A repository that shares tuning results of trained models generated by TensorFlow / Keras. Post-training quantization (Weight Quantization, Integer Quantization, Full Integer Quantization, Float16 Quantization), Quantization-aware training. TensorFlow Lite. OpenVINO. CoreML. TensorFlow.js. TF-TRT. MediaPipe. ONNX. [.tflite,.h5,.pb,saved_model,tfjs,tftrt,mlmodel,.xml/.bin, .onnx]

Overview

PINTO_model_zoo

Please read the contents of the LICENSE file located directly under each folder before using the model. My model conversion scripts are released under the MIT license, but the license of the source model itself is subject to the license of the provider repository.

A repository that shares tuning results of trained models generated by Tensorflow. Post-training quantization (Weight Quantization, Integer Quantization, Full Integer Quantization, Float16 Quantization), Quantization-aware training. I also try to convert it to OpenVINO's IR model as much as possible.

TensorFlow Lite, OpenVINO, CoreML, TensorFlow.js, TF-TRT, MediaPipe, ONNX [.tflite, .h5, .pb, saved_model, tfjs, tftrt, mlmodel, .xml/.bin, .onnx]

I have been working on quantization of various models as a hobby, but I have skipped the work of making sample code to check the operation because it takes a lot of time. I welcome a pull request from volunteers to provide sample code. 😄

[Note Jan 05, 2020] Currently, the MobileNetV3 backbone model and the Full Integer Quantization model do not return correctly.

[Note Jan 08, 2020] If you want the best performance with RaspberryPi4/3, install Ubuntu 19.10 aarch64 (64bit) instead of Raspbian armv7l (32bit). The official Tensorflow Lite is performance tuned for aarch64. On aarch64 OS, performance is about 4 times higher than on armv7l OS.

[Note Jun 22, 2020] I'm working on an issue where the final output of EfficientDet seems to have fewer detections. same issue

My article

List of pre-quantized models

* WQ = Weight Quantization ** OV = OpenVINO IR *** CM = CoreML **** DQ = Dynamic Range Quantization

1. Image Classification

No. Model Name Link FP32 FP16 INT8 TPU WQ OV CM TFJS TF-TRT ONNX Remarks
004 Efficientnet ■■■
010 Mobilenetv3 ■■■
011 Mobilenetv2 ■■■
016 Efficientnet-lite ■■■
070 age-gender-recognition ■■■
083 Person_Reidentification ■■■ 248,277,286,287,288,300
087 DeepSort ■■■
124 person-attributes-recognition-crossroad-0230 ■■■
125 person-attributes-recognition-crossroad-0234 ■■■
126 person-attributes-recognition-crossroad-0238 ■■■

2. 2D Object Detection

No. Model Name Link FP32 FP16 INT8 TPU WQ OV CM TFJS TF-TRT ONNX Remarks
002 Mobilenetv3-SSD ■■■
006 Mobilenetv2-SSDlite ■■■
008 Mask_RCNN_Inceptionv2 ■■■
018 EfficientDet ■■■
023 Yolov3-nano ■■■
024 Yolov3-lite ■■■
031 Yolov4 ■■■
034 SSD_Mobilenetv2_mnasfpn ■■■
038 SSDlite_MobileDet_edgetpu ■■■
039 SSDlite_MobileDet_cpu ■■■
042 Centernet ■■■
045 SSD_Mobilenetv2_oid_v4 ■■■
046 Yolov4-tiny ■■■
047 SpineNetMB_49 ■■■ Mobile RetinaNet
051 East_Text_Detection ■■■
054 KNIFT ■■■ MediaPipe
056 TextBoxes++ with dense blocks, separable convolution and Focal Loss ■■■
058 keras-retinanet ■■■ resnet50_coco_best_v2.1.0.h5,320x320
059 Yolov5 ■■■
072 NanoDet ■■■ issue #274
073 RetinaNet ■■■
074 Yolact ■■■
085 Yolact_Edge ■■■ WIP, MobileNetV2(256/320)
089 DETR ■■■ 256x256
103 EfficientDet_lite ■■■ lite0,lite1,lite2,lite3,lite4
116 DroNet ■■■ DroNet,DroNetV3
123 YOLOR ■■■ ssss_s2d/320x320,640x640,960x960,1280x1280
132 YOLOX ■■■ nano/tiny,320x320,416x416,480x640,544x960,736x1280,1088x1920
143 RAPiD ■■■ Fisheye, cepdof/habbof/mw_r, 608x608/1024x1024
145 text_detection_db ■■■ 480x640
151 object_detection_mobile_object_localizer ■■■ 192x192

3. 3D Object Detection

No. Model Name Link FP32 FP16 INT8 TPU WQ OV CM TFJS TF-TRT ONNX Remarks
036 Objectron ■■■ MediaPipe
063 3D BoundingBox estimation for autonomous driving ■■■ YouTube
107 SFA3D ■■■

4. 2D/3D Face Detection

No. Model Name Link FP32 FP16 INT8 TPU WQ OV CM TFJS TF-TRT ONNX Remarks
025 Head_Pose_Estimation ■■■
030 BlazeFace ■■■ MediaPipe
032 FaceMesh ■■■ MediaPipe
040 DSFD_vgg ■■■
041 DBFace ■■■ MobileNetV2/V3, 320x320,480x640,640x960,800x1280
043 Face_Landmark ■■■
049 Iris_Landmark ■■■ MediaPipe
095 CenterFace ■■■
096 RetinaFace ■■■
106 WHENet ■■■ Real-time Fine-Grained Estimation for Wide Range Head Pose
129 SCRFD ■■■ All types
130 YOLOv5_Face ■■■ yolov5n_0.5,yolov5n_face,yolov5s_face/256x320,480x640,736x1280
134 head-pose-estimation-adas-0001 ■■■
144 YuNet ■■■ 120x160

5. 2D/3D Hand Detection

No. Model Name Link FP32 FP16 INT8 TPU WQ OV CM TFJS TF-TRT ONNX Remarks
027 Minimal-Hand ■■■
033 Hand_Detection_and_Tracking ■■■ MediaPipe
094 hand_recrop ■■■ MediaPipe

6. 2D/3D Human Pose Estimation

No. Model Name Link FP32 FP16 INT8 TPU DQ WQ OV CM TFJS TF-TRT ONNX Remarks
003 Posenet ■■■
007 Mobilenetv2_Pose_Estimation ■■■
029 Human_Pose_Estimation_3D ■■■
053 BlazePose ■■■ MediaPipe,Integrate 058_BlazePose_Full_Keypoints
065 ThreeDPoseUnityBarracuda ■■■ YouTube
080 tf_pose_estimation ■■■
084 EfficientPose ■■■ SinglePose
088 Mobilenetv3_Pose_Estimation ■■■
115 MoveNet ■■■ lightning,thunder
137 MoveNet_MultiPose ■■■ lightning,192x192,192x256,256x256,256x320,320x320,480x640,720x1280,1280x1920
156 MobileHumanPose ■■■ 3D
157 3DMPPE_POSENET ■■■ 3D,192x192/256x256/320x320/416x416/480x640/512x512

7. Depth Estimation from Monocular/Stereo Images

No. Model Name Link FP32 FP16 INT8 TPU DQ WQ OV CM TFJS TF-TRT ONNX Remarks
009 Multi-Scale Local Planar Guidance for Monocular Depth Estimation ■■■
014 tf-monodepth2 ■■■
028 struct2depth ■■■
064 Dense Depth ■■■
066 Footprints ■■■
067 MiDaS ■■■
081 MiDaS v2 ■■■
135 CoEx ■■■ WIP, onnx/OpenVINO only
142 HITNET ■■■ WIP issue,flyingthings_finalpass_xl/eth3d/middlebury_d400,120x160/240x320/256x256/480x640/720x1280
146 FastDepth ■■■ 128x160,224x224,256x256,256x320,320x320,480x640,512x512,768x1280
147 PackNet-SfM ■■■ ddad/kitti,Convert all ResNet18 backbones only
148 LapDepth ■■■ kitti/nyu,192x320/256x320/368x640/480x640/720x1280
149 depth_estimation ■■■ nyu,180x320/240x320/360x640/480x640/720x1280
150 MobileStereoNet ■■■ WIP. Conversion script only.
153 MegaDepth ■■■ 192x256,384x512
158 HR-Depth ■■■

8. Semantic Segmentation

No. Model Name Link FP32 FP16 INT8 TPU WQ OV CM TFJS TF-TRT ONNX Remarks
001 deeplabv3 ■■■
015 Faster-Grad-CAM ■■■
020 EdgeTPU-Deeplab ■■■
021 EdgeTPU-Deeplab-slim ■■■
026 Mobile-Deeplabv3-plus ■■■
035 BodyPix ■■■ MediaPipe,MobileNet0.50/0.75/1.00,ResNet50
057 BiSeNetV2 ■■■
060 Hair Segmentation ■■■ WIP,MediaPipe
061 U^2-Net ■■■
069 ENet ■■■ Cityscapes,512x1024
075 ERFNet ■■■ Cityscapes,256x512,384x786,512x1024
078 MODNet ■■■ 128x128,192x192,256x256,512x512
082 MediaPipe_Meet_Segmentation ■■■ MediaPipe,128x128,144x256,96x160
104 DeeplabV3-plus ■■■ cityscapes,200x400,400x800,800x1600
109 Selfie_Segmentation ■■■ 256x256
136 road-segmentation-adas-0001 ■■■
138 BackgroundMattingV2 ■■■ 720x1280,2160x4096

9. Anomaly Detection

No. Model Name Link FP32 FP16 INT8 TPU WQ OV CM TFJS TF-TRT ONNX Remarks
005 One_Class_Anomaly_Detection ■■■
099 Efficientnet_Anomaly_Detection_Segmentation ■■■

10. Artistic

No. Model Name Link FP32 FP16 INT8 TPU WQ OV CM TFJS TF-TRT ONNX Remarks
017 Artistic-Style-Transfer ■■■
019 White-box-Cartoonization ■■■
037 First_Neural_Style_Transfer ■■■
044 Selfie2Anime ■■■
050 AnimeGANv2 ■■■
062 Facial Cartoonization ■■■
068 Colorful_Image_Colorization ■■■ experimental
101 arbitrary_image_stylization ■■■ magenta
113 Anime2Sketch ■■■

11. Super Resolution

No. Model Name Link FP32 FP16 INT8 TPU WQ OV CM TFJS TF-TRT ONNX Remarks
012 Fast_Accurate_and_Lightweight_Super-Resolution ■■■
022 Learning_to_See_Moving_Objects_in_the_Dark ■■■
071 Noise2Noise ■■■ srresnet/clear only
076 Deep_White_Balance ■■■
077 ESRGAN ■■■ 50x50->x4, 100x100->x4
079 MIRNet ■■■
086 Defocus Deblurring Using Dual-Pixel ■■■
090 Ghost-free_Shadow_Removal ■■■ 256x256
111 SRN-Deblur ■■■ 240x320,480x640,720x1280,1024x1280
112 DeblurGANv2 ■■■ inception/mobilenetv2:256x256,320x320,480x640,736x1280,1024x1280
114 Two-branch-dehazing ■■■ 240x320,480x640,720x1280
133 Real-ESRGAN ■■■ 16x16,32x32,64x64,128x128,240x320,256x256,320x320,480x640
152 DeepLPF ■■■

12. Sound Classifier

No. Model Name Link FP32 FP16 INT8 TPU WQ OV CM TFJS TF-TRT ONNX Remarks
013 ml-sound-classifier ■■■
097 YAMNet ■■■
098 SPICE ■■■
118 Speech-enhancement ■■■ WIP,EdgeTPU(LeakyLeRU)
120 FRILL ■■■ nofrontend

13. Natural Language Processing

No. Model Name Link FP32 FP16 INT8 TPU WQ OV CM TFJS TF-TRT ONNX Remarks
048 Mobile_BERT ■■■
121 GPT2/DistillGPT2 ■■■
122 DistillBert ■■■

14. Text Recognition

No. Model Name Link FP32 FP16 INT8 TPU WQ OV CM TFJS TF-TRT ONNX Remarks
052 Handwritten_Text_Recognition ■■■
055 Handwritten_Japanese_Recognition ■■■
093 ocr_japanese ■■■ 120x160

15. Action Recognition

No. Model Name Link FP32 FP16 INT8 TPU WQ OV CM TFJS TF-TRT ONNX Remarks
092 weld-porosity-detection-0001 ■■■

16. Inpainting

No. Model Name Link FP32 FP16 INT8 TPU WQ OV CM TFJS TF-TRT ONNX Remarks
100 HiFill ■■■

17. GAN

No. Model Name Link FP32 FP16 INT8 TPU WQ OV CM TFJS TF-TRT ONNX Remarks
105 MobileStyleGAN ■■■

18. Transformer

No. Model Name Link FP32 FP16 INT8 TPU WQ OV CM TFJS TF-TRT ONNX Remarks
127 dino ■■■ experimental,dino_deits8/dino_deits16

19. Others

No. Model Name Link FP32 FP16 INT8 TPU DQ WQ OV CM TFJS TF-TRT ONNX Remarks
091 gaze-estimation-adas-0002 ■■■
102 Coconet ■■■ magenta
108 HAWP ■■■ WIP
110 L-CNN ■■■ WIP
117 DTLN ■■■
119 M-LSD ■■■
131 CFNet ■■■ 256x256,512x768
139 PSD-Principled-Synthetic-to-Real-Dehazing-Guided-by-Physical-Priors ■■■
140 Ultra-Fast-Lane-Detection ■■■ 288x800
141 lanenet-lane-detection ■■■ 256x512
154 driver-action-recognition-adas-0002-encoder ■■■
155 driver-action-recognition-adas-0002-decoder ■■■

Sample.1 - Object detection by video file

$ cd 006_mobilenetv2-ssdlite/02_voc/03_integer_quantization
$ ./download.sh && cd ..
$ python3 mobilenetv2ssdlite_movie_sync.py

004

Sample.2 - Object detection by USB Camera

$ cd 006_mobilenetv2-ssdlite/02_voc/03_integer_quantization
$ ./download.sh && cd ..
$ python3 mobilenetv2ssdlite_usbcam_sync.py

005

Sample.3 - Head Pose Estimation, Multi-stage inference with multi-model

  • RaspberryPi4 (CPU only)
  • Ubuntu 19.10 64bit
  • Tensorflow / Tensorflow Lite with multi-thread acceleration tuning for PythonAPI
  • [Model.1] MobileNetV2-SSDLite dm=0.5 300x300, Integer Quantization
  • [Model.2] Head Pose Estimation 128x128, Integer Quantization
  • WIDERFACE
  • USB Camera, 640x480
  • IPS 1080p HDMI Display
  • Approximately 13FPS for all processes from pre-processing, inference, post-processing, and display
$ cd 025_head_pose_estimation/03_integer_quantization
$ ./download.sh
$ python3 head_pose_estimation.py

006

Sample.4 - Semantic Segmentation, DeeplabV3-plus 256x256

  • RaspberryPi4 (CPU only)
  • Ubuntu 19.10 64bit
  • Tensorflow / Tensorflow Lite with multi-thread acceleration tuning for PythonAPI
  • DeeplabV3-plus (MobileNetV2) Decoder 256x256, Integer Quantization
  • USB Camera, 640x480
  • IPS 1080p HDMI Display
  • Approximately 8.5 FPS for all processes from pre-processing, inference, post-processing, and display
$ cd 026_mobile-deeplabv3-plus/03_integer_quantization
$ ./download.sh
$ python3 deeplabv3plus_usbcam.py

007

Sample.5 - MediaPipe/FaceMesh, face_detection_front_128_weight_quant, face_landmark_192_weight_quant

