test-abtf-model

Automating ABTF model benchmarking using CM without Docker

This repository contains MLCommons CM automation recipes to make it easier to prepare and benchmark different versions of ABTF models (public or private) with MLPerf loadgen across different software and hardware.

Install MLCommons CM automation framework

Follow this online guide to install CM for your OS with minimal dependencies.

Install virtual environment

We tested these automation on Ubuntu and Windows.

We suggest you to create a virtual environment to avoid messing up your Python installation.

All CM repositories, artifacts and cache will be resided inside this virtual environment and can be removed at any time without influencing your own environment!

Linux

python3 -m venv ABTF
mkdir ABTF/work
. ABTF/bin/activate ; export CM_REPOS=$PWD/ABTF/CM ; cd ABTF/work

Windows

python -m venv ABTF
md ABTF\work
call ABTF\Scripts\activate.bat & set CM_REPOS=%CD%\ABTF\CM & cd ABTF\work

Install required CM automation recipes

CM repositories with automations

cm pull repo mlcommons@cm4mlops --checkout=dev
cm pull repo mlcommons@cm4abtf --checkout=dev

Show installed CM repositories

cm show repo

Find a specific repo

cm find repo mlcommons@cm4abtf

Update all repositories at any time

cm pull repo

Turn off debug info when running CM scripts

cm set cfg --key.script.silent

Download ABTF models and sample image

cmr "download file _wget" --url="https://www.dropbox.com/scl/fi/cwyi6ukwih5qjblgh2m7x/baseline_8MP_ss_scales_fm1_5x5_all_ep60.pth?rlkey=okaq1s32leqnbjzloru206t50&st=70yoyy89&dl=0" --verify_ssl=no --md5sum=26845c3b9573ce115ef29dca4ae5be14
cmr "download file _wget" --url="https://www.dropbox.com/scl/fi/ljdnodr4buiqqwo4rgetu/baseline_4MP_ss_all_ep60.pth?rlkey=zukpgfjsxcjvf4obl64e72rf3&st=umfnx8go&dl=0" --verify_ssl=no --md5sum=75e56779443f07c25501b8e43b1b094f
cmr "download file _wget" --url="https://www.dropbox.com/scl/fi/9un2i2169rgebui4xklnm/baseline_8MP_ss_scales_all_ep60.pth?rlkey=sez3dnjep4waa09s5uy4r3wmk&st=z859czgk&dl=0" --verify_ssl=no --md5sum=1ab66f523715f9564603626e94e59c8c
cmr "download file _wget" --url="https://www.dropbox.com/scl/fi/0n7rmxxwqvg04sxk7bbum/0000008766.png?rlkey=mhmr3ztrlsqk8oa67qtxoowuh&dl=0" --verify_ssl=no --md5sum=903306a7c8bfbe6c1ca68fad6e34fe52 -s

Download ABTF model code and register in CM cache

cmr "get ml-model abtf-ssd-pytorch"

or specific branch:

cmr "get ml-model abtf-ssd-pytorch" --model_code_git_branch=cognata

Check the state of the CM cache

cm show cache

Find model code

cm show cache --tags=ml-model,abtf

Prepare workflow to benchmark ABTF model on host CPU

Next commands prepare environment to benchmark host CPU. Check these docs to benchmark other devices:

• CUDA-based device

Detect python from virtual env

cmr "get python3" --quiet

Install or detect PyTorch

cmr "get generic-python-lib _torch"
cmr "get generic-python-lib _torchvision"

If you want to install a specific version of PyTorch, you can specify it as follows:

cmr "get generic-python-lib _torch" --version=2.2.0
cmr "get generic-python-lib _torchvision" --version=0.17.0

Test ABTF Model with a sample image and prepare for loadgen

cmr "demo abtf ssd-resnet50 cognata pytorch inference" --model=baseline_8MP_ss_scales_all_ep60.pth --config=baseline_8MP_ss_scales_all --input=0000008766.png --output=0000008766_prediction_test_8MP.jpg --num-classes=15
cmr "demo abtf ssd-resnet50 cognata pytorch inference" --model=baseline_4MP_ss_all_ep60.pth --config=baseline_4MP_ss_all --input=0000008766.png --output=0000008766_prediction_test_4MP.jpg --num-classes=15

CM will load a workflow described by this simple YAML, call other CM scripts to detect, download or build missing deps for a given platform, prepare all environment variables and run benchmark using the following Python code.

