rrivera1849/LUAR-CRUD

Author Style Representations using LUAR.

The LUAR training and evaluation repository can be found here.

This particular model was trained on a subsample of the Pushshift Reddit Dataset (5 million users) for comments published between January 2015 and October 2019 by authors publishing at least 100 comments during that period.

Usage

from transformers import AutoModel, AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("rrivera1849/LUAR-CRUD")
model = AutoModel.from_pretrained("rrivera1849/LUAR-CRUD")

# we embed `episodes`, a colletion of documents presumed to come from an author
# NOTE: make sure that `episode_length` consistent across `episode`
batch_size = 3
episode_length = 16
text = [
    ["Foo"] * episode_length,
    ["Bar"] * episode_length,
    ["Zoo"] * episode_length,
]
text = [j for i in text for j in i]
tokenized_text = tokenizer(
    text, 
    max_length=32,
    padding="max_length", 
    truncation=True,
    return_tensors="pt"
)
# inputs size: (batch_size, episode_length, max_token_length)
tokenized_text["input_ids"] = tokenized_text["input_ids"].reshape(batch_size, episode_length, -1)
tokenized_text["attention_mask"] = tokenized_text["attention_mask"].reshape(batch_size, episode_length, -1)
print(tokenized_text["input_ids"].size())       # torch.Size([3, 16, 32])
print(tokenized_text["attention_mask"].size())  # torch.Size([3, 16, 32])

out = model(**tokenized_text)
print(out.size())   # torch.Size([3, 512])

# to get the Transformer attentions:
out, attentions = model(**tokenized_text, output_attentions=True)
print(attentions[0].size())     # torch.Size([48, 12, 32, 32])

Usage (Batch)

Here's a more fleshed out example showing how to run LUAR across many batches of data:

import numpy as np
import torch
from termcolor import cprint
from transformers import AutoModel, AutoTokenizer
from tqdm import tqdm

def generate_data(num_batches: int = 100, batch_size: int = 32, num_samples_per_author: int = 16):
    """
    Generator that produces dummy data for testing.
    
    Args:
        num_batches (int): Total number of batches to yield.
        batch_size (int): Number of authors per batch.
        num_samples_per_author (int): Number of text samples per author.
        
    Yields:
        list: A batch of data structured as a list of lists of strings.
              Shape: (batch_size, num_samples_per_author)
    """
    s = "This is an example string."
    for batch in tqdm(range(num_batches)):
        # Create a batch where each element is a list of 's' repeated 'num_samples_per_author' times
        yield [[s] * num_samples_per_author for _ in range(batch_size)]

def flatten(l):
    """
    Helper function to flatten a 2D list into a 1D list.
    
    Args:
        l (list): List of lists.
        
    Returns:
        list: Flattened list.
    """
    return [item for sublist in l for item in sublist]

def main():
    cprint("Starting LUAR-MUD example script...", 'magenta')

    # --- Model Loading ---
    cprint("Loading model 'rrivera1849/LUAR-MUD'...", 'blue')
    # trust_remote_code=True is required for custom model architectures like LUAR-MUD
    model = AutoModel.from_pretrained("rrivera1849/LUAR-MUD", trust_remote_code=True)
    
    model.eval()
    
    # Check for CUDA availability and move model to appropriate device
    device = "cuda" if torch.cuda.is_available() else "cpu"
    cprint(f"Moving model to device: {device}", 'yellow')
    model.to(device)

    # --- Tokenizer Loading ---
    cprint("Loading tokenizer...", 'blue')
    tokenizer = AutoTokenizer.from_pretrained("rrivera1849/LUAR-MUD", trust_remote_code=True)

    # --- Configuration ---
    num_batches = 100
    batch_size = 32
    num_samples_per_author = 16
    max_length = 512

    cprint("\nConfiguration:", 'cyan')
    print(f"  Batch Size: {batch_size}")
    print(f"  Samples per Author: {num_samples_per_author}")
    print(f"  Max Length: {max_length}")
    print(f"  Device: {device}\n")

    all_outputs = []
    
    cprint("Starting inference loop...", 'green')
    
    # context manager for disabling gradient calculation to save memory/compute
    with torch.inference_mode():
        for i, batch in enumerate(generate_data(num_batches=num_batches, batch_size=batch_size, num_samples_per_author=num_samples_per_author)):
            if (i + 1) % 10 == 0:
                print(f"  Processing batch {i + 1}...")

            # Flatten the batch structure for tokenization: 
            # (batch_size, num_samples) -> (batch_size * num_samples)
            batch = flatten(batch)
            
            # Tokenize the flattened batch
            inputs = tokenizer(batch, return_tensors="pt", padding=True, max_length=max_length, truncation=True)
            
            # Move inputs to the same device as the model
            inputs = inputs.to(device)
            
            # Reshape input_ids and attention_mask to match the model's expected 3D input:
            # (batch_size, num_samples_per_author, sequence_length)
            inputs["input_ids"] = inputs["input_ids"].reshape(batch_size, num_samples_per_author, -1)
            inputs["attention_mask"] = inputs["attention_mask"].reshape(batch_size, num_samples_per_author, -1)
            
            # Forward pass through the model
            outputs = model(**inputs)
            
            # Move outputs back to CPU and convert to numpy for storage
            all_outputs.append(outputs.cpu().numpy())

    # Concatenate all batch results into a single array
    # axis=0 corresponds to the batch dimension
    all_outputs = np.concatenate(all_outputs, axis=0)
    
    cprint("\nInference complete!", 'green')
    cprint(f"Final output shape: {all_outputs.shape}", attrs=['bold'])

if __name__ == "__main__":
    main()

Citing & Authors

If you find this model helpful, feel free to cite our publication.

@inproceedings{uar-emnlp2021,
  author    = {Rafael A. Rivera Soto and Olivia Miano and Juanita Ordonez and Barry Chen and Aleem Khan and Marcus Bishop and Nicholas Andrews},
  title     = {Learning Universal Authorship Representations},
  booktitle = {EMNLP},
  year      = {2021},
}

License

LUAR is distributed under the terms of the Apache License (Version 2.0).

All new contributions must be made under the Apache-2.0 licenses.