{
  "id": 211920,
  "title": "[How to] FineTuning model",
  "url": "/competitions/cassava-leaf-disease-classification/discussion/211920",
  "author_name": "",
  "post_date": "2021-01-16T19:57:45.993074200Z",
  "votes": 46,
  "comment_count": 20,
  "views": 0,
  "content": "<h1>Introduction</h1>\n<p>For noisy labels, you can use finetuning other models to freezing layers.</p>\n<p><img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F3492127%2F7610c90265e4cab8665f870141060061%2Ffinetuning.png?generation=1610826648213026&amp;alt=media\" alt=\"\"></p>\n<h1>Add freeze &amp; unfreeze funcs to model</h1>\n<pre><code>    def freeze(self):\n        # To freeze the residual layers\n        for param in self.model.parameters():\n            param.requires_grad = False\n\n        for param in self.model.fc.parameters():\n            param.requires_grad = True\n\n    def unfreeze(self):\n        # Unfreeze all layers\n        for param in self.model.parameters():\n            param.requires_grad = True\n</code></pre>\n<h1>Try more</h1>\n<ul>\n<li>Use other models(e.g. ViT, DeiT, etc.)</li>\n<li>Use different Loss functions for Finetuning models</li>\n<li>Use different image size(e.g. orignal - 384x384, finetuning - 512x512)</li>\n<li>Train <code>Learning with Noisy Labels</code> and Finetune Original Labels</li>\n<li>Train Original Labels and <code>Learning with Noisy Labels</code></li>\n<li>etc.</li>\n</ul>\n<h1>Updated</h1>\n<ul>\n<li><p><code>V4</code> - Initial version</p></li>\n<li><p><code>V5</code> - Add <code>xm.set_rng_state(CFG.seed, device)</code> for reproducing results.(This requires more experimentation.)</p></li>\n<li><p><code>V6</code> - Add <code>map_location=torch.device('cpu')</code> to <code>load_state_dict</code> for TPU</p></li>\n<li><p><code>V7</code> - Test <code>xm.set_rng_state(CFG.seed, device)</code> and get same result for 1 epoch using same seed</p></li>\n</ul>\n<h1>End</h1>\n<p>Notebook is <a href=\"https://www.kaggle.com/piantic/how-to-finetuning-models-pytorch-xla-tpu?scriptVersionId=52023531\" target=\"_blank\">here</a>.<br>\n<a href=\"https://www.kaggle.com/piantic/how-to-finetuning-models-pytorch-xla-tpu?scriptVersionId=52023531\" target=\"_blank\">https://www.kaggle.com/piantic/how-to-finetuning-models-pytorch-xla-tpu?scriptVersionId=52023531</a></p>\n<p>Thank you!</p>",
  "messages": [
    {
      "id": "1155956",
      "postDate": "01/16/2021 19:57:45",
      "content": "<h1>Introduction</h1>\n<p>For noisy labels, you can use finetuning other models to freezing layers.</p>\n<p><img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F3492127%2F7610c90265e4cab8665f870141060061%2Ffinetuning.png?generation=1610826648213026&amp;alt=media\" alt=\"\"></p>\n<h1>Add freeze &amp; unfreeze funcs to model</h1>\n<pre><code>    def freeze(self):\n        # To freeze the residual layers\n        for param in self.model.parameters():\n            param.requires_grad = False\n\n        for param in self.model.fc.parameters():\n            param.requires_grad = True\n\n    def unfreeze(self):\n        # Unfreeze all layers\n        for param in self.model.parameters():\n            param.requires_grad = True\n</code></pre>\n<h1>Try more</h1>\n<ul>\n<li>Use other models(e.g. ViT, DeiT, etc.)</li>\n<li>Use different Loss functions for Finetuning models</li>\n<li>Use different image size(e.g. orignal - 384x384, finetuning - 512x512)</li>\n<li>Train <code>Learning with Noisy Labels</code> and Finetune Original Labels</li>\n<li>Train Original Labels and <code>Learning with Noisy Labels</code></li>\n<li>etc.</li>\n</ul>\n<h1>Updated</h1>\n<ul>\n<li><p><code>V4</code> - Initial version</p></li>\n<li><p><code>V5</code> - Add <code>xm.set_rng_state(CFG.seed, device)</code> for reproducing results.(This requires more experimentation.)</p></li>\n<li><p><code>V6</code> - Add <code>map_location=torch.device('cpu')</code> to <code>load_state_dict</code> for TPU</p></li>\n<li><p><code>V7</code> - Test <code>xm.set_rng_state(CFG.seed, device)</code> and get same result for 1 epoch using same seed</p></li>\n</ul>\n<h1>End</h1>\n<p>Notebook is <a href=\"https://www.kaggle.com/piantic/how-to-finetuning-models-pytorch-xla-tpu?scriptVersionId=52023531\" target=\"_blank\">here</a>.<br>\n<a href=\"https://www.kaggle.com/piantic/how-to-finetuning-models-pytorch-xla-tpu?scriptVersionId=52023531\" target=\"_blank\">https://www.kaggle.com/piantic/how-to-finetuning-models-pytorch-xla-tpu?scriptVersionId=52023531</a></p>\n<p>Thank you!</p>",
