{
  "id": 130276,
  "title": "Official External Data Thread",
  "url": "/competitions/flower-classification-with-tpus/discussion/130276",
  "author_name": "Julia Elliott",
  "post_date": "2020-02-13T07:24:00.463000",
  "votes": 4,
  "comment_count": 26,
  "views": 0,
  "content": "<p>Per the <a href=\"https://www.kaggle.com/c/flower-classification-with-tpus/rules\">Competition Rules</a>, freely and publicly available external data is permitted in this competition, but must be posted to this forum thread no later than the Entry Deadline (one week before competition close).</p>\n\n<p>Note for this competition you cannot use privately-held external datasets with the TPU integration. Instead, you'll want to upload a public dataset to Kaggle to use and declare those here.</p>\n\n<p>Once someone posts an external dataset to this thread, you do not need to re-post it if you are using the same one.</p>\n\n<p>You only need to declare the original dataset used; you do not need to declare re-labeled or augmented or otherwise processed versions of datasets. Pre-trained models can be declared, as well; however, models resulting from your own original work, that you have trained yourself offline do not need to be shared/declared.</p>",
  "messages": [
    {
      "id": 745340,
      "postDate": "2020-02-13T18:16:57.987Z",
      "content": "<p>And to get the GCS bucket path of a Kaggle dataset, in case your notebook uses two datasets or more, do this:\n<code>\ngcs_path = KaggleDatasets().get_gcs_path('kaggle_dataset_directory_name')\n</code>\nTo get the Kaggle dataset directory name, the easiest way is to attach the dataset to a notebook and then run this in order to list all attached dataset directories:\n<code>\n!ls /kaggle/input\n</code></p>",
      "rawMarkdown": "And to get the GCS bucket path of a Kaggle dataset, in case your notebook uses two datasets or more, do this:\n```\ngcs_path = KaggleDatasets().get_gcs_path('kaggle_dataset_directory_name')\n```\nTo get the Kaggle dataset directory name, the easiest way is to attach the dataset to a notebook and then run this in order to list all attached dataset directories:\n```\n!ls /kaggle/input\n```",
      "votes": 4
    },
    {
      "id": 744827,
      "postDate": "2020-02-13T07:24:00.463Z",
      "content": "<p>Per the <a href=\"https://www.kaggle.com/c/flower-classification-with-tpus/rules\">Competition Rules</a>, freely and publicly available external data is permitted in this competition, but must be posted to this forum thread no later than the Entry Deadline (one week before competition close).</p>\n\n<p>Note for this competition you cannot use privately-held external datasets with the TPU integration. Instead, you'll want to upload a public dataset to Kaggle to use and declare those here.</p>\n\n<p>Once someone posts an external dataset to this thread, you do not need to re-post it if you are using the same one.</p>\n\n<p>You only need to declare the original dataset used; you do not need to declare re-labeled or augmented or otherwise processed versions of datasets. Pre-trained models can be declared, as well; however, models resulting from your own original work, that you have trained yourself offline do not need to be shared/declared.</p>",
      "rawMarkdown": "Per the [Competition Rules](https://www.kaggle.com/c/flower-classification-with-tpus/rules), freely and publicly available external data is permitted in this competition, but must be posted to this forum thread no later than the Entry Deadline (one week before competition close).\n\nNote for this competition you cannot use privately-held external datasets with the TPU integration. Instead, you'll want to upload a public dataset to Kaggle to use and declare those here.\n\nOnce someone posts an external dataset to this thread, you do not need to re-post it if you are using the same one.\n\nYou only need to declare the original dataset used; you do not need to declare re-labeled or augmented or otherwise processed versions of datasets. Pre-trained models can be declared, as well; however, models resulting from your own original work, that you have trained yourself offline do not need to be shared/declared.",
      "votes": 4
    },
    {
      "id": 836352,
      "postDate": "2020-05-06T23:02:05.590Z",
      "content": "<p>Here is a clarification regarding use of external datasets that mirror the same images as the test set. I want to reiterate that the rules prohibit training on test set images: “Submissions may not use or incorporate information from hand labeling or human prediction of the validation dataset or test data records.” So, while you are permitted to use the 5 datasets that we used to source the training/test sets as <a href=\"https://www.kaggle.com/c/flower-classification-with-tpus/discussion/134701#769297\">I stated previously here</a>, <strong>it is prohibited to train your models using any mirrors/overlaps/duplication of the competition's test data within other flower datasets, as this conflicts with the above-stated rule.</strong></p>\n\n<p>All potential winners will be required to hand over their full code, including specifying the exact external datasets they've used. We will be cross-checking training images with those contained in the test set to confirm no overlaps as a condition to awarding prizes. If you've already committed this violation, you should ensure you do not select any submissions making use of prohibited data for your final 2 submissions that will be evaluated on the final Private Leaderboard.</p>",
      "rawMarkdown": "Here is a clarification regarding use of external datasets that mirror the same images as the test set. I want to reiterate that the rules prohibit training on test set images: “Submissions may not use or incorporate information from hand labeling or human prediction of the validation dataset or test data records.” So, while you are permitted to use the 5 datasets that we used to source the training/test sets as [I stated previously here](https://www.kaggle.com/c/flower-classification-with-tpus/discussion/134701#769297), **it is prohibited to train your models using any mirrors/overlaps/duplication of the competition's test data within other flower datasets, as this conflicts with the above-stated rule.**\n\nAll potential winners will be required to hand over their full code, including specifying the exact external datasets they've used. We will be cross-checking training images with those contained in the test set to confirm no overlaps as a condition to awarding prizes. If you've already committed this violation, you should ensure you do not select any submissions making use of prohibited data for your final 2 submissions that will be evaluated on the final Private Leaderboard.",
      "votes": 1,
      "replies": [
        {
          "id": 836391,
          "postDate": "2020-05-07T00:14:52.613Z",
          "content": "<p>Are we allowed to use pseudo labeled external data? (i.e. download external datasets then use a model trained on competition training data to predict the labels on external images?)</p>",
          "rawMarkdown": "Are we allowed to use pseudo labeled external data? (i.e. download external datasets then use a model trained on competition training data to predict the labels on external images?)"