You can run it in silent mode to skip CM workflow information using -s or --silent flag:

cmr "demo abtf ssd-resnet50 cognata pytorch inference" --model=baseline_4MP_ss_all_ep60.pth --config=baseline_4MP_ss_all --input=0000008766.png --output=0000008766_prediction_test_4MP.jpg --num-classes=15 -s

You can dump internal CM state with resolved dependencies, their versions and reproducibility README by adding flag --repro. You will find dump in the cm-repro directory of your current directory after script execution:

cmr "demo abtf ssd-resnet50 cognata pytorch inference" --model=baseline_4MP_ss_all_ep60.pth --config=baseline_4MP_ss_all --input=0000008766.png --output=0000008766_prediction_test_4MP.jpg --num-classes=15 -s --repro

Test ABTF model inference with loadgen

Build MLPerf loadgen

cmr "get mlperf inference loadgen _copy" --version=main

Run ABTF PyTorch model with loadgen

This example uses a universal python loadgen harness with PyTorch and ONNX backends with 1 real input saved as pickle:

cmr "demo abtf ssd-resnet50 cognata pytorch inference" --model=baseline_8MP_ss_scales_all_ep60.pth --config=baseline_8MP_ss_scales_all --input=0000008766.png --output=0000008766_prediction_test_8MP.jpg --num-classes=15
cmr "generic loadgen python _pytorch _custom _cmc" --samples=5 --modelsamplepath=0000008766.png.cpu.pickle --modelpath=baseline_8MP_ss_scales_all_ep60.pth --modelcfg.num_classes=15 --modelcfg.config=baseline_8MP_ss_scales_all

cmr "demo abtf ssd-resnet50 cognata pytorch inference" --model=baseline_4MP_ss_all_ep60.pth --config=baseline_4MP_ss_all --input=0000008766.png --output=0000008766_prediction_test_4MP.jpg --num-classes=15
cmr "generic loadgen python _pytorch _custom _cmc" --samples=5 --modelsamplepath=0000008766.png.cpu.pickle --modelpath=baseline_4MP_ss_all_ep60.pth --modelcfg.num_classes=15 --modelcfg.config=baseline_4MP_ss_all

Export PyTorch ABTF model to ONNX

cmr "demo abtf ssd-resnet50 cognata pytorch inference" --model=baseline_8MP_ss_scales_all_ep60.pth --config=baseline_8MP_ss_scales_all --input=0000008766.png --output=0000008766_prediction_test.jpg -s --export_model_to_onnx=baseline_8MP_ss_scales_all_ep60_opset17.onnx --export_model_to_onnx_opset=17
cmr "demo abtf ssd-resnet50 cognata pytorch inference" --model=baseline_4MP_ss_all_ep60.pth --config=baseline_4MP_ss_all --input=0000008766.png --output=0000008766_prediction_test.jpg -s --export_model_to_onnx=baseline_4MP_ss_all_ep60_opset17.onnx --export_model_to_onnx_opset=17

Run ABTF ONNX model with loadgen and random input

cm run script "generic loadgen python _onnxruntime" --samples=5 --modelpath=baseline_8MP_ss_scales_all_ep60_opset17.onnx --output_dir=results --repro -s
cm run script "generic loadgen python _onnxruntime" --samples=5 --modelpath=baseline_4MP_ss_all_ep60_opset17.onnx --output_dir=results --repro -s

Try quantization via HuggingFace quanto package

We added simple example to do basic and automatic quantization of the model to int8 using HugginFace's quanto package:

cm run script "demo abtf ssd-resnet50 cognata pytorch inference" 
     --ad.ml-model.model_code_git_branch=cognata \
     --model=baseline_8MP_ss_scales_all_ep60_state.pth \
     --config=baseline_8MP_ss_scales_all \
     --input=0000008766.png \
     --output=0000008766_prediction_test_8MP_quantized.jpg \
     --quantize_with_huggingface_quanto \
     --repro

Run inference with quantized model

cm run script "demo abtf ssd-resnet50 cognata pytorch inference" \
     --ad.ml-model.model_code_git_branch=cognata \
     --model=baseline_8MP_ss_scales_all_ep60_state_hf_quanto_qint8.pth \
     --config=baseline_8MP_ss_scales_all \
     --model_quanto \
     --input=0000008766.png \
     --output=0000008766_prediction_test_8MP_quantized_inference.jpg \
     --repro -s

Archive of exported models

You can find all exported models in this DropBox folder.

You can download individual files via CM as follows:

cmr "download file _wget" --url="https://www.dropbox.com/scl/fi/yj3m4vpudmlqgkdkaatk0/baseline_4MP_ss_all_ep60_opset17.onnx?rlkey=r9i3ew2hfajzyssvmwb0vytbg&st=vkb3fwc7&dl=0" --verify_ssl=no --md5sum=7fbd70f7e1a0c0fbc1464f85351bce4b

Test ABTF model with a Cognata sub-set

We have developed CM script to automate management and download of the MLCommons Cognata dataset. You just need to register here and obtain a private URL that you will need to enter once:

cmr "get raw dataset mlcommons-cognata" --serial_numbers=10002_Urban_Clear_Morning --group_names=Cognata_Camera_01_8M --file_names="Cognata_Camera_01_8M_ann.zip;Cognata_Camera_01_8M_ann_laneline.zip;Cognata_Camera_01_8M.zip"

cmr "demo abtf ssd-resnet50 cognata pytorch inference _dataset" --model=baseline_4MP_ss_all_ep60.pth --config=baseline_4MP_ss_all

cmr "demo abtf ssd-resnet50 cognata pytorch inference _dataset" --model=baseline_4MP_ss_all_ep60.pth --config=baseline_4MP_ss_all --visualize

Benchmark accuracy of ABTF model with MLPerf loadgen

Download Cognata data set via CM

We have developed a CM automation recipe to download different sub-sets of the MLCommons Cognata data set - you just need to provide a private URL after registering to access this dataset here:

cmr "get raw dataset mlcommons-cognata" --serial_numbers=10002_Urban_Clear_Morning --group_names=Cognata_Camera_01_8M --file_names="Cognata_Camera_01_8M_ann.zip;Cognata_Camera_01_8M_ann_laneline.zip;Cognata_Camera_01_8M.zip"

or in a more compact way via variation

cmr "get raw dataset mlcommons-cognata _abtf-demo"

and to some specific path

cmr "get raw dataset mlcommons-cognata _abtf-demo" --path=/cognata

or import the existing one

cmr "get raw dataset mlcommons-cognata _abtf-demo" --import=/cognata

Run basic MLPerf harness for ABTF model

We've prototyped a basic MLPerf harness with CM automation to benchmark ABTF model with Cognata data set. See CM script and MLPerf Python harness for Cognata and ABTF model. It should work on Linux and Windows with CPU and CUDA GPU.

You can run performance test with the ABTF model and Cognata datset on CPU as follows (use --count to select the number of samples):

cmr "download file _wget" --url="https://www.dropbox.com/scl/fi/9un2i2169rgebui4xklnm/baseline_8MP_ss_scales_all_ep60.pth?rlkey=sez3dnjep4waa09s5uy4r3wmk&st=z859czgk&dl=0" --verify_ssl=no --md5sum=1ab66f523715f9564603626e94e59c8c

cmr "get raw dataset mlcommons-cognata _abtf-demo" --import=/cognata

cm run script --tags=run-mlperf-inference,demo,abtf-model,_abtf-demo-model,_pytorch \
   --device=cpu \
   --dataset=cognata-8mp-pt \
   --model=$PWD/baseline_8MP_ss_scales_all_ep60.pth \
   --env.CM_ABTF_ML_MODEL_CONFIG=baseline_8MP_ss_scales_all \
   --env.CM_ABTF_NUM_CLASSES=15 \
   --env.CM_DATASET_MLCOMMONS_COGNATA_SERIAL_NUMBERS=10002_Urban_Clear_Morning \
   --env.CM_DATASET_MLCOMMONS_COGNATA_GROUP_NAMES=Cognata_Camera_01_8M \
   --env.CM_ABTF_ML_MODEL_TRAINING_FORCE_COGNATA_LABELS=yes \
   --env.CM_ABTF_ML_MODEL_SKIP_WARMUP=yes \
   --max_batchsize=1 \
   --count=2 \
   --precision=float32 \
   --implementation=mlcommons-python \
   --scenario=Offline \
   --mode=performance \
   --power=no \
   --adr.python.version_min=3.8 \
   --adr.compiler.tags=gcc \
   --output=$PWD/results \
   --rerun \
   --clean

You should see the following output:

======================================================================

Loading model ...


