How cseeyangchen/flora works
Flora is a research implementation for zero-shot skeleton action recognition. Its documented workflow prepares NTU or PKU-MMD skeleton data, optionally pretrains a Shift-GCN encoder or uses pre-extracted features, then runs the main training/evaluation entry point. The sampled mapping model combines motion and text semantics with learned encoders/decoders and uses CLIP-derived class embeddings. Trainer internals and several documented capabilities are not shown in the excerpts, so those links are left unspecified rather than inferred.
Main entry points: main.py Encoder pretraining · trainer.py Training and evaluation · fm.py Mapping network · get_raw_skes_data.py NTU extraction
Components 21
Data preparation
- NTU extraction
data/ntu60/get_raw_skes_data.py - NTU denoising
data/ntu60/get_raw_denoised_data.py - NTU transform
data/ntu60/seq_transformation.py - PKU extraction
data/pkummd/part1/pku_part1_skeleton.py - PKU generation
data/pkummd/part1/pku_part1_gendata.py - PKU preprocessing
data/pkummd/part1/preprocess_pku.py - PKU concatenation
data/pkummd/part1/pku_concat.py
Skeleton encoding
- Encoder pretraining
pretrain_skeleton/main.py - Skeleton feeder
pretrain_skeleton/feeders/feeder_ntu.py - Shift-GCN
pretrain_skeleton/model/shift_gcn.py
Zero-shot recognition
- Recognition entry
main.py - Training and evaluation
trainer.py - Feature feeder
feeders/feeder_ntu.py
Semantic mapping
- Mapping network
models/fm.py - Diffusion transformer
models/dit.py
Other components
- Researcher
- Raw NTU skeletons
- Raw PKU skeletons
- Skeleton features
- CLIP text model
- Recognition results
Connections 21
- Researcher → Raw NTU skeletons: provides
- Researcher → Raw PKU skeletons: provides
- Raw NTU skeletons → NTU extraction: reads skeletons
- NTU extraction → NTU denoising: extracts bodies
- NTU denoising → NTU transform: cleans sequences
- Raw PKU skeletons → PKU extraction: reads skeletons
- PKU extraction → PKU generation: generates data
- PKU generation → PKU preprocessing: creates sequences
- PKU preprocessing → PKU concatenation: downsamples data
- Researcher → Encoder pretraining: starts pretraining
- Encoder pretraining → Skeleton feeder: loads samples
- Encoder pretraining → Shift-GCN: trains encoder
- Researcher → Skeleton features: provides features
- Researcher → Recognition entry: starts experiment
- Recognition entry → Training and evaluation: starts training
- Researcher → Skeleton features: selects feature input
- Skeleton features → Feature feeder: supplies features
- Training and evaluation → Recognition results: produces evaluation
- CLIP text model → Mapping network: encodes prompts
- Mapping network → Diffusion transformer: uses diffusion model
- Training and evaluation → Mapping network: trains or evaluates