vit-swin-base-224-gpt2-image-captioning

This model is a fine-tuned VisionEncoderDecoder model on 60% of the COCO2014 dataset. It achieves the following results on the testing set:

• Loss: 0.7989
• Rouge1: 53.1153
• Rouge2: 24.2307
• Rougel: 51.5002
• Rougelsum: 51.4983
• Bleu: 17.7765

Model description

The model was initialized on microsoft/swin-base-patch4-window7-224-in22k as the vision encoder, the gpt2 as the decoder.

Intended uses & limitations

You can use this model for image captioning only.

How to use

You can either use the simple pipeline API:

from transformers import pipeline

image_captioner = pipeline("image-to-text", model="Abdou/vit-swin-base-224-gpt2-image-captioning")
# infer the caption
caption = image_captioner("http://images.cocodataset.org/test-stuff2017/000000000019.jpg")[0]['generated_text']
print(f"caption: {caption}")

Or initialize everything for more flexibility:

from transformers import VisionEncoderDecoderModel, GPT2TokenizerFast, ViTImageProcessor
import torch
import os
import urllib.parse as parse
from PIL import Image
import requests

# a function to determine whether a string is a URL or not
def is_url(string):
    try:
        result = parse.urlparse(string)
        return all([result.scheme, result.netloc, result.path])
    except:
        return False
    
# a function to load an image
def load_image(image_path):
    if is_url(image_path):
        return Image.open(requests.get(image_path, stream=True).raw)
    elif os.path.exists(image_path):
        return Image.open(image_path)

# a function to perform inference
def get_caption(model, image_processor, tokenizer, image_path):
    image = load_image(image_path)
    # preprocess the image
    img = image_processor(image, return_tensors="pt").to(device)
    # generate the caption (using greedy decoding by default)
    output = model.generate(**img)
    # decode the output
    caption = tokenizer.batch_decode(output, skip_special_tokens=True)[0]
    return caption

device = "cuda" if torch.cuda.is_available() else "cpu"
# load the fine-tuned image captioning model and corresponding tokenizer and image processor
model = VisionEncoderDecoderModel.from_pretrained("Abdou/vit-swin-base-224-gpt2-image-captioning").to(device)
tokenizer = GPT2TokenizerFast.from_pretrained("Abdou/vit-swin-base-224-gpt2-image-captioning")
image_processor = ViTImageProcessor.from_pretrained("Abdou/vit-swin-base-224-gpt2-image-captioning")

# target image
url = "http://images.cocodataset.org/test-stuff2017/000000000019.jpg"
# get the caption
caption = get_caption(model, image_processor, tokenizer, url)
print(f"caption: {caption}")

Output:

Two cows laying in a field with a sky background.

Training procedure

You can check this guide to learn how this model was fine-tuned.

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 5e-05
• train_batch_size: 64
• eval_batch_size: 64
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• num_epochs: 2

Training results

Total training time: ~5 hours on NVIDIA A100 GPU.

Framework versions

• Transformers 4.26.0
• Pytorch 1.13.1+cu116
• Datasets 2.9.0
• Tokenizers 0.13.2