This describes necessary steps to train models on Google Cloud
Switch branches/tags
Nothing to show
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Failed to load latest commit information.
README.md

README.md

Training Deep Learning Models on Google-Cloud

This describes necessary steps to train models on Google Cloud, assuming you are familiar with Google Cloud console and installed Google Cloud SDK

Enable following APIs in Google Cloud console

  • Compute
  • Billing
  • Storage
  • ML Engine

Upload your datasets faster using parallel transfer

Rename bucket name with your own wherever it says zubair-gc-bucket

gsutil -m cp -r datasets gs://zubair-gc-bucket/datasets

image

Create a **trainer** folder and move all your project files in it
  • Create empty init.py in a trainer folder under your project
  • Add Setup.py outside trainer folder

image

  • Add following in the cloudml-gpu.yaml configuration file
trainingInput:
    scaleTier: BASIC_GPU
    runtimeVersion: "1.4"
    pythonVersion: "3.5"

Open Setup.py and add following

'''Cloud ML Engine package configuration.'''
from setuptools import setup, find_packages

setup(name='shallownet_keras',
      version='1.0',
      packages=find_packages(),
      include_package_data=True,
      description='Model Training using Keras on Google Cloud',
      author='Zubair',
      author_email='your email',
      license='MIT',
      install_requires=[
          'keras',
          'h5py'],
      zip_safe=False)

Add a new argument in your model code before parsing

add.add_argument(
      '--job-dir',
      help='Cloud storage bucket to export the model and store temp files')

args = vars(ap.parse_args())

Include following for saving your model on Google cloud

if args["--job-dir"] != '':
	with file_io.FileIO(args["model"], mode='r') as input_f:
		with file_io.FileIO(args["--job-dir"] + '/' + args["model"], mode='w+') as output_f:
		  output_f.write(input_f.read())

Issue following command

On windows cmd doesn't work with multiline text

gcloud ml-engine jobs submit training job7 --package-path=./trainer --module-name=trainer.shallownet_train --job-dir=gs://zubair-gc-bucket/jobs/job7 --region=us-central1 --config=trainer/cloudml-gpu.yaml --runtime-version="1.4" -- --job_name="zubair-gc-job7"  --dataset=dataset/animals --model=shallownet_weights1.hdf5

Job starts running on Google Cloud

image