{
  "id": 301959,
  "title": "LB (TOP20) solutions SPEED 💥💥  analysis ",
  "url": "/competitions/tensorflow-great-barrier-reef/discussion/301959",
  "author_name": "",
  "post_date": "2022-01-20T08:35:58.918535900Z",
  "votes": 49,
  "comment_count": 17,
  "views": 0,
  "content": "<p>Recently, we had a lot of questions about whether 9h is enough time to run the competition notebook. Is Kaggle GPU enough to win the competition? Should I scale to huge resolutions ( it will have a significant impact on the speed of the solution)? </p>\n<p>I analyzed the times of the TOP20 solutions (TOP20 LB team submit time). For this, I wrote a Python script (using Kaggle API), ran it and … went drinking coffee waiting for …. results. Of course, I will not give you exact information about the team's results (due to respect for their work and I am not giving shortcuts).</p>\n<p>Assumption: </p>\n<ul>\n<li>This stats do not show the best solution time. I assume that people submit solution looking for improvement so time is correlated. To collect best solution time I sould look for team LB jump.</li>\n<li>data was collected on 2022.01.19</li>\n</ul>\n<p><strong>TOP5 (GREEN zone / paid zone)</strong></p>\n<ul>\n<li>minimum time: 130 min (2h 10min)</li>\n<li>maximum time: 464 min (7h 44min)</li>\n<li>avg time: 239 min (3h 59min)</li>\n</ul>\n<p><strong>TOP6-13 (GOLD ZONE / excluding TOP5)</strong></p>\n<ul>\n<li>minimum time: 55 min</li>\n<li>maximum time: 460 min (7h 40min)</li>\n<li>avg time: 213 min (3h 33 min)</li>\n</ul>\n<p><strong>TOP20 (ALL SOLUTIONS)</strong></p>\n<ul>\n<li>minimum time: 49 min</li>\n<li>maximum time: 464 min (7h 44min)</li>\n<li>avg time: 211 min (3h 31min)</li>\n</ul>\n<p>As we can see there are different solutions: </p>\n<ul>\n<li>from very light (49 min) to …. time demanding (7h 44min) … -&gt; think about current solution you use (Yolo5/YoloX/YoloR/FasterRCNN …., model size … nano/S/M/L, inference techniques - WBF/TTA/ blending (NMS)/Tracking etc. -&gt; this influence on speed),</li>\n<li>you do not need worry about 9h limit - to be in GOLD area you need only … or .. less then 60 min (for sure less then 4h), </li>\n<li>you do not need many GPUs to be in TOP10 … (I am almost sure that fast speed does not require many GPUs during training as well - these solution are just effect of thinking … - respect for those people),</li>\n<li>I suppose part of teams (TOP20) are going to take prize in two categories (general and speed) (this is my hypothesis only - not fact) - TF is required (is Yolo5 all you need?)</li>\n</ul>\n<p>This post is to motivate people who gave up having a thought in their head - is it require to have 10xGPU and I need over 9h to win …</p>",
  "messages": [
    {
      "id": "1657505",
      "postDate": "01/20/2022 08:35:58",
      "content": "<p>Recently, we had a lot of questions about whether 9h is enough time to run the competition notebook. Is Kaggle GPU enough to win the competition? Should I scale to huge resolutions ( it will have a significant impact on the speed of the solution)? </p>\n<p>I analyzed the times of the TOP20 solutions (TOP20 LB team submit time). For this, I wrote a Python script (using Kaggle API), ran it and … went drinking coffee waiting for …. results. Of course, I will not give you exact information about the team's results (due to respect for their work and I am not giving shortcuts).</p>\n<p>Assumption: </p>\n<ul>\n<li>This stats do not show the best solution time. I assume that people submit solution looking for improvement so time is correlated. To collect best solution time I sould look for team LB jump.</li>\n<li>data was collected on 2022.01.19</li>\n</ul>\n<p><strong>TOP5 (GREEN zone / paid zone)</strong></p>\n<ul>\n<li>minimum time: 130 min (2h 10min)</li>\n<li>maximum time: 464 min (7h 44min)</li>\n<li>avg time: 239 min (3h 59min)</li>\n</ul>\n<p><strong>TOP6-13 (GOLD ZONE / excluding TOP5)</strong></p>\n<ul>\n<li>minimum time: 55 min</li>\n<li>maximum time: 460 min (7h 40min)</li>\n<li>avg time: 213 min (3h 33 min)</li>\n</ul>\n<p><strong>TOP20 (ALL SOLUTIONS)</strong></p>\n<ul>\n<li>minimum time: 49 min</li>\n<li>maximum time: 464 min (7h 44min)</li>\n<li>avg time: 211 min (3h 31min)</li>\n</ul>\n<p>As we can see there are different solutions: </p>\n<ul>\n<li>from very light (49 min) to …. time demanding (7h 44min) … -&gt; think about current solution you use (Yolo5/YoloX/YoloR/FasterRCNN …., model size … nano/S/M/L, inference techniques - WBF/TTA/ blending (NMS)/Tracking etc. -&gt; this influence on speed),</li>\n<li>you do not need worry about 9h limit - to be in GOLD area you need only … or .. less then 60 min (for sure less then 4h), </li>\n<li>you do not need many GPUs to be in TOP10 … (I am almost sure that fast speed does not require many GPUs during training as well - these solution are just effect of thinking … - respect for those people),</li>\n<li>I suppose part of teams (TOP20) are going to take prize in two categories (general and speed) (this is my hypothesis only - not fact) - TF is required (is Yolo5 all you need?)</li>\n</ul>\n<p>This post is to motivate people who gave up having a thought in their head - is it require to have 10xGPU and I need over 9h to win …</p>",
      "rawMarkdown": "Recently, we had a lot of questions about whether 9h is enough time to run the competition notebook. Is Kaggle GPU enough to win the competition? Should I scale to huge resolutions ( it will have a significant impact on the speed of the solution)? \n\nI analyzed the times of the TOP20 solutions (TOP20 LB team submit time). For this, I wrote a Python script (using Kaggle API), ran it and ... went drinking coffee waiting for .... results. Of course, I will not give you exact information about the team's results (due to respect for their work and I am not giving shortcuts).\n\nAssumption: \n- This stats do not show the best solution time. I assume that people submit solution looking for improvement so time is correlated. To collect best solution time I sould look for team LB jump.\n- data was collected on 2022.01.19\n\n**TOP5 (GREEN zone / paid zone)**\n- minimum time: 130 min (2h 10min)\n- maximum time: 464 min (7h 44min)\n- avg time: 239 min (3h 59min)\n\n**TOP6-13 (GOLD ZONE / excluding TOP5)**\n- minimum time: 55 min\n- maximum time: 460 min (7h 40min)\n- avg time: 213 min (3h 33 min)\n\n**TOP20 (ALL SOLUTIONS)**\n- minimum time: 49 min\n- maximum time: 464 min (7h 44min)\n- avg time: 211 min (3h 31min)\n\nAs we can see there are different solutions: \n- from very light (49 min) to .... time demanding (7h 44min) ... -> think about current solution you use (Yolo5/YoloX/YoloR/FasterRCNN ...., model size ... nano/S/M/L, inference techniques - WBF/TTA/ blending (NMS)/Tracking etc. -> this influence on speed),\n- you do not need worry about 9h limit - to be in GOLD area you need only … or .. less then 60 min (for sure less then 4h), \n- you do not need many GPUs to be in TOP10 ... (I am almost sure that fast speed does not require many GPUs during training as well - these solution are just effect of thinking ... - respect for those people),\n- I suppose part of teams (TOP20) are going to take prize in two categories (general and speed) (this is my hypothesis only - not fact) - TF is required (is Yolo5 all you need?)\n\nThis post is to motivate people who gave up having a thought in their head - is it require to have 10xGPU and I need over 9h to win ...",
      "votes": null
    },
    {
      "id": "1657509",
      "postDate": "01/20/2022 08:39:30",
      "content": "<p>It will be interesting to see the winners of TF performance prizes. Most common fast models such as YOLO are all generally trained and implemented in PyTorch, I wonder if any teams have recreated these models in TF just for eligibility…</p>",
      "rawMarkdown": "It will be interesting to see the winners of TF performance prizes. Most common fast models such as YOLO are all generally trained and implemented in PyTorch, I wonder if any teams have recreated these models in TF just for eligibility...",
      "votes": null
    },
    {
      "id": "1657573",
      "postDate": "01/20/2022 09:48:27",
      "content": "<p>Wow! I didn't know that via kaggle API you have access to ALL teams submissions times </p>",
      "rawMarkdown": "Wow! I didn't know that via kaggle API you have access to ALL teams submissions times",
      "votes": null
    },
    {
      "id": "1657583",
      "postDate": "01/20/2022 09:58:24",
      "content": "<p>Kaggle API has access to leaderboard data - data which you can see on Leaderboard page. This is not a secret (data is avaliable and we can use it for analysis). This is competition so … strategy, tactics … ability to use forum, public solutions, leaderboard etc. is important :) </p>",
      "rawMarkdown": "Kaggle API has access to leaderboard data - data which you can see on Leaderboard page. This is not a secret (data is avaliable and we can use it for analysis). This is competition so ... strategy, tactics ... ability to use forum, public solutions, leaderboard etc. is important :)",
      "votes": null
    },
    {
      "id": "1657590",
      "postDate": "01/20/2022 10:06:05",
      "content": "<p>I agree, i'm still looking for TF implementations.👀</p>",
      "rawMarkdown": "I agree, i'm still looking for TF implementations.👀",
      "votes": null
    },
    {
      "id": "1657592",
      "postDate": "01/20/2022 10:06:36",
      "content": "<p>You can export pytorch model to tf … using onnx (<a href=\"https://github.com/onnx/onnx\" target=\"_blank\">https://github.com/onnx/onnx</a>) :) and as I understand rules it is ok 😄</p>",
      "rawMarkdown": "You can export pytorch model to tf ... using onnx (https://github.com/onnx/onnx) :) and as I understand rules it is ok 😄",
      "votes": null
    },
    {
      "id": "1657595",
      "postDate": "01/20/2022 10:08:12",
      "content": "<p>No it is not okay, it needs to be trained also in tensorflow.</p>\n<blockquote>\n  <p>Use TensorFlow 2.X on GPU (training and inference)</p>\n</blockquote>",
      "rawMarkdown": "No it is not okay, it needs to be trained also in tensorflow.\n\n> Use TensorFlow 2.X on GPU (training and inference)",
      "votes": null
    },
    {
      "id": "1657599",
      "postDate": "01/20/2022 10:10:45",
      "content": "<p>OK. Sorry for my short cut :) </p>",
      "rawMarkdown": "OK. Sorry for my short cut :)",
      "votes": null
    },
    {
      "id": "1657609",
      "postDate": "01/20/2022 10:18:48",
      "content": "<p><code>To collect best solution time I should look for team LB jump</code></p>\n<p>like the idea :) Thanks for sharing !</p>",
      "rawMarkdown": "`To collect best solution time I should look for team LB jump`\n\nlike the idea :) Thanks for sharing !",
      "votes": null
    },
    {
      "id": "1657617",
      "postDate": "01/20/2022 10:26:08",
      "content": "<p>I do not have time implement this (since we have object detection competition … not LB competition analysis) … but … collecting one data is enough to answer on many questions.   </p>",
      "rawMarkdown": "I do not have time implement this (since we have object detection competition ... not LB competition analysis) ... but ... collecting one data is enough to answer on many questions.",
      "votes": null
    },
    {
      "id": "1657790",
      "postDate": "01/20/2022 13:17:27",
      "content": "<p>so no TPU is allowed for training?  that's unfortunate considering it is TensorFlow and TPU on kaggle gives some options.  </p>",
      "rawMarkdown": "so no TPU is allowed for training?  that's unfortunate considering it is TensorFlow and TPU on kaggle gives some options.",
      "votes": null
    },
    {
      "id": "1657841",
      "postDate": "01/20/2022 13:56:53",
      "content": "<p>based on public notebooks,  TF implementations underperformance compared to PyTorch  model</p>",
      "rawMarkdown": "based on public notebooks,  TF implementations underperformance compared to PyTorch  model",
      "votes": null
    },
    {
      "id": "1658648",
      "postDate": "01/21/2022 07:42:38",
      "content": "<p>Is there information about submission time given in our submit page ? I can only find the timing for running public set. <br>\nHow did you get the submission time for hidden set ? </p>",
      "rawMarkdown": "Is there information about submission time given in our submit page ? I can only find the timing for running public set. \nHow did you get the submission time for hidden set ?",
      "votes": null
    },
    {
      "id": "1658653",
      "postDate": "01/21/2022 07:47:21",
      "content": "<p>Your submission runs on both the public and private sets when you submit.</p>",
      "rawMarkdown": "Your submission runs on both the public and private sets when you submit.",
      "votes": null
    },
    {
      "id": "1658695",
      "postDate": "01/21/2022 08:32:28",
      "content": "<p>Sorry, I think I don't make myself clear.  When we run the kernel, we only get 3 rows submission. The time shown in the kernel is this one.<br>\nAfter we press submit, it runs for a long time before we get to see the score (This is the one which has 9 hour limit).  Is it possible to get this timing , or this is obtained by guessing based on the 3 rows ? <br>\nI have the feeling <a href=\"https://www.kaggle.com/remekkinas\" target=\"_blank\">@remekkinas</a>  knows this number in exact</p>",
      "rawMarkdown": "Sorry, I think I don't make myself clear.  When we run the kernel, we only get 3 rows submission. The time shown in the kernel is this one.\nAfter we press submit, it runs for a long time before we get to see the score (This is the one which has 9 hour limit).  Is it possible to get this timing , or this is obtained by guessing based on the 3 rows ? \nI have the feeling @remekkinas  knows this number in exact",
      "votes": null
    },
    {
      "id": "1658699",
