We provide code to reproduce our paper on markup-to-image generation. Our code is built on top of HuggingFace diffusers and transformers.
Online Demo: https://huggingface.co/spaces/yuntian-deng/latex2im.
| Scheduled Sampling | Baseline | Ground Truth |
|---|---|---|
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
pip install transformers
pip install datasets
pip install accelerate
pip install -qU git+https://github.com/da03/diffusers
All datasets have been uploaded to Huggingface datasets.
To train the diffusion model,
python src/train.py --save_dir models/math
To train the diffusion model,
python src/train.py --dataset_name yuntian-deng/im2html-100k --save_dir models/tables
In our paper, we trained on the music dataset with 4 A100 GPUs. You might need to tune --batch_size and --gradient_accumulation_steps if you want to use a single GPU to train or if your GPUs have less memory.
We first run
accelerate config
to use 4 GPUs on a single machine. Note that we did not use fp16 or DeepSpeed.
Next, we launch multi-GPU training using accelerate:
accelerate launch src/train.py --dataset_name yuntian-deng/im2ly-35k-syn --save_dir models/music
To train the diffusion model,
python src/train.py --dataset_name yuntian-deng/im2smiles-20k --save_dir models/molecules
To generate,
python scripts/generate.py --model_path models/math/scheduled_sampling/model_e100_lr0.0001.pt.100 --output_dir outputs/math --save_intermediate_every -1
To generate,
python scripts/generate.py --dataset_name yuntian-deng/im2html-100k --model_path models/tables/scheduled_sampling/model_e100_lr0.0001.pt.100 --output_dir outputs/tables --save_intermediate_every -1
To generate,
python scripts/generate.py --dataset_name yuntian-deng/im2ly-35k-syn --model_path models/music/scheduled_sampling/model_e100_lr0.0001.pt.100 --output_dir outputs/music --save_intermediate_every -1
To generate,
python scripts/generate.py --dataset_name yuntian-deng/im2smiles-20k --model_path models/molecules/scheduled_sampling/model_e100_lr0.0001.pt.100 --output_dir outputs/molecules --save_intermediate_every -1
To visualize the generation process, we need to first use the following command to save the intermediate images during generation:
python scripts/generate.py --model_path models/math/scheduled_sampling/model_e100_lr0.0001.pt.100 --output_dir outputs/math/scheduled_sampling_visualization --save_intermediate_every 1 --num_batches 1
Next, we put together a gif image from the generated images:
python scripts/make_gif.py --input_dir outputs/math/scheduled_sampling_visualization/ --output_filename imgs/math_rendering.gif --select_filename 433d71b530.png --show_every 10
We can similarly visualize results from the baseline.
python scripts/generate.py --model_path models/math/baseline/model_e100_lr0.0001.pt.100 --output_dir outputs/math/baseline_visualization --save_intermediate_every 1 --num_batches 1
python scripts/make_gif.py --input_dir outputs/math/baseline_visualization/ --output_filename imgs/math_rendering_baseline.gif --select_filename 433d71b530.png --show_every 10
To visualize the generation process, we need to first use the following command to save the intermediate images during generation:
python scripts/generate.py --dataset_name yuntian-deng/im2html-100k --model_path models/tables/scheduled_sampling/model_e100_lr0.0001.pt.100 --output_dir outputs/tables/scheduled_sampling_visualization --save_intermediate_every 1 --num_batches 1
Next, we put together a gif image from the generated images:
python scripts/make_gif.py --input_dir outputs/tables/scheduled_sampling_visualization/ --output_filename imgs/tables_rendering.gif --select_filename 42725-full.png --show_every 10
We can similarly visualize results from the baseline.
python scripts/generate.py --dataset_name yuntian-deng/im2html-100k --model_path models/tables/baseline/model_e100_lr0.0001.pt.100 --output_dir outputs/tables/baseline_visualization --save_intermediate_every 1 --num_batches 1
python scripts/make_gif.py --input_dir outputs/tables/baseline_visualization/ --output_filename imgs/tables_rendering_baseline.gif --select_filename 42725-full.png --show_every 10
To visualize the generation process, we need to first use the following command to save the intermediate images during generation:
python scripts/generate.py --dataset_name yuntian-deng/im2ly-35k-syn --model_path models/music/scheduled_sampling/model_e100_lr0.0001.pt.100 --output_dir outputs/music/scheduled_sampling_visualization --save_intermediate_every 1 --num_batches 1
Next, we put together a gif image from the generated images:
python scripts/make_gif.py --input_dir outputs/music/scheduled_sampling_visualization/ --output_filename imgs/music_rendering.gif --select_filename comp.17342.png --show_every 10
We can similarly visualize results from the baseline.
python scripts/generate.py --dataset_name yuntian-deng/im2ly-35k-syn --model_path models/music/baseline/model_e100_lr0.0001.pt.100 --output_dir outputs/music/baseline_visualization --save_intermediate_every 1 --num_batches 1
python scripts/make_gif.py --input_dir outputs/music/baseline_visualization/ --output_filename imgs/music_rendering_baseline.gif --select_filename comp.17342.png --show_every 10
To visualize the generation process, we need to first use the following command to save the intermediate images during generation:
python scripts/generate.py --dataset_name yuntian-deng/im2smiles-20k --model_path models/molecules/scheduled_sampling/model_e100_lr0.0001.pt.100 --output_dir outputs/molecules/scheduled_sampling_visualization --save_intermediate_every 1 --num_batches 1
Next, we put together a gif image from the generated images:
python scripts/make_gif.py --input_dir outputs/molecules/scheduled_sampling_visualization/ --output_filename imgs/molecules_rendering.gif --select_filename B-1173.png --show_every 10
We can similarly visualize results from the baseline.
python scripts/generate.py --dataset_name yuntian-deng/im2smiles-20k --model_path models/molecules/baseline/model_e100_lr0.0001.pt.100 --output_dir outputs/molecules/baseline_visualization --save_intermediate_every 1 --num_batches 1
python scripts/make_gif.py --input_dir outputs/molecules/baseline_visualization/ --output_filename imgs/molecules_rendering_baseline.gif --select_filename B-1173.png --show_every 10
@inproceedings{
deng2023markuptoimage,
title={Markup-to-Image Diffusion Models with Scheduled Sampling},
author={Yuntian Deng and Noriyuki Kojima and Alexander M Rush},
booktitle={The Eleventh International Conference on Learning Representations },
year={2023},
url={https://openreview.net/forum?id=81VJDmOE2ol}
}