Reformer Enwik8

🚀 Reformer Language Model on Character Level (Trained on enwik8)

This project presents a Reformer-based language model operating at the character level, trained on the enwik8 dataset. The enwik8 dataset, sourced from Wikipedia, is commonly used to evaluate a model's data compression capabilities, such as in the context of the Hutter Prize (https://en.wikipedia.org/wiki/Hutter_Prize).

🚀 Quick Start

The reformer-enwik8 model was pretrained on the first 90M characters of the enwik8 dataset. The text was divided into batches, each containing 65536 characters (2^16). The model weights were retrieved from https://console.cloud.google.com/storage/browser/trax-ml/reformer/enwik8 and converted to Hugging Face's PyTorch ReformerLM model ReformerModelWithLMHead.

✨ Features

• Character-level Operation: This language model works directly on characters, eliminating the need for a tokenizer.
• Encoding and Decoding Functions: Custom functions are provided for encoding and decoding text.

💻 Usage Examples

Basic Usage

import torch

# Encoding
def encode(list_of_strings, pad_token_id=0):
    max_length = max([len(string) for string in list_of_strings])

    # create emtpy tensors
    attention_masks = torch.zeros((len(list_of_strings), max_length), dtype=torch.long)
    input_ids = torch.full((len(list_of_strings), max_length), pad_token_id, dtype=torch.long)

    for idx, string in enumerate(list_of_strings):
        # make sure string is in byte format
        if not isinstance(string, bytes):
            string = str.encode(string)

        input_ids[idx, :len(string)] = torch.tensor([x + 2 for x in string])
        attention_masks[idx, :len(string)] = 1

    return input_ids, attention_masks
    
# Decoding
def decode(outputs_ids):
    decoded_outputs = []
    for output_ids in outputs_ids.tolist():
        # transform id back to char IDs < 2 are simply transformed to ""
        decoded_outputs.append("".join([chr(x - 2) if x > 1 else "" for x in output_ids]))
    return decoded_outputs

Advanced Usage

from transformers import ReformerModelWithLMHead

model = ReformerModelWithLMHead.from_pretrained("google/reformer-enwik8")
encoded, attention_masks = encode(["In 1965, Brooks left IBM to found the Department of"])
decode(model.generate(encoded, do_sample=True, max_length=150))

# gives:
# In 1965, Brooks left IBM to found the Department of Journalism in 1968. IBM had jurisdiction himself in 1980, while Brooks resolved, nevertheless thro

⚠️ Important Note

Language generation using ReformerModelWithLMHead is not optimized yet and is rather slow.