vit-anime-caption

This model is the product of curiosity—imagine a choice that allows you to label anime images!

Disclaimer: The model has been trained on an entirely new dataset. Predictions made by the model prior to 2023 might be off. It's advisable to fine-tune the model according to your specific use case.

Quick setup guide:

from transformers.modeling_outputs import ImageClassifierOutput
from transformers import ViTImageProcessor, ViTForImageClassification
import torch
from PIL import Image

model_name_or_path = "Ojimi/vit-anime-caption"
processor = ViTImageProcessor.from_pretrained(model_name_or_path)
model = ViTForImageClassification.from_pretrained(model_name_or_path)
threshold = 0.3

device = torch.device('cuda')

image = Image.open(YOUR_IMAGE_PATH)

inputs = processor(image, return_tensors='pt')

model.to(device=device)
model.eval()


with torch.no_grad():
    pixel_values = inputs['pixel_values'].to(device=device)

    outputs : ImageClassifierOutput = model(pixel_values=pixel_values)

    logits = outputs.logits  # The raw scores before applying any activation
    sigmoid = torch.nn.Sigmoid()  # Sigmoid function to convert logits to probabilities
    logits : torch.FloatTensor = sigmoid(logits)  # Applying sigmoid activation

    predictions = []  # List to store predictions

    for idx, p in enumerate(logits[0]):
        if p > threshold:  # Applying a threshold of 0.3 to consider a class prediction
            predictions.append((model.config.id2label[idx], p.item()))  # Storing class label and probability

for tag in predictions:
    print(tag)  

Why the Sigmoid?

• Sigmoid turns boring scores into fun probabilities, so you can use thresholds and find more cool tags.
• It's like a wizard turning regular stuff into magic potions!

Training guide