Get Started with Intel MLPerf Inference Submission with Intel Optimized Docker Images

Please also check your CPU tempartures. The CPU temparture should not be higher than 50 degrees C.
Overheating will drop the CPU frequency and degrade the MLPerf performance.

From the results of svr-info, an example of unbalanced DIMM population is shown as follows,


An exmaple of Balanced DIMM population is shown as follows,


You should also see good numbers for memory NUMA bandwidth if you also benchmark memory via svr-info.
Here are some reference numbers from a 2S SPR system.

Get Started with Intel MLPerf v4.0 Submission with Intel Optimized Docker Images

MLPerf is a benchmark for measuring the performance of machine learning systems. It provides a set of performance metrics for a variety of machine learning tasks, including image classification, object detection, machine translation, and others. The benchmark is representative of real-world workloads and as a fair and useful way to compare the performance of different machine learning systems.

In this document, we'll show how to run Intel MLPerf v4.0 submission with Intel optimized Docker images and the automation scripts.

HW configuration:

BIOS settings:

Check System Health Using Intel® System Health Inspector:

Intel® System Health Inspector (aka svr-info) is a Linux OS utility for assessing the state and health of Intel Xeon computers. It is suggested to use svr-info first to check any system configuration issue before running any benchmark. Follow the Quick Start Guide for downloading and installation. The following are several key factors effecting the model performance.

CPU

Couple CPU features impact MLPerf performance via related BIOS knobs, so please double check the CPU features with your BIOS knobs. Some important CPU features are Hyperthreading, number of NUMA nodes, Prefetchers and Intel Turbo Boost.

Please also check your CPU tempartures. The CPU temparture should not be higher than 50 degrees C.
Overheating will drop the CPU frequency and degrade the MLPerf performance.

Power

We recommend the intel_pstate Frequency Driver.
For best performance, set the Frequency Governor and Power and Perf Policy to performance.
Here are related recommended power settings from svr-info.

Published Performance Numbers

All performance numbers are published in https://mlcommons.org/benchmarks/inference-datacenter/ Here are submitted numbers on Intel 5th gen Xeon scalable server processor (EMR) from different organizations.

Prerequisite

Get Intel MLPerf 4.0 Inferece package from MLCOMMONS

Users could get the Intel MLPerf 4.0 Inferece package from https://github.com/mlcommons/inferenceresultsv4.0 hosted by MLCOMMONS.
Please follow below command to get the whole MLPerf 4.0 Inference package

git clone https://github.com/mlcommons/inference_results_v4.0.git

The Intel MLPerf Inference package is under inferenceresultsv4.0/closed/Intel folder.
The model codes are under code folder, and we recommended users to use automation script to run all models.

Environment Setup for the automation scripts from Intel Inference package

We provides a kit to automate data ingestion, preprocessing, testing, log collection and submission checker. It requires the following software:

• python 3.9/3.10, pip, docker, wget, unzip, rclone, miniconda3/anaconda3

Note that the following approaches are supported for installing dependencies:

1. Install directly
2. Install in a conda environment
3. Install in a python virtual environment

For the option #2 and #3, you need to create the environment and activate it first before installation.

Install the dependencies by:

cd inference_results_v4.0/closed/Intel/code/automation
pip3 install -r requirements.txt

If your servers are working behind network proxy, please ensure docker proxy and the following environment parameters are well configured:

• httpproxy, HTTPPROXY, httpsproxy, HTTPSPROXY, noproxy, NOPROXY, ftpproxy, FTPPROXY, socksproxy, SOCKSPROXY

Downloading data and models by using automation scripts

First, go to automation folder to automate the data and model downloading.

cd inference_results_v4.0/closed/Intel/code/automation

Conda is required for downloading datasets and models. Please install Conda before proceeding. Once Conda is installed, you can download the datasets and models using the following commands:

model={resnet50,retinanet,rnnt,3d-unet,bert,gpt-j,dlrm_2,stable_diffusion,all} output_dir=<DATA_PATH> conda_path=<CONDA_ROOT_PATH> bash download_data.sh 

Parameters:

• model: specify a model name, all means all models.
• output_dir: the directory to persist data and model.
• conda_path: optional, ${HOME}/miniconda3 by default, root path of your conda, used by 3d-unet, retinanet and gpt-j only.
• dtype: optional, int8 by default, data preicision, currently unused

Download data for all models by using the following command:

model=all output_dir=<DATA_PATH> conda_path=<CONDA_ROOT_PATH> bash download_data.sh

For specific models, e.g. gpt-j, DLRMv2 and Stable Diffusion, Rclone is needed. If you don't have Rclone already installed, you can install it on Linux/MacOS with one simple command:

sudo -v ; curl https://rclone.org/install.sh | sudo bash  

For specific worload, e.g. gpt-j int4, use the following command:

model=gpt-j output_dir=<DATA_PATH> conda_path=<CONDA_ROOT_PATH> dtype=int4 bash download_data.sh  

Running Models by using automation scripts

First, go to automation folder to run models with scripts.

cd inference_results_v4.0/closed/Intel/code/automation

Customize your test by the following parameters:

DATA_DIR=<output_dir of download_data.sh>
OUTPUT_DIR=<The directory to save logs and results>
SUFFIX=<suffix to avoid duplicate container names>

Create output diretory if it does not exist:

mkdir -p ${OUTPUT_DIR}

Launch complete benchmark by using the following commands:

