How mpreda01/indus works
This repository trains dense-prediction models for semantic segmentation and depth estimation from driving-scene images. The principal workflow loads split datasets, applies paired image/label transforms, predicts selected tasks, and computes losses and metrics during training and validation; visualization and experiment tracking support monitoring. Synscapes preprocessing converts source data to the dataset layout. The experiment module’s sampled code establishes its dataset, transform, model, loss, metric, and visualization responsibilities; some subsystem wiring and checkpoint details are based on README documentation rather than sampled implementation.
Main entry points: experiment_multitask.py Multitask experiment · preprocess_synscapes.py Synscapes preprocessing · dataset_miniscapes.py Miniscapes loader · train.py Training entry point
Components 20
Data
- Synscapes preprocessing
source/datasets/preprocess_synscapes.py - Miniscapes loader
source/datasets/dataset_miniscapes.py - Task definitions
source/datasets/definitions.py - Image transforms
source/utils/transforms.py - Label mappings
labels.py
Prediction models
- DeepLab models
source/models/model_deeplab_v3_plus.py - Multitask model
source/models/model_deeplab_v3_plus_multitask.py - Adaptive depth model
source/models/model_adaptive_depth.py - Encoder and decoder
source/models/model_parts.py
Training
- Training entry point
source/scripts/train.py - Config handling
source/utils/config.py - Multitask experiment
source/experiments/experiment_multitask.py - Task losses
source/losses
Evaluation and monitoring
- Task metrics
source/utils/metrics.py - Prediction visualization
source/utils/visualization.py
Other components
- Researcher
- Synscapes source
- Dataset files
- Weights and Biases
- Checkpoints and submissions
Connections 23
- Researcher → Training entry point: launches
- Training entry point → Config handling: loads config
- Training entry point → Multitask experiment: starts training
- Synscapes source → Synscapes preprocessing: source images
- Synscapes preprocessing → Miniscapes loader: uses depth encoding
- Synscapes preprocessing → Label mappings: maps labels
- Synscapes preprocessing → Dataset files: writes layout
- Dataset files → Miniscapes loader: read samples
- Multitask experiment → Miniscapes loader: creates splits
- Multitask experiment → Image transforms: builds transforms
- Miniscapes loader → Image transforms: transforms samples
- Multitask experiment → Task definitions: selects tasks
- Multitask experiment → DeepLab models: resolves model
- Multitask experiment → Multitask model: resolves model
- Multitask experiment → Adaptive depth model: resolves model
- DeepLab models → Encoder and decoder: uses components
- Multitask model → Encoder and decoder: uses components
- Adaptive depth model → Encoder and decoder: uses components
- Multitask experiment → Task losses: computes task losses
- Multitask experiment → Task metrics: updates metrics
- Multitask experiment → Prediction visualization: renders predictions
- Multitask experiment → Weights and Biases: logs run
- Multitask experiment → Checkpoints and submissions: saves training artifacts