      "rawMarkdown": "# Introduction\nFor noisy labels, you can use finetuning other models to freezing layers.\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F3492127%2F7610c90265e4cab8665f870141060061%2Ffinetuning.png?generation=1610826648213026&alt=media)\n\n# Add freeze & unfreeze funcs to model\n```\n    def freeze(self):\n        # To freeze the residual layers\n        for param in self.model.parameters():\n            param.requires_grad = False\n\n        for param in self.model.fc.parameters():\n            param.requires_grad = True\n    \n    def unfreeze(self):\n        # Unfreeze all layers\n        for param in self.model.parameters():\n            param.requires_grad = True\n```\n\n# Try more\n - Use other models(e.g. ViT, DeiT, etc.)\n - Use different Loss functions for Finetuning models\n - Use different image size(e.g. orignal - 384x384, finetuning - 512x512)\n - Train `Learning with Noisy Labels` and Finetune Original Labels\n - Train Original Labels and `Learning with Noisy Labels`\n - etc.\n\n# Updated\n- `V4` - Initial version\n\n- `V5` - Add `xm.set_rng_state(CFG.seed, device)` for reproducing results.(This requires more experimentation.)\n\n- `V6` - Add `map_location=torch.device('cpu')` to `load_state_dict` for TPU\n\n- `V7` - Test `xm.set_rng_state(CFG.seed, device)` and get same result for 1 epoch using same seed\n\n# End\nNotebook is [here](https://www.kaggle.com/piantic/how-to-finetuning-models-pytorch-xla-tpu?scriptVersionId=52023531).\nhttps://www.kaggle.com/piantic/how-to-finetuning-models-pytorch-xla-tpu?scriptVersionId=52023531\n\nThank you!",
      "votes": null
    },
    {
      "id": "1156196",
      "postDate": "01/17/2021 02:22:27",
      "content": "<p>I've tried freezing layers in TF, but it didn't really help. Model couldn't converge well even after unfreezing. Any thoughts?</p>",
      "rawMarkdown": "I've tried freezing layers in TF, but it didn't really help. Model couldn't converge well even after unfreezing. Any thoughts?",
      "votes": null
    },
    {
      "id": "1156291",
      "postDate": "01/17/2021 04:24:54",
      "content": "<p>Thank you a lot for sharing.</p>",
      "rawMarkdown": "Thank you a lot for sharing.",
      "votes": null
    },
    {
      "id": "1157378",
      "postDate": "01/17/2021 21:04:15",
      "content": "<p>Thanks for sharing! The freezing and unfreezing the model layers helped me with convergence! Do you know any articles explaining why does finetuning improve model performance? Thanks!</p>",
      "rawMarkdown": "Thanks for sharing! The freezing and unfreezing the model layers helped me with convergence! Do you know any articles explaining why does finetuning improve model performance? Thanks!",
      "votes": null
    },
    {
      "id": "1157906",
      "postDate": "01/18/2021 08:41:30",
      "content": "<p>There are many articles and papers. So, It's hard to pick one for me.<br>\nI think that the articles related to <code>Transfer learning</code> and <code>finetuning</code> are helpful.</p>\n<p><a href=\"https://towardsdatascience.com/cnn-transfer-learning-fine-tuning-9f3e7c5806b2\" target=\"_blank\">https://towardsdatascience.com/cnn-transfer-learning-fine-tuning-9f3e7c5806b2</a></p>\n<p><a href=\"https://www.kaggle.com/capiru\" target=\"_blank\">@capiru</a> </p>",
      "rawMarkdown": "There are many articles and papers. So, It's hard to pick one for me.\nI think that the articles related to `Transfer learning` and `finetuning` are helpful.\n\nhttps://towardsdatascience.com/cnn-transfer-learning-fine-tuning-9f3e7c5806b2\n\n@capiru",
