OFA is a unified multimodal pretrained model that unifies modalities (i.e., cross-modality, vision, language) and tasks (e.g., image generation, visual grounding, image captioning, image classification, text generation, etc.) to a simple sequence-to-sequence learning framework. For more information, please refer to our paper: Unifying Architectures, Tasks, and Modalities Through a Simple Sequence-to-Sequence Learning Framework.
We welcome contributions to our project. Feel free to contact us or send us issues/PRs!
We provide online demo via Hugging Face Spaces for you to interact with our pretrained and finetuned models. Below are the links to the demos:
Also we provide Colab notebooks for you to better perceive the procedures. Click here to check them out!
- 2022.3.11: Released the finetuning & inference code/checkpoints for Gigaword.
- 2022.3.08: Released the pretrained checkpoint of OFA-Base in checkpoints.md. To use OFA-Base, you just need to load
ofa_base.ptand change--arch=ofa_largeto--arch=ofa_basein the training scripts. - 2022.3.07: Released the finetuning & inference code/checkpoints for Image Classification, which achieves 85.0 accuracy on ImageNet-1K, slightly better than reported in OFA paper.
- 2022.3.04: Released the finetuning & inference code/checkpoints for Text-to-Image Generation.
- 2022.3.03: Released the finetuning & inference code/checkpoints for SNLI-VE and GLUE.
- 2022.2.22: Released the finetuning & inference code/checkpoints for Visual Question Answering, which can reproduce the reported VQA accuracy in OFA paper (80.02 on test-std). We believe much accuracy improvement can still be achieved based on this codebase. Recently we have used this codebase to achieve a better result (80.45 on test-std) on the VQA Challenge.
- 2022.2.15: Released finetuning & inference code/checkpoints for Referring Expression Comprehension, as well as a Colab notebook and a demo in Hugging Face Spaces.
- 2022.2.10: Released the inference code & finetuned checkpoint for Image captioning, which can reproduce the results on COCO Karparthy test split (149.6 CIDEr). OFA also achieves No.1 on the COCO image captioning online leaderboard [Link] (marked as M6-Team).
- To release finetuning and inference codes for multimodal downstream tasks soon, including image captioning, VQA, text-to-image generation, SNLI-VE, referring expression, comprehension, etc.
- To release finetuning and inference codes for unimodal downstream tasks soon.
- To release codes for pretraining soon.
| Task | Image Captioning | Text-to-Image Generation | VQA | Visual Entailment | Referring Expression Comprehension | |||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Dataset | COCO | COCO | VQA v2 | SNLI-VE | RefCOCO | RefCOCO+ | RefCOCOg | |||||||||
| Split | Kaparthy test | test | test-dev | test-std | val | test | val | test-a | test-b | val | test-a | test-b | val-u | test-u | ||
| Metric | CIDEr | FID | CLIPSIM | IS | Acc. | Acc. | Acc. | |||||||||
| Score | 150.2 | 10.5 | 34.4 | 31.1 | 80.34 | 80.45 | 90.3 | 90.2 | 90.05 | 92.93 | 85.26 | 85.80 | 89.87 | 79.22 | 85.89 | 86.55 |
- python 3.7.4
- pytorch 1.8.1
- torchvision 0.9.1
- JAVA 1.8 (for COCO evaluation)
git clone https://github.com/OFA-Sys/OFA
pip install -r requirements.txtSee datasets.md and checkpoints.md.
To release soon:)
Below we provide methods for finetuning and inference on different downstream tasks. We provide both pretrained OFA-Large and OFA-Base in checkpoints.md. The scripts in this section are prepared for OFA-Large. If you want to use OFA-Base, just modify --restore-file to the path where ofa_base.pt is located and change --arch=ofa_large to --arch=ofa_base. Note that the optimal hyperparameters for the Base model may be different from the Large model and requires proper hyperparameter tuning.
We provide procedures to reproduce our results of image captioning on our paper below.