Sample.6 - MediaPipe/Objectron, object_detection_3d_chair_640x480_weight_quant

Sample.7 - MediaPipe/Objectron, object_detection_3d_chair_640x480_openvino_FP32

Sample.8 - MediaPipe/BlazeFace, face_detection_front_128_integer_quant

Sample.9 - MediaPipe/Hand_Detection_and_Tracking(3D Hand Pose), hand_landmark_3d_256_integer_quant.tflite + palm_detection_builtin_256_integer_quant.tflite

Sample.10 - DBFace, 640x480_openvino_FP32

Sample.11 - Human_Pose_Estimation_3D, 640x480, Tensorflow.js + WebGL + Browser

Sample.12 - BlazePose Full Body, 640x480, Tensorflow.js + WebGL + Browser

Sample.13 - Facial Cartoonization, 640x480, OpenVINO Corei7 CPU only

  • Ubuntu 18.04 x86_64
  • OpenVINO
  • USB Camera, 640x480
  • Test Code 015

1. Environment

  • Ubuntu 18.04 x86_64
  • RaspberryPi4 Raspbian Buster 32bit / Raspbian Buster 64bit / Ubuntu 19.10 aarch64
  • Tensorflow-GPU v1.15.2 or Tensorflow v2.3.1+
  • OpenVINO 2020.2+
  • PyTorch 1.6.0+
  • ONNX Opset12
  • Python 3.6.8
  • PascalVOC Dataset
  • COCO Dataset
  • Cityscapes Dataset
  • Imagenette Dataset
  • CelebA Dataset
  • Audio file (.wav)
  • WIDERFACE
  • Google Colaboratory

2. Procedure

Procedure examples

2-1. MobileNetV3+DeeplabV3+PascalVOC

2-1-1. Preparation

$ cd ~
$ mkdir deeplab;cd deeplab
$ git clone --depth 1 https://github.com/tensorflow/models.git
$ cd models/research/deeplab/datasets
$ mkdir pascal_voc_seg

$ curl -sc /tmp/cookie \
  "https://drive.google.com/uc?export=download&id=1rATNHizJdVHnaJtt-hW9MOgjxoaajzdh" > /dev/null
$ CODE="$(awk '/_warning_/ {print $NF}' /tmp/cookie)"
$ curl -Lb /tmp/cookie \
  "https://drive.google.com/uc?export=download&confirm=${CODE}&id=1rATNHizJdVHnaJtt-hW9MOgjxoaajzdh" \
  -o pascal_voc_seg/VOCtrainval_11-May-2012.tar

$ sed -i -e "s/python .\/remove_gt_colormap.py/python3 .\/remove_gt_colormap.py/g" \
      -i -e "s/python .\/build_voc2012_data.py/python3 .\/build_voc2012_data.py/g" \
      download_and_convert_voc2012.sh

$ sh download_and_convert_voc2012.sh

$ cd ../..
$ mkdir -p deeplab/datasets/pascal_voc_seg/exp/train_on_train_set/train
$ mkdir -p deeplab/datasets/pascal_voc_seg/exp/train_on_train_set/eval
$ mkdir -p deeplab/datasets/pascal_voc_seg/exp/train_on_train_set/vis

$ export PATH_TO_TRAIN_DIR=${HOME}/deeplab/models/research/deeplab/datasets/pascal_voc_seg/exp/train_on_train_set/train
$ export PATH_TO_DATASET=${HOME}/deeplab/models/research/deeplab/datasets/pascal_voc_seg/tfrecord
$ export PYTHONPATH=${HOME}/deeplab/models/research:${HOME}/deeplab/models/research/deeplab:${HOME}/deeplab/models/research/slim:${PYTHONPATH}
# See feature_extractor.network_map for supported model variants.
# models/research/deeplab/core/feature_extractor.py

networks_map = {
    'mobilenet_v2': _mobilenet_v2,
    'mobilenet_v3_large_seg': mobilenet_v3_large_seg,
    'mobilenet_v3_small_seg': mobilenet_v3_small_seg,
    'resnet_v1_18': resnet_v1_beta.resnet_v1_18,
    'resnet_v1_18_beta': resnet_v1_beta.resnet_v1_18_beta,
    'resnet_v1_50': resnet_v1_beta.resnet_v1_50,
    'resnet_v1_50_beta': resnet_v1_beta.resnet_v1_50_beta,
    'resnet_v1_101': resnet_v1_beta.resnet_v1_101,
    'resnet_v1_101_beta': resnet_v1_beta.resnet_v1_101_beta,
    'xception_41': xception.xception_41,
    'xception_65': xception.xception_65,
    'xception_71': xception.xception_71,
    'nas_pnasnet': nas_network.pnasnet,
    'nas_hnasnet': nas_network.hnasnet,
}

2-1-2. "mobilenet_v3_small_seg" Float32 regular training

$ python3 deeplab/train.py \
    --logtostderr \
    --training_number_of_steps=500000 \
    --train_split="train" \
    --model_variant="mobilenet_v3_small_seg" \
    --decoder_output_stride=16 \
    --train_crop_size="513,513" \
    --train_batch_size=8 \
    --dataset="pascal_voc_seg" \
    --save_interval_secs=300 \
    --save_summaries_secs=300 \
    --save_summaries_images=True \
    --log_steps=100 \
    --train_logdir=${PATH_TO_TRAIN_DIR} \
    --dataset_dir=${PATH_TO_DATASET}

2-1-3. "mobilenet_v3_large_seg" Float32 regular training

$ python3 deeplab/train.py \
    --logtostderr \
    --training_number_of_steps=1000000 \
    --train_split="train" \
    --model_variant="mobilenet_v3_large_seg" \
    --decoder_output_stride=16 \
    --train_crop_size="513,513" \
    --train_batch_size=8 \
    --dataset="pascal_voc_seg" \
    --save_interval_secs=300 \
    --save_summaries_secs=300 \
    --save_summaries_images=True \
    --log_steps=100 \
    --train_logdir=${PATH_TO_TRAIN_DIR} \
    --dataset_dir=${PATH_TO_DATASET}

2-1-4. Visualize training status

$ tensorboard \
  --logdir ${HOME}/deeplab/models/research/deeplab/datasets/pascal_voc_seg/exp/train_on_train_set/train

   

2-2. MobileNetV3+DeeplabV3+Cityscaps - Post-training quantization

2-2-1. Preparation

$ cd ~
$ mkdir -p git/deeplab && cd git/deeplab
$ git clone --depth 1 https://github.com/tensorflow/models.git
$ cd models/research/deeplab/datasets
$ mkdir cityscapes && cd cityscapes

# Clone the script to generate Cityscapes Dataset.
$ git clone --depth 1 https://github.com/mcordts/cityscapesScripts.git
$ mv cityscapesScripts cityscapesScripts_ && \
  mv cityscapesScripts_/cityscapesscripts . && \
  rm -rf cityscapesScripts_

# Download Cityscapes Dataset.
# https://www.cityscapes-dataset.com/
# You will need to sign up and issue a userID and password to download the data set.
$ wget --keep-session-cookies --save-cookies=cookies.txt \
  --post-data 'username=(userid)&password=(password)&submit=Login' \
  https://www.cityscapes-dataset.com/login/
$ wget --load-cookies cookies.txt \
  --content-disposition https://www.cityscapes-dataset.com/file-handling/?packageID=1
$ wget --load-cookies cookies.txt \
  --content-disposition https://www.cityscapes-dataset.com/file-handling/?packageID=3
$ unzip gtFine_trainvaltest.zip && rm gtFine_trainvaltest.zip
$ rm README && rm license.txt
$ unzip leftImg8bit_trainvaltest.zip && rm leftImg8bit_trainvaltest.zip
$ rm README && rm license.txt

# Convert Cityscapes Dataset to TFRecords format.
$ cd ..
$ sed -i -e "s/python/python3/g" convert_cityscapes.sh
$ export PYTHONPATH=${HOME}/git/deeplab/models/research/deeplab/datasets/cityscapes:${PYTHONPATH}
$ sh convert_cityscapes.sh

# Create a checkpoint storage folder for training. If training is not required,
# there is no need to carry out.
$ cd ../..
$ mkdir -p deeplab/datasets/cityscapes/exp/train_on_train_set/train && \
  mkdir -p deeplab/datasets/cityscapes/exp/train_on_train_set/eval && \
  mkdir -p deeplab/datasets/cityscapes/exp/train_on_train_set/vis

# Download the DeepLabV3 trained model of the MobileNetV3 backbone.
$ curl -sc /tmp/cookie \
  "https://drive.google.com/uc?export=download&id=1f5ccaJmJBYwBmHvRQ77yGIUcXnqQIRY_" > /dev/null
$ CODE="$(awk '/_warning_/ {print $NF}' /tmp/cookie)"
$ curl -Lb /tmp/cookie \
  "https://drive.google.com/uc?export=download&confirm=${CODE}&id=1f5ccaJmJBYwBmHvRQ77yGIUcXnqQIRY_" \
  -o deeplab_mnv3_small_cityscapes_trainfine_2019_11_15.tar.gz
$ tar -zxvf deeplab_mnv3_small_cityscapes_trainfine_2019_11_15.tar.gz
$ rm deeplab_mnv3_small_cityscapes_trainfine_2019_11_15.tar.gz

$ curl -sc /tmp/cookie \
  "https://drive.google.com/uc?export=download&id=1QxS3G55rUQvuiBF-hztQv5zCkfPfwlVU" > /dev/null
$ CODE="$(awk '/_warning_/ {print $NF}' /tmp/cookie)"
$ curl -Lb /tmp/cookie \
  "https://drive.google.com/uc?export=download&confirm=${CODE}&id=1QxS3G55rUQvuiBF-hztQv5zCkfPfwlVU" \
  -o deeplab_mnv3_large_cityscapes_trainfine_2019_11_15.tar.gz
$ tar -zxvf deeplab_mnv3_large_cityscapes_trainfine_2019_11_15.tar.gz
$ rm deeplab_mnv3_large_cityscapes_trainfine_2019_11_15.tar.gz

$ export PATH_TO_INITIAL_CHECKPOINT=${HOME}/git/deeplab/models/research/deeplab_mnv3_small_cityscapes_trainfine/model.ckpt
$ export PATH_TO_DATASET=${HOME}/git/deeplab/models/research/deeplab/datasets/cityscapes/tfrecord
$ export PYTHONPATH=${HOME}/git/deeplab/models/research:${HOME}/git/deeplab/models/research/deeplab:${HOME}/git/deeplab/models/research/slim:${PYTHONPATH}

# Fix a bug in the data generator.
$ sed -i -e \
  "s/splits_to_sizes={'train_fine': 2975,/splits_to_sizes={'train': 2975,/g" \
  deeplab/datasets/data_generator.py

# Back up the trained model.
$ cd ${HOME}/git/deeplab/models/research
$ cp deeplab/export_model.py deeplab/export_model.py_org
$ cp deeplab_mnv3_small_cityscapes_trainfine/frozen_inference_graph.pb \
  deeplab_mnv3_small_cityscapes_trainfine/frozen_inference_graph_org.pb
$ cp deeplab_mnv3_large_cityscapes_trainfine/frozen_inference_graph.pb \
  deeplab_mnv3_large_cityscapes_trainfine/frozen_inference_graph_org.pb

# Customize "export_model.py" according to the input resolution. Must be (multiple of 8 + 1).
#   (example.1) 769 = 8 * 96 + 1
#   (example.2) 512 = 8 * 64 + 1
#   (example.3) 320 = 8 * 40 + 1
# And it is necessary to change from tf.uint8 type to tf.float32 type.
$ sed -i -e \
  "s/tf.placeholder(tf.uint8, \[1, None, None, 3\], name=_INPUT_NAME)/tf.placeholder(tf.float32, \[1, 769, 769, 3\], name=_INPUT_NAME)/g" \
  deeplab/export_model.py

2-2-2. Parameter sheet

# crop_size and image_pooling_crop_size are multiples of --decoder_output_stride + 1
# 769 = 8 * 96 + 1
# 513 = 8 * 64 + 1
# 321 = 8 * 40 + 1

# --initialize_last_layer=True initializes the final layer with the weight of
# tf_initial_checkpoint (inherits the weight)

# Named tuple to describe the dataset properties.
# deeplab/datasets/data_generator.py
DatasetDescriptor = collections.namedtuple(
    'DatasetDescriptor',
    [
        'splits_to_sizes',  # Splits of the dataset into training, val and test.
        'num_classes',  # Number of semantic classes, including the
                        # background class (if exists). For example, there
                        # are 20 foreground classes + 1 background class in
                        # the PASCAL VOC 2012 dataset. Thus, we set
                        # num_classes=21.
        'ignore_label',  # Ignore label value.
    ])

_CITYSCAPES_INFORMATION = DatasetDescriptor(
    splits_to_sizes={'train': 2975,
                     'train_coarse': 22973,
                     'trainval_fine': 3475,
                     'trainval_coarse': 23473,
                     'val_fine': 500,
                     'test_fine': 1525},
    num_classes=19,
    ignore_label=255,
)

_PASCAL_VOC_SEG_INFORMATION = DatasetDescriptor(
    splits_to_sizes={
        'train': 1464,
        'train_aug': 10582,
        'trainval': 2913,
        'val': 1449,
    },
    num_classes=21,
    ignore_label=255,
)

_ADE20K_INFORMATION = DatasetDescriptor(
    splits_to_sizes={
        'train': 20210,  # num of samples in images/training
        'val': 2000,  # num of samples in images/validation
    },
    num_classes=151,
    ignore_label=0,
)

_DATASETS_INFORMATION = {
    'cityscapes': _CITYSCAPES_INFORMATION,
    'pascal_voc_seg': _PASCAL_VOC_SEG_INFORMATION,
    'ade20k': _ADE20K_INFORMATION,
}

# A map from network name to network function. model_variant.
# deeplab/core/feature_extractor.py
networks_map = {
    'mobilenet_v2': _mobilenet_v2,
    'mobilenet_v3_large_seg': mobilenet_v3_large_seg,
    'mobilenet_v3_small_seg': mobilenet_v3_small_seg,
    'resnet_v1_18': resnet_v1_beta.resnet_v1_18,
    'resnet_v1_18_beta': resnet_v1_beta.resnet_v1_18_beta,
    'resnet_v1_50': resnet_v1_beta.resnet_v1_50,
    'resnet_v1_50_beta': resnet_v1_beta.resnet_v1_50_beta,
    'resnet_v1_101': resnet_v1_beta.resnet_v1_101,
    'resnet_v1_101_beta': resnet_v1_beta.resnet_v1_101_beta,
    'xception_41': xception.xception_41,
    'xception_65': xception.xception_65,
    'xception_71': xception.xception_71,
    'nas_pnasnet': nas_network.pnasnet,
    'nas_hnasnet': nas_network.hnasnet,
}

2-2-3. "mobilenet_v3_small_seg" Export Model

Generate Freeze Graph (.pb) with INPUT Placeholder changed from checkpoint file (.ckpt).

$ python3 deeplab/export_model.py \
    --checkpoint_path=./deeplab_mnv3_small_cityscapes_trainfine/model.ckpt \
    --export_path=./deeplab_mnv3_small_cityscapes_trainfine/frozen_inference_graph.pb \
    --num_classes=19 \
    --crop_size=769 \
    --crop_size=769 \
    --model_variant="mobilenet_v3_small_seg" \
    --image_pooling_crop_size="769,769" \
    --image_pooling_stride=4,5 \
    --aspp_convs_filters=128 \
    --aspp_with_concat_projection=0 \
    --aspp_with_squeeze_and_excitation=1 \
    --decoder_use_sum_merge=1 \
    --decoder_filters=19 \
    --decoder_output_is_logits=1 \
    --image_se_uses_qsigmoid=1 \
    --image_pyramid=1 \
    --decoder_output_stride=8

2-2-4. "mobilenet_v3_large_seg" Export Model

Generate Freeze Graph (.pb) with INPUT Placeholder changed from checkpoint file (.ckpt).

