Note that this repository is outdated: we are now using the next generation of the MLCommons CK workflow automation meta-framework (Collective Mind aka CM) developed by the open working group. Feel free to join this community effort to learn how to modularize ML Systems and automate their benchmarking, optimization and deployment in the real world!

Collective Knowledge workflows for MLPerf

Linux/MacOS:

All CK components from the community are now aggregated in one CK repository.

News

April 2021 We are very excited to join forces with OctoML.ai! Contact Grigori Fursin for more details!
March 2021 For your convenience, all CK components for ML Systems are now aggregated in one GitHub repository! They can be also searched for at the cKnowledge.io portal!
March 2021 See our ACM TechTalk about the CK technology, reproducible research, FAIR principles and MLPerf.
March 2021 The overview of the CK technology has appeared in the Philosophical Transactions A, the world's longest-running journal where Newton published: DOI, ArXiv.

Installation
Inference v0.5
- Unofficial CK workflows
- CK workflows for official application with Docker
  - Datasets
  - Models
- CK workflows for official application without Docker
  - Prerequisites
  - Modify run_local.sh
  - Use run_local_sh

Installation

Install CK

$ python -m pip install ck --user
$ ck version
V1.11.1

Pull CK repositories

Pull repos (recursively, pulls ck-env, ck-tensorflow, etc.):

$ ck pull repo:ck-mlperf

MLPerf Inference v0.5

Using CK is optional for MLPerf Inference v0.5.

Unofficial CK workflows

We (unofficially) support two tasks out of three (i.e. except for Machine Translation). Full instructions are provided in the official MLPerf Inference repository:

CK workflows for official application with Docker

You can run the official vision application with CK model and dataset packages.

Install datasets

ImageNet 2012 validation dataset

Download the original dataset and auxiliaries:

$ ck install package --tags=image-classification,dataset,imagenet,val,original,full
$ ck install package --tags=image-classification,dataset,imagenet,aux

Copy the labels next to the images:

$ ck locate env --tags=image-classification,dataset,imagenet,val,original,full
/home/dvdt/CK-TOOLS/dataset-imagenet-ilsvrc2012-val
$ ck locate env --tags=image-classification,dataset,imagenet,aux
/home/dvdt/CK-TOOLS/dataset-imagenet-ilsvrc2012-aux
$ cp `ck locate env --tags=aux`/val.txt `ck locate env --tags=val`/val_map.txt

COCO 2017 validation dataset

$ ck install package --tags=object-detection,dataset,coco,2017,val,original
$ ck locate env --tags=object-detection,dataset,coco,2017,val,original
/home/dvdt/CK-TOOLS/dataset-coco-2017-val

Install and run TensorFlow models

NB: It is currently necessary to create symbolic links if a model's file name is different from the one hardcoded in the application for each profile. For example, for the tf-mobilenet profile (which can be used both for the non-quantized and quantized MobileNet TF models), the application specifies mobilenet_v1_1.0_224_frozen.pb , but the quantized model's file is mobilenet_v1_1.0_224_quant_frozen.pb.

ResNet

$ ck install package --tags=mlperf,image-classification,model,tf,resnet
$ export MODEL_DIR=`ck locate env --tags=model,tf,resnet`
$ export DATA_DIR=`ck locate env --tags=dataset,imagenet,val`
$ export EXTRA_OPS="--accuracy --count 50000 --scenario SingleStream"
$ ./run_and_time.sh tf resnet cpu
...
TestScenario.SingleStream qps=1089.79, mean=0.0455, time=45.880, acc=76.456, queries=50000, tiles=50.0:0.0447,80.0:0.0465,90.0:0.0481,95.0:0.0501,99.0:0.0564,99.9:0.0849

MobileNet non-quantized

$ ck install package --tags=mlperf,image-classification,model,tf,mobilenet,non-quantized
$ export MODEL_DIR=`ck locate env --tags=model,tf,mobilenet,non-quantized`
$ export DATA_DIR=`ck locate env --tags=dataset,imagenet,val`
$ export EXTRA_OPS="--accuracy --count 50000 --scenario Offline"
$ ./run_and_time.sh tf mobilenet cpu
...
TestScenario.Offline qps=352.92, mean=3.2609, time=4.534, acc=71.676, queries=1600, tiles=50.0:2.9725,80.0:4.0271,90.0:4.0907,95.0:4.3719,99.0:4.4811,99.9:4.5173

