{
  "id": 130034,
  "title": "Neural networks ? ",
  "url": "/competitions/cat-in-the-dat-ii/discussion/130034",
  "author_name": "",
  "post_date": "2020-02-11T19:38:31.593405800Z",
  "votes": 2,
  "comment_count": 24,
  "views": 0,
  "content": "<p>It seems that NN are the real way to go and, moreover, do not use Radam, nor Adam or what you want as «&nbsp;Adam familly&nbsp;»...</p>\n\n<p>All boosting algos are away from any solution...</p>\n\n<p>I have to dig more imputing and tell you then.</p>\n\n<p>Hard competition for 0€/$... just fun to compete ;-)</p>",
  "messages": [
    {
      "id": "743099",
      "postDate": "02/11/2020 19:38:31",
      "content": "<p>It seems that NN are the real way to go and, moreover, do not use Radam, nor Adam or what you want as «&nbsp;Adam familly&nbsp;»...</p>\n\n<p>All boosting algos are away from any solution...</p>\n\n<p>I have to dig more imputing and tell you then.</p>\n\n<p>Hard competition for 0€/$... just fun to compete ;-)</p>",
      "rawMarkdown": "It seems that NN are the real way to go and, moreover, do not use Radam, nor Adam or what you want as «&nbsp;Adam familly&nbsp;»...\n\nAll boosting algos are away from any solution...\n\nI have to dig more imputing and tell you then.\n\nHard competition for 0€/$... just fun to compete ;-)",
      "votes": null
    },
    {
      "id": "743210",
      "postDate": "02/11/2020 22:09:47",
      "content": "<p>Thank you!👍 </p>",
      "rawMarkdown": "Thank you!👍",
      "votes": null
    },
    {
      "id": "743256",
      "postDate": "02/11/2020 23:55:29",
      "content": "<p>Which NN? Scikit, Tensorflow ? And yes, difficult competition, indeed, (but therefore challenging). By the way, for me, until now, CatBoost worked the best, with various imputation strategies.  </p>",
      "rawMarkdown": "Which NN? Scikit, Tensorflow ? And yes, difficult competition, indeed, (but therefore challenging). By the way, for me, until now, CatBoost worked the best, with various imputation strategies.",
      "votes": null
    },
    {
      "id": "743283",
      "postDate": "02/12/2020 00:38:35",
      "content": "<p>Really? </p>",
      "rawMarkdown": "Really?",
      "votes": null
    },
    {
      "id": "743286",
      "postDate": "02/12/2020 00:41:14",
      "content": "<p>Well, just a guess.</p>",
      "rawMarkdown": "Well, just a guess.",
      "votes": null
    },
    {
      "id": "744260",
      "postDate": "02/12/2020 17:28:11",
      "content": "<p>I agree. Embeddings are especially good for categorical variables. Too bad they take forever to run.</p>",
      "rawMarkdown": "I agree. Embeddings are especially good for categorical variables. Too bad they take forever to run.",
      "votes": null
    },
    {
      "id": "744354",
      "postDate": "02/12/2020 19:04:04",
      "content": "<p>I would apologize for RAdam, it depends so much on data prep...</p>\n\n<p>Struggling with R Mice package that is a kind of never ending story.... 3 hours right now, and not finished yet.</p>\n\n<p>I’ll tell you more tomorrow or next week ;-)</p>",
      "rawMarkdown": "I would apologize for RAdam, it depends so much on data prep...\n\nStruggling with R Mice package that is a kind of never ending story.... 3 hours right now, and not finished yet.\n\nI’ll tell you more tomorrow or next week ;-)",
      "votes": null
    },
    {
      "id": "744356",
      "postDate": "02/12/2020 19:06:46",
      "content": "<p><a href=\"/catadanna\">@catadanna</a> </p>\n\n<p>Keras just Keras with Tensorflow.</p>",
      "rawMarkdown": "catadanna \n\nKeras just Keras with Tensorflow.",
      "votes": null
    },
    {
      "id": "744359",
      "postDate": "02/12/2020 19:07:38",
      "content": "<p><a href=\"/kailaicui\">@kailaicui</a> </p>\n\n<p>You’re Welcome !</p>",
      "rawMarkdown": "kailaicui \n\nYou’re Welcome !",
      "votes": null
    },
    {
      "id": "744367",
      "postDate": "02/12/2020 19:11:53",
      "content": "<p>I used MLPC of Scikit. In fact I tested several kernels of Scikit + CatBoost , just looped on them + Stratified Kfold crossed with different imputations. Did not try Tf yet. I worked a lot for data preparation and imputing. </p>",
      "rawMarkdown": "I used MLPC of Scikit. In fact I tested several kernels of Scikit + CatBoost , just looped on them + Stratified Kfold crossed with different imputations. Did not try Tf yet. I worked a lot for data preparation and imputing.",
      "votes": null
    },
    {
      "id": "744374",
