How yashvashist-iitbhu/enhanced-loss-function works
Project type and purpose
• This is an ML research/pipeline project aimed at implementing an Enhanced AM-Softmax loss for face-recognition tasks. It’s not a web or full-stack app but an offline data-processing and model-training pipeline, driven by Jupyter notebooks, CSV datasets, and a final report.
File structure
Top‐level files:
• Final_model_94%.ipynb – Main training & evaluation notebook.
• LFW.ipynb – Data preprocessing and exploration on the LFW dataset.
• lfw_allnames.csv, lfw_readme.csv – Dataset metadata.
• matchpairsDev*.csv / mismatchpairsDev*.csv – Positive/negative face-pair labels for dev/train/test splits.
• report (2).pdf – Final write-up of methods and results.
• README.md – Project overview and loss-function math.
README for additional insights
• Purpose: Modify AM-Softmax with an additive margin and exponential scaling factor to improve intra-class compactness and inter-class separability.
• Key components: data ingestion, custom loss module, model training, and evaluation.
a. Main components:
1. Data Ingestion & Storage
2. Data Preprocessing & Pair Generation
3. Model Definition & Enhanced Loss Module
4. Training Engine
5. Evaluation & Metrics
6. Reporting & Visualization
b. Relationships & data flow:
• “Data Ingestion” → “Preprocessing”
• “Preprocessing” → “Model Definition” & “Training Engine”
• “Model Definition” (Loss Module) → “Training Engine”
• “Training Engine” → “Evaluation & Metrics”
• “Evaluation & Metrics” → “Reporting & Visualization”
c. Architectural patterns/design principles:
• Pipeline pattern: sequential stages transforming raw CSV → features → trained model → results.
• Modularity: separate components for loss function, data prep, training, evaluation.
d. Technologies/frameworks:
• Jupyter Notebooks for orchestration (LFW.ipynb, Final_model_94%.ipynb)
• Python libraries (e.g., NumPy, Pandas, TensorFlow or PyTorch – implied in notebooks)
• CSV files as data store
• PDF report for final deliverable
Detailed element checklist
• Component labels: “Data Ingestion (lfw_allnames.csv, …)”, “Preprocessing (LFW.ipynb)”, “Enhanced AM-Softmax Loss Module”, etc.
• Parameter inputs (m, s) feeding into loss block
• Output artifacts: trained model checkpoint, evaluation metrics, PDF report