MobileNet quantized

$ ck install package --tags=mlperf,image-classification,model,tf,mobilenet,quantized
$ ln -s `ck locate env --tags=mobilenet,quantized`/mobilenet_v1_1.0_224{_quant,}_frozen.pb`
$ export MODEL_DIR=`ck locate env --tags=model,tf,mobilenet,quantized`
$ export DATA_DIR=`ck locate env --tags=dataset,imagenet,val`
$ export EXTRA_OPS="--accuracy --count 50000 --scenario Offline"
$ ./run_and_time.sh tf mobilenet cpu
...
TestScenario.Offline qps=128.83, mean=7.5497, time=12.419, acc=70.676, queries=1600, tiles=50.0:7.8294,80.0:11.1442,90.0:11.7616,95.0:12.1174,99.0:12.9126,99.9:13.1641

SSD-MobileNet non-quantized

$ ck install package --tags=mlperf,object-detection,model,tf,ssd-mobilenet,non-quantized
$ ln -s `ck locate env --tags=model,tf,ssd-mobilenet,non-quantized`/{frozen_inference_graph.pb,ssd_mobilenet_v1_coco_2018_01_28.pb}
$ export MODEL_DIR=`ck locate env --tags=model,tf,ssd-mobilenet,non-quantized`
$ export DATA_DIR=`ck locate env --tags=dataset,coco,2017,val`
$ export EXTRA_OPS="--accuracy --count 5000 --scenario Offline"
$ ./run_and_time.sh tf ssd-mobilenet cpu
...
TestScenario.Offline qps=5.82, mean=8.0406, time=27.497, acc=93.312, mAP=0.235, queries=160, tiles=50.0:6.7605,80.0:10.3870,90.0:10.4632,95.0:10.4788,99.0:10.4936,99.9:10.5068

SSD-MobileNet quantized

$ ck install package --tags=mlperf,object-detection,model,tf,ssd-mobilenet,quantized
$ ln -s `ck locate env --tags=model,tf,ssd-mobilenet,quantized`/{graph.pb,ssd_mobilenet_v1_coco_2018_01_28.pb}
$ export MODEL_DIR=`ck locate env --tags=model,tf,ssd-mobilenet,quantized`
$ export DATA_DIR=`ck locate env --tags=dataset,coco,2017,val`
$ export EXTRA_OPS="--accuracy --count 5000 --scenario Offline"
$ ./run_and_time.sh tf ssd-mobilenet cpu
...
TestScenario.Offline qps=5.46, mean=9.4975, time=29.310, acc=94.037, mAP=0.239, queries=160, tiles=50.0:7.9843,80.0:12.2297,90.0:12.3646,95.0:12.3965,99.0:12.4229,99.9:12.4351

SSD-ResNet

TODO

CK workflows for official application without Docker

Install prerequisites

To run the official vision app natively (i.e. without Docker), first install Python prerequisites such as OpenCV, TensorFlow and COCO Python API:

$ ck detect soft --tags=compiler,python --full_path=`which python3`
$ ck install package --tags=lib,tensorflow,v1.14,vcpu,vprebuilt
$ ck install package --tags=lib,python-package,cv2
$ ck install package --tags=tool,coco,api

Then, install the latest LoadGen package:

$ ck install package --tags=mlperf,inference,source,upstream.master
$ ck install package --tags=lib,python-package,absl
$ ck install package --tags=lib,python-package,mlperf,loadgen

NB: The most important thing during installation is to select the same version of Python 3 (if you have more than one registered with CK). Check that each package "needs" exactly the same version of Python 3 after installation:

$ ck show env --tags=lib,tensorflow,v1.14,vcpu,vprebuilt
Env UID:         Target OS: Bits: Name:                              Version: Tags:
087035468886d589   linux-64    64 TensorFlow library (prebuilt, cpu) 1.14.0   64bits,channel-stable,host-os-linux-64,lib,needs-python,needs-python-3.6.7,target-os-linux-64,tensorflow,tensorflow-cpu,tf,tf-cpu,v1,v1.14,v1.14.0,vcpu,vprebuilt

$ ck show env --tags=lib,python-package,cv2
Env UID:         Target OS: Bits: Name:                 Version: Tags:
5f31d16b444d6b8c   linux-64    64 Python OpenCV library 3.6.7    64bits,cv2,host-os-linux-64,lib,needs-python,needs-python-3.6.7,opencv,python-package,target-os-linux-64,v3,v3.6,v3.6.7

