eileen-code4fun

{
  "displayName": "e2e-tutorial-viz",
  "jobSpec": {
    "workerPoolSpecs": [
      {
        "replicaCount": 1,
        "machineSpec": {
          "machineType": "n1-standard-4",
          "acceleratorType": "NVIDIA_TESLA_V100",
          "acceleratorCount": 2

eileen-code4fun

import os

tensorboard_callback = tf.keras.callbacks.TensorBoard(
  log_dir=os.environ['AIP_TENSORBOARD_LOG_DIR'],
  histogram_freq=1
)

model.fit(train_dataset, epochs=args.epochs, validation_data=val_dataset, callbacks=[tensorboard_callback])

eileen-code4fun

studySpec:
  metrics:
  # Correspond to the metrics we use the hypertune library to report.
  - metricId: val_accuracy
    goal: MAXIMIZE
  parameters:
  # Correspond to the command line argument our Python code expects.
  - parameterId: dropout_rate
    doubleValueSpec:
      minValue: 0.01

eileen-code4fun

import tensorflow as tf
import hypertune

hpt = hypertune.HyperTune()

class CustomCallback(tf.keras.callbacks.Callback):
  def on_epoch_end(self, epoch, logs=None):
    hpt.report_hyperparameter_tuning_metric(
      hyperparameter_metric_tag='val_accuracy',
      metric_value=logs['val_accuracy'],

eileen-code4fun

from tensorflow.keras import layers, models, losses
import argparse

parser = argparse.ArgumentParser()
parser.add_argument('--epochs', dest='epochs', type=int, default=5)
parser.add_argument('--dropout_rate', dest='dropout_rate', type=float, default=0.1)

args = parser.parse_args()

def create_model():

eileen-code4fun

workerPoolSpecs:
  machineSpec:
    # Machines and GPUs: https://cloud.google.com/vertex-ai/docs/training/configure-compute#specifying_gpus
    machineType: n1-standard-4
    acceleratorType: NVIDIA_TESLA_V100
    acceleratorCount: 2
  replicaCount: 1
  pythonPackageSpec:
    # Executors: https://cloud.google.com/vertex-ai/docs/training/pre-built-containers
    executorImageUri: us-docker.pkg.dev/vertex-ai/training/tf-gpu.2-3:latest

eileen-code4fun

cloud_model = tf.keras.models.load_model(GCS_PATH_FOR_SAVED_MODEL)
# Compile the model to make sure it uses the correct accuracy metric.
cloud_model.compile(optimizer='adam', loss=tf.keras.losses.SparseCategoricalCrossentropy(), metrics=['accuracy'])
cloud_model.evaluate(test_dataset, verbose=2)
# Output loss: 0.917 - accuracy: 0.6620

eileen-code4fun

from setuptools import find_packages
from setuptools import setup

setup(
  name='trainer',
  version='0.1',
  packages=find_packages(),
  include_package_data=True,
  description='An E2E Tutorial of Vertex AI'
)

eileen-code4fun

import argparse

parser = argparse.ArgumentParser()
parser.add_argument('--epochs', dest='epochs', type=int, default=5)

args = parser.parse_args()

# Refer to the argument as args.epochs.

eileen-code4fun

# A distributed strategy to take advantage of available hardward.
# No-op otherwise.
mirrored_strategy = tf.distribute.MirroredStrategy()
with mirrored_strategy.scope():
  model = create_model()
  # Restore from the latest checkpoint if available.
  latest_ckpt = tf.train.latest_checkpoint(GCS_PATH_FOR_CHECKPOINTS)
  if latest_ckpt:
    model.load_weights(latest_ckpt)