      "postDate": "02/12/2020 19:18:16",
      "content": "<p><a href=\"/catadanna\">@catadanna</a> </p>\n\n<p>I do a lot on prep too... still waiting for hours right now with R mice... I have to find something quicker.</p>\n\n<p>My PC will burn :-(</p>",
      "rawMarkdown": "catadanna \n\nI do a lot on prep too... still waiting for hours right now with R mice... I have to find something quicker.\n\nMy PC will burn :-(",
      "votes": null
    },
    {
      "id": "744385",
      "postDate": "02/12/2020 19:32:50",
      "content": "<p>@mezoganet : I am sure you did a lot of prep! That is the key of a successful ML program generally. You work in R? I am in Python. I tried MICE for imputation but it took several hours and still running so I had to stop and find another way. And I spent a lot of time in encoding ... </p>",
      "rawMarkdown": "mezoganet : I am sure you did a lot of prep! That is the key of a successful ML program generally. You work in R? I am in Python. I tried MICE for imputation but it took several hours and still running so I had to stop and find another way. And I spent a lot of time in encoding ...",
      "votes": null
    },
    {
      "id": "744533",
      "postDate": "02/12/2020 23:10:29",
      "content": "<p><a href=\"/catadanna\">@catadanna</a> </p>\n\n<p>I use both Python and R.</p>\n\n<p>Generally I prepare data with R and the rest in Python, but I have to admit that R is my favorite.</p>\n\n<p>There are so many great books about Python that you have to know Python.</p>\n\n<p>FYI, in late 90´s, I discovered Python 1.5 (yes 1.5 !) and then swapped to Perl because of books available on Perl at that time.</p>\n\n<p>A computer language is not a problem when you know what you wan to do, what we call algorithm, don’t you think ?</p>",
      "rawMarkdown": "catadanna \n\nI use both Python and R.\n\nGenerally I prepare data with R and the rest in Python, but I have to admit that R is my favorite.\n\nThere are so many great books about Python that you have to know Python.\n\nFYI, in late 90´s, I discovered Python 1.5 (yes 1.5 !) and then swapped to Perl because of books available on Perl at that time.\n\nA computer language is not a problem when you know what you wan to do, what we call algorithm, don’t you think ?",
      "votes": null
    },
    {
      "id": "744542",
      "postDate": "02/12/2020 23:24:03",
      "content": "<p><a href=\"/catadanna\">@catadanna</a> </p>\n\n<p>I am trying to understand the meaning of nom_5...nom_9 features... I think there is something to dig from them.</p>",
      "rawMarkdown": "catadanna \n\nI am trying to understand the meaning of nom_5...nom_9 features... I think there is something to dig from them.",
      "votes": null
    },
    {
      "id": "744625",
      "postDate": "02/13/2020 02:05:40",
      "content": "<p>@mezoganet It seems they are hexadecimal numbers but I suppose we are to deal with them as strings. I tried to deal with them as numbers but had no better result.</p>",
      "rawMarkdown": "mezoganet It seems they are hexadecimal numbers but I suppose we are to deal with them as strings. I tried to deal with them as numbers but had no better result.",
      "votes": null
    },
    {
      "id": "744696",
      "postDate": "02/13/2020 03:29:38",
      "content": "<p>Yes. I tried also. No help ,so I gave up.</p>",
      "rawMarkdown": "Yes. I tried also. No help ,so I gave up.",
      "votes": null
    },
    {
      "id": "744758",
      "postDate": "02/13/2020 05:42:01",
      "content": "<p>@mezoganet I agree with a part of what you say. Yes the algorithm is important, but what is also important is the way it is implemented in different languages. For example Python has implementations for CNN and libraries which do not exist in most languages. And different implementations of NN may provide different results. </p>\n\n<p>By the way, yes I used to work in PERL 10 years ago, when I had to implement algos from scratch ;-p . PERL is very robust especially for text parsing (NLP). Now I prefer Python and its libs (Keras, Tf, &amp;co)</p>",
      "rawMarkdown": "mezoganet I agree with a part of what you say. Yes the algorithm is important, but what is also important is the way it is implemented in different languages. For example Python has implementations for CNN and libraries which do not exist in most languages. And different implementations of NN may provide different results. \n\nBy the way, yes I used to work in PERL 10 years ago, when I had to implement algos from scratch ;-p . PERL is very robust especially for text parsing (NLP). Now I prefer Python and its libs (Keras, Tf, &amp;co)",
      "votes": null
    },
    {
      "id": "745220",
      "postDate": "02/13/2020 16:12:16",
      "content": "<p>Yes NN seems the way. I'm using adam and also not doing any complex imputation. That's something I want to try as well but it's taking forever!\nDoes it change a lot for you if you use adam?</p>",
      "rawMarkdown": "Yes NN seems the way. I'm using adam and also not doing any complex imputation. That's something I want to try as well but it's taking forever!\nDoes it change a lot for you if you use adam?",
      "votes": null
    },
    {
      "id": "745274",
      "postDate": "02/13/2020 16:59:06",
      "content": "<p>See this one,score 0.78657\n<a href=\"https://www.kaggle.com/pavelvpster/cat-in-dat-2-stack\">https://www.kaggle.com/pavelvpster/cat-in-dat-2-stack</a></p>",
      "rawMarkdown": "See this one,score 0.78657\nhttps://www.kaggle.com/pavelvpster/cat-in-dat-2-stack",
      "votes": null
    },
    {
      "id": "745339",
      "postDate": "02/13/2020 18:14:53",
      "content": "<p>Apparently treating features from <code>nom_5</code> to <code>nom_9</code> by decomposing each string as a sequence of chars--as in \n<code>\nfor i in range(9):\n    col_name = 'nom_5_{}'.format(i)\n    train_features[col_name] = train_features['nom_5'].apply(lambda x: x[i] if type(x) is str else np.nan)\n    test_features[col_name] = test_features['nom_5'].apply(lambda x: x[i] if type(x) is str else np.nan)\ntrain_features.drop(columns=['nom_5'], inplace=True)\ntest_features.drop(columns=['nom_5'], inplace=True)\n</code>\ngets hilariously low score.</p>",
      "rawMarkdown": "Apparently treating features from `nom_5` to `nom_9` by decomposing each string as a sequence of chars--as in \n```\nfor i in range(9):\n    col_name = 'nom_5_{}'.format(i)\n    train_features[col_name] = train_features['nom_5'].apply(lambda x: x[i] if type(x) is str else np.nan)\n    test_features[col_name] = test_features['nom_5'].apply(lambda x: x[i] if type(x) is str else np.nan)\ntrain_features.drop(columns=['nom_5'], inplace=True)\ntest_features.drop(columns=['nom_5'], inplace=True)\n```\ngets hilariously low score.",
      "votes": null
    },
    {
      "id": "745726",
      "postDate": "02/14/2020 06:01:31",
      "content": "<p>Thanks, I came across it already. But it doesn't really answer the questions I had and what <a href=\"/pfichou\">@pfichou</a> was talking about.</p>",
      "rawMarkdown": "Thanks, I came across it already. But it doesn't really answer the questions I had and what @pfichou was talking about.",
      "votes": null
    },
    {
      "id": "746143",
      "postDate": "02/14/2020 17:15:04",
      "content": "<p><a href=\"/advaithasabnis\">@advaithasabnis</a> </p>\n\n<p>Not Adam, but RAdam... not exactly the same.</p>\n\n<p>I have had some issues today with Conda vs pip conflicts and had to fix some of them.</p>\n\n<p>I will be back soon  radam and lookahead, I really hope.</p>",
      "rawMarkdown": "advaithasabnis \n\nNot Adam, but RAdam... not exactly the same.\n\nI have had some issues today with Conda vs pip conflicts and had to fix some of them.\n\nI will be back soon  radam and lookahead, I really hope.",
      "votes": null
    },
    {
      "id": "746144",
      "postDate": "02/14/2020 17:16:55",
      "content": "<p>MICE with  R = 20 hours lost with no result...</p>",
      "rawMarkdown": "MICE with  R = 20 hours lost with no result...",
      "votes": null
    },
    {
      "id": "746148",
      "postDate": "02/14/2020 17:18:45",
      "content": "<p>I did not wait 20 hours, I stopped after one hour or less ... </p>",
      "rawMarkdown": "I did not wait 20 hours, I stopped after one hour or less ...",
      "votes": null
    },
    {
      "id": "746220",
      "postDate": "02/14/2020 18:58:03",
      "content": "<p><a href=\"/catadanna\">@catadanna</a> </p>\n\n<p>Age gives you more patience, I guess... wait for some years ;-)</p>",
      "rawMarkdown": "catadanna \n\nAge gives you more patience, I guess... wait for some years ;-)",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 743210,
      "author_name": "kailai",
      "author_url": "",
      "post_date": "02/11/2020 22:09:47",
      "content": "<p>Thank you!👍 </p>",
      "votes": null,
      "replies": [
        {
          "id": 744359,
          "author_name": "pfichou",
          "author_url": "",
