Training FairSeq Transformer on Cloud TPU using PyTorch

This tutorial specifically focuses on the FairSeq version of Transformer, and the WMT 18 translation task, translating English to German.

Objectives

Prepare the dataset.
Run the training job.
Verify the output results.

Costs

This tutorial uses the following billable components of Google Cloud:

Compute Engine
Cloud TPU

To generate a cost estimate based on your projected usage, use the pricing calculator. New Google Cloud users might be eligible for a free trial.

Before you begin

Before starting this tutorial, check that your Google Cloud project is correctly set up.

Sign in to your Google Cloud account. If you're new to Google Cloud, create an account to evaluate how our products perform in real-world scenarios. New customers also get $300 in free credits to run, test, and deploy workloads.

In the Google Cloud console, on the project selector page, select or create a Google Cloud project.

Go to project selector

Make sure that billing is enabled for your Cloud project. Learn how to check if billing is enabled on a project.

In the Google Cloud console, on the project selector page, select or create a Google Cloud project.

Go to project selector

Make sure that billing is enabled for your Cloud project. Learn how to check if billing is enabled on a project.

This walkthrough uses billable components of Google Cloud. Check the Cloud TPU pricing page to estimate your costs. Be sure to clean up resources you create when you've finished with them to avoid unnecessary charges.

Set up a Compute Engine instance

Open a Cloud Shell window.

Open Cloud Shell
Create a variable for your project's ID.
```
export PROJECT_ID=project-id
```
Configure Google Cloud CLI to use the project where you want to create Cloud TPU.
```
gcloud config set project ${PROJECT_ID}
```
The first time you run this command in a new Cloud Shell VM, an Authorize Cloud Shell page is displayed. Click Authorize at the bottom of the page to allow gcloud to make API calls with your credentials.

From the v, launch the Compute Engine resource required for this tutorial.

gcloud compute --project=${PROJECT_ID} instances create transformer-tutorial \
--zone=us-central1-a  \
--machine-type=n1-standard-16  \
--image-family=torch-xla \
--image-project=ml-images  \
--boot-disk-size=200GB \
--scopes=https://www.googleapis.com/auth/cloud-platform

Connect to the new Compute Engine instance.
```
gcloud compute ssh transformer-tutorial --zone=us-central1-a
```
Key Point: From this point on, a prefix of (vm) $ means you should run the command on the Compute Engine VM instance.

Launch a Cloud TPU resource

From the Compute Engine virtual machine, launch a Cloud TPU resource using the following command:

(vm) $ gcloud compute tpus create transformer-tutorial \
--zone=us-central1-a \
--network=default \
--version=pytorch-1.13 \
--accelerator-type=v3-8

Identify the IP address for the Cloud TPU resource.
```
(vm) $ gcloud compute tpus list --zone=us-central1-a
```
The IP address is located under the NETWORK_ENDPOINTS column. You will need this IP address when you create and configure the PyTorch environment.

Download the data

Create a directory, pytorch-tutorial-data to store the model data.
```
(vm) $ mkdir $HOME/pytorch-tutorial-data
```
Navigate to the pytorch-tutorial-data directory.
```
(vm) $ cd $HOME/pytorch-tutorial-data
```

Download the model data.

(vm) $ wget https://dl.fbaipublicfiles.com/fairseq/data/wmt18_en_de_bpej32k.zip

Extract the data.

(vm) $ sudo apt-get install unzip && \
unzip wmt18_en_de_bpej32k.zip

Create and configure the PyTorch environment

Start a conda environment.
```
(vm) $ conda activate torch-xla-1.13
```
Configure environmental variables for the Cloud TPU resource.
```
(vm) $ export TPU_IP_ADDRESS=ip-address; \
export XRT_TPU_CONFIG="tpu_worker;0;$TPU_IP_ADDRESS:8470"
```
Note: The TPU_IP_ADDRESS variable must equal the IP address of the Cloud TPU you identified in when you launched the Cloud TPU resource.

Train the model

To train the model, run the following script:

(vm) $ python /usr/share/torch-xla-1.13/tpu-examples/deps/fairseq/train.py \
  $HOME/pytorch-tutorial-data/wmt18_en_de_bpej32k \
  --save-interval=1 \
  --arch=transformer_vaswani_wmt_en_de_big \
  --max-target-positions=64 \
  --attention-dropout=0.1 \
  --no-progress-bar \
  --criterion=label_smoothed_cross_entropy \
  --source-lang=en \
  --lr-scheduler=inverse_sqrt \
  --min-lr 1e-09 \
  --skip-invalid-size-inputs-valid-test \
  --target-lang=de \
  --label-smoothing=0.1 \
  --update-freq=1 \
  --optimizer adam \
  --adam-betas '(0.9, 0.98)' \
  --warmup-init-lr 1e-07 \
  --lr 0.0005 \
  --warmup-updates 4000 \
  --share-all-embeddings \
  --dropout 0.3 \
  --weight-decay 0.0 \
  --valid-subset=valid \
  --max-epoch=25 \
  --input_shapes 128x64 \
  --num_cores=8 \
  --metrics_debug \
  --log_steps=100

Note: Changing the value for the `input_shapes` hyperparameter may lead to improved performance. for example:

--input_shapes 256x64 512x32

These changes can cause significantly slower initial compiles but faster epoch times after stabilization occurs.

To use these input shapes, you must enable bfloat16 use. To do so, run the following command:

(vm) $ export XLA_USE_BF16=1

Clean up

Perform a cleanup to avoid incurring unnecessary charges to your account after using the resources you created:

Disconnect from the Compute Engine instance, if you have not already done so:
```
(vm) $ exit
```
Your prompt should now be user@projectname, showing you are in the Cloud Shell.
In your Cloud Shell, use the Google Cloud CLI to delete the Compute Engine instance.
```
$  gcloud compute instances delete transformer-tutorial  --zone=us-central1-a
```

Use Google Cloud CLI to delete the Cloud TPU resource.

$  gcloud compute tpus delete transformer-tutorial --zone=us-central1-a

What's next

Try the PyTorch colabs: