Introduction

This directory contains a partial implementation of

Neural Data Augmentation via Example Extrapolation

Kenton Lee, Kelvin Guu, Luheng He, Tim Dozat, Hyung Won Chung*

(* equal contribution)

Getting started

We require Python 3.6+ and the latest versions of mesh_tensorflow and t5:

pip install -r requirements.txt

Data preprocessing

Create a Google Cloud Storage bucket to store the data that we will generate:

export CLINC150_DATA_DIR="gs://<YOUR_BUCKET>/<PATH_TO_DATA_DIR>"

Download the original data from https://github.com/clinc/oos-eval/blob/master/data/data_full.json, and run our preprocessing script to split the data and generate examples for both the Ex2 teacher and the CLINC150 student.

python -m preprocess_clinc150 \
  --data_path=<PATH_TO_DOWNLOADED_data_full.json> \
  --output_dir=${CLINC150_DATA_DIR}

Setting up T5 training

We provide a few minimal commands for using T5. We recommend taking a look at https://github.com/google-research/text-to-text-transfer-transformer for details and following the latest recommendations for how to run T5.

Please set the following environment variables based on your GCP project.

export PROJECT=<YOUR_PROJECT_NAME>
export ZONE=<YOUR_PROJECT_ZONE>
export TPU_NAME=ex2-tpu
export TPU_SIZE=v3-8
export MODELS_DIR="gs://<YOUR_BUCKET>/<PATH_TO_MODELS_DIR>"
export CLINC150_DATA_DIR="gs://<YOUR_BUCKET>/<PATH_TO_DATA_DIR>"
export TOTAL_TEACHER_TRAIN_STEPS=1001000
export TOTAL_STUDENT_TRAIN_STEPS=1020000

Note that the CLINC150_DATA_DIR environment variable is essential because tasks.py uses it to determine where to read the data from.

Launch a Cloud TPU instance (CTPU reference):

ctpu up --name=$TPU_NAME --project=$PROJECT --zone=$ZONE --tpu-size=$TPU_SIZE \
        --tpu-only --noconf

Ex2 teacher

First we train the Ex2 teacher on the many-shot data and use it to infer new synthetic data on the few-shot data.

Teacher training

python3 -m t5.models.mesh_transformer_main \
  --tpu="${TPU_NAME}" \
  --gcp_project="${PROJECT}" \
  --tpu_zone="${ZONE}" \
  --model_dir="${MODELS_DIR}/clinc150_teacher" \
  --module_import="tasks" \
  --gin_file="dataset.gin" \
  --gin_param="utils.tpu_mesh_shape.tpu_topology = '${TPU_SIZE}'" \
  --gin_param="utils.run.sequence_length = {'inputs': 512, 'targets': 128}" \
  --gin_param="utils.run.batch_size = ('sequences_per_batch', 128)" \
  --gin_param="MIXTURE_NAME = 'clinc150_teacher'" \
  --gin_file="gs://t5-data/pretrained_models/t5.1.1.xl/operative_config.gin" \
  --gin_param="run.train_steps = ${TOTAL_TEACHER_TRAIN_STEPS}"

Teacher inference

python3 -m t5.models.mesh_transformer_main \
  --tpu="${TPU_NAME}" \
  --gcp_project="${PROJECT}" \
  --tpu_zone="${ZONE}" \
  --model_dir="${MODELS_DIR}/clinc150_teacher" \
  --module_import="tasks" \
  --gin_file="${MODELS_DIR}/clinc150_teacher/operative_config.gin" \
  --gin_file="infer.gin" \
  --gin_file="sample_decode.gin" \
  --gin_param="input_filename = '${CLINC150_DATA_DIR}/teacher_inputs.txt'"\
  --gin_param="output_filename = '${CLINC150_DATA_DIR}/teacher_predictions.txt'"\
  --gin_param="utils.tpu_mesh_shape.tpu_topology = '${TPU_SIZE}'"\
  --gin_param="utils.run.sequence_length = {'inputs': 512, 'targets': 128}" \
  --gin_param="utils.run.batch_size = ('sequences_per_batch', 128)" \
  --gin_param="Bitransformer.decode.max_decode_length = 128" \
  --gin_param="infer_checkpoint_step = ${TOTAL_TEACHER_TRAIN_STEPS}"