          "post_date": "02/12/2020 19:07:38",
          "content": "<p><a href=\"/kailaicui\">@kailaicui</a> </p>\n\n<p>You’re Welcome !</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 743256,
      "author_name": "catadanna",
      "author_url": "",
      "post_date": "02/11/2020 23:55:29",
      "content": "<p>Which NN? Scikit, Tensorflow ? And yes, difficult competition, indeed, (but therefore challenging). By the way, for me, until now, CatBoost worked the best, with various imputation strategies.  </p>",
      "votes": null,
      "replies": [
        {
          "id": 743283,
          "author_name": "kailai",
          "author_url": "",
          "post_date": "02/12/2020 00:38:35",
          "content": "<p>Really? </p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 743286,
          "author_name": "catadanna",
          "author_url": "",
          "post_date": "02/12/2020 00:41:14",
          "content": "<p>Well, just a guess.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 744356,
          "author_name": "pfichou",
          "author_url": "",
          "post_date": "02/12/2020 19:06:46",
          "content": "<p><a href=\"/catadanna\">@catadanna</a> </p>\n\n<p>Keras just Keras with Tensorflow.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 744367,
          "author_name": "catadanna",
          "author_url": "",
          "post_date": "02/12/2020 19:11:53",
          "content": "<p>I used MLPC of Scikit. In fact I tested several kernels of Scikit + CatBoost , just looped on them + Stratified Kfold crossed with different imputations. Did not try Tf yet. I worked a lot for data preparation and imputing. </p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 744374,
          "author_name": "pfichou",
          "author_url": "",
          "post_date": "02/12/2020 19:18:16",
          "content": "<p><a href=\"/catadanna\">@catadanna</a> </p>\n\n<p>I do a lot on prep too... still waiting for hours right now with R mice... I have to find something quicker.</p>\n\n<p>My PC will burn :-(</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 744385,
          "author_name": "catadanna",
          "author_url": "",
          "post_date": "02/12/2020 19:32:50",
          "content": "<p>@mezoganet : I am sure you did a lot of prep! That is the key of a successful ML program generally. You work in R? I am in Python. I tried MICE for imputation but it took several hours and still running so I had to stop and find another way. And I spent a lot of time in encoding ... </p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 744533,
          "author_name": "pfichou",
          "author_url": "",
          "post_date": "02/12/2020 23:10:29",
          "content": "<p><a href=\"/catadanna\">@catadanna</a> </p>\n\n<p>I use both Python and R.</p>\n\n<p>Generally I prepare data with R and the rest in Python, but I have to admit that R is my favorite.</p>\n\n<p>There are so many great books about Python that you have to know Python.</p>\n\n<p>FYI, in late 90´s, I discovered Python 1.5 (yes 1.5 !) and then swapped to Perl because of books available on Perl at that time.</p>\n\n<p>A computer language is not a problem when you know what you wan to do, what we call algorithm, don’t you think ?</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 744542,
          "author_name": "pfichou",
          "author_url": "",
          "post_date": "02/12/2020 23:24:03",
          "content": "<p><a href=\"/catadanna\">@catadanna</a> </p>\n\n<p>I am trying to understand the meaning of nom_5...nom_9 features... I think there is something to dig from them.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 744625,
          "author_name": "catadanna",
          "author_url": "",
          "post_date": "02/13/2020 02:05:40",
          "content": "<p>@mezoganet It seems they are hexadecimal numbers but I suppose we are to deal with them as strings. I tried to deal with them as numbers but had no better result.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 744696,
          "author_name": "kailai",
          "author_url": "",
          "post_date": "02/13/2020 03:29:38",
          "content": "<p>Yes. I tried also. No help ,so I gave up.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 744758,
          "author_name": "catadanna",
          "author_url": "",
          "post_date": "02/13/2020 05:42:01",
