Training RetinaNet on Cloud TPU (TF 2.x)

This document describes an implementation of the RetinaNet object detection model. The code is available on GitHub.

The instructions below assume you are already familiar with running a model on Cloud TPU. If you are new to Cloud TPU, you can refer to the Quickstart for a basic introduction.

If you plan to train on a TPU Pod slice, review Training on TPU Pods to understand parameter changes required for Pod slices.

Objectives

Prepare the COCO dataset
Create a Cloud Storage bucket to hold your dataset and model output
Set up TPU resources for training and evaluation
Run training and evaluation on a single Cloud TPU or a Cloud TPU Pod

Costs

In this document, you use the following billable components of Google Cloud:

Compute Engine
Cloud TPU
Cloud Storage

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 Google Cloud 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 Google Cloud project.

This walkthrough uses billable components of Google Cloud. Check the Cloud TPU pricing page to estimate your costs.

Prepare the COCO dataset

This tutorial uses the COCO dataset. The dataset needs to be in TFRecord format on a Cloud Storage bucket to be used for the training.

If you already have the COCO dataset prepared on a Cloud Storage bucket that is located in the zone you will be using to train the model, you can go directly to single device training. Otherwise, use the following steps to prepare the dataset.

Open a Cloud Shell window.

Open Cloud Shell

In your Cloud Shell, configure gcloud with your project ID.

export PROJECT_ID=project-id
gcloud config set project ${PROJECT_ID}

In your Cloud Shell, 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.
```
gcloud storage buckets create gs://bucket-name --project=${PROJECT_ID} --location=us-central2
```

Create a Compute Engine VM to download and preprocess the dataset. For more information, see Create and start a Compute Engine instance.

$ gcloud compute instances create vm-name \
    --zone=us-central2-b \
    --image-family=ubuntu-2204-lts \
    --image-project=ubuntu-os-cloud \
    --machine-type=n1-standard-16 \
    --boot-disk-size=300GB

Connect to the Compute Engine VM using SSH:
```
$ gcloud compute ssh vm-name --zone=us-central2-b
```
When you connect to the VM, your shell prompt changes from username@projectname to username@vm-name.
Set up two variables, one for the storage bucket you created earlier and one for the directory that holds the training data (DATA_DIR) on the storage bucket.
```
(vm)$ export STORAGE_BUCKET=gs://bucket-name
```
```
(vm)$ export DATA_DIR=${STORAGE_BUCKET}/coco
```

Install the packages needed to pre-process the data.

(vm)$ sudo apt-get update && \
  sudo apt-get install python3-pip && \
  sudo apt-get install -y python3-tk && \
  pip3 install --user Cython matplotlib opencv-python-headless pyyaml Pillow numpy absl-py tensorflow && \
  pip3 install --user "git+https://github.com/cocodataset/cocoapi#egg=pycocotools&subdirectory=PythonAPI"

Run the download_and_preprocess_coco.sh script to convert the COCO dataset into a set of TFRecord files (*.tfrecord) that the training application expects.
```
(vm)$ git clone https://github.com/tensorflow/tpu.git
(vm)$ sudo bash tpu/tools/datasets/download_and_preprocess_coco.sh ./data/dir/coco
```
This installs the required libraries and then runs the preprocessing script. It outputs *.tfrecord files in your local data directory. The COCO download and conversion script takes approximately one hour to complete.
Copy the data to your Cloud Storage bucket.

After you convert the data into the TFRecord format, copy the data from local storage to your Cloud Storage bucket using the gcloud CLI. You must also copy the annotation files. These files help validate the model's performance.
```
(vm)$ gcloud storage cp ./data/dir/coco/*.tfrecord ${DATA_DIR}
(vm)$ gcloud storage cp ./data/dir/coco/raw-data/annotations/*.json ${DATA_DIR}
```
Disconnect from the Compute Engine VM:
```
(vm)$ exit
```
Your prompt should now be username@projectname, showing you are in the Cloud Shell.

