{
  "id": 166703,
  "title": "Using \"noisy-student\" is more powerful than \"imagenet\"",
  "url": "/competitions/siim-isic-melanoma-classification/discussion/166703",
  "author_name": "",
  "post_date": "2020-07-13T18:52:31.334366200Z",
  "votes": 15,
  "comment_count": 24,
  "views": 0,
  "content": "<p>Actually,  First I had used pre-trained weight from imagenet which give 93.8% accuracy but I change a little bit and use only noisy-student pre-trained weight got 94.1% accuracy.</p>\n\n<p><a href=\"https://www.kaggle.com/shivam17818/test-all-efficientnet-model-b0-b7\">Notebook Link</a></p>",
  "messages": [
    {
      "id": "928103",
      "postDate": "07/13/2020 18:52:31",
      "content": "<p>Actually,  First I had used pre-trained weight from imagenet which give 93.8% accuracy but I change a little bit and use only noisy-student pre-trained weight got 94.1% accuracy.</p>\n\n<p><a href=\"https://www.kaggle.com/shivam17818/test-all-efficientnet-model-b0-b7\">Notebook Link</a></p>",
      "rawMarkdown": "Actually,  First I had used pre-trained weight from imagenet which give 93.8% accuracy but I change a little bit and use only noisy-student pre-trained weight got 94.1% accuracy.\n\n[Notebook Link](https://www.kaggle.com/shivam17818/test-all-efficientnet-model-b0-b7)",
      "votes": null
    },
    {
      "id": "928132",
      "postDate": "07/13/2020 19:06:36",
      "content": "<p>Yeah. You can see the IMAGENET LB. Noisy-student is better</p>",
      "rawMarkdown": "Yeah. You can see the IMAGENET LB. Noisy-student is better",
      "votes": null
    },
    {
      "id": "928137",
      "postDate": "07/13/2020 19:09:40",
      "content": "<p>yup </p>",
      "rawMarkdown": "yup",
      "votes": null
    },
    {
      "id": "928142",
      "postDate": "07/13/2020 19:11:43",
      "content": "<p>you can add\n<code>\nRandomCropSize(img)\n</code>\nThis model will be better like FixEffNet is the best model now for IMAGENET</p>",
      "rawMarkdown": "you can add\n```\nRandomCropSize(img)\n```\nThis model will be better like FixEffNet is the best model now for IMAGENET",
      "votes": null
    },
    {
      "id": "928150",
      "postDate": "07/13/2020 19:16:42",
      "content": "<p>thanks for your  suggestion☺️</p>",
      "rawMarkdown": "thanks for your  suggestion☺️",
      "votes": null
    },
    {
      "id": "928152",
      "postDate": "07/13/2020 19:17:42",
      "content": "<p>Is there a pytorch version of these weithts?</p>",
      "rawMarkdown": "Is there a pytorch version of these weithts?",
      "votes": null
    },
    {
      "id": "928177",
      "postDate": "07/13/2020 19:34:20",
      "content": "<p>You guys can easily convert the files yourselves. There is a script in this very repo:</p>\n\n<p><a href=\"https://github.com/lukemelas/EfficientNet-PyTorch/tree/master/tf_to_pytorch\">https://github.com/lukemelas/EfficientNet-PyTorch/tree/master/tf_to_pytorch</a></p>\n\n<p>Use it to download tf weights and convert to .pth. For example, I navigated to <a href=\"https://github.com/tensorflow/tpu/tree/master/models/official/efficientnet\">https://github.com/tensorflow/tpu/tree/master/models/official/efficientnet</a>, and downloaded NoisyStudent + RA EfficientNet's B2 tensorflow .ckpt. Untar'd the file into the pretrained_tensorflow directory.</p>\n\n<p>Then ran this:</p>\n\n<p>python load_tf_weights.py --model_name efficientnet-b2 --tf_checkpoint ../pretrained_tensorflow/noisy-student-efficientnet-b2/ --output_file ../pretrained_pytorch/noisy-student-efficientnet-b2.pth which resulted in a .pth file ready to rock and roll.</p>\n\n<p>If you have errors with load_tf_weights.py due to it or .any of the other scripts complaining about being written in tf1.x, here's a copy of the code I've altered such that it runs just fine with tensorflow-2.2.0: <a href=\"https://drive.google.com/file/d/11jDyfRKoIhVpuNsFSeR_SAFe0e0R0fA1/view?usp=sharing\">https://drive.google.com/file/d/11jDyfRKoIhVpuNsFSeR_SAFe0e0R0fA1/view?usp=sharing</a></p>",
