Atlas800 Web Access

To access Atlas800, visit: https://atlas.cmc.ca/

Atlas 800 Training Cluster

The Atlas 800 cluster features 4 Atlas 800 training servers (model: 9000). Each server is powered by the Kunpeng 920 and Ascend 910 processors. It features the industry's highest computing density, ultra-high energy efficiency, and high network bandwidth. The server is widely used in deep learning model development and training scenarios and is an ideal option for computing-intensive industries, such as smart city, intelligent healthcare, astronomical exploration, and oil exploration.

The specification of each server node is summarized in the following table:

Form Factor	4U AI server
Processor	4 Kunpeng 920 processors
Processor Memory	Up to 32 DDR4 DIMM slots, supporting RDIMMs Up to 2933 MT/s 32 GB or 64 GB per DIMM
AI Processor	8 Ascend 910 processors
HBM	8 * 32 GB
AI Computing Power	2.56 / 2.24 / 2 PFLOPS FP16
**

Atlas 800 Software stack

The Atlas 800 cluster is managed by Apulis AI Studio, which is an end-to-end AI development platform that enables developers and data scientists of any skill level to rapidly build, train, and deploy models anywhere, from the cloud to the edge. As shown in the following figure, the Apulis AI studio accelerates life cycle of AI development and fosters AI innovation with key capabilities, including data preprocessing, semi-automated data labeling, distributed training, and automated model building.

Objectives

This tutorial walks you through all the necessary steps required to build, train and evaluate et deep neural network on the Atlas 800 cluster

Develop and Train Models in 'Code development' Project

• From the home page of the Atlas AI Platform, Click on AIArts

• Click on Projects under Project Management`` to display an overview of all current user group tasks or jobs as shown in the following Figure:

• Click on New Project, select the Code Development on the pop-window, setup a new code-development job. Give a name and a description for your new project as described in the following Figure:

• Click Submit to create your new project. In our case, the project name is cmctest

• From the list of the projects, click on the project you have just created (cmctest in our case) as shown in the following Figure:

• Start the code development environment in the project, and select the image, dataset, and NPU quota when starting the task

• When code development job is Running, You can directly startup jupyter-lab or setup the Remote-SSH connection in your local IDE (Pycharm, VS Code, etc.).

• In order to get the instructions for using the code development environment, including preset folder instructions, development instructions and model release instructions, click on the icon as shown in the following Figure:

• Click on Jupyter to start a Jupyter notebook session:

• Click the icon and upload the code package (1495.tar.gz) to the code development environment in the pop-up box.

• Drag and drop the file 1495.tar.gz to the submission window then click submit as shown in the following Figure:

• Use the following commands in the code development environment to unzip the code package, move the files to the code directory:

tar -xf 1495.tar.gz

mv /home/yhariri/code/1495/infer.yaml /home/yhariri/code/

mv /home/yhariri/code/1495/manifest.yaml /home/yhariri/code/

mv /home/yhariri/code/1495/code/* /home/yhariri/code/

• Execute the training script train.sh to start the training:

bash train.sh --batch_size=2 --epoch_size=5 --lr_init=0.01 --run_distribute=True --dataset=cifar10 --data_path=/home/yhariri/adhub/mindspore/0.1/ --output_path=/home/yhariri/outputs/

The execution status is shown in the following Figure

• You can also startup a Mindinsight interactive port to execute a lineage analysis of summary logs. In order to do that, use the following commands to start Mindinsight service in Jupyter-Terminal:

#!/bin/bash

# Dependence：Mindpsore <= v1.6, MindInsight <= v1.6.1

# Listening port

INTERACTIVE_PORT=8080

# Local summary log path

SUMMARY_PATH=$HOME/outputs

# Reload interval time

INTERVAL=5

# Update the Mindinsight

pip install -U mindinsight

# Setup the banding host and port

sed -i "s/HOST = '127.0.0.1'/HOST = '0.0.0.0'/g"  $HOME/.local/lib/python3.7/site-packages/mindinsight/conf/constants.py

# Startup service with specified local summary path

mindinsight start --port $INTERACTIVE_PORT --summary-base-dir $SUMMARY_PATH --reload-interval $INTERVAL

# Or if you want to stop

mindinsight stop --port $INTERACTIVE_PORT

• Create the interactive port on the development task board:

  • Click on New Port, then specify mindsight as a port name and 8080 as Port number as shown in the following Figure:

<!-- -->

• Launch the model evaluation process:

  • Withing the pre-set demo, execute the evaluation commands as following steps:

python3 eval.py --dataset=cifar10 --data_path=/home/yhariri/adhub/mindspore/0.1/ --output_path=/home/yhariri/outputs/ --checkpoint_path=/home/yhariri/outputs/

• Check the files in the outputs folder and copy the infer.yaml file from the code directory to the outputs folder.

cp /home/yhariri/code/infer.yaml /home/yhariri/outputs/

• Open Mindsight as shown in the following Figure:

• Click on the Training Dashboard as shown in the following Figure:

• A new window appears providing access to different visual information of the training:

• Click on Lineage analysis to access more details about the training process:

• The result should look like the following Figure:

• Submit the code under the code directory.

# The command to submit code

git add . && git commit -m "add tarin and eval" && git push

• Publish the model to Model Factory:

/start/aplab -trainOutPutPath /home/yhariri/outputs -modelName MSU-DEMO-V1

• Inspect the model trained in the code development environment in Model Factory

Resources

• Datasheet
• Ascend Community; 
• Video > Introduction;
• Ascend Open-Source (ModelZoo, > Samples); 
• Video > Introduction;
• Ascend Documentation (Model > Development)