Delete your Compute Engine VM:

$ gcloud compute instances delete vm-name \
--zone=us-central2-b

Cloud TPU single device training

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 Google Cloud API calls with your credentials.

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:

service-PROJECT_NUMBER@cloud-tpu.iam.gserviceaccount.com

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.
```
gcloud storage buckets create gs://bucket-name --project=${PROJECT_ID} --location=europe-west4
```
This Cloud Storage bucket stores the data you use to train your model and the training results. The gcloud command used in this tutorial to set up the TPU also 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.

Set up and start the Cloud TPU

Launch a Compute Engine VM and Cloud TPU using the gcloud command.
```
$ gcloud compute tpus tpu-vm create retinanet-tutorial \
--zone=europe-west4-a \
--accelerator-type=v3-8 \
--version=tpu-vm-tf-2.17.0-pjrt
```
Command flag descriptions

zone

The zone where you plan to create your Cloud TPU.

accelerator-type

The accelerator type specifies the version and size of the Cloud TPU you want to create. For more information about supported accelerator types for each TPU version, see TPU versions.

version

The Cloud TPU software version.

Note: If you have more than one project, you must specify the project ID with the --project flag.

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

Note: The first time you run gcloud compute tpus on a project it takes approximately 5 minutes to perform startup tasks such as SSH key propagation and API turnup.
Connect to the Compute Engine instance using SSH. When you are connected to the VM, your shell prompt changes from username@projectname to username@vm-name:
```
gcloud compute tpus tpu-vm ssh retinanet-tutorial --zone=europe-west4-a
```
Key Point: From this point on, a prefix of (vm) $ means you should run the command on the Compute Engine VM instance.

Install extra packages

The RetinaNet training application requires several extra packages. Install them now:

(vm)$ sudo apt-get install -y python3-tk
(vm)$ pip3 install --user Cython matplotlib opencv-python-headless pyyaml Pillow
(vm)$ pip3 install --user 'git+https://github.com/cocodataset/cocoapi#egg=pycocotools&subdirectory=PythonAPI'

Install TensorFlow requirements.

(vm)$ pip3 install -r /usr/share/tpu/models/official/requirements.txt

Set the Cloud TPU name variable.
```
(vm)$ export TPU_NAME=local
```

Add environment variables for the data and model directories.

(vm)$ export STORAGE_BUCKET=gs://bucket-name
(vm)$ export DATA_DIR=${STORAGE_BUCKET}/coco
(vm)$ export MODEL_DIR=${STORAGE_BUCKET}/retinanet-train
(vm)$ export PYTHONPATH="${PWD}/models:${PYTHONPATH}"

When creating your TPU, if you set the --version parameter to a version ending with -pjrt, set the following environment variables to enable the PJRT runtime:
```
  (vm)$ export NEXT_PLUGGABLE_DEVICE_USE_C_API=true
  (vm)$ export TF_PLUGGABLE_DEVICE_LIBRARY_PATH=/lib/libtpu.so
```

Change to directory that stores the model:

(vm)$ cd /usr/share/tpu/models/official/legacy/detection

Single Cloud TPU device training

The following training scripts were run on a Cloud TPU v3-8. It will take more time, but you can also run them on a Cloud TPU v2-8.

The following sample script trains for only 10 steps and takes less than 5 minutes to run on a v3-8 TPU. To train to convergence takes about 22,500 steps and approximately 1 1/2 hours on a Cloud TPU v3-8 TPU.

Set up the following environment variables:

(vm)$ export RESNET_CHECKPOINT=gs://cloud-tpu-checkpoints/retinanet/resnet50-checkpoint-2018-02-07
(vm)$ export TRAIN_FILE_PATTERN=${DATA_DIR}/train-*
(vm)$ export EVAL_FILE_PATTERN=${DATA_DIR}/val-*
(vm)$ export VAL_JSON_FILE=${DATA_DIR}/instances_val2017.json

Run the training script:
```
(vm)$ python3 main.py \
     --strategy_type=tpu \
     --tpu=${TPU_NAME} \
     --model_dir=${MODEL_DIR} \
     --mode="train" \
     --params_override="{ type: retinanet, train: { total_steps: 10, checkpoint: { path: ${RESNET_CHECKPOINT}, prefix: resnet50/ }, train_file_pattern: ${TRAIN_FILE_PATTERN} }, eval: { val_json_file: ${VAL_JSON_FILE}, eval_file_pattern: ${EVAL_FILE_PATTERN}, eval_samples: 5000 } }"
```
Command flag descriptions

strategy_type

To train the RetinaNet model on a TPU, you must set the distribution_strategy to tpu.

tpu

The name of the Cloud TPU. This is set using the TPU_NAME environment variable.

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.

mode

Set this to train to train the model or eval to evaluate the model.

params_override

A JSON string that overrides default script parameters. For more information on script parameters, see /usr/share/models/official/legacy/detection/main.py.

The model will train for 10 steps in about 5 minutes on a v3-8 TPU. When the training completes, you will see output similar to the following:

Train Step: 10/10  / loss = {
  'total_loss': 2.4581615924835205,
  'cls_loss': 1.4098565578460693,
  'box_loss': 0.012001709081232548,
  'model_loss': 2.0099422931671143,
  'l2_regularization_loss': 0.44821977615356445,
  'learning_rate': 0.008165999
}
/ training metric = {
  'total_loss': 2.4581615924835205,
  'cls_loss': 1.4098565578460693,
  'box_loss': 0.012001709081232548,
  'model_loss': 2.0099422931671143,
  'l2_regularization_loss': 0.44821977615356445,
 'learning_rate': 0.008165999
}

Single Cloud TPU device evaluation

The following procedure uses the COCO evaluation data. It takes about 10 minutes to run through the evaluation steps on a v3-8 TPU.

Set up the following environment variables:
```
(vm)$ export EVAL_SAMPLES=5000
```

Run the evaluation script:

(vm)$ python3 main.py \
      --strategy_type=tpu \
      --tpu=${TPU_NAME} \
      --model_dir=${MODEL_DIR} \
      --checkpoint_path=${MODEL_DIR} \
      --mode=eval_once \
      --params_override="{ type: retinanet, eval: { val_json_file: ${VAL_JSON_FILE}, eval_file_pattern: ${EVAL_FILE_PATTERN}, eval_samples: ${EVAL_SAMPLES} } }"

Command flag descriptions

strategy_type: The distribution strategy to use. Either tpu or multi_worker_gpu.
tpu: The name of the Cloud TPU. This is set using the TPU_NAME environment variable.
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.
mode: One of train, eval, or train_and_eval.
params_override: A JSON string that overrides default script parameters. For more information on script parameters, see /usr/share/models/official/legacy/detection/main.py.

At the end of the evaluation, you will see messages similar to the following on the console:

Accumulating evaluation results...
DONE (t=7.66s).
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.000
 Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=100 ] = 0.000
 Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets=100 ] = 0.000
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.000
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.000
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.000
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=  1 ] = 0.000
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets= 10 ] = 0.000
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.000
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.000
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.000
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.000

You have now completed single-device training and evaluation. Use the following steps to delete the current single-device TPU resources.

Disconnect from the Compute Engine instance:
```
(vm)$ exit
```
Your prompt should now be username@projectname, showing you are in the Cloud Shell.
Delete the TPU resource.
```
$ gcloud compute tpus tpu-vm delete retinanet-tutorial \
  --zone=europe-west4-a
```
Command flag descriptions

zone

The zone where your Cloud TPU resides.

Scale your model with Cloud TPU Pods

Training your model on Cloud TPU Pods may require some changes to your training script. For information, see Training on TPU Pods.

Training Retinanet on a TPU Pod

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 Google Cloud API calls with your Google Cloud credentials.
Create a Service Account for the Cloud TPU project.

Service accounts allow the Cloud TPU service to access other Google Cloud services.
```
gcloud beta services identity create --service tpu.googleapis.com --project $PROJECT_ID
```
The command returns a Cloud TPU Service Account with following format:
```
service-PROJECT_NUMBER@cloud-tpu.iam.gserviceaccount.com
```
Create a Cloud Storage bucket using the following command or use a bucket you created earlier for your project.

Important: Set up your Cloud TPU resources and your Cloud Storage bucket in the same region and zone to reduce network latency and network costs. Cloud TPUs are located in specific zones, which are subdivisions within a region.

In the following command, replace europe-west4 with the name of the region you will use to run the training. Replace bucket-name with the name you want to assign to your bucket.
```
gcloud storage buckets create gs://bucket-name \
  --project=${PROJECT_ID} \
  --location=europe-west4
```
This Cloud Storage bucket stores the data you use to train your model and the training results. The gcloud command 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 access level permissions.

The bucket location must be in the same region as your TPU resources.
If you previously prepared the COCO dataset and moved it to your storage bucket, you can use it again for Pod training. If you have not yet prepared the COCO dataset, prepare it now and return here to set up the training.
Set up and launch a Cloud TPU Pod

This tutorial specifies a v3-32 Pod. For other Pod options, see TPU versions.

Note: If there is not enough capacity currently available to create the TPU Pod, you can queue your request using queued resources. Queued resources allow you to receive capacity once it becomes available. To request your Cloud TPU resources as queued resources, use the gcloud compute tpus queued-resources create command instead. For more information, see Manage Queued Resources.

Launch a TPU VM Pod using the gcloud compute tpus tpu-vm command. This tutorial specifies a v3-32 Pod. For other Pod options, see the available TPU types page.
```
$ gcloud compute tpus tpu-vm create retinanet-tutorial \
  --zone=europe-west4-a \
  --accelerator-type=v3-32 \
  --version=tpu-vm-tf-2.17.0-pod-pjrt
 
```
Command flag descriptions

zone

The zone where you plan to create your Cloud TPU.

accelerator-type

The accelerator type specifies the version and size of the Cloud TPU you want to create. For more information about supported accelerator types for each TPU version, see TPU versions.

version

The Cloud TPU software version.

Note: The first time you run gcloud on a project it takes about 5 minutes to perform startup tasks such as SSH key propagation and API turnup.
Connect to the TPU VM instance using SSH. When you are connected to the VM, your shell prompt changes from username@projectname to username@vm-name:
```
gcloud compute tpus tpu-vm ssh retinanet-tutorial --zone=europe-west4-a
```
Key Point: From this point on, a prefix of (vm) $ means you should run the command on the Compute Engine VM instance. begins with (vm)$ in your VM session window.

Set the Cloud TPU name variable.

(vm)$ export TPU_NAME=retinanet-tutorial

Set Cloud Storage bucket variables

Set up the following environment variables, replacing bucket-name with the name of your Cloud Storage bucket:
```
(vm)$ export STORAGE_BUCKET=gs://bucket-name
```
```
(vm)$ export MODEL_DIR=${STORAGE_BUCKET}/retinanet-train
(vm)$ export DATA_DIR=${STORAGE_BUCKET}/coco
```
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.

Install extra packages

The RetinaNet training application requires several extra packages. Install them now:

(vm)$ sudo apt-get install -y python3-tk
(vm)$ pip3 install --user Cython matplotlib opencv-python-headless pyyaml Pillow
(vm)$ pip3 install --user 'git+https://github.com/cocodataset/cocoapi#egg=pycocotools&subdirectory=PythonAPI'

Install TensorFlow requirements.

