\nHere is my journey from score 0.038 ---&gt; 0.515 :</p>\n<ol>\n<li>I started with <a href=\"https://www.kaggle.com/mahipalsingh/detection-using-keras-retinanet\" target=\"_blank\"> 🌟🐟Detection using Keras-RetinaNet </a> but After a lot of fine-tuning, I was able to score 0.038 only.</li>\n<li>I found <a href=\"https://www.kaggle.com/remekkinas/yolox-inference-on-kaggle-for-cots-lb-0-507/notebook\" target=\"_blank\">YoloX inference on Kaggle for COTS [LB 0.507]</a> amazing work done by <a href=\"https://www.kaggle.com/remekkinas\" target=\"_blank\">@remekkinas</a>.</li>\n<li>I trained YoloX-S with a lot of different configurations and got a score of 0.246.</li>\n<li>I decided to go for yolox-l and see how it works..🤞</li>\n<li>Finally I kept trying with different parameters/config and my score was keep getting better and better. </li>\n</ol>\n<p>The below table shows the model with the score I got till now:- </p>\n<table>\n<thead>\n<tr>\n<th>Model</th>\n<th>Score</th>\n<th>Comment</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>Keras-retinaNet</td>\n<td>0.006</td>\n<td>default config</td>\n</tr>\n<tr>\n<td>Keras-retinaNet</td>\n<td>0.038</td>\n<td>Hyperparameter Tunned</td>\n</tr>\n<tr>\n<td>YoloX-S</td>\n<td>0.166</td>\n<td>default config</td>\n</tr>\n<tr>\n<td>YoloX-S</td>\n<td>0.246</td>\n<td>Hyperparameter Tunned</td>\n</tr>\n<tr>\n<td>YoloX-l</td>\n<td>0.269</td>\n<td>default config</td>\n</tr>\n<tr>\n<td>YoloX-l</td>\n<td>0.448</td>\n<td>setting hyperparameter and config</td>\n</tr>\n<tr>\n<td>YoloX-l</td>\n<td>0.502</td>\n<td>more setting hyperparameter and config</td>\n</tr>\n<tr>\n<td>YoloX-l</td>\n<td>0.515</td>\n<td>more and more setting up hyperparameter and config 👐</td>\n</tr>\n</tbody>\n</table>\n<p>Also, please do check out this work <a href=\"https://www.kaggle.com/soumya9977/learning-to-sea-underwater-img-enhancement-eda\" target=\"_blank\">Learning to Sea: Underwater img Enhancement + EDA</a></p>\n<p>I'm a beginner and here for learning, I really thank those people who make this platform beginner-friendly and easy. </p>\n<p>Thanks for reading.</p>\n<p><strong>Update:</strong></p>\n<table>\n<thead>\n<tr>\n<th>Model</th>\n<th>Score</th>\n<th>Comment</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>YoloX-l</td>\n<td>0.517</td>\n<td>change in conf threshold 🚀</td>\n</tr>\n<tr>\n<td>YoloX-l Fine tuning + Tracking</td>\n<td>0.540</td>\n<td>Thanks to <a href=\"https://www.kaggle.com/parapapapam/yolox-inference-tracking-on-cots-lb-0-539\" target=\"_blank\">this notebook</a> 🚀   🚀</td>\n</tr>\n<tr>\n<td>Yolov5</td>\n<td>0.57+</td>\n<td></td>\n</tr>\n<tr>\n<td>Yolov5 + yoloX</td>\n<td>0.63+</td>\n<td><a href=\"https://www.kaggle.com/mahipalsingh/gbr-yolox-yolov5-ensemble-2-o\" target=\"_blank\">GBR: YOLOX+YOLOv5 Ensemble 2.o</a></td>\n</tr>\n</tbody>\n</table>\n<p><strong>What I did:</strong></p>\n<ol>\n<li>Implemented RetinaNet: <a href=\"https://www.kaggle.com/mahipalsingh/detection-using-keras-retinanet-inference\" target=\"_blank\">🌟🐟Detection using Keras-RetinaNet [Train/Inference]</a></li>\n<li><a href=\"https://www.kaggle.com/mahipalsingh/gbr-yolox-yolov5-ensemble-2-o\" target=\"_blank\">GBR: YOLOX+YOLOv5 Ensemble 2.o</a></li>\n</ol>\n<p>(Sorry for grammatical mistakes if any😅)</p>",
  "messages": [
    {
      "id": 1612609,
      "postDate": "2021-12-09T06:40:30.493Z",
      "content": "<p>Hello, kagglers<br>\nMy last submission score 0.515 🔥. This is my first competition, Thanks to Kaggle, I'm learning a lot here. <br>\nHere is my journey from score 0.038 ---&gt; 0.515 :</p>\n<ol>\n<li>I started with <a href=\"https://www.kaggle.com/mahipalsingh/detection-using-keras-retinanet\" target=\"_blank\"> 🌟🐟Detection using Keras-RetinaNet </a> but After a lot of fine-tuning, I was able to score 0.038 only.</li>\n<li>I found <a href=\"https://www.kaggle.com/remekkinas/yolox-inference-on-kaggle-for-cots-lb-0-507/notebook\" target=\"_blank\">YoloX inference on Kaggle for COTS [LB 0.507]</a> amazing work done by <a href=\"https://www.kaggle.com/remekkinas\" target=\"_blank\">@remekkinas</a>.</li>\n<li>I trained YoloX-S with a lot of different configurations and got a score of 0.246.</li>\n<li>I decided to go for yolox-l and see how it works..🤞</li>\n<li>Finally I kept trying with different parameters/config and my score was keep getting better and better. </li>\n</ol>\n<p>The below table shows the model with the score I got till now:- </p>\n<table>\n<thead>\n<tr>\n<th>Model</th>\n<th>Score</th>\n<th>Comment</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>Keras-retinaNet</td>\n<td>0.006</td>\n<td>default config</td>\n</tr>\n<tr>\n<td>Keras-retinaNet</td>\n<td>0.038</td>\n<td>Hyperparameter Tunned</td>\n</tr>\n<tr>\n<td>YoloX-S</td>\n<td>0.166</td>\n<td>default config</td>\n</tr>\n<tr>\n<td>YoloX-S</td>\n<td>0.246</td>\n<td>Hyperparameter Tunned</td>\n</tr>\n<tr>\n<td>YoloX-l</td>\n<td>0.269</td>\n<td>default config</td>\n</tr>\n<tr>\n<td>YoloX-l</td>\n<td>0.448</td>\n<td>setting hyperparameter and config</td>\n</tr>\n<tr>\n<td>YoloX-l</td>\n<td>0.502</td>\n<td>more setting hyperparameter and config</td>\n</tr>\n<tr>\n<td>YoloX-l</td>\n<td>0.515</td>\n<td>more and more setting up hyperparameter and config 👐</td>\n</tr>\n</tbody>\n</table>\n<p>Also, please do check out this work <a href=\"https://www.kaggle.com/soumya9977/learning-to-sea-underwater-img-enhancement-eda\" target=\"_blank\">Learning to Sea: Underwater img Enhancement + EDA</a></p>\n<p>I'm a beginner and here for learning, I really thank those people who make this platform beginner-friendly and easy. </p>\n<p>Thanks for reading.</p>\n<p><strong>Update:</strong></p>\n<table>\n<thead>\n<tr>\n<th>Model</th>\n<th>Score</th>\n<th>Comment</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>YoloX-l</td>\n<td>0.517</td>\n<td>change in conf threshold 🚀</td>\n</tr>\n<tr>\n<td>YoloX-l Fine tuning + Tracking</td>\n<td>0.540</td>\n<td>Thanks to <a href=\"https://www.kaggle.com/parapapapam/yolox-inference-tracking-on-cots-lb-0-539\" target=\"_blank\">this notebook</a> 🚀   🚀</td>\n</tr>\n<tr>\n<td>Yolov5</td>\n<td>0.57+</td>\n<td></td>\n</tr>\n<tr>\n<td>Yolov5 + yoloX</td>\n<td>0.63+</td>\n<td><a href=\"https://www.kaggle.com/mahipalsingh/gbr-yolox-yolov5-ensemble-2-o\" target=\"_blank\">GBR: YOLOX+YOLOv5 Ensemble 2.o</a></td>\n</tr>\n</tbody>\n</table>\n<p><strong>What I did:</strong></p>\n<ol>\n<li>Implemented RetinaNet: <a href=\"https://www.kaggle.com/mahipalsingh/detection-using-keras-retinanet-inference\" target=\"_blank\">🌟🐟Detection using Keras-RetinaNet [Train/Inference]</a></li>\n<li><a href=\"https://www.kaggle.com/mahipalsingh/gbr-yolox-yolov5-ensemble-2-o\" target=\"_blank\">GBR: YOLOX+YOLOv5 Ensemble 2.o</a></li>\n</ol>\n<p>(Sorry for grammatical mistakes if any😅)</p>",
      "rawMarkdown": "Hello, kagglers\nMy last submission score 0.515 🔥. This is my first competition, Thanks to Kaggle, I'm learning a lot here. \nHere is my journey from score 0.038 ---> 0.515 :\n1. I started with [ 🌟🐟Detection using Keras-RetinaNet ](https://www.kaggle.com/mahipalsingh/detection-using-keras-retinanet) but After a lot of fine-tuning, I was able to score 0.038 only.\n2. I found [YoloX inference on Kaggle for COTS [LB 0.507]](https://www.kaggle.com/remekkinas/yolox-inference-on-kaggle-for-cots-lb-0-507/notebook) amazing work done by @remekkinas.\n3. I trained YoloX-S with a lot of different configurations and got a score of 0.246.\n4. I decided to go for yolox-l and see how it works..🤞\n5. Finally I kept trying with different parameters/config and my score was keep getting better and better. \n\nThe below table shows the model with the score I got till now:- \n| Model | Score | Comment |\n| --- | --- |\n| Keras-retinaNet | 0.006 | default config |\n| Keras-retinaNet | 0.038 | Hyperparameter Tunned |\n| YoloX-S | 0.166 | default config |\n| YoloX-S | 0.246 | Hyperparameter Tunned |\n| YoloX-l | 0.269 | default config |\n| YoloX-l | 0.448 | setting hyperparameter and config |\n| YoloX-l | 0.502 | more setting hyperparameter and config |\n| YoloX-l | 0.515 | more and more setting up hyperparameter and config 👐 |\n\nAlso, please do check out this work [Learning to Sea: Underwater img Enhancement + EDA](https://www.kaggle.com/soumya9977/learning-to-sea-underwater-img-enhancement-eda)\n\nI'm a beginner and here for learning, I really thank those people who make this platform beginner-friendly and easy. \n\nThanks for reading.\n\n**Update:**\n\n| Model | Score | Comment |\n| --- | --- |\n| YoloX-l | 0.517 | change in conf threshold 🚀  |\n| YoloX-l Fine tuning + Tracking | 0.540 | Thanks to [this notebook](https://www.kaggle.com/parapapapam/yolox-inference-tracking-on-cots-lb-0-539) 🚀   🚀  |\n| Yolov5 | 0.57+ |   |\n| Yolov5 + yoloX | 0.63+ | [GBR: YOLOX+YOLOv5 Ensemble 2.o](https://www.kaggle.com/mahipalsingh/gbr-yolox-yolov5-ensemble-2-o) |\n\n**What I did:**\n1. Implemented RetinaNet: [🌟🐟Detection using Keras-RetinaNet [Train/Inference]](https://www.kaggle.com/mahipalsingh/detection-using-keras-retinanet-inference)\n2. [GBR: YOLOX+YOLOv5 Ensemble 2.o](https://www.kaggle.com/mahipalsingh/gbr-yolox-yolov5-ensemble-2-o)\n\n(Sorry for grammatical mistakes if any😅)",
      "votes": 54
    },
    {
      "id": 1612762,
      "postDate": "2021-12-09T09:44:16.997Z",
      "content": "<p>yolox-l, with default setting, change input size to (960, 960), then LB 0.482.</p>",
      "rawMarkdown": "yolox-l, with default setting, change input size to (960, 960), then LB 0.482.",
      "votes": 6,
      "replies": [
        {
          "id": 1619552,
          "postDate": "2021-12-16T03:36:26.753Z",
          "rawMarkdown": "",
          "isDeleted": true
        },
        {
          "id": 1624973,
          "postDate": "2021-12-21T12:12:45.817Z",
          "content": "<p>Train more epochs <a href=\"https://www.kaggle.com/darshangovindaraj\" target=\"_blank\">@darshangovindaraj</a> </p>",
          "rawMarkdown": "Train more epochs @darshangovindaraj "
        }
      ]
    },
    {
      "id": 1631118,
      "postDate": "2021-12-28T03:57:37.643Z",
      "content": "<p>Hi! I am working on the YOLOV5 models!</p>\n<p>which have been I trained a lot of models in YOLOV5 like S, M, L, X, in that case, I cannot achieve the Score Compare with YOLOX. In default Config of YOLOV5 is perform better than the changed YOLOV5(Transfer Layer as a Head Layer ). When I am changing the Hyperparameters it has a little bit of increase in Accuracy.</p>\n<p>I cannot understand What is wrong with the concept I followed and studied in YOLOV5 GitHub. Any Suggestions to improve my model Accuracy :)</p>\n<p>I know how many hours to spend time in training &amp; Clearing the error! Thanks for sharing wonderful information brother )</p>",
      "rawMarkdown": "Hi! I am working on the YOLOV5 models!\n\nwhich have been I trained a lot of models in YOLOV5 like S, M, L, X, in that case, I cannot achieve the Score Compare with YOLOX. In default Config of YOLOV5 is perform better than the changed YOLOV5(Transfer Layer as a Head Layer ). When I am changing the Hyperparameters it has a little bit of increase in Accuracy.\n\nI cannot understand What is wrong with the concept I followed and studied in YOLOV5 GitHub. Any Suggestions to improve my model Accuracy :)\n\nI know how many hours to spend time in training & Clearing the error! Thanks for sharing wonderful information brother )\n\n",
      "votes": 3,
      "replies": [
        {
          "id": 1631159,
          "postDate": "2021-12-28T05:21:07.577Z",
          "content": "<p>Hi <a href=\"https://www.kaggle.com/balasubramaniamv\" target=\"_blank\">@balasubramaniamv</a>, <br>\nSame here, I tried Yolox-s, l, default Yolov5, and Yolov5 with transformer-module. In my case, Yolox is performing much better than yolov5.</p>",
          "rawMarkdown": "Hi @balasubramaniamv, \nSame here, I tried Yolox-s, l, default Yolov5, and Yolov5 with transformer-module. In my case, Yolox is performing much better than yolov5.",
          "votes": 1
        }
      ]
    },
    {
      "id": 1656592,
      "postDate": "2022-01-19T13:18:25.020Z",
      "content": "<p>Such a great journey. Thank you for sharing this :)</p>",
      "rawMarkdown": "Such a great journey. Thank you for sharing this :)",
      "votes": 1
    },
    {
      "id": 1632337,
      "postDate": "2021-12-29T14:05:08.760Z",
      "content": "<p>Good for you, I might say. Because not every beginner actually finds the platform beginner-friendly and easy. I guess it all boils down to each beginner's willingness to learn, not how easy or difficult the platform is.</p>",
      "rawMarkdown": "Good for you, I might say. Because not every beginner actually finds the platform beginner-friendly and easy. I guess it all boils down to each beginner's willingness to learn, not how easy or difficult the platform is.",
      "votes": 1,
      "replies": [
        {
          "id": 1632354,
          "postDate": "2021-12-29T14:34:17.930Z",
          "content": "<p>True, <a href=\"https://www.kaggle.com/adeyemiadewole\" target=\"_blank\">@adeyemiadewole</a> <br>\nWillingness is always a key here.<br>\nI saw a lot of people here, who started just like me, Now they are experts/grandmasters and still learning. They are always available to guide the beginners, share strategies and methods. Small support from experts increases enthusiasm. that's what I'm thankful for, to this platform. :)</p>",
          "rawMarkdown": "True, @adeyemiadewole \nWillingness is always a key here.\nI saw a lot of people here, who started just like me, Now they are experts/grandmasters and still learning. They are always available to guide the beginners, share strategies and methods. Small support from experts increases enthusiasm. that's what I'm thankful for, to this platform. :)"
        },
        {
          "id": 1632359,
          "postDate": "2021-12-29T14:38:35.480Z",
          "content": "<p>Good insides! I agree! </p>",
          "rawMarkdown": "Good insides! I agree! "
        }
      ]
    },
    {
      "id": 1630933,
      "postDate": "2021-12-27T21:19:42.533Z",
      "content": "<p>Hi Mahipal, thanks for sharing your progress.<br>\nMost of the improvements you shared are related to \"hyperparameter and config\". Could you explain what \"heuristic\" you use to explore hyperparams and config changes?<br>\nI'm trying to learn what the most efficient iteration practices are here, in order to avoid wasting resources and time</p>",
      "rawMarkdown": "Hi Mahipal, thanks for sharing your progress.\nMost of the improvements you shared are related to \"hyperparameter and config\". Could you explain what \"heuristic\" you use to explore hyperparams and config changes?\nI'm trying to learn what the most efficient iteration practices are here, in order to avoid wasting resources and time",
      "votes": 1,
      "replies": [
        {
          "id": 1631150,
          "postDate": "2021-12-28T05:00:50.930Z",
          "content": "<p>Hi, <a href=\"https://www.kaggle.com/diegoalejogm\" target=\"_blank\">@diegoalejogm</a> <br>\nI'm new to this, I think without wasting resources no one can actually learn😅, I wasted a lot of resources and time training models with different parameters, But also improve my understanding of the parameters and config. Although I don't blindly change parameters, There are a lot of articles available on \"parameters in YOLO-series\". For now, I just played around with different parameters like image_size, flip_prob, epochs. Some helped in improving scores :) and some not :( , still there are a lot of parameters to explore.</p>",
          "rawMarkdown": "Hi, @diegoalejogm \nI'm new to this, I think without wasting resources no one can actually learn😅, I wasted a lot of resources and time training models with different parameters, But also improve my understanding of the parameters and config. Although I don't blindly change parameters, There are a lot of articles available on \"parameters in YOLO-series\". For now, I just played around with different parameters like image_size, flip_prob, epochs. Some helped in improving scores :) and some not :( , still there are a lot of parameters to explore.",
          "votes": 2
        }
      ]
    },
    {
      "id": 1615201,
      "postDate": "2021-12-12T01:45:59.623Z",
      "content": "<p>Great work Mahipal</p>",
      "rawMarkdown": "Great work Mahipal",
      "votes": 1,
      "replies": [
        {
          "id": 1615271,
          "postDate": "2021-12-12T04:57:49.287Z",
          "content": "<p>Thanks… :)</p>",
          "rawMarkdown": "Thanks... :)",
          "votes": 1
        },
        {
          "id": 1616039,
          "postDate": "2021-12-13T04:45:34.650Z",
          "rawMarkdown": "",
          "isDeleted": true
        }
      ]
    },
    {
      "id": 1612622,
      "postDate": "2021-12-09T06:59:02.020Z",
      "content": "<p>I am really really happy! 👍👍🙏💪💪😍</p>",
      "rawMarkdown": "I am really really happy! 👍👍🙏💪💪😍",
      "votes": 2,
      "replies": [
        {
          "id": 1612632,
          "postDate": "2021-12-09T07:08:40.903Z",
          "content": "<p>Thanks, credit to your work <a href=\"https://www.kaggle.com/remekkinas\" target=\"_blank\">@remekkinas</a> 🤝🤍</p>",
          "rawMarkdown": "Thanks, credit to your work @remekkinas 🤝🤍",
          "votes": 1
        }
      ]
    },
    {
      "id": 1640470,
      "postDate": "2022-01-06T14:02:41.857Z",
      "content": "<p>Hello, your work <code>Learning to Sea: Underwater img Enhancement + EDA</code> is amazing 🔥. I tuned some hyperparameters but recall is too low. Is there any tip to increase recall. Thanks  </p>",
      "rawMarkdown": "Hello, your work `Learning to Sea: Underwater img Enhancement + EDA` is amazing 🔥. I tuned some hyperparameters but recall is too low. Is there any tip to increase recall. Thanks  "
    },
    {
      "id": 1638145,
      "postDate": "2022-01-04T13:42:38.233Z",
      "content": "<p>I'm also a beginner, can I ask The YoloX-l Fine tuning + Tracking goes to 0.540 (which base is 0.539) is comes from your custom training model or fine tune tracking hyperparameters (like distance_threshold, hit_inertia_mi, hit_inertia_max, initialization_delay)? i'm playing with those but still haven't got more than 0.539.  😐</p>",