$ python3 deeplab/export_model.py \
    --checkpoint_path=./deeplab_mnv3_large_cityscapes_trainfine/model.ckpt \
    --export_path=./deeplab_mnv3_large_cityscapes_trainfine/frozen_inference_graph.pb \
    --num_classes=19 \
    --crop_size=769 \
    --crop_size=769 \
    --model_variant="mobilenet_v3_large_seg" \
    --image_pooling_crop_size="769,769" \
    --image_pooling_stride=4,5 \
    --aspp_convs_filters=128 \
    --aspp_with_concat_projection=0 \
    --aspp_with_squeeze_and_excitation=1 \
    --decoder_use_sum_merge=1 \
    --decoder_filters=19 \
    --decoder_output_is_logits=1 \
    --image_se_uses_qsigmoid=1 \
    --image_pyramid=1 \
    --decoder_output_stride=8

If you follow the Google Colaboratory sample procedure, copy the "deeplab_mnv3_small_cityscapes_trainfine" folder and "deeplab_mnv3_large_cityscapes_trainfine" to your Google Drive "My Drive". It is not necessary if all procedures described in Google Colaboratory are performed in a PC environment. 001 002

2-2-5. Google Colaboratory - Post-training quantization - post_training_integer_quant.ipynb

  • Weight Quantization
  • Integer Quantization
  • Full Integer Quantization

https://colab.research.google.com/drive/1TtCJ-uMNTArpZxrf5DCNbZdn08DsiW8F    

2-3. MobileNetV3+DeeplabV3+Cityscaps - Quantization-aware training

2-3-1. "mobilenet_v3_small_seg" Quantization-aware training

$ cd ${HOME}/git/deeplab/models/research
$ export PATH_TO_TRAINED_FLOAT_MODEL=${HOME}/git/deeplab/models/research/deeplab_mnv3_small_cityscapes_trainfine/model.ckpt
$ export PATH_TO_TRAIN_DIR=${HOME}/git/deeplab/models/research/deeplab/datasets/cityscapes/exp/train_on_train_set/train
$ export PATH_TO_DATASET=${HOME}/git/deeplab/models/research/deeplab/datasets/cityscapes/tfrecord

# deeplab_mnv3_small_cityscapes_trainfine
$ python3 deeplab/train.py \
    --logtostderr \
    --training_number_of_steps=5000 \
    --train_split="train" \
    --model_variant="mobilenet_v3_small_seg" \
    --train_crop_size="769,769" \
    --train_batch_size=8 \
    --dataset="cityscapes" \
    --initialize_last_layer=False \
    --base_learning_rate=3e-5 \
    --quantize_delay_step=0 \
    --image_pooling_crop_size="769,769" \
    --image_pooling_stride=4,5 \
    --aspp_convs_filters=128 \
    --aspp_with_concat_projection=0 \
    --aspp_with_squeeze_and_excitation=1 \
    --decoder_use_sum_merge=1 \
    --decoder_filters=19 \
    --decoder_output_is_logits=1 \
    --image_se_uses_qsigmoid=1 \
    --image_pyramid=1 \
    --decoder_output_stride=8 \
    --save_interval_secs=300 \
    --save_summaries_secs=300 \
    --save_summaries_images=True \
    --log_steps=100 \
    --tf_initial_checkpoint=${PATH_TO_TRAINED_FLOAT_MODEL} \
    --train_logdir=${PATH_TO_TRAIN_DIR} \
    --dataset_dir=${PATH_TO_DATASET}

2-3-2. "mobilenet_v3_large_seg" Quantization-aware training

$ cd ${HOME}/git/deeplab/models/research
$ export PATH_TO_TRAINED_FLOAT_MODEL=${HOME}/git/deeplab/models/research/deeplab_mnv3_large_cityscapes_trainfine/model.ckpt
$ export PATH_TO_TRAIN_DIR=${HOME}/git/deeplab/models/research/deeplab/datasets/cityscapes/exp/train_on_train_set/train
$ export PATH_TO_DATASET=${HOME}/git/deeplab/models/research/deeplab/datasets/cityscapes/tfrecord

# deeplab_mnv3_large_cityscapes_trainfine
$ python3 deeplab/train.py \
    --logtostderr \
    --training_number_of_steps=4350 \
    --train_split="train" \
    --model_variant="mobilenet_v3_large_seg" \
    --train_crop_size="769,769" \
    --train_batch_size=8 \
    --dataset="cityscapes" \
    --initialize_last_layer=False \
    --base_learning_rate=3e-5 \
    --quantize_delay_step=0 \
    --image_pooling_crop_size="769,769" \
    --image_pooling_stride=4,5 \
    --aspp_convs_filters=128 \
    --aspp_with_concat_projection=0 \
    --aspp_with_squeeze_and_excitation=1 \
    --decoder_use_sum_merge=1 \
    --decoder_filters=19 \
    --decoder_output_is_logits=1 \
    --image_se_uses_qsigmoid=1 \
    --image_pyramid=1 \
    --decoder_output_stride=8 \
    --save_interval_secs=300 \
    --save_summaries_secs=300 \
    --save_summaries_images=True \
    --log_steps=100 \
    --tf_initial_checkpoint=${PATH_TO_TRAINED_FLOAT_MODEL} \
    --train_logdir=${PATH_TO_TRAIN_DIR} \
    --dataset_dir=${PATH_TO_DATASET}

The orange line is "deeplab_mnv3_small_cityscapes_trainfine" loss. The blue line is "deeplab_mnv3_large_cityscapes_trainfine" loss. 003    

2-4. MobileNetV2+DeeplabV3+coco/voc - Post-training quantization

2-4-1. Preparation

$ cd ${HOME}/git/deeplab/models/research

$ wget http://download.tensorflow.org/models/deeplabv3_mnv2_dm05_pascal_trainaug_2018_10_01.tar.gz
$ tar -zxvf deeplabv3_mnv2_dm05_pascal_trainaug_2018_10_01.tar.gz
$ rm deeplabv3_mnv2_dm05_pascal_trainaug_2018_10_01.tar.gz

$ wget http://download.tensorflow.org/models/deeplabv3_mnv2_dm05_pascal_trainval_2018_10_01.tar.gz
$ tar -zxvf deeplabv3_mnv2_dm05_pascal_trainval_2018_10_01.tar.gz
$ rm deeplabv3_mnv2_dm05_pascal_trainval_2018_10_01.tar.gz

$ wget http://download.tensorflow.org/models/deeplabv3_mnv2_pascal_train_aug_2018_01_29.tar.gz
$ tar -zxvf deeplabv3_mnv2_pascal_train_aug_2018_01_29.tar.gz
$ rm deeplabv3_mnv2_pascal_train_aug_2018_01_29.tar.gz

$ sed -i -e \
  "s/tf.placeholder(tf.uint8, \[1, None, None, 3\], name=_INPUT_NAME)/tf.placeholder(tf.float32, \[1, 257, 257, 3\], name=_INPUT_NAME)/g" \
  deeplab/export_model.py

$ export PYTHONPATH=${HOME}/git/deeplab/models/research:${HOME}/git/deeplab/models/research/deeplab:${HOME}/git/deeplab/models/research/slim:${PYTHONPATH}

$ python3 deeplab/export_model.py \
  --checkpoint_path=./deeplabv3_mnv2_dm05_pascal_trainaug/model.ckpt \
  --export_path=./deeplabv3_mnv2_dm05_pascal_trainaug/frozen_inference_graph.pb \
  --model_variant="mobilenet_v2" \
  --crop_size=257 \
  --crop_size=257 \
  --depth_multiplier=0.5

$ python3 deeplab/export_model.py \
  --checkpoint_path=./deeplabv3_mnv2_dm05_pascal_trainval/model.ckpt \
  --export_path=./deeplabv3_mnv2_dm05_pascal_trainval/frozen_inference_graph.pb \
  --model_variant="mobilenet_v2" \
  --crop_size=257 \
  --crop_size=257 \
  --depth_multiplier=0.5

$ python3 deeplab/export_model.py \
  --checkpoint_path=./deeplabv3_mnv2_pascal_train_aug/model.ckpt-30000 \
  --export_path=./deeplabv3_mnv2_pascal_train_aug/frozen_inference_graph.pb \
  --model_variant="mobilenet_v2" \
  --crop_size=257 \
  --crop_size=257

2-5. MobileNetV3-SSD+coco - Post-training quantization

2-5-1. Preparation

$ cd ~
$ sudo pip3 install tensorflow-gpu==1.15.0
$ git clone --depth 1 https://github.com/tensorflow/models.git
$ cd models/research

$ git clone https://github.com/cocodataset/cocoapi.git
$ cd cocoapi/PythonAPI
$ make
$ cp -r pycocotools ../..
$ cd ../..
$ wget -O protobuf.zip https://github.com/google/protobuf/releases/download/v3.0.0/protoc-3.0.0-linux-x86_64.zip
$ unzip protobuf.zip
$ ./bin/protoc object_detection/protos/*.proto --python_out=.

$ sudo apt-get install -y protobuf-compiler python3-pil python3-lxml python3-tk
$ sudo -H pip3 install Cython contextlib2 jupyter matplotlib

$ export PYTHONPATH=${PWD}:${PWD}/object_detection:${PWD}/slim:${PYTHONPATH}

$ mkdir -p ssd_mobilenet_v3_small_coco_2019_08_14 && cd ssd_mobilenet_v3_small_coco_2019_08_14
$ curl -sc /tmp/cookie "https://drive.google.com/uc?export=download&id=1uqaC0Y-yRtzkpu1EuZ3BzOyh9-i_3Qgi" > /dev/null
$ CODE="$(awk '/_warning_/ {print $NF}' /tmp/cookie)"
$ curl -Lb /tmp/cookie "https://drive.google.com/uc?export=download&confirm=${CODE}&id=1uqaC0Y-yRtzkpu1EuZ3BzOyh9-i_3Qgi" -o ssd_mobilenet_v3_small_coco_2019_08_14.tar.gz
$ tar -zxvf ssd_mobilenet_v3_small_coco_2019_08_14.tar.gz
$ rm ssd_mobilenet_v3_small_coco_2019_08_14.tar.gz
$ cd ..

$ mkdir -p ssd_mobilenet_v3_large_coco_2019_08_14 && cd ssd_mobilenet_v3_large_coco_2019_08_14
$ curl -sc /tmp/cookie "https://drive.google.com/uc?export=download&id=1NGLjKRWDQZ_kibQHlLZ7Eetuuz1waC7X" > /dev/null
$ CODE="$(awk '/_warning_/ {print $NF}' /tmp/cookie)"
$ curl -Lb /tmp/cookie "https://drive.google.com/uc?export=download&confirm=${CODE}&id=1NGLjKRWDQZ_kibQHlLZ7Eetuuz1waC7X" -o ssd_mobilenet_v3_large_coco_2019_08_14.tar.gz
$ tar -zxvf ssd_mobilenet_v3_large_coco_2019_08_14.tar.gz
$ rm ssd_mobilenet_v3_large_coco_2019_08_14.tar.gz
$ cd ..

2-5-2. Create a conversion script from checkpoint format to saved_model format

import tensorflow as tf
import os
import shutil
from tensorflow.python.saved_model import tag_constants
from tensorflow.python.tools import freeze_graph
from tensorflow.python import ops
from tensorflow.tools.graph_transforms import TransformGraph

def freeze_model(saved_model_dir, output_node_names, output_filename):
  output_graph_filename = os.path.join(saved_model_dir, output_filename)
  initializer_nodes = ''
  freeze_graph.freeze_graph(
      input_saved_model_dir=saved_model_dir,
      output_graph=output_graph_filename,
      saved_model_tags = tag_constants.SERVING,
      output_node_names=output_node_names,
      initializer_nodes=initializer_nodes,
      input_graph=None,
      input_saver=False,
      input_binary=False,
      input_checkpoint=None,
      restore_op_name=None,
      filename_tensor_name=None,
      clear_devices=True,
      input_meta_graph=False,
  )

def get_graph_def_from_file(graph_filepath):
  tf.reset_default_graph()
  with ops.Graph().as_default():
    with tf.gfile.GFile(graph_filepath, 'rb') as f:
      graph_def = tf.GraphDef()
      graph_def.ParseFromString(f.read())
      return graph_def

def optimize_graph(model_dir, graph_filename, transforms, input_name, output_names, outname='optimized_model.pb'):
  input_names = [input_name] # change this as per how you have saved the model
  graph_def = get_graph_def_from_file(os.path.join(model_dir, graph_filename))
  optimized_graph_def = TransformGraph(
      graph_def,
      input_names,
      output_names,
      transforms)
  tf.train.write_graph(optimized_graph_def,
                      logdir=model_dir,
                      as_text=False,
                      name=outname)
  print('Graph optimized!')

def convert_graph_def_to_saved_model(export_dir, graph_filepath, input_name, outputs):
  graph_def = get_graph_def_from_file(graph_filepath)
  with tf.Session(graph=tf.Graph()) as session:
    tf.import_graph_def(graph_def, name='')
    tf.compat.v1.saved_model.simple_save(
        session,
        export_dir,# change input_image to node.name if you know the name
        inputs={input_name: session.graph.get_tensor_by_name('{}:0'.format(node.name))
            for node in graph_def.node if node.op=='Placeholder'},
        outputs={t.rstrip(":0"):session.graph.get_tensor_by_name(t) for t in outputs}
    )
    print('Optimized graph converted to SavedModel!')

tf.compat.v1.enable_eager_execution()

# Look up the name of the placeholder for the input node
graph_def=get_graph_def_from_file('./ssd_mobilenet_v3_small_coco_2019_08_14/frozen_inference_graph.pb')
input_name_small=""
for node in graph_def.node:
    if node.op=='Placeholder':
        print("##### ssd_mobilenet_v3_small_coco_2019_08_14 - Input Node Name #####", node.name) # this will be the input node
        input_name_small=node.name

# Look up the name of the placeholder for the input node
graph_def=get_graph_def_from_file('./ssd_mobilenet_v3_large_coco_2019_08_14/frozen_inference_graph.pb')
input_name_large=""
for node in graph_def.node:
    if node.op=='Placeholder':
        print("##### ssd_mobilenet_v3_large_coco_2019_08_14 - Input Node Name #####", node.name) # this will be the input node
        input_name_large=node.name

# ssd_mobilenet_v3 output names
output_node_names = ['raw_outputs/class_predictions','raw_outputs/box_encodings']
outputs = ['raw_outputs/class_predictions:0','raw_outputs/box_encodings:0']

# Optimizing the graph via TensorFlow library
transforms = []
optimize_graph('./ssd_mobilenet_v3_small_coco_2019_08_14', 'frozen_inference_graph.pb', transforms, input_name_small, output_node_names, outname='optimized_model_small.pb')
optimize_graph('./ssd_mobilenet_v3_large_coco_2019_08_14', 'frozen_inference_graph.pb', transforms, input_name_large, output_node_names, outname='optimized_model_large.pb')

# convert this to a s TF Serving compatible mode - ssd_mobilenet_v3_small_coco_2019_08_14
shutil.rmtree('./ssd_mobilenet_v3_small_coco_2019_08_14/0', ignore_errors=True)
convert_graph_def_to_saved_model('./ssd_mobilenet_v3_small_coco_2019_08_14/0',
                                 './ssd_mobilenet_v3_small_coco_2019_08_14/optimized_model_small.pb', input_name_small, outputs)

