Training transformer on Cloud TPU (TF 2.x)

Transformer is a neural network architecture that solves sequence to sequence problems using attention mechanisms. Unlike traditional neural seq2seq models, Transformer does not involve recurrent connections. The attention mechanism learns dependencies between tokens in two sequences. Since attention weights apply to all tokens in the sequences, the Transformer model is able to easily capture long-distance dependencies.

Transformer's overall structure follows the standard encoder-decoder pattern. The encoder uses self-attention to compute a representation of the input sequence. The decoder generates the output sequence one token at a time, taking the encoder output and previous decoder-outputted tokens as inputs.

The model also applies embeddings on the input and output tokens, and adds a constant positional encoding. The positional encoding adds information about the position of each token.

Warning: This tutorial uses a third-party dataset. Google provides no representation, warranty, or other guarantees about the validity, or any other aspects of this dataset.

Costs

This tutorial uses billable components of Google Cloud, including:

• Compute Engine
• Cloud TPU

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

Before you begin

If you plan to train on a TPU Pod slice, please make sure you go over this document that explains the special considerations when training on a Pod slice.

Before starting this tutorial, follow the steps below to check that your Google Cloud project is correctly set up.

Important: To reduce network latency and network costs, set up your Compute Engine VM and your Cloud TPU node in the same time zone.

1. Open a Cloud Shell window.

   Open Cloud Shell

2. Create a variable for your project's ID.

   export PROJECT_ID=project-id
   

3. Configure gcloud command-line tool 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 GCP API calls with your credentials.

4. Create a Service Account for the Cloud TPU project.

   gcloud beta services identity create --service tpu.googleapis.com --project $PROJECT_ID
   

   The command returns a Cloud TPU Service Account with following format:

   [e-mail]
   

5. Create a Cloud Storage bucket using the following command:

   Note: In the following command, replace bucket-name with the name you want to assign to your bucket.

   gsutil mb -p ${PROJECT_ID} -c standard -l europe-west4 -b on gs://bucket-name
   

   This Cloud Storage bucket stores the data you use to train your model and the training results. The gcloud tool used in this tutorial sets up default permissions for the Cloud TPU Service Account you set up in the previous step. If you want finer-grain permissions, review the access level permissions.

6. Launch a Compute Engine VM using the gcloud command.

   $ gcloud compute tpus execution-groups create \
    --vm-only \
    --name=transformer-tutorial \
    --zone=europe-west4-a \
    --disk-size=300 \
    --machine-type=n1-standard-8 \
    --tf-version=2.4.1
   

   Command flag descriptions

   vm-only

   Create a VM only. By default the gcloud compute tpus execution-groups command creates a VM and a Cloud TPU.

   name

   The name of the Cloud TPU to create.

   zone

   The zone where you plan to create your Cloud TPU.

   disk-size

   The size of the hard disk in GB of the VM created by the gcloud compute tpus execution-groups command.

   machine-type

   The machine type of the Compute Engine VM to create.

   tf-version

   The version of Tensorflow gcloud compute tpus execution-groups installs on the VM.

   For more information on the gcloud command, see the gcloud Reference.

7. The configuration you specified appears. Enter y to approve or n to cancel.

   Note: The first time you run gcloud compute tpus execution-groups on a project it takes about 5 minutes to perform startup tasks such as SSH key propagation and API turnup.

   When the gcloud compute tpus execution-groups command has finished executing, verify that your shell prompt has changed from username@projectname to username@vm-name. This change shows that you are now logged into your Compute Engine VM.

   Note: If you are not connected to the Compute Engine instance, you can connect by running the following command:

   gcloud compute ssh transformer-tutorial --zone=europe-west4-a
   

   As you continue these instructions, run each command that begins with (vm)$ in your Compute Engine instance.

Generate the training dataset

On your Compute Engine VM:

1. Create the following environment variables. Replace bucket-name with your bucket name:

   (vm)$ export STORAGE_BUCKET=gs://bucket-name
   

   (vm)$ export GCS_DATA_DIR=${STORAGE_BUCKET}/data/transformer
   (vm)$ export PARAM_SET=big
   (vm)$ export MODEL_DIR=${STORAGE_BUCKET}/transformer/model_${PARAM_SET}
   (vm)$ export PYTHONPATH="$PYTHONPATH:/usr/share/models"
   (vm)$ export DATA_DIR=${HOME}/transformer/data
   (vm)$ export VOCAB_FILE=${DATA_DIR}/vocab.ende.32768
   