      "rawMarkdown": "I'm also a beginner, can I ask The YoloX-l Fine tuning + Tracking goes to 0.540 (which base is 0.539) is comes from your custom training model or fine tune tracking hyperparameters (like distance_threshold, hit_inertia_mi, hit_inertia_max, initialization_delay)? i'm playing with those but still haven't got more than 0.539.  😐",
      "replies": [
        {
          "id": 1639231,
          "postDate": "2022-01-05T13:54:12.343Z",
          "content": "<p>nvm, i just reached 0.540. happy learning. 💪💪💪</p>",
          "rawMarkdown": "nvm, i just reached 0.540. happy learning. 💪💪💪"
        }
      ]
    },
    {
      "id": 1630966,
      "postDate": "2021-12-27T21:58:39.873Z",
      "content": "<p>Did you get any results by splitting in more/less folds?</p>",
      "rawMarkdown": "Did you get any results by splitting in more/less folds?",
      "replies": [
        {
          "id": 1631152,
          "postDate": "2021-12-28T05:03:17.920Z",
          "content": "<p>Using more data for training, less for validation improves the score. </p>",
          "rawMarkdown": "Using more data for training, less for validation improves the score. "
        }
      ]
    },
    {
      "id": 1628518,
      "postDate": "2021-12-25T02:02:53.940Z",
      "content": "<p>I am a beginner too. Love to see a fellow beginner succeed. 🎉</p>",
      "rawMarkdown": "I am a beginner too. Love to see a fellow beginner succeed. 🎉",
      "replies": [
        {
          "id": 1628659,
          "postDate": "2021-12-25T06:00:04.650Z",
          "content": "<p>Thanks.. <a href=\"https://www.kaggle.com/lightmk\" target=\"_blank\">@lightmk</a> </p>",
          "rawMarkdown": "Thanks.. @lightmk "
        }
      ]
    },
    {
      "id": 1619487,
      "postDate": "2021-12-16T01:57:05.400Z",
      "content": "<p>when I use yolox-l, it occurred 'CUDA out of memory'.<br>\nHow did you trained yolox-l on kaggle notebook? </p>",
      "rawMarkdown": "when I use yolox-l, it occurred 'CUDA out of memory'.\nHow did you trained yolox-l on kaggle notebook? ",
      "replies": [
        {
          "id": 1619554,
          "postDate": "2021-12-16T03:36:54.420Z",
          "rawMarkdown": "",
          "votes": 2,
          "isDeleted": true
        }
      ]
    },
    {
      "id": 1616696,
      "postDate": "2021-12-13T15:21:12.887Z",
      "content": "<p>Awesome! Great to see the learning journey. </p>\n<p>Also \"Learning to Sea\" 😀is probably the best EDA title for this competition easily. It's also quite a well-done notebook. </p>\n<p>Keep up the awesome work!</p>",
      "rawMarkdown": "Awesome! Great to see the learning journey. \n\nAlso \"Learning to Sea\" 😀is probably the best EDA title for this competition easily. It's also quite a well-done notebook. \n\nKeep up the awesome work!",
      "replies": [
        {
          "id": 1616907,
          "postDate": "2021-12-13T17:44:34.923Z",
          "content": "<p>Thanks.. Yeah, there are also a lot more great notbook. </p>",
          "rawMarkdown": "Thanks.. Yeah, there are also a lot more great notbook. "
        }
      ]
    },
    {
      "id": 1614882,
      "postDate": "2021-12-11T14:38:24.553Z",
      "content": "<p>Interesting post! I also got a score of 0.27 with YoloX-l with standard configs. Therefore, I think you also use this train/test split: <a href=\"https://www.kaggle.com/awsaf49/great-barrier-reef-yolov5-train\" target=\"_blank\">https://www.kaggle.com/awsaf49/great-barrier-reef-yolov5-train</a> ?</p>\n<p>Do you use an input size of (960, 960)? And on what configs did you focus to get from 0.27 to 0.448 (big increase)?</p>",
      "rawMarkdown": "Interesting post! I also got a score of 0.27 with YoloX-l with standard configs. Therefore, I think you also use this train/test split: https://www.kaggle.com/awsaf49/great-barrier-reef-yolov5-train ?\n\nDo you use an input size of (960, 960)? And on what configs did you focus to get from 0.27 to 0.448 (big increase)?",
      "replies": [
        {
          "id": 1615276,
          "postDate": "2021-12-12T05:06:14.863Z",
          "content": "<p>Yes, I used 960 sizes, I suggest keep changing training parameters you will find a way. setting lower nmsthre helped!</p>",
          "rawMarkdown": "Yes, I used 960 sizes, I suggest keep changing training parameters you will find a way. setting lower nmsthre helped!",
          "votes": 3
        },
        {
          "id": 1617253,
          "postDate": "2021-12-14T01:40:42.883Z",
          "rawMarkdown": "",
          "isDeleted": true
        },
        {
          "id": 1619555,
          "postDate": "2021-12-16T03:38:00.947Z",
          "rawMarkdown": "",
          "isDeleted": true
        }
      ]
    },
    {
      "id": 1614139,
      "postDate": "2021-12-10T17:04:04.370Z",
      "content": "<p>good job! Have you tried img Enhancement and get improvements?</p>",
      "rawMarkdown": "good job! Have you tried img Enhancement and get improvements?",
      "replies": [
        {
          "id": 1614434,
          "postDate": "2021-12-11T05:10:16.867Z",
          "content": "<p>No, I didn't try yet </p>",
          "rawMarkdown": "No, I didn't try yet "
        },
        {
          "id": 1630935,
          "postDate": "2021-12-27T21:20:35.197Z",
          "content": "<p>What image enhancements are you referring to?<br>\nI see you're 3rd in the PB, congrats! Trying to learn as much as I can</p>",
          "rawMarkdown": "What image enhancements are you referring to?\nI see you're 3rd in the PB, congrats! Trying to learn as much as I can"
        },
        {
          "id": 1630968,
          "postDate": "2021-12-27T22:01:21.757Z",
          "content": "<p>Rule #467 - TOP guys only ask question 😄😄😂😆<br>\nI am interested in tip as well :)</p>",
          "rawMarkdown": "Rule #467 - TOP guys only ask question 😄😄😂😆\nI am interested in tip as well :)"
        },
        {
          "id": 1631051,
          "postDate": "2021-12-28T01:38:35.907Z",
          "content": "<p>Hahaha…………………</p>",
          "rawMarkdown": "Hahaha....................."
