Smithery is now a part of Arcade.dev! Read more in our announcement here
Complete context for TAR UMT Data Science FYP implementing CrossViT for COVID-19 chest X-ray classification...
Student: Tan Ming Kai (24PMR12003)
Program: Bachelor of Computer Science (Honours) in Data Science
University: TAR UMT, Penang Branch, Malaysia
Academic Year: 2025/26
Supervisor: Angkay A/P Subramaniam
Project Title: Multi-Scale Vision Transformer (CrossViT) for COVID-19 Chest X-ray Classification Using Dual-Branch Architecture
Core Philosophy: COMPLETION OVER PERFECTION
# Exact specifications from Chapter 4
1. Load image (grayscale/RGB handling)
2. CLAHE enhancement:
- clip_limit = 2.0
- tile_grid_size = (8, 8) # Creates 64 contextual regions
3. Resize to 240×240 (CrossViT requirement)
4. Normalize to ImageNet stats:
- mean = [0.485, 0.456, 0.406]
- std = [0.229, 0.224, 0.225]
5. Data Augmentation (training only):
- Random rotation: ±10°
- Horizontal flip: 50% probability
- Translation: ±5% in both axes
- Brightness/contrast: ±10%
- NO vertical flipping (anatomically incorrect)
- NO aggressive elastic deformation
# Hardware-optimized for NVIDIA RTX 6000 Ada Generation (51GB VRAM)
batch_size = 8 # Maximum 16 with gradient accumulation
gradient_accumulation_steps = 4 # Effective batch = 32
mixed_precision = True # Use FP16 (automatic mixed precision)
optimizer = AdamW(lr=5e-5, weight_decay=0.05)
scheduler = CosineAnnealingWarmRestarts(T_0=10, T_mult=2)
criterion = nn.CrossEntropyLoss(weight=[1.47, 0.52, 0.88, 3.95]) # Class weights
max_epochs = 50
early_stopping_patience = 15 # Monitor validation loss
seed = 42 # For reproducibility
Primary Metrics:
Medical Metrics:
Statistical Validation:
H₀ (Null): No significant difference between CrossViT and CNN baselines (p≥0.05)
H₁ (Primary): CrossViT achieves significantly higher accuracy than CNN baselines (p<0.05)
H₂ (Multi-scale): Dual-branch processing improves accuracy by ≥5% vs single-scale
H₃ (CLAHE): Contrast enhancement improves performance by ≥2% vs no CLAHE
H₄ (Augmentation): Conservative augmentation improves generalization without degrading accuracy
Available Resources:
Critical Memory Tactics:
import torch
# 1. Clear cache frequently
torch.cuda.empty_cache()
# 2. Monitor memory usage
print(f"Allocated: {torch.cuda.memory_allocated()/1e9:.2f} GB")
print(f"Reserved: {torch.cuda.memory_reserved()/1e9:.2f} GB")
# 3. Use gradient checkpointing for large models
model.gradient_checkpointing_enable()
# 4. Delete unnecessary tensors
del outputs, loss
torch.cuda.empty_cache()
# 5. Use DataLoader with num_workers=4, pin_memory=True
train_loader = DataLoader(
dataset,
batch_size=8,
num_workers=4,
pin_memory=True,
persistent_workers=True # Faster epoch transitions
)
00_Environment_Setup.ipynb # Verify GPU, packages, dependencies
01_Data_Exploration.ipynb # EDA with visualizations (Chapter 4 evidence)
02_Data_Preprocessing.ipynb # CLAHE optimization, class balance
03_Data_Augmentation.ipynb # Test augmentation strategies
04_Baseline_Models.ipynb # Train all 5 baselines
05_CrossViT_Training.ipynb # Main model training
06_Results_Analysis.ipynb # Statistical tests, 95% CI, hypothesis testing
07_Ablation_Studies.ipynb # Test H2, H3, H4 hypotheses
08_Flask_Demo_Prep.ipynb # Export model for web interface
Every notebook MUST include:
"""
Notebook: XX_Name.ipynb
Purpose: [Clear single-sentence description]
Author: Tan Ming Kai (24PMR12003)
Date: 2025-XX-XX
FYP: CrossViT for COVID-19 Classification
Hardware: NVIDIA RTX 6000 Ada Generation (51GB VRAM) VRAM
Relates to: Chapter 4, Section X.X
Key Outputs: [List deliverables]
"""
