{
  "id": 142254,
  "title": "MultiFit - Fast.ai model for multilingual text classification",
  "url": "/competitions/jigsaw-multilingual-toxic-comment-classification/discussion/142254",
  "author_name": "",
  "post_date": "2020-04-09T16:24:31.263614800Z",
  "votes": 4,
  "comment_count": 4,
  "views": 0,
  "content": "<p>Hello everyone,</p>\n\n<p>I found this model regarding multilingual text classification. The model performs pretty well on Portuguese and French.</p>\n\n<p>You can implement it using the fast.ai library.</p>\n\n<p>MultiFit: <a href=\"https://arxiv.org/abs/1909.04761\">https://arxiv.org/abs/1909.04761</a></p>",
  "messages": [
    {
      "id": "802581",
      "postDate": "04/09/2020 16:24:31",
      "content": "<p>Hello everyone,</p>\n\n<p>I found this model regarding multilingual text classification. The model performs pretty well on Portuguese and French.</p>\n\n<p>You can implement it using the fast.ai library.</p>\n\n<p>MultiFit: <a href=\"https://arxiv.org/abs/1909.04761\">https://arxiv.org/abs/1909.04761</a></p>",
      "rawMarkdown": "Hello everyone,\n\nI found this model regarding multilingual text classification. The model performs pretty well on Portuguese and French.\n\nYou can implement it using the fast.ai library.\n\nMultiFit: [https://arxiv.org/abs/1909.04761](https://arxiv.org/abs/1909.04761)",
      "votes": null
    },
    {
      "id": "803012",
      "postDate": "04/10/2020 04:09:06",
      "content": "<p>I will try to implement. I have already implemented one with Fast.ai using \"bert-base-multilingual-uncased\" on small sample Dataset as the data don't fit in GPU memory. It will be good if Fast.ai support TPU for faster processing with whole Dataset.</p>",
      "rawMarkdown": "I will try to implement. I have already implemented one with Fast.ai using \"bert-base-multilingual-uncased\" on small sample Dataset as the data don't fit in GPU memory. It will be good if Fast.ai support TPU for faster processing with whole Dataset.",
      "votes": null
    },
    {
      "id": "803260",
      "postDate": "04/10/2020 09:50:24",
      "content": "<p>Since Fast.ai is built on top of PyTorch, can't you tweak the implementation to enable TPU support ?</p>",
      "rawMarkdown": "Since Fast.ai is built on top of PyTorch, can't you tweak the implementation to enable TPU support ?",
      "votes": null
    },
    {
      "id": "803266",
      "postDate": "04/10/2020 10:03:17",
      "content": "<p>I have started to understand the Fast.ai Learner Function. I think I need to modify it. </p>",
      "rawMarkdown": "I have started to understand the Fast.ai Learner Function. I think I need to modify it.",
      "votes": null
    },
    {
      "id": "809582",
      "postDate": "04/16/2020 10:02:19",
      "content": "<p>This looks very powerful for this problem, my (quick) reading of the paper is that the model first works hard on the central problem (toxicity detection) in the main (mono-) language, and then works hard on translation - but a reduced sort of translation that is constrained to toxicity.</p>\n\n<p>So there is a lot less ... stuff ... coming at the model, and the parameter space is greatly reduced.</p>",
      "rawMarkdown": "This looks very powerful for this problem, my (quick) reading of the paper is that the model first works hard on the central problem (toxicity detection) in the main (mono-) language, and then works hard on translation - but a reduced sort of translation that is constrained to toxicity.\n\nSo there is a lot less ... stuff ... coming at the model, and the parameter space is greatly reduced.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 803012,
      "author_name": "lucca9211",
      "author_url": "",
      "post_date": "04/10/2020 04:09:06",
      "content": "<p>I will try to implement. I have already implemented one with Fast.ai using \"bert-base-multilingual-uncased\" on small sample Dataset as the data don't fit in GPU memory. It will be good if Fast.ai support TPU for faster processing with whole Dataset.</p>",
      "votes": null,
      "replies": [
        {
          "id": 803260,
          "author_name": "rftexas",
          "author_url": "",
          "post_date": "04/10/2020 09:50:24",
          "content": "<p>Since Fast.ai is built on top of PyTorch, can't you tweak the implementation to enable TPU support ?</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 803266,
          "author_name": "lucca9211",
          "author_url": "",
          "post_date": "04/10/2020 10:03:17",
          "content": "<p>I have started to understand the Fast.ai Learner Function. I think I need to modify it. </p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 809582,
      "author_name": "iridiumblue",
      "author_url": "",
      "post_date": "04/16/2020 10:02:19",
      "content": "<p>This looks very powerful for this problem, my (quick) reading of the paper is that the model first works hard on the central problem (toxicity detection) in the main (mono-) language, and then works hard on translation - but a reduced sort of translation that is constrained to toxicity.</p>\n\n<p>So there is a lot less ... stuff ... coming at the model, and the parameter space is greatly reduced.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "802581": "Hello everyone,\n\nI found this model regarding multilingual text classification. The model performs pretty well on Portuguese and French.\n\nYou can implement it using the fast.ai library.\n\nMultiFit: [https://arxiv.org/abs/1909.04761](https://arxiv.org/abs/1909.04761)",
    "803012": "I will try to implement. I have already implemented one with Fast.ai using \"bert-base-multilingual-uncased\" on small sample Dataset as the data don't fit in GPU memory. It will be good if Fast.ai support TPU for faster processing with whole Dataset.",
    "803260": "Since Fast.ai is built on top of PyTorch, can't you tweak the implementation to enable TPU support ?",
    "803266": "I have started to understand the Fast.ai Learner Function. I think I need to modify it.",
    "809582": "This looks very powerful for this problem, my (quick) reading of the paper is that the model first works hard on the central problem (toxicity detection) in the main (mono-) language, and then works hard on translation - but a reduced sort of translation that is constrained to toxicity.\n\nSo there is a lot less ... stuff ... coming at the model, and the parameter space is greatly reduced."
  },
  "source": "meta"
}