quick_start_en.md

English | 简体中文

Quick Start

1 Installation

Please follow the instructions here.

2 Making Predictions by Pre-trained Models

2.1 Download Pre-trained Weights and Demo Data

First, download pre-trained weights (model.pdparams) from here. Then, acquire the image for demonstration (demo.png) here. Put the model weights and the image in the root directory of this project.

2.2 Infer with Pre-trained Models

Execute the following instructions to make inference with the pre-trained model:

mkdir -p vis
python tools/predict.py \
    --config configs/panoptic_deeplab/panoptic_deeplab_resnet50_os32_cityscapes_1025x513_bs8_90k.yml \
    --model_path model.pdparams \
    --image_path demo.png \
    --save_dir vis

Note that we have used a dynamic-graph model for inference, which is usually of less efficiency. For more efficient model inference, please refer to the deployment documentation.

2.3 See Visualization Results

There are three output images for each input image (with the same prefix), which can be found in vis.

The image demo_sem.png visualizes the network predictions from a semantic segmentation perspective:

The image demo_ins.png visualizes the network predictions from an instance segmentation perspective:

The image demo_pan.png combines the first two images and visualizes the network predictions from a panoptic segmentation perspective:

A detailed description of can be found here.

3 Training and Evaluating Models

3.1 Prepare Datasets

This toolkit provides a few scripts to accommodate public panoptic segmentation datasets in order to use them for model training. For details of use, please see the data preparation tutorial.

3.2 Train Models

First, find the model configuration file that you plan to use in configs, e.g. configs/panoptic_deeplab/panoptic_deeplab_resnet50_os32_cityscapes_1025x513_bs8_90k.yml. Then, execute the following instructions:

python tools/train.py \
    --config configs/panoptic_deeplab/panoptic_deeplab_resnet50_os32_cityscapes_1025x513_bs8_90k.yml \
    --do_eval \
    --save_dir output

Note that some models have prerequisites. For example, you may have to compile an external C++/CUDA operator before performing any further steps. You can find the detailed instructions in the documentation for each model in config.

3.3 Evaluate Models

During and after training, you can find model checkpoints that store model weights and (possibly) optimizer parameters in output (specified by the --save_dir option in tools/train.py). You can evaluate the model weights that achieve the highest accuracy on the validation set (i.e. the checkpoint output/best_model) by running:

python tools/val.py \
    --config configs/panoptic_deeplab/panoptic_deeplab_resnet50_os32_cityscapes_1025x513_bs8_90k.yml \
    --model_path output/best_model/model.pdparams \
    --eval_sem \
    --eval_ins

By default the model is evaluated on the validation set, and the setting can be changed by modifying the configuration file. With --eval_sem and --eval_ins, semantic segmentation metrics (such as mIoU) and instance segmentation metrics (such as mAP) are also calculated and reported.