Task-based modeling provides model building capabilities in a wizard-like training task submission method and supports task submission based on multiple algorithm sources. It allows you to directly bind mainstream high-performance, distributed training frameworks through code package to quickly submit and start training tasks. The following uses a simple PyTorch MPI job as an example to demonstrate how to quickly create a task in task-based modeling.
This document uses the MNIST database available here.
The training script in this document is written by using the PyTorch framework. Its code package can be downloaded here.
Enter the following information on the basic information page:
After entering the required information, click Next to go to the task configuration page.
Enter the following information on the task configuration page:
mnist.pytorchfolder generated after decompressing the code package to the bucket, and select the path of the code package.
Data Source: select COS.
ti-imagesfolder generated after decompressing the dataset, and select the folder path as shown below:
Tuning Parameter: none.
Training Output: click Select. In the COS pop-up window, select the target bucket path for saving the training output data as shown below:
After completing the configuration, view the hourly rate of the training task at the bottom of the page and click OK.