tianzhi0549/fcos

How tianzhi0549/fcos works

FCOS is a PyTorch, anchor-free object-detection implementation. Its principal workflow takes images through a configured detector to bounding-box predictions; training builds datasets and a model, then optimizes and checkpoints it, while evaluation aggregates predictions and evaluates datasets. The graph also includes the RetinaNet alternative, ROI heads, ONNX export, and shared box/layer utilities. Some subsystem links rely on the README or SOURCE INDEX references; unsampled implementation details are not inferred as direct calls.

Main entry points: generalized_rcnn.py Detector assembly · train_net.py Training command · fcos.py Image detector API · inference.py FCOS postprocessing

Components 24

User entry points

Training command tools/train_net.py
Image detector API fcos/fcos.py
ONNX export onnx/export_model_to_onnx.py

Training data and loop

COCO and VOC data fcos_core/data/datasets
Data loader fcos_core/data/build.py
Image transforms fcos_core/data/transforms
Training loop fcos_core/engine/trainer.py
Optimizer scheduler fcos_core/solver/build.py

Detection model

Detector assembly fcos_core/modeling/detector/generalized_rcnn.py
Backbone and FPN fcos_core/modeling/backbone
FCOS prediction head fcos_core/modeling/rpn/fcos/fcos.py
RetinaNet head fcos_core/modeling/rpn/retinanet/retinanet.py
Box mask keypoints fcos_core/modeling/roi_heads

Inference and results

FCOS postprocessing fcos_core/modeling/rpn/fcos/inference.py
Inference evaluator fcos_core/engine/inference.py
Dataset evaluation fcos_core/data/datasets/evaluation

Shared runtime support

Model configuration fcos_core/config/defaults.py
Box encoding fcos_core/modeling/box_coder.py
Box data structure fcos_core/structures/bounding_box.py
Box filtering NMS fcos_core/structures/boxlist_ops.py
Tensor and vision ops fcos_core/layers/misc.py
Checkpoint loading fcos_core/utils/checkpoint.py
Runtime utilities fcos_core/utils

Other components

Image user

Connections 26

Image user → Image detector API: submits image README.md
Image detector API → Model configuration: loads settings fcos/fcos.py
Image detector API → Detector assembly: builds detector fcos/fcos.py
Image detector API → Checkpoint loading: loads weights fcos/fcos.py
Training command → Model configuration: reads settings tools/train_net.py
Training command → Detector assembly: builds detector tools/train_net.py
Training command → Data loader: builds batches tools/train_net.py
Training command → Optimizer scheduler: builds optimizer tools/train_net.py
Training command → Training loop: starts training tools/train_net.py
Data loader → COCO and VOC data: constructs datasets fcos_core/data/build.py
Data loader → Image transforms: builds transforms fcos_core/data/build.py
Detector assembly → Backbone and FPN: assembles model
Detector assembly → FCOS prediction head: assembles model
Detector assembly → RetinaNet head: supports alternative
Detector assembly → Box mask keypoints: supports heads
FCOS prediction head → FCOS postprocessing: produces predictions
FCOS postprocessing → Box encoding: decodes boxes fcos_core/modeling/rpn/fcos/inference.py
FCOS postprocessing → Box data structure: creates detections fcos_core/modeling/rpn/fcos/inference.py
FCOS postprocessing → Box filtering NMS: filters boxes fcos_core/modeling/rpn/fcos/inference.py
Box filtering NMS → Box data structure: operates on boxes fcos_core/structures/boxlist_ops.py
Box filtering NMS → Tensor and vision ops: uses NMS ops fcos_core/structures/boxlist_ops.py
Inference evaluator → Dataset evaluation: evaluates results fcos_core/engine/inference.py
Inference evaluator → Runtime utilities: gathers predictions fcos_core/engine/inference.py
ONNX export → Detector assembly: builds detector onnx/export_model_to_onnx.py
ONNX export → Checkpoint loading: loads weights onnx/export_model_to_onnx.py
ONNX export → Inference evaluator: imports evaluator onnx/export_model_to_onnx.py

More Python diagrams: adrienverge/yamllint · pgmpy/pgmpy · sophomoresty/gemini-web2api · deepseek-ai/deepspec · samugit83/redamon

Drawn by GitDiagram on October 1, 2026 from the repository's file tree, README and source files. Read as Markdown ·