/Dockship-Attentive-Ai-1st-place-Solution

My approach for the 1st place solution in Attentive AI hiring challenge

Primary LanguageJupyter Notebook

Note:

  1. The dataset (tfrecords)[299x299] can be found here
  2. Original images are here

Approach:

  1. Converted Original data to TFRecords, for using the Kaggle's TPU using this
  2. All images were scaled to 299x299 considering majority of the images had similar resolution.
  3. Model used EfficientNet-B4 (noisy student weights) freezed BatchNorm layers.
  4. Augmentations used:
    • Horizontal Fips
    • Shearing
    • H/W Zoom
    • H/W Shift
  5. Custom Cosine Annealed Learning Schedule
  6. 5-Fold Cross Validation pipeline.
  7. Class weights to combat class-imbalance.

Will be making the kaggle notebook public soon!