{
  "id": 97756,
  "title": "18th place solution ",
  "url": "/competitions/freesound-audio-tagging-2019/discussion/97756",
  "author_name": "Ilia Larchenko",
  "post_date": "2019-06-28T18:23:48.251000",
  "votes": 3,
  "comment_count": 0,
  "views": 0,
  "content": "<p>A concise description of my solution:\n-   Curated subset only\n-   Data preparation: raw data -&gt; augmentation (random shift + random noise) -&gt; 4 versions of augmented spectrograms for every sample \n-   Model: 10 conv layers with skip connections -&gt; 2 fully connected layers\n-   Augmentations: zoom, random crop, lighting, warp, cutout \n-   Training using fast.ai\n-   Inference: a blend of 24 models (trained on different random subsets) with TTA30</p>",
  "messages": [
    {
      "id": 563842,
      "postDate": "2019-06-28T18:23:48.250Z",
      "content": "<p>A concise description of my solution:\n-   Curated subset only\n-   Data preparation: raw data -&gt; augmentation (random shift + random noise) -&gt; 4 versions of augmented spectrograms for every sample \n-   Model: 10 conv layers with skip connections -&gt; 2 fully connected layers\n-   Augmentations: zoom, random crop, lighting, warp, cutout \n-   Training using fast.ai\n-   Inference: a blend of 24 models (trained on different random subsets) with TTA30</p>",
      "rawMarkdown": "A concise description of my solution:\n-\tCurated subset only\n-\tData preparation: raw data -&gt; augmentation (random shift + random noise) -&gt; 4 versions of augmented spectrograms for every sample \n-\tModel: 10 conv layers with skip connections -&gt; 2 fully connected layers\n-\tAugmentations: zoom, random crop, lighting, warp, cutout \n-\tTraining using fast.ai\n-\tInference: a blend of 24 models (trained on different random subsets) with TTA30",
      "votes": 3
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "563842": "A concise description of my solution:\n-\tCurated subset only\n-\tData preparation: raw data -&gt; augmentation (random shift + random noise) -&gt; 4 versions of augmented spectrograms for every sample \n-\tModel: 10 conv layers with skip connections -&gt; 2 fully connected layers\n-\tAugmentations: zoom, random crop, lighting, warp, cutout \n-\tTraining using fast.ai\n-\tInference: a blend of 24 models (trained on different random subsets) with TTA30"
  }
}