More options to customize your tests:

usage: run.py [-h] -n {rnnt,bert-99,3d-unet-99.9,resnet50,retinanet,dlrm-v2-99.9}
              [-i IMPLEMENTATION] [-y {int8,bf16,fp16,int4,fp32,mix}] -d DATASET -m MODEL_DIR
              -t OUTPUT -x CONTAINER_NAME_SUFFIX [-p] [-a] [-o] [-s] [-c] [-b] [-r] [-z]

options:
  -h, --help            show this help message and exit
  -n {rnnt,bert-99,3d-unet-99.9,resnet50,retinanet,dlrm-v2-99.9}, --model {rnnt,bert-99,3d-unet-99.9,resnet50,retinanet,dlrm-v2-99.9}
                        Benchmarking model
  -i IMPLEMENTATION, --implementation IMPLEMENTATION
                        Implementation id
  -y {int8,bf16,fp16,int4,fp32,mix}, --dtype {int8,bf16,fp16,int4,fp32,mix}
                        Precision
  -d DATASET, --dataset-dir DATASET_DIR
                        path of the datasets
  -m MODEL_DIR, --model-dir MODEL_DIR
                        path of the models
  -t OUTPUT, --output OUTPUT
                        path of the outputs
  -x CONTAINER_NAME_SUFFIX, --container-name-suffix CONTAINER_NAME_SUFFIX
                        The suffix of docker container name, used for avoiding name conflicts.
  -p, --performance-only
                        The option of running performance test only.
  -a, --accuracy-only   The option of running accuracy test only.
  -o, --offline-only    The option of running offline scenario only.
  -s, --server-only     The option of running server scenario only.
  -c, --compliance-only
                        The option of running compliance test only.
  -b, --skip-docker-build
                        The option of skipping building docker image.
  -u, --skip-create-container
                        The option of skipping docker build and container creation.  
  -r, --skip-data-preprocess
                        The option of skipping data preprocessing.
  -z, --ci-run
                        The option of running ci testings

To save time, rerun a benchmark directly within a running container without rebuilding the Docker image, creating the container and data preprocessing. For example, to rerun the bert-99 benchmark, use this command:

python3 run.py -n bert-99 -d ${DATA_DIR}/bert/dataset -m ${DATA_DIR}/bert/model -t ${OUTPUT_DIR} -x ${SUFFIX} -u -r

You can run a specific test of a benchmark in a specific mode. E.g. run bert-99 accuracy test at Offline mode by using the following command:

python3 run.py -n bert-99 -d ${DATA_DIR}/bert/dataset -m ${DATA_DIR}/bert/model -t ${OUTPUT_DIR} -x ${SUFFIX} -o -a

Test outputs

• Runtime log: mlperf.log
• Preprocessing log: ${OUTPUT_DIR}/preproc_<benchmark>_<implementation>_<precision>.log
• Results of Performance/Accuracy run: ${OUTPUT_DIR}/<division>/Intel/results/<system_desc_id>/<benchmark>/<scenario>
• Results of compliance test: ${OUTPUT_DIR}/<division>/Intel/compliance/<system_desc_id>/<benchmark>/<scenario>/<test_id>
• Measurments: ${OUTPUT_DIR}/<division>/Intel/measurements/<system_desc_id>/<benchmark>/<scenario>
• Runtime environment: ${OUTPUT_DIR}/env_<benchmark>_<impl>_<dtype>.log

Performing Submission Checker

1. Create ${OUTPUT_DIR}/<division>/<orgnization>/systems directory, and within it, create the <system_desc_id>.json and <system_desc_id>_<implementation_id>_<scenario>.json files according to submission rules.
2. Perform submission checker by running:

cd inference_results_v4.0/closed/Intel/code/automation
export TRUNCATED_OUTPUT_DIR=<a new directory to save truncated logs>
python3 submission_checker.py -i ${OUTPUT_DIR} -o ${TRUNCATED_OUTPUT_DIR}

Please make sure you have read permission of ${OUTPUT_DIR}, and r/w permission for ${TRUNCATED_OUTPUT_DIR} and /tmp

Trouble Shooting

TypeError: __init__() missing 1 required positional argument: 'status'

Solution: start docker serivce before launching the kit.

Performance regaresssion due to "Round Robin" MADT core enumeration

Solution: change "Round Robin" scheduling to "Linear"

Get the latest MLPerf 3.1 release

MLPerf is a benchmark for measuring the performance of machine learning systems. It provides a set of performance metrics for a variety of machine learning tasks, including image classification, object detection, machine translation, and others. The benchmark is representative of real-world workloads and as a fair and useful way to compare the performance of different machine learning systems.

In this document, we'll show how to use the publicly accessible codes and scritps on GitHub, which was published by Mlcommons, to run Intel MLPerf v3.1 submission with Intel optimized Docker images. The following contents will refer to this GitHub repository as .

Intel Docker Images for MLPerf

The Intel optimized Docker images for MLPerf v3.1 can be built using the Dockerfiles. Example for building docker image with Dockerfile:

Notice : please contact aice.mlperf@intel.com or aice@intel.com to get the latest patches for Intel mlperf inference release from mlcomm github repo

# Get the mlperf v3.1 workloads scritps from GitHub
git clone https://github.com/mlcommons/inference_results_v3.1.git
cd <THIS_REPO>/

# Get the latest patch from Intel to update Dockerfiles
git am 0001-update-dockerfiles-from-4.0-version.patch

# Use resnet50 docker build as an example to build all docker images
cd <THIS_REPO>/closed/Intel/code/resnet50/pytorch-cpu/docker/
bash build_resnet50_contanier.sh

HW configuration:

Recommmended BIOS Knobs:

Please also refer to Eagle Stream Platform Performance & Power Optimization Guide for more details.

