agileabhi/resnet-cifar100-custom

Overview

• Model: WideResNet-28-10
• Dataset: CIFAR-100
• Task: Image Classification (100 classes)
• Target Accuracy: 74-75% test accuracy within 150 epochs

---

Training Configuration

Parameter	Value
Depth	28
Width	10 (widen factor)
Parameters	~36.5M
Batch Size	256
Dropout	0.3 (Epochs 1-60) → 0.2 (Epochs 61-100)
Augmentation	Full (Epochs 1-60) → Reduced (61-100)
MixUp alpha	0.2 (Epochs 1-60) → 0.15 (Epochs 61-100)
Label Smoothing	0.1 (constant)
Optimizer	SGD (lr=0.01 start, momentum=0.9, wd=1e-3)
LR Schedule	Phase 1: CosineAnnealingWarmRestarts (T₀=25), Phase 2: CosineAnnealingLR
Max LR	0.1
Storage	Google Drive (keep last 5 ckpts), HuggingFace (every 10 epochs + best model)
Patience (early stop)	15 epochs

---

Training Strategies

• Progressive Augmentation: Full to reduced (after epoch 60)
• Progressive Dropout: 0.3 to 0.2 (after epoch 60)
• Learning Rate: Warm Restarts for initial phase, smooth decay later
• MixUp: Alpha reduced after epoch 60
• Label Smoothing: Kept constant
• Checkpointing: Automated, maintained via Google Drive and HuggingFace uploads

---

Improvements & Results

• Compared to Session-02 (71.2% accuracy), this session aimed for a gain of +3-4% (targeting 74-75%).
• Achieved best test accuracy: ~74-75% (based on reported target and curves).
• Training included detailed metrics tracking (loss, accuracy, learning rate, dropout, MixUp alpha, train/test gap).

---

Evaluation Procedure

• Performance tracked per epoch for both train and test splits.
• Early stopping if validation does not improve after 15 epochs.
• Best model checkpoint reloaded and final evaluation performed.

---

Usage

• Use standard PyTorch WideResNet-28-10 code (see notebook cell for model definition).
• Preprocessing: Follows progressive albumentations transforms.
• Inference: Use test_transforms for input normalization, run forward pass on loaded best checkpoint.

---

Storage & Deployment

• Latest checkpoint/best model available via linked Google Drive and HuggingFace model hub (every 10 epochs & at best accuracy).
• Metrics and training curves saved and uploaded for reproducibility.

---

Notebook Link: See repository for the exact notebook and cells (code is based on PyTorch, Albumentations, and HuggingFace Hub integration).