This model is built on the EfficientNet-B4 backbone with a custom classifier head, trained using a novel 6-Phase Progressive Training strategy. The training progressively increases image resolution (280px → 400px) and augmentation difficulty (None → MixUp → CutMix → Hybrid).

Developed by: fth2745
Model type: Image Classification (CNN)
License: MIT
Finetuned from: EfficientNet-B4 (ImageNet pretrained)

Performance

Metric	Test Set	Validation Set
Top-1 Accuracy	94.49%	97.45%
Top-3 Accuracy	97.61%	98.82%
Top-5 Accuracy	98.49%	99.31%
Macro F1-Score	94.75%	97.13%

Training Details

Training Data

Oxford Flowers-102 dataset with offline data augmentation (tier-based augmentation for class balancing).

Training Procedure

6-Phase Progressive Training

Phase	Epochs	Resolution	Augmentation	Dropout
1. Basic	1-5	280×280	Basic Preprocessing	0.4
2. MixUp Soft	6-10	320×320	MixUp α=0.2	0.2
3. MixUp Hard	11-15	320×320	MixUp α=0.4	0.2
4. CutMix Soft	16-20	380×380	CutMix α=0.2	0.2
5. CutMix Hard	21-30	380×380	CutMix α=0.5	0.2
6. Grand Finale	31-40	400×400	Hybrid	0.2

Preprocessing

Resize → RandomCrop → HorizontalFlip → Rotation (±20°) → Affine → ColorJitter → Normalize (ImageNet)

Training Hyperparameters

Optimizer: AdamW
Learning Rate: 1e-3
Weight Decay: 1e-4
Scheduler: CosineAnnealingWarmRestarts (T_0=5, T_mult=2)
Loss: CrossEntropyLoss (label_smoothing=0.1)
Batch Size: 8
Training Regime: fp16 mixed precision (AMP)

🎯 6-Phase Progressive Training

Phase 1 ──→ Phase 2 ──→ Phase 3 ──→ Phase 4 ──→ Phase 5 ──→ Phase 6
 280px      320px       320px       380px       380px       400px
  None      MixUp       MixUp      CutMix      CutMix      Hybrid
           α=0.2       α=0.4       α=0.2       α=0.5      MixUp+Cut

Phase Details

Phase	Epochs	Resolution	Technique	Alpha	Dropout	Purpose
1️⃣ Basic	1-5	280×280	Basic Preprocessing	-	0.4	Learn fundamental features
2️⃣ MixUp Soft	6-10	320×320	MixUp	0.2	0.2	Gentle texture blending
3️⃣ MixUp Hard	11-15	320×320	MixUp	0.4	0.2	Strong texture mixing
4️⃣ CutMix Soft	16-20	380×380	CutMix	0.2	0.2	Learn partial structures
5️⃣ CutMix Hard	21-30	380×380	CutMix	0.5	0.2	Handle occlusions
6️⃣ Grand Finale	31-40	400×400	Hybrid	0.1-0.3	0.2	Final polish with both

💡 Why Progressive Training? Starting with low resolution helps the model learn general shapes first. Gradual augmentation increase builds robustness incrementally.

🖼️ Preprocessing Pipeline (All Phases)

⚠️ Note: These preprocessing steps are applied in ALL PHASES. Only img_size changes per phase.

Complete Training Flow

┌─────────────────────────────────────────────────────────────┐
│              📷 RAW IMAGE INPUT                             │
└─────────────────────────────────────────────────────────────┘
                              ↓
┌─────────────────────────────────────────────────────────────┐
│     🔄 STEP 1: IMAGE-LEVEL PREPROCESSING (Per image)        │
├─────────────────────────────────────────────────────────────┤
│  1️⃣ Resize         │ (img_size + 32) × (img_size + 32)     │
│  2️⃣ RandomCrop     │ img_size × img_size                   │
│  3️⃣ HorizontalFlip │ p=0.5                                 │
│  4️⃣ RandomRotation │ ±20°                                  │
│  5️⃣ RandomAffine   │ scale=(0.8, 1.2)                      │
│  6️⃣ ColorJitter    │ brightness, contrast, saturation=0.2  │
│  7️⃣ ToTensor       │ [0-255] → [0.0-1.0]                   │
│  8️⃣ Normalize      │ ImageNet mean/std                     │
└─────────────────────────────────────────────────────────────┘
                              ↓
┌─────────────────────────────────────────────────────────────┐
│        🎲 STEP 2: BATCH-LEVEL AUGMENTATION (Phase-specific) │
├─────────────────────────────────────────────────────────────┤
│  Phase 1: None (Preprocessing only)                         │
│  Phase 2-3: MixUp (λ×ImageA + (1-λ)×ImageB)                 │
│  Phase 4-5: CutMix (Patch swap between images)              │
│  Phase 6: Hybrid (MixUp + CutMix combined)                  │
└─────────────────────────────────────────────────────────────┘
                              ↓
┌─────────────────────────────────────────────────────────────┐
│              🎯 READY FOR MODEL TRAINING                    │
└─────────────────────────────────────────────────────────────┘

