Skip to content

Latest commit

 

History

History
323 lines (263 loc) · 19.3 KB

README.md

File metadata and controls

323 lines (263 loc) · 19.3 KB
Raw

FBNet Model Inference

FBNet is a convolutional neural network architecture that utilises depthwise convolutions and an inverted residual structure for image classification. This directory contains a sample implementation of image classification with FBNet. It is targeted to run on Intel Discrete Graphics platforms (XPUs) by leveraging Intel® Extension for Pytorch.

The input to the model is a tensor representing a batch of input images that is fed to the FBNet model for inference. The output is a tensor representing the inputs classification results for all possible classes.

The sample supports two modes of execution:

  • A performance benchmarking mode where the sample executes FBNet inference based on a dummy tensor of the requested batch size over a specified number of input frames. This dummy tensor dataset is itself looped over repeatedly until a minimum test duration has been reached. Average throughput and latency values are reported.
  • A accuracy-check mode which takes in frames from the ImageNet 2012 validation dataset and measures the accuracy of the resulting classification against references contained in the dataset. A percentage score - both top 1 accuracy and top 5 accuracy values - of passing frames along with average throughput and latency values are reported.

The rest of this document covers more details about the model, dataset, and the control knobs for each mode of execution. Further, instructions are provided on how to use the scripts in this directory for execution in bare-metal and docker container environments.

Model and Sources

The sample uses FBNet model implementations from HuggingFace:

Model Documentation Weights
fbnetc_100 HuggingFace README fbnetc_100.rmsp_in1k

The model is downloaded from HuggingFace at runtime and a custom performance optimization patch for Intel® Data Center GPU Flex Series is automatically applied. This optimization replaces the models default BatchNormAct2d operation - nn.functional.bach_norm - which uses pytorch's backend implementation instead of IPEX with nn.BatchNorm2d which is supported by IPEX. This patch is applied in the file loader_utils. Its impact on performance when running with CUDA GPUs has not been verified.

Dataset

Note

Throughput and latency benchmarking can be done with dummy data (./run_model.sh --dummy). In such a case dataset setup can be skipped. As a downside expect to see low accuracy on the dummy data.

Note

~13.3 GB of free disk space is required to download and extract ImageNet dataset.

ImageNet validation dataset is required to measure accuracy during inference. Visit ImageNet site and download the following files:

Note

Both dataset components must be downloaded to the same folder. This folder must be the $DATASET_DIR referenced in the following sections.

get_dataset.sh script can be used to download these files. There is no need to extract and format these files before running this sample. On the first run sample script will extract the archive with torchvision.datasets.ImageNet. Consequent runs will skip extraction.

Prerequisites

Hardware:

Software:

Run the model under container

Note

Sample requires network connection to download model from the network via HTTPS. Make sure to set https_proxy under running container if you work behind the proxy.

Pull pre-built image with the sample:

docker pull intel/image-recognition:pytorch-flex-gpu-fbnet-inference

or build it locally:

docker build \
  $(env | grep -E '(_proxy=|_PROXY)' | sed 's/^/--build-arg /') \
  -f docker/flex-gpu/pytorch-fbnet-inference/pytorch-flex-series-fbnet-inference.Dockerfile \
  -t intel/image-recognition:pytorch-flex-gpu-fbnet-inference .

Run sample as follows:

  • With dummy data:

    • Running with dummy data is recommended for performance benchmarking (throughput and latency measurements)
    • Use higher NUM_INPUTS values for more precise peak performance results. NUM_INPUTS will be rounded to a multiple of BATCH_SIZE.
    • NOTE: Accuracy will be zero when using dummy data
    mkdir -p /tmp/output && rm -f /tmp/output/* && chmod -R 777 /tmp/output
export BATCH_SIZE=1
docker run -it --rm --ipc=host \
  $(env | grep -E '(_proxy=|_PROXY)' | sed 's/^/-e /') \
  --cap-add SYS_NICE \
  --device /dev/dri/ \
  -e MODEL_NAME=fbnetc_100 \
  -e PLATFORM=Flex \
  -e MAX_TEST_DURATION=60 \
  -e MIN_TEST_DURATION=60 \
  -e NUM_INPUTS=1000 \
  -e BATCH_SIZE=${BATCH_SIZE} \
  -e OUTPUT_DIR=/tmp/output \
  -v /tmp/output:/tmp/output \
  intel/image-recognition:pytorch-flex-gpu-fbnet-inference \
    /bin/bash -c "./run_model.sh --dummy"

  • With ImageNet dataset (assumes that dataset was downloaded to the $DATASET_DIR folder):