Check System Health Using Intel® System Health Inspector:

Intel® System Health Inspector (aka svr-info) is a Linux OS utility for assessing the state and health of Intel Xeon computers. It is suggested to use svr-info first to check any system configuration issue before running any benchmark. Follow the Quick Start Guide for downloading and installation. The following are several key factors effecting the model performance.

Best Known Configurations:

sudo bash run_clean.sh

Known Issues

sysctl -r fs.file-max

sysctl -w fs.file-max=980000

Benchmarking using automation scripts

For your convinience, we prepare a set of automation scritps to help you download data, create docker, do data and model preprocessing, run accuracy, performance and compliance test in a batch. Please refer to ./automation/README.md for details about the usage. Example on for using automation scripts:

# Get the mlperf v3.1 workloads scritps from GitHub
git clone https://github.com/mlcommons/inference_results_v3.1.git

# Go to directory of automation scripts
cd <THIS_REPO>/closed/Intel/code/automation/

# Download dataset
bash download_dataset.sh <model> <location>
# <model> can be resnet50, retinanet, rnnt, 3d-unet-99.9, bert-99, gptj-99, or dlrm2-99.9
# <location> is where you save the data, which can be /data/mlperf_data

# Test model performance
PerformanceOnly="True" bash run.sh <model> <location>

# Test model Auccuracy
# Suppose you have done running the performance test workload, you can skip launching docker container and processing the data
Skip_docker_build="True" Skip_data_proprocess="True" AccuracyOnly="True" bash run.sh <model> <location>

For more details, please refer to the instructions in https://github.com/mlcommons/inferenceresultsv3.1/blob/main/closed/Intel/code/automation/README.md.

If you prefer to understand what the automation scripts do for you, we also provide instructions on how to run model performance/accuracy benchmarking step-by-step in the following sections.

Running models step-by-step

In the following sections, we'll show you how to set up and run each of the seven models:

• DLRM2
• GPT-J
• 3DUNET
• BERT
• RESNET50
• RETINANET
• RNNT

Note: All the codes and scripts are publicly accissible and can be downloaded from GitHub. The following sessions will refer this GitHub repository as .

Get started with DLRM2

cd <THIS_REPO>/closed/Intel/code/dlrm-99.9/pytorch-cpu-int8/docker
# Please firstly refer to the prerequisite file in the current directory to download the compiler before building the Docker image. 
bash build_dlrm-99.9_container.sh

cd /data/   # or path to where you want to store the data
mkdir -p /data/mlperf_data/dlrm_2/model/bf16
mkdir -p /data/mlperf_data/dlrm_2/data_npy

# Prepare DLRM dataset
# Create a directory (such as /data/mlperf_data/dlrm_2/data_npy) which contain:
#        day_23_dense.npy
#     day_23_sparse_multi_hot.npz
#     day_23_labels.npy
#
# Learn how to get the dataset from:
#     https://github.com/mlcommons/inference/tree/master/recommendation/dlrm_v2/pytorch

# Prepare pre-trained DLRM model
cd /data/mlperf_data/dlrm_2/model/bf16
wget https://cloud.mlcommons.org/index.php/s/XzfSeLgW8FYfR3S/download
unzip weights.zip

cd <THIS_REPO>/closed/Intel/code/dlrm-99.9/pytorch-cpu/
export MODEL_DIR=/data/mlperf_data/dlrm_2/model/bf16


# dump model from snapshot to torch
bash run_dump_torch_model.sh

docker run --name intel_inference_dlrm_int8 --privileged -itd --net=host --ipc=host \
  -v /path/of/dlrm:/data/dlrm_2_dataset mlperf_inference_dlrm2:3.1

docker exec -it intel_inference_dlrm_int8 bash

export http_proxy="your host proxy"
export https_proxy="your host proxy"

cd /opt/workdir/code/dlrm2-99.9/pytorch-cpu-int8
bash ./run_calibration.sh

# export model directory to saved model path
export MODEL_DIR=/data/mlperf_data/dlrm_2/model/bf16

# export dataset directory to saved dataset path where .npy .npz are stored.
export DATA_DIR=/data/mlperf_data/dlrm_2/data_npy

# offline performance
source setup_env_offline.sh
bash run_main.sh offline int8                   

# offline accuracy
source setup_env_offline.sh
bash run_main.sh offline accuracy int8              

# server performance
source setup_env_server.sh
bash run_main.sh server int8

# server accuracy
source setup_env_server.sh   
bash run_main.sh server accuracy int8    

Get Started with GPT-J

export WORKLOAD_DATA=/data/mlperf_data/gpt-j
mkdir -p ${WORKLOAD_DATA}

cd <THIS_REPO>/closed/Intel/code/gptj-99/pytorch-cpu/
python download-calibration-dataset.py --calibration-list-file calibration-list.txt --output-dir ${WORKLOAD_DATA}/calibration-data

python download-dataset.py --split validation --output-dir ${WORKLOAD_DATA}/validation-data

