How nvlabs/spade works
SPADE is a research implementation for semantic image synthesis: it turns paired semantic labels (and optional instance maps) into generated images, supporting dataset preparation, model training, and checkpoint-based inference. Training is initiated by the training entry point, which loads data and delegates optimization to the trainer; testing uses the model to generate result images. Dataset variants, SPADE network components, and optional logging/output capabilities are represented. Edges follow shown source references or README statements; unsampled entry-point details are limited to documented relationships.
Main entry points: train.py Training entry · pix2pix_model.py Pix2Pix model · test.py Testing entry · visualizer.py Progress output
Components 24
Run workflow
- Training entry
train.py - Testing entry
test.py - Run options
options/base_options.py - Training harness
trainers/pix2pix_trainer.py
Semantic data
- Dataset selection
data/__init__.py - Dataset adapters
data/coco_dataset.py - Paired samples
data/pix2pix_dataset.py - Image folder
data/image_folder.py - Image transforms
data/base_dataset.py
Synthesis model
- Pix2Pix model
models/pix2pix_model.py - SPADE generator
models/networks/generator.py - Discriminator
models/networks/discriminator.py - VAE encoder
models/networks/encoder.py - Network blocks
models/networks/architecture.py - SPADE normalization
models/networks/normalization.py - GAN losses
models/networks/loss.py
Training support
- Instance maps
datasets/coco_generate_instance_map.py - Iteration tracking
util/iter_counter.py - Progress output
util/visualizer.py - Image utilities
util/util.py - HTML results
util/html.py
Other components
- Researcher
- TensorBoard
- Generated images
Connections 26
- Researcher → Training entry: starts training
README.md - Researcher → Testing entry: starts inference
README.md - Training entry → Run options: reads options
README.md - Training entry → Dataset selection: loads dataset
README.md - Training entry → Training harness: runs training
README.md - Training entry → Iteration tracking: tracks epochs
README.md - Training entry → Progress output: reports progress
README.md - Testing entry → Dataset selection: loads dataset
README.md - Testing entry → Pix2Pix model: generates images
README.md - Testing entry → Generated images: writes results
README.md - Run options → Dataset selection: configures dataset
options/base_options.py - Dataset selection → Dataset adapters: selects adapter
data/__init__.py - Dataset adapters → Paired samples: extends paired loader
data/coco_dataset.py - Paired samples → Image transforms: transforms samples
data/pix2pix_dataset.py - Paired samples → Image folder: discovers images
data/coco_dataset.py - Instance maps → Dataset adapters: prepares COCO maps
README.md - Training harness → Pix2Pix model: optimizes model
trainers/pix2pix_trainer.py - Pix2Pix model → SPADE generator: runs generator
models/pix2pix_model.py - Pix2Pix model → Discriminator: runs discriminator
models/pix2pix_model.py - Pix2Pix model → VAE encoder: encodes VAE input
models/pix2pix_model.py - Pix2Pix model → GAN losses: computes GAN loss
models/pix2pix_model.py - SPADE generator → Network blocks: uses blocks
models/networks/generator.py - Network blocks → SPADE normalization: applies SPADE
models/networks/architecture.py - Progress output → Image utilities: converts visuals
util/visualizer.py - Progress output → HTML results: renders galleries
util/visualizer.py - Progress output → TensorBoard: logs summaries
util/visualizer.py