{
  "id": 226478,
  "title": "(Help me)My yolov5 baseline scores too low. what might it be a problem?",
  "url": "/competitions/vinbigdata-chest-xray-abnormalities-detection/discussion/226478",
  "author_name": "kaggler",
  "post_date": "2021-03-16T15:31:55.560000",
  "votes": 9,
  "comment_count": 38,
  "views": 0,
  "content": "<p>Thank you for reading this<br>\nI am the newbie of object-detection <br>\nI have been studying object detection for school homework</p>\n<p>My yolov5 scores too low</p>\n<ul>\n<li><p>14classes(training with just abnormal)</p></li>\n<li><p>Image size : 1024 x 1024 resized</p></li>\n<li><p>100 epochs.</p></li>\n<li><p>Stratified with iterative-stratification</p></li>\n<li><p>No 2 class filter</p></li>\n<li><p>VALIDATION MAP@.5 scores 0.35, </p></li>\n<li><p>submission -&gt; yolo detect.py with conf 0.05, iou 0.4</p></li>\n</ul>\n<p><strong>But, my submission just scores around 0.11</strong><br>\njust 512 image size baseline scores above 0.15<br>\nwhat can be a problem?</p>\n<p>Thank you for reading this. </p>",
  "messages": [
    {
      "id": 1240719,
      "postDate": "2021-03-16T15:31:55.560Z",
      "content": "<p>Thank you for reading this<br>\nI am the newbie of object-detection <br>\nI have been studying object detection for school homework</p>\n<p>My yolov5 scores too low</p>\n<ul>\n<li><p>14classes(training with just abnormal)</p></li>\n<li><p>Image size : 1024 x 1024 resized</p></li>\n<li><p>100 epochs.</p></li>\n<li><p>Stratified with iterative-stratification</p></li>\n<li><p>No 2 class filter</p></li>\n<li><p>VALIDATION MAP@.5 scores 0.35, </p></li>\n<li><p>submission -&gt; yolo detect.py with conf 0.05, iou 0.4</p></li>\n</ul>\n<p><strong>But, my submission just scores around 0.11</strong><br>\njust 512 image size baseline scores above 0.15<br>\nwhat can be a problem?</p>\n<p>Thank you for reading this. </p>",
      "rawMarkdown": "Thank you for reading this\nI am the newbie of object-detection \nI have been studying object detection for school homework\n\nMy yolov5 scores too low\n\n* 14classes(training with just abnormal)\n* Image size : 1024 x 1024 resized\n* 100 epochs.\n* Stratified with iterative-stratification\n* No 2 class filter\n\n* VALIDATION MAP@.5 scores 0.35, \n* submission -> yolo detect.py with conf 0.05, iou 0.4\n\n**But, my submission just scores around 0.11**\njust 512 image size baseline scores above 0.15\nwhat can be a problem?\n\nThank you for reading this. \n\n\n",
      "votes": 9
    },
    {
      "id": 1243357,
      "postDate": "2021-03-18T06:49:27.920Z",
      "content": "<p>For yolov5, I suggest u should start with hyp.scratch.yaml file, set <code>mixup</code> to 0. With yolov5x, I see  it converge before 50 epochs. I can get baseline score above 0.15 in with a fold, use 2 class filter to greatly improve your LB score </p>",
      "rawMarkdown": "For yolov5, I suggest u should start with hyp.scratch.yaml file, set `mixup` to 0. With yolov5x, I see  it converge before 50 epochs. I can get baseline score above 0.15 in with a fold, use 2 class filter to greatly improve your LB score ",
      "votes": 3,
      "replies": [
        {
          "id": 1243694,
          "postDate": "2021-03-18T12:25:04.463Z",
          "content": "<p>ok. I am gonna try 2 classifier. thanks a lot! good luck!</p>",
          "rawMarkdown": "ok. I am gonna try 2 classifier. thanks a lot! good luck!"
        },
        {
          "id": 1250393,
          "postDate": "2021-03-24T02:04:27.290Z",
          "content": "<p>What do you mean by 2 class filter? I'm pretty new to this…</p>",
          "rawMarkdown": "What do you mean by 2 class filter? I'm pretty new to this..."
