Paper. The online evaluation pipeline is scheduled to release.
Results
Results are available in answer_save folder. It should be noted that for BLIP2OPT, when using the inference code on Hugging Face, the accuracy of text recognition is high, but the model outputs nothing for the VQA tasks. Conversely, when using the LAVIS library for inference, the accuracy of text recognition is low, while the VQA accuracy is normal. We believe that the inference process of BLIP2OPT still needs to be optimized. In our experiments, we take the maximum value of the two methods as the final result.
Visualization results
Data Download
| Data file | Size |
|---|---|
| text recognition code:iwyn | 1.37GB |
| STVQA End-to-End Task-3 and Training images | 1.88GB |
| ocrVQA | -- |
| textVQA val set | 6.6GB |
| docVQA Task 1 Validation set | 0.8GB |
| ESTVQA | 5.2GB |
| SROIE Task 3 test set | 0.19GB |
| FUNSD | 16MB |
| POIE | 0.43GB |
| HME100K | 0.69GB |
TextVQA, KIE and HME will be updated soon.
We assume that your symlinked data directory has the following structure:
data
|_ IC13_857
|_ IC15_1811
|_ ...
|_ ESTVQA
|_ textVQA
|_ ...
|_ FUNSD
|_ POIE
Usage
eval on all datasets
python eval.py --model_name LLaVA --eval_all
eval on one dataset
python eval.py --model_name LLaVA --eval_textVQA
python eval.py --model_name LLaVA --eval_ocr --ocr_dataset_name "ct80 IIIT5K"
The results will be saved at answer folder.
If you want to add a new model, please write its inference function under the folder "models", and update the get_model function in eval.py. An example inference code is as follows：
import torch
from PIL import Image
from lavis.models import load_model_and_preprocess
from ..process import pad_image, resize_image
class lavis:
def __init__(self, model_name, model_type, device) -> None:
model, vis_processors, txt_processors = load_model_and_preprocess(name = model_name, model_type = model_type, is_eval=True, device=device)
self.model_name = model_name
self.model = model
self.vis_processors = vis_processors
self.txt_processors = txt_processors
self.device = device
def generate(self, image, question, name='resize'):
if 'opt' in self.model_name:
prompt = f'Question: {question} Answer:'
elif 't5' in self.model_name:
prompt = f'Question: {question} Short answer:'
else:
prompt = f'Question: {question} Answer:'
image = Image.open(image).convert("RGB")
if name == "pad":
image = pad_image(image, (224,224))
elif name == "resize":
image = resize_image(image, (224,224))
image = self.vis_processors["eval"](image).unsqueeze(0).to(self.device)
prompt = self.txt_processors["eval"](prompt)
answer = self.model.predict_answers(samples={"image": image, "text_input": prompt}, inference_method="generate", max_len=48, min_len=1)[0]
return answer