Running Inception on Cloud TPU

This tutorial shows you how to train the Inception model on Cloud TPU.

Disclaimer

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.

Model description

Inception v3 is a widely-used image recognition model that can attain significant accuracy. The model is the culmination of many ideas developed by multiple researchers over the years. It is based on the original paper: "Rethinking the Inception Architecture for Computer Vision" by Szegedy, et. al.

The model has a mixture of symmetric and asymmetric building blocks, including:

• convolutions
• average pooling
• max pooling
• concats
• dropouts
• fully connected layers

Loss is computed via Softmax.

The following picture shows the model at a high level:

You can find more information about the model at GitHub.

The model is built using the high-level Estimator API.

This API greatly simplifies model creation by encapsulating most low-level functions, allowing users to focus on model development, not the inner workings of the underlying hardware that runs things.

Objectives

• Create a Cloud Storage bucket to hold your dataset and model output.
• Run the training job.
• Verify the output results.

Costs

This tutorial uses billable components of Google Cloud, including:

• Compute Engine
• Cloud TPU
• Cloud Storage

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

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

1. Sign in to your Google Account.

   If you don't already have one, sign up for a new account.

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

   Note: If you don't plan to keep the resources that you create in this procedure, create a project instead of selecting an existing project. After you finish these steps, you can delete the project, removing all resources associated with the project.

   Go to the project selector page

3. Make sure that billing is enabled for your Cloud project. Learn how to confirm that billing is enabled for your 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 your resources

This section provides information on setting up Cloud Storage, VM, and Cloud TPU resources for tutorials.

Important: Set up all resources in the same region/zone to reduce network latency and network costs.

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. Replace bucket-name with a name for your bucket.

   gsutil mb -p ${PROJECT_ID} -c standard -l us-central1 -b on gs://bucket-name
   

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

   The bucket location must be in the same region as your virtual machine (VM) and your TPU node. VMs and TPU nodes are located in specific zones, which are subdivisions within a region.

6. Launch the Compute Engine resources required for this tutorial using the ctpu up command.

   Note: This tutorial uses TensorFlow 1.15.5 which is based on the Debian 10 image and defaults to Python 3.6. Earlier versions of TensorFlow 1.15.x were based on Debian 9 and default to Python 3.5.

   ctpu up --project=${PROJECT_ID} \
    --zone=us-central1-b \
    --vm-only \
    --machine-type=n1-standard-8 \
    --tf-version=1.15.5 \
    --name=inception-tutorial
   

   Command flag descriptions

   project

   Your GCP project ID

   zone

   The zone where you plan to create your Cloud TPU.

   vm-only

   Creates the VM without creating a Cloud TPU. By default the ctpu up command creates a VM and a Cloud TPU.

   machine-type

   The machine type of the Compute Engine VM to create.

   tf-version

   The version of Tensorflow ctpu installs on the VM.

   name

   The name of the Cloud TPU to create.

   For more information on the CTPU utility, see CTPU Reference.

7. When prompted, press y to create your Cloud TPU resources.

   Note: The first time you run ctpu up on a project it takes about 5 minutes to perform startup tasks such as SSH key propagation and API turnup.

   When the ctpu up 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. If you are not connected to the Compute Engine instance, you can do so by running the following command:

   gcloud compute ssh inception-tutorial --zone=us-central1-b
   

   From this point on, a prefix of (vm)$ means you should run the command on the Compute Engine VM instance.

8. Create an environment variable for the storage bucket. Replace bucket-name with the name of your Cloud Storage bucket.

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

9. Create an environment variable for the TPU name.

   (vm)$ export TPU_NAME=inception-tutorial

Training dataset

The training application expects your training data to be accessible in Cloud Storage. The training application also uses your Cloud Storage bucket to store checkpoints during training.

ImageNet is an image database. The images in the database are organized into a hierarchy, with each node of the hierarchy depicted by hundreds and thousands of images.

This tutorial uses a demonstration version of the full ImageNet dataset, referred to as the fake_imagenet dataset. This demonstration version allows you to test out the tutorial, without requiring the storage or time that required to download and run a model against the full ImageNet database. Below are the instructions for using the randomly generated fake_imagenet dataset to test the model. Alternatively, you can use the full ImageNet dataset.

A DATA_DIR environment variable described below is used to specify which dataset to train on.

Note that the fake_imagenet dataset is only useful for understanding how to use a Cloud TPU, and validating end-to-end performance. The accuracy numbers and saved model will not be meaningful.

The fake_imagenet dataset is at this location on Cloud Storage:

gs://cloud-tpu-test-datasets/fake_imagenet

Note: For this tutorial, make sure you don't set the STORAGE_BUCKET environment variable to the path of the fake_imagenet dataset. You can read from gs://cloud-tpu-test-datasets but you can't write to it. As a result, you can't use it to write out training logs. Make sure the STORAGE_BUCKET environment variable is set to your own Cloud Storage bucket, as shown above.

(Optional) Set up TensorBoard

TensorBoard offers a suite of tools designed to present TensorFlow data visually. When used for monitoring, TensorBoard can help identify bottlenecks in processing and suggest ways to improve performance.

If you don't need to monitor the model's output at this time, you can skip the TensorBoard setup steps.

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

Run the model

You are now ready to train and evaluate the Inception v3 model using ImageNet data.

The Inception v3 model is pre-installed on your Compute Engine VM, in the /usr/share/tpu/models/experimental/inception/ directory.

In the following steps, a prefix of (vm)$ means you should run the command on your Compute Engine VM:

1. Set up a DATA_DIR environment variable containing one of the following values:

   • If you are using the fake_imagenet dataset:

     (vm)$ export DATA_DIR=gs://cloud-tpu-test-datasets/fake_imagenet
     

   • If you have uploaded a set of training data to your Cloud Storage bucket:

     (vm)$ export DATA_DIR=${STORAGE_BUCKET}/data
     

2. Run the Inception v3 model:

   (vm)$ python /usr/share/tpu/models/experimental/inception/inception_v3.py \
       --tpu=$TPU_NAME \
       --learning_rate=0.165 \
       --train_steps=250000 \
       --iterations=500 \
       --use_tpu=True \
       --use_data=real \
       --mode=train_and_eval \
       --train_steps_per_eval=2000 \
       --data_dir=${DATA_DIR} \
       --model_dir=${STORAGE_BUCKET}/inception

   • --tpu specifies the name of the Cloud TPU. Note that ctpu passes this name to the Compute Engine VM as an environment variable (TPU_NAME).
   • --use_data specifies which type of data the program must use during training, either fake or real. The default value is fake.
   • --data_dir specifies the Cloud Storage path for training input. The application ignores this parameter when you're using fake_imagenet data.
   • --model_dir specifies the directory where checkpoints and summaries are stored during model training. If the folder is missing, the program creates one. When using a Cloud TPU, the model_dir must be a Cloud Storage path (gs://...). You can reuse an existing folder to load current checkpoint data and to store additional checkpoints as long as the previous checkpoints were created using TPU of the same size and Tensorflow version.

What to expect

Inception v3 operates on 299x299 images. The default training batchsize is 1024, which means that each iteration operates on 1024 of those images.

You can use the --mode flag to select one of three modes of operation: train, eval, and train_and_eval:

• --mode=train or --mode=eval specifies either a training-only or an evaluation-only job.
• --mode=train_and_eval specifies a hybrid job that does both training and evaluation.

Train-only jobs run for the specified number of steps defined in train_steps and can go through the entire training set, if desired.

Train_and_eval jobs cycle though training and evaluation segments. Each training cycle runs for train_steps_per_eval and is followed by an evaluation job (using the weights that have been trained up to that point).

The number of training cycles is defined by the floor function of train_steps divided by train_steps_per_eval.

floor(train_steps / train_steps_per_eval)

Note: The train_steps_per_eval flag must be a multiple of iterations.

By default, Estimator API-based models report loss values every certain number of steps. The reporting format is along the lines of:

step = 15440, loss = 12.6237

Discussion: TPU-specific modifications to the model

The specific modifications required to get Estimator API-based models ready for TPUs are surprisingly minimal. The program imports the following libraries:

from google.third_party.tensorflow.contrib.tpu.python.tpu import tpu_config
from google.third_party.tensorflow.contrib.tpu.python.tpu import tpu_estimator
from google.third_party.tensorflow.contrib.tpu.python.tpu import tpu_optimizer

The CrossShardOptimizer function wraps the optimizer, as in:

if FLAGS.use_tpu:
  optimizer = tpu_optimizer.CrossShardOptimizer(optimizer)

The function that defines the model returns an Estimator specification using:

return tpu_estimator.TPUEstimatorSpec(
    mode=mode, loss=loss, train_op=train_op, eval_metrics=eval_metrics)

The main function defines an Estimator-compatible configuration using:

run_config = tpu_config.RunConfig(
    master=tpu_grpc_url,
    evaluation_master=tpu_grpc_url,
    model_dir=FLAGS.model_dir,
    save_checkpoints_secs=FLAGS.save_checkpoints_secs,
    save_summary_steps=FLAGS.save_summary_steps,
    session_config=tf.ConfigProto(
        allow_soft_placement=True,
        log_device_placement=FLAGS.log_device_placement),
    tpu_config=tpu_config.TPUConfig(
        iterations_per_loop=iterations,
        num_shards=FLAGS.num_shards,
        per_host_input_for_training=per_host_input_for_training))

The program uses this defined configuration and a model definition function to create an Estimator object:

inception_classifier = tpu_estimator.TPUEstimator(
    model_fn=inception_model_fn,
    use_tpu=FLAGS.use_tpu,
    config=run_config,
    params=params,
    train_batch_size=FLAGS.train_batch_size,
    eval_batch_size=eval_batch_size,
    batch_axis=(batch_axis, 0))

Train-only jobs need only to call the train function:

inception_classifier.train(
    input_fn=imagenet_train.input_fn, steps=FLAGS.train_steps)

Evaluation-only jobs get their data from available checkpoints and wait until a new one becomes available:

for checkpoint in get_next_checkpoint():
  eval_results = inception_classifier.evaluate(
      input_fn=imagenet_eval.input_fn,
      steps=eval_steps,
      hooks=eval_hooks,
      checkpoint_path=checkpoint)

When you choose the option train_and_eval, the training and the evaluation jobs run in parallel. During evaluation, trainable variables are loaded from the latest available checkpoint. Training and evaluation cycles repeat as you specify in the flags::

for cycle in range(FLAGS.train_steps // FLAGS.train_steps_per_eval):
  inception_classifier.train(
      input_fn=imagenet_train.input_fn, steps=FLAGS.train_steps_per_eval)

  eval_results = inception_classifier.evaluate(
      input_fn=imagenet_eval.input_fn, steps=eval_steps, hooks=eval_hooks)

If you used the fake_imagenet dataset to train the model, proceed to clean up.

Using the full Imagenet dataset

The ImageNet dataset consists of three parts, training data, validation data, and image labels.

The training data contains 1000 categories and 1.2 million images, packaged for easy downloading. The validation and test data are not contained in the ImageNet training data (duplicates have been removed).

The validation and test data consists of 150,000 photographs, collected from flickr and other search engines, hand labeled with the presence or absence of 1000 object categories. The 1000 object categories contain both internal nodes and leaf nodes of ImageNet, but do not overlap with each other. A random subset of 50,000 of the images with labels has been released as validation data along with a list of the 1000 categories. The remaining images are used for evaluation and have been released without labels.

Steps to pre-processing the full ImageNet dataset

There are five steps to preparing the full ImageNet dataset for use by a Machine Learning model:

1. Verify that you have space on the download target.
2. Set up the target directories.
3. Register on the ImageNet site and request download permission.

4. Download the dataset to local disk or Compute Engine VM.

   Note: Downloading to a Compute Engine VM using Cloud Shell takes considerably longer than downloading to your local machine (approximately 40 hours versus 7 hours). However, if you download to your local machine, you then need to copy the files to a Compute Engine VM to pre-process them and upload them to Cloud Storage before using them to run your model. Copying all of the training and validation files from you local machine to the VM takes about 13 hours. The least complicated, and recommended, approach is to download to the VM.

5. Run the pre-processing and upload script.

Verify space requirements

Whether you download the dataset to your local machine or to a Compute Engine VM, you need about 300GB of space available on the download target. On a VM, you can check your available storage with the df -ha command.

Note: If you use ctpu up to set up your VM, it will allocate 250GB by default.

You can increase the size of the VM disk using one of the following methods:

• Specify the --disk-size-gb flag on the ctpu up command line with the size, in GB, that you want allocated.
• Follow the Compute Engine guide to add a disk to your VM.
  • Set When deleting instance to Delete disk to ensure that the disk is removed when you remove the VM.
  • Make a note of the path to your new disk. For example: /mnt/disks/mnt-dir.

Set up the target directories

On your local machine or Compute Engine VM, set up the directory structure to store the downloaded data.

• Create and export a home directory for the ImageNet dataset.

  Create a directory, for example, imagenet under your home directory on your local machine or VM. Under this directory, create two sub-directories: train and validation. Export the home directory as IMAGENET_HOME:

  export IMAGENET_HOME=~/imagenet
  

Register and request permission to download the dataset

• Register on the Imagenet website. You cannot download the dataset until ImageNet confirms your registration and sends you a confirmation email. If you do not get the confirmation email within a couple of days, contact ImageNet support to see why your registration has not been confirmed. Once your registration is comfirmed, go to the download site.

Download the ImageNet dataset

1. From the download site, go to the Images section on the page and right click on "Training images (Task 1 & 2)". This will give you the URL needed to download the largest part of the training set. Save the URL.

   Right click on "Training images (Task 3)" to get the URL for the second training set. Save the URL.

   Right click on "Validation images (all tasks)" to get the URL for the validation data set. Save the URL.

   If you download the ImageNet files to your local machine, after the download completes, you need to copy the directories on your local machine to the corresponding $IMAGENET_HOME directory on your Compute Engine VM. Copying the ImageNet dataset from local host to your VM takes approximately 13 hours.

   For example the following command copies all of the files under $IMAGENET_HOME on your local machine to your VM that displays the shell prompt username@vm-name:

   gcloud compute scp --recurse $IMAGENET_HOME username@vm-name:~/imagenet
   

2. From $IMAGENET_HOME, use wget to download the training and validation files using the saved URLs.

   The "Training images (Task 1 & 2)" file is the large training set. It is 138GB and if you are downloading to a Compute Engine VM using the Cloud Shell, the estimated time to download is approximately 40 hours. For this particularly large file, you can prepend nohup to the command or use screen to avoid having the download exit if the Cloud Shell disconnects.

   cd $IMAGENET_HOME \
   nohup wget http://image-net.org/challenges/LSVRC/2012/dd31405981ef5f776aa17412e1f0c112/ILSVRC2012_img_train.tar
   

   This downloads a large tar file: ILSVRC2012_img_train.tar.

   From $IMAGENET_HOME on the VM, extract the individual training directories into the $IMAGENET_HOME/train directory using the following command. The extraction takes between 1 - 3 hours.

   tar xf ILSVRC2012_img_train.tar
   

   The "Training images (Task 3)" file is 728 MB and takes just a few minutes to download so you do not need to take precautions against losing the Cloud Shell connection.

   When you download this file, it extracts the individual training directories into the existing $IMAGENET_HOME/train directory.

   wget http://www.image-net.org/challenges/LSVRC/2012/dd31405981ef5f776aa17412e1f0c112/ILSVRC2012_img_train_t3.tar
   

   The "Validation images (all tasks)" file is 6GB, so you might want to use nohup or screen to avoid having the download exit if the Cloud Shell disconnects.

   wget http://www.image-net.org/challenges/LSVRC/2012/dd31405981ef5f776aa17412e1f0c112/ILSVRC2012_img_val.tar
   

   This download takes about 30 minutes. When you download this file, it extracts the individual validation directories into the $IMAGENET_HOME/validation directory.

   If you downloaded the validation files to your local machine, you need to copy the $IMAGENET_HOME/validation directory on your local machine to the $IMAGENET_HOME/validation directory on your Compute Engine VM. This copy operation takes about 30 minutes.

   Download the labels file. This take just seconds.

   wget -O $IMAGENET_HOME/synset_labels.txt \
   https://raw.githubusercontent.com/tensorflow/models/master/research/inception/inception/data/imagenet_2012_validation_synset_labels.txt
   

   If you downloaded the labels file to your local machine, you need to copy it to the $IMAGENET_HOME directory on your local machine to $IMAGENET_HOME on your Compute Engine VM. This copy operation takes a few seconds.

   The training subdirectory names (for example, n03062245) are "WordNet IDs" (wnid). The ImageNet API shows the mapping of WordNet IDs to their associated validation labels in the synset_labels.txt file. A synset in this context is a visually-similar group of images.

Process the Imagenet dataset and, optionally, upload to Cloud Storage

1. Download the imagenet_to_gcs.py script from GitHub:

   wget https://raw.githubusercontent.com/tensorflow/tpu/master/tools/datasets/imagenet_to_gcs.py
   

2. If you are uploading the dataset to Cloud Storage, specify the storage bucket location to upload the ImageNet dataset:

   export STORAGE_BUCKET=gs://bucket-name
   

3. If you are uploading the dataset to Cloud Storage, specify a storage bucket data directory to hold the dataset:

   (vm)$ export DATA_DIR=$STORAGE_BUCKET/dataset-directory
   

4. Run the script to pre-process the raw dataset as TFRecords and upload it to Cloud Storage using the following command:

   Note: If you don't want to upload to Cloud Storage, specify --nogcs_upload as another parameter and leave off the --project and --gcs_outpu_path parameters.

    python3 imagenet_to_gcs.py \
     --project=$PROJECT \
     --gcs_output_path=$DATA_DIR  \
     --raw_data_dir=$IMAGENET_HOME \
     --local_scratch_dir=$IMAGENET_HOME/tf_records
   

Note: Downloading and preprocessing the data can take 10 or more hours, depending on your network and computer speed. Do not interrupt the script.

The script generates a set of directories (for both training and validation) of the form:

${DATA_DIR}/train-00000-of-01024
${DATA_DIR}/train-00001-of-01024
 ...
${DATA_DIR}/train-01023-of-01024

and

${DATA_DIR}/validation-00000-of-00128
S{DATA_DIR}/validation-00001-of-00128
 ...
${DATA_DIR}/validation-00127-of-00128

After the data has been uploaded to your Cloud bucket, run your model and set --data_dir=${DATA_DIR}.

Clean up

To avoid incurring charges to your GCP account for the resources used in this topic:

1. Disconnect from the Compute Engine VM:

   (vm)$ exit
   

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

2. In your Cloud Shell, run ctpu delete with the --zone flag you used when you set up the Cloud TPU to delete your Compute Engine VM and your Cloud TPU:

   $ ctpu delete [optional: --zone]
   

   Important: If you set the TPU resources name when you ran ctpu up, you must specify that name with the --name flag when you run ctpu delete in order to shut down your TPU resources.

3. Run ctpu status to make sure you have no instances allocated to avoid unnecessary charges for TPU usage. The deletion might take several minutes. A response like the one below indicates there are no more allocated instances:

   $ ctpu status --zone=europe-west4-a
   

   2018/04/28 16:16:23 WARNING: Setting zone to "--zone=europe-west4-a"
   No instances currently exist.
       Compute Engine VM:     --
       Cloud TPU:             --
   

4. Run gsutil as shown, replacing bucket-name with the name of the Cloud Storage bucket you created for this tutorial:

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

Note: For free storage limits and other pricing information, see the Cloud Storage pricing guide.

Inception v4

The Inception v4 model is a deep neural network model that uses Inception v3 building blocks to achieve higher accuracy than Inception v3. It is described in the paper "Inception-v4, Inception-ResNet and the Impact of Residual Connections on Learning" by Szegedy et. al.

The Inception v4 model is pre-installed on your Compute Engine VM, in the /usr/share/tpu/models/experimental/inception/ directory.

In the following steps, a prefix of (vm)$ means you should run the command on your Compute Engine VM:

1. If you have TensorBoard running in your Cloud Shell tab, you need another tab to work in. Open another tab in your Cloud Shell, and use ctpu in the new shell to connect to your Compute Engine VM:

   $ ctpu up --project=${PROJECT_ID} 

2. Set up a DATA_DIR environment variable containing one of the following values:

   • If you are using the fake_imagenet dataset:

     (vm)$ export DATA_DIR=gs://cloud-tpu-test-datasets/fake_imagenet
     

   • If you have uploaded a set of training data to your Cloud Storage bucket:

     (vm)$ export DATA_DIR=${STORAGE_BUCKET}/data
     

3. Run the Inception v4 model:

   (vm)$ python /usr/share/tpu/models/experimental/inception/inception_v4.py \
       --tpu=$TPU_NAME \
       --learning_rate=0.36 \
       --train_steps=1000000 \
       --iterations=500 \
       --use_tpu=True \
       --use_data=real \
       --train_batch_size=256 \
       --mode=train_and_eval \
       --train_steps_per_eval=2000 \
       --data_dir=${DATA_DIR} \
       --model_dir=${STORAGE_BUCKET}/inception

   • --tpu specifies the name of the Cloud TPU. Note that ctpu passes this name to the Compute Engine VM as an environment variable (TPU_NAME).
   • --use_data specifies which type of data the program must use during training, either fake or real. The default value is fake.
   • --train_batch_size specifies the train batch size to be 256. As the Inception v4 model is larger than Inception v3, it must be run at a smaller batch size per TPU core.
   • --data_dir specifies the Cloud Storage path for training input. The application ignores this parameter when you're using fake_imagenet data.
   • --model_dir specifies the directory where checkpoints and summaries are stored during model training. If the folder is missing, the program creates one. When using a Cloud TPU, the model_dir must be a Cloud Storage path (gs://...). You can reuse an existing folder to load current checkpoint data and to store additional checkpoints as long as the previous checkpoints were created using TPU of the same size and Tensorflow version.

Cleaning up

What's next

## What's next {: #whats-next } In this tutorial you have trained the Inception 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.

• Go in depth with an advanced view of Inception v3 on Cloud TPU.
• Learn more about ctpu, including how to install it on a local machine.
• Explore the TPU tools in TensorBoard.