        },
        {
          "id": 1251798,
          "postDate": "2021-03-25T06:26:01.423Z",
          "content": "<p>2 class filter is used to filter the no-finding images on testset</p>",
          "rawMarkdown": "2 class filter is used to filter the no-finding images on testset",
          "votes": 1
        }
      ]
    },
    {
      "id": 1241249,
      "postDate": "2021-03-17T00:56:06.463Z",
      "content": "<p>hmm, same problem.</p>",
      "rawMarkdown": "hmm, same problem.",
      "votes": 1,
      "replies": [
        {
          "id": 1241326,
          "postDate": "2021-03-17T02:05:17.110Z",
          "content": "<p>My single fold improved from 0.111 to 0.162<br>\n<a href=\"https://www.kaggle.com/adrielcabral\" target=\"_blank\">@adrielcabral</a> told me  \"multi-label\" in general.py file (def non_max_suppression(prediction, conf_thres=0.25, iou_thres=0.45, classes=None, agnostic=False, multi_label=True) line 393</p>\n<p>After setting multi-label = True, my score improved</p>",
          "rawMarkdown": "My single fold improved from 0.111 to 0.162\n@adrielcabral told me  \"multi-label\" in general.py file (def non_max_suppression(prediction, conf_thres=0.25, iou_thres=0.45, classes=None, agnostic=False, multi_label=True) line 393\n\nAfter setting multi-label = True, my score improved",
          "votes": 1
        },
        {
          "id": 1245746,
          "postDate": "2021-03-20T05:30:24.767Z",
          "content": "<p>I tried, still very low lb</p>",
          "rawMarkdown": "I tried, still very low lb",
          "votes": 1
        },
        {
          "id": 1252527,
          "postDate": "2021-03-25T18:44:40.420Z",
          "rawMarkdown": "",
          "isDeleted": true
        },
        {
          "id": 1252573,
          "postDate": "2021-03-25T19:32:01.220Z",
          "content": "<p>Actually you do have to.. because False &amp; True = False :-)</p>",
          "rawMarkdown": "Actually you do have to.. because False & True = False :-)"
        },
        {
          "id": 1252596,
          "postDate": "2021-03-25T20:17:11.807Z",
          "rawMarkdown": "",
          "votes": 1,
          "isDeleted": true
        }
      ]
    },
    {
      "id": 1240742,
      "postDate": "2021-03-16T15:49:03.633Z",
      "content": "<ul>\n<li>which model u are using ?</li>\n<li>u trained from scratch ?</li>\n<li>make any pre-processing ?</li>\n<li>try diferents thresholds to iou and conf</li>\n<li>i think that large imgsize will not always help (large -&gt; more information -&gt; overfitting), mainly on larger models, (i am new in DL please correct me if i am wrong)</li>\n<li>try diferents hyperparameters in hyp.scratch.yaml file (default hyperparameters don't work for me)</li>\n</ul>",
      "rawMarkdown": "- which model u are using ?\n- u trained from scratch ?\n- make any pre-processing ?\n- try diferents thresholds to iou and conf\n- i think that large imgsize will not always help (large -> more information -> overfitting), mainly on larger models, (i am new in DL please correct me if i am wrong)\n- try diferents hyperparameters in hyp.scratch.yaml file (default hyperparameters don't work for me)",
      "votes": 2,
      "replies": [
        {
          "id": 1240754,
          "postDate": "2021-03-16T15:52:49.373Z",
          "content": "<ol>\n<li>yolov5x model</li>\n<li>maybe i trained from scratch</li>\n<li>no preprocessing</li>\n<li>i am planning to apply your hyperparameters</li>\n</ol>",
          "rawMarkdown": "1. yolov5x model\n2. maybe i trained from scratch\n3. no preprocessing\n4. i am planning to apply your hyperparameters",
          "votes": 1
        },
        {
          "id": 1240769,
          "postDate": "2021-03-16T16:01:11.697Z",
          "content": "<p>I remember that train from scratch don't work for me (after 50 epochs the model is overfitted) and get a lb score around of 0.18 - 0.2 (with 2 class filter)</p>",
          "rawMarkdown": "I remember that train from scratch don't work for me (after 50 epochs the model is overfitted) and get a lb score around of 0.18 - 0.2 (with 2 class filter)",