CHECKPOINT_DIR=${WORKLOAD_DATA}/gpt-j-checkpoint
wget https://cloud.mlcommons.org/index.php/s/QAZ2oM94MkFtbQx/download -O gpt-j-checkpoint.zip
unzip gpt-j-checkpoint.zip
mv gpt-j/checkpoint-final/ ${CHECKPOINT_DIR}

cd <THIS_REPO>/closed/Intel/code/gptj-99/pytorch-cpu/docker
bash build_gpt-j_container.sh

docker run --name intel_gptj --privileged -itd --net=host --ipc=host -v ${WORKLOAD_DATA}:/opt/workdir/code/gptj-99/pytorch-cpu/data mlperf_inference_gptj:3.1
docker exec -it intel_gptj bash

cd code/gptj-99/pytorch-cpu

source setup_env.sh
bash run_quantization.sh

bash run_offline.sh

bash run_offline_accuracy.sh

bash run_server.sh

bash run_server_accuracy.sh

Get Started with 3DUNET

cd <THIS_REPO>/closed/Intel/code/3d-unet-99.9/pytorch-cpu/docker
bash build_3dunet_container.sh

mkdir 3dunet
cd 3dunet
git clone https://github.com/neheller/kits19
cd kits19
pip3 install -r requirements.txt
python3 -m starter_code.get_imaging
cd ..

docker run --name intel_3dunet --privileged -itd -v /path/to/3dunet:/root/mlperf_data/3dunet-kits --net=host --ipc=host mlperf_inference_3dunet:3.1

docker exec -it intel_3dunet bash

export http_proxy="your host proxy"
export https_proxy="your host proxy"

cd code/3d-unet-99.9/pytorch-cpu/
bash process_data_model.sh 

# 3dunet only has offline mode
bash run.sh perf # offline performance
bash run.sh acc  # offline accuracy

Get started with BERT

cd <THIS_REPO>/closed/Intel/code/bert-99/pytorch-cpu/docker/

bash build_bert-99_contanier.sh

cd /data/mlperf_data   # or path to where you want to store the data
mkdir bert
cd bert
mkdir dataset
wget https://rajpurkar.github.io/SQuAD-explorer/dataset/dev-v1.1.json -O dataset/dev-v1.1.json
git clone https://huggingface.co/bert-large-uncased model
cd model
wget https://zenodo.org/record/4792496/files/pytorch_model.bin?download=1 -O pytorch_model.bin

docker run --name bert_3-1 --privileged -itd --net=host --ipc=host -v /path/of/bert:/data/mlperf_data/bert <bert docker image ID>

docker exec -it bert_3-1 bash

export http_proxy="your host proxy"
export https_proxy="your host proxy"

cd code/bert-99/pytorch-cpu
export DATA_PATH=/data/mlperf_data/bert
bash convert.sh

bash run.sh                    #offline performance
bash run.sh --accuracy         #offline accuracy
bash run_server.sh             #server performance
bash run_server.sh --accuracy  #server accuracy

Get Started with ResNet50

bash download_imagenet.sh

cd <THIS_REPO>/closed/Intel/code/resnet50/pytorch-cpu/docker/

bash build_resnet50_contanier.sh

docker run -v </path/to/ILSVRC2012_img_val>:/opt/workdir/code/resnet50/pytorch-cpu/ILSVRC2012_img_val -it --privileged <resnet docker image ID> /bin/bash

cd code/resnet50/pytorch-cpu

export http_proxy="your host proxy"
export https_proxy="your host proxy"

bash prepare_calibration_dataset.sh

bash download_model.sh

export DATA_CAL_DIR=calibration_dataset
export CHECKPOINT=resnet50-fp32-model.pth

bash generate_torch_model.sh

export DATA_DIR=${PWD}/ILSVRC2012_img_val
export RN50_START=models/resnet50-start-int8-model.pth
export RN50_END=models/resnet50-end-int8-model.pth
export RN50_FULL=models/resnet50-full.pth

bash run_offline.sh <batch_size>

bash run_server.sh

bash run_offline_accuracy.sh <batch_size>

bash run_server_accuracy.sh

Get Started with Retinanet

pip3 install --upgrade pip --user
pip3 install opencv-python-headless==4.5.3.56 pycocotools==2.0.2 fiftyone==0.16.5

CUR_DIR=$(pwd)
export WORKLOAD_DATA=${CUR_DIR}/data
mkdir -p ${WORKLOAD_DATA}

export ENV_DEPS_DIR=${CUR_DIR}/retinanet-env

bash openimages_mlperf.sh --dataset-path ${WORKLOAD_DATA}/openimages

bash openimages_calibration_mlperf.sh --dataset-path ${WORKLOAD_DATA}/openimages-calibration

wget --no-check-certificate 'https://zenodo.org/record/6617981/files/resnext50_32x4d_fpn.pth' -O 'retinanet-model.pth'
mv 'retinanet-model.pth' ${WORKLOAD_DATA}/

cd <THIS_REPO>/closed/Intel/code/retinanet/pytorch-cpu/docker/

bash build_retinanet_contanier.sh

docker run --name intel_retinanet --privileged -itd --net=host --ipc=host -v ${WORKLOAD_DATA}:/opt/workdir/code/retinanet/pytorch-cpu/data <resnet docker image ID> 