          "content": "<p>@mezoganet I agree with a part of what you say. Yes the algorithm is important, but what is also important is the way it is implemented in different languages. For example Python has implementations for CNN and libraries which do not exist in most languages. And different implementations of NN may provide different results. </p>\n\n<p>By the way, yes I used to work in PERL 10 years ago, when I had to implement algos from scratch ;-p . PERL is very robust especially for text parsing (NLP). Now I prefer Python and its libs (Keras, Tf, &amp;co)</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 745339,
          "author_name": "davidbnn92",
          "author_url": "",
          "post_date": "02/13/2020 18:14:53",
          "content": "<p>Apparently treating features from <code>nom_5</code> to <code>nom_9</code> by decomposing each string as a sequence of chars--as in \n<code>\nfor i in range(9):\n    col_name = 'nom_5_{}'.format(i)\n    train_features[col_name] = train_features['nom_5'].apply(lambda x: x[i] if type(x) is str else np.nan)\n    test_features[col_name] = test_features['nom_5'].apply(lambda x: x[i] if type(x) is str else np.nan)\ntrain_features.drop(columns=['nom_5'], inplace=True)\ntest_features.drop(columns=['nom_5'], inplace=True)\n</code>\ngets hilariously low score.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 744260,
      "author_name": "davidbnn92",
      "author_url": "",
      "post_date": "02/12/2020 17:28:11",
      "content": "<p>I agree. Embeddings are especially good for categorical variables. Too bad they take forever to run.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 744354,
      "author_name": "pfichou",
      "author_url": "",
      "post_date": "02/12/2020 19:04:04",
      "content": "<p>I would apologize for RAdam, it depends so much on data prep...</p>\n\n<p>Struggling with R Mice package that is a kind of never ending story.... 3 hours right now, and not finished yet.</p>\n\n<p>I’ll tell you more tomorrow or next week ;-)</p>",
      "votes": null,
      "replies": [
        {
          "id": 746144,
          "author_name": "pfichou",
          "author_url": "",
          "post_date": "02/14/2020 17:16:55",
          "content": "<p>MICE with  R = 20 hours lost with no result...</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 746148,
          "author_name": "catadanna",
          "author_url": "",
          "post_date": "02/14/2020 17:18:45",
          "content": "<p>I did not wait 20 hours, I stopped after one hour or less ... </p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 746220,
          "author_name": "pfichou",
          "author_url": "",
          "post_date": "02/14/2020 18:58:03",
          "content": "<p><a href=\"/catadanna\">@catadanna</a> </p>\n\n<p>Age gives you more patience, I guess... wait for some years ;-)</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 745220,
      "author_name": "advaithasabnis",
      "author_url": "",
      "post_date": "02/13/2020 16:12:16",
      "content": "<p>Yes NN seems the way. I'm using adam and also not doing any complex imputation. That's something I want to try as well but it's taking forever!\nDoes it change a lot for you if you use adam?</p>",
      "votes": null,
      "replies": [
        {
          "id": 745274,
          "author_name": "kailai",
          "author_url": "",
          "post_date": "02/13/2020 16:59:06",
          "content": "<p>See this one,score 0.78657\n<a href=\"https://www.kaggle.com/pavelvpster/cat-in-dat-2-stack\">https://www.kaggle.com/pavelvpster/cat-in-dat-2-stack</a></p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 745726,
          "author_name": "advaithasabnis",
          "author_url": "",
          "post_date": "02/14/2020 06:01:31",
          "content": "<p>Thanks, I came across it already. But it doesn't really answer the questions I had and what <a href=\"/pfichou\">@pfichou</a> was talking about.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 746143,
          "author_name": "pfichou",
          "author_url": "",
          "post_date": "02/14/2020 17:15:04",
          "content": "<p><a href=\"/advaithasabnis\">@advaithasabnis</a> </p>\n\n<p>Not Adam, but RAdam... not exactly the same.</p>\n\n<p>I have had some issues today with Conda vs pip conflicts and had to fix some of them.</p>\n\n<p>I will be back soon  radam and lookahead, I really hope.</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "743099": "It seems that NN are the real way to go and, moreover, do not use Radam, nor Adam or what you want as «&nbsp;Adam familly&nbsp;»...\n\nAll boosting algos are away from any solution...\n\nI have to dig more imputing and tell you then.\n\nHard competition for 0€/$... just fun to compete ;-)",