          "votes": 1
        },
        {
          "id": 1240796,
          "postDate": "2021-03-16T16:16:14.617Z",
          "content": "<p>Namespace(adam=False, batch_size=16, bucket='', cache_images=False, cfg='', data='vinbigdata.yaml', device='', entity=None, epochs=120, evolve=False, exist_ok=False, fold=0, global_rank=-1, hyp='data/hyp.scratch.yaml', image_weights=False, img_size=[640, 640], linear_lr=False, local_rank=-1, log_artifacts=False, log_imgs=16, multi_scale=False, name='exp', noautoanchor=False, nosave=False, notest=False, project='runs/train', quad=False, rect=False, resume=False, save_dir='runs/train/exp3', single_cls=False, sync_bn=False, total_batch_size=16, weights='yolov5x.pt', workers=8, world_size=1)<br>\nwandb: Install Weights &amp; Biases for YOLOv5 logging with 'pip install wandb' (recommended)</p>\n<p>am i training from scratch?</p>\n<p>i thought that i was using coco pretrained yolov5x  <br>\nfar from scratch, do i need to set up some options to <strong>finetune</strong>?</p>",
          "rawMarkdown": "Namespace(adam=False, batch_size=16, bucket='', cache_images=False, cfg='', data='vinbigdata.yaml', device='', entity=None, epochs=120, evolve=False, exist_ok=False, fold=0, global_rank=-1, hyp='data/hyp.scratch.yaml', image_weights=False, img_size=[640, 640], linear_lr=False, local_rank=-1, log_artifacts=False, log_imgs=16, multi_scale=False, name='exp', noautoanchor=False, nosave=False, notest=False, project='runs/train', quad=False, rect=False, resume=False, save_dir='runs/train/exp3', single_cls=False, sync_bn=False, total_batch_size=16, weights='yolov5x.pt', workers=8, world_size=1)\nwandb: Install Weights & Biases for YOLOv5 logging with 'pip install wandb' (recommended)\n\nam i training from scratch?\n\ni thought that i was using coco pretrained yolov5x  \nfar from scratch, do i need to set up some options to **finetune**?",
          "votes": 1
        },
        {
          "id": 1240805,
          "postDate": "2021-03-16T16:23:42.550Z",
          "content": "<p>it's ok, u are not training from scratch, you're right.</p>",
          "rawMarkdown": "it's ok, u are not training from scratch, you're right.",
          "votes": 1
        },
        {
          "id": 1240816,
          "postDate": "2021-03-16T16:32:46.347Z",
          "content": "<p>I really appreciate your kind guidance. Thank you so much  <br>\nI am investigating what the problem is<br>\nYou had reported your model's all folds exceeded CV 0.4<br>\nDoes that CV mean  yolo MAP@0.5 or CUSTOM CV(after detection, compare validation predictions with answers)<br>\nIf you use CUSTOM CV, Do you use WBF or NMS for CV CHECK?</p>",
          "rawMarkdown": "I really appreciate your kind guidance. Thank you so much  \nI am investigating what the problem is\nYou had reported your model's all folds exceeded CV 0.4\nDoes that CV mean  yolo MAP@0.5 or CUSTOM CV(after detection, compare validation predictions with answers)\nIf you use CUSTOM CV, Do you use WBF or NMS for CV CHECK?",
          "votes": 1
        },
        {
          "id": 1240832,
          "postDate": "2021-03-16T16:44:13.307Z",
          "content": "<p>Mean yolo MAP@0.5</p>",
          "rawMarkdown": "Mean yolo MAP@0.5",
          "votes": 1
        },
        {
          "id": 1240838,
          "postDate": "2021-03-16T16:49:40.367Z",
          "content": "<p>My MAP@0.5 fluctuates near 0.35<br>\nI'm gonna try different settings. Thank you very much. </p>",
          "rawMarkdown": "My MAP@0.5 fluctuates near 0.35\nI'm gonna try different settings. Thank you very much. "
        },
        {
          "id": 1240842,
          "postDate": "2021-03-16T16:53:01.473Z",
          "content": "<p>I see that u don't cache images, how long does your training take?<br>\nYou’re welcome !</p>\n<p>Before i forgot enabled the \"multi-label\" in general.py file (def non_max_suppression(prediction, conf_thres=0.25, iou_thres=0.45, classes=None, agnostic=False, multi_label=True) line 393</p>",