# convert this to a s TF Serving compatible mode - ssd_mobilenet_v3_large_coco_2019_08_14
shutil.rmtree('./ssd_mobilenet_v3_large_coco_2019_08_14/0', ignore_errors=True)
convert_graph_def_to_saved_model('./ssd_mobilenet_v3_large_coco_2019_08_14/0',
                                 './ssd_mobilenet_v3_large_coco_2019_08_14/optimized_model_large.pb', input_name_large, outputs)

2-5-3. Confirm the structure of saved_model 【ssd_mobilenet_v3_small_coco_2019_08_14】

$ saved_model_cli show --dir ./ssd_mobilenet_v3_small_coco_2019_08_14/0 --all

MetaGraphDef with tag-set: 'serve' contains the following SignatureDefs:

signature_def['serving_default']:
  The given SavedModel SignatureDef contains the following input(s):
    inputs['normalized_input_image_tensor'] tensor_info:
        dtype: DT_FLOAT
        shape: (1, 320, 320, 3)
        name: normalized_input_image_tensor:0
  The given SavedModel SignatureDef contains the following output(s):
    outputs['raw_outputs/box_encodings'] tensor_info:
        dtype: DT_FLOAT
        shape: (1, 2034, 4)
        name: raw_outputs/box_encodings:0
    outputs['raw_outputs/class_predictions'] tensor_info:
        dtype: DT_FLOAT
        shape: (1, 2034, 91)
        name: raw_outputs/class_predictions:0
  Method name is: tensorflow/serving/predict

2-5-4. Confirm the structure of saved_model 【ssd_mobilenet_v3_large_coco_2019_08_14】

$ saved_model_cli show --dir ./ssd_mobilenet_v3_large_coco_2019_08_14/0 --all

MetaGraphDef with tag-set: 'serve' contains the following SignatureDefs:

signature_def['serving_default']:
  The given SavedModel SignatureDef contains the following input(s):
    inputs['normalized_input_image_tensor'] tensor_info:
        dtype: DT_FLOAT
        shape: (1, 320, 320, 3)
        name: normalized_input_image_tensor:0
  The given SavedModel SignatureDef contains the following output(s):
    outputs['raw_outputs/box_encodings'] tensor_info:
        dtype: DT_FLOAT
        shape: (1, 2034, 4)
        name: raw_outputs/box_encodings:0
    outputs['raw_outputs/class_predictions'] tensor_info:
        dtype: DT_FLOAT
        shape: (1, 2034, 91)
        name: raw_outputs/class_predictions:0
  Method name is: tensorflow/serving/predict

2-5-5. Creating the destination path for the calibration test dataset 6GB

$ curl -sc /tmp/cookie "https://drive.google.com/uc?export=download&id=1Uk9F4Tc-9UgnvARIVkloSoePUynyST6E" > /dev/null
$ CODE="$(awk '/_warning_/ {print $NF}' /tmp/cookie)"
$ curl -Lb /tmp/cookie "https://drive.google.com/uc?export=download&confirm=${CODE}&id=1Uk9F4Tc-9UgnvARIVkloSoePUynyST6E" -o TFDS.tar.gz
$ tar -zxvf TFDS.tar.gz
$ rm TFDS.tar.gz

2-5-6. Quantization

2-5-6-1. ssd_mobilenet_v3_small_coco_2019_08_14
import tensorflow as tf
import tensorflow_datasets as tfds
import numpy as np

def representative_dataset_gen():
  for data in raw_test_data.take(100):
    image = data['image'].numpy()
    image = tf.image.resize(image, (320, 320))
    image = image[np.newaxis,:,:,:]
    yield [image]

tf.compat.v1.enable_eager_execution()

# Generating a calibration data set
#raw_test_data, info = tfds.load(name="coco/2017", with_info=True, split="test", data_dir="./TFDS")
raw_test_data, info = tfds.load(name="coco/2017", with_info=True, split="test", data_dir="./TFDS", download=False)
print(info)

# Weight Quantization - Input/Output=float32
converter = tf.lite.TFLiteConverter.from_saved_model('./ssd_mobilenet_v3_small_coco_2019_08_14/0')
converter.optimizations = [tf.lite.Optimize.OPTIMIZE_FOR_SIZE]
tflite_quant_model = converter.convert()
with open('./ssd_mobilenet_v3_small_coco_2019_08_14/mobilenet_v3_small_weight_quant.tflite', 'wb') as w:
    w.write(tflite_quant_model)
print("Weight Quantization complete! - mobilenet_v3_small_weight_quant.tflite")

# Integer Quantization - Input/Output=float32
converter = tf.lite.TFLiteConverter.from_saved_model('./ssd_mobilenet_v3_small_coco_2019_08_14/0')
converter.optimizations = [tf.lite.Optimize.DEFAULT]
converter.representative_dataset = representative_dataset_gen
tflite_quant_model = converter.convert()
with open('./ssd_mobilenet_v3_small_coco_2019_08_14/mobilenet_v3_small_integer_quant.tflite', 'wb') as w:
    w.write(tflite_quant_model)
print("Integer Quantization complete! - mobilenet_v3_small_integer_quant.tflite")

# Full Integer Quantization - Input/Output=int8
converter = tf.lite.TFLiteConverter.from_saved_model('./ssd_mobilenet_v3_small_coco_2019_08_14/0')
converter.optimizations = [tf.lite.Optimize.DEFAULT]
converter.representative_dataset = representative_dataset_gen
converter.target_spec.supported_ops = [tf.lite.OpsSet.TFLITE_BUILTINS_INT8]
converter.inference_input_type = tf.uint8
converter.inference_output_type = tf.uint8
tflite_quant_model = converter.convert()
with open('./ssd_mobilenet_v3_small_coco_2019_08_14/mobilenet_v3_small_full_integer_quant.tflite', 'wb') as w:
    w.write(tflite_quant_model)
print("Full Integer Quantization complete! - mobilenet_v3_small_full_integer_quant.tflite")
2-5-6-2. ssd_mobilenet_v3_large_coco_2019_08_14
import tensorflow as tf
import tensorflow_datasets as tfds
import numpy as np

def representative_dataset_gen():
  for data in raw_test_data.take(100):
    image = data['image'].numpy()
    image = tf.image.resize(image, (320, 320))
    image = image[np.newaxis,:,:,:]
    yield [image]

tf.compat.v1.enable_eager_execution()

# Generating a calibration data set
#raw_test_data, info = tfds.load(name="coco/2017", with_info=True, split="test", data_dir="./TFDS")
raw_test_data, info = tfds.load(name="coco/2017", with_info=True, split="test", data_dir="./TFDS", download=False)

# Weight Quantization - Input/Output=float32
converter = tf.lite.TFLiteConverter.from_saved_model('./ssd_mobilenet_v3_large_coco_2019_08_14/0')
converter.optimizations = [tf.lite.Optimize.OPTIMIZE_FOR_SIZE]
tflite_quant_model = converter.convert()
with open('./ssd_mobilenet_v3_large_coco_2019_08_14/mobilenet_v3_large_weight_quant.tflite', 'wb') as w:
    w.write(tflite_quant_model)
print("Weight Quantization complete! - mobilenet_v3_large_weight_quant.tflite")

# Integer Quantization - Input/Output=float32
converter = tf.lite.TFLiteConverter.from_saved_model('./ssd_mobilenet_v3_large_coco_2019_08_14/0')
converter.optimizations = [tf.lite.Optimize.DEFAULT]
converter.representative_dataset = representative_dataset_gen
tflite_quant_model = converter.convert()
with open('./ssd_mobilenet_v3_large_coco_2019_08_14/mobilenet_v3_large_integer_quant.tflite', 'wb') as w:
    w.write(tflite_quant_model)
print("Integer Quantization complete! - mobilenet_v3_large_integer_quant.tflite")

# Full Integer Quantization - Input/Output=int8
converter = tf.lite.TFLiteConverter.from_saved_model('./ssd_mobilenet_v3_large_coco_2019_08_14/0')
converter.optimizations = [tf.lite.Optimize.DEFAULT]
converter.representative_dataset = representative_dataset_gen
converter.target_spec.supported_ops = [tf.lite.OpsSet.TFLITE_BUILTINS_INT8]
converter.inference_input_type = tf.uint8
converter.inference_output_type = tf.uint8
tflite_quant_model = converter.convert()
with open('./ssd_mobilenet_v3_large_coco_2019_08_14/mobilenet_v3_large_full_integer_quant.tflite', 'wb') as w:
    w.write(tflite_quant_model)
print("Full Integer Quantization complete! - mobilenet_v3_large_full_integer_quant.tflite")

2-6. MobileNetV2-SSDLite+VOC - Training -> Integer Quantization

2-6-1. Training

Learning with the MobileNetV2-SSDLite Pascal-VOC dataset [Remake of Docker version]

2-6-2. Export model (--add_postprocessing_op=True)

06_mobilenetv2-ssdlite/02_voc/01_float32/00_export_tflite_model.txt

2-6-3. Integer Quantization

06_mobilenetv2-ssdlite/02_voc/01_float32/03_integer_quantization_with_postprocess.py

3. TFLite Model Benchmark

$ sudo apt-get install python-future

## Bazel for Ubuntu18.04 x86_64 install
$ wget https://github.com/bazelbuild/bazel/releases/download/2.0.0/bazel-2.0.0-installer-linux-x86_64.sh
$ sudo chmod +x bazel-2.0.0-installer-linux-x86_64.sh
$ ./bazel-2.0.0-installer-linux-x86_64.sh
$ sudo apt-get install -y openjdk-8-jdk

## Bazel for RaspberryPi3/4 Raspbian/Debian Buster armhf install
$ wget https://github.com/PINTO0309/Bazel_bin/raw/master/3.1.0/Raspbian_Debian_Buster_armhf/openjdk-8-jdk/install.sh
$ ./install.sh
$ curl -sc /tmp/cookie \
  "https://drive.google.com/uc?export=download&id=1LQUSal55R6fmawZS9zZuk6-5ZFOdUqRK" > /dev/null
$ CODE="$(awk '/_warning_/ {print $NF}' /tmp/cookie)"
$ curl -Lb /tmp/cookie \
  "https://drive.google.com/uc?export=download&confirm=${CODE}&id=1LQUSal55R6fmawZS9zZuk6-5ZFOdUqRK" \
  -o adoptopenjdk-8-hotspot_8u222-b10-2_armhf.deb
$ sudo apt-get install -y ./adoptopenjdk-8-hotspot_8u222-b10-2_armhf.deb

## Bazel for RaspberryPi3/4 Raspbian/Debian Buster aarch64 install
$ wget https://github.com/PINTO0309/Bazel_bin/raw/master/3.1.0/Raspbian_Debian_Buster_aarch64/openjdk-8-jdk/install.sh
$ ./install.sh
$ curl -sc /tmp/cookie \
  "https://drive.google.com/uc?export=download&id=1VwLxzT3EOTbhSzwvRF2H4ChTQyTQBt3x" > /dev/null
$ CODE="$(awk '/_warning_/ {print $NF}' /tmp/cookie)"
$ curl -Lb /tmp/cookie \
  "https://drive.google.com/uc?export=download&confirm=${CODE}&id=1VwLxzT3EOTbhSzwvRF2H4ChTQyTQBt3x" \
  -o adoptopenjdk-8-hotspot_8u222-b10-2_arm64.deb
$ sudo apt-get install -y ./adoptopenjdk-8-hotspot_8u222-b10-2_arm64.deb

## Clone Tensorflow v2.1.0+
$ git clone --depth 1 https://github.com/tensorflow/tensorflow.git
$ cd tensorflow

## Build and run TFLite Model Benchmark Tool
$ bazel run -c opt tensorflow/lite/tools/benchmark:benchmark_model -- \
  --graph=${HOME}/Downloads/deeplabv3_257_mv_gpu.tflite \
  --num_threads=4 \
  --warmup_runs=1 \
  --enable_op_profiling=true

$ bazel run -c opt tensorflow/lite/tools/benchmark:benchmark_model -- \
  --graph=${HOME}/Downloads/deeplabv3_257_mv_gpu.tflite \
  --num_threads=4 \
  --warmup_runs=1 \
  --use_xnnpack=true \
  --enable_op_profiling=true

$ bazel run \
  -c opt \
  --config=noaws \
  --config=nohdfs \
  --config=nonccl \
  tensorflow/lite/tools/benchmark:benchmark_model_plus_flex -- \
  --graph=${HOME}/git/tf-monodepth2/monodepth2_flexdelegate_weight_quant.tflite \
  --num_threads=4 \
  --warmup_runs=1 \
  --enable_op_profiling=true

$ bazel run \
  -c opt \
  --config=noaws \
  --config=nohdfs \
  --config=nonccl \
  tensorflow/lite/tools/benchmark:benchmark_model_plus_flex -- \
  --graph=${HOME}/git/tf-monodepth2/monodepth2_flexdelegate_weight_quant.tflite \
  --num_threads=4 \
  --warmup_runs=1 \
  --use_xnnpack=true \
  --enable_op_profiling=true
x86_64 deeplab_mnv3_small_weight_quant_769.tflite Benchmark
Number of nodes executed: 171
============================== Summary by node type ==============================
	             [Node type]	  [count]	  [avg ms]	    [avg %]	    [cdf %]	  [mem KB]	[times called]
	                 CONV_2D	       45	  1251.486	    67.589%	    67.589%	     0.000	        0
	       DEPTHWISE_CONV_2D	       11	   438.764	    23.696%	    91.286%	     0.000	        0
	              HARD_SWISH	       16	    54.855	     2.963%	    94.248%	     0.000	        0
	                 ARG_MAX	        1	    24.850	     1.342%	    95.591%	     0.000	        0
	         RESIZE_BILINEAR	        5	    23.805	     1.286%	    96.876%	     0.000	        0
	                     MUL	       30	    14.914	     0.805%	    97.682%	     0.000	        0
	                     ADD	       18	    10.646	     0.575%	    98.257%	     0.000	        0
	       SPACE_TO_BATCH_ND	        7	     9.567	     0.517%	    98.773%	     0.000	        0
	       BATCH_TO_SPACE_ND	        7	     7.431	     0.401%	    99.175%	     0.000	        0
	                     SUB	        2	     6.131	     0.331%	    99.506%	     0.000	        0
	         AVERAGE_POOL_2D	       10	     5.435	     0.294%	    99.799%	     0.000	        0
	                 RESHAPE	        6	     2.171	     0.117%	    99.916%	     0.000	        0
	                     PAD	        1	     0.660	     0.036%	    99.952%	     0.000	        0
	                    CAST	        2	     0.601	     0.032%	    99.985%	     0.000	        0
	           STRIDED_SLICE	        1	     0.277	     0.015%	   100.000%	     0.000	        0
	        Misc Runtime Ops	        1	     0.008	     0.000%	   100.000%	    33.552	        0
	              DEQUANTIZE	        8	     0.000	     0.000%	   100.000%	     0.000	        0