      "votes": null
    },
    {
      "id": "1159676",
      "postDate": "01/19/2021 11:47:46",
      "content": "<p>Thanks for sharing!</p>",
      "rawMarkdown": "Thanks for sharing!",
      "votes": null
    },
    {
      "id": "1161712",
      "postDate": "01/20/2021 17:48:55",
      "content": "<p>Thanks for sharing!</p>",
      "rawMarkdown": "Thanks for sharing!",
      "votes": null
    },
    {
      "id": "1163583",
      "postDate": "01/21/2021 19:20:17",
      "content": "<p>Hello <a href=\"https://www.kaggle.com/piantic\" target=\"_blank\">@piantic</a>,<br>\nI have a question. For finetuning is usually used same <strong>LR</strong> as for previous normal training process or should I reduce or increase it?<br>\nbtw: thank you for your sharing 👍</p>",
      "rawMarkdown": "Hello @piantic,\nI have a question. For finetuning is usually used same **LR** as for previous normal training process or should I reduce or increase it?\nbtw: thank you for your sharing 👍",
      "votes": null
    },
    {
      "id": "1163590",
      "postDate": "01/21/2021 19:27:31",
      "content": "<p>Usually set to a lower value than the initial lr or last train lr.</p>\n<p><a href=\"https://www.kaggle.com/filchy\" target=\"_blank\">@filchy</a> </p>",
      "rawMarkdown": "Usually set to a lower value than the initial lr or last train lr.\n\n@filchy",
      "votes": null
    },
    {
      "id": "1163592",
      "postDate": "01/21/2021 19:28:41",
      "content": "<p>Thank you very much.<br>\n<a href=\"https://www.kaggle.com/piantic\" target=\"_blank\">@piantic</a> </p>",
      "rawMarkdown": "Thank you very much.\n@piantic",
      "votes": null
    },
    {
      "id": "1164040",
      "postDate": "01/22/2021 05:17:22",
      "content": "<p>Thanks for sharing <a href=\"https://www.kaggle.com/piantic\" target=\"_blank\">@piantic</a> <br>\nyes, I think you are right. I have also tried freezing layers. I trained firstly the layers that I added to the top of the pre-trained model and after that, I tried several options by refreezing layers from 100, then 50. unfortunately, it didn't help. any suggestion?</p>",
      "rawMarkdown": "Thanks for sharing @piantic \nyes, I think you are right. I have also tried freezing layers. I trained firstly the layers that I added to the top of the pre-trained model and after that, I tried several options by refreezing layers from 100, then 50. unfortunately, it didn't help. any suggestion?",
      "votes": null
    },
    {
      "id": "1164426",
      "postDate": "01/22/2021 11:15:05",
      "content": "<p>thanks for sharing!</p>",
      "rawMarkdown": "thanks for sharing!",
      "votes": null
    },
    {
      "id": "1171596",
      "postDate": "01/27/2021 01:29:11",
      "content": "<p>Thank you this will definitely help me!</p>",
      "rawMarkdown": "Thank you this will definitely help me!",
      "votes": null
    },
    {
      "id": "1179181",
      "postDate": "01/31/2021 11:35:43",
      "content": "<p>WOW! IT IS REALLY INTRESTING<br>\nUPVOTED!<br>\n<a href=\"https://www.kaggle.com/ashokkumarbibbab/covid-19-vaccination-process\" target=\"_blank\">https://www.kaggle.com/ashokkumarbibbab/covid-19-vaccination-process</a><br>\nSEE THIS NOTEBOOK🙄</p>",
      "rawMarkdown": "WOW! IT IS REALLY INTRESTING\nUPVOTED!\nhttps://www.kaggle.com/ashokkumarbibbab/covid-19-vaccination-process\nSEE THIS NOTEBOOK🙄",
      "votes": null
    },
    {
      "id": "1179678",
      "postDate": "01/31/2021 18:03:21",
      "content": "<p>this is really nice </p>",
      "rawMarkdown": "this is really nice",
      "votes": null
    },
    {
      "id": "1179736",
      "postDate": "01/31/2021 19:12:41",
      "content": "<p>nice stuff here</p>",
      "rawMarkdown": "nice stuff here",
      "votes": null
    },
    {
      "id": "1193235",
      "postDate": "02/09/2021 14:29:37",
