How hust-slow/seas works
SeaS is a PyTorch research system for few-shot industrial anomaly and normal-image generation, anomaly-mask prediction, and downstream anomaly-detection/segmentation use. The principal workflow prepares industrial data, trains the SeaS generation model, runs inference to produce images and masks, and evaluates or uses those outputs downstream. Its custom pipeline builds on Diffusers models and schedulers. README describes three implementation stages, but the provided excerpt ends before detailing them; links involving unsampled inference and mask code are therefore marked unverified.
Main entry points: _SeaSpipeline.py SeaS pipeline · seas.py Generator training · seas_mask.py Mask prediction · seas_infer.py Inference runner
Components 18
Data preparation
- Dataset loader
datasets/seas_dataset.py
Generation model
- Generator training
models/seas.py - SeaS pipeline
models/modules/_SeaSpipeline.py - Conditioned U-Net
models/modules/_unet_2d_condition.py - Autoencoder
models/modules/_autoencoder_kl.py - Inference runner
models/seas_infer.py - Generated images
models/modules/_SeaSpipeline.py
Mask prediction
- Mask prediction
models/seas_mask.py - Anomaly masks
models/seas_mask.py
Diffusion runtime
- Diffusion scheduler
diffusers/schedulers/scheduling_dpmsolver_multistep.py - Image processing
diffusers/image_processor.py - Pipeline utilities
diffusers/pipelines/pipeline_utils.py - Model configuration
diffusers/configuration_utils.py
Evaluation and downstream
- Generation metrics
utils/metrics/metrics_lpips.py - Downstream tasks
- Metric dataset
utils/metrics/dataset.py
Other components
- Researcher
- Industrial datasets
Connections 23
- Researcher → Dataset loader: selects data
- Industrial datasets → Dataset loader: loads samples
- Dataset loader → Generator training: supplies samples
models/seas.py - Generator training → Conditioned U-Net: trains
models/seas.py - Generator training → Autoencoder: uses
models/seas.py - Generator training → SeaS pipeline: uses
- Inference runner → SeaS pipeline: runs inference
- SeaS pipeline → Conditioned U-Net: denoises with
models/modules/_SeaSpipeline.py - SeaS pipeline → Autoencoder: encodes and decodes
models/modules/_SeaSpipeline.py - SeaS pipeline → Diffusion scheduler: steps diffusion
models/modules/_SeaSpipeline.py - SeaS pipeline → Pipeline utilities: extends
models/modules/_SeaSpipeline.py - SeaS pipeline → Image processing: preprocesses images
models/modules/_SeaSpipeline.py - SeaS pipeline → Generated images: produces
models/modules/_SeaSpipeline.py - Generator training → Mask prediction: trains mask stage
- Inference runner → Mask prediction: predicts masks
- Mask prediction → Generated images: uses generated images
- Mask prediction → Anomaly masks: produces
- Generated images → Generation metrics: evaluates
- Metric dataset → Generation metrics: supplies comparisons
- Generated images → Downstream tasks: supports
- Anomaly masks → Downstream tasks: supports
- Researcher → Inference runner: requests generation
- Model configuration → Pipeline utilities: configures components
diffusers/pipelines/pipeline_utils.py