Timings (microseconds): count=52 first=224 curr=1869070 min=224 max=2089397 avg=1.85169e+06 std=373988
Memory (bytes): count=0
171 nodes observed
x86_64 deeplab_mnv3_large_weight_quant_769.tflite Benchmark
Number of nodes executed: 194
============================== Summary by node type ==============================
	             [Node type]	  [count]	  [avg ms]	    [avg %]	    [cdf %]	  [mem KB]	[times called]
	                 CONV_2D	       51	  4123.348	    82.616%	    82.616%	     0.000	        0
	       DEPTHWISE_CONV_2D	       15	   628.139	    12.586%	    95.202%	     0.000	        0
	              HARD_SWISH	       15	    90.448	     1.812%	    97.014%	     0.000	        0
	                     MUL	       32	    29.393	     0.589%	    97.603%	     0.000	        0
	                 ARG_MAX	        1	    22.866	     0.458%	    98.061%	     0.000	        0
	                     ADD	       25	    22.860	     0.458%	    98.519%	     0.000	        0
	         RESIZE_BILINEAR	        5	    22.494	     0.451%	    98.970%	     0.000	        0
	       SPACE_TO_BATCH_ND	        8	    18.518	     0.371%	    99.341%	     0.000	        0
	       BATCH_TO_SPACE_ND	        8	    15.522	     0.311%	    99.652%	     0.000	        0
	         AVERAGE_POOL_2D	        9	     7.855	     0.157%	    99.809%	     0.000	        0
	                     SUB	        2	     5.896	     0.118%	    99.928%	     0.000	        0
	                 RESHAPE	        6	     2.133	     0.043%	    99.970%	     0.000	        0
	                     PAD	        1	     0.631	     0.013%	    99.983%	     0.000	        0
	                    CAST	        2	     0.575	     0.012%	    99.994%	     0.000	        0
	           STRIDED_SLICE	        1	     0.260	     0.005%	   100.000%	     0.000	        0
	        Misc Runtime Ops	        1	     0.012	     0.000%	   100.000%	    38.304	        0
	              DEQUANTIZE	       12	     0.003	     0.000%	   100.000%	     0.000	        0

Timings (microseconds): count=31 first=193 curr=5276579 min=193 max=5454605 avg=4.99104e+06 std=1311782
Memory (bytes): count=0
194 nodes observed
Ubuntu 19.10 aarch64 + RaspberryPi4 deeplab_v3_plus_mnv3_decoder_256_integer_quant.tflite Benchmark
Number of nodes executed: 180
============================== Summary by node type ==============================
	             [Node type]	  [count]	  [avg ms]	    [avg %]	    [cdf %]	  [mem KB]	[times called]
	                 CONV_2D	       38	    37.595	    45.330%	    45.330%	     0.000	       38
	                     ADD	       37	    12.319	    14.854%	    60.184%	     0.000	       37
	       DEPTHWISE_CONV_2D	       17	    11.424	    13.774%	    73.958%	     0.000	       17
	         RESIZE_BILINEAR	        4	     7.336	     8.845%	    82.804%	     0.000	        4
	                     MUL	        9	     4.204	     5.069%	    87.873%	     0.000	        9
	                QUANTIZE	       13	     3.976	     4.794%	    92.667%	     0.000	       13
	         AVERAGE_POOL_2D	        9	     1.809	     2.181%	    94.848%	     0.000	        9
	                     DIV	        9	     1.167	     1.407%	    96.255%	     0.000	        9
	                 ARG_MAX	        1	     1.137	     1.371%	    97.626%	     0.000	        1
	           CONCATENATION	        2	     0.780	     0.940%	    98.566%	     0.000	        2
	         FULLY_CONNECTED	       16	     0.715	     0.862%	    99.428%	     0.000	       16
	              DEQUANTIZE	        9	     0.473	     0.570%	    99.999%	     0.000	        9
	                 RESHAPE	       16	     0.001	     0.001%	   100.000%	     0.000	       16

Timings (microseconds): count=50 first=83065 curr=82874 min=82675 max=85743 avg=83036 std=499
Memory (bytes): count=0
180 nodes observed
Ubuntu 19.10 aarch64 + RaspberryPi4 deeplab_v3_plus_mnv2_decoder_256_integer_quant.tflite Benchmark
Number of nodes executed: 81
============================== Summary by node type ==============================
	             [Node type]	  [count]	  [avg ms]	    [avg %]	    [cdf %]	  [mem KB]	[times called]
	                 CONV_2D	       41	    47.427	    65.530%	    65.530%	     0.000	       41
	       DEPTHWISE_CONV_2D	       19	    11.114	    15.356%	    80.887%	     0.000	       19
	         RESIZE_BILINEAR	        4	     7.342	    10.145%	    91.031%	     0.000	        4
	                QUANTIZE	        3	     2.953	     4.080%	    95.112%	     0.000	        3
	                     ADD	       10	     1.633	     2.256%	    97.368%	     0.000	       10
	                 ARG_MAX	        1	     1.137	     1.571%	    98.939%	     0.000	        1
	           CONCATENATION	        2	     0.736	     1.017%	    99.956%	     0.000	        2
	         AVERAGE_POOL_2D	        1	     0.032	     0.044%	   100.000%	     0.000	        1

Timings (microseconds): count=50 first=72544 curr=72425 min=72157 max=72745 avg=72412.9 std=137
Memory (bytes): count=0
81 nodes observed
Ubuntu 19.10 aarch64 + RaspberryPi4 mobilenet_v3_small_full_integer_quant.tflite Benchmark
Number of nodes executed: 176
============================== Summary by node type ==============================
	             [Node type]	  [count]	  [avg ms]	    [avg %]	    [cdf %]	  [mem KB]	[times called]
	                 CONV_2D	       61	    10.255	    36.582%	    36.582%	     0.000	       61
	       DEPTHWISE_CONV_2D	       27	     5.058	    18.043%	    54.625%	     0.000	       27
	                     MUL	       26	     5.056	    18.036%	    72.661%	     0.000	       26
	                     ADD	       14	     4.424	    15.781%	    88.442%	     0.000	       14
	                QUANTIZE	       13	     1.633	     5.825%	    94.267%	     0.000	       13
	              HARD_SWISH	       10	     0.918	     3.275%	    97.542%	     0.000	       10
	                LOGISTIC	        1	     0.376	     1.341%	    98.883%	     0.000	        1
	         AVERAGE_POOL_2D	        9	     0.199	     0.710%	    99.593%	     0.000	        9
	           CONCATENATION	        2	     0.084	     0.300%	    99.893%	     0.000	        2
	                 RESHAPE	       13	     0.030	     0.107%	   100.000%	     0.000	       13

Timings (microseconds): count=50 first=28827 curr=28176 min=27916 max=28827 avg=28121.2 std=165
Memory (bytes): count=0
176 nodes observed
Ubuntu 19.10 aarch64 + RaspberryPi4 mobilenet_v3_small_weight_quant.tflite Benchmark
Number of nodes executed: 186
============================== Summary by node type ==============================
	             [Node type]	  [count]	  [avg ms]	    [avg %]	    [cdf %]	  [mem KB]	[times called]
	                 CONV_2D	       61	    82.600	    79.265%	    79.265%	     0.000	       61
	       DEPTHWISE_CONV_2D	       27	     8.198	     7.867%	    87.132%	     0.000	       27
	                     MUL	       26	     4.866	     4.670%	    91.802%	     0.000	       26
	                     ADD	       14	     4.863	     4.667%	    96.469%	     0.000	       14
	                LOGISTIC	        1	     1.645	     1.579%	    98.047%	     0.000	        1
	         AVERAGE_POOL_2D	        9	     0.761	     0.730%	    98.777%	     0.000	        9
	              HARD_SWISH	       10	     0.683	     0.655%	    99.433%	     0.000	       10
	           CONCATENATION	        2	     0.415	     0.398%	    99.831%	     0.000	        2
	                 RESHAPE	       13	     0.171	     0.164%	    99.995%	     0.000	       13
	              DEQUANTIZE	       23	     0.005	     0.005%	   100.000%	     0.000	       23

Timings (microseconds): count=50 first=103867 curr=103937 min=103708 max=118926 avg=104299 std=2254
Memory (bytes): count=0
186 nodes observed
Ubuntu 19.10 aarch64 + RaspberryPi4 Posenet model-mobilenet_v1_101_257_integer_quant.tflite Benchmark
Number of nodes executed: 38
============================== Summary by node type ==============================
	             [Node type]	  [count]	  [avg ms]	    [avg %]	    [cdf %]	  [mem KB]	[times called]
	                 CONV_2D	       18	    31.906	    83.360%	    83.360%	     0.000	        0
	       DEPTHWISE_CONV_2D	       13	     5.959	    15.569%	    98.929%	     0.000	        0
	                QUANTIZE	        1	     0.223	     0.583%	    99.511%	     0.000	        0
	        Misc Runtime Ops	        1	     0.148	     0.387%	    99.898%	    96.368	        0
	              DEQUANTIZE	        4	     0.030	     0.078%	    99.976%	     0.000	        0
	                LOGISTIC	        1	     0.009	     0.024%	   100.000%	     0.000	        0

Timings (microseconds): count=70 first=519 curr=53370 min=519 max=53909 avg=38296 std=23892
Memory (bytes): count=0
38 nodes observed
Ubuntu 19.10 aarch64 + RaspberryPi4 MobileNetV2-SSDLite ssdlite_mobilenet_v2_coco_300_integer_quant.tflite Benchmark
Number of nodes executed: 128
============================== Summary by node type ==============================
	             [Node type]	  [count]	  [avg ms]	    [avg %]	    [cdf %]	  [mem KB]	[times called]
	                 CONV_2D	       55	    27.253	    71.185%	    71.185%	     0.000	        0
	       DEPTHWISE_CONV_2D	       33	     8.024	    20.959%	    92.143%	     0.000	        0
	                     ADD	       10	     1.565	     4.088%	    96.231%	     0.000	        0
	                QUANTIZE	       11	     0.546	     1.426%	    97.657%	     0.000	        0
	        Misc Runtime Ops	        1	     0.368	     0.961%	    98.618%	   250.288	        0
	                LOGISTIC	        1	     0.253	     0.661%	    99.279%	     0.000	        0
	              DEQUANTIZE	        2	     0.168	     0.439%	    99.718%	     0.000	        0
	           CONCATENATION	        2	     0.077	     0.201%	    99.919%	     0.000	        0
	                 RESHAPE	       13	     0.031	     0.081%	   100.000%	     0.000	        0

Timings (microseconds): count=70 first=1289 curr=53049 min=1289 max=53590 avg=38345.2 std=23436
Memory (bytes): count=0
128 nodes observed
Ubuntu 19.10 aarch64 + RaspberryPi4 ml-sound-classifier mobilenetv2_fsd2018_41cls_weight_quant.tflite Benchmark
Number of nodes executed: 111
============================== Summary by node type ==============================
	             [Node type]	  [count]	  [avg ms]	    [avg %]	    [cdf %]	  [mem KB]	[times called]
	                 MINIMUM	       35	    10.020	    45.282%	    45.282%	     0.000	       35
	                 CONV_2D	       34	     8.376	    37.852%	    83.134%	     0.000	       34
	       DEPTHWISE_CONV_2D	       18	     1.685	     7.615%	    90.749%	     0.000	       18
	                    MEAN	        1	     1.422	     6.426%	    97.176%	     0.000	        1
	         FULLY_CONNECTED	        2	     0.589	     2.662%	    99.837%	     0.000	        2
	                     ADD	       10	     0.031	     0.140%	    99.977%	     0.000	       10
	                 SOFTMAX	        1	     0.005	     0.023%	   100.000%	     0.000	        1
	              DEQUANTIZE	       10	     0.000	     0.000%	   100.000%	     0.000	       10

Timings (microseconds): count=50 first=22417 curr=22188 min=22041 max=22417 avg=22182 std=70
Memory (bytes): count=0
111 nodes observed
Ubuntu 19.10 aarch64 + RaspberryPi4 ml-sound-classifier mobilenetv2_fsd2018_41cls_integer_quant.tflite Benchmark
Number of nodes executed: 173
============================== Summary by node type ==============================
	             [Node type]	  [count]	  [avg ms]	    [avg %]	    [cdf %]	  [mem KB]	[times called]
	                QUANTIZE	       70	     1.117	    23.281%	    23.281%	     0.000	        0
	                 MINIMUM	       35	     1.104	    23.010%	    46.290%	     0.000	        0
	                 CONV_2D	       34	     0.866	    18.049%	    64.339%	     0.000	        0
	                    MEAN	        1	     0.662	    13.797%	    78.137%	     0.000	        0
	       DEPTHWISE_CONV_2D	       18	     0.476	     9.921%	    88.058%	     0.000	        0
	         FULLY_CONNECTED	        2	     0.251	     5.231%	    93.289%	     0.000	        0
	        Misc Runtime Ops	        1	     0.250	     5.211%	    98.499%	    71.600	        0
	                     ADD	       10	     0.071	     1.480%	    99.979%	     0.000	        0
	                 SOFTMAX	        1	     0.001	     0.021%	   100.000%	     0.000	        0
	              DEQUANTIZE	        1	     0.000	     0.000%	   100.000%	     0.000	        0

Timings (microseconds): count=198 first=477 curr=9759 min=477 max=10847 avg=4876.6 std=4629
Memory (bytes): count=0
173 nodes observed
Raspbian Buster aarch64 + RaspberryPi4 deeplabv3_mnv2_pascal_trainval_257_integer_quant.tflite Benchmark
Number of nodes executed: 82
============================== Summary by node type ==============================
	             [Node type]	  [count]	  [avg ms]	    [avg %]	    [cdf %]	  [mem KB]	[times called]
	                 CONV_2D	       38	   103.576	    56.077%	    56.077%	     0.000	       38
	       DEPTHWISE_CONV_2D	       17	    33.151	    17.948%	    74.026%	     0.000	       17
	         RESIZE_BILINEAR	        3	    15.143	     8.199%	    82.224%	     0.000	        3
	                     SUB	        2	    10.908	     5.906%	    88.130%	     0.000	        2
	                     ADD	       11	     9.821	     5.317%	    93.447%	     0.000	       11
	                 ARG_MAX	        1	     8.824	     4.777%	    98.225%	     0.000	        1
	                     PAD	        1	     1.024	     0.554%	    98.779%	     0.000	        1
	                QUANTIZE	        2	     0.941	     0.509%	    99.289%	     0.000	        2
	                     MUL	        1	     0.542	     0.293%	    99.582%	     0.000	        1
	           CONCATENATION	        1	     0.365	     0.198%	    99.780%	     0.000	        1
	         AVERAGE_POOL_2D	        1	     0.150	     0.081%	    99.861%	     0.000	        1
	                 RESHAPE	        2	     0.129	     0.070%	    99.931%	     0.000	        2
	             EXPAND_DIMS	        2	     0.128	     0.069%	   100.000%	     0.000	        2

Timings (microseconds): count=50 first=201226 curr=176476 min=176476 max=201226 avg=184741 std=4791
Memory (bytes): count=0
82 nodes observed
Ubuntu 18.04 x86_64 + XNNPACK enabled + 10 Threads deeplabv3_257_mv_gpu.tflite Benchmark
Number of nodes executed: 8
============================== Summary by node type ==============================
	             [Node type]	  [count]	  [avg ms]	    [avg %]	    [cdf %]	  [mem KB]	[times called]
	                DELEGATE	        3	     6.716	    61.328%	    61.328%	     0.000	        3
	         RESIZE_BILINEAR	        3	     3.965	    36.207%	    97.534%	     0.000	        3
	           CONCATENATION	        1	     0.184	     1.680%	    99.215%	     0.000	        1
	         AVERAGE_POOL_2D	        1	     0.086	     0.785%	   100.000%	     0.000	        1

Timings (microseconds): count=91 first=11051 curr=10745 min=10521 max=12552 avg=10955.4 std=352
Memory (bytes): count=0
8 nodes observed

