dll-bench is a command line tool that makes it easy for you to benchmark the model on different platforms.
With djl-bench, you can easily compare your model’s behavior in different use cases, such as:
djl-bench currently support benchmark the following type of models:
You can build djl-bench from source if you need to benchmark fastText/BlazingText/Sentencepiece models.
brew install cask djl-bench
sudo snap install djlbench --classic sudo snap alias djlbench djl-bench
curl -O https://publish.djl.ai/djl-bench/0.14.0/djl-bench_0.13.0-1_all.deb sudo dpkg -i djl-bench_0.13.0-1_all.deb
For centOS or Amazon Linux 2
You can download djl-bench zip file from here.
curl -O https://publish.djl.ai/djl-bench/0.14.0/benchmark-0.13.0.zip unzip benchmark-0.13.0.zip rm benchmark-0.13.0.zip sudo ln -s $PWD/benchmark-0.13.0/bin/benchmark /usr/bin/djl-bench
We are considering to create a
chocolatey package for Windows. For the time being, you can
download djl-bench zip file from here.
Or you can run benchmark using gradle:
cd djl gradlew benchmark --args="--help"
Please ensure Java 8+ is installed and you are using an OS that DJL supported with.
After that, you need to clone the djl project and
cd into the folder.
DJL supported OS:
If you are trying to use GPU, please ensure the CUDA driver is installed. You can verify that through:
to checkout the version. For different Deep Learning engine you are trying to run the benchmark, they have different CUDA version to support. Please check the individual Engine documentation to ensure your CUDA version is supported.
Here is a few sample benchmark script for you to refer. You can also skip this and directly follow the 4-step instructions for your own model.
Benchmark on a Tensorflow model from tfhub url with all-zeros NDArray input for 10 times:
djl-bench -e TensorFlow -u https://tfhub.dev/tensorflow/resnet_50/classification/1 -c 10 -s 1,224,224,3
Similarly, this is for PyTorch
djl-bench -e PyTorch -u https://alpha-djl-demos.s3.amazonaws.com/model/djl-blockrunner/pytorch_resnet18.zip -n traced_resnet18 -c 10 -s 1,3,224,224
Benchmark a model from ONNX Model Zoo
djl-bench -e OnnxRuntime -u https://s3.amazonaws.com/onnx-model-zoo/resnet/resnet18v1/resnet18v1.tar.gz -s 1,3,224,224 -n resnet18v1/resnet18v1 -c 10
Resnet50 image classification model:
djl-bench -c 2 -s 1,3,224,224 -u djl://ai.djl.mxnet/resnet/0.0.1/resnet50_v2
SSD object detection model:
djl-bench -e PyTorch -c 2 -s 1,3,300,300 -u djl://ai.djl.pytorch/ssd/0.0.1/ssd_300_resnet50
To start your benchmarking, we need to make sure we provide the following information.
The benchmark script located here.
Just do the following:
This will print out the possible arguments to pass in:
usage: djl-bench [-p MODEL-PATH] -s INPUT-SHAPES [OPTIONS] -c,--iteration <ITERATION> Number of total iterations. -d,--duration <DURATION> Duration of the test in minutes. -e,--engine <ENGINE-NAME> Choose an Engine for the benchmark. -g,--gpus <NUMBER_GPUS> Number of GPUS to run multithreading inference. -h,--help Print this help. -l,--delay <DELAY> Delay of incremental threads. --model-arguments <MODEL-ARGUMENTS> Specify model loading arguments. --model-options <MODEL-OPTIONS> Specify model loading options. -n,--model-name <MODEL-NAME> Specify model file name. --neuron-cores <NEURON-CORES> Number of neuron cores to run multithreading inference, See https://awsdocs-neuron.readthedocs-hosted.com. -o,--output-dir <OUTPUT-DIR> Directory for output logs. -p,--model-path <MODEL-PATH> Model directory file path. -s,--input-shapes <INPUT-SHAPES> Input data shapes for the model. -t,--threads <NUMBER_THREADS> Number of inference threads. -u,--model-url <MODEL-URL> Model archive file URL.
By default, the above script will use MXNet as the default Engine, but you can always change that by adding the followings:
-e TensorFlow # TensorFlow -e PyTorch # PyTorch -e MXNet # Apache MXNet -e PaddlePaddle # PaddlePaddle -e OnnxRuntime # pytorch -e TFLite # TFLite -e TensorRT # TensorRT -e DLR # Neo DLR -e XGBoost # XGBoost -e Python # Python script
DJL accept variety of models came from different places.
--model-url option to load a model from a URL. The URL must point to an archive file.
The following is a pytorch model
We would recommend to make model files in a zip for better file tracking.
--model-path option to load model from a local directory or an archive file.
-p /home/ubuntu/models/pytorch_resnet18 or -p /home/ubuntu/models/pytorch_resnet18.zip
-p C:\models\pytorch_resnet18 or -p C:\models\pytorch_resnet18.zip
If the model file name is different from the parent folder name (or the archive file name), you need
--model-name in the
-c inside with a number
This will run 1000 times inference.
The benchmark script uses dummy NDArray inputs.
It will make fake NDArrays (like
NDArray.ones) to feed in the model for inference.
If we would like to fake an image:
This will create a NDArray (DataType FLOAT32) of shape(1, 3, 224, 224).
If your model requires multiple inputs like three NDArrays with shape 1, 384 and 384. You can do the followings:
If you input
DataType is not FLOAT32, you can specify the data type with suffix:
You can also do multi-threading inference with DJL. For example, if you would like to run the inference with 10 threads:
Best thread number for your system: The same number of cores your system have or double of the total cores.
You can also add
-l to simulate the increment load for your inference server. It will add threads with the delay of time.
-t 10 -l 100
The above code will create 10 threads with the wait time of 100ms.
For different purposes, we designed different mode you can play with. Such as the following arg:
This will ask the benchmark script repeatedly running the designed task for 86400 seconds (24 hour). If you would like to make sure DJL is stable in the long run, you can do that.
You can also keep monitoring the DJL memory usages by enable the following flag:
The memory report will be made available in