1. Prepare the Dataset & Checkpoints
Download data (see datasets.md) and models (see checkpoints.md) and put them in the correct directory. The dataset zipfile caption_data.zip contains caption_stage1_train.tsv, caption_stage2_train.tsv, caption_val.tsv and caption_test.tsv. Each image corresponds to only 1 caption in caption_stage1_train.tsv and corresponds to multiple captions in other TSV files (about 5 captions per image). Each line of the dataset represents a caption sample with the following format. The information of uniq-id, image-id, caption, predicted object labels (taken from VinVL, not used), image base64 string are separated by tabs.
162365 12455 the sun sets over the trees beyond some docks. sky&&water&&dock&&pole /9j/4AAQSkZJ....UCP/2Q==
2. Finetuning
Following previous standard practice, we divide the finetuning process of image captioning into two stages. In stage 1, we finetune OFA with cross-entropy loss on 4 NVIDIA-V100 GPUs with 32GB memory (expected to obtain ~139.5 CIDEr on the validation set at this stage). In stage 2, we select the best checkpoint of stage 1 and train with CIDEr optimization on 8 NVIDIA-V100 GPUs (expected to get ~149.4 CIDEr on the validation set at this stage). Note that CIDEr optimization is very unstable and requires careful hyperparameter tuning. If you encounter training errors in the stage2 finetuning, you can increase the batch size or reduce the learning rate. If neither of these works, you can directly set --freeze-resnet to freeze the inner states of batch normalization.
cd run_scripts/caption nohup sh train_caption_stage1.sh > train_stage1.out & # stage 1, train with cross-entropy loss nohup sh train_caption_stage2.sh > train_stage2.out & # stage 2, load the best ckpt of stage1 and train with CIDEr optimization
3. Inference
Run the following commands to get your results and evaluate your model.
cd run_scripts/caption ; sh evaluate_caption.sh # inference & evaluate
This part provides procedures for the finetuning and inference of text-to-image generation. See below.
1. Prepare the Dataset & Checkpoints
Download data (see datasets.md) and models (see checkpoints.md) and put them in the correct directory. The dataset zipfile coco_image_gen.zip contains coco_vqgan_train.tsv, coco_vqgan_dev.tsv and coco_vqgan_full_test.tsv. Each line of the dataset represents a sample with the following format. The information of uniq-id, image-code (produced by vqgan, a list of integers separated by single-whitespaces), lowercased caption are separated by tabs.
1 6674 4336 4532 5334 3251 5461 3615 2469 ...4965 4190 1846 the people are posing for a group photo.
The checkpoint zipfile image_gen_large_best.zip contains image_gen_large_best.pt, vqgan/last.ckpt, vqgan/model.yaml and clip/Vit-B-16.pt.
2. Shuffle the Training Data
(Optional, but achieves better result): If the disk storage is sufficient, we recommend to prepare the shuffled training data for each epoch in advance.
cd dataset/image_gen
ln coco_vqgan_train.tsv coco_vqgan_train_1.tsv
for idx in `seq 1 9`;do shuf coco_vqgan_train_${idx}.tsv > coco_vqgan_train_$[${idx}+1].tsv;done # each file is used for an epoch
3. Finetuning
Following previous practice, we divide the finetuning process of image generating into two stages. In stage 1, we finetune OFA with cross-entropy loss on 4 8-V100-32G-GPU servers (expected to obtain ~32.5+ CLIP Score on the validation set at this stage). In stage 2, we select the last checkpoint of stage 1 and train with CLIP Score optimization on 4 8-V100-32G-GPU servers (expected to obtain ~34.0+ CLIP Score on the validation set at this stage). During the validation, the generated image will be dumped into _GEN_IMAGE_PATH_.