    • Running with dataset images is recommended for accuracy measurements
    • In this mode, the test duration can be controlled by using the NUM_INPUTS parameter. The app tests a number of batches equal to max(1, NUM_INPUTS // BATCH_SIZE)
    • NOTE: Performance results (throughput and latency measurements) may be impacted due to data handling overhead
    mkdir -p /tmp/output && rm -f /tmp/output/* && chmod -R 777 /tmp/output
export BATCH_SIZE=1
docker run -it --rm --ipc=host \
  $(env | grep -E '(_proxy=|_PROXY)' | sed 's/^/-e /') \
  --cap-add SYS_NICE \
  --device /dev/dri/ \
  -e MODEL_NAME=fbnetc_100 \
  -e PLATFORM=Flex \
  -e NUM_INPUTS=50000 \
  -e BATCH_SIZE=${BATCH_SIZE} \
  -e OUTPUT_DIR=/tmp/output \
  -v /tmp/output:/tmp/output \
  -e DATASET_DIR=/dataset \
  -v $DATASET_DIR:/dataset \
  intel/image-recognition:pytorch-flex-gpu-fbnet-inference \
    /bin/bash -c "./run_model.sh"

Mind the following docker run arguments:

  • HTTPS proxy is required to download model over network (-e https_proxy=<...>)
  • --cap-add SYS_NICE is required for numactl
  • --device /dev/dri is required to expose GPU device to running container
  • --ipc=host is required for multi-stream benchmarking (./run_model.sh --dummy --streams 2) or large dataset cases
  • -v $DATASET_DIR:/dataset in case where dataset is used. $DATASET_DIR should be replaced with the actual path to the ImageNet dataset.

Run the model on baremetal

Note

Sample requires network connection to download model from the network via HTTPS. Make sure to set https_proxy before running run_model.sh if you work behind proxy.

  1. Download the sample:

    git clone https://github.com/IntelAI/models.git
cd models/models_v2/pytorch/fbnet/inference/gpu

  2. Create virtual environment venv and activate it:

    python3 -m venv venv
. ./venv/bin/activate

  3. Install sample python dependencies:

    python3 -m pip install -r requirements.txt

  4. Install Intel® Extension for PyTorch

  5. Add path to common python modules in the repo:

    export PYTHONPATH=$(pwd)/../../../../common

  6. Setup required environment variables and run the sample with ./run_model.sh:

    • With dummy data:

      • Running with dummy data is recommended for performance benchmarking (throughput and latency measurements)
      • Use higher NUM_INPUTS values for more precise peak performance results. NUM_INPUTS will be rounded to a multiple of BATCH_SIZE.
      • NOTE: Accuracy will be zero when using dummy data
      export MODEL_NAME=fbnetc_100
export PLATFORM=Flex
export BATCH_SIZE=1
export MAX_TEST_DURATION=60
export MIN_TEST_DURATION=60
export NUM_INPUTS=1000
export OUTPUT_DIR=/tmp/output
./run_model.sh --dummy

  • With ImageNet dataset (assumes that dataset was downloaded to the $DATASET_DIR folder):

    • Running with dataset images is recommended for accuracy measurements
    • In this mode, the test duration can be controlled by using the NUM_INPUTS parameter. The app tests a number of batches equal to max(1, NUM_INPUTS // BATCH_SIZE)
    • NOTE: Performance results (throughput and latency measurements) may be impacted due to data handling overhead
    export MODEL_NAME=fbnetc_100
export PLATFORM=Flex
export BATCH_SIZE=1
export NUM_INPUTS=50000
export OUTPUT_DIR=/tmp/output
export DATASET_DIR=$DATASET_DIR
./run_model.sh

Runtime arguments and environment variables

run_model.sh accepts a number of arguments to tune behavior. run_model.sh supports the use of environment variables as well as command line arguments for specifying these arguments (see the table below for details).

Before running run_model.sh script, user is required to:

  • Set OUTPUT_DIR environment variable (or use --output-dir) where script should write logs.
  • Use --dummy data or set DATASET_DIR environment variable (or use --data) pointing to ImageNet dataset.

Other arguments and/or environment variables are optional and should be used according to the actual needs (see examples above).