          "rawMarkdown": "I see that u don't cache images, how long does your training take?\nYou’re welcome !\n\nBefore i forgot enabled the \"multi-label\" in general.py file (def non_max_suppression(prediction, conf_thres=0.25, iou_thres=0.45, classes=None, agnostic=False, multi_label=True) line 393",
          "votes": 1
        },
        {
          "id": 1240860,
          "postDate": "2021-03-16T17:03:33.297Z",
          "content": "<p>I use google colab. Training takes around 4 hours with 640 * 640 image size and 16 batch size</p>",
          "rawMarkdown": "I use google colab. Training takes around 4 hours with 640 * 640 image size and 16 batch size"
        },
        {
          "id": 1240882,
          "postDate": "2021-03-16T17:31:53.877Z",
          "content": "<p>multi-label flag is a killing point<br>\nIn training phase, multi-label flag is turned on<br>\nbut In detect phase, multi-label flag is turned off in Yolov5 newest version.<br>\nThank you very much</p>",
          "rawMarkdown": "multi-label flag is a killing point\nIn training phase, multi-label flag is turned on\nbut In detect phase, multi-label flag is turned off in Yolov5 newest version.\nThank you very much",
          "votes": 1
        },
        {
          "id": 1240916,
          "postDate": "2021-03-16T18:23:17.540Z",
          "rawMarkdown": "",
          "isDeleted": true
        },
        {
          "id": 1242009,
          "postDate": "2021-03-17T10:44:31.780Z",
          "rawMarkdown": "",
          "isDeleted": true
        },
        {
          "id": 1242092,
          "postDate": "2021-03-17T11:52:41.623Z",
          "content": "<p>i think that is the only way </p>",
          "rawMarkdown": "i think that is the only way ",
          "votes": 1
        },
        {
          "id": 1242371,
          "postDate": "2021-03-17T15:15:41.487Z",
          "content": "<p>My current training on YOLOv5 fluctuates a lot. Some of the single model are 0.17+, but many below 0.12. Couldn't figure out the reason. 🤔</p>\n<p>I am also interested to see a comparison between 604 vs 1024. Plan to do a test on it. </p>",
          "rawMarkdown": "My current training on YOLOv5 fluctuates a lot. Some of the single model are 0.17+, but many below 0.12. Couldn't figure out the reason. 🤔\n\nI am also interested to see a comparison between 604 vs 1024. Plan to do a test on it. "
        },
        {
          "id": 1242393,
          "postDate": "2021-03-17T15:29:44.563Z",
          "content": "<p>open general.py in the yolov5 repository and you will find that multi_label flag is set False</p>\n<p>Before i forgot enabled the \"multi-label\" in general.py file (def non_max_suppression(prediction, conf_thres=0.25, iou_thres=0.45, classes=None, agnostic=False, multi_label=True) line 393</p>",
          "rawMarkdown": "open general.py in the yolov5 repository and you will find that multi_label flag is set False\n\nBefore i forgot enabled the \"multi-label\" in general.py file (def non_max_suppression(prediction, conf_thres=0.25, iou_thres=0.45, classes=None, agnostic=False, multi_label=True) line 393"
        },
        {
          "id": 1242396,
          "postDate": "2021-03-17T15:31:32.073Z",
          "content": "<p>1024 image size looks like better than 640 size</p>",
          "rawMarkdown": "1024 image size looks like better than 640 size"
        },
        {
          "id": 1246180,
          "postDate": "2021-03-20T14:02:02.493Z",
          "content": "<p>Yeah, I didn’t mention that the multi-label had been already enabled before my fluctuating observations.</p>",
          "rawMarkdown": "Yeah, I didn’t mention that the multi-label had been already enabled before my fluctuating observations.",
          "votes": 1
        },
        {
          "id": 1252538,
          "postDate": "2021-03-25T19:00:19.567Z",
          "rawMarkdown": "",
          "isDeleted": true
        }
      ]
    },
    {
      "id": 1245938,
      "postDate": "2021-03-20T11:12:47.067Z",
      "content": "<p>My ensemble prediction is worse than a single model. Are you using --augment and IOU nsm of 0.4?</p>",
      "rawMarkdown": "My ensemble prediction is worse than a single model. Are you using --augment and IOU nsm of 0.4?",