# run on each worker after the distributed and data configs have been correctly set following the guide in train_image_gen_stage1_distributed.sh cd run_scripts/image_gen nohup sh train_image_gen_stage1_distributed.sh # stage 1, train with cross-entropy loss nohup sh train_image_gen_stage2_distributed.sh # stage 2, load the last ckpt of stage1 and train with CLIP Score optimization
4. Inference
Run the command below to generate your images.
cd run_scripts/image_gen ; sh evaluate_image_gen.sh # inference & evaluate (FID, IS and CLIP Score)
Here we provide the finetuning and inference codes to reproduce the VQAv2 result reported in our paper (test-std 80.02). We believe much improvement on accuracy can still be achieved based on this codebase :)
1. Prepare the Dataset & Checkpoints
Download data (see datasets.md) and models (see checkpoints.md) and put them in the correct directory. The dataset zipfile vqa_data.zip is around 100G and the decompressed data costs around 135G disk storage, which contains the training, validation and testing samples together with other necessary data resources. Following common practice, VG-QA samples are also included in the training data. To adapt to the seq2seq paradigm of OFA, we transform original VQA training questions with multiple golden answers into multiple training samples. For the original VQA validation set, we keep around 10k samples for our validation and utilize the other samples for training. Each line of the dataset represents a VQA sample with the following format. The information of question-id, image-id, question, answer (with confidence), predicted object labels (taken from VinVL, slightly brings around +0.1 accuracy improvement), image base64 string are separated by tabs.
79459 79459 is this person wearing shorts? 0.6|!+no house&&short&&...&&sky /9j/4AAQS...tigZ/9k=
2. Shuffle the Training Data
(Optional, but achieves better finetuning accuracy): If the disk storage is sufficient, we recommend to prepare the shuffled training data for each epoch in advance. In our experiments, we use shuffling which brings around +0.3 improvement on VQA accuracy.
cd dataset/vqa_data
ln vqa_train.tsv vqa_train_1.tsv
for idx in `seq 1 9`;do shuf vqa_train_${idx}.tsv > vqa_train_$[${idx}+1].tsv;done # each file is used for an epoch
3. Finetuning
In our experiments, the VQA finetuning is performed on 4 8-A100-GPU servers (with RDMA). Here provides the finetuning script train_vqa_distributed.sh, which supports multi-server distributed training (as well as single-server training). Please refer to the comments in the beginning of the script and set the configs correctly according to your distribution environment. If you have shuffled the training data in the previous step, please correctly specify the training data path following the guide in the script comments. The command should be run on each worker.
# run on each worker after the distributed and data configs have been correctly set following the guide in train_vqa_distributed.sh cd run_scripts/vqa bash train_vqa_distributed.sh
In our experiments, the finetuning costs around 36 hours (for 12 epochs). After each epoch, an evaluation on validation set is performed. The best validation accuracy during finetuning will be around 80.8. The log is saved in ${log_dir}.
4. Inference
We provide 2 types of inference, beam-search (much faster but gets sub-optimal accuracy) and all-candidate evaluation (slower but best accuracy).
For beam-search inference, use the script evaluate_vqa_beam.sh. Refer to the command below. The inference on test set costs around 16 GPU hours. After inference on test set, the result JSON file will be dumped in the ${result_path} defined in the shell script. You can submit the result test_predict.json to EvalAI. Using our released finetuned checkpoint, beam-search inference will get 80.15 validation accuracy, 79.36 test-dev accuracy and 79.48 test-std accuracy (around 0.6 lower than all-candidate evaluation).
cd run_scripts/vqa bash evaluate_vqa_beam.sh val # specify 'val' or 'test'
For all-candidate evaluation, we recommend to use the distributed script evaluate_vqa_allcand_distributed.sh. Please refer to the guide in the script to set the distributed configs before running. The result JSON file will be dumped in the ${result_path} defined in the shell script of rank-0 server. All-candidate evaluation computes scores on all the candidate answers in the VQA dataset, which achieves 80.82 validation accuracy, 79.87 test-dev accuracy and 80.02 test-std accuracy, reproducing our reported results in the paper. However, the inference on test set costs around 1k GPU hours, which is much slower.
# run on each worker after the distributed configs have been correctly set following the guide in evaluate_vqa_allcand_distributed.sh cd run_scripts/vqa bash evaluate_vqa_allcand_distributed.sh val # specify 'val' or 'test'
Here provides procedures for you to prepare data, train, and evaluate your model on visual grounding.
