Boonichi / MacroWoodClassification

Macroscopic Wood Classification

Macroscopic Wood Classification

Statement

Description

Pipeline (Developed)

• Data Split: Stratified K-Fold (num_folds = 5)
• Data Augmentation:
  • Train Dataset:
    • Resize: 224x224
    • Color_jitter : 0.4
    • Vertical flip : 0.5
    • Horizontal flip: 0.5
    • Interpolation: bicubic
    • Random Erase params:
      • reprob: 0.25
      • remode: pixel
      • recount: 1
    • RandomCrop (img_size > 32): padding = 4
    • Normalization (Mean, std): ((0.4875, 0.4194, 0.3976),(0.0332, 0.0276, 0.0245))
  • Validation Dataset:
    • Resize: 224x224
    • Interpolation: bicubic
    • Normalize (Mean, std)
• Sampler techniques:
  • Train Dataset: DistributedSampler
  • Val Dataset: SequentialSampler
• Criterion (Loss): PolyLoss, LabelSmoothingCrossEntropy, CrossEntropy
• Logging: TensorBoard, Wandb
• Callbacks: Early Stopping, LRMonitor
• Optimizers: Adamw, Nadam, radam (easy to add more options)
• Simple Models (Resnet, VGG, densenet, inception, ...)

Undeveloped

• Models: Transformer (Attentionhead)
• Processes: Finetune, predict
• Augmentation: Mixup
• Optimization: LR/Layer decay (Cosine Scheduler), ModelEMA, LossScaler
• Save Model