/RALF

[CVPR24 Oral] Official repository for RALF: Retrieval-Augmented Layout Transformer for Content-Aware Layout Generation

Primary LanguagePythonApache License 2.0Apache-2.0

Retrieval-Augmented Layout Transformer for Content-Aware Layout Generation

Daichi Horita1  Naoto Inoue2  Kotaro Kikuchi2  Kota Yamaguchi2  Kiyoharu Aizawa1 
1The University of Tokyo, 2CyberAgent

CVPR 2024 (Oral)

arxiv paper License

Content-aware graphic layout generation aims to automatically arrange visual elements along with a given content, such as an e-commerce product image. This repository aims to provide all-in-one package for content-aware layout generation. If you like this repository, please give it a star!

teaser In this paper, we propose Retrieval-augmented content-aware layout generation. We retrieve nearest neighbor examples based on the input image and use them as a reference to augment the generation process.

Content

  • Setup
  • Dataset splits
  • Pre-processing Dataset
  • Training
  • Inference & Evaluation
  • Inference using a canvas

Overview of Benchmark

We provide not only our method (RALF / Autoreg Baseline) but also other state-of-the-art methods for content-aware layout generation. The following methods are included in this repository:

Setup

We recommend using Docker to easily try our code.

1. Requirements

  • Python3.9+
  • PyTorch 1.13.1

We recommend using Poetry (all settings and dependencies in pyproject.toml).

2. How to install

Local environment

  1. Install poetry (see official docs).
curl -sSL https://install.python-poetry.org | python3 -
  1. Install dependencies (it may be slow..)
poetry install

Docker environment

  1. Build a Docker image
bash scripts/docker/build.sh
  1. Attach the container to your shell.
bash scripts/docker/exec.sh
  1. Install dependencies in the container
poetry install

3. Setup global environment variables

Some variables should be set. Please make scripts/bin/setup.sh on your own. At least these three variables should be set. If you download the provided zip, please ignore the setup.

DATA_ROOT="./cache/dataset"

Some variables might be set (e.g., OMP_NUM_THREADS)

4. Check Checkpoints and experimental results

The checkpoints and generated layouts of the Autoreg Baseline and our RALF for the unconstrained and constrained tasks are available at google drive or Microsoft OneDrive. After downloading it, please run unzip cache.zip in this directory. Note that the file size is 13GB.

cache directory contains:

  1. the preprocessed CGL dataset in cache/dataset.
  2. the weights of the layout encoder and ResNet50 in cache/PRECOMPUTED_WEIGHT_DIR.
  3. the pre-computed layout feature of CGL in cache/eval_gt_features.
  4. the relationship of elements for a relationship task in cache/pku_cgl_relationships_dic_using_canvas_sort_label_lexico.pt.
  5. the checkpoints and evaluation results of both the Autoreg Baseline and our RALF in cache/training_logs.

Dataset splits

Train / Test / Val / Real data splits

We perform preprocessing on the PKU and CGL datasets by partitioning the training set into validation and test subsets, as elaborated in Section 4.1. The CGL dataset, as distributed, is already segmented into these divisions. For replication of our results, we furnish details of the filenames within the data_splits/splits/<DATASET_NAME> directory. We encourage the use of these predefined splits when conducting experiments based on our setting and using our reported scores such as CGL-GAN and DS-GAN.

IDs of retrieved samples

We use the training split as a retrieval source. For example, when RALF is trained with the PKU, the training split of PKU is used for training and evaluation. We provide the pre-computed correspondense using DreamSim [Fu+ NeurIPS23] in data_splits/retrieval/<DATASET_NAME>. The data structure follows below

FILENAME:
    - FILENAME top1
    - FILENAME top2
    ...
    - FILENAME top16

You can load an image from <IMAGE_ROOT>/<FILENAME>.png.

Pre-processing Dataset

We highly recommend to pre-process datasets since you can run your experiments as quick as possible!!
Each script can be used for processing both PKU and CGL by specifying --dataset_type (pku|cgl)

Dataset setup

Folder names with parentheses will be generated by this pipeline.