      "replies": [
        {
          "id": 1246278,
          "postDate": "2021-03-20T15:52:31.660Z",
          "content": "<p>I set IOU and NMS Threshold 0.4 and I don't use --augment</p>",
          "rawMarkdown": "I set IOU and NMS Threshold 0.4 and I don't use --augment\n",
          "votes": 1
        }
      ]
    },
    {
      "id": 1245001,
      "postDate": "2021-03-19T11:44:45.827Z",
      "content": "<p>My yolov5 size 640 is better than 768……</p>",
      "rawMarkdown": "My yolov5 size 640 is better than 768......",
      "replies": [
        {
          "id": 1245693,
          "postDate": "2021-03-20T03:34:54.797Z",
          "content": "<p>I am only training with 1024 size. i will try 640 and notify the result</p>",
          "rawMarkdown": "I am only training with 1024 size. i will try 640 and notify the result"
        },
        {
          "id": 1245725,
          "postDate": "2021-03-20T04:49:29.830Z",
          "content": "<p>Where did you trained?I try it on kaggle but failed.</p>",
          "rawMarkdown": "Where did you trained?I try it on kaggle but failed.",
          "votes": 1
        },
        {
          "id": 1246276,
          "postDate": "2021-03-20T15:51:31.893Z",
          "content": "<p>On google colab. when i train 1024 image size, i usually set batch size=4</p>",
          "rawMarkdown": "On google colab. when i train 1024 image size, i usually set batch size=4\n"
        },
        {
          "id": 1246675,
          "postDate": "2021-03-21T01:54:59.790Z",
          "content": "<p>Thank you,I will have a try.</p>",
          "rawMarkdown": "Thank you,I will have a try."
        }
      ]
    },
    {
      "id": 1242173,
      "postDate": "2021-03-17T13:01:58.667Z",
      "content": "<p>Hi,did u train it on your machine?</p>",
      "rawMarkdown": "Hi,did u train it on your machine?",
      "replies": [
        {
          "id": 1242397,
          "postDate": "2021-03-17T15:31:38.450Z",
          "content": "<p>No. i use google colab</p>",
          "rawMarkdown": "No. i use google colab"
        }
      ]
    }
  ],
  "comments": [
    {
      "id": 1243357,
      "author_name": "nOpe",
      "author_url": "",
      "post_date": "2021-03-18T06:49:27.920000",
      "content": "<p>For yolov5, I suggest u should start with hyp.scratch.yaml file, set <code>mixup</code> to 0. With yolov5x, I see  it converge before 50 epochs. I can get baseline score above 0.15 in with a fold, use 2 class filter to greatly improve your LB score </p>",
      "votes": 3,
      "replies": [
        {
          "id": 1243694,
          "author_name": "kaggler",
          "author_url": "",
          "post_date": "2021-03-18T12:25:04.463000",
          "content": "<p>ok. I am gonna try 2 classifier. thanks a lot! good luck!</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1250393,
          "author_name": "Joshua Schaaf",
          "author_url": "",
          "post_date": "2021-03-24T02:04:27.290000",
          "content": "<p>What do you mean by 2 class filter? I'm pretty new to this…</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1251798,
          "author_name": "David",
          "author_url": "",
          "post_date": "2021-03-25T06:26:01.423000",
          "content": "<p>2 class filter is used to filter the no-finding images on testset</p>",
          "votes": 1,
          "replies": []
        }
      ]
    },
    {
      "id": 1241249,
      "author_name": "Tian",
      "author_url": "",
      "post_date": "2021-03-17T00:56:06.463000",
      "content": "<p>hmm, same problem.</p>",
      "votes": 1,
      "replies": [
        {
          "id": 1241326,
          "author_name": "kaggler",
          "author_url": "",
          "post_date": "2021-03-17T02:05:17.110000",
          "content": "<p>My single fold improved from 0.111 to 0.162<br>\n<a href=\"https://www.kaggle.com/adrielcabral\" target=\"_blank\">@adrielcabral</a> told me  \"multi-label\" in general.py file (def non_max_suppression(prediction, conf_thres=0.25, iou_thres=0.45, classes=None, agnostic=False, multi_label=True) line 393</p>\n<p>After setting multi-label = True, my score improved</p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 1245746,