# 1. REPRODUCIBILITY (ALWAYS FIRST)
import random
import numpy as np
import torch
SEED = 42
random.seed(SEED)
np.random.seed(SEED)
torch.manual_seed(SEED)
torch.cuda.manual_seed_all(SEED)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
# 2. IMPORTS (organized by category)
# Standard library
import os
from pathlib import Path
# Data science
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import seaborn as sns
# PyTorch
import torch
import torch.nn as nn
from torch.utils.data import DataLoader
# Computer Vision
import cv2
from PIL import Image
import timm
# 3. CONFIGURATION (single source of truth)
CONFIG = {
'seed': 42,
'device': 'cuda' if torch.cuda.is_available() else 'cpu',
'data_dir': '/path/to/covid19_radiography_database',
'output_dir': './outputs',
'batch_size': 8,
'num_workers': 4,
# ... all parameters here
}
# 4. HARDWARE VERIFICATION
print(f"PyTorch Version: {torch.__version__}")
print(f"CUDA Available: {torch.cuda.is_available()}")
if torch.cuda.is_available():
print(f"GPU: {torch.cuda.get_device_name(0)}")
print(f"CUDA Version: {torch.version.cuda}")
print(f"Total VRAM: {torch.cuda.get_device_properties(0).total_memory/1e9:.2f} GB")
# Set publication style
import matplotlib.pyplot as plt
import seaborn as sns
plt.style.use('seaborn-v0_8-paper')
sns.set_palette("husl")
plt.rcParams['figure.figsize'] = (10, 6)
plt.rcParams['figure.dpi'] = 300
plt.rcParams['font.size'] = 12
plt.rcParams['font.family'] = 'serif'
# Always label axes, add titles, legends
# Always save high-res versions for report
plt.savefig('figure.png', dpi=300, bbox_inches='tight')
def train_epoch(model, loader, criterion, optimizer, device):
model.train()
total_loss = 0.0
for batch_idx, (images, labels) in enumerate(loader):
# Move to device
images = images.to(device, non_blocking=True)
labels = labels.to(device, non_blocking=True)
# Forward pass with mixed precision
with torch.cuda.amp.autocast():
outputs = model(images)
loss = criterion(outputs, labels)
# Backward pass
optimizer.zero_grad(set_to_none=True) # More efficient
loss.backward()
optimizer.step()
total_loss += loss.item()
# Clear cache every N batches
if batch_idx % 10 == 0:
torch.cuda.empty_cache()
# Memory monitoring (can be commented out in production)
if batch_idx % 50 == 0:
print(f"Batch {batch_idx}: VRAM {torch.cuda.memory_allocated()/1e9:.2f} GB")
return total_loss / len(loader)
This project supports multiple UN Sustainable Development Goals:
Primary: SDG 3 (Good Health and Well-being)
Secondary: SDG 9 (Industry, Innovation, Infrastructure)
Tertiary: SDG 10 (Reduced Inequalities)
For detailed technical specifications, implementation details, and academic context:
references/technical_specs.md - Complete technical specificationsreferences/academic_context.md - Full academic background from Chapters 1-4For utility scripts:
scripts/memory_monitor.py - GPU memory tracking utilityscripts/check_dataset.py - Validate dataset integrityscripts/setup_env.py - Environment verification script# Minimal working example to verify setup
import torch
import timm
# Check GPU
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
print(f"Using device: {device}")
# Load CrossViT-Tiny
model = timm.create_model('crossvit_tiny_240', pretrained=True, num_classes=4)
model = model.to(device)
# Test forward pass
dummy_input = torch.randn(1, 3, 240, 240).to(device)
with torch.no_grad():
output = model(dummy_input)
print(f"Output shape: {output.shape}") # Should be [1, 4]
print("✅ Setup verified! Ready to start FYP implementation.")
You PASS when:
You DON'T need:
Remember: DONE > PERFECT. This is about graduation, not publication.
