pali-torch 0.0.9

Pali - PyTorch

PALI: A JOINTLY-SCALED MULTILINGUAL LANGUAGE-IMAGE MODEL

The open source implementation of the Multi-Modality AI model from "PaLI: Scaling Language-Image Learning in 100+ Languages" The model architecture is text -> encoder -> decoder -> logits -> text. The Vision architecture is image -> vit -> embeddings -> encoder -> decoder -> logits -> text

NOTE

• This is the base model architecture, no tokenizer or pretrained weights
• To train, find tokenizer, like tokenmonster and patchify the images to make it compatible with example.py
• We're utilizing an Encoder/Decoder as UL2 and a VIT model that embeds the image which is then injected into the text encoder decoder
• If you would like to help train this model and release it open source please click on the agora banner and join the lab!

🌟 Appreciation

Big bear hugs 🐻💖 to LucidRains for the fab x_transformers and for championing the open source AI cause.

🚀 Install

pip install pali-torch

---

🧙 Usage

import torch  # Importing the torch library for tensor operations
from pali import Pali  # Importing the Pali class from the pali module

model = Pali()  # Creating an instance of the Pali class and assigning it to the variable 'model'

img = torch.randn(1, 3, 256, 256)  # Creating a random image tensor with shape (1, 3, 256, 256)
# The shape represents (batch_size, channels, height, width)

prompt = torch.randint(0, 256, (1, 1024))  # Creating a random text integer tensor with shape (1, 1024)
# The shape represents (batch_size, sequence_length)

output_text = torch.randint(0, 256, (1, 1024))  # Creating a random target text integer tensor with shape (1, 1024)
# The shape represents (batch_size, sequence_length)

out = model.forward(img, prompt, output_text, mask=None)  # Calling the forward method of the 'model' instance
# The forward method takes the image tensor, prompt tensor, output_text tensor, and an optional mask tensor as inputs
# It performs computations and returns the output tensor

print(out)  # Printing the output tensor

Vit Image Embedder

• To embed your images, you can use the vit model:

from PIL import Image
from torchvision import transforms

from pali.model import VitModel


def img_to_tensor(img: str = "pali.png", img_size: int = 256):
    # Load image
    image = Image.open(img)

    # Define a transforms to convert the image to a tensor and apply preprocessing
    transform = transforms.Compose(
        [
            transforms.Lambda(lambda image: image.convert("RGB")),
            transforms.Resize((img_size, img_size)),  # Resize the image to 256x256
            transforms.ToTensor(),  # Convert the image to a tensor,
            transforms.Normalize(
                mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]
            ),  # Normalize the pixel values
        ]
    )

    # apply transforms to the image
    x = transform(image)

    # print(f"Image shape: {x.shape}")

    # Add batch dimension
    x = x.unsqueeze(0)
    print(x.shape)

    return x


# Convert image to tensor
x = img_to_tensor()

# # Initialize model
model = VitModel()

# Forward pass
out = model(x)

# Print output shape
print(out)

---

Datasets Strategy

Dataset strategy as closely shown in the paper.

Here is a markdown table with metadata and links to the datasets on HuggingFace for the datasets used:

Dataset	Description	Size	Languages	Link
WebLI	Large-scale web crawled image-text dataset	10B images, 12B captions	109 languages	Private
CC3M	Conceptual Captions dataset	3M image-text pairs	English	Link
CC3M-35L	Translated version of CC3M to 35 languages	105M image-text pairs	36 languages	Private
VQAv2	VQA dataset built on COCO images	204K images, 1.1M QA pairs	English	Link
VQ2A-CC3M	VQA dataset built from CC3M	3M image-text pairs	English	Private
VQ2A-CC3M-35L	Translated version of VQ2A-CC3M to 35 languages	105M image-text pairs	36 languages	Private
Open Images	Large scale image dataset	9M images with labels	English	Link
Visual Genome	Image dataset with dense annotations	108K images with annotations	English	Link
Object365	Image dataset for object detection	500K images with labels	English	Private

The key datasets used for pre-training PaLI include:

• WebLI: A large-scale multilingual image-text dataset crawled from the web, comprising 10B images and 12B captions in 109 languages.

• CC3M-35L: CC3M Conceptual Captions dataset machine translated into 35 additional languages, totaling 105M image-text pairs in 36 languages.

• VQ2A-CC3M-35L: VQA dataset based on CC3M, also translated into 35 languages.

The model was evaluated on diverse tasks using standard datasets like VQAv2, Open Images, COCO Captions etc. Links and details provided above.

---

---

🎉 Features

• Double the Power: MT5 for text and ViT for images - Pali's the superhero we didn't know we needed! 💪📖🖼️
• Winning Streak: With roots in the tried-and-true MT5 & ViT, success is in Pali's DNA. 🏆
• Ready, Set, Go: No fuss, no muss! Get Pali rolling in no time. ⏱️
• Easy-Peasy: Leave the heavy lifting to Pali and enjoy your smooth sailing. 🛳️

🌆 Real-World Use-Cases

• E-commerce: Jazz up those recs! Understand products inside-out with images & descriptions. 🛍️
• Social Media: Be the smart reply guru for posts with pics & captions. 📱
• Healthcare: Boost diagnostics with insights from images & textual data. 🏥

---

📚 Citation

@inproceedings{chen2022pali,
  title={PaLI: Scaling Language-Image Learning in 100+ Languages},
  author={Chen, Xi and Wang, Xiao},
  booktitle={Conference on Neural Information Processing Systems (NeurIPS)},
  year={2022}
}

Todo

• Make a table of datasets used in paper,
• Provide tokenizer integration
• Provide training script
• Provide usage/inference scripts

---

📜 License

MIT