Skip to content

siliconflow/nexeval

Repository files navigation

nexeval

  1. Introduction 🌟
  2. Installation 🛠️
  3. Quick Start
  4. Qualitative Evaluation 🎨
  5. References 📚
  6. Citing 📖

Introduction

This repository is used for evaluating the quality of generation after compilation acceleration using OneDiff.

It can also serve as a benchmark for evaluating the performance of different text-to-image models.

Installation

  1. Prepare the OneDiff environment.

    Follow the instructions to install OneDiff and other dependencies:

  2. Prepare Benchmark environment.

    pip3 install -r requirements.txt
    pip3 install -e .
    

Quick Start

Evaluating the use of all generative models is divided into two steps, taking the kolors model as an example:

1. Generate benchmark images.

  • On MS COCO-30K:

    Assume that the folders kolors_torch_coco, kolors_oneflow_coco, and kolors_nexfort_coco respectively store the original images, images compiled by the onediff's oneflow backend, and images compiled by the nexfort backend.

    # Create a path to store the generated images.
    mkdir /path/to/your/kolors_torch_coco
    
    # Original pytorch generates reference images.
    python3 models/kolors/text_to_image_kolors_quality_benchmark.py \
    --dataset coco \
    --csv-file resources/MS-COCO_val2014_30k_captions.csv \
    --output-dir /path/to/your/kolors_torch_coco
    
    # Accelerate using onediff's oneflow backend.
    python3 models/kolors/text_to_image_kolors_quality_benchmark.py \
    --compiler oneflow \
    --dataset coco \
    --csv-file resources/MS-COCO_val2014_30k_captions.csv \
    --output-dir /path/to/your/kolors_oneflow_coco
    
    # Accelerate using onediff's nexfort backend.
    python3 models/kolors/text_to_image_kolors_quality_benchmark.py \
    --compiler nexfort \
    --compiler-config '{"mode": "max-optimize:max-autotune:low-precision", "memory_format": "channels_last"}' \
    --dataset coco \
    --csv-file resources/MS-COCO_val2014_30k_captions.csv \
    --output-dir /path/to/your/kolors_nexfort_coco
    
  • On Human Preference Dataset v2 (HPD v2):

    Simply modify the --dataset parameters, do not read prompts from the --csv-file parameter, and customize the --output-dir for generating images. For example:

    python3 models/kolors/text_to_image_kolors_quality_benchmark.py \
    --dataset hps \
    --output-dir /path/to/your/kolors_torch_hps
    

2. Test using multiple indicators with scripts.

bash scripts/run_kolors_tests.sh coco
bash scripts/run_kolors_tests.sh hps

A quality report can refer to: models/kolors/README.md

Qualitative evaluation

We collected several typical prompts to visualize the generated images for qualitative evaluation. These prompts reflect the model's semantic understanding, long text, detail, spatial relationships, diversity, clarity, and text embedding capabilities.

References

Citing

@misc{odeval,
  author = {Xiang Li and others},
  title = {odeval: A Library for benchmarking the accelerated generation quality},
  year = {2023},
  publisher = {SiliconFlow},
  howpublished = {\url{https://github.com/siliconflow/odeval}},
  note = {Accessed: 2024-07-26}
}

About

Benchmarking the accelerated generation quality of OneDiff.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published