{
  "id": 314156,
  "title": "How good are you doing with Tensorflow Keras ?",
  "url": "/competitions/birdclef-2022/discussion/314156",
  "author_name": "",
  "post_date": "2022-03-21T09:26:15.822486100Z",
  "votes": 16,
  "comment_count": 6,
  "views": 0,
  "content": "<p>Hi Everyone, I hope you are doing well. Like past BirdCall competitions it seems that pytorch users are outperforming tensorflow users. So in this year I'd like to know how far we can push performances with tensorflow. As far as I'm concerned my best performace is achieved with TF (a simple Conv2D model) with <strong>LB=0.54</strong>. if you're willing just share your best achievable score with pytorch and more specifically with tensorflow.</p>\n<p><strong>Update</strong>: I just reached <strong>LB=0.60</strong> with tensorflow by changing some parameters of my whole pipeline.</p>",
  "messages": [
    {
      "id": "1730456",
      "postDate": "03/21/2022 09:26:15",
      "content": "<p>Hi Everyone, I hope you are doing well. Like past BirdCall competitions it seems that pytorch users are outperforming tensorflow users. So in this year I'd like to know how far we can push performances with tensorflow. As far as I'm concerned my best performace is achieved with TF (a simple Conv2D model) with <strong>LB=0.54</strong>. if you're willing just share your best achievable score with pytorch and more specifically with tensorflow.</p>\n<p><strong>Update</strong>: I just reached <strong>LB=0.60</strong> with tensorflow by changing some parameters of my whole pipeline.</p>",
      "rawMarkdown": "Hi Everyone, I hope you are doing well. Like past BirdCall competitions it seems that pytorch users are outperforming tensorflow users. So in this year I'd like to know how far we can push performances with tensorflow. As far as I'm concerned my best performace is achieved with TF (a simple Conv2D model) with **LB=0.54**. if you're willing just share your best achievable score with pytorch and more specifically with tensorflow.\n\n**Update**: I just reached **LB=0.60** with tensorflow by changing some parameters of my whole pipeline.",
      "votes": null
    },
    {
      "id": "1730663",
      "postDate": "03/21/2022 14:04:52",
      "content": "<p>With TensorFlow/Keras I'm at <strong>0.67</strong> currently (could probably get it up to 0.75 if all goes well… but we shall see)</p>\n<ul>\n<li>Fairly complicated pipeline</li>\n<li>No augmentation or TTA</li>\n<li>EfficientNetB0 model architecture(s) w/ a single Dense layer</li>\n</ul>",
      "rawMarkdown": "With TensorFlow/Keras I'm at **0.67** currently (could probably get it up to 0.75 if all goes well... but we shall see)\n* Fairly complicated pipeline\n* No augmentation or TTA\n* EfficientNetB0 model architecture(s) w/ a single Dense layer",
      "votes": null
    },
    {
      "id": "1730732",
      "postDate": "03/21/2022 15:10:24",
      "content": "<p>Wow that's impressive <a href=\"https://www.kaggle.com/dschettler8845\" target=\"_blank\">@dschettler8845</a> . My dummy CNN is better than my EfficientNetB0 which is predicting a single bird. Probably there is an error in my pipeline or i'm using poor parameters to fit my network. But it encourages me to keep pushing with tensorflow.</p>",
      "rawMarkdown": "Wow that's impressive @dschettler8845 . My dummy CNN is better than my EfficientNetB0 which is predicting a single bird. Probably there is an error in my pipeline or i'm using poor parameters to fit my network. But it encourages me to keep pushing with tensorflow.",
      "votes": null
    },
    {
      "id": "1730759",
      "postDate": "03/21/2022 15:32:37",
      "content": "<p>Glad to hear! I want to clean things up a bit more (I'm really focusing on readability and cleanliness) as this competition is a bit more complicated (training, scoring etc.). But then I will opensource a small MVP version to show how everything works from E2E. :) I'll update here if I crack 0.7 today.</p>",
      "rawMarkdown": "Glad to hear! I want to clean things up a bit more (I'm really focusing on readability and cleanliness) as this competition is a bit more complicated (training, scoring etc.). But then I will opensource a small MVP version to show how everything works from E2E. :) I'll update here if I crack 0.7 today.",
      "votes": null
    },
    {
      "id": "1730787",
      "postDate": "03/21/2022 16:00:56",
      "content": "<p>Nice! Look forward to it.</p>",
      "rawMarkdown": "Nice! Look forward to it.",
      "votes": null
    },
    {
      "id": "1760790",
      "postDate": "04/19/2022 14:15:30",
      "content": "<p>I did not succeed to have a decent TF pipeline. With same architecture the results are much more worse with my tensorflow approach (I am trying to use a fully TF pipeline including for the generation of spectrogram on the fly)</p>",
      "rawMarkdown": "I did not succeed to have a decent TF pipeline. With same architecture the results are much more worse with my tensorflow approach (I am trying to use a fully TF pipeline including for the generation of spectrogram on the fly)",