      "content": "<p>This is really helpful, Thanks for sharing. Increasing image size worked for me.</p>",
      "rawMarkdown": "This is really helpful, Thanks for sharing. Increasing image size worked for me.",
      "votes": null
    },
    {
      "id": "1199971",
      "postDate": "02/14/2021 09:55:11",
      "content": "<p>Hi <a href=\"https://www.kaggle.com/piantic\" target=\"_blank\">@piantic</a>,<br>\nDo you think that finetuning models is good idea for such a noisy test set as we have in this competition?</p>",
      "rawMarkdown": "Hi @piantic,\nDo you think that finetuning models is good idea for such a noisy test set as we have in this competition?",
      "votes": null
    },
    {
      "id": "1200244",
      "postDate": "02/14/2021 14:11:00",
      "content": "<p>Maybe, the 1st place team knows that. :)</p>",
      "rawMarkdown": "Maybe, the 1st place team knows that. :)",
      "votes": null
    },
    {
      "id": "1211103",
      "postDate": "02/20/2021 00:26:04",
      "content": "<p>This is really helpful, Thanks for sharing. </p>",
      "rawMarkdown": "This is really helpful, Thanks for sharing.",
      "votes": null
    },
    {
      "id": "1211179",
      "postDate": "02/20/2021 03:15:10",
      "content": "<p>Thanks for sharing. Finetuning is also retraining the entire model based on the pretrained weights? Or finetuning is just the re-training of the Fully Connected Layers?</p>",
      "rawMarkdown": "Thanks for sharing. Finetuning is also retraining the entire model based on the pretrained weights? Or finetuning is just the re-training of the Fully Connected Layers?",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1156196,
      "author_name": "junyingsg",
      "author_url": "",
      "post_date": "01/17/2021 02:22:27",
      "content": "<p>I've tried freezing layers in TF, but it didn't really help. Model couldn't converge well even after unfreezing. Any thoughts?</p>",
      "votes": null,
      "replies": [
        {
          "id": 1164040,
          "author_name": "kutaykutlu",
          "author_url": "",
          "post_date": "01/22/2021 05:17:22",
          "content": "<p>Thanks for sharing <a href=\"https://www.kaggle.com/piantic\" target=\"_blank\">@piantic</a> <br>\nyes, I think you are right. I have also tried freezing layers. I trained firstly the layers that I added to the top of the pre-trained model and after that, I tried several options by refreezing layers from 100, then 50. unfortunately, it didn't help. any suggestion?</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1156291,
      "author_name": "durbin164",
      "author_url": "",
      "post_date": "01/17/2021 04:24:54",
      "content": "<p>Thank you a lot for sharing.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1157378,
      "author_name": "capiru",
      "author_url": "",
      "post_date": "01/17/2021 21:04:15",
      "content": "<p>Thanks for sharing! The freezing and unfreezing the model layers helped me with convergence! Do you know any articles explaining why does finetuning improve model performance? Thanks!</p>",
      "votes": null,
      "replies": [
        {
          "id": 1157906,
          "author_name": "piantic",
          "author_url": "",
          "post_date": "01/18/2021 08:41:30",
          "content": "<p>There are many articles and papers. So, It's hard to pick one for me.<br>\nI think that the articles related to <code>Transfer learning</code> and <code>finetuning</code> are helpful.</p>\n<p><a href=\"https://towardsdatascience.com/cnn-transfer-learning-fine-tuning-9f3e7c5806b2\" target=\"_blank\">https://towardsdatascience.com/cnn-transfer-learning-fine-tuning-9f3e7c5806b2</a></p>\n<p><a href=\"https://www.kaggle.com/capiru\" target=\"_blank\">@capiru</a> </p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1159676,
      "author_name": "khushi14",
      "author_url": "",
      "post_date": "01/19/2021 11:47:46",
      "content": "<p>Thanks for sharing!</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1161712,
