Skip to content

Latest commit

 

History

History
 
 

object_detection

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
models
models
 
 
README.md
README.md
 
 
__init__.py
__init__.py
 
 
base_run_inference.sh
base_run_inference.sh
 
 
infer.py
infer.py
 
 

README.md

TensorRT / TensorFlow Object Detection

This package provides scripts to benchmark performance and accuracy of object detection models using TensorRT integration in TensorFlow 2.0.

The input to the script is a SavedModel direcotry that includes a pre-trained model. Passing a data directory (e.g. COCO) is also necessary in case of validating accuracy (e.g. mAP).

Setup

Install object detection dependencies (from tftrt/benchmarking-python/object_detection)

git submodule update --init
../helper_scripts/install_pycocotools.sh

Usage

Run python infer.py --help to see what arguments are available.

Example:

python infer.py \
    --saved_model_dir input_saved_model \
    --data_dir /data/coco/val2017 \
    --annotation_path /data/coco/annotations/instances_val2017.json \
    --input_size 640 \
    --batch_size 8 \
    --use_tftrt \
    --precision FP16

Ready to Use Scripts:

1. faster_rcnn_resnet50_coco

# Tensorflow - FP32
./models/faster_rcnn_resnet50_coco/run_inference.sh \
    --use_xla --no_tf32 \
    --data_dir=/data/coco2017 --input_saved_model_dir=/models

# Tensorflow - TF32 (identical to FP32 on an NVIDIA Turing GPU or older)
./models/faster_rcnn_resnet50_coco/run_inference.sh \
    --use_xla \
    --data_dir=/data/coco2017 --input_saved_model_dir=/models

# TF-TRT - FP32
./models/faster_rcnn_resnet50_coco/run_inference.sh \
    --use_xla --no_tf32 \
    --data_dir=/data/coco2017 --input_saved_model_dir=/models \
    --use_tftrt --precision="FP32"    

# TF-TRT - TF32 (identical to FP32 on an NVIDIA Turing GPU or older)
./models/faster_rcnn_resnet50_coco/run_inference.sh \
    --use_xla \
    --data_dir=/data/coco2017 --input_saved_model_dir=/models \
    --use_tftrt --precision="FP32"

# TF-TRT - FP16
./models/faster_rcnn_resnet50_coco/run_inference.sh \
    --use_xla \
    --data_dir=/data/coco2017 --input_saved_model_dir=/models \
    --use_tftrt --precision="FP16"

# TF-TRT - INT8
./models/faster_rcnn_resnet50_coco/run_inference.sh \
    --use_xla \
    --data_dir=/data/coco2017 --input_saved_model_dir=/models \
    --use_tftrt --precision="INT8"

2. ssd_inception_v2_coco

# Tensorflow - FP32
./models/ssd_inception_v2_coco/run_inference.sh \
    --use_xla --no_tf32 \
    --data_dir=/data/coco2017 --input_saved_model_dir=/models

# Tensorflow - TF32 (identical to FP32 on an NVIDIA Turing GPU or older)
./models/ssd_inception_v2_coco/run_inference.sh \
    --use_xla \
    --data_dir=/data/coco2017 --input_saved_model_dir=/models

# TF-TRT - FP32
./models/ssd_inception_v2_coco/run_inference.sh \
    --use_xla --no_tf32 \
    --data_dir=/data/coco2017 --input_saved_model_dir=/models \
    --use_tftrt --precision="FP32"    

# TF-TRT - TF32 (identical to FP32 on an NVIDIA Turing GPU or older)
./models/ssd_inception_v2_coco/run_inference.sh \
    --use_xla \
    --data_dir=/data/coco2017 --input_saved_model_dir=/models \
    --use_tftrt --precision="FP32"

# TF-TRT - FP16
./models/ssd_inception_v2_coco/run_inference.sh \
    --use_xla \
    --data_dir=/data/coco2017 --input_saved_model_dir=/models \
    --use_tftrt --precision="FP16"

