{
  "id": 326112,
  "title": "Trained Model - 3+ hours scoring time",
  "url": "/competitions/birdclef-2022/discussion/326112",
  "author_name": "",
  "post_date": "2022-05-20T09:08:28.286873900Z",
  "votes": 1,
  "comment_count": 1,
  "views": 0,
  "content": "<p>Hi,</p>\n<p>It seems like many people have scorings that take 1-2 hours, but my current scoring attempt is taking over 4 hours. </p>\n<p>I'm using a model that I've trained on an outside computer so the pipeline is relatively straightforward (simplified):  read test_soundscapes -&gt; glob and sort *.ogg files -&gt; create tf.Dataset and convert ogg files to spectrograms via tf.stft -&gt; load model -&gt; input into the model by model.predict(BATCH_SIZE). I've checked, tested and converted the predictions into the form as given by test_submission.csv</p>\n<p>There are no complex procedures, \"run all\" takes the notebook around 40s</p>\n<p>Does anyone have any general advice or is this expected?</p>",
  "messages": [
    {
      "id": "1795929",
      "postDate": "05/20/2022 09:08:28",
      "content": "<p>Hi,</p>\n<p>It seems like many people have scorings that take 1-2 hours, but my current scoring attempt is taking over 4 hours. </p>\n<p>I'm using a model that I've trained on an outside computer so the pipeline is relatively straightforward (simplified):  read test_soundscapes -&gt; glob and sort *.ogg files -&gt; create tf.Dataset and convert ogg files to spectrograms via tf.stft -&gt; load model -&gt; input into the model by model.predict(BATCH_SIZE). I've checked, tested and converted the predictions into the form as given by test_submission.csv</p>\n<p>There are no complex procedures, \"run all\" takes the notebook around 40s</p>\n<p>Does anyone have any general advice or is this expected?</p>",
      "rawMarkdown": "Hi,\n\nIt seems like many people have scorings that take 1-2 hours, but my current scoring attempt is taking over 4 hours. \n\nI'm using a model that I've trained on an outside computer so the pipeline is relatively straightforward (simplified):  read test_soundscapes -> glob and sort *.ogg files -> create tf.Dataset and convert ogg files to spectrograms via tf.stft -> load model -> input into the model by model.predict(BATCH_SIZE). I've checked, tested and converted the predictions into the form as given by test_submission.csv\n\nThere are no complex procedures, \"run all\" takes the notebook around 40s\n\nDoes anyone have any general advice or is this expected?",
      "votes": null
    },
    {
      "id": "1795988",
      "postDate": "05/20/2022 10:39:52",
      "content": "<p>Ran it again with GPU toggled on - success by 2 hours in. Model was overfitted and most likely learned which audio chunk belonged to which file 😂</p>",
      "rawMarkdown": "Ran it again with GPU toggled on - success by 2 hours in. Model was overfitted and most likely learned which audio chunk belonged to which file 😂",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1795988,
      "author_name": "kevenr",
      "author_url": "",
      "post_date": "05/20/2022 10:39:52",
      "content": "<p>Ran it again with GPU toggled on - success by 2 hours in. Model was overfitted and most likely learned which audio chunk belonged to which file 😂</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1795929": "Hi,\n\nIt seems like many people have scorings that take 1-2 hours, but my current scoring attempt is taking over 4 hours. \n\nI'm using a model that I've trained on an outside computer so the pipeline is relatively straightforward (simplified):  read test_soundscapes -> glob and sort *.ogg files -> create tf.Dataset and convert ogg files to spectrograms via tf.stft -> load model -> input into the model by model.predict(BATCH_SIZE). I've checked, tested and converted the predictions into the form as given by test_submission.csv\n\nThere are no complex procedures, \"run all\" takes the notebook around 40s\n\nDoes anyone have any general advice or is this expected?",
    "1795988": "Ran it again with GPU toggled on - success by 2 hours in. Model was overfitted and most likely learned which audio chunk belonged to which file 😂"
  },
  "source": "meta"
}