2. Change directory to the training directory:

   (vm)$ cd /usr/share/models/official/nlp/transformer

   Note: if you are running TensorFlow 2.1, use the path: /usr/share/models/official/transformer

3. Set up the following dataset environment variables:

   (vm)$ export GCS_DATA_DIR=${STORAGE_BUCKET}/data/transformer
   (vm)$ export MODEL_DIR=${STORAGE_BUCKET}/transformer/model_${PARAM_SET}
   

4. Download and prepare the datasets

   (vm)$ python3 data_download.py --data_dir=${DATA_DIR}
   (vm)$ gsutil cp -r ${DATA_DIR} ${GCS_DATA_DIR}
   

   Note: if you are running TensorFlow 2.1, change to the v2 directory: cd /usr/share/models/official/transformer/v2

   data_download.py downloads and preprocesses the training and evaluation WMT datasets. After the data is downloaded and extracted, the training data is used to generate a vocabulary of subtokens. The evaluation and training strings are tokenized, and the resulting data is sharded, shuffled, and saved as TFRecords.

   1.75GB of compressed data is downloaded. In total, the raw files (compressed, extracted, and combined files) take up 8.4GB of disk space. The resulting TFRecord and vocabulary files are 722MB. The script takes around 40 minutes to run, with the bulk of the time spent downloading and ~15 minutes spent on preprocessing.

Train an English-German translation model on a single Cloud TPU

Note: If you want to monitor the model's output and performance, follow the guide to setting up TensorBoard.

Run the following commands on your Compute Engine VM:

1. Run the following command to create your Cloud TPU.

   (vm)$ gcloud compute tpus execution-groups create --tpu-only \
     --accelerator-type=v3-8  \
     --zone=europe-west4-a \
     --tf-version=2.4.1 \
     --name=transformer-tutorial
   

   Command flag descriptions

   project

   Your GCP project ID

   tpu-only

   Create a Cloud TPU only. By default the gcloud compute tpus execution-groups command creates a VM and a Cloud TPU.

   tpu-size

   Specifies the type of Cloud TPU, for example v3-8.

   zone

   The zone where you plan to create your Cloud TPU. This should be the same zone you used for the Compute Engine VM. For example, europe-west4-a.

   tf-version

   The version of Tensorflow ctpu installs on the VM.

   name

   The name of the Cloud TPU to create.

   The configuration you specified appears. Enter y to approve or n to cancel.

   You will see a message: Operation success; not ssh-ing to Compute Engine VM due to --tpu-only flag. Since you previously completed SSH key propagation, you can ignore this message.

2. Set the Cloud TPU name variable. This will either be a name you specified with the --name parameter to gcloud compute tpus execution-groups or the default, your username:

   (vm)$ export TPU_NAME=transformer-tutorial
   

3. Run the training script:

   (vm)$ python3 transformer_main.py \
       --tpu=${TPU_NAME} \
       --model_dir=${MODEL_DIR} \
       --data_dir=${GCS_DATA_DIR} \
       --vocab_file=${GCS_DATA_DIR}/vocab.ende.32768 \
       --bleu_source=${GCS_DATA_DIR}/newstest2014.en \
       --bleu_ref=${GCS_DATA_DIR}/newstest2014.de \
       --batch_size=6144 \
       --train_steps=2000 \
       --static_batch=true \
       --use_ctl=true \
       --param_set=big \
       --max_length=64 \
       --decode_batch_size=32 \
       --decode_max_length=97 \
       --padded_decode=true \
       --distribution_strategy=tpu

   Command flag descriptions

   tpu

   The name of the Cloud TPU. This is set by specifying the environment variable (TPU_NAME).

   model_dir

   The Cloud Storage bucket where checkpoints and summaries are stored during training. You can use an existing folder to load previously generated checkpoints created on a TPU of the same size and TensorFlow version.

   data_dir

   The Cloud Storage path of training input. It is set to the fake_imagenet dataset in this example.

   vocab_file

   A file that contains the vocabulary for translation.

   bleu_source

   A file that contains source sentences for translation.

   bleu_ref

   A file that contains the reference for the translation sentences.

   train_steps

   The number of steps to train the model. One step processes one batch of data. This includes both a forward pass and back propagation.

   batch_size

   The training batch size.

   static_batch

   Specifies whether the batches in the dataset has static shapes.

   use_ctl

   Specifies whether the script runs with a custom training loop.

   param_set

   The parameter set to use when creating and training the model. The parameters define the input shape, model configuration, and other settings.

   max_length

   The maximum length of an example in the dataset.

   decode_batch_size

   The global batch size used for Transformer auto-regressive decoding on a Cloud TPU.

   decode_max_length

   The maximum sequence length of the decode/eval data. This is used by the Transformer auto-regressive decoding on a Cloud TPU to minimize the amount of required data padding.

   padded_decode

   Specifies whether the auto-regressive decoding runs with input data padded to the decode_max_length. Tor TPU/XLA-GPU runs, this flag must be set due to the static shape requirement.

   distribution_strategy

   To train the ResNet model on a Cloud TPU, set distribution_strategy to tpu.

By default, the model will evaluate after every 2000 steps. In order to train to convergence, change train_steps to 200000. You can increase the number of training steps or specify how often to run evaluations by setting these parameters:

• --train_steps: Sets the total number of training steps to run.
• --steps_between_evals: Number of training steps to run between evaluations.

Training and evaluation takes approximately 7 minutes on a v3-8 Cloud TPU. When the training and evaluation complete, a message similar to the following appears:

INFO:tensorflow:Writing to file /tmp/tmpf2gn8vpa
I1125 21:22:30.734232 140328028010240 translate.py:182] Writing to file /tmp/tmpf2gn8vpa
I1125 21:22:42.785628 140328028010240 transformer_main.py:121] Bleu score (uncased): 0.01125154594774358
I1125 21:22:42.786558 140328028010240 transformer_main.py:122] Bleu score (cased): 0.01123994225054048

Compute BLEU score during model evaluation

Use these flags to compute the BLEU when the model evaluates:

• --bleu_source: Path to file containing text to translate.
• --bleu_ref: Path to file containing the reference translation.

From here, you can either conclude this tutorial and clean up your GCP resources, or you can further explore running the model on a Cloud TPU Pod.

Scaling your model with Cloud TPU Pods

You can get results faster by scaling your model with Cloud TPU Pods. The fully supported Transformer model can work with the following Pod slices:

• v2-32
• v3-32

1. Delete the Cloud TPU resource you created for training the model on a single device.

   (vm)$ gcloud compute tpus execution-groups delete transformer-tutorial \
     --zone=europe-west4-a \
     --tpu-only

2. Run the gcloud compute tpus execution-groups command, using the accelerator-type parameter to specify the Pod slice you want to use. For example, the following command uses a v3-32 Pod slice.

   (vm)$ gcloud compute tpus execution-groups create --name=transformer-tutorial \
     --accelerator-type=v2-32  \
     --zone=europe-west4-a \
     --tf-version=2.4.1 \
     --tpu-only
   

   Command flag descriptions

   name

   The name of the Cloud TPU to create.

   accelerator-type

   The type of the Cloud TPU to create.

   zone

   The zone where you plan to create your Cloud TPU.

   tf-version

   The version of Tensorflow gcloud installs on the VM.

   tpu-only

   Create a Cloud TPU only. By default the gcloud command creates a VM and a Cloud TPU.

   Note: If you are not connected to the Compute Engine instance, connect by running the following command:

   gcloud compute ssh transformer-tutorial --zone=europe-west4-a
   

3. Export the TPU name:

   (vm)$ export TPU_NAME=transformer-tutorial
   

4. Export the model directory variable:

   (vm)$ export MODEL_DIR=${STORAGE_BUCKET}/transformer/model_${PARAM_SET}_pod
   

5. Change directory to the training directory:

   (vm)$ cd /usr/share/models/official/nlp/transformer
   

   Note: If you are using TensorFlow 2.1, use the path: /usr/share/models/official/transformer/v2

6. Run the Pod training script:

   (vm)$ python3 transformer_main.py \
        --tpu=${TPU_NAME} \
        --model_dir=${MODEL_DIR} \
        --data_dir=${GCS_DATA_DIR} \
        --vocab_file=${GCS_DATA_DIR}/vocab.ende.32768 \
        --bleu_source=${GCS_DATA_DIR}/newstest2014.en \
        --bleu_ref=${GCS_DATA_DIR}/newstest2014.de \
        --batch_size=24576 \
        --train_steps=2000 \
        --static_batch=true \
        --use_ctl=true \
        --param_set=big \
        --max_length=64 \
        --decode_batch_size=32 \
        --decode_max_length=97 \
        --padded_decode=true \
        --steps_between_evals=2000 \
        --distribution_strategy=tpu
   

   Command flag descriptions

   tpu

   The name of the Cloud TPU. This is set by specifying the environment variable (TPU_NAME).

   model_dir

   The Cloud Storage bucket where checkpoints and summaries are stored during training. You can use an existing folder to load previously generated checkpoints created on a TPU of the same size and TensorFlow version.

   data_dir

   The Cloud Storage path of training input. It is set to the fake_imagenet dataset in this example.

   vocab_file

   A file that contains the vocabulary for translation.

   bleu_source

   A file that contains source sentences for translation.

   bleu_ref

   A file that contains the reference for the translation sentences.

   batch_size

   The training batch size.

   train_steps

   The number of steps to train the model. One step processes one batch of data. This includes both a forward pass and back propagation.

   static_batch

   Specifies whether the batches in the dataset has static shapes.

   use_ctl

   Specifies whether the script runs with a custom training loop.

   param_set

   The parameter set to use when creating and training the model. The parameters define the input shape, model configuration, and other settings.

   max_length

   The maximum length of an example in the dataset.

   decode_batch_size

   The global batch size used for Transformer auto-regressive decoding on a Cloud TPU.

   decode_max_length

   The maximum sequence length of the decode/eval data. This is used by the Transformer auto-regressive decoding on a Cloud TPU to minimize the amount of required data padding.

   padded_decode

   Specifies whether the auto-regressive decoding runs with input data padded to the decode_max_length. Tor TPU/XLA-GPU runs, this flag must be set due to the static shape requirement.

   steps_between_evals

   The number of training steps to run between evaluations.

   distribution_strategy

   To train the ResNet model on a TPU, set distribution_strategy to tpu.

This training script trains for 2000 steps and runs evaluation every 2000 steps. This particular training and evaluation takes approximately 8 minutes on a v2-32 Cloud TPU Pod.

In order to train to convergence, change train_steps to 200000. You can increase the number of training steps or specify how often to run evaluations by setting these parameters:

• --train_steps: Sets the total number of training steps to run.
• --steps_between_evals: Number of training steps to run between evaluations.

When the training and evaluation complete, a message similar to the following appears:

0509 00:27:59.984464 140553148962624 translate.py:184] Writing to file /tmp/tmp_rk3m8jp
I0509 00:28:11.189308 140553148962624 transformer_main.py:119] Bleu score (uncased): 1.3239131309092045
I0509 00:28:11.189623 140553148962624 transformer_main.py:120] Bleu score (cased): 1.2855342589318752

Cleaning up

To avoid incurring charges to your Google Cloud account for the resources used in this tutorial, either delete the project that contains the resources, or keep the project and delete the individual resources.

1. Disconnect from the Compute Engine instance, if you have not already done so:

  (vm)$ exit
  

Your prompt should now be username@projectname, showing you are in the Cloud Shell.

1. In your Cloud Shell, use the following command to delete your Compute Engine VM and Cloud TPU:

   $ gcloud compute tpus execution-groups delete transformer-tutorial \
     --zone=europe-west4-a
   

2. Verify the resources have been deleted by running gcloud compute tpus execution-groups list. The deletion might take several minutes. A response like the one below indicates your instances have been successfully deleted.

   $ gcloud compute tpus execution-groups list \
    --zone=europe-west4-a
   

   You should see an empty list of TPUs like the following:

      NAME             STATUS
   

3. Delete your Cloud Storage bucket using gsutil as shown below. Replace bucket-name with the name of your Cloud Storage bucket.

   Caution: All training data will be lost when you delete your bucket, so only do this step when you are finished running the tutorial.

   $ gsutil rm -r gs://bucket-name
   

What's next

In this tutorial you have trained the Transformer model using a sample dataset. The results of this training are (in most cases) not usable for inference. To use a model for inference you can train the data on a publicly available dataset or your own data set. Models trained on Cloud TPUs require datasets to be in TFRecord format.

You can use the dataset conversion tool sample to convert an image classification dataset into TFRecord format. If you are not using an image classification model, you will have to convert your dataset to TFRecord format yourself. For more information, see TFRecord and tf.Example

Hyperparameter tuning

To improve the model's performance with your dataset, you can tune the model's hyperparameters. You can find information about hyperparameters common to all TPU supported models on GitHub. Information about model-specific hyperparameters can be found in the source code for each model. For more information on hyperparameter tuning, see Overview of hyperparameter tuning, Using the Hyperparameter tuning service and Tune hyperparameters.

Inference

Once you have trained your model you can use it for inference (also called prediction). AI Platform is a cloud-based solution for developing, training, and deploying machine learning models. Once a model is deployed, you can use the AI Platform Prediction service.

• Learn more about ctpu, including how to install it on a local machine.
• Explore more Tensor2Tensor models for TPU.
• Experiment with more TPU samples.
• Explore the TPU tools in TensorBoard.