<DATASET_ROOT>
| - annotation
| | (for PKU)
| | - train_csv_9973.csv
| | - [test_csv_905.csv](https://drive.google.com/file/d/19BIHOdOzVPBqf26SZY0hu1bImIYlRqVd/view?usp=sharing)
| |  (for CGL)
| | - layout_train_6w_fixed_v2.json
| | - layout_test_6w_fixed_v2.json
| | - yinhe.json
| - image
| | - train
| | | - original: image with layout elements
| | | - (input): image without layout elements (by inpainting)
| | | - (saliency)
| | | - (saliency_sub)
| | - test
| | | - input: image without layout elements
| | | - (saliency)
| | | - (saliency_sub)

Image inpainting

poetry run python image2layout/hfds_builder/inpainting.py --dataset_root <DATASET_ROOT>

Saliency detection

poetry run python image2layout/hfds_builder/saliency_detection.py --input_dir <INPUT_DIR> --output_dir <OUTPUT_DIR> (--algorithm (isnet|basnet))

Aggregate data and dump to HFDS

poetry run python image2layout/hfds_builder/dump_dataset.py --dataset_root <DATASET_ROOT> --output_dir <OUTPUT_DIR>

Training

Tips

configs/<METHOD>_<DATASET>.sh contains the hyperparameters and settings for each method and dataset. Please refer to the file for the details. In particular, please check whether the debugging mode DEBUG=True or False.

Autoreg Baseline with CGL

Please run

bash scripts/train/autoreg_cgl.sh <GPU_ID> <TASK_NAME>
# If you wanna run train and eval, please run
bash scripts/run_job/end_to_end.sh <GPU_ID e.g. 0> autoreg cgl <TASK_NAME e.g. uncond>

where TASK_NAME indicates the unconstrained and constrained tasks. Please refer to the below task list:

  1. uncond: Unconstraint generation
  2. c: Category → Size + Position
  3. cwh: Category + Size → Position
  4. partial: Completion
  5. refinement: Refinement
  6. relation: Relationship

RALF with CGL

The dataset with inpainting.

Please run

bash scripts/train/ralf_cgl.sh <GPU_ID> <TASK_NAME>
# If you wanna run train and eval, please run
bash scripts/run_job/end_to_end.sh <GPU_ID e.g. 0> ralf cgl <TASK_NAME e.g. uncond>

Other methods

For example, these scripts are helpful. end_to_end.sh is a wrapper script for training, inference, and evaluation.

# DS-GAN with CGL dataset
bash scripts/run_job/end_to_end.sh 0 dsgan cgl uncond
# LayoutDM with CGL dataset
bash scripts/run_job/end_to_end.sh 2 layoutdm cgl uncond
# CGL-GAN + Retrieval Augmentation with CGL dataset
bash scripts/run_job/end_to_end.sh 2 cglgan_ra cgl uncond

Inference & Evaluation

Experimental results are provided in cache/training_logs. For example, a directory of autoreg_c_cgl, which the results of the Autoreg Baseline with Category → Size + Position task, includes:

  1. test_<SEED>.pkl: the generated layouts
  2. layout_test_<SEED>.png: the rendered layouts, in which top sample is ground truth and bottom sample is a predicted sample
  3. gen_final_model.pt: the final checkpoint
  4. scores_test.tex: summarized qualitative results

Annotated split

Please see and run

bash scripts/eval_inference/eval_inference.sh <GPU_ID> <JOB_DIR> <COND_TYPE> cgl

For example,

# Autoreg Baseline with Unconstraint generation
bash scripts/eval_inference/eval_inference.sh 0 "cache/training_logs/autoreg_uncond_cgl" uncond cgl

Unannotated split

The dataset with real canvas i.e. no inpainting.

Please see and run

bash scripts/eval_inference/eval_inference_all.sh <GPU_ID>

Inference using a canvas

Please run

bash scripts/run_job/inference_single_data.sh <GPU_ID> <JOB_DIR> cgl <SAMPLE_ID>

where SAMPLE_ID can optionally be set as a dataset index.

For example,

bash scripts/run_job/inference_single_data.sh 0 "./cache/training_logs/ralf_uncond_cgl" cgl

Inference using your personal data

Please customize image2layout/train/inference_single_data.py to load your data.

Citation

If you find our work useful in your research, please consider citing:

@article{horita2024retrievalaugmented,
    title={{Retrieval-Augmented Layout Transformer for Content-Aware Layout Generation}},
    author={Daichi Horita and Naoto Inoue and Kotaro Kikuchi and Kota Yamaguchi and Kiyoharu Aizawa},
    booktitle={CVPR},
    year={2024}
}