        }
      ]
    },
    {
      "id": 1613081,
      "postDate": "2021-12-09T15:21:09.377Z",
      "content": "<p>Where are you training your Model?<br>\nif it's on kaggle notebook how much time it is requiring to get trained</p>",
      "rawMarkdown": "Where are you training your Model?\nif it's on kaggle notebook how much time it is requiring to get trained",
      "replies": [
        {
          "id": 1613182,
          "postDate": "2021-12-09T17:42:14.913Z",
          "content": "<p>I trained the model on Kaggle. I Save weight to avoid time limit error</p>",
          "rawMarkdown": "I trained the model on Kaggle. I Save weight to avoid time limit error",
          "votes": 1
        }
      ]
    },
    {
      "id": 1612641,
      "postDate": "2021-12-09T07:17:39.417Z",
      "content": "<p><a href=\"https://www.kaggle.com/mahipalsingh\" target=\"_blank\">@mahipalsingh</a> how did you split the data for <code>train</code> and <code>validation</code> ?</p>",
      "rawMarkdown": "@mahipalsingh how did you split the data for `train` and `validation` ?",
      "replies": [
        {
          "id": 1612654,
          "postDate": "2021-12-09T07:28:37.333Z",
          "content": "<p>It is the same from [<a href=\"https://www.kaggle.com/remekkinas/yolox-inference-on-kaggle-for-cots-lb-0-507/notebook\" target=\"_blank\">https://www.kaggle.com/remekkinas/yolox-inference-on-kaggle-for-cots-lb-0-507/notebook</a>]</p>",
          "rawMarkdown": "It is the same from [https://www.kaggle.com/remekkinas/yolox-inference-on-kaggle-for-cots-lb-0-507/notebook]",
          "votes": 2
        },
        {
          "id": 1612729,
          "postDate": "2021-12-09T09:03:01.287Z",
          "content": "<p>As I can see <a href=\"https://www.kaggle.com/awsaf49\" target=\"_blank\">@awsaf49</a> it is your strategy :) </p>",
          "rawMarkdown": "As I can see @awsaf49 it is your strategy :) ",
          "votes": 1
        },
        {
          "id": 1612733,
          "postDate": "2021-12-09T09:04:40.070Z",
          "content": "<p>haha xD <a href=\"https://www.kaggle.com/remekkinas\" target=\"_blank\">@remekkinas</a> </p>",
          "rawMarkdown": "haha xD @remekkinas "
        },
        {
          "id": 1612739,
          "postDate": "2021-12-09T09:11:16.180Z",
          "content": "<p>Lack of data …. As I can see many models learn good but then require more more more data …. to go up.</p>",
          "rawMarkdown": "Lack of data .... As I can see many models learn good but then require more more more data .... to go up."
