Experiments

Complete guide to running experiments on PatchCamelyon and CAMELYON16 datasets.

PatchCamelyon Benchmark

The PatchCamelyon (PCam) dataset consists of 96×96 pixel patches extracted from histopathologic scans of lymph node sections.

Generate Synthetic Data

python scripts/generate_synthetic_pcam.py

Train Model

python experiments/train_pcam.py --config experiments/configs/pcam.yaml

Evaluate

python experiments/evaluate_pcam.py \
  --checkpoint checkpoints/pcam/best_model.pth \
  --data-root data/pcam \
  --output-dir results/pcam

Results

Synthetic subset (8 epochs):

• Test Accuracy: 94.0%
• Test AUC: 1.0
• Training Time: ~40 seconds (CPU)

See PCAM_BENCHMARK_RESULTS.html for detailed analysis.

CAMELYON16 Slide-Level Classification

Slide-level classification using pre-extracted patch features with attention-based aggregation.

Generate Synthetic Data

python scripts/generate_synthetic_camelyon.py

Train Slide Classifier

# Quick test (1 epoch)
python experiments/train_camelyon.py \
  --config experiments/configs/camelyon_quick_test.yaml

# Full training (50 epochs)
python experiments/train_camelyon.py \
  --config experiments/configs/camelyon.yaml

Evaluate with CSV Export

python experiments/evaluate_camelyon.py \
  --checkpoint checkpoints/camelyon/best_model.pth \
  --data-root data/camelyon \
  --output-dir results/camelyon \
  --save-predictions-csv

Features

• Multiple aggregation strategies (mean, max pooling)
• Variable-length slide handling with masking
• CSV export for downstream analysis
• ROC curves and confusion matrices

See CAMELYON_TRAINING_STATUS.html for detailed guide.

Pretrained Model Integration

Load and fine-tune pretrained encoders:

from src.models.pretrained import load_pretrained_encoder

# Load ResNet50 with ImageNet weights
encoder = load_pretrained_encoder(
    model_name='resnet50',
    source='torchvision',
    pretrained=True,
    num_classes=2
)

# Access feature dimension
feature_dim = encoder.feature_dim  # 2048 for ResNet50

Supported Sources:

• torchvision: ResNet, DenseNet, EfficientNet, VGG, MobileNet
• timm: 1000+ models including Vision Transformers, ConvNeXt, Swin

Model Profiling

Analyze model performance and export for deployment:

# Profile inference latency
python scripts/model_profiler.py \
  --checkpoint models/best_model.pth \
  --profile-type time

# Export to ONNX format
python scripts/export_onnx.py \
  --checkpoint models/best_model.pth \
  --output models/model.onnx

Baseline Comparisons

Compare multiple model architectures:

# Quick test (3 epochs)
python experiments/compare_pcam_baselines.py \
  --configs experiments/configs/pcam_comparison/*.yaml \
  --quick-test

# Full training
python experiments/compare_pcam_baselines.py \
  --configs experiments/configs/pcam_comparison/*.yaml

See PCAM_COMPARISON_GUIDE.html for detailed comparison guide.

Configuration

Example training configuration:

# experiments/configs/pcam.yaml
data:
  root_dir: "data/pcam"
  batch_size: 32
  num_workers: 4

model:
  architecture: "resnet18"
  num_classes: 2
  dropout: 0.5

training:
  epochs: 10
  learning_rate: 0.001
  weight_decay: 0.0001
  optimizer: "adam"

Next Steps

• Getting Started Guide - Installation and setup
• API Reference - Complete API documentation
• Architecture Guide - System design details
• Documentation Index - All documentation