      "postDate": "01/21/2022 08:38:31",
      "content": "<p><a href=\"https://ibb.co/2PT2vsW\"><img src=\"https://i.ibb.co/82H3rdN/runtime.jpg\" alt=\"runtime\"></a></p>\n<p>34 seconds run. No info about the submission time  (I am measuring the time I click submit until the time score appear, but usually I'm already sleeping when the score appear)</p>",
      "rawMarkdown": "<a href=\"https://ibb.co/2PT2vsW\"><img src=\"https://i.ibb.co/82H3rdN/runtime.jpg\" alt=\"runtime\" border=\"0\" /></a>\n\n34 seconds run. No info about the submission time  (I am measuring the time I click submit until the time score appear, but usually I'm already sleeping when the score appear)",
      "votes": null
    },
    {
      "id": "1658708",
      "postDate": "01/21/2022 08:44:38",
      "content": "<p>Kaggle API provides function for Leaderboard and Submission. I use Leaderboard (competition_leaderboard_view) to get information about \"submissionDate\" (this is actually DateTime). On picture you can see Submission.</p>",
      "rawMarkdown": "Kaggle API provides function for Leaderboard and Submission. I use Leaderboard (competition_leaderboard_view) to get information about \"submissionDate\" (this is actually DateTime). On picture you can see Submission.",
      "votes": null
    },
    {
      "id": "1658742",
      "postDate": "01/21/2022 09:09:36",
      "content": "<p>If I want to know how long my submission run, any way to do it instead of timing it manually?<br>\nI'm searching here : <a href=\"https://www.kaggle.com/docs/api#interacting-with-notebooks\" target=\"_blank\">https://www.kaggle.com/docs/api#interacting-with-notebooks</a><br>\nI can't find any function that returns submission time length. </p>",
      "rawMarkdown": "If I want to know how long my submission run, any way to do it instead of timing it manually?\nI'm searching here : https://www.kaggle.com/docs/api#interacting-with-notebooks\nI can't find any function that returns submission time length.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1657509,
      "author_name": "maxvandijck",
      "author_url": "",
      "post_date": "01/20/2022 08:39:30",
      "content": "<p>It will be interesting to see the winners of TF performance prizes. Most common fast models such as YOLO are all generally trained and implemented in PyTorch, I wonder if any teams have recreated these models in TF just for eligibility…</p>",
      "votes": null,
      "replies": [
        {
          "id": 1657590,
          "author_name": "locbaop",
          "author_url": "",
          "post_date": "01/20/2022 10:06:05",
          "content": "<p>I agree, i'm still looking for TF implementations.👀</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1657592,
          "author_name": "remekkinas",
          "author_url": "",
          "post_date": "01/20/2022 10:06:36",
          "content": "<p>You can export pytorch model to tf … using onnx (<a href=\"https://github.com/onnx/onnx\" target=\"_blank\">https://github.com/onnx/onnx</a>) :) and as I understand rules it is ok 😄</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1657595,
          "author_name": "philippsinger",
          "author_url": "",
          "post_date": "01/20/2022 10:08:12",
          "content": "<p>No it is not okay, it needs to be trained also in tensorflow.</p>\n<blockquote>\n  <p>Use TensorFlow 2.X on GPU (training and inference)</p>\n</blockquote>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1657599,
          "author_name": "remekkinas",
          "author_url": "",
          "post_date": "01/20/2022 10:10:45",
          "content": "<p>OK. Sorry for my short cut :) </p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1657790,
          "author_name": "something4kag",
          "author_url": "",
          "post_date": "01/20/2022 13:17:27",
          "content": "<p>so no TPU is allowed for training?  that's unfortunate considering it is TensorFlow and TPU on kaggle gives some options.  </p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1657841,
          "author_name": "dragonzhang",
          "author_url": "",
          "post_date": "01/20/2022 13:56:53",
          "content": "<p>based on public notebooks,  TF implementations underperformance compared to PyTorch  model</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1657573,
      "author_name": "imeintanis",
      "author_url": "",
      "post_date": "01/20/2022 09:48:27",
      "content": "<p>Wow! I didn't know that via kaggle API you have access to ALL teams submissions times </p>",
      "votes": null,
      "replies": [
        {
          "id": 1657583,
          "author_name": "remekkinas",