      "votes": null
    },
    {
      "id": "1761808",
      "postDate": "04/20/2022 07:36:17",
      "content": "<p>It is the same for me, the pipeline that scores 0.60 with TF scores 0.73 with pytorch. I'm puzzled by that. I will try tuning again the gradient policy and see if TF can matcj pytorch.</p>",
      "rawMarkdown": "It is the same for me, the pipeline that scores 0.60 with TF scores 0.73 with pytorch. I'm puzzled by that. I will try tuning again the gradient policy and see if TF can matcj pytorch.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1730663,
      "author_name": "dschettler8845",
      "author_url": "",
      "post_date": "03/21/2022 14:04:52",
      "content": "<p>With TensorFlow/Keras I'm at <strong>0.67</strong> currently (could probably get it up to 0.75 if all goes well… but we shall see)</p>\n<ul>\n<li>Fairly complicated pipeline</li>\n<li>No augmentation or TTA</li>\n<li>EfficientNetB0 model architecture(s) w/ a single Dense layer</li>\n</ul>",
      "votes": null,
      "replies": [
        {
          "id": 1730732,
          "author_name": "ulrich07",
          "author_url": "",
          "post_date": "03/21/2022 15:10:24",
          "content": "<p>Wow that's impressive <a href=\"https://www.kaggle.com/dschettler8845\" target=\"_blank\">@dschettler8845</a> . My dummy CNN is better than my EfficientNetB0 which is predicting a single bird. Probably there is an error in my pipeline or i'm using poor parameters to fit my network. But it encourages me to keep pushing with tensorflow.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1730759,
          "author_name": "dschettler8845",
          "author_url": "",
          "post_date": "03/21/2022 15:32:37",
          "content": "<p>Glad to hear! I want to clean things up a bit more (I'm really focusing on readability and cleanliness) as this competition is a bit more complicated (training, scoring etc.). But then I will opensource a small MVP version to show how everything works from E2E. :) I'll update here if I crack 0.7 today.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1730787,
          "author_name": "ulrich07",
          "author_url": "",
          "post_date": "03/21/2022 16:00:56",
          "content": "<p>Nice! Look forward to it.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1760790,
      "author_name": "ludovick",
      "author_url": "",
      "post_date": "04/19/2022 14:15:30",
      "content": "<p>I did not succeed to have a decent TF pipeline. With same architecture the results are much more worse with my tensorflow approach (I am trying to use a fully TF pipeline including for the generation of spectrogram on the fly)</p>",
      "votes": null,
      "replies": [
        {
          "id": 1761808,
          "author_name": "ulrich07",
          "author_url": "",
          "post_date": "04/20/2022 07:36:17",
          "content": "<p>It is the same for me, the pipeline that scores 0.60 with TF scores 0.73 with pytorch. I'm puzzled by that. I will try tuning again the gradient policy and see if TF can matcj pytorch.</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1730456": "Hi Everyone, I hope you are doing well. Like past BirdCall competitions it seems that pytorch users are outperforming tensorflow users. So in this year I'd like to know how far we can push performances with tensorflow. As far as I'm concerned my best performace is achieved with TF (a simple Conv2D model) with **LB=0.54**. if you're willing just share your best achievable score with pytorch and more specifically with tensorflow.\n\n**Update**: I just reached **LB=0.60** with tensorflow by changing some parameters of my whole pipeline.",
    "1730663": "With TensorFlow/Keras I'm at **0.67** currently (could probably get it up to 0.75 if all goes well... but we shall see)\n* Fairly complicated pipeline\n* No augmentation or TTA\n* EfficientNetB0 model architecture(s) w/ a single Dense layer",
    "1730732": "Wow that's impressive @dschettler8845 . My dummy CNN is better than my EfficientNetB0 which is predicting a single bird. Probably there is an error in my pipeline or i'm using poor parameters to fit my network. But it encourages me to keep pushing with tensorflow.",
    "1730759": "Glad to hear! I want to clean things up a bit more (I'm really focusing on readability and cleanliness) as this competition is a bit more complicated (training, scoring etc.). But then I will opensource a small MVP version to show how everything works from E2E. :) I'll update here if I crack 0.7 today.",
    "1730787": "Nice! Look forward to it.",
    "1760790": "I did not succeed to have a decent TF pipeline. With same architecture the results are much more worse with my tensorflow approach (I am trying to use a fully TF pipeline including for the generation of spectrogram on the fly)",
    "1761808": "It is the same for me, the pipeline that scores 0.60 with TF scores 0.73 with pytorch. I'm puzzled by that. I will try tuning again the gradient policy and see if TF can matcj pytorch."
  },
  "source": "meta"
}