        }
      ]
    },
    {
      "id": 1650425,
      "postDate": "2022-01-15T03:29:43.680Z",
      "content": "<p>Thank you for sharing your journey!</p>",
      "rawMarkdown": "Thank you for sharing your journey!",
      "votes": 1
    }
  ],
  "comments": [
    {
      "id": 1612762,
      "author_name": "AndyYuan",
      "author_url": "",
      "post_date": "2021-12-09T09:44:16.997000",
      "content": "<p>yolox-l, with default setting, change input size to (960, 960), then LB 0.482.</p>",
      "votes": 6,
      "replies": [
        {
          "id": 1619552,
          "author_name": "",
          "author_url": "",
          "post_date": "2021-12-16T03:36:26.753000",
          "content": "",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1624973,
          "author_name": "Mahipal Singh",
          "author_url": "",
          "post_date": "2021-12-21T12:12:45.817000",
          "content": "<p>Train more epochs <a href=\"https://www.kaggle.com/darshangovindaraj\" target=\"_blank\">@darshangovindaraj</a> </p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 1631118,
      "author_name": "Balasubramaniam",
      "author_url": "",
      "post_date": "2021-12-28T03:57:37.643000",
      "content": "<p>Hi! I am working on the YOLOV5 models!</p>\n<p>which have been I trained a lot of models in YOLOV5 like S, M, L, X, in that case, I cannot achieve the Score Compare with YOLOX. In default Config of YOLOV5 is perform better than the changed YOLOV5(Transfer Layer as a Head Layer ). When I am changing the Hyperparameters it has a little bit of increase in Accuracy.</p>\n<p>I cannot understand What is wrong with the concept I followed and studied in YOLOV5 GitHub. Any Suggestions to improve my model Accuracy :)</p>\n<p>I know how many hours to spend time in training &amp; Clearing the error! Thanks for sharing wonderful information brother )</p>",
      "votes": 3,
      "replies": [
        {
          "id": 1631159,
          "author_name": "Mahipal Singh",
          "author_url": "",
          "post_date": "2021-12-28T05:21:07.577000",
          "content": "<p>Hi <a href=\"https://www.kaggle.com/balasubramaniamv\" target=\"_blank\">@balasubramaniamv</a>, <br>\nSame here, I tried Yolox-s, l, default Yolov5, and Yolov5 with transformer-module. In my case, Yolox is performing much better than yolov5.</p>",
          "votes": 1,
          "replies": []
        }
      ]
    },
    {
      "id": 1656592,
      "author_name": "Eugene J. Ryu",
      "author_url": "",
      "post_date": "2022-01-19T13:18:25.020000",
      "content": "<p>Such a great journey. Thank you for sharing this :)</p>",
      "votes": 1,
      "replies": []
    },
    {
      "id": 1632337,
      "author_name": "Adeyemi Adewole",
      "author_url": "",
      "post_date": "2021-12-29T14:05:08.760000",
      "content": "<p>Good for you, I might say. Because not every beginner actually finds the platform beginner-friendly and easy. I guess it all boils down to each beginner's willingness to learn, not how easy or difficult the platform is.</p>",
      "votes": 1,
      "replies": [
        {
          "id": 1632354,
          "author_name": "Mahipal Singh",
          "author_url": "",
          "post_date": "2021-12-29T14:34:17.930000",
          "content": "<p>True, <a href=\"https://www.kaggle.com/adeyemiadewole\" target=\"_blank\">@adeyemiadewole</a> <br>\nWillingness is always a key here.<br>\nI saw a lot of people here, who started just like me, Now they are experts/grandmasters and still learning. They are always available to guide the beginners, share strategies and methods. Small support from experts increases enthusiasm. that's what I'm thankful for, to this platform. :)</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1632359,
          "author_name": "Remek Kinas",
          "author_url": "",
          "post_date": "2021-12-29T14:38:35.480000",
          "content": "<p>Good insides! I agree! </p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 1630933,
      "author_name": "Diego Gomez",
      "author_url": "",
      "post_date": "2021-12-27T21:19:42.533000",
      "content": "<p>Hi Mahipal, thanks for sharing your progress.<br>\nMost of the improvements you shared are related to \"hyperparameter and config\". Could you explain what \"heuristic\" you use to explore hyperparams and config changes?<br>\nI'm trying to learn what the most efficient iteration practices are here, in order to avoid wasting resources and time</p>",
      "votes": 1,
      "replies": [
        {
          "id": 1631150,
          "author_name": "Mahipal Singh",
          "author_url": "",
          "post_date": "2021-12-28T05:00:50.930000",
          "content": "<p>Hi, <a href=\"https://www.kaggle.com/diegoalejogm\" target=\"_blank\">@diegoalejogm</a> <br>\nI'm new to this, I think without wasting resources no one can actually learn😅, I wasted a lot of resources and time training models with different parameters, But also improve my understanding of the parameters and config. Although I don't blindly change parameters, There are a lot of articles available on \"parameters in YOLO-series\". For now, I just played around with different parameters like image_size, flip_prob, epochs. Some helped in improving scores :) and some not :( , still there are a lot of parameters to explore.</p>",
          "votes": 2,
          "replies": []
        }
      ]
    },
    {
      "id": 1615201,
      "author_name": "Chris Deotte",
      "author_url": "",
      "post_date": "2021-12-12T01:45:59.623000",
      "content": "<p>Great work Mahipal</p>",
      "votes": 1,
      "replies": [
        {
          "id": 1615271,
          "author_name": "Mahipal Singh",
          "author_url": "",
          "post_date": "2021-12-12T04:57:49.287000",
          "content": "<p>Thanks… :)</p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 1616039,
          "author_name": "",
          "author_url": "",
          "post_date": "2021-12-13T04:45:34.650000",
          "content": "",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 1612622,
      "author_name": "Remek Kinas",
      "author_url": "",
      "post_date": "2021-12-09T06:59:02.020000",
      "content": "<p>I am really really happy! 👍👍🙏💪💪😍</p>",
      "votes": 2,
      "replies": [
        {
          "id": 1612632,
          "author_name": "Mahipal Singh",
          "author_url": "",
          "post_date": "2021-12-09T07:08:40.903000",
          "content": "<p>Thanks, credit to your work <a href=\"https://www.kaggle.com/remekkinas\" target=\"_blank\">@remekkinas</a> 🤝🤍</p>",
          "votes": 1,
          "replies": []
        }
      ]
    },
    {
      "id": 1640470,
      "author_name": "Tanjirooo",
      "author_url": "",
      "post_date": "2022-01-06T14:02:41.857000",
      "content": "<p>Hello, your work <code>Learning to Sea: Underwater img Enhancement + EDA</code> is amazing 🔥. I tuned some hyperparameters but recall is too low. Is there any tip to increase recall. Thanks  </p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 1638145,