      "author_name": "rmtrickster",
      "author_url": "",
      "post_date": "01/20/2021 17:48:55",
      "content": "<p>Thanks for sharing!</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1163583,
      "author_name": "filchy",
      "author_url": "",
      "post_date": "01/21/2021 19:20:17",
      "content": "<p>Hello <a href=\"https://www.kaggle.com/piantic\" target=\"_blank\">@piantic</a>,<br>\nI have a question. For finetuning is usually used same <strong>LR</strong> as for previous normal training process or should I reduce or increase it?<br>\nbtw: thank you for your sharing 👍</p>",
      "votes": null,
      "replies": [
        {
          "id": 1163590,
          "author_name": "piantic",
          "author_url": "",
          "post_date": "01/21/2021 19:27:31",
          "content": "<p>Usually set to a lower value than the initial lr or last train lr.</p>\n<p><a href=\"https://www.kaggle.com/filchy\" target=\"_blank\">@filchy</a> </p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1163592,
          "author_name": "filchy",
          "author_url": "",
          "post_date": "01/21/2021 19:28:41",
          "content": "<p>Thank you very much.<br>\n<a href=\"https://www.kaggle.com/piantic\" target=\"_blank\">@piantic</a> </p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1164426,
      "author_name": "freyablue",
      "author_url": "",
      "post_date": "01/22/2021 11:15:05",
      "content": "<p>thanks for sharing!</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1171596,
      "author_name": "meckanical",
      "author_url": "",
      "post_date": "01/27/2021 01:29:11",
      "content": "<p>Thank you this will definitely help me!</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1179181,
      "author_name": "ashokkumarbibbab",
      "author_url": "",
      "post_date": "01/31/2021 11:35:43",
      "content": "<p>WOW! IT IS REALLY INTRESTING<br>\nUPVOTED!<br>\n<a href=\"https://www.kaggle.com/ashokkumarbibbab/covid-19-vaccination-process\" target=\"_blank\">https://www.kaggle.com/ashokkumarbibbab/covid-19-vaccination-process</a><br>\nSEE THIS NOTEBOOK🙄</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1179678,
      "author_name": "anuravgupta",
      "author_url": "",
      "post_date": "01/31/2021 18:03:21",
      "content": "<p>this is really nice </p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1179736,
      "author_name": "abhiest",
      "author_url": "",
      "post_date": "01/31/2021 19:12:41",
      "content": "<p>nice stuff here</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1193235,
      "author_name": "shanmukh05",
      "author_url": "",
      "post_date": "02/09/2021 14:29:37",
      "content": "<p>This is really helpful, Thanks for sharing. Increasing image size worked for me.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1199971,
      "author_name": "filchy",
      "author_url": "",
      "post_date": "02/14/2021 09:55:11",
      "content": "<p>Hi <a href=\"https://www.kaggle.com/piantic\" target=\"_blank\">@piantic</a>,<br>\nDo you think that finetuning models is good idea for such a noisy test set as we have in this competition?</p>",
      "votes": null,
      "replies": [
        {
          "id": 1200244,
          "author_name": "piantic",
          "author_url": "",
          "post_date": "02/14/2021 14:11:00",
          "content": "<p>Maybe, the 1st place team knows that. :)</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1211103,
      "author_name": "berkayaytac",
      "author_url": "",
      "post_date": "02/20/2021 00:26:04",
      "content": "<p>This is really helpful, Thanks for sharing. </p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1211179,
      "author_name": "antoreepjana",
      "author_url": "",
      "post_date": "02/20/2021 03:15:10",