        },
        {
          "id": 836608,
          "postDate": "2020-05-07T05:30:58.330Z",
          "rawMarkdown": "",
          "votes": 1,
          "isDeleted": true
        },
        {
          "id": 837430,
          "postDate": "2020-05-07T19:04:55.063Z",
          "content": "<p>A re-post of <a href=\"https://www.kaggle.com/c/flower-classification-with-tpus/discussion/148329#837427\">my reply to a separate thread</a> which addresses these:</p>\n\n<blockquote>\n  <ul>\n  <li>Pseudolabeling is permitted unless you're pseudolabeling using images from or duplicative of images in the test set.</li>\n  <li>Using publicly-available pre-trained models is allowed, including those with components of the test set underlying the pre-trained models. To be super clear, this means you're not permitted to be \"pre-training\" models using the test set and then calling that \"pre-trained,\" but you can use the host of ImageNet or other canned publicly-available pre-trained models mentioned in preceding posts.</li>\n  <li>The currently published datasets can be used (including the 5 datasets from which the training/test sets were sourced), but images which overlap with or are the same as the test set cannot be used (should be removed). Correct, no new external datasets can be announced.</li>\n  </ul>\n</blockquote>\n\n<p>Lastly, a reiteration that this is a <strong>playground competition</strong>. As with all playgrounds, it was created in the spirit of learning (specifically for the launch of our TPU integration), more than the pursuit of winning. Hence, it is not awarding points/medals, and prizes are non-monetary &amp; centered around TPU use. We launched this with the knowledge there would be vulnerabilities and more grey area than we'd prefer in a featured competition, hence sending it to the playground. I hope sound judgment is used by participants that aligns with this competition's spirit, so that we can continue to offer opportunities like this.</p>",
          "rawMarkdown": "A re-post of [my reply to a separate thread](https://www.kaggle.com/c/flower-classification-with-tpus/discussion/148329#837427) which addresses these:\n\n&gt; - Pseudolabeling is permitted unless you're pseudolabeling using images from or duplicative of images in the test set.\n- Using publicly-available pre-trained models is allowed, including those with components of the test set underlying the pre-trained models. To be super clear, this means you're not permitted to be \"pre-training\" models using the test set and then calling that \"pre-trained,\" but you can use the host of ImageNet or other canned publicly-available pre-trained models mentioned in preceding posts.\n- The currently published datasets can be used (including the 5 datasets from which the training/test sets were sourced), but images which overlap with or are the same as the test set cannot be used (should be removed). Correct, no new external datasets can be announced.\n\nLastly, a reiteration that this is a **playground competition**. As with all playgrounds, it was created in the spirit of learning (specifically for the launch of our TPU integration), more than the pursuit of winning. Hence, it is not awarding points/medals, and prizes are non-monetary &amp; centered around TPU use. We launched this with the knowledge there would be vulnerabilities and more grey area than we'd prefer in a featured competition, hence sending it to the playground. I hope sound judgment is used by participants that aligns with this competition's spirit, so that we can continue to offer opportunities like this.",
          "votes": 1
        },
        {
          "id": 837817,
          "postDate": "2020-05-08T04:14:51.437Z",
          "content": "<p>Hi <a href=\"/juliaelliott\">@juliaelliott</a>，I have removed all test images (using imagehash and image similarity) from all five external datasets and uploaded this corrected data to the dataset mentioned before in this thread\n<a href=\"https://www.kaggle.com/kirillblinov/tf-flower-photo-tfrec\">https://www.kaggle.com/kirillblinov/tf-flower-photo-tfrec</a>\nCould you please confirm may or not we use this corrected data in the competition? </p>",
          "rawMarkdown": "Hi @juliaelliott，I have removed all test images (using imagehash and image similarity) from all five external datasets and uploaded this corrected data to the dataset mentioned before in this thread\nhttps://www.kaggle.com/kirillblinov/tf-flower-photo-tfrec\nCould you please confirm may or not we use this corrected data in the competition? ",
          "votes": 3
        },
        {
          "id": 837844,
          "postDate": "2020-05-08T04:52:43.977Z",
          "content": "<p>Thanks for the update <a href=\"/kirillblinov\">@kirillblinov</a> </p>",
          "rawMarkdown": "Thanks for the update @kirillblinov "
        },
        {
          "id": 838521,
          "postDate": "2020-05-08T16:12:08.747Z",
          "content": "<p>Hi <a href=\"/kirillblinov\">@kirillblinov</a> Thanks for your wonderful dataset and thanks for trying to clean it. I'm double checking your work and unfortunately you've missed many test images. In the following example, Kaggle rotated the external image by 90 degrees before using it. This is image 1126 in your <code>imagenet_no_test</code> images and 797 in the test set.</p>\n\n<p><img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1723677%2Ff6ab8988f3d5c29933bac417429596a2%2Frotate.png?generation=1588954161325113&amp;alt=media\" alt=\"\"></p>\n\n<p>The index 1126 into your images are when reading your files as follows\n    ['../../../Flowers/external2/imagenet_no_test//tfrecords-jpeg-224x224/0-224x224-1785.tfrec',\n '../../../Flowers/external2/imagenet_no_test//tfrecords-jpeg-224x224/5-224x224-1785.tfrec',\n '../../../Flowers/external2/imagenet_no_test//tfrecords-jpeg-224x224/2-224x224-1785.tfrec',\n '../../../Flowers/external2/imagenet_no_test//tfrecords-jpeg-224x224/12-224x224-1785.tfrec',\n '../../../Flowers/external2/imagenet_no_test//tfrecords-jpeg-224x224/13-224x224-1785.tfrec',\n '../../../Flowers/external2/imagenet_no_test//tfrecords-jpeg-224x224/9-224x224-1785.tfrec',\n '../../../Flowers/external2/imagenet_no_test//tfrecords-jpeg-224x224/7-224x224-1785.tfrec',\n '../../../Flowers/external2/imagenet_no_test//tfrecords-jpeg-224x224/14-224x224-1785.tfrec',\n '../../../Flowers/external2/imagenet_no_test//tfrecords-jpeg-224x224/15-224x224-1774.tfrec',\n '../../../Flowers/external2/imagenet_no_test//tfrecords-jpeg-224x224/6-224x224-1785.tfrec',\n '../../../Flowers/external2/imagenet_no_test//tfrecords-jpeg-224x224/11-224x224-1785.tfrec',\n '../../../Flowers/external2/imagenet_no_test//tfrecords-jpeg-224x224/8-224x224-1785.tfrec',\n '../../../Flowers/external2/imagenet_no_test//tfrecords-jpeg-224x224/10-224x224-1785.tfrec',\n '../../../Flowers/external2/imagenet_no_test//tfrecords-jpeg-224x224/1-224x224-1785.tfrec',\n '../../../Flowers/external2/imagenet_no_test//tfrecords-jpeg-224x224/4-224x224-1785.tfrec',\n '../../../Flowers/external2/imagenet_no_test//tfrecords-jpeg-224x224/3-224x224-1785.tfrec']</p>",