      "rawMarkdown": "You guys can easily convert the files yourselves. There is a script in this very repo:\n\nhttps://github.com/lukemelas/EfficientNet-PyTorch/tree/master/tf_to_pytorch\n\nUse it to download tf weights and convert to .pth. For example, I navigated to https://github.com/tensorflow/tpu/tree/master/models/official/efficientnet, and downloaded NoisyStudent + RA EfficientNet's B2 tensorflow .ckpt. Untar'd the file into the pretrained_tensorflow directory.\n\nThen ran this:\n\npython load_tf_weights.py --model_name efficientnet-b2 --tf_checkpoint ../pretrained_tensorflow/noisy-student-efficientnet-b2/ --output_file ../pretrained_pytorch/noisy-student-efficientnet-b2.pth which resulted in a .pth file ready to rock and roll.\n\nIf you have errors with load_tf_weights.py due to it or .any of the other scripts complaining about being written in tf1.x, here's a copy of the code I've altered such that it runs just fine with tensorflow-2.2.0: https://drive.google.com/file/d/11jDyfRKoIhVpuNsFSeR_SAFe0e0R0fA1/view?usp=sharing",
      "votes": null
    },
    {
      "id": "928182",
      "postDate": "07/13/2020 19:35:32",
      "content": "<p>Thank you my friend! I also found: <a href=\"https://github.com/rwightman/gen-efficientnet-pytorch\">https://github.com/rwightman/gen-efficientnet-pytorch</a>\nThe weights are already converted!</p>",
      "rawMarkdown": "Thank you my friend! I also found: https://github.com/rwightman/gen-efficientnet-pytorch\nThe weights are already converted!",
      "votes": null
    },
    {
      "id": "928208",
      "postDate": "07/13/2020 19:46:28",
      "content": "<p>FixEfficientNet  pre-trained weights for pytorch---<a href=\"https://github.com/facebookresearch/FixRes\">https://github.com/facebookresearch/FixRes</a></p>",
      "rawMarkdown": "FixEfficientNet  pre-trained weights for pytorch---https://github.com/facebookresearch/FixRes",
      "votes": null
    },
    {
      "id": "928215",
      "postDate": "07/13/2020 19:49:23",
      "content": "<p>Nice. I love PyTorch but in this competition TF look better. </p>",
      "rawMarkdown": "Nice. I love PyTorch but in this competition TF look better.",
      "votes": null
    },
    {
      "id": "928221",
      "postDate": "07/13/2020 19:51:30",
      "content": "<p><a href=\"/shivam17818\">@shivam17818</a> Cool! Thanks</p>",
      "rawMarkdown": "shivam17818 Cool! Thanks",
      "votes": null
    },
    {
      "id": "928233",
      "postDate": "07/13/2020 19:57:21",
      "content": "<p>you're welcome!</p>",
      "rawMarkdown": "you're welcome!",
      "votes": null
    },
    {
      "id": "928377",
      "postDate": "07/13/2020 23:56:53",
      "content": "<p>don't trust one experiment. I found (and I am astonished) that the imagenet weights work better in this competition.\nDon't know why...</p>",
      "rawMarkdown": "don't trust one experiment. I found (and I am astonished) that the imagenet weights work better in this competition.\nDon't know why...",
      "votes": null
    },
    {
      "id": "928545",
      "postDate": "07/14/2020 04:35:40",
      "content": "<p>Gonna try this tomorrow, thanks!</p>",
      "rawMarkdown": "Gonna try this tomorrow, thanks!",
      "votes": null
    },
    {
      "id": "928612",
      "postDate": "07/14/2020 05:19:53",
      "content": "<p>Offcourse, you can be right but in my all experiment noisy-student is better.\nYou can suggest, in which way imagenet is better, please!😊</p>",
      "rawMarkdown": "Offcourse, you can be right but in my all experiment noisy-student is better.\nYou can suggest, in which way imagenet is better, please!😊",
      "votes": null
    },
    {
      "id": "928850",
      "postDate": "07/14/2020 09:06:01",
      "content": "<p>In all of my experiments (10 noisy trys) imagenet performed clearly better :)\nmaybe I should recheck this...</p>",
      "rawMarkdown": "In all of my experiments (10 noisy trys) imagenet performed clearly better :)\nmaybe I should recheck this...",
      "votes": null
    },
    {
      "id": "929221",
      "postDate": "07/14/2020 14:39:22",
