How hl-yuan/pmimc works
Project type and purpose
• This is a research-oriented machine-learning codebase implementing “Prototype Matching Learning” for incomplete multi-view clustering.
• It provides modules for loading multi-view data, defining a neural network, computing prototype-matching losses, running a clustering algorithm (k-means GPU), and evaluating clustering performance.
File structure
Top-level files & directories:
• Dataset/
– Caltech101_7.mat, handwritten1031_v73.mat – raw multi-view datasets
• datasets.py – dataset wrappers & preprocessing pipelines
• load_data.py – functions to split views and simulate missing views
• network.py – PyTorch module definitions for the clustering network
• loss.py – prototype matching loss functions
• Nmetrics.py – evaluation metrics (e.g., clustering accuracy, NMI)
• kmeans_gpu.py – GPU-accelerated k-means clustering
• main.py – orchestration script: parses arguments, calls all pieces in order
• utils.py – shared utilities (logging, checkpointing, argument parsing)
• Dataset/ – storage of .mat files
Extract architectural insights from the README
• Core dependencies: PyTorch, NumPy, scikit-learn, munkres (Hungarian algorithm)
• Entry point: python main.py --i_d 0 --missrate 0.5
• Workflow stages mentioned: data loading, training, evaluation
Major system components
a. Data Storage & Input
– Raw .mat files under Dataset/
– Configuration (hyperparameters, missing-rate) passed to main.py
b. Data Loader & Preprocessor (datasets.py, load_data.py)
– Loads views, applies missing-view simulation
– Outputs batched multi-view tensors
c. Model Definition (network.py)
– Neural network encoders per view
– Prototype representation layers
d. Loss Computation (loss.py)
– Prototype Matching Loss
– Reconstruction or contrastive terms if any
e. Clustering Module (kmeans_gpu.py)
– Runs k-means on learned features to update prototypes
f. Training Loop (main.py)
– Orchestrates forward/backward passes, optimizer steps, prototype updates
g. Evaluation (Nmetrics.py)
– Calculates clustering metrics (ACC, NMI) after each epoch or final
h. Utilities (utils.py)
– Logging, checkpoint save/load, argument parsing
Relationships & data flows
• Data Loader outputs → Model Definition (forward pass)
• Model outputs & Prototypes → Loss Computation
• Loss → Optimizer step (inside Training Loop)
• Feature embeddings → Clustering Module to recompute prototypes
• Clustering labels + ground truth → Evaluation Module
• Utilities wrap around Training Loop for config, logging, checkpoints
Architectural patterns and design principles
• Modular separation: data, model, loss, clustering, evaluation
• Iterative EM-like loop: feature learning ↔ prototype update (k-means)
• Dependency injection: main.py wires all components based on config
Technologies & frameworks call-outs
• PyTorch: network, autograd, optim
• NumPy/Scipy: data manipulation, .mat loading
• scikit-learn & munkres: metrics and assignment algorithms
• CUDA (via PyTorch & kmeans_gpu.py) for acceleration