          "rawMarkdown": "Hi @kirillblinov Thanks for your wonderful dataset and thanks for trying to clean it. I'm double checking your work and unfortunately you've missed many test images. In the following example, Kaggle rotated the external image by 90 degrees before using it. This is image 1126 in your `imagenet_no_test` images and 797 in the test set.\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1723677%2Ff6ab8988f3d5c29933bac417429596a2%2Frotate.png?generation=1588954161325113&amp;alt=media)\n\nThe index 1126 into your images are when reading your files as follows\n    ['../../../Flowers/external2/imagenet_no_test//tfrecords-jpeg-224x224/0-224x224-1785.tfrec',\n '../../../Flowers/external2/imagenet_no_test//tfrecords-jpeg-224x224/5-224x224-1785.tfrec',\n '../../../Flowers/external2/imagenet_no_test//tfrecords-jpeg-224x224/2-224x224-1785.tfrec',\n '../../../Flowers/external2/imagenet_no_test//tfrecords-jpeg-224x224/12-224x224-1785.tfrec',\n '../../../Flowers/external2/imagenet_no_test//tfrecords-jpeg-224x224/13-224x224-1785.tfrec',\n '../../../Flowers/external2/imagenet_no_test//tfrecords-jpeg-224x224/9-224x224-1785.tfrec',\n '../../../Flowers/external2/imagenet_no_test//tfrecords-jpeg-224x224/7-224x224-1785.tfrec',\n '../../../Flowers/external2/imagenet_no_test//tfrecords-jpeg-224x224/14-224x224-1785.tfrec',\n '../../../Flowers/external2/imagenet_no_test//tfrecords-jpeg-224x224/15-224x224-1774.tfrec',\n '../../../Flowers/external2/imagenet_no_test//tfrecords-jpeg-224x224/6-224x224-1785.tfrec',\n '../../../Flowers/external2/imagenet_no_test//tfrecords-jpeg-224x224/11-224x224-1785.tfrec',\n '../../../Flowers/external2/imagenet_no_test//tfrecords-jpeg-224x224/8-224x224-1785.tfrec',\n '../../../Flowers/external2/imagenet_no_test//tfrecords-jpeg-224x224/10-224x224-1785.tfrec',\n '../../../Flowers/external2/imagenet_no_test//tfrecords-jpeg-224x224/1-224x224-1785.tfrec',\n '../../../Flowers/external2/imagenet_no_test//tfrecords-jpeg-224x224/4-224x224-1785.tfrec',\n '../../../Flowers/external2/imagenet_no_test//tfrecords-jpeg-224x224/3-224x224-1785.tfrec']\n",
          "votes": 1
        },
        {
          "id": 838529,
          "postDate": "2020-05-08T16:18:43.377Z",
          "content": "<p>Here's another one they rotated 5 degrees counter clockwise. (The image below that says \"Train 23525\" is image 23525 in external dataset <code>imagenet_no_test</code>).\n<img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1723677%2F0ed1d2be6ddc5b8466b2d029d47ce37d%2Frotate2.png?generation=1588954720569101&amp;alt=media\" alt=\"\"></p>",
          "rawMarkdown": "Here's another one they rotated 5 degrees counter clockwise. (The image below that says \"Train 23525\" is image 23525 in external dataset `imagenet_no_test`).\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1723677%2F0ed1d2be6ddc5b8466b2d029d47ce37d%2Frotate2.png?generation=1588954720569101&amp;alt=media)\n"
        },
        {
          "id": 838535,
          "postDate": "2020-05-08T16:23:03.737Z",
          "content": "<p>Here's an interesting one. It's a different photo but it's of the same flower just taking from a slightly different angle. I'm not sure if we can include this one or not. (The image below that says \"Train 25322\" is image 25322 in external dataset <code>imagenet_no_test</code>).\n<img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1723677%2F719ed5ff4a9b465532c1cee9b96f2f6c%2Fphoto.png?generation=1588954910061954&amp;alt=media\" alt=\"\"></p>",
          "rawMarkdown": "Here's an interesting one. It's a different photo but it's of the same flower just taking from a slightly different angle. I'm not sure if we can include this one or not. (The image below that says \"Train 25322\" is image 25322 in external dataset `imagenet_no_test`).\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1723677%2F719ed5ff4a9b465532c1cee9b96f2f6c%2Fphoto.png?generation=1588954910061954&amp;alt=media)\n"
        },
        {
          "id": 838565,
          "postDate": "2020-05-08T16:39:28.413Z",
          "content": "<p>In your <code>imagenet_no_test</code>, I also found the following duplicates. Each pair of number is <code>test-yours</code>. \n    3680 - 24276\n    4383 - 3893\n    4489 - 3426\n    4615 - 5015\n    4931 - 25806\n    5358 - 22559\n    5465 - 19257\n    6243 - 23909\n    6777 - 19279\n    7029 - 21922\n    7264 - 51</p>\n\n<p>The first row of the above list is displayed below. You can see that they applied a color altering filter. The green leaf in the test set is not as green as the original image (The image below that says \"Train 24276\" is image 24276 in external dataset <code>imagenet_no_test</code>).\n<img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1723677%2F06b5422dc47c79467c23683dbd442d35%2FScreen%20Shot%202020-05-08%20at%209.16.04%20AM.png?generation=1588955943801734&amp;alt=media\" alt=\"\"></p>",
          "rawMarkdown": "In your `imagenet_no_test`, I also found the following duplicates. Each pair of number is `test-yours`. \n    3680 - 24276\n    4383 - 3893\n    4489 - 3426\n    4615 - 5015\n    4931 - 25806\n    5358 - 22559\n    5465 - 19257\n    6243 - 23909\n    6777 - 19279\n    7029 - 21922\n    7264 - 51\n\nThe first row of the above list is displayed below. You can see that they applied a color altering filter. The green leaf in the test set is not as green as the original image (The image below that says \"Train 24276\" is image 24276 in external dataset `imagenet_no_test`).\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1723677%2F06b5422dc47c79467c23683dbd442d35%2FScreen%20Shot%202020-05-08%20at%209.16.04%20AM.png?generation=1588955943801734&amp;alt=media)\n"
        },
        {
          "id": 838573,
          "postDate": "2020-05-08T16:43:03.063Z",
          "content": "<p>Ok, i'm done double checking your <code>imagenet_no_test</code>. I've listed all the duplicates I found above. But even after we do all this work, I'm sure we've missed at least one, so I'm not sure if it is worth the time and risk for us to spend more hours double checking more datasets. </p>\n\n<p>None-the-less, if i have time later i will try. The method I am using is like <code>cosine-similarity</code>. I'm comparing the similarity of hidden activations of a trained (on competition train data) EfficientNetB4 between test images and external images.</p>",
