Skip to content

Pretrained Models

This page documents the pretrained checkpoints provided in Pretrained_Model/.

The repository currently provides pretrained models for:

  • GAN-based baselines
  • UNet-based baselines

GAN

GAN checkpoints are organized as:

Pretrained_Model/GAN/<task_id>/
  best_G.pth
  best_D.pth
  config.json
  eval_results.json
  epoch_history.json
  batch_history.json

Files

best_G.pth

Best generator checkpoint.

This is the main checkpoint used for:

  • inference
  • evaluation
  • pretrained result reproduction

best_D.pth

Best discriminator checkpoint.

This is mainly provided for:

  • adversarial training continuation
  • full training-state reproduction

config.json

Stores the task configuration used for training/evaluation.

This may include settings such as:

  • split strategy
  • input mode
  • sparse/dense mode
  • number of samples
  • training hyperparameters

eval_results.json

Stores evaluation results for the pretrained checkpoint.

epoch_history.json

Stores epoch-level training history.

batch_history.json

Stores batch-level training history.

Available GAN Tasks

The currently available pretrained GAN tasks are:

  • beam_dense_encoding
  • beam_dense_feature
  • random_dense_encoding
  • random_dense_feature
  • random_sparse_encoding_samples819
  • random_sparse_feature_samples819
  • scene_dense_encoding
  • scene_dense_feature

These task IDs are aligned with the benchmark naming in benchmark.md.

For evaluation or inference, users typically only need:

  • best_G.pth
  • config.json

The folder naming directly indicates the benchmark setting, e.g.:

Pretrained_Model/GAN/random_dense_feature/
Pretrained_Model/GAN/scene_dense_encoding/
Pretrained_Model/GAN/random_sparse_feature_samples819/

This makes it straightforward to map:

  • benchmark task
  • pretrained checkpoint
  • evaluation configuration

UNet

UNet checkpoints are stored under:

Pretrained_Model/Unet/
  *.pt
  *.png
  *.csv

Files

*.pt

Serialized UNet checkpoints.

These files contain pretrained model weights for different task settings.

*.png

Visualization files for qualitative result inspection.

These may include:

  • prediction visualizations
  • comparison plots
  • example outputs

*.csv

Summary tables or recorded evaluation results.

Notes

UNet checkpoints currently use a different naming style from the GAN task folders.

For example, checkpoint names may encode:

  • dense vs sparse setting
  • featuremap vs environment input
  • first-stage vs second-stage network
  • random seed

A dedicated mapping table from UNet checkpoint names to benchmark tasks will be provided in a future update.

Current Status

  • GAN pretrained folders are already organized with task-aligned naming
  • UNet checkpoints are available, but their naming will be further standardized

This means:

  • GAN checkpoints can already be used directly with the benchmark task names
  • UNet checkpoints may require manual mapping based on filename conventions

Suggested Starting Points

For most users, the following pretrained GAN models are recommended as starting points:

  • random_dense_feature — standard dense baseline
  • scene_dense_feature — cross-environment generalization benchmark
  • random_sparse_feature_samples819 — sparse reconstruction benchmark

These provide representative coverage of:

  • standard supervised prediction
  • generalization to unseen scenes
  • sparse observation settings