      "content": "<p>same situation</p>",
      "rawMarkdown": "same situation",
      "votes": null
    },
    {
      "id": "929731",
      "postDate": "07/14/2020 22:40:42",
      "content": "<p>same situation for me. But, I found that the correlation between imagenet and noisy-student is relatively low (0.4 ~ 0.6) and the ensemble with both of them looks better.</p>",
      "rawMarkdown": "same situation for me. But, I found that the correlation between imagenet and noisy-student is relatively low (0.4 ~ 0.6) and the ensemble with both of them looks better.",
      "votes": null
    },
    {
      "id": "930675",
      "postDate": "07/15/2020 16:33:50",
      "content": "<p>Where add that code?\nImageDataGenerator??</p>",
      "rawMarkdown": "Where add that code?\nImageDataGenerator??",
      "votes": null
    },
    {
      "id": "930917",
      "postDate": "07/15/2020 20:03:30",
      "content": "<pre><code>    img = transform(img,DIM=dim)\n    img = tf.image.random_flip_left_right(img)\n    img = tf.image.random_saturation(img, 0.7, 1.3)\n\n    img = transform(img,DIM=256)\n    img = tf.image.random_flip_left_right(img)\n    img = tf.image.random_saturation(img, 0.7, 1.3)\n    img = tf.image.random_contrast(img, 0.8, 1.2)\n    img = tf.image.random_brightness(img, 0.1)\n    **img = tf.image.random_crop(img, size=[256, 256, 3])**\n</code></pre>",
      "rawMarkdown": "img = transform(img,DIM=dim)\n        img = tf.image.random_flip_left_right(img)\n        img = tf.image.random_saturation(img, 0.7, 1.3)\n      \n        img = transform(img,DIM=256)\n        img = tf.image.random_flip_left_right(img)\n        img = tf.image.random_saturation(img, 0.7, 1.3)\n        img = tf.image.random_contrast(img, 0.8, 1.2)\n        img = tf.image.random_brightness(img, 0.1)\n        **img = tf.image.random_crop(img, size=[256, 256, 3])**",
      "votes": null
    },
    {
      "id": "931700",
      "postDate": "07/16/2020 11:41:10",
      "content": "<p>I too have tried. IN some tests it seems that with noisy-student the NN continue to learn after 10 epochs (with imagenet rarely it happens to me). But at the end, I have not seen any real improvement.</p>",
      "rawMarkdown": "I too have tried. IN some tests it seems that with noisy-student the NN continue to learn after 10 epochs (with imagenet rarely it happens to me). But at the end, I have not seen any real improvement.",
      "votes": null
    },
    {
      "id": "933063",
      "postDate": "07/17/2020 13:04:35",
      "content": "<p>What your CV score <a href=\"/shivam17818\">@shivam17818</a> ?</p>",
      "rawMarkdown": "What your CV score @shivam17818 ?",
      "votes": null
    },
    {
      "id": "933436",
      "postDate": "07/17/2020 17:42:53",
      "content": "<p>In my public notebook, my CV was 93.8 but I recently run a kernel, in which my CV is ~95.6 but LB is approx ~95.</p>",
      "rawMarkdown": "In my public notebook, my CV was 93.8 but I recently run a kernel, in which my CV is ~95.6 but LB is approx ~95.",
      "votes": null
    },
    {
      "id": "933894",
      "postDate": "07/18/2020 05:28:36",
      "content": "<p>I have checked carefully and see that noisy-student is a bit worse than imagenet for B2.</p>",
      "rawMarkdown": "I have checked carefully and see that noisy-student is a bit worse than imagenet for B2.",
      "votes": null
    },
    {
      "id": "933964",
      "postDate": "07/18/2020 07:01:57",
      "content": "<p>I also found that in some cases,noise-student gives worse accuracy than imagenet.</p>",
      "rawMarkdown": "I also found that in some cases,noise-student gives worse accuracy than imagenet.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 928132,
      "author_name": "doanquanvietnamca",
      "author_url": "",
      "post_date": "07/13/2020 19:06:36",
      "content": "<p>Yeah. You can see the IMAGENET LB. Noisy-student is better</p>",
      "votes": null,
      "replies": [
        {
          "id": 928137,
          "author_name": "shivam17818",
          "author_url": "",