(vm)$ pip3 install -r /usr/share/tpu/models/official/requirements.txt

Set some required environment variables:

(vm)$ export RESNET_PRETRAIN_DIR=gs://cloud-tpu-checkpoints/retinanet/resnet50-checkpoint-2018-02-07
(vm)$ export TRAIN_FILE_PATTERN=${DATA_DIR}/train-*
(vm)$ export EVAL_FILE_PATTERN=${DATA_DIR}/val-*
(vm)$ export VAL_JSON_FILE=${DATA_DIR}/instances_val2017.json
(vm)$ export PYTHONPATH="${PWD}/models:${PYTHONPATH}"
(vm)$ export TPU_LOAD_LIBRARY=0

Change to directory that stores the model:

(vm)$ cd /usr/share/tpu/models/official/legacy/detection

Train the model
```
(vm)$ python3 main.py \
  --strategy_type=tpu \
  --tpu=${TPU_NAME} \
  --model_dir=${MODEL_DIR} \
  --mode=train \
  --model=retinanet \
  --params_override="{architecture: {use_bfloat16: true}, eval: {batch_size: 40, eval_file_pattern: ${EVAL_FILE_PATTERN}, val_json_file: ${VAL_JSON_FILE}}, postprocess: {pre_nms_num_boxes: 1000}, predict: {batch_size: 40}, train: {batch_size: 256, checkpoint: {path: ${RESNET_PRETRAIN_DIR}, prefix: resnet50/}, iterations_per_loop: 5000, total_steps: 5625, train_file_pattern: ${TRAIN_FILE_PATTERN}, } }" 
```
Command flag descriptions

tpu

The name of your TPU.

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 Cloud TPU of the same size and TensorFlow version.

params_override

A JSON string that overrides default script parameters. For more information on script parameters, see /usr/share/tpu/models/official/legacy/detection/main.py.

This procedure trains the model on the COCO dataset for 5625 training steps. This training takes approximately 20 minutes on a v3-32 TPU. When the training completes, a message similar to the following appears:

When the training completes, a message similar to the following appears:

   Train Step: 5625/5625  / loss = {'total_loss': 0.730501651763916,
   'cls_loss': 0.3229793608188629, 'box_loss': 0.003082591574639082,
   'model_loss': 0.4771089553833008, 'l2_regularization_loss': 0.2533927261829376,
   'learning_rate': 0.08} / training metric = {'total_loss': 0.730501651763916,
   'cls_loss': 0.3229793608188629, 'box_loss': 0.003082591574639082,
   'model_loss': 0.4771089553833008, 'l2_regularization_loss': 0.2533927261829376,
   'learning_rate': 0.08}

Clean 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.

Disconnect from the Compute Engine VM:
```
(vm)$ exit
```
Your prompt should now be username@projectname, showing you are in the Cloud Shell.

Delete your Cloud TPU and Compute Engine resources.

$ gcloud compute tpus tpu-vm delete retinanet-tutorial \
  --zone=europe-west4-a

Verify the resources have been deleted by running gcloud compute tpus tpu-vm list. The deletion might take several minutes. A response like the following indicates your instances have been successfully deleted.
```
$ gcloud compute tpus tpu-vm list --zone=europe-west4-a
```
```
Listed 0 items.
```
Delete your Cloud Storage bucket. Replace bucket-name with the name of your Cloud Storage bucket.
Delete your Cloud Storage bucket using the gcloud CLI as shown in the following example. Replace bucket-name with the name of your Cloud Storage bucket.
```
$ gcloud storage rm gs://bucket-name --recursive
```

What's next

The TensorFlow Cloud TPU tutorials generally train the model using a sample dataset. The results of this training are not usable for inference. To use a model for inference, you can train the data on a publicly available dataset or your own dataset. TensorFlow models trained on Cloud TPUs generally 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 and Tune hyperparameters.

Inference

Once you have trained your model, you can use it for inference (also called prediction). You can use the Cloud TPU inference converter tool to prepare and optimize a TensorFlow model for inference on Cloud TPU v5e. For more information about inference on Cloud TPU v5e, see Cloud TPU v5e inference introduction.

Train with different image sizes

You can explore using a larger backbone network (for example, ResNet-101 instead of ResNet-50). A larger input image and a more powerful backbone will yield a slower but more precise model.

Use a different basis

Alternatively, you can explore pre-training a ResNet model on your own dataset and using it as a basis for your RetinaNet model. With some more work, you can also swap in an alternative backbone network in place of ResNet. Finally, if you are interested in implementing your own object detection models, this network may be a good basis for further experimentation.