{
  "id": 240029,
  "title": "Batched Beam Search in TensorFlow?",
  "url": "/competitions/bms-molecular-translation/discussion/240029",
  "author_name": "",
  "post_date": "2021-05-18T10:58:26.479008500Z",
  "votes": 4,
  "comment_count": 4,
  "views": 0,
  "content": "<p>Has anyone managed to implement Batched Beam search that's compilable in a TensorFlow Graph. I can't seem to do it.</p>",
  "messages": [
    {
      "id": "1312975",
      "postDate": "05/18/2021 10:58:26",
      "content": "<p>Has anyone managed to implement Batched Beam search that's compilable in a TensorFlow Graph. I can't seem to do it.</p>",
      "rawMarkdown": "Has anyone managed to implement Batched Beam search that's compilable in a TensorFlow Graph. I can't seem to do it.",
      "votes": null
    },
    {
      "id": "1313003",
      "postDate": "05/18/2021 11:12:46",
      "content": "<p>You may try to integrate <a href=\"https://www.tensorflow.org/addons/api_docs/python/tfa/seq2seq/BeamSearchDecoder\" target=\"_blank\">this</a> in your decoder</p>\n<p>A Colab tutorial in the conext of NMT is <a href=\"https://colab.research.google.com/github/tensorflow/addons/blob/master/docs/tutorials/networks_seq2seq_nmt.ipynb\" target=\"_blank\">here</a> . The beam search is done batch wise</p>",
      "rawMarkdown": "You may try to integrate [this](https://www.tensorflow.org/addons/api_docs/python/tfa/seq2seq/BeamSearchDecoder) in your decoder\n\nA Colab tutorial in the conext of NMT is [here](https://colab.research.google.com/github/tensorflow/addons/blob/master/docs/tutorials/networks_seq2seq_nmt.ipynb) . The beam search is done batch wise",
      "votes": null
    },
    {
      "id": "1313044",
      "postDate": "05/18/2021 11:35:08",
      "content": "<p>Does it work for transformers or only LSTMS?</p>",
      "rawMarkdown": "Does it work for transformers or only LSTMS?",
      "votes": null
    },
    {
      "id": "1313077",
      "postDate": "05/18/2021 11:52:45",
      "content": "<p>Yes !  It's a beam seach wrapper.  You have to integrate it in your decoder class.  But the documentation is clearly poor.</p>",
      "rawMarkdown": "Yes !  It's a beam seach wrapper.  You have to integrate it in your decoder class.  But the documentation is clearly poor.",
      "votes": null
    },
    {
      "id": "1317936",
      "postDate": "05/21/2021 20:09:57",
      "content": "<p>Our team is working on it. It’s a lower priority task for us. If we get it implemented we will open source it.</p>",
      "rawMarkdown": "Our team is working on it. It’s a lower priority task for us. If we get it implemented we will open source it.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1313003,
      "author_name": "serigne",
      "author_url": "",
      "post_date": "05/18/2021 11:12:46",
      "content": "<p>You may try to integrate <a href=\"https://www.tensorflow.org/addons/api_docs/python/tfa/seq2seq/BeamSearchDecoder\" target=\"_blank\">this</a> in your decoder</p>\n<p>A Colab tutorial in the conext of NMT is <a href=\"https://colab.research.google.com/github/tensorflow/addons/blob/master/docs/tutorials/networks_seq2seq_nmt.ipynb\" target=\"_blank\">here</a> . The beam search is done batch wise</p>",
      "votes": null,
      "replies": [
        {
          "id": 1313044,
          "author_name": "andrewshao05",
          "author_url": "",
          "post_date": "05/18/2021 11:35:08",
          "content": "<p>Does it work for transformers or only LSTMS?</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1313077,
          "author_name": "serigne",
          "author_url": "",
          "post_date": "05/18/2021 11:52:45",
          "content": "<p>Yes !  It's a beam seach wrapper.  You have to integrate it in your decoder class.  But the documentation is clearly poor.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1317936,
      "author_name": "dschettler8845",
      "author_url": "",
      "post_date": "05/21/2021 20:09:57",
      "content": "<p>Our team is working on it. It’s a lower priority task for us. If we get it implemented we will open source it.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1312975": "Has anyone managed to implement Batched Beam search that's compilable in a TensorFlow Graph. I can't seem to do it.",
    "1313003": "You may try to integrate [this](https://www.tensorflow.org/addons/api_docs/python/tfa/seq2seq/BeamSearchDecoder) in your decoder\n\nA Colab tutorial in the conext of NMT is [here](https://colab.research.google.com/github/tensorflow/addons/blob/master/docs/tutorials/networks_seq2seq_nmt.ipynb) . The beam search is done batch wise",
    "1313044": "Does it work for transformers or only LSTMS?",
    "1313077": "Yes !  It's a beam seach wrapper.  You have to integrate it in your decoder class.  But the documentation is clearly poor.",
    "1317936": "Our team is working on it. It’s a lower priority task for us. If we get it implemented we will open source it."
  },
  "source": "meta"
}