          "author_name": "Tian",
          "author_url": "",
          "post_date": "2021-03-20T05:30:24.767000",
          "content": "<p>I tried, still very low lb</p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 1252527,
          "author_name": "",
          "author_url": "",
          "post_date": "2021-03-25T18:44:40.420000",
          "content": "",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1252573,
          "author_name": "Hannes Öhler",
          "author_url": "",
          "post_date": "2021-03-25T19:32:01.220000",
          "content": "<p>Actually you do have to.. because False &amp; True = False :-)</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1252596,
          "author_name": "",
          "author_url": "",
          "post_date": "2021-03-25T20:17:11.807000",
          "content": "",
          "votes": 1,
          "replies": []
        }
      ]
    },
    {
      "id": 1240742,
      "author_name": "adriel cabral",
      "author_url": "",
      "post_date": "2021-03-16T15:49:03.633000",
      "content": "<ul>\n<li>which model u are using ?</li>\n<li>u trained from scratch ?</li>\n<li>make any pre-processing ?</li>\n<li>try diferents thresholds to iou and conf</li>\n<li>i think that large imgsize will not always help (large -&gt; more information -&gt; overfitting), mainly on larger models, (i am new in DL please correct me if i am wrong)</li>\n<li>try diferents hyperparameters in hyp.scratch.yaml file (default hyperparameters don't work for me)</li>\n</ul>",
      "votes": 2,
      "replies": [
        {
          "id": 1240754,
          "author_name": "kaggler",
          "author_url": "",
          "post_date": "2021-03-16T15:52:49.373000",
          "content": "<ol>\n<li>yolov5x model</li>\n<li>maybe i trained from scratch</li>\n<li>no preprocessing</li>\n<li>i am planning to apply your hyperparameters</li>\n</ol>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 1240769,
          "author_name": "adriel cabral",
          "author_url": "",
          "post_date": "2021-03-16T16:01:11.697000",
          "content": "<p>I remember that train from scratch don't work for me (after 50 epochs the model is overfitted) and get a lb score around of 0.18 - 0.2 (with 2 class filter)</p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 1240796,
          "author_name": "kaggler",
          "author_url": "",
          "post_date": "2021-03-16T16:16:14.617000",
          "content": "<p>Namespace(adam=False, batch_size=16, bucket='', cache_images=False, cfg='', data='vinbigdata.yaml', device='', entity=None, epochs=120, evolve=False, exist_ok=False, fold=0, global_rank=-1, hyp='data/hyp.scratch.yaml', image_weights=False, img_size=[640, 640], linear_lr=False, local_rank=-1, log_artifacts=False, log_imgs=16, multi_scale=False, name='exp', noautoanchor=False, nosave=False, notest=False, project='runs/train', quad=False, rect=False, resume=False, save_dir='runs/train/exp3', single_cls=False, sync_bn=False, total_batch_size=16, weights='yolov5x.pt', workers=8, world_size=1)<br>\nwandb: Install Weights &amp; Biases for YOLOv5 logging with 'pip install wandb' (recommended)</p>\n<p>am i training from scratch?</p>\n<p>i thought that i was using coco pretrained yolov5x  <br>\nfar from scratch, do i need to set up some options to <strong>finetune</strong>?</p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 1240805,
          "author_name": "adriel cabral",
          "author_url": "",
          "post_date": "2021-03-16T16:23:42.550000",
          "content": "<p>it's ok, u are not training from scratch, you're right.</p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 1240816,
          "author_name": "kaggler",
          "author_url": "",
          "post_date": "2021-03-16T16:32:46.347000",