          "rawMarkdown": "Ok, i'm done double checking your `imagenet_no_test`. I've listed all the duplicates I found above. But even after we do all this work, I'm sure we've missed at least one, so I'm not sure if it is worth the time and risk for us to spend more hours double checking more datasets. \n\nNone-the-less, if i have time later i will try. The method I am using is like `cosine-similarity`. I'm comparing the similarity of hidden activations of a trained (on competition train data) EfficientNetB4 between test images and external images.",
          "votes": 4
        },
        {
          "id": 838878,
          "postDate": "2020-05-08T21:26:56.213Z",
          "content": "<p>Hi <a href=\"/cdeotte\">@cdeotte</a>  Thank you very much for the update. This is very helpful, because it is clearly shows that we really not able to use external datasets as we not completely sure that we can find all test images in external datasets. \nSo, I agree with your <a href=\"https://www.kaggle.com/c/flower-classification-with-tpus/discussion/134701\">suggestion </a>to completely ban all external data in this competition.</p>\n\n<p>BTW. Original training data contains at least 27 test images.</p>",
          "rawMarkdown": "Hi @cdeotte  Thank you very much for the update. This is very helpful, because it is clearly shows that we really not able to use external datasets as we not completely sure that we can find all test images in external datasets. \nSo, I agree with your [suggestion ](https://www.kaggle.com/c/flower-classification-with-tpus/discussion/134701)to completely ban all external data in this competition.\n\nBTW. Original training data contains at least 27 test images.",
          "votes": 5
        },
        {
          "id": 839826,
          "postDate": "2020-05-09T16:05:26.173Z",
          "content": "<p><a href=\"/kirillblinov\">@kirillblinov</a> you meant the kaggle provided train dataset, or your external dataset has atleast 27 test images?</p>",
          "rawMarkdown": "@kirillblinov you meant the kaggle provided train dataset, or your external dataset has atleast 27 test images?"
        },
        {
          "id": 840001,
          "postDate": "2020-05-09T17:48:51.583Z",
          "content": "<p>External dataset has plenty of images that are similar to test images. Using this is not allowed.</p>\n\n<p>Default competition training set also has images that are similar to test images. Using this is ok. Because it is part of competition dataset. It will be difficult for anyone to separate this out and not even expected. In fact it could have been even added intentionally. And using it is ok.</p>",
          "rawMarkdown": "External dataset has plenty of images that are similar to test images. Using this is not allowed.\n\nDefault competition training set also has images that are similar to test images. Using this is ok. Because it is part of competition dataset. It will be difficult for anyone to separate this out and not even expected. In fact it could have been even added intentionally. And using it is ok.",
          "votes": 1
        },
        {
          "id": 840644,
          "postDate": "2020-05-10T08:28:03.740Z",
          "content": "<p><a href=\"/kurianbenoy\">@kurianbenoy</a>, I mean the Kaggle provided train dataset.</p>",
          "rawMarkdown": "@kurianbenoy, I mean the Kaggle provided train dataset."
        }
      ]
    },
    {
      "id": 799030,
      "postDate": "2020-04-06T05:10:34.567Z",
      "content": "<p>kaggle dataset\n<a href=\"https://www.kaggle.com/hengck23/externaldatasettpuflower\">https://www.kaggle.com/hengck23/externaldatasettpuflower</a></p>\n\n<p>see discussion on: <a href=\"https://www.kaggle.com/c/flower-classification-with-tpus/discussion/140866\">https://www.kaggle.com/c/flower-classification-with-tpus/discussion/140866</a></p>",
      "rawMarkdown": "kaggle dataset\nhttps://www.kaggle.com/hengck23/externaldatasettpuflower\n\nsee discussion on: https://www.kaggle.com/c/flower-classification-with-tpus/discussion/140866",
      "votes": 1
    },
    {
      "id": 796820,
      "postDate": "2020-04-04T00:31:11.127Z",
      "content": "<p>Oxford 102 category Flowers dataset from Kaggle datset <a href=\"https://www.kaggle.com/cdeotte/oxford-flowers-tfrecords\">here</a>. Kaggle said <a href=\"https://www.kaggle.com/c/flower-classification-with-tpus/discussion/134701#769297\">here</a> that it is ok to use this dataset.</p>",
      "rawMarkdown": "Oxford 102 category Flowers dataset from Kaggle datset [here][1]. Kaggle said [here][2] that it is ok to use this dataset.\n\n[1]: https://www.kaggle.com/cdeotte/oxford-flowers-tfrecords\n[2]: https://www.kaggle.com/c/flower-classification-with-tpus/discussion/134701#769297",
      "votes": 1
    },
    {
      "id": 825270,
      "postDate": "2020-04-28T22:18:23.257Z",
      "content": "<p>Heng CherKeng dataset converted to tfrecords\n<a href=\"https://www.kaggle.com/kirillblinov/tf-flower-photo-tfrec\">https://www.kaggle.com/kirillblinov/tf-flower-photo-tfrec</a></p>",
      "rawMarkdown": "Heng CherKeng dataset converted to tfrecords\nhttps://www.kaggle.com/kirillblinov/tf-flower-photo-tfrec",
      "votes": 2
    },
    {
      "id": 816814,
      "postDate": "2020-04-22T15:56:33.387Z",
      "content": "<p><a href=\"https://www.kaggle.com/ronyroy/flowers-test-pseudo-90-512\">https://www.kaggle.com/ronyroy/flowers-test-pseudo-90-512</a>\n<a href=\"https://www.kaggle.com/ronyroy/flowers-test-pseudo-99-512\">https://www.kaggle.com/ronyroy/flowers-test-pseudo-99-512</a> </p>\n\n<p>The idea was to use test data from a weak classifier to build a stronger classifier.\nImages from the test set that score more than 0.9 /0.99 512*512 resolution\nuse\n    LABELED_TFREC_FORMAT = {\n        \"image\": tf.io.FixedLenFeature([], tf.string), # tf.string means bytestring\n        \"class\": tf.io.FixedLenFeature([], tf.int64),  # shape [] means single element\n    }</p>\n\n<p>Disclosure: Didn't work out well for me... hope it helps someone else who knows how to do this right...</p>",
      "rawMarkdown": "https://www.kaggle.com/ronyroy/flowers-test-pseudo-90-512\nhttps://www.kaggle.com/ronyroy/flowers-test-pseudo-99-512 \n\nThe idea was to use test data from a weak classifier to build a stronger classifier.\nImages from the test set that score more than 0.9 /0.99 512*512 resolution\nuse\n    LABELED_TFREC_FORMAT = {\n        \"image\": tf.io.FixedLenFeature([], tf.string), # tf.string means bytestring\n        \"class\": tf.io.FixedLenFeature([], tf.int64),  # shape [] means single element\n    }\n\nDisclosure: Didn't work out well for me... hope it helps someone else who knows how to do this right..."