# TF-TRT - INT8
./models/ssd_inception_v2_coco/run_inference.sh \
    --use_xla \
    --data_dir=/data/coco2017 --input_saved_model_dir=/models \
    --use_tftrt --precision="INT8"

3. ssd_mobilenet_v1_coco

# Tensorflow - FP32
./models/ssd_mobilenet_v1_coco/run_inference.sh \
    --use_xla --no_tf32 \
    --data_dir=/data/coco2017 --input_saved_model_dir=/models

# Tensorflow - TF32 (identical to FP32 on an NVIDIA Turing GPU or older)
./models/ssd_mobilenet_v1_coco/run_inference.sh \
    --use_xla \
    --data_dir=/data/coco2017 --input_saved_model_dir=/models

# TF-TRT - FP32
./models/ssd_mobilenet_v1_coco/run_inference.sh \
    --use_xla --no_tf32 \
    --data_dir=/data/coco2017 --input_saved_model_dir=/models \
    --use_tftrt --precision="FP32"    

# TF-TRT - TF32 (identical to FP32 on an NVIDIA Turing GPU or older)
./models/ssd_mobilenet_v1_coco/run_inference.sh \
    --use_xla \
    --data_dir=/data/coco2017 --input_saved_model_dir=/models \
    --use_tftrt --precision="FP32"

# TF-TRT - FP16
./models/ssd_mobilenet_v1_coco/run_inference.sh \
    --use_xla \
    --data_dir=/data/coco2017 --input_saved_model_dir=/models \
    --use_tftrt --precision="FP16"

# TF-TRT - INT8
./models/ssd_mobilenet_v1_coco/run_inference.sh \
    --use_xla \
    --data_dir=/data/coco2017 --input_saved_model_dir=/models \
    --use_tftrt --precision="INT8"

4. ssd_mobilenet_v1_fpn_coco

# Tensorflow - FP32
./models/ssd_mobilenet_v1_fpn_coco/run_inference.sh \
    --use_xla --no_tf32 \
    --data_dir=/data/coco2017 --input_saved_model_dir=/models

# Tensorflow - TF32 (identical to FP32 on an NVIDIA Turing GPU or older)
./models/ssd_mobilenet_v1_fpn_coco/run_inference.sh \
    --use_xla \
    --data_dir=/data/coco2017 --input_saved_model_dir=/models

# TF-TRT - FP32
./models/ssd_mobilenet_v1_fpn_coco/run_inference.sh \
    --use_xla --no_tf32 \
    --data_dir=/data/coco2017 --input_saved_model_dir=/models \
    --use_tftrt --precision="FP32"    

# TF-TRT - TF32 (identical to FP32 on an NVIDIA Turing GPU or older)
./models/ssd_mobilenet_v1_fpn_coco/run_inference.sh \
    --use_xla \
    --data_dir=/data/coco2017 --input_saved_model_dir=/models \
    --use_tftrt --precision="FP32"

# TF-TRT - FP16
./models/ssd_mobilenet_v1_fpn_coco/run_inference.sh \
    --use_xla \
    --data_dir=/data/coco2017 --input_saved_model_dir=/models \
    --use_tftrt --precision="FP16"

# TF-TRT - INT8
./models/ssd_mobilenet_v1_fpn_coco/run_inference.sh \
    --use_xla \
    --data_dir=/data/coco2017 --input_saved_model_dir=/models \
    --use_tftrt --precision="INT8"

5. ssd_mobilenet_v2_coco

# Tensorflow - FP32
./models/ssd_mobilenet_v2_coco/run_inference.sh \
    --use_xla --no_tf32 \
    --data_dir=/data/coco2017 --input_saved_model_dir=/models

# Tensorflow - TF32 (identical to FP32 on an NVIDIA Turing GPU or older)
./models/ssd_mobilenet_v2_coco/run_inference.sh \
    --use_xla \
    --data_dir=/data/coco2017 --input_saved_model_dir=/models

# TF-TRT - FP32
./models/ssd_mobilenet_v2_coco/run_inference.sh \
    --use_xla --no_tf32 \
    --data_dir=/data/coco2017 --input_saved_model_dir=/models \
    --use_tftrt --precision="FP32"    

