How intellindust-ai-lab/gtr works
GTR is a PyTorch vision system for dense prediction: its recurrent vision backbone feeds task-specific models for detection, instance and semantic segmentation, pose, depth, and oriented-box detection. The principal workflow configures a task, loads and transforms datasets, runs training or evaluation through solver engines, and produces predictions or metrics. The README also documents ONNX/TensorRT deployment. Several subsystem paths and capabilities come from the tree and README; unsampled wiring is shown only where documented, not inferred as direct calls.
Main entry points: hybrid_encoder.py Feature encoder · train.py Training entry · vit_adapter.py GTR backbone · depth_solver.py Depth solver
Components 22
Data preparation
- Task datasets
engine/data/dataset - Transforms and batching
engine/data/transforms
GTR models
- GTR backbone
engine/gtr/backbone/vit_adapter.py - Feature encoder
engine/gtr/hybrid_encoder.py - Detection model
engine/gtr/det/decoder.py - Instance segmentation
engine/gtr/det/segmentation_head.py - Oriented detection
engine/gtr/obb/decoder.py - Pose model
engine/gtr/pose/decoder.py - Semantic segmentation
engine/gtr/semseg/gtrsemseg.py - Depth model
engine/gtr/depth/gtrdepth.py
Training and evaluation
- Config and workspace
engine/core/workspace.py - Training entry
train.py - Detection solver
engine/solver/det_solver.py - Depth solver
engine/solver/depth_solver.py - Segmentation solver
engine/solver/semseg_solver.py - Depth evaluation
engine/solver/depth_engine.py - PyTorch inference
tools/inference/torch_inf.py
Deployment
- ONNX export
tensorrt_plugin/export_onnx.py - ONNX model patches
tensorrt_plugin/gtr_onnx_patches.py - TensorRT verification
tensorrt_plugin/verify_trt.py
Other components
- Researcher
- Predictions and metrics
Connections 25
- Researcher → Config and workspace: configures
- Config and workspace → Training entry: configures
- Task datasets → Transforms and batching: loads and transforms
- Training entry → Task datasets: uses
- Training entry → Detection solver: starts training
- Training entry → Depth solver: starts training
- Training entry → Segmentation solver: starts training
- Detection solver → GTR backbone: runs model
- Depth solver → GTR backbone: runs model
- Segmentation solver → GTR backbone: runs model
- GTR backbone → Feature encoder: features
- Feature encoder → Detection model: features
- Feature encoder → Instance segmentation: features
- Feature encoder → Oriented detection: features
- Feature encoder → Pose model: features
- Feature encoder → Semantic segmentation: features
- Feature encoder → Depth model: features
- Depth solver → Depth evaluation: evaluates
- Depth evaluation → Predictions and metrics: returns metrics
- Researcher → PyTorch inference: requests inference
- PyTorch inference → Predictions and metrics: produces
- Researcher → ONNX export: requests export
- ONNX export → ONNX model patches: applies patches
- ONNX export → TensorRT verification: exports model for
- TensorRT verification → Predictions and metrics: verifies outputs