    },
    {
      "id": 804955,
      "postDate": "2020-04-12T07:23:05.130Z",
      "content": "<p>Oxford Flowers but with filtered out mutual images (basing on high image SSI).\n<a href=\"https://www.kaggle.com/szacho/oxford-102-for-tpu-competition\">https://www.kaggle.com/szacho/oxford-102-for-tpu-competition</a></p>",
      "rawMarkdown": "Oxford Flowers but with filtered out mutual images (basing on high image SSI).\nhttps://www.kaggle.com/szacho/oxford-102-for-tpu-competition",
      "replies": [
        {
          "id": 824502,
          "postDate": "2020-04-28T12:16:22.507Z",
          "content": "<p>I want to make an external dataset for the images that <a href=\"/hengck23\">@hengck23</a> posted above. Would you be able to post the scripts you used to process Oxford 102?</p>",
          "rawMarkdown": "I want to make an external dataset for the images that @hengck23 posted above. Would you be able to post the scripts you used to process Oxford 102?"
        },
        {
          "id": 824766,
          "postDate": "2020-04-28T15:37:37.030Z",
          "content": "<p>Sure, but it's not documented since I didn't know I'm going to share it. You will need to create 3 directories of flowers. The first one contains low-resolution images you want to filter out (like oxford dataset) and the second one the same resolution of images from kaggle competition. The third directory is a directory you want to remove images from. So, you compare 2 low-resolution (in my case 224x224) images to remove the image of higher resolution (saves a lot of time). Also, this script groups flowers by their classes, so they are compared only within their own species. I used aspect aware resizer for resizing external images and named every .jpg file in this pattern 'INDEX_CLASSNAME.jpg' (e. g. '00038_passion flower.jpg'). Here is the code: <a href=\"https://bit.ly/2KIBPE8\">https://bit.ly/2KIBPE8</a> . </p>",
          "rawMarkdown": "Sure, but it's not documented since I didn't know I'm going to share it. You will need to create 3 directories of flowers. The first one contains low-resolution images you want to filter out (like oxford dataset) and the second one the same resolution of images from kaggle competition. The third directory is a directory you want to remove images from. So, you compare 2 low-resolution (in my case 224x224) images to remove the image of higher resolution (saves a lot of time). Also, this script groups flowers by their classes, so they are compared only within their own species. I used aspect aware resizer for resizing external images and named every .jpg file in this pattern 'INDEX_CLASSNAME.jpg' (e. g. '00038_passion flower.jpg'). Here is the code: https://bit.ly/2KIBPE8 . "
        }
      ]
    },
    {
      "id": 755672,
      "postDate": "2020-02-25T02:30:42.903Z",
      "rawMarkdown": ""
    },
    {
      "id": 840635,
      "postDate": "2020-05-10T08:21:26.963Z",
      "rawMarkdown": "",
      "votes": -2,
      "isDeleted": true
    },
    {
      "id": 836699,
      "postDate": "2020-05-07T06:54:25.530Z",
      "rawMarkdown": "",
      "isDeleted": true
    }
  ],
  "comments": [
    {
      "id": 745340,
      "author_name": "Martin Görner",
      "author_url": "",
      "post_date": "2020-02-13T18:16:57.987000",
      "content": "<p>And to get the GCS bucket path of a Kaggle dataset, in case your notebook uses two datasets or more, do this:\n<code>\ngcs_path = KaggleDatasets().get_gcs_path('kaggle_dataset_directory_name')\n</code>\nTo get the Kaggle dataset directory name, the easiest way is to attach the dataset to a notebook and then run this in order to list all attached dataset directories:\n<code>\n!ls /kaggle/input\n</code></p>",
      "votes": 4,
      "replies": []
    },
    {
      "id": 836352,
      "author_name": "Julia Elliott",
      "author_url": "",
      "post_date": "2020-05-06T23:02:05.590000",
      "content": "<p>Here is a clarification regarding use of external datasets that mirror the same images as the test set. I want to reiterate that the rules prohibit training on test set images: “Submissions may not use or incorporate information from hand labeling or human prediction of the validation dataset or test data records.” So, while you are permitted to use the 5 datasets that we used to source the training/test sets as <a href=\"https://www.kaggle.com/c/flower-classification-with-tpus/discussion/134701#769297\">I stated previously here</a>, <strong>it is prohibited to train your models using any mirrors/overlaps/duplication of the competition's test data within other flower datasets, as this conflicts with the above-stated rule.</strong></p>\n\n<p>All potential winners will be required to hand over their full code, including specifying the exact external datasets they've used. We will be cross-checking training images with those contained in the test set to confirm no overlaps as a condition to awarding prizes. If you've already committed this violation, you should ensure you do not select any submissions making use of prohibited data for your final 2 submissions that will be evaluated on the final Private Leaderboard.</p>",
      "votes": 1,
      "replies": [
        {
          "id": 836391,
          "author_name": "Chris Deotte",
          "author_url": "",
          "post_date": "2020-05-07T00:14:52.613000",
          "content": "<p>Are we allowed to use pseudo labeled external data? (i.e. download external datasets then use a model trained on competition training data to predict the labels on external images?)</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 836608,
          "author_name": "",
          "author_url": "",
          "post_date": "2020-05-07T05:30:58.330000",
          "content": "",
          "votes": 1,
          "replies": []
        },
        {
          "id": 837430,
          "author_name": "Julia Elliott",
          "author_url": "",
          "post_date": "2020-05-07T19:04:55.063000",
