Upgrade to Pro — share decks privately, control downloads, hide ads and more …

Great Barrier Reef Model Pipeline: 15th place

Maxwell
February 16, 2022

Great Barrier Reef Model Pipeline: 15th place

Maxwell

February 16, 2022
Tweet

More Decks by Maxwell

Other Decks in Science

Transcript

  1. Copyright 2022 Maxwell_110
    Validation strategy
    - Sequence-based 4 fold CV
    - The number of CoTS is close in each fold
    - Training data is frames with CoTs
    - Validation data includes frames w/o CoTs
    Resize up to 2.75 times
    using progressive learning
    1280
    720
    Augmentation
    Increasing probability of applying
    augmentation as progressive
    learning progresses.
    - Default YOLO-X augmentations
    - random resize: (-5, 5)
    - mosaic / MixUp / hsv / flip:
    p = 0.6 -> 0.8
    - degrees: Not used
    - translate: 0.1
    - mosaic / MixUp scale: (0.5, 1.5)
    - RandomGamma
    - RGBShift
    - Sharpen
    - GaussNoise
    Batch Size: 4
    GeForce RTX 3080 (x 2)
    Solution description in Kaggle discussion
    https://www.kaggle.com/c/tensorflow-great-barrier-reef/discussion/307691
    Learning strategy
    - Progressive learning
    - Optimizer: default SGD
    (decay: 5e-4, momentum: 0.9)
    - LR: .000625
    - Scheduler: yoloxwarmcos
    - min_lr_ratio: 0.1
    - EMA: on
    - warmup_epochs: 5
    - max_epoch: 30
    TTA Seq-NMS
    https://arxiv.org/abs/1602.08465
    https://github.com/tmoopenn/seq-nms
    n_frames: 2
    confidence threshold: 0.07
    linkage threshold: 0.1
    nms th: 0.4
    Weighted Box Fusion
    skip box threshold: 0.05
    wbf IoU threshold: 0.45
    Final confidence threshold: .08
    Public LB : 0.607
    Private LB : 0.714

    View full-size slide