Use other deep-learning frameworks

Brief introduction

Xiaomi Cloud-ML supports almost all mainstream deep-learning frameworks. Users only need to write the corresponding application code and to specify the selected framework and version when submitting tasks.

To query the deep-learning frameworks supported by the current service, the following command can be used.

cloudml framework list

Use other frameworks

For example, having compiled the code for the MXNet model, we referred to the method above in packaging it and uploading it to FDS. Then we submitted it using the command below.

cloudml jobs submit -n linear -m trainer.task -u fds://cloud-ml/linear/trainer-1.0.tar.gz -F mxnet -V 0.9.0

Or compile the mxnet_job.json file as defined by the JSON file.

{
  "job_name": "mxnet",
  "module_name": "trainer.task",
  "trainer_uri": "fds://cloud-ml/trainer-1.0.tar.gz",
  "job_args": "",
  "cpu_limit": "0.5",
  "memory_limit": "100M",
  "framework": "mxnet",
  "framework_version": "0.9.0"
}

Then use the command line to submit it. You can also use the GPU, the NodeSelector or other such parameters.

cloudml jobs submit -f ./mxnet_job.json

Use a different version

The TensorFlow deep-learning framework currently supports versions 0.12.0 and 1.0.0, and it may support additional versions of the API in the future. The corresponding frameworks and versions can be specified directly when submitting tasks.

cloudml jobs submit -n linear -m trainer.task -u fds://cloud-ml/linear/trainer-1.0.tar.gz -F tensorflow -V 1.0.0

Parameters introduction

  • -F indicates which deep-learning framework is used.
  • -V indicates which versions use the deep-learning framework.