Note: as the benchmark tool itself affects memory footprint, the following is only APPROXIMATE to the actual memory footprint of the model at runtime. Take the information at your discretion.
Peak memory footprint (MB): init=3.58203 overall=56.0703
Ubuntu 18.04 x86_64 + XNNPACK disabled + 10 Threads deeplabv3_257_mv_gpu.tflite Benchmark
Number of nodes executed: 70
============================== Summary by node type ==============================
	             [Node type]	  [count]	  [avg ms]	    [avg %]	    [cdf %]	  [mem KB]	[times called]
	       DEPTHWISE_CONV_2D	       17	    41.704	    68.372%	    68.372%	     0.000	       17
	                 CONV_2D	       38	    15.932	    26.120%	    94.491%	     0.000	       38
	         RESIZE_BILINEAR	        3	     3.060	     5.017%	    99.508%	     0.000	        3
	                     ADD	       10	     0.149	     0.244%	    99.752%	     0.000	       10
	           CONCATENATION	        1	     0.109	     0.179%	    99.931%	     0.000	        1
	         AVERAGE_POOL_2D	        1	     0.042	     0.069%	   100.000%	     0.000	        1

Timings (microseconds): count=50 first=59929 curr=60534 min=59374 max=63695 avg=61031.6 std=1182
Memory (bytes): count=0
70 nodes observed

Note: as the benchmark tool itself affects memory footprint, the following is only APPROXIMATE to the actual memory footprint of the model at runtime. Take the information at your discretion.
Peak memory footprint (MB): init=0 overall=13.7109
Ubuntu 18.04 x86_64 + XNNPACK enabled + 4 Threads Faster-Grad-CAM weights_weight_quant.tflite Benchmark
umber of nodes executed: 74
============================== Summary by node type ==============================
	             [Node type]	  [count]	  [avg ms]	    [avg %]	    [cdf %]	  [mem KB]	[times called]
	                 CONV_2D	       31	     4.947	    77.588%	    77.588%	     0.000	       31
	                DELEGATE	       17	     0.689	    10.806%	    88.394%	     0.000	       17
	       DEPTHWISE_CONV_2D	       10	     0.591	     9.269%	    97.663%	     0.000	       10
	                    MEAN	        1	     0.110	     1.725%	    99.388%	     0.000	        1
	                     PAD	        5	     0.039	     0.612%	   100.000%	     0.000	        5
	              DEQUANTIZE	       10	     0.000	     0.000%	   100.000%	     0.000	       10

Timings (microseconds): count=155 first=6415 curr=6443 min=6105 max=6863 avg=6409.22 std=69
Memory (bytes): count=0
74 nodes observed
RaspberryPi4 + Ubuntu 19.10 aarch64 + 4 Threads Faster-Grad-CAM weights_integer_quant.tflite Benchmark
Number of nodes executed: 72
============================== Summary by node type ==============================
	             [Node type]	  [count]	  [avg ms]	    [avg %]	    [cdf %]	  [mem KB]	[times called]
	                 CONV_2D	       35	     0.753	    34.958%	    34.958%	     0.000	        0
	                     PAD	        5	     0.395	    18.338%	    53.296%	     0.000	        0
	                    MEAN	        1	     0.392	    18.199%	    71.495%	     0.000	        0
	        Misc Runtime Ops	        1	     0.282	    13.092%	    84.587%	    89.232	        0
	       DEPTHWISE_CONV_2D	       17	     0.251	    11.653%	    96.240%	     0.000	        0
	                     ADD	       10	     0.054	     2.507%	    98.747%	     0.000	        0
	                QUANTIZE	        1	     0.024	     1.114%	    99.861%	     0.000	        0
	              DEQUANTIZE	        2	     0.003	     0.139%	   100.000%	     0.000	        0

Timings (microseconds): count=472 first=564 curr=3809 min=564 max=3950 avg=2188.51 std=1625
Memory (bytes): count=0
72 nodes observed
Ubuntu 18.04 x86_64 + XNNPACK enabled + 4 Threads EfficientNet-lite efficientnet-lite0-fp32.tflite Benchmark
Number of nodes executed: 5
============================== Summary by node type ==============================
	             [Node type]	  [count]	  [avg ms]	    [avg %]	    [cdf %]	  [mem KB]	[times called]
	                DELEGATE	        2	     5.639	    95.706%	    95.706%	     0.000	        2
	         FULLY_CONNECTED	        1	     0.239	     4.056%	    99.762%	     0.000	        1
	         AVERAGE_POOL_2D	        1	     0.014	     0.238%	   100.000%	     0.000	        1
	                 RESHAPE	        1	     0.000	     0.000%	   100.000%	     0.000	        1

Timings (microseconds): count=168 first=5842 curr=5910 min=5749 max=6317 avg=5894.55 std=100
Memory (bytes): count=0
5 nodes observed
Ubuntu 18.04 x86_64 + XNNPACK enabled + 4 Threads EfficientNet-lite efficientnet-lite4-fp32.tflite Benchmark
Number of nodes executed: 5
============================== Summary by node type ==============================
	             [Node type]	  [count]	  [avg ms]	    [avg %]	    [cdf %]	  [mem KB]	[times called]
	                DELEGATE	        2	    33.720	    99.235%	    99.235%	     0.000	        2
	         FULLY_CONNECTED	        1	     0.231	     0.680%	    99.915%	     0.000	        1
	         AVERAGE_POOL_2D	        1	     0.029	     0.085%	   100.000%	     0.000	        1
	                 RESHAPE	        1	     0.000	     0.000%	   100.000%	     0.000	        1

Timings (microseconds): count=50 first=32459 curr=34867 min=31328 max=35730 avg=33983.5 std=1426
Memory (bytes): count=0
5 nodes observed
Ubuntu 18.04 x86_64 + XNNPACK enabled + 4 Threads White-box-Cartoonization white_box_cartoonization_weight_quant.tflite Benchmark
Number of nodes executed: 47
============================== Summary by node type ==============================
	             [Node type]	  [count]	  [avg ms]	    [avg %]	    [cdf %]	  [mem KB]	[times called]
	                 CONV_2D	       18	 10731.842	    97.293%	    97.293%	     0.000	       18
	              LEAKY_RELU	       13	   236.792	     2.147%	    99.440%	     0.000	       13
	   TfLiteXNNPackDelegate	       10	    45.534	     0.413%	    99.853%	     0.000	       10
	         RESIZE_BILINEAR	        2	    11.237	     0.102%	    99.954%	     0.000	        2
	                     SUB	        3	     4.053	     0.037%	    99.991%	     0.000	        3
	                     DIV	        1	     0.977	     0.009%	   100.000%	     0.000	        1

Timings (microseconds): count=14 first=10866837 curr=11292015 min=10697744 max=12289882 avg=1.10305e+07 std=406791
Memory (bytes): count=0
47 nodes observed
RaspberryPi4 + Ubuntu 19.10 aarch64 + 4 Threads edgetpu_deeplab_257_os16_integer_quant.tflite Benchmark
Number of nodes executed: 91
============================== Summary by node type ==============================
	             [Node type]	  [count]	  [avg ms]	    [avg %]	    [cdf %]	  [mem KB]	[times called]
	                 CONV_2D	       49	    54.679	    58.810%	    58.810%	     0.000	       49
	                     SUB	        2	    11.043	    11.877%	    70.687%	     0.000	        2
	                     ADD	       16	     8.909	     9.582%	    80.269%	     0.000	       16
	                 ARG_MAX	        1	     7.184	     7.727%	    87.996%	     0.000	        1
	         RESIZE_BILINEAR	        3	     6.654	     7.157%	    95.153%	     0.000	        3
	       DEPTHWISE_CONV_2D	       13	     3.409	     3.667%	    98.819%	     0.000	       13
	                     MUL	        1	     0.548	     0.589%	    99.408%	     0.000	        1
	                QUANTIZE	        2	     0.328	     0.353%	    99.761%	     0.000	        2
	                 RESHAPE	        2	     0.162	     0.174%	    99.935%	     0.000	        2
	         AVERAGE_POOL_2D	        1	     0.043	     0.046%	    99.982%	     0.000	        1
	           CONCATENATION	        1	     0.017	     0.018%	   100.000%	     0.000	        1

Timings (microseconds): count=50 first=92752 curr=93058 min=92533 max=94478 avg=93021.2 std=274
Memory (bytes): count=0
91 nodes observed
RaspberryPi4 + Ubuntu 19.10 aarch64 + 4 Threads edgetpu_deeplab_257_os32_integer_quant.tflite Benchmark
Number of nodes executed: 91
============================== Summary by node type ==============================
	             [Node type]	  [count]	  [avg ms]	    [avg %]	    [cdf %]	  [mem KB]	[times called]
	                 CONV_2D	       49	    39.890	    52.335%	    52.335%	     0.000	       49
	                     SUB	        2	    11.043	    14.488%	    66.823%	     0.000	        2
	                     ADD	       16	     8.064	    10.580%	    77.403%	     0.000	       16
	                 ARG_MAX	        1	     7.011	     9.198%	    86.601%	     0.000	        1
	         RESIZE_BILINEAR	        3	     6.623	     8.689%	    95.290%	     0.000	        3
	       DEPTHWISE_CONV_2D	       13	     2.503	     3.284%	    98.574%	     0.000	       13
	                     MUL	        1	     0.544	     0.714%	    99.288%	     0.000	        1
	                QUANTIZE	        2	     0.313	     0.411%	    99.698%	     0.000	        2
	                 RESHAPE	        2	     0.178	     0.234%	    99.932%	     0.000	        2
	         AVERAGE_POOL_2D	        1	     0.041	     0.054%	    99.986%	     0.000	        1
	           CONCATENATION	        1	     0.011	     0.014%	   100.000%	     0.000	        1

Timings (microseconds): count=50 first=75517 curr=75558 min=75517 max=97776 avg=76262.5 std=3087
Memory (bytes): count=0
91 nodes observed
RaspberryPi4 + Ubuntu 19.10 aarch64 + 4 Threads human_pose_estimation_3d_0001_256x448_integer_quant.tflite Benchmark
Number of nodes executed: 165
============================== Summary by node type ==============================
	             [Node type]	  [count]	  [avg ms]	    [avg %]	    [cdf %]	  [mem KB]	[times called]
	                 CONV_2D	       69	   343.433	    78.638%	    78.638%	     0.000	       69
	                     PAD	       38	    51.637	    11.824%	    90.462%	     0.000	       38
	       DEPTHWISE_CONV_2D	       14	    15.306	     3.505%	    93.967%	     0.000	       14
	                     ADD	       15	    14.535	     3.328%	    97.295%	     0.000	       15
	                     ELU	        6	     5.071	     1.161%	    98.456%	     0.000	        6
	                QUANTIZE	       11	     4.481	     1.026%	    99.482%	     0.000	       11
	              DEQUANTIZE	        9	     1.851	     0.424%	    99.906%	     0.000	        9
	           CONCATENATION	        3	     0.410	     0.094%	   100.000%	     0.000	        3

Timings (microseconds): count=50 first=425038 curr=423469 min=421348 max=969226 avg=436808 std=77255
Memory (bytes): count=0
165 nodes observed
RaspberryPi4 + Ubuntu 19.10 aarch64 + 4 Threads + BlazeFace face_detection_front_128_integer_quant.tflite Benchmark
Number of nodes executed: 79
============================== Summary by node type ==============================
	             [Node type]	  [count]	  [avg ms]	    [avg %]	    [cdf %]	  [mem KB]	[times called]
	                     ADD	       16	     2.155	    34.120%	    34.120%	     0.000	       16
	                 CONV_2D	       21	     2.017	    31.935%	    66.054%	     0.000	       21
	                     PAD	       11	     1.014	    16.054%	    82.109%	     0.000	       11
	       DEPTHWISE_CONV_2D	       16	     0.765	    12.112%	    94.221%	     0.000	       16
	                QUANTIZE	        4	     0.186	     2.945%	    97.166%	     0.000	        4
	             MAX_POOL_2D	        3	     0.153	     2.422%	    99.588%	     0.000	        3
	              DEQUANTIZE	        2	     0.017	     0.269%	    99.857%	     0.000	        2
	           CONCATENATION	        2	     0.006	     0.095%	    99.952%	     0.000	        2
	                 RESHAPE	        4	     0.003	     0.047%	   100.000%	     0.000	        4

Timings (microseconds): count=144 first=6415 curr=6319 min=6245 max=6826 avg=6359.12 std=69
Memory (bytes): count=0
79 nodes observed
RaspberryPi4 + Ubuntu 19.10 aarch64 + 4 Threads + ssd_mobilenet_v2_mnasfpn_shared_box_predictor_320_coco_integer_quant.tflite Benchmark
Number of nodes executed: 588
============================== Summary by node type ==============================
	             [Node type]	  [count]	  [avg ms]	    [avg %]	    [cdf %]	  [mem KB]	[times called]
	                 CONV_2D	      119	   109.253	    52.671%	    52.671%	     0.000	      119
	       DEPTHWISE_CONV_2D	       61	    33.838	    16.313%	    68.984%	     0.000	       61
	TFLite_Detection_PostProcess	        1	    22.711	    10.949%	    79.933%	     0.000	        1
	                LOGISTIC	        1	    17.696	     8.531%	    88.465%	     0.000	        1
	                     ADD	       59	    12.300	     5.930%	    94.395%	     0.000	       59
	                 RESHAPE	        8	     4.175	     2.013%	    96.407%	     0.000	        8
	           CONCATENATION	        2	     3.416	     1.647%	    98.054%	     0.000	        2
	 RESIZE_NEAREST_NEIGHBOR	       12	     1.873	     0.903%	    98.957%	     0.000	       12
	             MAX_POOL_2D	       13	     1.363	     0.657%	    99.614%	     0.000	       13
	                     MUL	       16	     0.737	     0.355%	    99.970%	     0.000	       16
	              DEQUANTIZE	      296	     0.063	     0.030%	   100.000%	     0.000	      296

Timings (microseconds): count=50 first=346007 curr=196005 min=192539 max=715157 avg=207709 std=75605
Memory (bytes): count=0
588 nodes observed
RaspberryPi4 + Ubuntu 19.10 aarch64 + 4 Threads + object_detection_3d_chair_640x480_integer_quant.tflite Benchmark
Number of nodes executed: 126
============================== Summary by node type ==============================
	             [Node type]	  [count]	  [avg ms]	    [avg %]	    [cdf %]	  [mem KB]	[times called]
	                 CONV_2D	       60	   146.537	    63.805%	    63.805%	     0.000	       60
	       DEPTHWISE_CONV_2D	       26	    45.022	    19.604%	    83.409%	     0.000	       26
	                     ADD	       23	    23.393	    10.186%	    93.595%	     0.000	       23
	          TRANSPOSE_CONV	        3	     9.930	     4.324%	    97.918%	     0.000	        3
	                QUANTIZE	        5	     3.103	     1.351%	    99.269%	     0.000	        5
	           CONCATENATION	        4	     1.541	     0.671%	    99.940%	     0.000	        4
	              DEQUANTIZE	        3	     0.117	     0.051%	    99.991%	     0.000	        3
	                     EXP	        1	     0.018	     0.008%	    99.999%	     0.000	        1
	                     NEG	        1	     0.002	     0.001%	   100.000%	     0.000	        1

Timings (microseconds): count=50 first=218224 curr=217773 min=217174 max=649357 avg=229732 std=62952
Memory (bytes): count=0
126 nodes observed
RaspberryPi4 + Ubuntu 19.10 aarch64 + 4 Threads + ssdlite_mobiledet_cpu_320x320_coco_integer_quant.tflite Benchmark
Number of nodes executed: 288
============================== Summary by node type ==============================
	             [Node type]	  [count]	  [avg ms]	    [avg %]	    [cdf %]	  [mem KB]	[times called]
	                 CONV_2D	       96	    22.996	    33.342%	    33.342%	     0.000	       96
	              HARD_SWISH	       57	    11.452	    16.604%	    49.946%	     0.000	       57
	                     MUL	       19	     9.423	    13.662%	    63.608%	     0.000	       19
	         AVERAGE_POOL_2D	       19	     8.439	    12.236%	    75.843%	     0.000	       19
	       DEPTHWISE_CONV_2D	       35	     7.810	    11.324%	    87.167%	     0.000	       35
	TFLite_Detection_PostProcess	        1	     5.650	     8.192%	    95.359%	     0.000	        1
	                     ADD	       12	     1.690	     2.450%	    97.809%	     0.000	       12
	                QUANTIZE	       12	     0.879	     1.274%	    99.084%	     0.000	       12
	                LOGISTIC	       20	     0.277	     0.402%	    99.485%	     0.000	       20
	              DEQUANTIZE	        2	     0.234	     0.339%	    99.825%	     0.000	        2
	           CONCATENATION	        2	     0.079	     0.115%	    99.939%	     0.000	        2
	                 RESHAPE	       13	     0.042	     0.061%	   100.000%	     0.000	       13