          "content": "<p>A re-post of <a href=\"https://www.kaggle.com/c/flower-classification-with-tpus/discussion/148329#837427\">my reply to a separate thread</a> which addresses these:</p>\n\n<blockquote>\n  <ul>\n  <li>Pseudolabeling is permitted unless you're pseudolabeling using images from or duplicative of images in the test set.</li>\n  <li>Using publicly-available pre-trained models is allowed, including those with components of the test set underlying the pre-trained models. To be super clear, this means you're not permitted to be \"pre-training\" models using the test set and then calling that \"pre-trained,\" but you can use the host of ImageNet or other canned publicly-available pre-trained models mentioned in preceding posts.</li>\n  <li>The currently published datasets can be used (including the 5 datasets from which the training/test sets were sourced), but images which overlap with or are the same as the test set cannot be used (should be removed). Correct, no new external datasets can be announced.</li>\n  </ul>\n</blockquote>\n\n<p>Lastly, a reiteration that this is a <strong>playground competition</strong>. As with all playgrounds, it was created in the spirit of learning (specifically for the launch of our TPU integration), more than the pursuit of winning. Hence, it is not awarding points/medals, and prizes are non-monetary &amp; centered around TPU use. We launched this with the knowledge there would be vulnerabilities and more grey area than we'd prefer in a featured competition, hence sending it to the playground. I hope sound judgment is used by participants that aligns with this competition's spirit, so that we can continue to offer opportunities like this.</p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 837817,
          "author_name": "Kirill Blinov",
          "author_url": "",
          "post_date": "2020-05-08T04:14:51.437000",
          "content": "<p>Hi <a href=\"/juliaelliott\">@juliaelliott</a>，I have removed all test images (using imagehash and image similarity) from all five external datasets and uploaded this corrected data to the dataset mentioned before in this thread\n<a href=\"https://www.kaggle.com/kirillblinov/tf-flower-photo-tfrec\">https://www.kaggle.com/kirillblinov/tf-flower-photo-tfrec</a>\nCould you please confirm may or not we use this corrected data in the competition? </p>",
          "votes": 3,
          "replies": []
        },
        {
          "id": 837844,
          "author_name": "Kurian Benoy",
          "author_url": "",
          "post_date": "2020-05-08T04:52:43.977000",
          "content": "<p>Thanks for the update <a href=\"/kirillblinov\">@kirillblinov</a> </p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 838521,
          "author_name": "Chris Deotte",
          "author_url": "",
          "post_date": "2020-05-08T16:12:08.747000",
          "content": "<p>Hi <a href=\"/kirillblinov\">@kirillblinov</a> Thanks for your wonderful dataset and thanks for trying to clean it. I'm double checking your work and unfortunately you've missed many test images. In the following example, Kaggle rotated the external image by 90 degrees before using it. This is image 1126 in your <code>imagenet_no_test</code> images and 797 in the test set.</p>\n\n<p><img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1723677%2Ff6ab8988f3d5c29933bac417429596a2%2Frotate.png?generation=1588954161325113&amp;alt=media\" alt=\"\"></p>\n\n<p>The index 1126 into your images are when reading your files as follows\n    ['../../../Flowers/external2/imagenet_no_test//tfrecords-jpeg-224x224/0-224x224-1785.tfrec',\n '../../../Flowers/external2/imagenet_no_test//tfrecords-jpeg-224x224/5-224x224-1785.tfrec',\n '../../../Flowers/external2/imagenet_no_test//tfrecords-jpeg-224x224/2-224x224-1785.tfrec',\n '../../../Flowers/external2/imagenet_no_test//tfrecords-jpeg-224x224/12-224x224-1785.tfrec',\n '../../../Flowers/external2/imagenet_no_test//tfrecords-jpeg-224x224/13-224x224-1785.tfrec',\n '../../../Flowers/external2/imagenet_no_test//tfrecords-jpeg-224x224/9-224x224-1785.tfrec',\n '../../../Flowers/external2/imagenet_no_test//tfrecords-jpeg-224x224/7-224x224-1785.tfrec',\n '../../../Flowers/external2/imagenet_no_test//tfrecords-jpeg-224x224/14-224x224-1785.tfrec',\n '../../../Flowers/external2/imagenet_no_test//tfrecords-jpeg-224x224/15-224x224-1774.tfrec',\n '../../../Flowers/external2/imagenet_no_test//tfrecords-jpeg-224x224/6-224x224-1785.tfrec',\n '../../../Flowers/external2/imagenet_no_test//tfrecords-jpeg-224x224/11-224x224-1785.tfrec',\n '../../../Flowers/external2/imagenet_no_test//tfrecords-jpeg-224x224/8-224x224-1785.tfrec',\n '../../../Flowers/external2/imagenet_no_test//tfrecords-jpeg-224x224/10-224x224-1785.tfrec',\n '../../../Flowers/external2/imagenet_no_test//tfrecords-jpeg-224x224/1-224x224-1785.tfrec',\n '../../../Flowers/external2/imagenet_no_test//tfrecords-jpeg-224x224/4-224x224-1785.tfrec',\n '../../../Flowers/external2/imagenet_no_test//tfrecords-jpeg-224x224/3-224x224-1785.tfrec']</p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 838529,
          "author_name": "Chris Deotte",
          "author_url": "",
          "post_date": "2020-05-08T16:18:43.377000",
          "content": "<p>Here's another one they rotated 5 degrees counter clockwise. (The image below that says \"Train 23525\" is image 23525 in external dataset <code>imagenet_no_test</code>).\n<img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1723677%2F0ed1d2be6ddc5b8466b2d029d47ce37d%2Frotate2.png?generation=1588954720569101&amp;alt=media\" alt=\"\"></p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 838535,
          "author_name": "Chris Deotte",
          "author_url": "",
          "post_date": "2020-05-08T16:23:03.737000",
          "content": "<p>Here's an interesting one. It's a different photo but it's of the same flower just taking from a slightly different angle. I'm not sure if we can include this one or not. (The image below that says \"Train 25322\" is image 25322 in external dataset <code>imagenet_no_test</code>).\n<img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1723677%2F719ed5ff4a9b465532c1cee9b96f2f6c%2Fphoto.png?generation=1588954910061954&amp;alt=media\" alt=\"\"></p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 838565,
          "author_name": "Chris Deotte",
          "author_url": "",
          "post_date": "2020-05-08T16:39:28.413000",