$ ck show env --tags=tool,coco,api
Env UID:         Target OS: Bits: Name:            Version: Tags:
885a8f71bf1219da   linux-64    64 COCO dataset API master   64bits,api,coco,compiled-by-gcc,compiled-by-gcc-8.3.0,host-os-linux-64,needs-python,needs-python-3.6.7,target-os-linux-64,tool,v0,vmaster,vtrunk

$ ck show env --tags=lib,python-package,mlperf,loadgen
Env UID:         Target OS: Bits: Name:                            Version: Tags:
462592cb2beeaf63   linux-64    64 MLPerf Inference LoadGen library master   64bits,host-os-linux-64,lib,loadgen,mlperf,mlperf-loadgen,mlperf_loadgen,needs-python,needs-python-3.6.7,python-package,target-os-linux-64,v0,vmaster

Modify `run_local.sh`

Modify the run_local.sh script under v0.5/classification_and_detection as follows:

$ git diff
diff --git a/v0.5/classification_and_detection/run_local.sh b/v0.5/classification_and_detection/run_local.sh
index 1262991..7597403 100755
--- a/v0.5/classification_and_detection/run_local.sh
+++ b/v0.5/classification_and_detection/run_local.sh
@@ -9,5 +9,5 @@ if [ ! -d $OUTPUT_DIR ]; then
     mkdir -p $OUTPUT_DIR
 fi
 
-python python/main.py --profile $profile $common_opt --model $model_path $dataset \
-    --output $OUTPUT_DIR $EXTRA_OPS $@
+ck virtual env --tag_groups="lib,tensorflow-cpu,v1.14,vcpu,vprebuilt lib,python-package,cv2 tool,coco lib,python-package,mlperf,loadgen" \
+--shell_cmd="python3.6 python/main.py --profile $profile $common_opt --model $model_path $dataset --output $OUTPUT_DIR $EXTRA_OPS $@"

NB: Use exactly the same Python version as your prerequisites "need" (only the major and minor version numbers e.g. 3.6, not 3.6.7).

Use `run_local.sh`

See above for how to specify datasets and models.

Example: MobileNet non-quantized

$ ck install package --tags=mlperf,image-classification,model,tf,mobilenet,non-quantized
$ export MODEL_DIR=`ck locate env --tags=model,tf,mobilenet,non-quantized`
$ export DATA_DIR=`ck locate env --tags=dataset,imagenet,val`
$ export EXTRA_OPS="--count 1024 --scenario Offline"
$ ./run_local.sh tf mobilenet cpu
...
TestScenario.Offline qps=237.10, mean=3.3406, time=4.319, queries=1024, tiles=50.0:2.9683,80.0:4.2340,90.0:4.2692,95.0:4.2827,99.0:4.2932,99.9:4.2932

Cheap and reliable Node.js hosting starts at $3/month, and $1/month static HTML hosting

ctuning / ck-mlperf

Programming Languages

Collective Knowledge workflows for MLPerf

News

Table of Contents

Installation

Install CK

Pull CK repositories

MLPerf Inference v0.5

Unofficial CK workflows

CK workflows for official application with Docker

Install datasets

ImageNet 2012 validation dataset

COCO 2017 validation dataset

Install and run TensorFlow models

ResNet

MobileNet non-quantized

MobileNet quantized

SSD-MobileNet non-quantized

SSD-MobileNet quantized

SSD-ResNet

CK workflows for official application without Docker

Install prerequisites

Modify `run_local.sh`

Use `run_local.sh`

Example: MobileNet non-quantized

Cheap and reliable Node.js hosting starts at $3/month, and $1/month static HTML hosting

ctuning / ck-mlperf

Programming Languages

Collective Knowledge workflows for MLPerf

News

Table of Contents

Installation

Install CK

Pull CK repositories

MLPerf Inference v0.5

Unofficial CK workflows

CK workflows for official application with Docker

Install datasets

ImageNet 2012 validation dataset

COCO 2017 validation dataset

Install and run TensorFlow models

ResNet

MobileNet non-quantized

MobileNet quantized

SSD-MobileNet non-quantized

SSD-MobileNet quantized

SSD-ResNet

CK workflows for official application without Docker

Install prerequisites

Modify run_local.sh

Use run_local.sh

Example: MobileNet non-quantized

Modify `run_local.sh`

Use `run_local.sh`