Skip to content

Latest commit

 

History

History
93 lines (57 loc) · 4.41 KB

README.md

File metadata and controls

93 lines (57 loc) · 4.41 KB
Raw

ResNet50 Benchmark

This benchmark performs image classifications using the ResNet-50 network and the ImageNet dataset.

⚠️ IMPORTANT: Please use closed/Dell as the working directory when running the below commands. ⚠️

Dataset

Downloading / obtaining the dataset

The dataset used for this benchmark is ImageNet 2012 validation set. Please manually download the dataset and unzip the images to $MLPERF_SCRATCH_PATH/data/imagenet/. You can run bash code/resnet50/tensorrt/download_data.sh to verify if the images are in the expected locations.

Preprocessing the dataset for usage

To process the input images to INT8 NCHW format, please run python3 code/resnet50/tensorrt/preprocess_data.py. The preprocessed data will be saved to $MLPERF_SCRATCH_PATH/preprocessed_data/imagenet/ResNet50.

Model

Downloading / obtaining the model

The ONNX model resnet50_v1.onnx is downloaded from the zenodo link provided by the MLPerf inference repository.

This can also be downloaded by running bash code/resnet50/tensorrt/download_model.sh.

Optimizations

Plugins

The following TensorRT plugins are used to optimize ResNet50 benchmark:

  • RES2_FULL_FUSION: fuses all the res2* layers into one CUDA kernel These plugins are available in TensorRT starting with the TensorRT 7.2 release.

Lower Precision

To further optimize performance, with minimal impact on classification accuracy, we run the computations in INT8 precision.

Removal of Softmax

Softmax layer is removed since it does not affect the predicted label.

Calibration

ResNet50 INT8 is calibrated on a subset of the ImageNet validation set. The indices of this subset can be found at data_maps/imagenet/cal_map.txt. We use TensorRT symmetric calibration, and store the scaling factors in code/resnet50/tensorrt/calibrator.cache.

Instructions for Audits

Run Inference through LoadGen

Run the following commands from within the container to run inference through LoadGen:

make run RUN_ARGS="--benchmarks=resnet50 --scenarios=<SCENARIO> --test_mode=PerformanceOnly"
make run RUN_ARGS="--benchmarks=resnet50 --scenarios=<SCENARIO> --test_mode=AccuracyOnly"

To run inference through Triton Inference Server and LoadGen:

make run RUN_ARGS="--benchmarks=resnet50 --scenarios=<SCENARIO> --config_ver=triton --test_mode=PerformanceOnly"
make run RUN_ARGS="--benchmarks=resnet50 --scenarios=<SCENARIO> --config_ver=triton --test_mode=AccuracyOnly"

The performance and the accuracy results will be printed to stdout, and the LoadGen logs can be found in build/logs.

Run with New Weights

Follow these steps to run inference with new weights:

  1. If the new weights are in TensorFlow frozen graph format, please use resnet50-to-onnx.sh in the official MLPerf repository to convert it to ONNX format.
  2. Replace build/models/ResNet50/resnet50_v1.onnx with new ONNX model.
  3. Run make calibrate RUN_ARGS="--benchmarks=resnet50" to generate a new calibration cache.
  4. Run inference by make run RUN_ARGS="--benchmarks=resnet50 --scenarios=<SCENARIO>".

Run with New Validation Dataset

Follow these steps to run inference with new validation dataset:

  1. Put the validation dataset under build/data/imagenet.
  2. Modify data_maps/imagenet/val_map.txt to contain all the file names and the corresponding labels of the new validation dataset.
  3. Preprocess data by python3 code/resnet50/tensorrt/preprocess_data.py --val_only.
  4. Run inference by make run RUN_ARGS="--benchmarks=resnet50 --scenarios=<SCENARIO>".

Run with New Calibration Dataset

Follow these steps to generate a new calibration cache with new calibration dataset:

  1. Put the calibration dataset under build/data/imagenet.
  2. Modify data_maps/imagenet/cal_map.txt to contain all the file names of the new calibration dataset.
  3. Preprocess data by python3 code/resnet50/tensorrt/preprocess_data.py --cal_only.
  4. Run make calibrate RUN_ARGS="--benchmarks=resnet50" to generate a new calibration cache.