Transformer Chatbot with TensorFlow 2

Build an end-to-end chatbot with Transformer in TensorFlow 2. Checkout my tutorial on blog.tensorflow.org.

Updates

• 16 June 2022:
  • Update setup.sh script to install Apple Silicon version of TensorFlow 2.9 (only use this if you're feeling adventurous).
  • Updated the two custom layers, PositionalEncoding and MultiHeadAttentionLayer, to allow model saving via model.save() or tf.keras.models.save_model().
  • train.py showcase how to call model.save() and tf.keras.models.load_model().
• 8 Dec 2020: Updated support to TensorFlow 2.3.1 and TensorFlow Datasets 4.1.0
• 18 Jan 2020: Added notebook with Google Colab TPU support in TensorFlow 2.1.

Packages

• TensorFlow 2.9.1
• TensorFlow Datasets

Setup

• create new anaconda environment and initialize environment chatbot

  conda create -n chatbot python=3.8
  conda activate chatbot
  

• run installation script

  sh setup.sh
  

• Note: the script would install CUDA and cuDNN via conda if installing on a Linux system, or tensorflow-metal for devices with Apple Silicon (Note that there are tons of bugs with TensorFlow on Apple Silicon GPU, e.g. Adam optimizer does not work).

Dataset

• We will use the conversations in movies and TV shows provided by Cornell Movie-Dialogs Corpus, which contains more than 220 thousands conversational exchanges between more than 10k pairs of movie characters, as our dataset.
• We pre-process our dataset in the following order:
  • Extract max_samples conversation pairs into list of questions and answers.
  • Pre-process each sentence by removing special characters in each sentence.
  • Build tokenizer (map text to ID and ID to text) using TensorFlow Datasets SubwordTextEncoder.
  • Tokenize each sentence and add start_token and end_token to indicate the start and end of each sentence.
  • Filter out sentence that has more than max_length tokens.
  • Pad tokenized sentences to max_length
• Check dataset.py implementation.

Model

• check model.py for the implementation of Multi-Headed Attention, Positional Encoding and Transformer.

Run

• check all available flags and hyper-parameters python main.py --help

python train.py --output_dir runs/save_model --batch_size 256 --epochs 50 --max_samples 50000

• the final trained model will be saved to runs/save_model.

Samples

input:		where have you been?
output:		i m not talking about that .

input:		it's a trap!
output:		no , it s not .