      "content": "<p>Thanks for sharing. Finetuning is also retraining the entire model based on the pretrained weights? Or finetuning is just the re-training of the Fully Connected Layers?</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1155956": "# Introduction\nFor noisy labels, you can use finetuning other models to freezing layers.\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F3492127%2F7610c90265e4cab8665f870141060061%2Ffinetuning.png?generation=1610826648213026&alt=media)\n\n# Add freeze & unfreeze funcs to model\n```\n    def freeze(self):\n        # To freeze the residual layers\n        for param in self.model.parameters():\n            param.requires_grad = False\n\n        for param in self.model.fc.parameters():\n            param.requires_grad = True\n    \n    def unfreeze(self):\n        # Unfreeze all layers\n        for param in self.model.parameters():\n            param.requires_grad = True\n```\n\n# Try more\n - Use other models(e.g. ViT, DeiT, etc.)\n - Use different Loss functions for Finetuning models\n - Use different image size(e.g. orignal - 384x384, finetuning - 512x512)\n - Train `Learning with Noisy Labels` and Finetune Original Labels\n - Train Original Labels and `Learning with Noisy Labels`\n - etc.\n\n# Updated\n- `V4` - Initial version\n\n- `V5` - Add `xm.set_rng_state(CFG.seed, device)` for reproducing results.(This requires more experimentation.)\n\n- `V6` - Add `map_location=torch.device('cpu')` to `load_state_dict` for TPU\n\n- `V7` - Test `xm.set_rng_state(CFG.seed, device)` and get same result for 1 epoch using same seed\n\n# End\nNotebook is [here](https://www.kaggle.com/piantic/how-to-finetuning-models-pytorch-xla-tpu?scriptVersionId=52023531).\nhttps://www.kaggle.com/piantic/how-to-finetuning-models-pytorch-xla-tpu?scriptVersionId=52023531\n\nThank you!",
    "1156196": "I've tried freezing layers in TF, but it didn't really help. Model couldn't converge well even after unfreezing. Any thoughts?",
    "1156291": "Thank you a lot for sharing.",
    "1157378": "Thanks for sharing! The freezing and unfreezing the model layers helped me with convergence! Do you know any articles explaining why does finetuning improve model performance? Thanks!",
    "1157906": "There are many articles and papers. So, It's hard to pick one for me.\nI think that the articles related to `Transfer learning` and `finetuning` are helpful.\n\nhttps://towardsdatascience.com/cnn-transfer-learning-fine-tuning-9f3e7c5806b2\n\n@capiru",
    "1159676": "Thanks for sharing!",
    "1161712": "Thanks for sharing!",
    "1163583": "Hello @piantic,\nI have a question. For finetuning is usually used same **LR** as for previous normal training process or should I reduce or increase it?\nbtw: thank you for your sharing 👍",
    "1163590": "Usually set to a lower value than the initial lr or last train lr.\n\n@filchy",
    "1163592": "Thank you very much.\n@piantic",
    "1164040": "Thanks for sharing @piantic \nyes, I think you are right. I have also tried freezing layers. I trained firstly the layers that I added to the top of the pre-trained model and after that, I tried several options by refreezing layers from 100, then 50. unfortunately, it didn't help. any suggestion?",
    "1164426": "thanks for sharing!",
    "1171596": "Thank you this will definitely help me!",
    "1179181": "WOW! IT IS REALLY INTRESTING\nUPVOTED!\nhttps://www.kaggle.com/ashokkumarbibbab/covid-19-vaccination-process\nSEE THIS NOTEBOOK🙄",
    "1179678": "this is really nice",
    "1179736": "nice stuff here",
    "1193235": "This is really helpful, Thanks for sharing. Increasing image size worked for me.",
    "1199971": "Hi @piantic,\nDo you think that finetuning models is good idea for such a noisy test set as we have in this competition?",
    "1200244": "Maybe, the 1st place team knows that. :)",
    "1211103": "This is really helpful, Thanks for sharing.",
    "1211179": "Thanks for sharing. Finetuning is also retraining the entire model based on the pretrained weights? Or finetuning is just the re-training of the Fully Connected Layers?"
  },
  "source": "meta"
}