          "post_date": "07/13/2020 19:09:40",
          "content": "<p>yup </p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 928142,
          "author_name": "doanquanvietnamca",
          "author_url": "",
          "post_date": "07/13/2020 19:11:43",
          "content": "<p>you can add\n<code>\nRandomCropSize(img)\n</code>\nThis model will be better like FixEffNet is the best model now for IMAGENET</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 928150,
          "author_name": "shivam17818",
          "author_url": "",
          "post_date": "07/13/2020 19:16:42",
          "content": "<p>thanks for your  suggestion☺️</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 930675,
          "author_name": "zxzxs9182",
          "author_url": "",
          "post_date": "07/15/2020 16:33:50",
          "content": "<p>Where add that code?\nImageDataGenerator??</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 930917,
          "author_name": "shivam17818",
          "author_url": "",
          "post_date": "07/15/2020 20:03:30",
          "content": "<pre><code>    img = transform(img,DIM=dim)\n    img = tf.image.random_flip_left_right(img)\n    img = tf.image.random_saturation(img, 0.7, 1.3)\n\n    img = transform(img,DIM=256)\n    img = tf.image.random_flip_left_right(img)\n    img = tf.image.random_saturation(img, 0.7, 1.3)\n    img = tf.image.random_contrast(img, 0.8, 1.2)\n    img = tf.image.random_brightness(img, 0.1)\n    **img = tf.image.random_crop(img, size=[256, 256, 3])**\n</code></pre>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 928152,
      "author_name": "yannmajewski",
      "author_url": "",
      "post_date": "07/13/2020 19:17:42",
      "content": "<p>Is there a pytorch version of these weithts?</p>",
      "votes": null,
      "replies": [
        {
          "id": 928177,
          "author_name": "doanquanvietnamca",
          "author_url": "",
          "post_date": "07/13/2020 19:34:20",
          "content": "<p>You guys can easily convert the files yourselves. There is a script in this very repo:</p>\n\n<p><a href=\"https://github.com/lukemelas/EfficientNet-PyTorch/tree/master/tf_to_pytorch\">https://github.com/lukemelas/EfficientNet-PyTorch/tree/master/tf_to_pytorch</a></p>\n\n<p>Use it to download tf weights and convert to .pth. For example, I navigated to <a href=\"https://github.com/tensorflow/tpu/tree/master/models/official/efficientnet\">https://github.com/tensorflow/tpu/tree/master/models/official/efficientnet</a>, and downloaded NoisyStudent + RA EfficientNet's B2 tensorflow .ckpt. Untar'd the file into the pretrained_tensorflow directory.</p>\n\n<p>Then ran this:</p>\n\n<p>python load_tf_weights.py --model_name efficientnet-b2 --tf_checkpoint ../pretrained_tensorflow/noisy-student-efficientnet-b2/ --output_file ../pretrained_pytorch/noisy-student-efficientnet-b2.pth which resulted in a .pth file ready to rock and roll.</p>\n\n<p>If you have errors with load_tf_weights.py due to it or .any of the other scripts complaining about being written in tf1.x, here's a copy of the code I've altered such that it runs just fine with tensorflow-2.2.0: <a href=\"https://drive.google.com/file/d/11jDyfRKoIhVpuNsFSeR_SAFe0e0R0fA1/view?usp=sharing\">https://drive.google.com/file/d/11jDyfRKoIhVpuNsFSeR_SAFe0e0R0fA1/view?usp=sharing</a></p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 928182,
          "author_name": "yannmajewski",
          "author_url": "",
          "post_date": "07/13/2020 19:35:32",
          "content": "<p>Thank you my friend! I also found: <a href=\"https://github.com/rwightman/gen-efficientnet-pytorch\">https://github.com/rwightman/gen-efficientnet-pytorch</a>\nThe weights are already converted!</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 928208,
          "author_name": "shivam17818",
          "author_url": "",
          "post_date": "07/13/2020 19:46:28",
          "content": "<p>FixEfficientNet  pre-trained weights for pytorch---<a href=\"https://github.com/facebookresearch/FixRes\">https://github.com/facebookresearch/FixRes</a></p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 928215,