The predictions need to be stitched back together with the original data using our postprocessing script:

python3 -m postprocess_teacher_predictions \
  --teacher_intents_path="${CLINC150_DATA_DIR}/teacher_intents.txt" \
  --teacher_predictions_path="${CLINC150_DATA_DIR}/teacher_predictions.txt-${TOTAL_TEACHER_TRAIN_STEPS}" \
  --original_train_path="${CLINC150_DATA_DIR}/student_train.tfr" \
  --output_train_path="${CLINC150_DATA_DIR}/student_train_with_ex2.tfr"

CLINC150 Student

Finally, we train the CLINC150 student on either the original training data or the augmented training data.

Student training

python3 -m t5.models.mesh_transformer_main \
  --tpu="${TPU_NAME}" \
  --gcp_project="${PROJECT}" \
  --tpu_zone="${ZONE}" \
  --model_dir="${MODELS_DIR}/clinc150_student_with_ex2" \
  --module_import="tasks" \
  --gin_file="dataset.gin" \
  --gin_param="utils.tpu_mesh_shape.tpu_topology = '${TPU_SIZE}'" \
  --gin_param="utils.run.sequence_length = {'inputs': 128, 'targets': 32}" \
  --gin_param="utils.run.batch_size = ('sequences_per_batch', 128)" \
  --gin_param="MIXTURE_NAME = 'clinc150_student_with_ex2'" \
  --gin_file="gs://t5-data/pretrained_models/t5.1.1.xl/operative_config.gin" \
  --gin_param="run.train_steps = ${TOTAL_STUDENT_TRAIN_STEPS}"

Student evaluation

python3 -m t5.models.mesh_transformer_main \
  --tpu="${TPU_NAME}" \
  --gcp_project="${PROJECT}" \
  --tpu_zone="${ZONE}" \
  --model_dir="${MODELS_DIR}/clinc150_student_with_ex2" \
  --gin_file="${MODELS_DIR}/clinc150_student_with_ex2/operative_config.gin" \
  --module_import="tasks" \
  --gin_file="eval.gin" \
  --gin_file="greedy_decode.gin" \
  --gin_param="run.dataset_split = 'fewshot_validation'" \
  --gin_param="utils.tpu_mesh_shape.tpu_topology = '${TPU_SIZE}'" \
  --gin_param="utils.run.sequence_length = {'inputs': 128, 'targets': 32}" \
  --gin_param="utils.run.batch_size = ('sequences_per_batch', 128)" \
  --gin_param="Bitransformer.decode.max_decode_length = 128" \
  --gin_param="MIXTURE_NAME = 'clinc150_student'" \
  --gin_param="eval_checkpoint_step = ${TOTAL_STUDENT_TRAIN_STEPS}"

Replace all instances of clinc150_student_with_ex2 above with clinc150_student to run the baseline that does not use any synthetic data.

Replace fewshot_validation above with overall_validation to compute overall results.

Replicating results

For conceptual simplicity, this directory only contains the minimal code necessary to demonstrate the Ex2 recipe for CLINC150.

In order to replicate the CLINC150 results from the paper, the following additional steps are neccessary:

• Continuous student and teacher evaluation for early stopping.
  • We simply use a fixed number of steps.
• Preprocessing and evaluating on the test set for the student model.
  • We only provide code for evaluating on the validation set.
• Aggregating results by cross-validating the few-shot domain.
  • We only provide code for experimenting with a single few-shot domain.

For reference, in the released implementation, when the banking domain is the few-shot domain, you should expect the following results: