Practical overview
When it fits
Use Cvpods when you want a code-based research workflow for computer-vision experiments, with task configurations and command-line training and testing tools.
Who it is for
- Computer-vision researchers building and comparing model experiments.
- Developers who want a configurable PyTorch codebase for training and testing models across multiple vision tasks.
Documented uses
- Configure and train an object-detection experiment on COCO.
- Test a trained model using supplied weights and an output directory.
- Explore model implementations and experiments across several computer-vision tasks.
- Set up a multi-machine training command with machine count, rank and distributed URL.
Requirements & limits
- Requires Linux, Python 3.6 or later, PyTorch 1.3 or later, and matching torchvision.
- The README's GPU-oriented installation guidance requires an available GPU; OpenCV is optional for demos and visualization.
- The COCO example expects the dataset at datasets/coco and uses an eight-GPU command.
- The project is licensed under Apache 2.0 and provides tutorials, API documentation and a model-zoo index.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Train and test the COCO object-detection example
Use the documented COCO RetinaNet example to run a training and testing workflow.
Before you start
- Linux with Python 3.6 or later.
- PyTorch 1.3 or later and a matching torchvision installation.
- A GPU for the README's GPU-oriented installation and example.
- A COCO dataset and the cvpods repository.
Steps
Prepare the dataset path by linking your COCO dataset to datasets/coco.
Change to playground/retinanet/retinanet.res50.fpn.coco.multiscale.1x.
Run pods_train --num-gpus 8.
Run pods_test --num-gpus 8, supplying MODEL.WEIGHTS and OUTPUT_DIR as needed.
Expected outcome
A documented object-detection training and testing workflow using the COCO example.
Guide limitations
- The example assumes the dataset is prepared and uses eight GPUs.
- The README provides commands but does not establish results for a particular machine.
Configure a multi-machine training command
Set the documented machine-count, rank and distributed URL arguments for multi-node training.
Before you start
- A configured cvpods experiment and the training dataset.
- Multiple machines with GPU resources.
- Values for machine count, each machine's rank and the master address and port.
Steps
Set --num-machines to the number of machines.
Assign each machine its --machine-rank, from 0 through N-1.
Set --dist-url to the shared tcp://MASTER_IP:port address.
Run pods_train with the configured values and the documented --num-gpus argument.
Expected outcome
A multi-node training command configured according to the README example.
Guide limitations
- The README demonstrates the command but does not establish performance or compatibility for a specific cluster.
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 16 October 2026
- Original website title
- Megvii-BaseDetection/cvpods · Official homepage title · retained in its source language