# TF-TRT - TF32 (identical to FP32 on an NVIDIA Turing GPU or older)
./models/ssd_mobilenet_v2_coco/run_inference.sh \
    --use_xla \
    --data_dir=/data/coco2017 --input_saved_model_dir=/models \
    --use_tftrt --precision="FP32"

# TF-TRT - FP16
./models/ssd_mobilenet_v2_coco/run_inference.sh \
    --use_xla \
    --data_dir=/data/coco2017 --input_saved_model_dir=/models \
    --use_tftrt --precision="FP16"

# TF-TRT - INT8
./models/ssd_mobilenet_v2_coco/run_inference.sh \
    --use_xla \
    --data_dir=/data/coco2017 --input_saved_model_dir=/models \
    --use_tftrt --precision="INT8"

6. ssd_resnet_50_fpn_coco

# Tensorflow - FP32
./models/ssd_resnet_50_fpn_coco/run_inference.sh \
    --use_xla --no_tf32 \
    --data_dir=/data/coco2017 --input_saved_model_dir=/models

# Tensorflow - TF32 (identical to FP32 on an NVIDIA Turing GPU or older)
./models/ssd_resnet_50_fpn_coco/run_inference.sh \
    --use_xla \
    --data_dir=/data/coco2017 --input_saved_model_dir=/models

# TF-TRT - FP32
./models/ssd_resnet_50_fpn_coco/run_inference.sh \
    --use_xla --no_tf32 \
    --data_dir=/data/coco2017 --input_saved_model_dir=/models \
    --use_tftrt --precision="FP32"    

# TF-TRT - TF32 (identical to FP32 on an NVIDIA Turing GPU or older)
./models/ssd_resnet_50_fpn_coco/run_inference.sh \
    --use_xla \
    --data_dir=/data/coco2017 --input_saved_model_dir=/models \
    --use_tftrt --precision="FP32"

# TF-TRT - FP16
./models/ssd_resnet_50_fpn_coco/run_inference.sh \
    --use_xla \
    --data_dir=/data/coco2017 --input_saved_model_dir=/models \
    --use_tftrt --precision="FP16"

# TF-TRT - INT8
./models/ssd_resnet_50_fpn_coco/run_inference.sh \
    --use_xla \
    --data_dir=/data/coco2017 --input_saved_model_dir=/models \
    --use_tftrt --precision="INT8"

7. ssdlite_mobilenet_v2_coco

# Tensorflow - FP32
./models/ssdlite_mobilenet_v2_coco/run_inference.sh \
    --use_xla --no_tf32 \
    --data_dir=/data/coco2017 --input_saved_model_dir=/models

# Tensorflow - TF32 (identical to FP32 on an NVIDIA Turing GPU or older)
./models/ssdlite_mobilenet_v2_coco/run_inference.sh \
    --use_xla \
    --data_dir=/data/coco2017 --input_saved_model_dir=/models

# TF-TRT - FP32
./models/ssdlite_mobilenet_v2_coco/run_inference.sh \
    --use_xla --no_tf32 \
    --data_dir=/data/coco2017 --input_saved_model_dir=/models \
    --use_tftrt --precision="FP32"    

# TF-TRT - TF32 (identical to FP32 on an NVIDIA Turing GPU or older)
./models/ssdlite_mobilenet_v2_coco/run_inference.sh \
    --use_xla \
    --data_dir=/data/coco2017 --input_saved_model_dir=/models \
    --use_tftrt --precision="FP32"

# TF-TRT - FP16
./models/ssdlite_mobilenet_v2_coco/run_inference.sh \
    --use_xla \
    --data_dir=/data/coco2017 --input_saved_model_dir=/models \
    --use_tftrt --precision="FP16"

# TF-TRT - INT8
./models/ssdlite_mobilenet_v2_coco/run_inference.sh \
    --use_xla \
    --data_dir=/data/coco2017 --input_saved_model_dir=/models \
    --use_tftrt --precision="INT8"