Distributed training NasNet with tensorflow_cloud and Google Cloud

View on TensorFlow.org Run in Google Colab View on GitHub Download notebook Kaggle logoRun in Kaggle

This example is based on Image classification via fine-tuning with EfficientNet to demonstrate how to train a NasNetMobile model using tensorflow_cloud and Google Cloud Platform at scale using distributed training.

Import required modules

import tensorflow as tf
tf.version.VERSION
'2.6.0'
! pip install -q tensorflow-cloud

import tensorflow_cloud as tfc
tfc.__version__
import sys

Project Configurations

Set project parameters. For Google Cloud Specific parameters refer to Google Cloud Project Setup Instructions.

# Set Google Cloud Specific parameters

# TODO: Please set GCP_PROJECT_ID to your own Google Cloud project ID.
GCP_PROJECT_ID = 'YOUR_PROJECT_ID'

# TODO: set GCS_BUCKET to your own Google Cloud Storage (GCS) bucket.
GCS_BUCKET = 'YOUR_GCS_BUCKET_NAME'

# DO NOT CHANGE: Currently only the 'us-central1' region is supported.
REGION = 'us-central1'

# OPTIONAL: You can change the job name to any string.
JOB_NAME = 'nasnet'

# Setting location were training logs and checkpoints will be stored
GCS_BASE_PATH = f'gs://{GCS_BUCKET}/{JOB_NAME}'
TENSORBOARD_LOGS_DIR = os.path.join(GCS_BASE_PATH,"logs")
MODEL_CHECKPOINT_DIR = os.path.join(GCS_BASE_PATH,"checkpoints")
SAVED_MODEL_DIR = os.path.join(GCS_BASE_PATH,"saved_model")

Authenticating the notebook to use your Google Cloud Project

For Kaggle Notebooks click on "Add-ons"->"Google Cloud SDK" before running the cell below.

# Using tfc.remote() to ensure this code only runs in notebook
if not tfc.remote():

    # Authentication for Kaggle Notebooks
    if "kaggle_secrets" in sys.modules:
        from kaggle_secrets import UserSecretsClient
        UserSecretsClient().set_gcloud_credentials(project=GCP_PROJECT_ID)

    # Authentication for Colab Notebooks
    if "google.colab" in sys.modules:
        from google.colab import auth
        auth.authenticate_user()
        os.environ["GOOGLE_CLOUD_PROJECT"] = GCP_PROJECT_ID

Load and prepare data

Read raw data and split to train and test data sets.

(x_train, y_train), (x_test, y_test) = tf.keras.datasets.cifar10.load_data()

# Setting input specific parameters
# The model expects input of dimension (INPUT_IMG_SIZE, INPUT_IMG_SIZE, 3)
INPUT_IMG_SIZE = 32
NUM_CLASSES = 10

Add preprocessing layers APIs for image augmentation.

from tensorflow.keras.layers.experimental import preprocessing
from tensorflow.keras.models import Sequential


img_augmentation = Sequential(
    [
        # Resizing input to better match ImageNet size
        preprocessing.Resizing(256, 256),
        preprocessing.RandomRotation(factor=0.15),
        preprocessing.RandomFlip(),
        preprocessing.RandomContrast(factor=0.1),
    ],
    name="img_augmentation",
)

Load the model and prepare for training

We will load a NASNetMobile pretrained model (with weights) and unfreeze a few layers for fine tuning the model to better match the dataset.

from tensorflow.keras import layers

def build_model(num_classes, input_image_size):
    inputs = layers.Input(shape=(input_image_size, input_image_size, 3))
    x = img_augmentation(inputs)

    model = tf.keras.applications.NASNetMobile(
        input_shape=None,
        include_top=False,
        weights="imagenet",
        input_tensor=x,
        pooling=None,
        classes=num_classes,
    )

    # Freeze the pretrained weights
    model.trainable = False

    # We unfreeze the top 20 layers while leaving BatchNorm layers frozen
    for layer in model.layers[-20:]:
        if not isinstance(layer, layers.BatchNormalization):
            layer.trainable = True

    # Rebuild top
    x = layers.GlobalAveragePooling2D(name="avg_pool")(model.output)
    x = layers.BatchNormalization()(x)

    x = layers.Dense(128, activation="relu")(x)
    x = layers.Dense(64, activation="relu")(x)
    outputs = layers.Dense(num_classes, activation="softmax", name="pred")(x)

    # Compile
    model = tf.keras.Model(inputs, outputs, name="NASNetMobile")
    optimizer = tf.keras.optimizers.Adam(learning_rate=3e-4)
    model.compile(
        optimizer=optimizer,
        loss="sparse_categorical_crossentropy",
        metrics=["accuracy"]
    )
    return model
model = build_model(NUM_CLASSES, INPUT_IMG_SIZE)

if tfc.remote():
    # Configure Tensorboard logs
    callbacks=[
        tf.keras.callbacks.TensorBoard(log_dir=TENSORBOARD_LOGS_DIR),
        tf.keras.callbacks.ModelCheckpoint(
            MODEL_CHECKPOINT_DIR,
            save_best_only=True),
        tf.keras.callbacks.EarlyStopping(
            monitor='loss',
            min_delta =0.001,
            patience=3)]

    model.fit(x=x_train, y=y_train, epochs=100,
              validation_split=0.2, callbacks=callbacks)

    model.save(SAVED_MODEL_DIR)

else:
    # Run the training for 1 epoch and a small subset of the data to validate setup
    model.fit(x=x_train[:100], y=y_train[:100], validation_split=0.2, epochs=1)

Start the remote training

This step will prepare your code from this notebook for remote execution and starts a distributed training remotely on Google Cloud Platform to train the model. Once the job is submitted you can go to the next step to monitor the jobs progress via Tensorboard.

# If you are using a custom image you can install modules via requirements
# txt file.
with open('requirements.txt','w') as f:
    f.write('tensorflow-cloud\n')

# Optional: Some recommended base images. If you provide none the system
# will choose one for you.
TF_GPU_IMAGE= "tensorflow/tensorflow:latest-gpu"
TF_CPU_IMAGE= "tensorflow/tensorflow:latest"

# Submit a distributed training job using GPUs.
tfc.run(
    distribution_strategy='auto',
    requirements_txt='requirements.txt',
    docker_config=tfc.DockerConfig(
        parent_image=TF_GPU_IMAGE,
        image_build_bucket=GCS_BUCKET
        ),
    chief_config=tfc.COMMON_MACHINE_CONFIGS['K80_1X'],
      worker_config=tfc.COMMON_MACHINE_CONFIGS['K80_1X'],
      worker_count=3,
    job_labels={'job': JOB_NAME}
)

Training Results

Reconnect your Colab instance

Most remote training jobs are long running. If you are using Colab, it may time out before the training results are available. In that case, rerun the following sections to reconnect and configure your Colab instance to access the training results. Run the following sections in order:

  1. Import required modules
  2. Project Configurations
  3. Authenticating the notebook to use your Google Cloud Project

Load Tensorboard

While the training is in progress you can use Tensorboard to view the results. Note the results will show only after your training has started. This may take a few minutes.

%load_ext tensorboard
%tensorboard --logdir $TENSORBOARD_LOGS_DIR

Load your trained model

trained_model = tf.keras.models.load_model(SAVED_MODEL_DIR)
trained_model.summary()