convert-tf

ConveRT (Conversational Representations from Transformers) TF model based on https://arxiv.org/pdf/1911.03688.pdf.

Usage

Training

Below is the code to train a new model with the same default settings as described in the paper.

from model import convert

model = convert.get_compiled_model(vocab_path, max_steps)
model.fit(inputs)

where:

• vocab_path (str): path to any WordPiece compatible vocab file with "##" suffix indicator, or build one by running tools/build_vocab.py on the train input corpus.
• max_steps (int): number of training steps before stopping learning rate decay.
• inputs: a dictionary mapping "context" and "response" keys to the corresponding input string array/tensors or a tf.data.Dataset of the same format (check Keras Model.fit API).

Input specs:

{
  "context": tf.TensorSpec(shape=(None,), dtype=tf.string),
  "response": tf.TensorSpec(shape=(None,), dtype=tf.string)
}

A sample training code can be found in tools/test_train.py.

Inference

An ExportSavedModel callback is provided in model/utils.py to export the model for inference.

The saved model can then be loaded in the following ways for inference:

import tensorflow as tf
import tensorflow_text as tf_text
from model import convert

inputs = {"context": ..., "response": ...}

# Option 1
model = tf.saved_model.load(saved_model)
serve_fn = model.signatures["serve"]
output = serve_fn(context=inputs["context"], response=inputs["response"])

# Option 2
model = tf.keras.models.load_model(
  saved_model,
  custom_objects={"ConveRT": convert.ConveRT}
)
output = model(inputs)

Note

1. Model supports single context currently.
2. Only 1 OOV bucket.