docker exec -it intel_retinanet bash 

cd code/retinanet/pytorch-cpu/

CUR_DIR=$(pwd)
export WORKLOAD_DATA=${CUR_DIR}/data
export CALIBRATION_DATA_DIR=${WORKLOAD_DATA}/openimages-calibration/train/data
export MODEL_CHECKPOINT=${WORKLOAD_DATA}/retinanet-model.pth
export CALIBRATION_ANNOTATIONS=${WORKLOAD_DATA}/openimages-calibration/annotations/openimages-mlperf-calibration.json
bash run_calibration.sh

export http_proxy="your host proxy"
export https_proxy="your host proxy"

source setup_env.sh

# Run one of these performance or accuracy scripts at a time
# since the log files will be overwritten on each run

# for offline performance
bash run_offline.sh

# for server performance
bash run_server.sh

# for offline accuracy
bash run_offline_accuracy.sh

# for server accuracy
bash run_server_accuracy.sh

Get Started with RNNT

cd <THIS_REPO>/closed/Intel/code/rnnt/pytorch-cpu/docker/
bash build_rnnt-99_container.sh

docker run --name intel_rnnt --privileged -itd -v /data/mlperf_data:/data/mlperf_data --net=host --ipc=host mlperf_inference_rnnt:3.1

docker ps -a #get container "id"
docker exec -it <id> bash
cd /opt/workdir/code/rnnt/pytorch-cpu

export LD_LIBRARY_PATH=/opt/workdir/code/rnnt/pytorch-cpu/third_party/lib:$LD_LIBRARY_PATH

export http_proxy="your host proxy"
export https_proxy="your host proxy"

 SKIP_BUILD=1 STAGE=0 bash run.sh

 SKIP_BUILD=1 STAGE=5 bash run.sh

Complinace Test

To run compliance test please follow https://github.com/mlcommons/inference/tree/master/compliance/nvidia or use automation script introduced in next section.

Get the latest MLPerf 3.0 release

Please follow the below commands to get the latest mlperf 3.0 release.

git clone https://github.com/mlcommons/inference_results_v3.0.git
cd inference_results_v3.0
wget https://raw.githubusercontent.com/intel-ai-tce/ai-documents/mlperf_patches/AEM/mlperf/patches/0001-updates-for-3.0-submission.patch
git am 0001-updates-for-3.0-submission.patch

Intel Docker Images for MLPerf

The Intel optimized Docker images for MLPerf v3.0 can be built using the Dockerfiles.
Please refer to "Build & Run Docker container from Dockerfile" sub-section in each model section.

Example for building docker image with Dockerfile:

cd inference_results_v3.0/closed/Intel/code/resnet50/pytorch-cpu/docker/

bash build_resnet50_contanier.sh

Validated HW configuration:

Recommmended BIOS Knobs:

Please also refer to Eagle Stream Platform Performance & Power Optimization Guide for more details.

Check System Health Using Intel® System Health Inspector:

Intel® System Health Inspector (aka svr-info) is a Linux OS utility for assessing the state and health of Intel Xeon computers. It is suggested to use svr-info first to check any system configuration issue before running any benchmark. Follow the Quick Start Guide for downloading and installation. The following are several key factors effecting the model performance.

CPU

Couple CPU features impact MLPerf performance via related BIOS knobs, so please double check the CPU features with your BIOS knobs. Some important CPU features are Hyperthreading, number of NUMA nodes, Prefetchers and Intel Turbo Boost.

Please also check your CPU tempartures. The CPU temparture should not be higher than 50 degrees C.
Overheating will drop the CPU frequency and degrade the MLPerf performance.

Power

We recommend the intel_pstate Frequency Driver.
For best performance, set the Frequency Governor and Power and Perf Policy to performance.
Here are related recommended power settings from svr-info.

Best Known Configurations:

sudo bash run_clean.sh

Running models:

In the following sections, we'll show you how to set up and run each of the six models:

• 3DUNET
• BERT
• DLRM
• RESNET50
• RETINANET
• RNNT

Get Started with 3DUNET

Build & Run Docker container from Dockerfile

If you haven't already done so, build the Intel optimized Docker image for 3DUNET using:

cd inference_results_v3.0/closed/Intel/code/3d-unet-99.9/pytorch-cpu/docker
bash build_3dunet_container.sh

Prerequisites

Use these commands to prepare the 3DUNET dataset and model on your host system:

mkdir 3dunet
cd 3dunet
git clone https://github.com/neheller/kits19
cd kits19
pip3 install -r requirements.txt
python3 -m starter_code.get_imaging
cd ..

Set Up Environment

Follow these steps to set up the docker instance and preprocess the data.

Start a Container

Use docker run to start a container with the optimized Docker image we pulled earlier. Replace /path/of/3dunet with the 3dunet folder path created earlier:

docker run --name intel_3dunet --privileged -itd -v /path/to/3dunet:/root/mlperf_data/3dunet-kits --net=host --ipc=host mlperf_inference_3dunet:3.0

Login to Docker Instance

Login into a bashrc shell in the Docker instance.

docker exec -it intel_3dunet bash

Preprocess Data

If you need a proxy to access the internet, replace your host proxy with the proxy server for your environment. If no proxy is needed, you can skip this step:

export http_proxy="your host proxy"
export https_proxy="your host proxy"

Preprocess the data and download the model using the provided script:

pip install numpy==1.23.5
cd code/3d-unet-99.9/pytorch-cpu/
bash process_data_model.sh 

Run the Benchmark

# 3dunet only has offline mode
bash run.sh perf # offline performance
bash run.sh acc  # offline accuracy

Get the Results

• Check log file. Performance results are in ./output/mlperf_log_summary.txt. Verify that you see results is: valid.

