How ldkong1205/transvae works
TranSVAE is a research framework for unsupervised video domain adaptation in action recognition. Its user workflow prepares video datasets or uses extracted features, then trains a model on source and unlabeled target domains and evaluates target-domain action recognition. The model separates domain-related and temporal/semantic latent factors to support reconstruction and cross-domain transfer. The tree and excerpts support the training, model, and feature-extraction components; wiring details for unsampled dataset loaders and experiment variants are therefore kept conservative.
Main entry points: TranSVAE.py TranSVAE video model · dataset.py Training datasets · train_hmdb_ucf.py UCF-HMDB trainer · train_sprites.py Sprites trainer
Components 21
Data preparation
- Dataset splitting
dataset_preparation/dataset2split.py - List generation
dataset_preparation/video_dataset2list.py - I3D features
dataset_preparation/video2I3D_ucf_hmdb.py - Epic features
dataset_preparation/video2I3D_epic_kitchens.py - I3D backbone
dataset_preparation/pytorch_i3d.py
Adaptation training
- Training datasets
exp/dataset.py - UCF-HMDB trainer
exp/train_hmdb_ucf.py - Jester trainer
exp/train_jester.py - Epic trainer
exp/train_epic_kitchens.py - Sprites trainer
exp/train_sprites.py - Training objectives
exp/utils.py
TranSVAE model
- TranSVAE video model
exp/TranSVAE.py - Temporal aggregation
exp/TRNmodule.py - Image encoders
models/dcgan_64.py - Model primitives
models/base.py
Data and results
- Dataset lists
dataset - Video features
dataset - Target predictions
- Transferred videos
Other components
- Researcher
- Video datasets
Connections 32
- Researcher → Dataset splitting: prepares data
- Researcher → List generation: creates lists
- Researcher → UCF-HMDB trainer: starts training
- Video datasets → Dataset splitting: input videos
- Video datasets → List generation: indexes videos
- Video datasets → I3D features: video frames
- Video datasets → Epic features: video frames
- Dataset splitting → Dataset lists: writes splits
- List generation → Dataset lists: writes lists
- I3D features → I3D backbone: extracts with
- Epic features → I3D backbone: extracts with
- I3D features → Video features: writes features
- Epic features → Video features: writes features
- Dataset lists → Training datasets: loads samples
- Video features → Training datasets: loads features
- Training datasets → UCF-HMDB trainer: supplies batches
- Training datasets → Jester trainer: supplies batches
- Training datasets → Epic trainer: supplies batches
- Training datasets → Sprites trainer: supplies batches
- UCF-HMDB trainer → TranSVAE video model: trains model
- Jester trainer → TranSVAE video model: trains model
- Epic trainer → TranSVAE video model: trains model
- Sprites trainer → TranSVAE video model: trains model
- UCF-HMDB trainer → Training objectives: computes objectives
- Jester trainer → Training objectives: computes objectives
- Epic trainer → Training objectives: computes objectives
- Sprites trainer → Training objectives: computes objectives
- TranSVAE video model → Temporal aggregation: uses aggregation
- TranSVAE video model → Target predictions: predicts actions
- TranSVAE video model → Transferred videos: reconstructs domains
- Image encoders → Sprites trainer: supports images
- Researcher → Image encoders: experiments with