Skip to content

Latest commit

 

History

History
108 lines (91 loc) · 5.05 KB

README.md

File metadata and controls

108 lines (91 loc) · 5.05 KB
Raw

Swin-transformer Inference

Swin-transformer Inference best known configurations with Intel® Extension for PyTorch.

Model Information

Use Case Framework Model Repo Branch/Commit/Tag Optional Patch
Inference PyTorch https://github.com/microsoft/Swin-Transformer main/afeb877fba1139dfbc186276983af2abb02c2196 -

Pre-Requisite

Prepare Dataset

Dataset: imagenet

ImageNet is recommended, the download link is https://image-net.org/challenges/LSVRC/2012/2012-downloads.php.

We recommend downloading ILSVRC2012_img_train_t3.tar. To create the training director you can extract like so:

mkdir imagenet
cp ILSVRC2012_img_train_t3.tar imagenet/
cd imagenet
tar -xvf ILSVRC2012_img_train_t3.tar --one-top-level=images
find images/ -name '*.tar' -execdir tar -xvf '{}' --one-top-level \;

You will need to reorganize the files in this package into training, testing and validation directories. The easiest way to do that is with the split-packages python tool.

pip install split-folders
rm images/*.tar
split-folders --ratio .8 .1 .1 --output dataset images

The imagenet/dataset directory is your dataset directory.

Dataset: dummy

Alternatively, you can edit run_models.sh to include the parameter --dummy when it invokes main_no_ddp.py. This flag will use randomly generated data instead. This will not create any actual useful machine learning, but allows you to get up and running quickly and easily without any datasets downloaded and run split-packages or anything like that.

Inference

  1. git clone https://github.com/IntelAI/models.git
  2. cd models/models_v2/pytorch/swin-transformer/inference/gpu
  3. Create virtual environment venv and activate it: 
    python3 -m venv venv
. ./venv/bin/activate
  4. Run setup.sh 
    source ./setup.sh
  5. Install the latest GPU versions of torch, torchvision and intel_extension_for_pytorch: 
    python -m pip install torch==<torch_version> torchvision==<torchvision_version> intel-extension-for-pytorch==<ipex_version> --extra-index-url https://pytorch-extension.intel.com/release-whl-aitools/
  6. Set environment variables for Intel® oneAPI Base Toolkit: Default installation location {ONEAPI_ROOT} is /opt/intel/oneapi for root account, ${HOME}/intel/oneapi for other accounts 
    source {ONEAPI_ROOT}/compiler/latest/env/vars.sh
source {ONEAPI_ROOT}/mkl/latest/env/vars.sh
source {ONEAPI_ROOT}/tbb/latest/env/vars.sh
source {ONEAPI_ROOT}/mpi/latest/env/vars.sh
source {ONEAPI_ROOT}/ccl/latest/env/vars.sh
  7. Setup required environment paramaters
Parameter export command
MULTI_TILE export MULTI_TILE=False
PLATFORM export PLATFORM=Flex (Flex)
DATASET_DIR export DATASET_DIR=
OUTPUT_DIR export OUTPUT_DIR=$PWD
BATCH_SIZE (optional) export BATCH_SIZE=512
NUM_ITERATIONS (optional) export NUM_ITERATIONS=500
  1. Run run_model.sh

Output

Single-tile output will typically looks like:

[2023-12-28 07:17:30 swin_base_patch4_window7_224](main_no_ddp.py 383): INFO Latency: 1.199571
[2023-12-28 07:17:30 swin_base_patch4_window7_224](main_no_ddp.py 384): INFO Throughput: 426.819398

Final results of the inference run can be found in results.yaml file.

results:
 - key: throughput
   value: 426.8194
   unit: fps
 - key: latency
   value: 1.1995705968359012
   unit: s
 - key: accuracy
   value: 7.535
   unit: loss