Timings (microseconds): count=50 first=69091 curr=68590 min=68478 max=83971 avg=69105.3 std=2147
Memory (bytes): count=0
288 nodes observed
RaspberryPi4 + Ubuntu 19.10 aarch64 + 4 Threads + mobilenet_v2_pose_256_256_dm100_integer_quant.tflite Benchmark
Number of nodes executed: 189
============================== Summary by node type ==============================
	             [Node type]	  [count]	  [avg ms]	    [avg %]	    [cdf %]	  [mem KB]	[times called]
	                 CONV_2D	       86	    51.819	    70.575%	    70.575%	     0.000	       86
	       DEPTHWISE_CONV_2D	       73	    18.207	    24.797%	    95.372%	     0.000	       73
	                     ADD	        8	     1.243	     1.693%	    97.065%	     0.000	        8
	                QUANTIZE	       13	     1.132	     1.542%	    98.607%	     0.000	       13
	           CONCATENATION	        7	     0.607	     0.827%	    99.433%	     0.000	        7
	         RESIZE_BILINEAR	        1	     0.354	     0.482%	    99.916%	     0.000	        1
	              DEQUANTIZE	        1	     0.062	     0.084%	   100.000%	     0.000	        1

Timings (microseconds): count=50 first=73752 curr=73430 min=73191 max=75764 avg=73524.8 std=485
Memory (bytes): count=0
189 nodes observed
RaspberryPi4 + Ubuntu 19.10 aarch64 + 4 Threads + mobilenet_v2_pose_368_432_dm100_integer_quant.tflite Benchmark
Number of nodes executed: 189
============================== Summary by node type ==============================
	             [Node type]	  [count]	  [avg ms]	    [avg %]	    [cdf %]	  [mem KB]	[times called]
	                 CONV_2D	       86	   141.296	    69.289%	    69.289%	     0.000	       86
	       DEPTHWISE_CONV_2D	       73	    53.244	    26.110%	    95.399%	     0.000	       73
	                QUANTIZE	       13	     3.059	     1.500%	    96.899%	     0.000	       13
	                     ADD	        8	     3.014	     1.478%	    98.377%	     0.000	        8
	           CONCATENATION	        7	     2.302	     1.129%	    99.506%	     0.000	        7
	         RESIZE_BILINEAR	        1	     0.852	     0.418%	    99.924%	     0.000	        1
	              DEQUANTIZE	        1	     0.155	     0.076%	   100.000%	     0.000	        1

Timings (microseconds): count=50 first=189613 curr=579873 min=189125 max=579873 avg=204021 std=70304
Memory (bytes): count=0
189 nodes observed
RaspberryPi4 + Ubuntu 19.10 aarch64 + 4 Threads + mobilenet_v2_pose_256_256_dm050_integer_quant.tflite Benchmark
Number of nodes executed: 189
============================== Summary by node type ==============================
	             [Node type]	  [count]	  [avg ms]	    [avg %]	    [cdf %]	  [mem KB]	[times called]
	                 CONV_2D	       86	    40.952	    71.786%	    71.786%	     0.000	       86
	       DEPTHWISE_CONV_2D	       73	    13.508	    23.679%	    95.465%	     0.000	       73
	                QUANTIZE	       13	     1.123	     1.969%	    97.434%	     0.000	       13
	                     ADD	        8	     0.710	     1.245%	    98.678%	     0.000	        8
	           CONCATENATION	        7	     0.498	     0.873%	    99.551%	     0.000	        7
	         RESIZE_BILINEAR	        1	     0.193	     0.338%	    99.890%	     0.000	        1
	              DEQUANTIZE	        1	     0.063	     0.110%	   100.000%	     0.000	        1

Timings (microseconds): count=50 first=57027 curr=57048 min=56773 max=58042 avg=57135 std=229
Memory (bytes): count=0
189 nodes observed
RaspberryPi4 + Ubuntu 19.10 aarch64 + 4 Threads + mobilenet_v2_pose_368_432_dm050_integer_quant.tflite Benchmark
Number of nodes executed: 189
============================== Summary by node type ==============================
	             [Node type]	  [count]	  [avg ms]	    [avg %]	    [cdf %]	  [mem KB]	[times called]
	                 CONV_2D	       86	   104.618	    71.523%	    71.523%	     0.000	       86
	       DEPTHWISE_CONV_2D	       73	    34.527	    23.605%	    95.128%	     0.000	       73
	                QUANTIZE	       13	     2.572	     1.758%	    96.886%	     0.000	       13
	           CONCATENATION	        7	     2.257	     1.543%	    98.429%	     0.000	        7
	                     ADD	        8	     1.683	     1.151%	    99.580%	     0.000	        8
	         RESIZE_BILINEAR	        1	     0.460	     0.314%	    99.894%	     0.000	        1
	              DEQUANTIZE	        1	     0.155	     0.106%	   100.000%	     0.000	        1

Timings (microseconds): count=50 first=172545 curr=146065 min=145260 max=172545 avg=146362 std=3756
Memory (bytes): count=0
189 nodes observed
RaspberryPi4 + Ubuntu 19.10 aarch64 + 4 Threads + yolov4_tiny_voc_416x416_integer_quant.tflite Benchmark
Number of nodes executed: 71
============================== Summary by node type ==============================
	             [Node type]	  [count]	  [avg ms]	    [avg %]	    [cdf %]	  [mem KB]	[times called]
	                 CONV_2D	       21	   149.092	    61.232%	    61.232%	     0.000	       21
	              LEAKY_RELU	       19	    77.644	    31.888%	    93.121%	     0.000	       19
	                     PAD	        2	     8.036	     3.300%	    96.421%	     0.000	        2
	                QUANTIZE	       10	     4.580	     1.881%	    98.302%	     0.000	       10
	           CONCATENATION	        7	     2.415	     0.992%	    99.294%	     0.000	        7
	             MAX_POOL_2D	        3	     0.982	     0.403%	    99.697%	     0.000	        3
	                   SPLIT	        3	     0.615	     0.253%	    99.950%	     0.000	        3
	              DEQUANTIZE	        2	     0.082	     0.034%	    99.984%	     0.000	        2
	 RESIZE_NEAREST_NEIGHBOR	        1	     0.032	     0.013%	    99.997%	     0.000	        1
	           STRIDED_SLICE	        1	     0.004	     0.002%	    99.998%	     0.000	        1
	                     MUL	        1	     0.004	     0.002%	   100.000%	     0.000	        1
	                   SHAPE	        1	     0.000	     0.000%	   100.000%	     0.000	        1

Timings (microseconds): count=50 first=233307 curr=233318 min=232446 max=364068 avg=243522 std=33354
Memory (bytes): count=0
71 nodes observed

4. Reference articles

  1. [deeplab] what's the parameters of the mobilenetv3 pretrained model?
  2. When you want to fine-tune DeepLab on other datasets, there are a few cases
  3. [deeplab] Training deeplab model with ADE20K dataset
  4. Running DeepLab on PASCAL VOC 2012 Semantic Segmentation Dataset
  5. Quantize DeepLab model for faster on-device inference
  6. https://github.com/tensorflow/models/blob/master/research/deeplab/g3doc/model_zoo.md
  7. https://github.com/tensorflow/models/blob/master/research/deeplab/g3doc/quantize.md
  8. the quantized form of Shape operation is not yet implemented
  9. Post-training quantization
  10. Converter command line reference
  11. Quantization-aware training
  12. Converting a .pb file to .meta in TF 1.3
  13. Minimal code to load a trained TensorFlow model from a checkpoint and export it with SavedModelBuilder
  14. How to restore Tensorflow model from .pb file in python?
  15. Error with tag-sets when serving model using tensorflow_model_server tool
  16. ValueError: No 'serving_default' in the SavedModel's SignatureDefs. Possible values are 'name_of_my_model'
  17. kerasのモデルをデプロイする手順 - Signature作成方法解説
  18. TensorFlow で学習したモデルのグラフを tf.train.import_meta_graph でロードする
  19. Tensorflowのグラフ操作 Part1
  20. Configure input_map when importing a tensorflow model from metagraph file
  21. TFLite Model Benchmark Tool
  22. How to install Ubuntu 19.10 aarch64 (64bit) on RaspberryPi4
  23. https://github.com/rwightman/posenet-python.git
  24. https://github.com/sayakpaul/Adventures-in-TensorFlow-Lite.git
Comments
  • Movenet: error on loading model with Openvino

    Movenet: error on loading model with Openvino

    1. OS Ubuntu 18.04

    2. OS Architecture x86_64

    3. Version of OpenVINO 2021.3.394

    9. Movenet from your model zoo

    Ha ha it's me again ;-) I saw you have already converted Movenet ! Naturally I wanted to give it a try. I get this error message when loading the 'lightning' (or 'thunder') model:

    openvino@ubuntu:/workdir$ python3 MovenetOpenvino.py -m lightning
    Video FPS: 30
    Loading Inference Engine
    Device info:
            CPU
            MKLDNNPlugin version ......... 2.1
            Build ........... 2021.3.0-2787-60059f2c755-releases/2021/3
    Pose Detection model - Reading network files:
            /workdir/models/movenet_lightning_FP32.xml
            /workdir/models/movenet_lightning_FP32.bin
    Traceback (most recent call last):
      File "MovenetOpenvino.py", line 569, in <module>
        output=args.output)
      File "MovenetOpenvino.py", line 99, in __init__
        self.load_model(xml, device)
      File "MovenetOpenvino.py", line 131, in load_model
        self.pd_net = self.ie.read_network(model=xml_path, weights=bin_path)
      File "ie_api.pyx", line 293, in openvino.inference_engine.ie_api.IECore.read_network
      File "ie_api.pyx", line 315, in openvino.inference_engine.ie_api.IECore.read_network
    RuntimeError: Check 'element::Type::merge(inputs_et, inputs_et, get_input_element_type(i))' failed at core/src/op/concat.cpp:62:
    While validating node 'v0::Concat Concat_1866 (stack_2_StatefulPartitionedCall/stack_2_1/Unsqueeze/Output_0/Data__const[0]:i32{1,1}, stack_2_StatefulPartitionedCall/stack_2_1/Unsqueeze503[0]:i64{1,1}, stack_2_StatefulPartitionedCall/stack_2_1/Unsqueeze505[0]:i64{1,1}) -> ()' with friendly_name 'Concat_1866':
    Argument element types are inconsistent.
    
    
    opened by geaxgx 88
  •  Is there an easy way to convert ONNX or PB from (NCHW) to (NHWC)?

    Is there an easy way to convert ONNX or PB from (NCHW) to (NHWC)?

    @PINTO0309 Hi, Nice work with YOLOv4 / tiny!

    As I see you use:

    • NCHW for: OpenVINO (xml / bin), Darknet (cfg / weights)

    • NHWC for: TFLite, Keras (yolov4_tiny_voc.json / yolov4_tiny_voc.h5), TF1 (pb), TF2 (saved_models.json / saved_models.pb)

    I have several questions:

    • Is there an easy way to convert ONNX or PB from (NCHW) to (NHWC)? I've seen converters that add transpose before and after each layer, but this seems to slow things down a lot. Is it possible to do this transformation without slowing down the inference?

    • Is there an easy way to convert TF1-pb to TF2-saved_models.pb ?

    • Is NHWC slowing down execution on the GPU?

    • How many FPS do you get on Google Coral TPU-Edge and RaspberryPi4 for yolov4-tiny (int8)?

    • What script did you use to get yolov4_tiny_voc.json ?

    opened by AlexeyAB 54
  • Model conversion error

    Model conversion error

    After running download.sh, I am trying to convert face_detection_front.pb to quantized tflite model. I have tried all scripts inside 30_BlazeFace/01_float32 directory but all gets failed with the following error:

    ValueError: This converter can only convert a single ConcreteFunction. Converting multiple functions is under development.

    I am using TensorFlow 2.2.0 on MacOS. Also, tried with a Linux machine with 2.1.0 and 2.2.0.

    NOTE: I am trying to rebuild quantized model to run on microcontroller. The quantized model for the blazeface provided in your repo download.sh has error while running on the microcontroller: "Didn't find op for builtin opcode 'CONV_2D' version '3'" or "Didn't find op for builtin opcode 'QUANTIZE' version '2'".

    enhancement 
    opened by metanav 33
  • Request: FaceMesh-with-Attention model conversion (unsupported custom ops)

    Request: FaceMesh-with-Attention model conversion (unsupported custom ops)

    MediaPipe has released a new FaceMesh-with-Attention model

    That's basically an old FaceMesh model augmented with 3 additional new attention models that refine results, all inside single TFlite model:

    I've tried converting it:

    tflite2tensorflow --model_path face_landmark_with_attention.tflite --flatc_path ./flatc --schema_path schema.fbs --output_pb
    

    but it fails with

    RuntimeError: Encountered unresolved custom op: Landmarks2TransformMatrix.Node number 192 (Landmarks2TransformMatrix) failed to prepare.
    

    It seems that TFLite model is using custom ops to link different execution paths inside it - that is beyond me...

    opened by vladmandic 31
  • posenet versions and resnet.

    posenet versions and resnet.

    Hi, first of all, thanks for this great repository of tensor flow models!, I am learning Tensor Flow and it's very useful!.

    I am trying to use the models from Posenet, and the results I am getting don't look very good compared to what we can see in the online posenet demo.

    Also, in the TensorFlow.JS repository, they say they're using the new PoseNet 2.0, which is only available for TensorFlow.JS.... and it comes in two modes: MobileNet and ResNet.

    My questions are:

    The Posenet models available in your repository, are based on the old Posenet models? or are they based on the new Posenet 2.0 advertised by tensorflow.js?

    Could it be possible for you to include the new Posenet ResNet models in your repository?

    Thanks in advance!

    research 
    opened by vpenades 20
  • 033_Hand_Detection_and_Tracking : handedness ?

    033_Hand_Detection_and_Tracking : handedness ?

    Congratulations and thank you for the great job you are doing !

    I have downloaded the Openvino models from your repository (https://github.com/PINTO0309/PINTO_model_zoo/blob/master/033_Hand_Detection_and_Tracking/07_openvino/download.sh), and I am writing some python code to use them with Openvino. Palm detection model is working fine. But for the hand landmarks model, I don't find the handedness output described in the google paper MediaPipe Hands: On-device Real-time Hand Tracking (https://arxiv.org/pdf/2006.10214.pdf). When I use Netron on the current model from mediapipe repo (https://github.com/google/mediapipe/blob/master/mediapipe/models/hand_landmark.tflite), I can see the output named 'output_handedness', which does not exist in the model from your repo. Is it because google published several versions of this model and you are using an older version ? If yes, do you know if the improvements in the last version are worth using it and if you plan to convert it ?

