vit-base-patch16-224-turkish-gpt2-medium

This vision encoder-decoder model utilizes the google/vit-base-patch16-224 as the encoder and ytu-ce-cosmos/turkish-gpt2-medium as the decoder, and it has been fine-tuned on the flickr8k-turkish dataset to generate image captions in Turkish.

Usage

import torch
from transformers import VisionEncoderDecoderModel, ViTImageProcessor, AutoTokenizer
from PIL import Image

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model_id = "atasoglu/vit-base-patch16-224-turkish-gpt2-medium"
img = Image.open("example.jpg")

feature_extractor = ViTImageProcessor.from_pretrained(model_id)
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = VisionEncoderDecoderModel.from_pretrained(model_id)
model.to(device)

features = feature_extractor(images=[img], return_tensors="pt")
pixel_values = features.pixel_values.to(device)

generated_captions = tokenizer.batch_decode(
    model.generate(pixel_values, max_new_tokens=20),
    skip_special_tokens=True,
)

print(generated_captions)