README.md

RFCN FP32 inference

This document has instructions for running RFCN FP32 inference using Intel-optimized TensorFlow.

Dataset

The COCO validation dataset is used in these RFCN quickstart scripts. The inference quickstart scripts use raw images, and the accuracy quickstart scripts require the dataset to be converted into the TF records format. See the COCO dataset for instructions on downloading and preprocessing the COCO validation dataset.

Quick Start Scripts

Script name	Description
fp32_inference.sh	Runs inference on a directory of raw images for 500 steps and outputs performance metrics.
fp32_accuracy.sh	Processes the TF records to run inference and check accuracy on the results.

Run the model

Setup your environment using the instructions below, depending on if you are using AI Kit:

Setup using AI Kit on Linux	Setup without AI Kit on Linux	Setup without AI Kit on Windows
To run using AI Kit on Linux you will need: git numactl wget build-essential Cython contextlib2 jupyter lxml matplotlib pillow>=8.1.2 protobuf-compiler pycocotools Activate the `tensorflow` conda environment conda activate tensorflow	To run without AI Kit on Linux you will need: Python 3 git numactl wget intel-tensorflow>=2.5.0 build-essential Cython contextlib2 jupyter lxml matplotlib pillow>=8.1.2 protobuf-compiler pycocotools A clone of the Model Zoo repo git clone https://github.com/IntelAI/models.git	To run without AI Kit on Windows you will need: Intel Model Zoo on Windows Systems prerequisites build-essential Cython contextlib2 jupyter lxml matplotlib pillow>=8.1.2 protobuf-compiler pycocotools A clone of the Model Zoo repo git clone https://github.com/IntelAI/models.git

For more information on the required dependencies, see the documentation on prerequisites in the TensorFlow models repo.

Download and extract the pretrained model and set the PRETRAINED_MODEL environment variable to point to the frozen graph file. If you run on Windows, please use a browser to download the pretrained model using the link below. For Linux, run:

wget https://storage.googleapis.com/intel-optimized-tensorflow/models/v1_8/rfcn_resnet101_fp32_coco_pretrained_model.tar.gz
tar -xzvf rfcn_resnet101_fp32_coco_pretrained_model.tar.gz
export PRETRAINED_MODEL=$(pwd)/rfcn_resnet101_coco_2018_01_28/frozen_inference_graph.pb

RFCN uses the object detection code from the TensorFlow Model Garden. Clone this repo with the SHA specified below and apply the patch from the model zoo directory. Set the TF_MODELS_DIR environment variable to the path of your clone of the TF Model Garden.

# Clone the TensorFlow Model Garden
git clone https://github.com/tensorflow/models.git tensorflow-models
cd tensorflow-models
git checkout 6c21084503b27a9ab118e1db25f79957d5ef540b

# Apply the TF2 patch from the model zoo repo directory
git apply --ignore-space-change --ignore-whitespace <model zoo directory>/models/object_detection/tensorflow/rfcn/inference/tf-2.0.patch

# Set the TF_MODELS_DIR env var
export TF_MODELS_DIR=$(pwd)

Run on Linux

Download and install Google Protobuf version 3.3.0, and run protobuf compilation.

cd TF_MODELS_DIR/research
protoc object_detection/protos/*.proto --python_out=.
cd ../..

Once your environment is setup, navigate back to your Model Zoo directory. Ensure that you have set environment variables pointing to the TensorFlow Model Garden repo, the dataset, and output directories, and then run a quickstart script.

To run inference with performance metrics:

# cd to your model zoo directory
cd models

export DATASET_DIR=<path to the coco val2017 raw image directory (ex: /home/user/coco_dataset/val2017)>
export OUTPUT_DIR=<directory where log files will be written>
export TF_MODELS_DIR=<directory where TensorFlow Model Garden is cloned>

./quickstart/object_detection/tensorflow/rfcn/inference/cpu/fp32/fp32_inference.sh

To get accuracy metrics:

# cd to your model zoo directory
cd models

export DATASET_DIR=<path to TF record file (ex: /home/user/coco_output/coco_val.record)>
export OUTPUT_DIR=<directory where log files will be written>
export TF_MODELS_DIR=<directory where TensorFlow Model Garden is cloned>

./quickstart/object_detection/tensorflow/rfcn/inference/cpu/fp32/fp32_accuracy.sh

Run on Windows

• Download and install Google Protobuf version 3.4.0 for Windows in addition to the above listed dependencies. Download and extract protoc-3.4.0-win32.zip Navigate to the research directory in TF_MODELS_DIR and install Google Protobuf:

set TF_MODELS_DIR=<directory where TensorFlow Model Garden is cloned>
cd %TF_MODELS_DIR%\research
“C:\<user>\protoc-3.4.0-win32\bin\protoc.exe” object_detection/protos/*.proto --python_out=.

After installing the prerequisites and cloning the TensorFlow models repo, and downloading the pretrained model, set the environment variables for the paths to your PRETRAINED_MODEL, an OUTPUT_DIR where log files will be written, TF_MODELS_DIR, and DATASET_DIR for COCO raw dataset directory or tf_records file based on whether you run inference or accuracy scripts. Navigate to your model zoo directory and then run a quickstart script.

# cd to your model zoo directory
cd models

set PRETRAINED_MODEL=<path to the frozen graph downloaded above>
set DATASET_DIR=<path to COCO raw dataset directory or tf_records file based on whether you run inference or accuracy scripts>
set OUTPUT_DIR=<directory where log files will be written>
set TF_MODELS_DIR=<directory where TensorFlow Model Garden is cloned>

bash quickstart\object_detection\tensorflow\rfcn\inference\cpu\fp32\<script name>.sh

Note: You may use cygpath to convert the Windows paths to Unix paths before setting the environment variables. As an example, if the dataset location on Windows is D:\user\coco_dataset\val2017, convert the Windows path to Unix as shown:

cygpath D:\user\coco_dataset\val2017
/d/user/coco_dataset/val2017

Then, set the DATASET_DIR environment variable set DATASET_DIR=/d/user/coco_dataset/val2017.

Additional Resources

• To run more advanced use cases, see the instructions here for calling the launch_benchmark.py script directly.
• To run the model using docker, please see the oneContainer workload container:
  https://software.intel.com/content/www/us/en/develop/articles/containers/rfcn-fp32-inference-tensorflow-container.html.