Thai Image Captioning

Encoder-decoder style image captioning model using Swin-L and Wangchanberta. Trained on Thai language MSCOCO and IPU24 dataset.

Usage

With VisionEncoderDecoderModel.

from transformers import VisionEncoderDecoderModel, AutoImageProcessor, AutoTokenizer
device = 'cuda'
gen_kwargs = {"max_length": 120, "num_beams": 4}
model_path = 'Natthaphon/thaicapgen-swin-wangchan'
feature_extractor = AutoImageProcessor.from_pretrained(model_path)
tokenizer = AutoTokenizer.from_pretrained(model_path)
model = VisionEncoderDecoderModel.from_pretrained(model_path).to(device)
pixel_values = feature_extractor(images=[Image.open(image_path)], return_tensors="pt").pixel_values
pixel_values = pixel_values.to(device)
output_ids = model.generate(pixel_values, **gen_kwargs)
preds = tokenizer.batch_decode(output_ids, skip_special_tokens=True)

You can also use AutoModel to load it. But this requires trust_remote_code=True.

from transformers import AutoModel
model_path = 'Natthaphon/thaicapgen-swin-wangchan'
model = AutoModel.from_pretrained(model_path, trust_remote_code=True).to(device)

Acknowledgement

This work is partially supported by the Program Management Unit for Human Resources & Institutional Development, Research and Innovation (PMU-B) [Grant number B04G640107]