    "743210": "Thank you!👍",
    "743256": "Which NN? Scikit, Tensorflow ? And yes, difficult competition, indeed, (but therefore challenging). By the way, for me, until now, CatBoost worked the best, with various imputation strategies.",
    "743283": "Really?",
    "743286": "Well, just a guess.",
    "744260": "I agree. Embeddings are especially good for categorical variables. Too bad they take forever to run.",
    "744354": "I would apologize for RAdam, it depends so much on data prep...\n\nStruggling with R Mice package that is a kind of never ending story.... 3 hours right now, and not finished yet.\n\nI’ll tell you more tomorrow or next week ;-)",
    "744356": "catadanna \n\nKeras just Keras with Tensorflow.",
    "744359": "kailaicui \n\nYou’re Welcome !",
    "744367": "I used MLPC of Scikit. In fact I tested several kernels of Scikit + CatBoost , just looped on them + Stratified Kfold crossed with different imputations. Did not try Tf yet. I worked a lot for data preparation and imputing.",
    "744374": "catadanna \n\nI do a lot on prep too... still waiting for hours right now with R mice... I have to find something quicker.\n\nMy PC will burn :-(",
    "744385": "mezoganet : I am sure you did a lot of prep! That is the key of a successful ML program generally. You work in R? I am in Python. I tried MICE for imputation but it took several hours and still running so I had to stop and find another way. And I spent a lot of time in encoding ...",
    "744533": "catadanna \n\nI use both Python and R.\n\nGenerally I prepare data with R and the rest in Python, but I have to admit that R is my favorite.\n\nThere are so many great books about Python that you have to know Python.\n\nFYI, in late 90´s, I discovered Python 1.5 (yes 1.5 !) and then swapped to Perl because of books available on Perl at that time.\n\nA computer language is not a problem when you know what you wan to do, what we call algorithm, don’t you think ?",
    "744542": "catadanna \n\nI am trying to understand the meaning of nom_5...nom_9 features... I think there is something to dig from them.",
    "744625": "mezoganet It seems they are hexadecimal numbers but I suppose we are to deal with them as strings. I tried to deal with them as numbers but had no better result.",
    "744696": "Yes. I tried also. No help ,so I gave up.",
    "744758": "mezoganet I agree with a part of what you say. Yes the algorithm is important, but what is also important is the way it is implemented in different languages. For example Python has implementations for CNN and libraries which do not exist in most languages. And different implementations of NN may provide different results. \n\nBy the way, yes I used to work in PERL 10 years ago, when I had to implement algos from scratch ;-p . PERL is very robust especially for text parsing (NLP). Now I prefer Python and its libs (Keras, Tf, &amp;co)",
    "745220": "Yes NN seems the way. I'm using adam and also not doing any complex imputation. That's something I want to try as well but it's taking forever!\nDoes it change a lot for you if you use adam?",
    "745274": "See this one,score 0.78657\nhttps://www.kaggle.com/pavelvpster/cat-in-dat-2-stack",
    "745339": "Apparently treating features from `nom_5` to `nom_9` by decomposing each string as a sequence of chars--as in \n```\nfor i in range(9):\n    col_name = 'nom_5_{}'.format(i)\n    train_features[col_name] = train_features['nom_5'].apply(lambda x: x[i] if type(x) is str else np.nan)\n    test_features[col_name] = test_features['nom_5'].apply(lambda x: x[i] if type(x) is str else np.nan)\ntrain_features.drop(columns=['nom_5'], inplace=True)\ntest_features.drop(columns=['nom_5'], inplace=True)\n```\ngets hilariously low score.",
    "745726": "Thanks, I came across it already. But it doesn't really answer the questions I had and what @pfichou was talking about.",
    "746143": "advaithasabnis \n\nNot Adam, but RAdam... not exactly the same.\n\nI have had some issues today with Conda vs pip conflicts and had to fix some of them.\n\nI will be back soon  radam and lookahead, I really hope.",
    "746144": "MICE with  R = 20 hours lost with no result...",
    "746148": "I did not wait 20 hours, I stopped after one hour or less ...",
    "746220": "catadanna \n\nAge gives you more patience, I guess... wait for some years ;-)"
  },
  "source": "meta"
}