1. Prepare the Dataset & Checkpoints
Download data (see datasets.md) and models (see checkpoints.md) and put them in the correct directory. We provide RefCOCO (split by UNC), RefCOCO+ (split by UNC) and RefCOCOg (split by UMD) datasets. See RefCOCO and Refer for more details. Note that in the original dataset, each region-coord (or bounding box) may corresponds to multiple descriptive texts. We split these texts into multiple samples so that the region-coord in each sample corresponds to only one text. Each line of the processed dataset represents a sample with the following format. The information of uniq-id, image-id, text, region-coord (separated by commas), image base64 string are separated by tabs.
79_1 237367 A woman in a white blouse holding a glass of wine. 230.79,121.75,423.66,463.06 9j/4AAQ...1pAz/9k=
2. Finetuning
Unlike the original paper, we finetune OFA with a drop-path rate of 0.2, and found that training with this hyper-parameter achieves better results. We will update the reported results of the paper later.
cd run_scripts/refcoco nohup sh train_refcoco.sh > train_refcoco.out & # finetune for refcoco nohup sh train_refcocoplus.sh > train_refcocoplus.out & # finetune for refcoco+ nohup sh train_refcocog.sh > train_refcocog.out & # finetune for refcocog
3. Inference
Run the following commands for the evaluation.
cd run_scripts/refcoco ; sh evaluate_refcoco.sh # inference & evaluate for refcoco/refcoco+/refcocog
We provide steps for you to reproduce our results in visual entailment. See the details below.
1. Prepare the Dataset & Checkpoints
Download data (see datasets.md) and models (see checkpoints.md) and put them in the correct directory. Each line of the processed dataset represents a sample with the following format. The information of uniq-id, image-id, image base64 string, hypothesis, caption (or text premise), label are separated by tabs.
252244149.jpg#1r1n 252244149 /9j/4AAQ...MD/2Q== a man in pink and gold is chewing on a wooden toothpick. a man in pink is chewing a toothpick on the subway. neutral
2. Finetuning
In our experiments, the SNLI-VE finetuning is performed on 8 NVIDIA-V100 GPUs with 32GB memory. In this task, we experimented with only a few sets of hyperparameters. We believe that proper hyperparameter tuning can lead to further accuracy improvement.
cd run_scripts/snli_ve nohup sh train_snli_ve.sh > train_snli_ve.out & # finetune for snli_ve
3. Inference
Run the following command to obtain the results.
cd run_scripts/snli_ve ; sh evaluate_snli_ve.sh # inference & evaluate for snli_ve
Here we provide steps for you to finetune and evaluate our model on language understanding tasks. We demonstrate our practice for the GLUE benchmark.
1. Prepare the Dataset & Checkpoints
Download data (see datasets.md) and models (see checkpoints.md) and put them in the correct directory. we provide 7 language understanding datasets from GLUE benchmark, including COLA, MNLI, MRPC, QNLI, QQP, RTE and SST2. More details about these datasets can be found in this link.
2. Finetuning
For each task, we have tried multiple sets of hyperparameters (including learning rate, batch size, training epochs). The results under different sets of hyperparameters can be found in ${log_dir}.
cd run_scripts/glue nohup sh train_cola.sh > train_cola.out & # finetune for cola nohup sh train_mnli.sh > train_mnli.out & # finetune for mnli nohup sh train_mrpc.sh > train_mrpc.out & # finetune for mrpc nohup sh train_qnli.sh > train_qnli.out & # finetune for qnli nohup sh train_qqp.sh > train_qqp.out & # finetune for qqp nohup sh train_rte.sh > train_rte.out & # finetune for rte nohup sh train_sst2.sh > train_sst2.out & # finetune for sst2
We provide the finetuning and inference codes which reproduce 85.0 ImageNet-1K accuracy, slightly better than reported in our paper.