• For offline mode performance, check the field Samples per second:

• Accuracy results are in ./output/accuracy.txt. Check the field mean =.

• The performance result is controled by the value of "targetqps" in usersocket.conf file. The scripts will automatically select usersocket.conf file according to the number of sockets on customer's platform. Customers can also manully change the value of "targetqps" in corresponding usersocket.conf files.

Save these output log files elsewhere when each test is completed as they will be overwritten by the next test.

Get started with BERT

The docker container can be created either by building it using the Dockerfile or pulling the image from Dockerhub (if available).

Build & Run Docker container from Dockerfile

If you haven't already done so, build and run the Intel optimized Docker image for BERT using:

cd inference_results_v3.0/closed/Intel/code/bert-99/pytorch-cpu/docker/

bash build_bert-99_contanier.sh

Prerequisites

Use these commands to prepare the BERT dataset and model on your host system:

cd /data/mlperf_data   # or path to where you want to store the data
mkdir bert
cd bert
mkdir dataset
wget https://rajpurkar.github.io/SQuAD-explorer/dataset/dev-v1.1.json -O dataset/dev-v1.1.json
git clone https://huggingface.co/bert-large-uncased model
cd model
wget https://zenodo.org/record/4792496/files/pytorch_model.bin?download=1 -O pytorch_model.bin

Set Up Environment

Follow these steps to set up the docker instance and preprocess the data.

Start a Container

Use docker run to start a container with the optimized Docker image we pulled or built earlier. Replace /path/of/bert with the bert folder path created earlier (i.e. /data/mlperf_data/bert):

docker run --name bert_3-0 --privileged -itd --net=host --ipc=host \
  -v /path/of/bert:/data/mlperf_data/bert <bert docker image ID>

Login to Docker Instance

Login into a bashrc shell in the Docker instance.

docker exec -it bert_3-0 bash

Convert Dataset and Model

If you need a proxy to access the internet, replace your host proxy with the proxy server for your environment. If no proxy is needed, you can skip this step:

export http_proxy="your host proxy"
export https_proxy="your host proxy"

cd code/bert-99/pytorch-cpu
export DATA_PATH=/data/mlperf_data/bert
bash convert.sh

Run the Benchmark

bash run.sh                    #offline performance
bash run.sh --accuracy         #offline accuracy
bash run_server.sh             #server performance
bash run_server.sh --accuracy  #server accuracy

Get the Results

Check the performance log file ./test_log/mlperf_log_summary.txt:

• Verify you see results is: valid.
• For offline mode performance, check the field Samples per second:
• For server mode performance, check the field Scheduled samples per second:
• The performance results are controled by the value of "targetqps" in usersocket.conf file. The scripts will automatically select usersocket.conf file according to the number of sockets on customer's platform. Customers can also manully change the value of "targetqps" in corresponding usersocket.conf files.

Check the accuracy log file ./test_log/accuracy.txt.

• Check the field f1

Save these output log files elsewhere when each test is completed as they will be overwritten by the next test.

Get started with DLRM

Build & Run Docker container from Dockerfile

If you haven't already done so, build the Intel optimized Docker image for DLRM using:

# Please get compiler first.
cd inference_results_v3.0/closed/Intel/code/dlrm-99.9
wget https://registrationcenter-download.intel.com/akdlm/irc_nas/18679/l_HPCKit_p_2022.2.0.191.sh

# Build docker image
cd inference_results_v3.0/closed/Intel/code/dlrm-99.9/pytorch-cpu/docker
bash build_dlrm-99.9_container.sh

Prerequisites

Use these commands to prepare the Deep Learning Recommendation Model (DLRM) dataset and model on your host system:

cd /data/   # or path to where you want to store the data
mkdir -p /data/dlrm/model
mkdir -p /data/dlrm/terabyte_input

# download dataset
# Create a directory (such as /data/dlrm/terabyte_input) which contain:
#        day_fea_count.npz
#        terabyte_processed_test.bin
#
# Learn how to get the dataset from:
#     https://github.com/facebookresearch/dlrm
# You can also copy it using:
#     scp -r mlperf@10.112.230.156:/home/mlperf/dlrm_data/* /data/dlrm/terabyte_input
#
# download model
# Create a directory (such as /data/dlrm/model):
cd /data/dlrm/model
wget https://dlrm.s3-us-west-1.amazonaws.com/models/tb00_40M.pt -O dlrm_terabyte.pytorch

Set Up Environment

Follow these steps to set up the docker instance.