          "author_url": "",
          "post_date": "01/20/2022 09:58:24",
          "content": "<p>Kaggle API has access to leaderboard data - data which you can see on Leaderboard page. This is not a secret (data is avaliable and we can use it for analysis). This is competition so … strategy, tactics … ability to use forum, public solutions, leaderboard etc. is important :) </p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1657609,
          "author_name": "seshurajup",
          "author_url": "",
          "post_date": "01/20/2022 10:18:48",
          "content": "<p><code>To collect best solution time I should look for team LB jump</code></p>\n<p>like the idea :) Thanks for sharing !</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1657617,
          "author_name": "remekkinas",
          "author_url": "",
          "post_date": "01/20/2022 10:26:08",
          "content": "<p>I do not have time implement this (since we have object detection competition … not LB competition analysis) … but … collecting one data is enough to answer on many questions.   </p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1658648,
      "author_name": "nyleve",
      "author_url": "",
      "post_date": "01/21/2022 07:42:38",
      "content": "<p>Is there information about submission time given in our submit page ? I can only find the timing for running public set. <br>\nHow did you get the submission time for hidden set ? </p>",
      "votes": null,
      "replies": [
        {
          "id": 1658653,
          "author_name": "maxvandijck",
          "author_url": "",
          "post_date": "01/21/2022 07:47:21",
          "content": "<p>Your submission runs on both the public and private sets when you submit.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1658695,
          "author_name": "nyleve",
          "author_url": "",
          "post_date": "01/21/2022 08:32:28",
          "content": "<p>Sorry, I think I don't make myself clear.  When we run the kernel, we only get 3 rows submission. The time shown in the kernel is this one.<br>\nAfter we press submit, it runs for a long time before we get to see the score (This is the one which has 9 hour limit).  Is it possible to get this timing , or this is obtained by guessing based on the 3 rows ? <br>\nI have the feeling <a href=\"https://www.kaggle.com/remekkinas\" target=\"_blank\">@remekkinas</a>  knows this number in exact</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1658699,
          "author_name": "nyleve",
          "author_url": "",
          "post_date": "01/21/2022 08:38:31",
          "content": "<p><a href=\"https://ibb.co/2PT2vsW\"><img src=\"https://i.ibb.co/82H3rdN/runtime.jpg\" alt=\"runtime\"></a></p>\n<p>34 seconds run. No info about the submission time  (I am measuring the time I click submit until the time score appear, but usually I'm already sleeping when the score appear)</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1658708,
          "author_name": "remekkinas",
          "author_url": "",
          "post_date": "01/21/2022 08:44:38",
          "content": "<p>Kaggle API provides function for Leaderboard and Submission. I use Leaderboard (competition_leaderboard_view) to get information about \"submissionDate\" (this is actually DateTime). On picture you can see Submission.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1658742,
          "author_name": "nyleve",
          "author_url": "",
          "post_date": "01/21/2022 09:09:36",
          "content": "<p>If I want to know how long my submission run, any way to do it instead of timing it manually?<br>\nI'm searching here : <a href=\"https://www.kaggle.com/docs/api#interacting-with-notebooks\" target=\"_blank\">https://www.kaggle.com/docs/api#interacting-with-notebooks</a><br>\nI can't find any function that returns submission time length. </p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1657505": "Recently, we had a lot of questions about whether 9h is enough time to run the competition notebook. Is Kaggle GPU enough to win the competition? Should I scale to huge resolutions ( it will have a significant impact on the speed of the solution)? \n\nI analyzed the times of the TOP20 solutions (TOP20 LB team submit time). For this, I wrote a Python script (using Kaggle API), ran it and ... went drinking coffee waiting for .... results. Of course, I will not give you exact information about the team's results (due to respect for their work and I am not giving shortcuts).\n\nAssumption: \n- This stats do not show the best solution time. I assume that people submit solution looking for improvement so time is correlated. To collect best solution time I sould look for team LB jump.\n- data was collected on 2022.01.19\n\n**TOP5 (GREEN zone / paid zone)**\n- minimum time: 130 min (2h 10min)\n- maximum time: 464 min (7h 44min)\n- avg time: 239 min (3h 59min)\n\n**TOP6-13 (GOLD ZONE / excluding TOP5)**\n- minimum time: 55 min\n- maximum time: 460 min (7h 40min)\n- avg time: 213 min (3h 33 min)\n\n**TOP20 (ALL SOLUTIONS)**\n- minimum time: 49 min\n- maximum time: 464 min (7h 44min)\n- avg time: 211 min (3h 31min)\n\nAs we can see there are different solutions: \n- from very light (49 min) to .... time demanding (7h 44min) ... -> think about current solution you use (Yolo5/YoloX/YoloR/FasterRCNN ...., model size ... nano/S/M/L, inference techniques - WBF/TTA/ blending (NMS)/Tracking etc. -> this influence on speed),\n- you do not need worry about 9h limit - to be in GOLD area you need only … or .. less then 60 min (for sure less then 4h), \n- you do not need many GPUs to be in TOP10 ... (I am almost sure that fast speed does not require many GPUs during training as well - these solution are just effect of thinking ... - respect for those people),\n- I suppose part of teams (TOP20) are going to take prize in two categories (general and speed) (this is my hypothesis only - not fact) - TF is required (is Yolo5 all you need?)\n\nThis post is to motivate people who gave up having a thought in their head - is it require to have 10xGPU and I need over 9h to win ...",
    "1657509": "It will be interesting to see the winners of TF performance prizes. Most common fast models such as YOLO are all generally trained and implemented in PyTorch, I wonder if any teams have recreated these models in TF just for eligibility...",
    "1657573": "Wow! I didn't know that via kaggle API you have access to ALL teams submissions times",
    "1657583": "Kaggle API has access to leaderboard data - data which you can see on Leaderboard page. This is not a secret (data is avaliable and we can use it for analysis). This is competition so ... strategy, tactics ... ability to use forum, public solutions, leaderboard etc. is important :)",
    "1657590": "I agree, i'm still looking for TF implementations.👀",
    "1657592": "You can export pytorch model to tf ... using onnx (https://github.com/onnx/onnx) :) and as I understand rules it is ok 😄",
    "1657595": "No it is not okay, it needs to be trained also in tensorflow.\n\n> Use TensorFlow 2.X on GPU (training and inference)",
    "1657599": "OK. Sorry for my short cut :)",
    "1657609": "`To collect best solution time I should look for team LB jump`\n\nlike the idea :) Thanks for sharing !",
    "1657617": "I do not have time implement this (since we have object detection competition ... not LB competition analysis) ... but ... collecting one data is enough to answer on many questions.",
    "1657790": "so no TPU is allowed for training?  that's unfortunate considering it is TensorFlow and TPU on kaggle gives some options.",
    "1657841": "based on public notebooks,  TF implementations underperformance compared to PyTorch  model",
    "1658648": "Is there information about submission time given in our submit page ? I can only find the timing for running public set. \nHow did you get the submission time for hidden set ?",
    "1658653": "Your submission runs on both the public and private sets when you submit.",
    "1658695": "Sorry, I think I don't make myself clear.  When we run the kernel, we only get 3 rows submission. The time shown in the kernel is this one.\nAfter we press submit, it runs for a long time before we get to see the score (This is the one which has 9 hour limit).  Is it possible to get this timing , or this is obtained by guessing based on the 3 rows ? \nI have the feeling @remekkinas  knows this number in exact",
    "1658699": "<a href=\"https://ibb.co/2PT2vsW\"><img src=\"https://i.ibb.co/82H3rdN/runtime.jpg\" alt=\"runtime\" border=\"0\" /></a>\n\n34 seconds run. No info about the submission time  (I am measuring the time I click submit until the time score appear, but usually I'm already sleeping when the score appear)",
    "1658708": "Kaggle API provides function for Leaderboard and Submission. I use Leaderboard (competition_leaderboard_view) to get information about \"submissionDate\" (this is actually DateTime). On picture you can see Submission.",
    "1658742": "If I want to know how long my submission run, any way to do it instead of timing it manually?\nI'm searching here : https://www.kaggle.com/docs/api#interacting-with-notebooks\nI can't find any function that returns submission time length."
  },
  "source": "meta"
}