Phase-Specific Image Sizes

Phase	img_size	Resize To	RandomCrop To
1️⃣ Basic	280	312×312	280×280
2️⃣ MixUp Soft	320	352×352	320×320
3️⃣ MixUp Hard	320	352×352	320×320
4️⃣ CutMix Soft	380	412×412	380×380
5️⃣ CutMix Hard	380	412×412	380×380
6️⃣ Grand Finale	400	432×432	400×400

Preprocessing Details (All Phases)

Step	Transform	Parameters	Purpose
1️⃣	Resize	(size+32, size+32)	Prepare for random crop
2️⃣	RandomCrop	(size, size)	Random position augmentation
3️⃣	RandomHorizontalFlip	p=0.5	Left-right invariance
4️⃣	RandomRotation	degrees=20	Rotation invariance
5️⃣	RandomAffine	scale=(0.8, 1.2)	Scale variation
6️⃣	ColorJitter	(0.2, 0.2, 0.2)	Brightness/Contrast/Saturation
7️⃣	ToTensor	-	Convert to PyTorch tensor
8️⃣	Normalize	mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]	ImageNet normalization

Test/Validation Preprocessing

Step	Transform	Parameters
1️⃣	Resize	(size, size)
2️⃣	ToTensor	-
3️⃣	Normalize	ImageNet mean/std

💡 Key Insight: Preprocessing (8 steps) is applied per image in every phase. MixUp/CutMix is applied AFTER preprocessing as batch-level augmentation.

🔀 Batch-Level Augmentation Techniques (Phase-Specific)

MixUp

Image A (Rose) + Image B (Sunflower) 
    ↓
λ = Beta(α, α)  →  New Image = λ×A + (1-λ)×B
    ↓
Blended Image (70% Rose + 30% Sunflower features)

Benefits: ✅ Smoother decision boundaries ✅ Reduces overconfidence ✅ Better generalization

CutMix

Image A (Rose) + Random BBox from Image B (Sunflower)
    ↓
Paste B's region onto A
    ↓
Composite Image (Rose background + Sunflower patch)

Benefits: ✅ Object completion ability ✅ Occlusion robustness ✅ Localization skills

Hybrid (Grand Finale)

Apply MixUp (blend two images)
Apply CutMix (cut on blended image)
Result: Maximum augmentation challenge

🛡️ Smart Training Features

Two-Layer Early Stopping

Layer	Condition	Patience	Action
Phase-level	Train↓ + Val↑ (Overfitting)	2 epochs	Skip to next phase
Global	Val loss not improving	8 epochs	Stop training

Smart Dropout Mechanism

Signal	Condition	Action
⚠️ Overfitting	Train↓ + Val↑	Dropout += 0.05
🚑 Underfitting	Train↑ + Val↑	Dropout -= 0.05
✅ Normal	Train↓ + Val↓	No change

Bounds: min=0.10, max=0.50

Model Architecture

EfficientNet-B4 (pretrained)
    └── Custom Classifier Head
        ├── BatchNorm1d (1792)
        ├── Dropout
        ├── Linear (1792 → 512)
        ├── GELU
        ├── BatchNorm1d (512)
        ├── Dropout
        └── Linear (512 → 102)

Total Parameters: ~19M (all trainable)

Supported Flower Classes

102 flower species including: Rose, Sunflower, Tulip, Orchid, Lily, Daisy, Hibiscus, Lotus, Magnolia, and 93 more.

Limitations

Trained only on Oxford Flowers-102 dataset
Best performance at 400×400 resolution
May not generalize well to flowers outside the 102 trained classes

Citation

bibtex

@misc{efficientnet-b4-flowers102,
  title={EfficientNet-B4 Flower Classifier with 6-Phase Progressive Training},
  author={fth2745},
  year={2024},
  url={https://huggingface.co/fth2745/efficientnet-b4-flowers102}
}