          "content": "<p>In your <code>imagenet_no_test</code>, I also found the following duplicates. Each pair of number is <code>test-yours</code>. \n    3680 - 24276\n    4383 - 3893\n    4489 - 3426\n    4615 - 5015\n    4931 - 25806\n    5358 - 22559\n    5465 - 19257\n    6243 - 23909\n    6777 - 19279\n    7029 - 21922\n    7264 - 51</p>\n\n<p>The first row of the above list is displayed below. You can see that they applied a color altering filter. The green leaf in the test set is not as green as the original image (The image below that says \"Train 24276\" is image 24276 in external dataset <code>imagenet_no_test</code>).\n<img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1723677%2F06b5422dc47c79467c23683dbd442d35%2FScreen%20Shot%202020-05-08%20at%209.16.04%20AM.png?generation=1588955943801734&amp;alt=media\" alt=\"\"></p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 838573,
          "author_name": "Chris Deotte",
          "author_url": "",
          "post_date": "2020-05-08T16:43:03.063000",
          "content": "<p>Ok, i'm done double checking your <code>imagenet_no_test</code>. I've listed all the duplicates I found above. But even after we do all this work, I'm sure we've missed at least one, so I'm not sure if it is worth the time and risk for us to spend more hours double checking more datasets. </p>\n\n<p>None-the-less, if i have time later i will try. The method I am using is like <code>cosine-similarity</code>. I'm comparing the similarity of hidden activations of a trained (on competition train data) EfficientNetB4 between test images and external images.</p>",
          "votes": 4,
          "replies": []
        },
        {
          "id": 838878,
          "author_name": "Kirill Blinov",
          "author_url": "",
          "post_date": "2020-05-08T21:26:56.213000",
          "content": "<p>Hi <a href=\"/cdeotte\">@cdeotte</a>  Thank you very much for the update. This is very helpful, because it is clearly shows that we really not able to use external datasets as we not completely sure that we can find all test images in external datasets. \nSo, I agree with your <a href=\"https://www.kaggle.com/c/flower-classification-with-tpus/discussion/134701\">suggestion </a>to completely ban all external data in this competition.</p>\n\n<p>BTW. Original training data contains at least 27 test images.</p>",
          "votes": 5,
          "replies": []
        },
        {
          "id": 839826,
          "author_name": "Kurian Benoy",
          "author_url": "",
          "post_date": "2020-05-09T16:05:26.173000",
          "content": "<p><a href=\"/kirillblinov\">@kirillblinov</a> you meant the kaggle provided train dataset, or your external dataset has atleast 27 test images?</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 840001,
          "author_name": "haveri",
          "author_url": "",
          "post_date": "2020-05-09T17:48:51.583000",
          "content": "<p>External dataset has plenty of images that are similar to test images. Using this is not allowed.</p>\n\n<p>Default competition training set also has images that are similar to test images. Using this is ok. Because it is part of competition dataset. It will be difficult for anyone to separate this out and not even expected. In fact it could have been even added intentionally. And using it is ok.</p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 840644,
          "author_name": "Kirill Blinov",
          "author_url": "",
          "post_date": "2020-05-10T08:28:03.740000",
          "content": "<p><a href=\"/kurianbenoy\">@kurianbenoy</a>, I mean the Kaggle provided train dataset.</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 799030,
      "author_name": "hengck23",
      "author_url": "",
      "post_date": "2020-04-06T05:10:34.567000",
      "content": "<p>kaggle dataset\n<a href=\"https://www.kaggle.com/hengck23/externaldatasettpuflower\">https://www.kaggle.com/hengck23/externaldatasettpuflower</a></p>\n\n<p>see discussion on: <a href=\"https://www.kaggle.com/c/flower-classification-with-tpus/discussion/140866\">https://www.kaggle.com/c/flower-classification-with-tpus/discussion/140866</a></p>",
      "votes": 1,
      "replies": []
    },
    {
      "id": 796820,
      "author_name": "Chris Deotte",
      "author_url": "",
      "post_date": "2020-04-04T00:31:11.127000",
      "content": "<p>Oxford 102 category Flowers dataset from Kaggle datset <a href=\"https://www.kaggle.com/cdeotte/oxford-flowers-tfrecords\">here</a>. Kaggle said <a href=\"https://www.kaggle.com/c/flower-classification-with-tpus/discussion/134701#769297\">here</a> that it is ok to use this dataset.</p>",
      "votes": 1,
      "replies": []
    },
    {
      "id": 825270,
      "author_name": "Kirill Blinov",
      "author_url": "",
      "post_date": "2020-04-28T22:18:23.257000",
      "content": "<p>Heng CherKeng dataset converted to tfrecords\n<a href=\"https://www.kaggle.com/kirillblinov/tf-flower-photo-tfrec\">https://www.kaggle.com/kirillblinov/tf-flower-photo-tfrec</a></p>",
      "votes": 2,
      "replies": []
    },
    {
      "id": 816814,
      "author_name": "Rony K Roy",
      "author_url": "",
      "post_date": "2020-04-22T15:56:33.387000",
      "content": "<p><a href=\"https://www.kaggle.com/ronyroy/flowers-test-pseudo-90-512\">https://www.kaggle.com/ronyroy/flowers-test-pseudo-90-512</a>\n<a href=\"https://www.kaggle.com/ronyroy/flowers-test-pseudo-99-512\">https://www.kaggle.com/ronyroy/flowers-test-pseudo-99-512</a> </p>\n\n<p>The idea was to use test data from a weak classifier to build a stronger classifier.\nImages from the test set that score more than 0.9 /0.99 512*512 resolution\nuse\n    LABELED_TFREC_FORMAT = {\n        \"image\": tf.io.FixedLenFeature([], tf.string), # tf.string means bytestring\n        \"class\": tf.io.FixedLenFeature([], tf.int64),  # shape [] means single element\n    }</p>\n\n<p>Disclosure: Didn't work out well for me... hope it helps someone else who knows how to do this right...</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 804955,
      "author_name": "Michał Szachniewicz",
      "author_url": "",
      "post_date": "2020-04-12T07:23:05.130000",