Loading dataset and preprocessing if needed ...
* Dataset path: /cognata
* Preprocessed cache path: /home/cmuser/CM/repos/mlcommons@cm4abtf/script/demo-ml-model-abtf-cognata-pytorch-loadgen/tmp-preprocessed-dataset


Cognata folders: ['10002_Urban_Clear_Morning']
Cognata cameras: ['Cognata_Camera_01_8M']


Scanning Cognata dataset ...
  Number of files found: 1020
  Time: 2.36 sec.

Preloading and preprocessing Cognata dataset on the fly ...
  Time: 0.62 sec.

TestScenario.Offline qps=3.73, mean=0.5024, time=0.537, queries=2, tiles=50.0:0.5024,80.0:0.5056,90.0:0.5067,95.0:0.5072,99.0:0.5076,99.9:0.5077

You can run MLPerf accuracy test by changing --mode=performance to --mode=accuracy in the above command line and rerun it. You should the output similar to the following:

======================================================================

Loading model ...


Loading dataset and preprocessing if needed ...
* Dataset path: /cognata
* Preprocessed cache path: /home/cmuser/CM/repos/mlcommons@cm4abtf/script/demo-ml-model-abtf-cognata-pytorch-loadgen/tmp-preprocessed-dataset


Cognata folders: ['10002_Urban_Clear_Morning']
Cognata cameras: ['Cognata_Camera_01_8M']


Scanning Cognata dataset ...
  Number of files found: 1020
  Time: 2.36 sec.

Preloading and preprocessing Cognata dataset on the fly ...
  Time: 0.64 sec.

=================================================
{   'classes': tensor([ 2,  3,  4,  5,  9, 10], dtype=torch.int32),
    'map': tensor(0.3814),
    'map_50': tensor(0.6143),
    'map_75': tensor(0.4557),
    'map_large': tensor(0.5470),
    'map_medium': tensor(0.4010),
    'map_per_class': tensor([0.0000, 0.1002, 0.6347, 0.7505, 0.1525, 0.6505]),
    'map_small': tensor(0.0037),
    'mar_1': tensor(0.2909),
    'mar_10': tensor(0.3833),
    'mar_100': tensor(0.3833),
    'mar_100_per_class': tensor([0.0000, 0.1000, 0.6500, 0.7500, 0.1500, 0.6500]),
    'mar_large': tensor(0.5500),
    'mar_medium': tensor(0.4000),
    'mar_small': tensor(0.0071)}
=================================================
TestScenario.Offline qps=1.81, mean=0.7691, time=1.107, acc=0.000%, mAP=38.140%, mAP_classes={'Props': 0.0, 'TrafficSign': 0.10024752467870712, 'Car': 0.6346888542175293, 'Van': 0.7504950761795044, 'Pedestrian': 0.1524752527475357, 'Truck': 0.6504950523376465}, queries=2, tiles=50.0:0.7691,80.0:0.7713,90.0:0.7721,95.0:0.7725,99.0:0.7728,99.9:0.7728
======================================================================

Please check CM CUDA docs to benchmark ABTF model using MLPerf loadgen.

Using Docker

We prepared a demo to use Docker with CM automation for ABTF. You can find and run scripts to build and run Docker (CPU & CUDA) with CM here.

You can then use above commands to benchmark ABTF model. Don't forget to set up your private GitHub tocken to be able to pull a [private GitHub repo with ABTF model](https://github.com/mlcommons/abtf-ssd-pytorch/tree/cognata

export CM_GH_TOKEN="YOUR TOKEN"

Further developments

Current CM for ABTF roadmap here. MLCommons will fund a CM engineer to automate MLPerf inference until the end of 2024 - further developments will be done in sync with the MLCommons Inference Working Group.

Feedback

Join MLCommons discord server with ABTF channel

Acknowledgments

CM automation and MLPerf harness prototyping was sponsored by cKnowledge.org based on feedback from ABTF members.