genn 0.7.7

GeNN (Generative Neural Networks) is a high-level interface for text applications using PyTorch

GeNN

GeNN (generative neural networks) is a high-level interface for text applications using PyTorch RNN's.

Features

1. Preprocessing:
   • Parsing txt, json, and csv files.
   • NLTK, regex and spacy tokenization support.
   • GloVe and fastText pretrained embeddings, with the ability to fine-tune for your data.
2. Architectures and customization:
   • GPT2 with small, medium, and large variants.
   • LSTM and GRU, with variable size.
   • Variable number of layers and batches.
   • Dropout.
3. Text generation:
   • Random seed sampling from the n first tokens in all instances, or the most frequent token.
   • Top-K sampling for next token prediction with variable K.
   • Nucleus sampling for next token prediction with variable probability threshold.
4. Text Summarization:
   • All GPT2 variants can be trained to perform text summarization.

Getting started

How to install

pip install genn

Prerequisites

• PyTorch 1.4.0

pip install torch==1.4.0

• Pytorch Transformers

pip install pytorch_transformers

• NumPy

pip install numpy

• fastText

pip install fasttext

Use the package manager pip to install genn.

Usage

Text Generation:

RNNs (You can switch LSTMGenerator with GRUGenerator:

from genn import Preprocessing, LSTMGenerator, GRUGenerator
#LSTM example
ds = Preprocessing("data.txt")
gen = LSTMGenerator(ds, nLayers = 2,
                        batchSize = 16,
                        embSize = 64,
                        lstmSize = 16,
                        epochs = 20)
			
#Train the model
gen.run()

# Generate 5 new documents
print(gen.generate_document(5))

GPT2 Generator:

#GPT2 example
gen = GPT2("data.txt",
 	    taskToken = "Movie:",
	    epochs = 7,
	    variant = "medium")
#Train the model
gen.run()

#Generate 10 new documents
print(gen.generate_document(10))

Text Summarization:

GPT2 Summarizer:

#GPT2 Summarizer example
from genn import GPT2Summarizer
summ = GPT2Summarizer("data.txt",
			epochs=3,
			batch_size=8)

#Train the model
summ.run()

#Create 5 summaries of a source document
src_doc = "This is the source document to summarize"
print(summ.summarize_document(n=5, source = src_doc))

For more examples on how to use Preprocessing, please refer to this file.

For more examples on how to use LSTMGenerator and GRUGenerator, please refer to this file.

For more examples on how to use GPT2, please refer to this file.

For more examples on how to use GPT2Summarizer, please refer to this file.

Contributing

Pull requests are welcome. For major changes, please open an issue first to discuss what you would like to change.

License

Distributed under the MIT License. See LICENSE for more information.