PaSSPI: Parallel Semantic Segmentation with Pipeline and Image Slicing

Install

• Pippy

git clone https://github.com/pytorch/PiPPy.git

• Install Environment

pip install -r requirements.txt

Run

Baseline Model: nvidia/segformer-b1-finetuned-cityscapes-1024-1024

Run Segformer with PiPPy

python3 hf_generate.py
# measure latency
torchrun --nproc_per_node 8 pippy_evaluate.py

Slicing Images

Parallel Processing of Divided Photos

Divide the photo into four equal parts and sixteen equal parts, then utilize multi-threading for parallelization to process them simultaneously.

python3 slicing_evaluate.py

Parallel Processing of Multiple Photos

Utilize both multi-thread and multi-process to concurrently process multiple images, enhancing overall throughput and resource utilization.

• multi-thread

python3 multi_images_handling/test_threading.py

• multi-process

python3 multi_images_handling/test_multiprocessing.py

Usage of cityscape dataset

1. Download and unzip the gtFine_trainvaltest.zip and leftImg8bit_trainvaltest.zip from cityscapes official websites
2. Use Validation set from cityscapes dataset

Parallel GPU inference

To run parallel_gpu.py with multiple GPUs, use the command:

accelerate config

Follow the settings in the terminal to finish editing the config file.

After that, run the code by:

accelerate launch parallel_gpu.py