          "content": "<p>I really appreciate your kind guidance. Thank you so much  <br>\nI am investigating what the problem is<br>\nYou had reported your model's all folds exceeded CV 0.4<br>\nDoes that CV mean  yolo MAP@0.5 or CUSTOM CV(after detection, compare validation predictions with answers)<br>\nIf you use CUSTOM CV, Do you use WBF or NMS for CV CHECK?</p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 1240832,
          "author_name": "adriel cabral",
          "author_url": "",
          "post_date": "2021-03-16T16:44:13.307000",
          "content": "<p>Mean yolo MAP@0.5</p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 1240838,
          "author_name": "kaggler",
          "author_url": "",
          "post_date": "2021-03-16T16:49:40.367000",
          "content": "<p>My MAP@0.5 fluctuates near 0.35<br>\nI'm gonna try different settings. Thank you very much. </p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1240842,
          "author_name": "adriel cabral",
          "author_url": "",
          "post_date": "2021-03-16T16:53:01.473000",
          "content": "<p>I see that u don't cache images, how long does your training take?<br>\nYou’re welcome !</p>\n<p>Before i forgot enabled the \"multi-label\" in general.py file (def non_max_suppression(prediction, conf_thres=0.25, iou_thres=0.45, classes=None, agnostic=False, multi_label=True) line 393</p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 1240860,
          "author_name": "kaggler",
          "author_url": "",
          "post_date": "2021-03-16T17:03:33.297000",
          "content": "<p>I use google colab. Training takes around 4 hours with 640 * 640 image size and 16 batch size</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1240882,
          "author_name": "kaggler",
          "author_url": "",
          "post_date": "2021-03-16T17:31:53.877000",
          "content": "<p>multi-label flag is a killing point<br>\nIn training phase, multi-label flag is turned on<br>\nbut In detect phase, multi-label flag is turned off in Yolov5 newest version.<br>\nThank you very much</p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 1240916,
          "author_name": "",
          "author_url": "",
          "post_date": "2021-03-16T18:23:17.540000",
          "content": "",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1242009,
          "author_name": "",
          "author_url": "",
          "post_date": "2021-03-17T10:44:31.780000",
          "content": "",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1242092,
          "author_name": "adriel cabral",
          "author_url": "",
          "post_date": "2021-03-17T11:52:41.623000",
          "content": "<p>i think that is the only way </p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 1242371,
          "author_name": "Kuan Zhang",
          "author_url": "",
          "post_date": "2021-03-17T15:15:41.487000",
          "content": "<p>My current training on YOLOv5 fluctuates a lot. Some of the single model are 0.17+, but many below 0.12. Couldn't figure out the reason. 🤔</p>\n<p>I am also interested to see a comparison between 604 vs 1024. Plan to do a test on it. </p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1242393,
          "author_name": "kaggler",
          "author_url": "",
          "post_date": "2021-03-17T15:29:44.563000",
          "content": "<p>open general.py in the yolov5 repository and you will find that multi_label flag is set False</p>\n<p>Before i forgot enabled the \"multi-label\" in general.py file (def non_max_suppression(prediction, conf_thres=0.25, iou_thres=0.45, classes=None, agnostic=False, multi_label=True) line 393</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1242396,
          "author_name": "kaggler",
          "author_url": "",
          "post_date": "2021-03-17T15:31:32.073000",
          "content": "<p>1024 image size looks like better than 640 size</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1246180,
          "author_name": "Kuan Zhang",
          "author_url": "",
          "post_date": "2021-03-20T14:02:02.493000",
          "content": "<p>Yeah, I didn’t mention that the multi-label had been already enabled before my fluctuating observations.</p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 1252538,
