TensorflowLite-bin

Prebuilt binary for TensorflowLite's standalone installer. For RaspberryPi. I provide a FlexDelegate, XNNPACK enabled binary.

Here is the Tensorflow's official README.

If you want the best performance with RaspberryPi4/3, install Ubuntu 18.04+ 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. How to install Ubuntu 19.10 aarch64 (64bit) on RaspberryPi4 - Qiita - PINTO

The full build package for Tensorflow can be found here (Tensorflow-bin).

TensorFlow Lite will continue to have TensorFlow Lite builtin ops optimized for mobile and embedded devices. However, TensorFlow Lite models can now use a subset of TensorFlow ops when TFLite builtin ops are not sufficient. 1. TensorflowLite-flexdelegate (Tensorflow Select Ops) - Github - PINTO0309 2. Select TensorFlow operators to use in TensorFlow Lite

A repository that shares tuning results of trained models generated by Tensorflow. Post-training quantization (Weight Quantization, Integer Quantization, Full Integer Quantization), Quantization-aware training. PINTO_model_zoo - Github - PINTO0309

Reference articles

• My article. Tensorflow Lite v1.14.0 / v1.15.0-rc0 armhf (armv7l) is tuned for MultiThread acceleration and cross-compiled for RaspberryPi on Ubuntu

• Please refer to the following URL for details of performance. Post-training quantization with TF2.0 Keras - nb.o’s Diary. The performance evaluation article was created by @Nextremer_nb_o / Github. Thank you.

• [Japanese ver.] [Tensorflow Lite] Various Neural Network Model quantization methods for Tensorflow Lite (Weight Quantization, Integer Quantization, Full Integer Quantization, Float16 Quantization, EdgeTPU). As of May 05, 2020.

• [English ver.] [Tensorflow Lite] Various Neural Network Model quantization methods for Tensorflow Lite (Weight Quantization, Integer Quantization, Full Integer Quantization, Float16 Quantization, EdgeTPU). As of May 05, 2020.

• A standalone 2MB installer for TensorflowLite v2.4.0-rc4 and a libedgetpu.so.1 build intended to run on a Yocto-generated environment

• Add a custom OP to the TFLite runtime to build the whl installer (for Python), MaxPoolingWithArgmax2D, MaxUnpooling2D, Convolution2DTransposeBias

Python API packages

Usage

$ sudo apt install -y \
swig libjpeg-dev zlib1g-dev python3-dev \
unzip wget python3-pip curl git cmake make
$ sudo pip3 install numpy==1.24.3

$ TFVER=2.12.0

$ PYVER=39
or
$ PYVER=38
or
$ PYVER=310
or
$ PYVER=311

$ ARCH=aarch64
or
$ ARCH=armhf

$ sudo -H pip3 install \
--no-cache-dir \
https://github.com/PINTO0309/TensorflowLite-bin/releases/download/v${TFVER}/tflite_runtime-${TFVER/-/}-cp${PYVER}-none-linux_${ARCH}.whl

Note

Unlike tensorflow this will be installed to a tflite_runtime namespace. You can then use the Tensorflow Lite interpreter as.

from tflite_runtime.interpreter import Interpreter
### Tensorflow v2.2.0
interpreter = Interpreter(model_path="foo.tflite")
### Tensorflow v2.3.0+
interpreter = Interpreter(model_path="foo.tflite", num_threads=4)

Build parameter

• git clone

  TFVER=2.12.0
  git clone -b v${TFVER} --depth 1 https://github.com/tensorflow/tensorflow.git
  cd tensorflow

• Adjustment of Dockerfile and Header file

  sed -i -e 's/RUN yes/#RUN yes/g' tensorflow/lite/tools/pip_package/Dockerfile.py3
  sed -i -e 's/RUN pip3 install numpy~=1.19.2/RUN pip3 install numpy==1.24.3/g' tensorflow/lite/tools/pip_package/Dockerfile.py3
  sed -i '19a ENV DEBIAN_FRONTEND=noninteractive' tensorflow/lite/tools/pip_package/Dockerfile.py3
  sed -i '20a ENV TZ=Etc/UTC' tensorflow/lite/tools/pip_package/Dockerfile.py3
  sed -i '29a \      tzdata \\' tensorflow/lite/tools/pip_package/Dockerfile.py3
  sed -i -e 's/RUN DEBIAN_FRONTEND=noninteractive/#RUN DEBIAN_FRONTEND=noninteractive/g' tensorflow/lite/tools/pip_package/Dockerfile.py3
  sed -i '50a \      python-is-python3 \\' tensorflow/lite/tools/pip_package/Dockerfile.py3
  sed -i '15a #include <assert.h>' tensorflow/tsl/framework/fixedpoint/MatMatProductAVX2.h

• Added FlexDelegate and XNNPACK as build options. If you want to keep the binary size as small as possible, change tflite_pip_with_flex and tflite_with_xnnpack to false and build. The --copt=-fpermissive option of armhf is deprecated.

  nano tensorflow/lite/tools/pip_package/build_pip_package_with_bazel.sh
  
  # Build python interpreter_wrapper.
  cd "${BUILD_DIR}"
  case "${TENSORFLOW_TARGET}" in
    armhf)
      BAZEL_FLAGS="--config=elinux_armhf
        --copt=-march=armv7-a --copt=-mfpu=neon-vfpv4
        --copt=-O3 --copt=-fno-tree-pre --copt=-fpermissive
        --define tensorflow_mkldnn_contraction_kernel=0
        --define=raspberry_pi_with_neon=true"
      ;;
    aarch64)
      BAZEL_FLAGS="--config=elinux_aarch64
        --define tensorflow_mkldnn_contraction_kernel=0
        --copt=-O3"
      ;;
    native)
      BAZEL_FLAGS="--copt=-O3 --copt=-march=native"
      ;;
    *)
      BAZEL_FLAGS="--copt=-O3"
      ;;
  esac
  