Start a Container

Use docker run to start a container with the optimized Docker image we pulled earlier. Replace /path/of/dlrm with the dlrm folder path created earlier (/data/dlrm for example):

docker run --name intel_inference_dlrm --privileged -itd --net=host --ipc=host \
  -v /path/of/dlrm:/data/mlperf_data/raw_dlrm mlperf_inference_dlrm:3.0

Login to Docker Container

Login into a bashrc shell in the Docker instance.

docker exec -it intel_inference_dlrm bash

Preprocess model and dataset

If you need a proxy to access the internet, replace your host proxy with the proxy server for your environment. If no proxy is needed, you can skip this step:

export http_proxy="your host proxy"
export https_proxy="your host proxy"

cd /opt/workdir/code/dlrm/pytorch-cpu
export MODEL=/data/mlperf_data/raw_dlrm/model
export DATASET=/data/mlperf_data/raw_dlrm/terabyte_input
export DUMP_PATH=/data/mlperf_data/dlrm
bash dump_model_dataset.sh

Run the Benchmark

export MODEL_DIR=/data/mlperf_data/dlrm
export DATA_DIR=/data/mlperf_data/dlrm

bash runcppsut                     # offline performance
bash runcppsut accuracy               # offline accuracy
bash runcppsut performance server  # server performance
bash runcppsut accuracy server     # server accuracy

Get the Results

Check the appropriate offline or server performance log file, either ./output/PerformanceOnly/Offline/mlperf_log_summary.txt or ./output/PerformanceOnly/Server/mlperf_log_summary.txt:

• Verify you see results is: valid.
• For offline mode performance, check the field Samples per second:
• For server mode performance, check the field Scheduled samples per second:
• The performance result is controled by the value of "targetqps" in usersocket.conf file. The scripts will automatically select usersocket.conf file according to the number of sockets on customer's platform. Customers can also manully change the value of "targetqps" in corresponding usersocket.conf files.

Check the appropriate offline or server accuracy log file, either ./output/AccuracyOnly/Offline/accuracy.txt or ./output/AccuracyOnly/Server/accuracy.txt:

• Check the field AUC

Save these output log files elsewhere when each test is completed as they will be overwritten by the next test.

Get Started with ResNet50

The docker container can be created either by building it using the Dockerfile or pulling the image from Dockerhub (if available). Please download the Imagenet dataset on the host system before starting the container.

Download Imagenet Dataset for Calibration

Download ImageNet (50000) dataset

bash download_imagenet.sh

Build & Run Docker container from Dockerfile

If you haven't already done so, build and run the Intel optimized Docker image for ResNet50 using:

cd inference_results_v3.0/closed/Intel/code/resnet50/pytorch-cpu/docker/

bash build_resnet50_contanier.sh

docker run -v </path/to/ILSVRC2012_img_val>:/opt/workdir/code/resnet50/pytorch-cpu/ILSVRC2012_img_val -it --privileged <resnet docker image ID> /bin/bash

cd code/resnet50/pytorch-cpu

Prepare Calibration Dataset & Download Model ( Inside Container )

If you need a proxy to access the internet, replace your host proxy with the proxy server for your environment. If no proxy is needed, you can skip this step:

export http_proxy="your host proxy"
export https_proxy="your host proxy"

Prepare calibration 500 images into folders

cd /opt/workdir/code/resnet50/pytorch-cpu
bash prepare_calibration_dataset.sh

Download the model

bash download_model.sh

The downloaded model will be saved as resnet50-fp32-model.pth

Quantize Torchscript Model and Check Accuracy

• Set the following paths:

export DATA_CAL_DIR=calibration_dataset
export CHECKPOINT=resnet50-fp32-model.pth

• Generate scales and models

bash generate_torch_model.sh

The start and end parts of the model are also saved (respectively named) in models

Run Benchmark (Common for Docker & Baremetal)

export DATA_DIR=${PWD}/ILSVRC2012_img_val
export RN50_START=models/resnet50-start-int8-model.pth
export RN50_END=models/resnet50-end-int8-model.pth
export RN50_FULL=models/resnet50-full.pth

Performance

• Offline

bash run_offline.sh <batch_size>

• Server

bash run_server.sh

Accuracy

• Offline

bash run_offline_accuracy.sh <batch_size>

• Server

bash run_server_accuracy.sh

Get the Results

Check the ./mlperf_log_summary.txt log file:

• Verify you see results is: valid.
• For offline mode performance, check the field Samples per second:
• For server mode performance, check the field Scheduled samples per second:
• The performance result is controled by the value of "targetqps" in ./src/usersocket.conf file. The scripts will automatically select usersocket.conf file according to the number of sockets on customer's platform. Customers can also manully change the value of "targetqps" in corresponding usersocket.conf files.

Check the ./offline_accuracy.txt or ./server_accuracy.txt log file:

• Check the field accuracy

Save these output log files elsewhere when each test is completed as they will be overwritten by the next test.

Get Started with Retinanet

The docker container can be created either by building it using the Dockerfile or pulling the image from Dockerhub (if available). Please download the Imagenet dataset on the host system before starting the container.

Download the dataset

• Install dependencies (python3.9 or above)

pip3 install --upgrade pip --user
pip3 install opencv-python-headless==4.5.3.56 pycocotools==clear2.0.2 fiftyone==0.16.5

• Setup env vars

CUR_DIR=$(pwd)
export WORKLOAD_DATA=${CUR_DIR}/data
mkdir -p ${WORKLOAD_DATA}

export ENV_DEPS_DIR=${CUR_DIR}/retinanet-env

• Download OpenImages (264) dataset

bash openimages_mlperf.sh --dataset-path ${WORKLOAD_DATA}/openimages

Images are downloaded to ${WORKLOAD_DATA}/openimages

• Download Calibration images

bash openimages_calibration_mlperf.sh --dataset-path ${WORKLOAD_DATA}/openimages-calibration

Calibration dataset downloaded to ${WORKLOAD_DATA}/openimages-calibration

Note: If you meet any obstacles on downloading the dataset, please try again in the docker container to be launched after [Build & Run Docker container from Dockerfile](Build & Run Docker container from Dockerfile).