      "author_name": "Bao Loc Pham",
      "author_url": "",
      "post_date": "2022-01-04T13:42:38.233000",
      "content": "<p>I'm also a beginner, can I ask The YoloX-l Fine tuning + Tracking goes to 0.540 (which base is 0.539) is comes from your custom training model or fine tune tracking hyperparameters (like distance_threshold, hit_inertia_mi, hit_inertia_max, initialization_delay)? i'm playing with those but still haven't got more than 0.539.  😐</p>",
      "votes": 0,
      "replies": [
        {
          "id": 1639231,
          "author_name": "Bao Loc Pham",
          "author_url": "",
          "post_date": "2022-01-05T13:54:12.343000",
          "content": "<p>nvm, i just reached 0.540. happy learning. 💪💪💪</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 1630966,
      "author_name": "Diego Gomez",
      "author_url": "",
      "post_date": "2021-12-27T21:58:39.873000",
      "content": "<p>Did you get any results by splitting in more/less folds?</p>",
      "votes": 0,
      "replies": [
        {
          "id": 1631152,
          "author_name": "Mahipal Singh",
          "author_url": "",
          "post_date": "2021-12-28T05:03:17.920000",
          "content": "<p>Using more data for training, less for validation improves the score. </p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 1628518,
      "author_name": "Lin Myat Ko",
      "author_url": "",
      "post_date": "2021-12-25T02:02:53.940000",
      "content": "<p>I am a beginner too. Love to see a fellow beginner succeed. 🎉</p>",
      "votes": 0,
      "replies": [
        {
          "id": 1628659,
          "author_name": "Mahipal Singh",
          "author_url": "",
          "post_date": "2021-12-25T06:00:04.650000",
          "content": "<p>Thanks.. <a href=\"https://www.kaggle.com/lightmk\" target=\"_blank\">@lightmk</a> </p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 1619487,
      "author_name": "Andy Laer",
      "author_url": "",
      "post_date": "2021-12-16T01:57:05.400000",
      "content": "<p>when I use yolox-l, it occurred 'CUDA out of memory'.<br>\nHow did you trained yolox-l on kaggle notebook? </p>",
      "votes": 0,
      "replies": [
        {
          "id": 1619554,
          "author_name": "",
          "author_url": "",
          "post_date": "2021-12-16T03:36:54.420000",
          "content": "",
          "votes": 2,
          "replies": []
        }
      ]
    },
    {
      "id": 1616696,
      "author_name": "Darien Schettler",
      "author_url": "",
      "post_date": "2021-12-13T15:21:12.887000",
      "content": "<p>Awesome! Great to see the learning journey. </p>\n<p>Also \"Learning to Sea\" 😀is probably the best EDA title for this competition easily. It's also quite a well-done notebook. </p>\n<p>Keep up the awesome work!</p>",
      "votes": 0,
      "replies": [
        {
          "id": 1616907,
          "author_name": "Mahipal Singh",
          "author_url": "",
          "post_date": "2021-12-13T17:44:34.923000",
          "content": "<p>Thanks.. Yeah, there are also a lot more great notbook. </p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 1614882,
      "author_name": "Meesz9",
      "author_url": "",
      "post_date": "2021-12-11T14:38:24.553000",
      "content": "<p>Interesting post! I also got a score of 0.27 with YoloX-l with standard configs. Therefore, I think you also use this train/test split: <a href=\"https://www.kaggle.com/awsaf49/great-barrier-reef-yolov5-train\" target=\"_blank\">https://www.kaggle.com/awsaf49/great-barrier-reef-yolov5-train</a> ?</p>\n<p>Do you use an input size of (960, 960)? And on what configs did you focus to get from 0.27 to 0.448 (big increase)?</p>",
      "votes": 0,
      "replies": [
        {
          "id": 1615276,
          "author_name": "Mahipal Singh",
          "author_url": "",
          "post_date": "2021-12-12T05:06:14.863000",
          "content": "<p>Yes, I used 960 sizes, I suggest keep changing training parameters you will find a way. setting lower nmsthre helped!</p>",
          "votes": 3,
          "replies": []
        },
        {
          "id": 1617253,
          "author_name": "",
          "author_url": "",
          "post_date": "2021-12-14T01:40:42.883000",
          "content": "",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1619555,
          "author_name": "",
          "author_url": "",
          "post_date": "2021-12-16T03:38:00.947000",
          "content": "",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 1614139,
      "author_name": "clwclw",
      "author_url": "",
      "post_date": "2021-12-10T17:04:04.370000",
      "content": "<p>good job! Have you tried img Enhancement and get improvements?</p>",
      "votes": 0,
      "replies": [
        {
          "id": 1614434,
          "author_name": "Mahipal Singh",
          "author_url": "",
          "post_date": "2021-12-11T05:10:16.867000",
          "content": "<p>No, I didn't try yet </p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1630935,
          "author_name": "Diego Gomez",
          "author_url": "",
          "post_date": "2021-12-27T21:20:35.197000",
          "content": "<p>What image enhancements are you referring to?<br>\nI see you're 3rd in the PB, congrats! Trying to learn as much as I can</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1630968,
          "author_name": "Remek Kinas",
          "author_url": "",
          "post_date": "2021-12-27T22:01:21.757000",
          "content": "<p>Rule #467 - TOP guys only ask question 😄😄😂😆<br>\nI am interested in tip as well :)</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1631051,
          "author_name": "DeepInvolution",
          "author_url": "",
          "post_date": "2021-12-28T01:38:35.907000",
          "content": "<p>Hahaha…………………</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 1613081,
      "author_name": "DeepUnderstanding",
      "author_url": "",
      "post_date": "2021-12-09T15:21:09.377000",
      "content": "<p>Where are you training your Model?<br>\nif it's on kaggle notebook how much time it is requiring to get trained</p>",
      "votes": 0,
      "replies": [
        {
          "id": 1613182,
          "author_name": "Mahipal Singh",
          "author_url": "",
          "post_date": "2021-12-09T17:42:14.913000",
          "content": "<p>I trained the model on Kaggle. I Save weight to avoid time limit error</p>",
          "votes": 1,
          "replies": []
        }
      ]
    },
    {
      "id": 1612641,
      "author_name": "Awsaf",
      "author_url": "",
      "post_date": "2021-12-09T07:17:39.417000",
      "content": "<p><a href=\"https://www.kaggle.com/mahipalsingh\" target=\"_blank\">@mahipalsingh</a> how did you split the data for <code>train</code> and <code>validation</code> ?</p>",
      "votes": 0,
      "replies": [
        {
          "id": 1612654,
          "author_name": "Mahipal Singh",
          "author_url": "",
          "post_date": "2021-12-09T07:28:37.333000",