Argument Environment variable Valid Values Purpose
--ipex IPEX yes Use Intel® Extension for Pytorch for XPU support (default: yes)
no Use PyTorch XPU backend instead of Intel® Extension for Pytorch. Requires PyTorch version 2.4.0a or later.
--amp AMP yes Use AMP on model conversion to the desired precision (default: yes)
no
--arch MODEL_NAME fbnetc_100 HuggingFace model to run (default: fbnetc_100)
--batch-size BATCH_SIZE >=1 Batch size to use (default: 1)
--data DATASET_DIR String Location to load images from
--dummy DUMMY Use randomly generated dummy dataset in place of --data argument
--jit JIT none JIT method to use (default: trace)
compile
script
trace
--load LOAD_PATH Local path to load model from (default: disabled)
--max-test-duration MAX_TEST_DURATION >=0 Maximum duration in seconds to run benchmark. Testing will be truncated once maximum test duration has been reached. (default: disabled)
--min-test-duration MIN_TEST_DURATION >=0 Minimum duration in seconds to run benchmark. Images will be repeated until minimum test duration has been reached. (default: disabled)
--num-inputs NUM_INPUTS >=1 Number of images to load (default: 1)
--output-dir OUTPUT_DIR String Location to write output
--proxy https_proxy String System proxy
--precision PRECISION bp16 Precision to use for the model (default: fp32)
fp16
fp32
--save SAVE_PATH Local path to save model to (default: disabled)
--streams STREAMS >=1 Number of parallel streams to do inference on (default: 1)
--socket SOCKET String Socket to control telemetry capture (default: disabled)

Note

  • If --dummy is not specified (i.e. Quality Check mode), --min/max-test-duration settings are ignored. Test length is limited by minimum of num-inputs and the size of the dataset.

For more details, check help with run_model.sh --help

Example output

Script output is written to the console as well as to the output directory in the file output.log.

For multi-stream cases per-stream results can be found in the results_[STREAM_INSTANCE].json files.

Final results of the inference run can be found in results.yaml file. More verbose results summaries are in results.json file.

The yaml file contents will look like:

results:
 - key: throughput
   value: 9199.48
   unit: img/s
 - key: latency
   value: 31.394199
   unit: ms
 - key: accuracy
   value: 76.06
   unit: percents

Performance Benchmarking

benchmark.sh script can be used to benchmark FBNet performance for the predefined use cases. The benchmark.sh script is a tiny FBNet specific wrapper on top of benchmark.py script. The workflow for running a benchmark is as follows:

  • (optional) Specify path to svr-info:

    export PATH_TO_SVR_INFO=/path/to/svrinfo

  • Specify path to output benchmark results (folder must be creatable/writable under root):

    export OUTPUT_DIR=/opt/output

  • Run the benchmark script (assumes intel/image-recognition:pytorch-flex-gpu-fbnet-inference has already been pulled or built locally):

    sudo \
  PATH=$PATH_TO_SVR_INFO:$PATH \
  IMAGE=intel/image-recognition:pytorch-flex-gpu-fbnet-inference \
  OUTPUT_DIR=$OUTPUT_DIR \
  PROFILE=$(pwd)/models_v2/pytorch/fbnet/inference/gpu/profiles/c100.bf16.csv \
  PYTHONPATH=$(pwd)/models_v2/common \
  $(env | grep -E '(_proxy=|_PROXY)' | sed 's/^//') \
    $(pwd)/models_v2/pytorch/fbnet/inference/gpu/benchmark.sh

  • Final output will be written to $OUTPUT_DIR.

Note

Additonal arguments that arent specified in the benchmark profile (b0.bf16.csv in the example above) can be specified through environment variables as described in previous sections.

Usage With CUDA GPU

Scripts have a matching degree of functionality for usage on CUDA GPU's. However, this is significantly less validated and so may not work as smoothly. The primary difference for using these scripts with CUDA is building the associated docker image. We will not cover CUDA on baremetal here. In addition Intel does not provide pre-built dockers for CUDA. These must be built locally.

docker build \
  $(env | grep -E '(_proxy=|_PROXY)' | sed 's/^/--build-arg /') \
  -f docker/cuda-gpu/pytorch-fbnet-inference/pytorch-cuda-series-fbnet-inference.Dockerfile \
  -t intel/image-recognition:pytorch-cuda-gpu-fbnet-inference .

All other usage outlined in this README should be identical, with the exception of referencing this CUDA docker image in place of the for Intel GPU when running docker run as well as needing to add the --gpus all argument.

Example usage with dummy data is shown below:

mkdir -p /tmp/output && rm -f /tmp/output/* && chmod -R 777 /tmp/output
export BATCH_SIZE=1
docker run -it --rm --gpus all --ipc=host \
  $(env | grep -E '(_proxy=|_PROXY)' | sed 's/^/-e /') \
  --cap-add SYS_NICE \
  --device /dev/dri/ \
  -e MODEL_NAME=fbnetc_100 \
  -e PLATFORM=CUDA \
  -e NUM_ITERATIONS=32 \
  -e NUM_IMAGES=${BATCH_SIZE} \
  -e BATCH_SIZE=${BATCH_SIZE} \
  -e OUTPUT_DIR=/tmp/output \
  -v /tmp/output:/tmp/output \
  intel/image-recognition:pytorch-cuda-gpu-fbnet-inference \
    /bin/bash -c "./run_model.sh --dummy"