          "author_name": "",
          "author_url": "",
          "post_date": "2021-03-25T19:00:19.567000",
          "content": "",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 1245938,
      "author_name": "Phat Tran",
      "author_url": "",
      "post_date": "2021-03-20T11:12:47.067000",
      "content": "<p>My ensemble prediction is worse than a single model. Are you using --augment and IOU nsm of 0.4?</p>",
      "votes": 0,
      "replies": [
        {
          "id": 1246278,
          "author_name": "kaggler",
          "author_url": "",
          "post_date": "2021-03-20T15:52:31.660000",
          "content": "<p>I set IOU and NMS Threshold 0.4 and I don't use --augment</p>",
          "votes": 1,
          "replies": []
        }
      ]
    },
    {
      "id": 1245001,
      "author_name": "Zekun",
      "author_url": "",
      "post_date": "2021-03-19T11:44:45.827000",
      "content": "<p>My yolov5 size 640 is better than 768……</p>",
      "votes": 0,
      "replies": [
        {
          "id": 1245693,
          "author_name": "kaggler",
          "author_url": "",
          "post_date": "2021-03-20T03:34:54.797000",
          "content": "<p>I am only training with 1024 size. i will try 640 and notify the result</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1245725,
          "author_name": "Zekun",
          "author_url": "",
          "post_date": "2021-03-20T04:49:29.830000",
          "content": "<p>Where did you trained?I try it on kaggle but failed.</p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 1246276,
          "author_name": "kaggler",
          "author_url": "",
          "post_date": "2021-03-20T15:51:31.893000",
          "content": "<p>On google colab. when i train 1024 image size, i usually set batch size=4</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1246675,
          "author_name": "Zekun",
          "author_url": "",
          "post_date": "2021-03-21T01:54:59.790000",
          "content": "<p>Thank you,I will have a try.</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 1242173,
      "author_name": "Zekun",
      "author_url": "",
      "post_date": "2021-03-17T13:01:58.667000",
      "content": "<p>Hi,did u train it on your machine?</p>",
      "votes": 0,
      "replies": [
        {
          "id": 1242397,
          "author_name": "kaggler",
          "author_url": "",
          "post_date": "2021-03-17T15:31:38.450000",
          "content": "<p>No. i use google colab</p>",
          "votes": 0,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1240719": "Thank you for reading this\nI am the newbie of object-detection \nI have been studying object detection for school homework\n\nMy yolov5 scores too low\n\n* 14classes(training with just abnormal)\n* Image size : 1024 x 1024 resized\n* 100 epochs.\n* Stratified with iterative-stratification\n* No 2 class filter\n\n* VALIDATION MAP@.5 scores 0.35, \n* submission -> yolo detect.py with conf 0.05, iou 0.4\n\n**But, my submission just scores around 0.11**\njust 512 image size baseline scores above 0.15\nwhat can be a problem?\n\nThank you for reading this. \n\n\n",
    "1243357": "For yolov5, I suggest u should start with hyp.scratch.yaml file, set `mixup` to 0. With yolov5x, I see  it converge before 50 epochs. I can get baseline score above 0.15 in with a fold, use 2 class filter to greatly improve your LB score ",
    "1241249": "hmm, same problem.",
    "1240742": "- which model u are using ?\n- u trained from scratch ?\n- make any pre-processing ?\n- try diferents thresholds to iou and conf\n- i think that large imgsize will not always help (large -> more information -> overfitting), mainly on larger models, (i am new in DL please correct me if i am wrong)\n- try diferents hyperparameters in hyp.scratch.yaml file (default hyperparameters don't work for me)",
    "1245938": "My ensemble prediction is worse than a single model. Are you using --augment and IOU nsm of 0.4?",
    "1245001": "My yolov5 size 640 is better than 768......",
    "1242173": "Hi,did u train it on your machine?"
  }
}