          "author_name": "doanquanvietnamca",
          "author_url": "",
          "post_date": "07/13/2020 19:49:23",
          "content": "<p>Nice. I love PyTorch but in this competition TF look better. </p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 928221,
          "author_name": "yannmajewski",
          "author_url": "",
          "post_date": "07/13/2020 19:51:30",
          "content": "<p><a href=\"/shivam17818\">@shivam17818</a> Cool! Thanks</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 928233,
          "author_name": "shivam17818",
          "author_url": "",
          "post_date": "07/13/2020 19:57:21",
          "content": "<p>you're welcome!</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 928377,
      "author_name": "romanweilguny",
      "author_url": "",
      "post_date": "07/13/2020 23:56:53",
      "content": "<p>don't trust one experiment. I found (and I am astonished) that the imagenet weights work better in this competition.\nDon't know why...</p>",
      "votes": null,
      "replies": [
        {
          "id": 928612,
          "author_name": "shivam17818",
          "author_url": "",
          "post_date": "07/14/2020 05:19:53",
          "content": "<p>Offcourse, you can be right but in my all experiment noisy-student is better.\nYou can suggest, in which way imagenet is better, please!😊</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 928850,
          "author_name": "romanweilguny",
          "author_url": "",
          "post_date": "07/14/2020 09:06:01",
          "content": "<p>In all of my experiments (10 noisy trys) imagenet performed clearly better :)\nmaybe I should recheck this...</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 929221,
          "author_name": "ragnar123",
          "author_url": "",
          "post_date": "07/14/2020 14:39:22",
          "content": "<p>same situation</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 929731,
          "author_name": "syumei",
          "author_url": "",
          "post_date": "07/14/2020 22:40:42",
          "content": "<p>same situation for me. But, I found that the correlation between imagenet and noisy-student is relatively low (0.4 ~ 0.6) and the ensemble with both of them looks better.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 931700,
          "author_name": "luigisaetta",
          "author_url": "",
          "post_date": "07/16/2020 11:41:10",
          "content": "<p>I too have tried. IN some tests it seems that with noisy-student the NN continue to learn after 10 epochs (with imagenet rarely it happens to me). But at the end, I have not seen any real improvement.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 928545,
      "author_name": "santiviquez",
      "author_url": "",
      "post_date": "07/14/2020 04:35:40",
      "content": "<p>Gonna try this tomorrow, thanks!</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 933063,
      "author_name": "kurianbenoy",
      "author_url": "",
      "post_date": "07/17/2020 13:04:35",
      "content": "<p>What your CV score <a href=\"/shivam17818\">@shivam17818</a> ?</p>",
      "votes": null,
      "replies": [
        {
          "id": 933436,
          "author_name": "shivam17818",
          "author_url": "",
          "post_date": "07/17/2020 17:42:53",
          "content": "<p>In my public notebook, my CV was 93.8 but I recently run a kernel, in which my CV is ~95.6 but LB is approx ~95.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 933894,
      "author_name": "hainamnguyen",
      "author_url": "",
      "post_date": "07/18/2020 05:28:36",
      "content": "<p>I have checked carefully and see that noisy-student is a bit worse than imagenet for B2.</p>",
      "votes": null,
      "replies": [
        {
          "id": 933964,
          "author_name": "shivam17818",
          "author_url": "",
          "post_date": "07/18/2020 07:01:57",