1. Prepare the Dataset & Checkpoints
Download data (see datasets.md) and models (see checkpoints.md) and put them in the correct directory. Our provided data is derived from the original ImageNet-1K (ILSVRC2012 train & validation) dataset and shares the same data split with it. To formulate the classification task into seq2seq paradigm, we use the synset words provided by Caffe as the generation target for each image class. Each line of the processed dataset represents a sample with the following format. The information of image base64 string, classification label (1-indexed, conform to the order in synset_words.txt), synset words of the label are separated by tabs.
_9j_4AAQS...fzX__Z 769 rugby ball
2. Shuffle the Training Data
(Optional, but achieves better finetuning accuracy): If the disk storage is sufficient, we recommend to prepare the shuffled training data for each epoch in advance. In our experiments, we use shuffling which brings around +0.2 improvement on ImageNet-1K accuracy.
cd dataset/imagenet_1k_data
ln imagenet_1k_train.tsv imagenet_1k_train_1.tsv
for idx in `seq 1 9`;do shuf imagenet_1k_train_${idx}.tsv > imagenet_1k_train_$[${idx}+1].tsv;done # each file is used for an epoch one by one
3. Finetuning
In our experiments, the ImageNet-1K finetuning is performed on 2 8-A100-GPU servers (with RDMA). Here provides the finetuning script train_imagenet_distributed.sh, which supports multi-server distributed training (as well as single-server training). Please refer to the comments in the beginning of the script and set the configs correctly according to your distribution environment. If you have shuffled the training data in the previous step, please correctly specify the training data path following the guide in the script comments. The command should be run on each worker. For quick evaluation during finetuning, by default we sample 20% of the original validation split and report accuracy on this subset after each epoch. The accuracy on the validation subset is generally ±0.1 relative to accuracy on the whole validation split.
# run on each worker after the distributed and data configs have been correctly set following the guide in train_imagenet_distributed.sh cd run_scripts/image_classify bash train_imagenet_distributed.sh
In our experiments, the finetuning costs around 80 hours (for 32 epochs). The best accuracy on validation subset during finetuning will be around 85.0. The log is saved in ${log_dir}.
4. Inference
To get the validation accuracy on the whole ImageNet-1K validation set, run the following command. The evaluation costs around 10 GPU hours. The accuracy will be reported in the stdout (expected to be around 85.0).
cd run_scripts/image_classify ; sh evaluate_imagenet.sh # inference & evaluate for imagenet-1k
We provide steps for you to reproduce our results in Gigaword. See the details below.
1. Prepare the Dataset & Checkpoints
Download data (see datasets.md) and models (see checkpoints.md) and put them in the correct directory. The original dataset is taken from UniLM and we organized the data into the tsv format. Each line of the processed dataset represents a sample with the following format. The information of source and target texts are separated by tabs.
factory orders for manufactured goods rose #.# percent in september... us september factory orders up #.# percent
2. Finetuning
Run the following command to train the model.
cd run_scripts/gigaword nohup sh train_gigaword.sh > train_gigaword.out & # finetune for gigaword
3. Inference
Run the following command to obtain the results (~36.43 rougeL).
cd run_scripts/gigaword ; sh evaluate_gigaword.sh # inference & evaluate for gigaword
Below we provide examples of OFA in text-to-image generation and open-ended VQA. Also, we demonstrate its performance in unseen task (Grounded QA) as well as unseen domain (Visual Grounding on images from unseen domains).
Feel free to submit Github issues or pull requests. Welcome to contribute to our project!
To contact us, never hestitate to send an email to [email protected] or [email protected]!
Please cite our paper if you find it helpful :)
@article{wang2022OFA,
title={Unifying Architectures, Tasks, and Modalities Through a Simple Sequence-to-Sequence Learning Framework},
author={Wang, Peng and Yang, An and Men, Rui and Lin, Junyang and Bai, Shuai and Li, Zhikang and Ma, Jianxin and Zhou, Chang and Zhou, Jingren and Yang, Hongxia},
journal={arXiv preprint arXiv:2202.03052},
year={2022}
}