How jirkapivrnec/projconv2d works
ProjConv2D is a PyTorch image-classification experiment. A researcher selects FashionMNIST or CIFAR-10, applies normalization, loads batches, trains SlidingProjectionNet, validates predictions and metrics, then exports feature-map and error visualizations. The principal model path is a learnable patch-projection layer followed by conventional convolutional blocks and a classifier. The graph reflects sampled runtime code and documented capabilities; some experiment variants are described but not independently wired in the excerpts.
Components 18
Data Preparation
- Experiment Entry
train.py - Image Transforms
train.py - Dataset Loaders
train.py
Model Architecture
- Model Orchestration
extendedCNN.py - Projection Operator
ProjConv2d.py - CNN Backbone
extendedCNN.py - Classifier Head
extendedCNN.py
Training Evaluation
- Training Loop
train.py - Validation Loop
train.py - Loss Function
train.py - Optimizer Scheduler
train.py - Validation Metrics
train.py
Analysis Outputs
- Feature Analysis
train.py - Feature Map Artifacts
train.py - Evaluation Artifacts
train.py - Model Checkpoints
Other components
- Researcher
- Torchvision Datasets
Connections 22
- Researcher → Experiment Entry: launches
- Experiment Entry → Image Transforms: creates
- Experiment Entry → Dataset Loaders: creates
- Image Transforms → Dataset Loaders: configures
- Torchvision Datasets → Dataset Loaders: supplies data
- Dataset Loaders → Training Loop: batches samples
- Dataset Loaders → Validation Loop: batches samples
- Experiment Entry → Model Checkpoints: loads state
- Training Loop → Model Orchestration: calls model
- Validation Loop → Model Orchestration: calls model
- Model Orchestration → Projection Operator: dispatches patches
- Model Orchestration → CNN Backbone: dispatches features
- Projection Operator → CNN Backbone: returns features
- CNN Backbone → Classifier Head: returns features
- Training Loop → Loss Function: computes loss
- Validation Loop → Loss Function: computes loss
- Training Loop → Optimizer Scheduler: steps optimizer
- Validation Loop → Optimizer Scheduler: steps scheduler
- Validation Loop → Validation Metrics: calculates metrics
- Validation Loop → Feature Analysis: provides maps
- Feature Analysis → Feature Map Artifacts: writes images
- Validation Metrics → Evaluation Artifacts: writes reports