          "content": "<p>I also found that in some cases,noise-student gives worse accuracy than imagenet.</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "928103": "Actually,  First I had used pre-trained weight from imagenet which give 93.8% accuracy but I change a little bit and use only noisy-student pre-trained weight got 94.1% accuracy.\n\n[Notebook Link](https://www.kaggle.com/shivam17818/test-all-efficientnet-model-b0-b7)",
    "928132": "Yeah. You can see the IMAGENET LB. Noisy-student is better",
    "928137": "yup",
    "928142": "you can add\n```\nRandomCropSize(img)\n```\nThis model will be better like FixEffNet is the best model now for IMAGENET",
    "928150": "thanks for your  suggestion☺️",
    "928152": "Is there a pytorch version of these weithts?",
    "928177": "You guys can easily convert the files yourselves. There is a script in this very repo:\n\nhttps://github.com/lukemelas/EfficientNet-PyTorch/tree/master/tf_to_pytorch\n\nUse it to download tf weights and convert to .pth. For example, I navigated to https://github.com/tensorflow/tpu/tree/master/models/official/efficientnet, and downloaded NoisyStudent + RA EfficientNet's B2 tensorflow .ckpt. Untar'd the file into the pretrained_tensorflow directory.\n\nThen ran this:\n\npython load_tf_weights.py --model_name efficientnet-b2 --tf_checkpoint ../pretrained_tensorflow/noisy-student-efficientnet-b2/ --output_file ../pretrained_pytorch/noisy-student-efficientnet-b2.pth which resulted in a .pth file ready to rock and roll.\n\nIf you have errors with load_tf_weights.py due to it or .any of the other scripts complaining about being written in tf1.x, here's a copy of the code I've altered such that it runs just fine with tensorflow-2.2.0: https://drive.google.com/file/d/11jDyfRKoIhVpuNsFSeR_SAFe0e0R0fA1/view?usp=sharing",
    "928182": "Thank you my friend! I also found: https://github.com/rwightman/gen-efficientnet-pytorch\nThe weights are already converted!",
    "928208": "FixEfficientNet  pre-trained weights for pytorch---https://github.com/facebookresearch/FixRes",
    "928215": "Nice. I love PyTorch but in this competition TF look better.",
    "928221": "shivam17818 Cool! Thanks",
    "928233": "you're welcome!",
    "928377": "don't trust one experiment. I found (and I am astonished) that the imagenet weights work better in this competition.\nDon't know why...",
    "928545": "Gonna try this tomorrow, thanks!",
    "928612": "Offcourse, you can be right but in my all experiment noisy-student is better.\nYou can suggest, in which way imagenet is better, please!😊",
    "928850": "In all of my experiments (10 noisy trys) imagenet performed clearly better :)\nmaybe I should recheck this...",
    "929221": "same situation",
    "929731": "same situation for me. But, I found that the correlation between imagenet and noisy-student is relatively low (0.4 ~ 0.6) and the ensemble with both of them looks better.",
    "930675": "Where add that code?\nImageDataGenerator??",
    "930917": "img = transform(img,DIM=dim)\n        img = tf.image.random_flip_left_right(img)\n        img = tf.image.random_saturation(img, 0.7, 1.3)\n      \n        img = transform(img,DIM=256)\n        img = tf.image.random_flip_left_right(img)\n        img = tf.image.random_saturation(img, 0.7, 1.3)\n        img = tf.image.random_contrast(img, 0.8, 1.2)\n        img = tf.image.random_brightness(img, 0.1)\n        **img = tf.image.random_crop(img, size=[256, 256, 3])**",
    "931700": "I too have tried. IN some tests it seems that with noisy-student the NN continue to learn after 10 epochs (with imagenet rarely it happens to me). But at the end, I have not seen any real improvement.",
    "933063": "What your CV score @shivam17818 ?",
    "933436": "In my public notebook, my CV was 93.8 but I recently run a kernel, in which my CV is ~95.6 but LB is approx ~95.",
    "933894": "I have checked carefully and see that noisy-student is a bit worse than imagenet for B2.",
    "933964": "I also found that in some cases,noise-student gives worse accuracy than imagenet."
  },
  "source": "meta"
}