2

https://cloud.google.com/ml/docs/concepts/training-overview mentions the following:

If your trainer application has any dependencies that are not already on the default virtual machines that Cloud ML uses, you must package them and upload them to a Google Cloud Storage location as well.

What is "already on the default virtual machines that Cloud ML uses"? I couldn't find this info anywhere.

Incidentally, are there any published specs of the machine types here? https://cloud.google.com/ml/reference/rest/v1beta1/projects.jobs#ScaleTier

jwayne
  • 706
  • 2
  • 7
  • 20

1 Answers1

4

What is pre-installed on the CloudML machines is in the process of being documented. In the meantime, this is an informal list of packages with their versions:

numpy==1.10.4
pandas==0.17.1
scipy==0.17.0
scikit-learn==0.17.0
sympy==0.7.6.1
statsmodels==0.6.1
oauth2client==2.2.0
httplib2==0.9.2
python-dateutil==2.5.0
argparse==1.2.1
six==1.10.0
PyYAML==3.11
wrapt==1.10.8
crcmod==1.7
google-api-python-client==1.5.1
python-json-logger==0.1.5
gcloud==0.18.1
subprocess32==3.2.7
wheel==0.30.0a0
WebOb==1.6.2
Paste==2.0.3
tornado==4.3
grpcio==1.0.1
requests==2.9.1
webapp2==3.0.0b1
bs4==0.0.1
Pillow==3.4.1
nltk==3.2.1
python-snappy==0.5
google-cloud-dataflow==0.5.1
google-cloud-logging==0.22.0

In terms of published specs of machine types, those are not available.

Jing Jing Long
  • 191
  • 1
  • 6
  • What about the Tensorflow version? It seems like it's 0.11, but is it possible to ask for Tensorflow 0.12? Or alternately, when will TF 0.12 be available? We've been developing using 0.12 locally, since Tensorboard with TF 0.11 no longer runs well on graph checkpoints hosted on GCS (due to auth upgrades in TF 0.12). But then, when we deploy our local code to Cloud ML, a lot of code no longer works (due to function renames in TF 0.12) – jwayne Feb 08 '17 at 13:59
  • @jwayne based on instruction here https://www.kaggle.com/c/youtube8m#getting-started-with-google-cloud , you have to install it. – Josan Feb 22 '17 at 19:44
  • @AbnerChou, those instructions tell you how to set up your Google Cloud Shell environment so it has Tensorflow, not how to fix a version of Tensorflow on Cloud ML. Even when you deploy a job on Cloud ML from a Cloud Shell environment with a fixed TF version installed, that installed version doesn't control the version of TF being run on the Cloud ML machines. I actually ended up finding an answer to that here: http://stackoverflow.com/questions/42140031/google-cloud-ml-tensorflow-version – jwayne Feb 23 '17 at 22:31