      "content": "<p>Oxford Flowers but with filtered out mutual images (basing on high image SSI).\n<a href=\"https://www.kaggle.com/szacho/oxford-102-for-tpu-competition\">https://www.kaggle.com/szacho/oxford-102-for-tpu-competition</a></p>",
      "votes": 0,
      "replies": [
        {
          "id": 824502,
          "author_name": "Sapphire Brand",
          "author_url": "",
          "post_date": "2020-04-28T12:16:22.507000",
          "content": "<p>I want to make an external dataset for the images that <a href=\"/hengck23\">@hengck23</a> posted above. Would you be able to post the scripts you used to process Oxford 102?</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 824766,
          "author_name": "Michał Szachniewicz",
          "author_url": "",
          "post_date": "2020-04-28T15:37:37.030000",
          "content": "<p>Sure, but it's not documented since I didn't know I'm going to share it. You will need to create 3 directories of flowers. The first one contains low-resolution images you want to filter out (like oxford dataset) and the second one the same resolution of images from kaggle competition. The third directory is a directory you want to remove images from. So, you compare 2 low-resolution (in my case 224x224) images to remove the image of higher resolution (saves a lot of time). Also, this script groups flowers by their classes, so they are compared only within their own species. I used aspect aware resizer for resizing external images and named every .jpg file in this pattern 'INDEX_CLASSNAME.jpg' (e. g. '00038_passion flower.jpg'). Here is the code: <a href=\"https://bit.ly/2KIBPE8\">https://bit.ly/2KIBPE8</a> . </p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 755672,
      "author_name": "SwordFaith",
      "author_url": "",
      "post_date": "2020-02-25T02:30:42.903000",
      "content": "",
      "votes": 0,
      "replies": []
    },
    {
      "id": 840635,
      "author_name": "",
      "author_url": "",
      "post_date": "2020-05-10T08:21:26.963000",
      "content": "",
      "votes": -2,
      "replies": []
    },
    {
      "id": 836699,
      "author_name": "",
      "author_url": "",
      "post_date": "2020-05-07T06:54:25.530000",
      "content": "",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "745340": "And to get the GCS bucket path of a Kaggle dataset, in case your notebook uses two datasets or more, do this:\n```\ngcs_path = KaggleDatasets().get_gcs_path('kaggle_dataset_directory_name')\n```\nTo get the Kaggle dataset directory name, the easiest way is to attach the dataset to a notebook and then run this in order to list all attached dataset directories:\n```\n!ls /kaggle/input\n```",
    "744827": "Per the [Competition Rules](https://www.kaggle.com/c/flower-classification-with-tpus/rules), freely and publicly available external data is permitted in this competition, but must be posted to this forum thread no later than the Entry Deadline (one week before competition close).\n\nNote for this competition you cannot use privately-held external datasets with the TPU integration. Instead, you'll want to upload a public dataset to Kaggle to use and declare those here.\n\nOnce someone posts an external dataset to this thread, you do not need to re-post it if you are using the same one.\n\nYou only need to declare the original dataset used; you do not need to declare re-labeled or augmented or otherwise processed versions of datasets. Pre-trained models can be declared, as well; however, models resulting from your own original work, that you have trained yourself offline do not need to be shared/declared.",
    "836352": "Here is a clarification regarding use of external datasets that mirror the same images as the test set. I want to reiterate that the rules prohibit training on test set images: “Submissions may not use or incorporate information from hand labeling or human prediction of the validation dataset or test data records.” So, while you are permitted to use the 5 datasets that we used to source the training/test sets as [I stated previously here](https://www.kaggle.com/c/flower-classification-with-tpus/discussion/134701#769297), **it is prohibited to train your models using any mirrors/overlaps/duplication of the competition's test data within other flower datasets, as this conflicts with the above-stated rule.**\n\nAll potential winners will be required to hand over their full code, including specifying the exact external datasets they've used. We will be cross-checking training images with those contained in the test set to confirm no overlaps as a condition to awarding prizes. If you've already committed this violation, you should ensure you do not select any submissions making use of prohibited data for your final 2 submissions that will be evaluated on the final Private Leaderboard.",
    "799030": "kaggle dataset\nhttps://www.kaggle.com/hengck23/externaldatasettpuflower\n\nsee discussion on: https://www.kaggle.com/c/flower-classification-with-tpus/discussion/140866",
    "796820": "Oxford 102 category Flowers dataset from Kaggle datset [here][1]. Kaggle said [here][2] that it is ok to use this dataset.\n\n[1]: https://www.kaggle.com/cdeotte/oxford-flowers-tfrecords\n[2]: https://www.kaggle.com/c/flower-classification-with-tpus/discussion/134701#769297",
    "825270": "Heng CherKeng dataset converted to tfrecords\nhttps://www.kaggle.com/kirillblinov/tf-flower-photo-tfrec",
    "816814": "https://www.kaggle.com/ronyroy/flowers-test-pseudo-90-512\nhttps://www.kaggle.com/ronyroy/flowers-test-pseudo-99-512 \n\nThe idea was to use test data from a weak classifier to build a stronger classifier.\nImages from the test set that score more than 0.9 /0.99 512*512 resolution\nuse\n    LABELED_TFREC_FORMAT = {\n        \"image\": tf.io.FixedLenFeature([], tf.string), # tf.string means bytestring\n        \"class\": tf.io.FixedLenFeature([], tf.int64),  # shape [] means single element\n    }\n\nDisclosure: Didn't work out well for me... hope it helps someone else who knows how to do this right...",
    "804955": "Oxford Flowers but with filtered out mutual images (basing on high image SSI).\nhttps://www.kaggle.com/szacho/oxford-102-for-tpu-competition",
    "755672": "",
    "840635": "",
    "836699": ""
  }
}