Download Model

wget --no-check-certificate 'https://zenodo.org/record/6617981/files/resnext50_32x4d_fpn.pth' -O 'retinanet-model.pth'
mv 'retinanet-model.pth' ${WORKLOAD_DATA}/

Build & Run Docker container from Dockerfile

If you haven't already done so, build and run the Intel optimized Docker image for Retinanet using:

cd inference_results_v3.0/closed/Intel/code/retinanet/pytorch-cpu/docker/

bash build_retinanet_contanier.sh

docker run --name intel_retinanet --privileged -itd --net=host --ipc=host -v ${WORKLOAD_DATA}:/opt/workdir/code/retinanet/pytorch-cpu/data <retinanet docker image ID> 

docker exec -it intel_retinanet bash 

cd code/retinanet/pytorch-cpu/

Calibrate and generate torchscript model

If you need a proxy to access the internet, replace your host proxy with the proxy server for your environment. If no proxy is needed, you can skip this step:

export http_proxy="your host proxy"
export https_proxy="your host proxy"

Run Calibration

CUR_DIR=$(pwd)
export WORKLOAD_DATA=${CUR_DIR}/data
export CALIBRATION_DATA_DIR=${WORKLOAD_DATA}/openimages-calibration/train/data
export MODEL_CHECKPOINT=${WORKLOAD_DATA}/retinanet-model.pth
export CALIBRATION_ANNOTATIONS=${WORKLOAD_DATA}/openimages-calibration/annotations/openimages-mlperf-calibration.json

cd /opt/workdir/code/retinanet/pytorch-cpu/retinanet-env/vision
git checkout 8e078971b8aebdeb1746fea58851e3754f103053
python setup.py install && python setup.py develop

cd /opt/workdir/code/retinanet/pytorch-cpu
bash run_calibration.sh

Set Up Environment

Export the environment settings

source setup_env.sh

Run the Benchmark

# Run one of these performance or accuracy scripts at a time
# since the log files will be overwritten on each run

# for offline performance
bash run_offline.sh

# for server performance
bash run_server.sh

# for offline accuracy
bash run_offline_accuracy.sh

# for server accuracy
bash run_server_accuracy.sh

Get the results

Check the ./mlperf_log_summary.txt log file:

• Verify you see results is: valid.
• For offline mode performance, check the field Samples per second:
• For server mode performance, check the field Scheduled samples per second:
• The performance result is controled by the value of "targetqps" in usersocket.conf file. The scripts will automatically select usersocket.conf file according to the number of sockets on customer's platform. Customers can also manully change the value of "targetqps" in corresponding usersocket.conf files.

Check the ./accuracy.txt log file:

• Check the field mAP

Save these output log files elsewhere when each test is completed as they will be overwritten by the next test.

Get Started with RNNT

Build & Run Docker container from Dockerfile

If you haven't already done so, build the Intel optimized Docker image for RNNT using:

cd inference_results_v3.0/closed/Intel/code/rnnt/pytorch-cpu/docker/
bash build_rnnt-99_container.sh

Set Up Environment

Follow these steps to set up the docker instance.

Start a Container

Use docker run to start a container with the optimized Docker image we built earlier.

docker run --name intel_rnnt --privileged -itd -v /data/mlperf_data:/data/mlperf_data \
--net=host --ipc=host mlperf_inference_rnnt:3.0

Login to Docker Container

Get the Docker container ID and login into a bashrc shell in the Docker instance using docker exec.

docker ps -a #get container "id"
docker exec -it <id> bash
cd /opt/workdir/code/rnnt/pytorch-cpu

• Setup env vars

export LD_LIBRARY_PATH=/opt/workdir/code/rnnt/pytorch-cpu/third_party/lib:$LD_LIBRARY_PATH

If you need a proxy to access the internet, replace your host proxy with the proxy server for your environment. If no proxy is needed, you can skip this step:

export http_proxy="your host proxy"
export https_proxy="your host proxy"

Run the Benchmark

The provided run.sh script abstracts the end-to-end process for RNNT:

Run run.sh with STAGE=0 to invoke all the steps requried to run the benchmark (i.e download the model & dataset, preprocess the data, calibrate and build the model):

 SKIP_BUILD=1 STAGE=0 bash run.sh

or to skip to stage 5 without previous steps: Offline/Server accuracy and benchmark:

 SKIP_BUILD=1 STAGE=5 bash run.sh

Get the Results

Check the appropriate offline or server performance log files, either ./logs/Server/performance/.../mlperf_log_summary.txt or ./logs/Offline/performance/.../mlperf_log_summary.txt:

• Verify you see results is: valid.
• For offline mode performance, check the field Samples per second:
• For server mode performance, check the field Scheduled samples per second:
• The performance result is controled by the value of "targetqps" in ./configs/usersocket.conf file. The scripts will automatically select usersocket.conf file according to the number of sockets on customer's platform. Customers can also manully change the value of "targetqps" in corresponding usersocket.conf files.

Check the appropriate offline or server accuracy log file, either ./logs/Server/accuracy/.../mlperf_log_summary.txt or ./logs/Offline/accuracy/.../mlperf_log_summary.txt:

Save these output log files elsewhere when each test is completed as they will be overwritten by the next test.