Kaggle Bengali.AI 6 th place solution

Higher resolution model pipeline picture

Transcript

1. Bengali.AI
   Handwritten Grapheme Classification
   ইঁদুর এবং ভালুক
   

   View Slide

2. Train Images
   - 200,840 images
   - 168 / 11 / 7 classes
   - 137 x 236 x 1
   Test Images
   - About 200,000 images
   - 137 x 236 x 1
   - 4 parquets
   Copyright 2020 @ Maxwell_110
   Bengali Model Pipeline
   Customized SE-ResNet 50
   - `NOT` pretrained
   - Iterative Stratified 5 folds
   - 137 x 236 x 1 Input Image Size
   - Images divided by 255
   - Adam
   - 3 Stage Learning
   1. 13 - 18 CyclicLR (Triangle, 8 epoch/cycle, 1e-3~-4), Xentropy
   2. LRonP (36 epoch, 3 pat, 6 ES, 1e-5), Reduced Focal Loss
   3. LRonP (36 epoch, 3 pat, 6 ES, 1e-5), Xentropy
   - Batch Size: 64
   - Augmenatation:
   Width Shift (20%), Erosion, CutOut (holes=8, 3 types of Size),
   GridMask (rotate=15deg)
   - CutMixUp
   CutMix (p=1/3, alpha=0.5) / MixUp (p=1/3, alpha=0.2)
   Inception ResNet V2
   - Pretrained
   - Iterative Stratified 5 folds ( same folds as Maxwell )
   - 180 x 180 x 3 Input Image Size
   - Images divided by 255
   - Adam
   - 2 Stage Cyclic Learning (4 epoch/cycle, 5e-5 ~ 2e-3)
   1. Weighted Reduced Focal Loss, 40 epoch,
   weight = 1 / observation counts
   2. LRonPlateau (1e-5), Xentropy
   - Batch Size: 64
   - Augmenatation:
   Rotate (8deg), Zoom (1.2), height/width shift (15%),
   CutOut (holes=20, max_h=25, max_w=40)
   - MixUp
   MixUp (p=1, alpha=0.4)
   137
   236
   137
   236
   Prediction 1
   (Maxwell)
   Prediction 2
   (Nejumi)
   * 0.75 +
   - Resize to (180, 180)
   * 0.25 = Blended
   Prediction
   Public : 0.9871 / 78 th
   Private : 0.9557 / 6 th
   Submission with Post Processing
   - Multiply coefficients to each predicted probability
   - Coefficients for each label (186 types)
   - Use NM solver to calculate optimal coeffcients
   grapheme_root : [c_g_1, c_g_2, ..., c_g_168]
   vowel_diacritic : [c_v_1, c_v_2, ... , c_v_11]
   consonant_diacritic : [c_c_1, c_c_2, ... , c_c_7]
   Apply correction coefficients
   to blended predictions
   Inference Limitations:
   - Inference on kernel
   - GPU inference time <= 2 hours
   - Memory Limit <= 13 GB
   Our Resources
   Maxwell: TITAN RTX, Geforce 1080Ti x 2, GCP
   Nejumi: TITAN RTX, Geforce 1080Ti x 2,
   Vast.ai ( https://vast.ai/ )*1
   *1 Nejumi is cloud addict
   Prediction 1
   CV(w/ pp)
   : 0.9888
   Public : 0.9864
   Private : 0.9527 (12th)
   Input (1ch)
   GeM 2D
   512 fc
   Grapheme
   168 nodes
   Vowel
   11 nodes
   Consonant
   7 nodes
   SoftMax
   + +
   +
   Add
   1280 fc 512 fc
   512 fc 512 fc
   Customized
   SE-ResNet 50
   Block
   ( 3 x 3 bottom kernel )
   Inception ResNet V2
   Block
   Input (1ch)
   Input (3ch)
   Prediction 2
   CV (w/ pp)
   : 0.9845
   Public : 0.9810
   Private : 0.9449 (19th)
   Convert to
   3 ch image
   Separable Conv 2D
   BN
   ReLu
   GAP 2D
   Grapheme
   168 nodes
   Vowel
   11 nodes
   Consonant
   7 nodes
   SoftMax
   180
   180
   180
   180
   

   View Slide

3. View Slide