    Thanks !

    opened by geaxgx 18
  • SCRFD tflite infer error: INT64 != INT32

    SCRFD tflite infer error: INT64 != INT32

    Tanks for your great work!!! when I test SCRFD tflite model which get from https://hub.fastgit.org/PINTO0309/PINTO_model_zoo/blob/main/129_SCRFD/download.sh , it occurs error, the details is as follows:

    1. OS using Ubuntu 18.04

    2. OS Architecture x86_64

    3. Version of TensorFlow 2.6.0

    4. URL of the repository from which the transformed model was taken, https://hub.fastgit.org/PINTO0309/PINTO_model_zoo/tree/main/129_SCRFD, and the model download from https://drive.google.com/uc?export=download&id=1QRmCB2d_5MUcxRD3Zs0iigUuJGybGiKZ。

    5 source code for simple inference testing code

    import numpy as np
    from tensorflow.lite.python.interpreter import Interpreter
    from pprint import pprint
    
    MODEL='scrfd_500m'
    H=240
    W=320
    input_data = np.ones((1,H,W,3), dtype=np.float32)
    
    # tflite ==========================================================
    interpreter = Interpreter(model_path='./model_float32.tflite', num_threads=4)
    interpreter.allocate_tensors()
    input_blob = interpreter.get_input_details()
    output_blob = interpreter.get_output_details()
    interpreter.set_tensor(input_blob[0]['index'], input_data)
    interpreter.invoke()
    output_float32 = interpreter.get_tensor(output_blob[0]['index'])
    print(f'tflite sum output(float32): {np.sum(output_float32)}')
    

    When I run this code, it occurs: Traceback (most recent call last): File "infer_tflite.py", line 15, in interpreter.allocate_tensors() File "/usr/local/lib/python3.6/dist-packages/tensorflow/lite/python/interpreter.py", line 423, in allocate_tensors return self._interpreter.AllocateTensors() RuntimeError: tensorflow/lite/kernels/reduce.cc:223 op_context.axis->type != kTfLiteInt32 (INT64 != INT32)Node number 57 (MEAN) failed to prepare.

    opened by asonee 16
  • BlazePose frozen model

    BlazePose frozen model

    Sorry to bother you again with this... I guess you're already working on converting the BlazePose TFLite models to Frozen model as long as saved model?

    I've successfully ran the TfLite models, and even in desktop and using only CPU the models are very fast and responsive, it's definitely a much better solution than the PoseNet models.

    My understanding is that PoseNet models are better suited for one shot detection, whereas BlazePose is better for continuous tracking, which makes it very good for realtime applications.

    opened by vpenades 15
  • DBFace 480x640 models don't work

    DBFace 480x640 models don't work

    Hi, thanks for your interesting works! I tried to run 041_DBFace/dbface_infer_tflite.py using dbface_keras_480x640_integer_quant_nhwc.tflite and also other 480x640 models, but they created wrong boxes. Other models, like dbface_keras_256x256_integer_quant_nhwc.tflite, detected in a right way. Could you create 480x640 models correctly? And it seems there are bugs in dbface_infer_tflite.py so I also created PR. Please check it out!

    opened by m0ka-Lv98 14
  • EfficientPose conversion

    EfficientPose conversion

    There seems to be something wrong with the EfficientPose models (or at least EfficientPoseII, didnt test the others yet) The output is very small and fuzzy, not at all like the expected keypoint heatmap. For comparison I converted the current model myself with tf2onnx, and that one works as expected.

    moreover, I wanted to compare the .pb from your repo with the original .pb, but wasn't able to load/run your version..

    I see the repo also provides torch models, did you use those for conversion to ONNX? Or was it an older, buggy version? Either way, I think the efficientpose-folder might need an update :)

    Windows10, x86_64 example model EfficientPoseII onnx 1.8.1, onnxruntime 1.5.2 with DirectML

    If you want to test, following pre-processing should do the trick:

    sess = rt.InferenceSession("model_float32_opt.onnx")
    input_name = sess.get_inputs()[0].name
    
    # resize    (test_frame is the result of cv2 videocapture)
    f = min(368 / test_frame.shape[0], 368 / test_frame.shape[1])
    test_frame = cv2.resize(test_frame, (0, 0), fx=f, fy=f)
    
    test_frame = test_frame.astype('float32') / 255  # scale to range 0..1
    test_frame = test_frame[..., ::-1]  # assumes channels last! ->  x = x[::-1, ...]  # BGR->RGB
    
    # pad
    blob = np.zeros((368,368,3), dtype='float32')  # channels last
    d1 = int((368-test_frame.shape[0])/2)
    d2 = int((368-test_frame.shape[1])/2)
    blob[d1:368-d1,d2:368-d2] = test_frame
    
    mean=[0.485, 0.456, 0.406]
    std=[0.229, 0.224, 0.225]
    for i in range(3):
        blob[..., i] -= mean[i]
        blob[..., i] /= std[i]
    
    out = sess.run(None, {input_name: [blob]})
    
    bug 
    opened by Laubeee 14
  • DeepLabv3+ and OpenVINO

    DeepLabv3+ and OpenVINO

    Hi @PINTO0309 ! Thanks for this amazing repo. Congrats!.

    I'm looking for a way to run DeepLabv3+ with OpenVINO. I see that in your repository you have several folders related to this model. It seems that the one that best fits my needs is Sample.4 - Semantic Segmentation, DeeplabV3-plus 256x256. But I would like to do the development simply by calling the executable files and the XML, as in the typical OpenVINO examples. This example requires TensorFlow installation. What about OpenVINO-bin? I did not find a lot of documentation about this latter one. I am using RPi4 and a OAK-D camera.

    Thanks for your help.

    opened by paularamo 14
  • Empty prior bboxes in 227_face-detection-adas-0001

    Empty prior bboxes in 227_face-detection-adas-0001

    Issue Type

    Support

    OS

    Ubuntu

    OS architecture

    x86_64

    Programming Language

    Python

    Framework

    TensorFlowLite

    Model name and Weights/Checkpoints URL

    https://github.com/PINTO0309/PINTO_model_zoo/tree/main/227_face-detection-adas-0001

    Description

    Hi PINTO, I'm trying to run the face detection model followed by the issue response and the sample code. The output bboxes seem ok but the corresponding prior bboxes is always zero (i.e. [0,0,0,0]) which causes zero decoded bbox for the final result. I had checked the prior bboxes loaded from 0.npy where totally number is 10112 and found 10105 zero bboxes. Is this prior bboxes file 0.npy correct for this model?

    Relevant Log Output

    No response

    URL or source code for simple inference testing code

    No response

    opened by PCTsai 2
  • HiFill inpainting conversion to CoreML

    HiFill inpainting conversion to CoreML

    Issue Type

    Support

    OS

    Mac OS

    OS architecture

    aarch64

    Programming Language

    Python

    Framework

    CoreML

    Model name and Weights/Checkpoints URL

    https://github.com/PINTO0309/PINTO_model_zoo/tree/main/100_HiFill

    Description

    I am trying to convert the tensorflow model to coreML to use it on iOS devices.

    Is this code correct for the conversion?

    import tensorflow as tf
    import coremltools as ct
    import numpy as np
    
    path = "hifill.pb"
    
    def wrap_frozen_graph(graph_def, inputs, outputs):
        def _imports_graph_def():
            tf.compat.v1.import_graph_def(graph_def, name="")
        wrapped_import = tf.compat.v1.wrap_function(_imports_graph_def, [])
        import_graph = wrapped_import.graph
        return wrapped_import.prune(
            tf.nest.map_structure(import_graph.as_graph_element, inputs),
            tf.nest.map_structure(import_graph.as_graph_element, outputs))
    
    def tf1_tf2(model_path):
        # path = "/content/sample-imageinpainting-HiFill/GPU_CPU/pb/hifill.pb"
        graph_def = tf.compat.v1.GraphDef()
        loaded = graph_def.ParseFromString(open(model_path, 'rb').read())
        inception_func = wrap_frozen_graph(
            graph_def, inputs=['img:0', 'mask:0'],
            outputs=['inpainted:0', 'attention:0', 'mask_processed:0'])
        return graph_def
    
    imgSize = np.random.rand(1, 512, 512, 3)
    maskSize = np.random.rand(1, 512, 512, 1)
    tf_model = tf1_tf2(path)
    mlmodel = ct.convert(tf_model,
                        source="tensorflow",
                        inputs=[ct.ImageType(name="img"), ct.ImageType(name="mask")])
    
    mlmodel.save("hifill.mlmodel")
    

    I obtain the mlmodel but the output is strange :(

    If it helps I could provide the Xcode project I am running the test on :)

    Any support on this point? image

    Relevant Log Output

    No response

    URL or source code for simple inference testing code

    No response

    opened by DanielZanchi 0
  • 060_hair_segmentation demo?

    060_hair_segmentation demo?

    Issue Type

    Performance

    OS

    Windows

    OS architecture

    x86_64

    Programming Language

    Python

    Framework

    PyTorch

    Model name and Weights/Checkpoints URL

    https://github.com/PINTO0309/PINTO_model_zoo/tree/main/060_hair_segmentation

    Description

    Could you please release the infer code to show the performance of hair segmentation? I have tried to write some but the results seems to be not correct. Thanks a lot.

    Relevant Log Output

    No response

    URL or source code for simple inference testing code

    No response

    opened by OberstWB 0
  • Port GPEN to modelzoo / coreml

    Port GPEN to modelzoo / coreml

    Issue Type

    Feature Request

    OS

    Mac OS

    OS architecture

    aarch64

    Programming Language

    Other

    Framework

    CoreML

    Model name and Weights/Checkpoints URL

    https://github.com/yangxy/GPEN

    Description

    Would be awesome to port GPEN to modelzoo for face restoration in coreml. It provides great results.

    Relevant Log Output

    No response

    URL or source code for simple inference testing code

    No response

    feature_request 
    opened by yousifa 2
  • SwinIR super resolution coreml, tflite conversion

    SwinIR super resolution coreml, tflite conversion

    Issue Type

    Feature Request

    OS

    Other

    OS architecture

    Other

    Programming Language

    Python

    Framework

    PyTorch

    Model name and Weights/Checkpoints URL

    https://github.com/JingyunLiang/SwinIR

    Description

    It would be nice to add the current sota for image super resolution and denoising

    Relevant Log Output

    No response

    URL or source code for simple inference testing code

    No response

    feature_request 
    opened by Vozf 1
Owner
Katsuya Hyodo
Hobby programmer. Intel Software Innovator Program member.
Katsuya Hyodo
WHENet - ONNX, OpenVINO, TFLite, TensorRT, EdgeTPU, CoreML, TFJS, YOLOv4/YOLOv4-tiny-3L

HeadPoseEstimation-WHENet-yolov4-onnx-openvino ONNX, OpenVINO, TFLite, TensorRT, EdgeTPU, CoreML, TFJS, YOLOv4/YOLOv4-tiny-3L 1. Usage $ git clone htt

Katsuya Hyodo 49 Sep 21, 2022
YOLOv5 in PyTorch > ONNX > CoreML > TFLite

This repository represents Ultralytics open-source research into future object detection methods, and incorporates lessons learned and best practices evolved over thousands of hours of training and evolution on anonymized client datasets. All code and models are under active development, and are subject to modification or deletion without notice.

Ultralytics 34.1k Dec 31, 2022
YOLOv3 in PyTorch > ONNX > CoreML > TFLite

This repository represents Ultralytics open-source research into future object detection methods, and incorporates lessons learned and best practices

Ultralytics 9.3k Jan 7, 2023
MMdnn is a set of tools to help users inter-operate among different deep learning frameworks. E.g. model conversion and visualization. Convert models between Caffe, Keras, MXNet, Tensorflow, CNTK, PyTorch Onnx and CoreML.

MMdnn MMdnn is a comprehensive and cross-framework tool to convert, visualize and diagnose deep learning (DL) models. The "MM" stands for model manage

Microsoft 5.7k Jan 9, 2023
tf2onnx - Convert TensorFlow, Keras and Tflite models to ONNX.

tf2onnx converts TensorFlow (tf-1.x or tf-2.x), tf.keras and tflite models to ONNX via command line or python api.

Open Neural Network Exchange 1.8k Jan 8, 2023
Quantized tflite models for ailia TFLite Runtime

ailia-models-tflite Quantized tflite models for ailia TFLite Runtime About ailia TFLite Runtime ailia TF Lite Runtime is a TensorFlow Lite compatible

ax Inc. 13 Dec 23, 2022
Repo for parser tensorflow(.pb) and tflite(.tflite)

tfmodel_parser .pb file is the format of tensorflow model .tflite file is the format of tflite model, which usually used in mobile devices before star

null 1 Dec 23, 2021
Robust Video Matting in PyTorch, TensorFlow, TensorFlow.js, ONNX, CoreML!

Robust Video Matting in PyTorch, TensorFlow, TensorFlow.js, ONNX, CoreML!

Peter Lin 6.5k Jan 4, 2023
Robust Video Matting in PyTorch, TensorFlow, TensorFlow.js, ONNX, CoreML!

Robust Video Matting (RVM) English | 中文 Official repository for the paper Robust High-Resolution Video Matting with Temporal Guidance. RVM is specific

flow-dev 2 Aug 21, 2022
A few stylization coreML models that I've trained with CreateML

CoreML-StyleTransfer A few stylization coreML models that I've trained with CreateML You can open and use the .mlmodel files in the "models" folder in

Doron Adler 8 Aug 18, 2022
KSAI Lite is a deep learning inference framework of kingsoft, based on tensorflow lite

KSAI Lite is a deep learning inference framework of kingsoft, based on tensorflow lite

null 80 Dec 27, 2022
Saeed Lotfi 28 Dec 12, 2022
Convert Pytorch model to onnx or tflite, and the converted model can be visualized by Netron

Convert Pytorch model to onnx or tflite, and the converted model can be visualized by Netron

Roxbili 5 Nov 19, 2022
A high-performance anchor-free YOLO. Exceeding yolov3~v5 with ONNX, TensorRT, NCNN, and Openvino supported.

YOLOX is an anchor-free version of YOLO, with a simpler design but better performance! It aims to bridge the gap between research and industrial communities. For more details, please refer to our report on Arxiv.

null 7.7k Jan 6, 2023
YOLOX is a high-performance anchor-free YOLO, exceeding yolov3~v5 with ONNX, TensorRT, ncnn, and OpenVINO supported.

Introduction YOLOX is an anchor-free version of YOLO, with a simpler design but better performance! It aims to bridge the gap between research and ind

null 7.7k Jan 3, 2023
With this package, you can generate mixed-integer linear programming (MIP) models of trained artificial neural networks (ANNs) using the rectified linear unit (ReLU) activation function

With this package, you can generate mixed-integer linear programming (MIP) models of trained artificial neural networks (ANNs) using the rectified linear unit (ReLU) activation function. At the moment, only TensorFlow sequential models are supported. Interfaces to either the Pyomo or Gurobi modeling environments are offered.

ChemEngAI 40 Dec 27, 2022
DiffQ performs differentiable quantization using pseudo quantization noise. It can automatically tune the number of bits used per weight or group of weights, in order to achieve a given trade-off between model size and accuracy.

Differentiable Model Compression via Pseudo Quantization Noise DiffQ performs differentiable quantization using pseudo quantization noise. It can auto

Facebook Research 145 Dec 30, 2022
Official code for paper "Demystifying Local Vision Transformer: Sparse Connectivity, Weight Sharing, and Dynamic Weight"

Demysitifing Local Vision Transformer, arxiv This is the official PyTorch implementation of our paper. We simply replace local self attention by (dyna

null 138 Dec 28, 2022
Convert weight file.pth to weight file.blob

CONVERT YOUR MODEL TO IR FORMAT INSTALLATION OpenVino Toolkit Download openvinotoolkit 2021.3 version : Link Instruction of installation : Link Pytorc

Tran Anh Tuan 3 Nov 18, 2021