          "content": "<p>It is the same from [<a href=\"https://www.kaggle.com/remekkinas/yolox-inference-on-kaggle-for-cots-lb-0-507/notebook\" target=\"_blank\">https://www.kaggle.com/remekkinas/yolox-inference-on-kaggle-for-cots-lb-0-507/notebook</a>]</p>",
          "votes": 2,
          "replies": []
        },
        {
          "id": 1612729,
          "author_name": "Remek Kinas",
          "author_url": "",
          "post_date": "2021-12-09T09:03:01.287000",
          "content": "<p>As I can see <a href=\"https://www.kaggle.com/awsaf49\" target=\"_blank\">@awsaf49</a> it is your strategy :) </p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 1612733,
          "author_name": "Awsaf",
          "author_url": "",
          "post_date": "2021-12-09T09:04:40.070000",
          "content": "<p>haha xD <a href=\"https://www.kaggle.com/remekkinas\" target=\"_blank\">@remekkinas</a> </p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1612739,
          "author_name": "Remek Kinas",
          "author_url": "",
          "post_date": "2021-12-09T09:11:16.180000",
          "content": "<p>Lack of data …. As I can see many models learn good but then require more more more data …. to go up.</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 1650425,
      "author_name": "jellybeanz",
      "author_url": "",
      "post_date": "2022-01-15T03:29:43.680000",
      "content": "<p>Thank you for sharing your journey!</p>",
      "votes": 1,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1612609": "Hello, kagglers\nMy last submission score 0.515 🔥. This is my first competition, Thanks to Kaggle, I'm learning a lot here. \nHere is my journey from score 0.038 ---> 0.515 :\n1. I started with [ 🌟🐟Detection using Keras-RetinaNet ](https://www.kaggle.com/mahipalsingh/detection-using-keras-retinanet) but After a lot of fine-tuning, I was able to score 0.038 only.\n2. I found [YoloX inference on Kaggle for COTS [LB 0.507]](https://www.kaggle.com/remekkinas/yolox-inference-on-kaggle-for-cots-lb-0-507/notebook) amazing work done by @remekkinas.\n3. I trained YoloX-S with a lot of different configurations and got a score of 0.246.\n4. I decided to go for yolox-l and see how it works..🤞\n5. Finally I kept trying with different parameters/config and my score was keep getting better and better. \n\nThe below table shows the model with the score I got till now:- \n| Model | Score | Comment |\n| --- | --- |\n| Keras-retinaNet | 0.006 | default config |\n| Keras-retinaNet | 0.038 | Hyperparameter Tunned |\n| YoloX-S | 0.166 | default config |\n| YoloX-S | 0.246 | Hyperparameter Tunned |\n| YoloX-l | 0.269 | default config |\n| YoloX-l | 0.448 | setting hyperparameter and config |\n| YoloX-l | 0.502 | more setting hyperparameter and config |\n| YoloX-l | 0.515 | more and more setting up hyperparameter and config 👐 |\n\nAlso, please do check out this work [Learning to Sea: Underwater img Enhancement + EDA](https://www.kaggle.com/soumya9977/learning-to-sea-underwater-img-enhancement-eda)\n\nI'm a beginner and here for learning, I really thank those people who make this platform beginner-friendly and easy. \n\nThanks for reading.\n\n**Update:**\n\n| Model | Score | Comment |\n| --- | --- |\n| YoloX-l | 0.517 | change in conf threshold 🚀  |\n| YoloX-l Fine tuning + Tracking | 0.540 | Thanks to [this notebook](https://www.kaggle.com/parapapapam/yolox-inference-tracking-on-cots-lb-0-539) 🚀   🚀  |\n| Yolov5 | 0.57+ |   |\n| Yolov5 + yoloX | 0.63+ | [GBR: YOLOX+YOLOv5 Ensemble 2.o](https://www.kaggle.com/mahipalsingh/gbr-yolox-yolov5-ensemble-2-o) |\n\n**What I did:**\n1. Implemented RetinaNet: [🌟🐟Detection using Keras-RetinaNet [Train/Inference]](https://www.kaggle.com/mahipalsingh/detection-using-keras-retinanet-inference)\n2. [GBR: YOLOX+YOLOv5 Ensemble 2.o](https://www.kaggle.com/mahipalsingh/gbr-yolox-yolov5-ensemble-2-o)\n\n(Sorry for grammatical mistakes if any😅)",
    "1612762": "yolox-l, with default setting, change input size to (960, 960), then LB 0.482.",
    "1631118": "Hi! I am working on the YOLOV5 models!\n\nwhich have been I trained a lot of models in YOLOV5 like S, M, L, X, in that case, I cannot achieve the Score Compare with YOLOX. In default Config of YOLOV5 is perform better than the changed YOLOV5(Transfer Layer as a Head Layer ). When I am changing the Hyperparameters it has a little bit of increase in Accuracy.\n\nI cannot understand What is wrong with the concept I followed and studied in YOLOV5 GitHub. Any Suggestions to improve my model Accuracy :)\n\nI know how many hours to spend time in training & Clearing the error! Thanks for sharing wonderful information brother )\n\n",
    "1656592": "Such a great journey. Thank you for sharing this :)",
    "1632337": "Good for you, I might say. Because not every beginner actually finds the platform beginner-friendly and easy. I guess it all boils down to each beginner's willingness to learn, not how easy or difficult the platform is.",
    "1630933": "Hi Mahipal, thanks for sharing your progress.\nMost of the improvements you shared are related to \"hyperparameter and config\". Could you explain what \"heuristic\" you use to explore hyperparams and config changes?\nI'm trying to learn what the most efficient iteration practices are here, in order to avoid wasting resources and time",
    "1615201": "Great work Mahipal",
    "1612622": "I am really really happy! 👍👍🙏💪💪😍",
    "1640470": "Hello, your work `Learning to Sea: Underwater img Enhancement + EDA` is amazing 🔥. I tuned some hyperparameters but recall is too low. Is there any tip to increase recall. Thanks  ",
    "1638145": "I'm also a beginner, can I ask The YoloX-l Fine tuning + Tracking goes to 0.540 (which base is 0.539) is comes from your custom training model or fine tune tracking hyperparameters (like distance_threshold, hit_inertia_mi, hit_inertia_max, initialization_delay)? i'm playing with those but still haven't got more than 0.539.  😐",
    "1630966": "Did you get any results by splitting in more/less folds?",
    "1628518": "I am a beginner too. Love to see a fellow beginner succeed. 🎉",
    "1619487": "when I use yolox-l, it occurred 'CUDA out of memory'.\nHow did you trained yolox-l on kaggle notebook? ",
    "1616696": "Awesome! Great to see the learning journey. \n\nAlso \"Learning to Sea\" 😀is probably the best EDA title for this competition easily. It's also quite a well-done notebook. \n\nKeep up the awesome work!",
    "1614882": "Interesting post! I also got a score of 0.27 with YoloX-l with standard configs. Therefore, I think you also use this train/test split: https://www.kaggle.com/awsaf49/great-barrier-reef-yolov5-train ?\n\nDo you use an input size of (960, 960)? And on what configs did you focus to get from 0.27 to 0.448 (big increase)?",
    "1614139": "good job! Have you tried img Enhancement and get improvements?",
    "1613081": "Where are you training your Model?\nif it's on kaggle notebook how much time it is requiring to get trained",
    "1612641": "@mahipalsingh how did you split the data for `train` and `validation` ?",
    "1650425": "Thank you for sharing your journey!"
  }
}