   ↓
  
  # Build python interpreter_wrapper.
  cd "${BUILD_DIR}"
  case "${TENSORFLOW_TARGET}" in
    armhf)
      BAZEL_FLAGS="--config=elinux_armhf
        --copt=-march=armv7-a --copt=-mfpu=neon-vfpv4
        --copt=-O3 --copt=-fno-tree-pre
        --define tensorflow_mkldnn_contraction_kernel=0
        --define=raspberry_pi_with_neon=true
        --define=tflite_pip_with_flex=true
        --define=tflite_with_xnnpack=false
        --define=tflite_kernel_use_xnnpack=false
        --ui_actions_shown=20"
      ;;
    aarch64)
      BAZEL_FLAGS="--config=elinux_aarch64
        --define tensorflow_mkldnn_contraction_kernel=0
        --define=tflite_pip_with_flex=true
        --define=tflite_with_xnnpack=true
        --copt=-O3
        --ui_actions_shown=20"
      ;;
    native)
      BAZEL_FLAGS="--copt=-O3 --copt=-march=native
        --define=tflite_pip_with_flex=true
        --define=tflite_with_xnnpack=true
        --ui_actions_shown=20"
      ;;
    *)
      BAZEL_FLAGS="--copt=-O3
        --define=tflite_pip_with_flex=true
        --define=tflite_with_xnnpack=true
        --ui_actions_shown=20"
      ;;
  esac

• Build

  cd tensorflow/lite/tools/pip_package
  
  make BASE_IMAGE=ubuntu:20.04 PYTHON=python3 PYTHON_VERSION=3.8 TENSORFLOW_TARGET=aarch64 docker-build
  make BASE_IMAGE=debian:bullseye PYTHON=python3 PYTHON_VERSION=3.9 TENSORFLOW_TARGET=aarch64 docker-build
  make BASE_IMAGE=ubuntu:22.04 PYTHON=python3 PYTHON_VERSION=3.10 TENSORFLOW_TARGET=aarch64 docker-build
  make BASE_IMAGE=debian:bookworm PYTHON=python3 PYTHON_VERSION=3.11 TENSORFLOW_TARGET=aarch64 docker-build
  
  make BASE_IMAGE=ubuntu:20.04 PYTHON=python3 PYTHON_VERSION=3.8 TENSORFLOW_TARGET=armhf docker-build
  make BASE_IMAGE=debian:bullseye PYTHON=python3 PYTHON_VERSION=3.9 TENSORFLOW_TARGET=armhf docker-build
  make BASE_IMAGE=ubuntu:22.04 PYTHON=python3 PYTHON_VERSION=3.10 TENSORFLOW_TARGET=armhf docker-build
  make BASE_IMAGE=debian:bookworm PYTHON=python3 PYTHON_VERSION=3.11 TENSORFLOW_TARGET=armhf docker-build
  
  make BASE_IMAGE=ubuntu:20.04 PYTHON=python3 PYTHON_VERSION=3.8 TENSORFLOW_TARGET=native docker-build

Operation check 【Classification】

Sample of MultiThread x4 by Tensorflow Lite [MobileNetV1 / 75ms]

Sample of MultiThread x4 by Tensorflow Lite [MobileNetV2 / 68ms]

• Environmental preparation

$ cd ~;mkdir test
$ curl https://raw.githubusercontent.com/tensorflow/tensorflow/master/tensorflow/lite/examples/label_image/testdata/grace_hopper.bmp -o ~/test/grace_hopper.bmp
$ curl https://storage.googleapis.com/download.tensorflow.org/models/mobilenet_v1_1.0_224_frozen.tgz | tar xzv -C ~/test mobilenet_v1_1.0_224/labels.txt
$ mv ~/test/mobilenet_v1_1.0_224/labels.txt ~/test/
$ curl http://download.tensorflow.org/models/mobilenet_v1_2018_02_22/mobilenet_v1_1.0_224_quant.tgz | tar xzv -C ~/test
$ cd ~/test

• label_image.py

import argparse
import numpy as np
import time

from PIL import Image

from tflite_runtime.interpreter import Interpreter

def load_labels(filename):
  my_labels = []
  input_file = open(filename, 'r')
  for l in input_file:
    my_labels.append(l.strip())
  return my_labels
if __name__ == "__main__":
  floating_model = False
  parser = argparse.ArgumentParser()
  parser.add_argument(
    "-i",
    "--image",
    default="/tmp/grace_hopper.bmp", \
    help="image to be classified"
  )
  parser.add_argument(
    "-m",
    "--model_file", \
    default="/tmp/mobilenet_v1_1.0_224_quant.tflite", \
    help=".tflite model to be executed"
  )
  parser.add_argument(
    "-l",
    "--label_file",
    default="/tmp/labels.txt", \
    help="name of file containing labels"
  )
  parser.add_argument(
    "--input_mean",
    default=127.5,
    help="input_mean"
  )
  parser.add_argument(
    "--input_std",
    default=127.5, \
    help="input standard deviation"
  )
  parser.add_argument(
    "--num_threads",
    default=1,
    help="number of threads"
  )
  args = parser.parse_args()

  interpreter = Interpreter(
    model_path="foo.tflite",
    num_threads=args.num_threads
  )
  try:
    interpreter.allocate_tensors()
  except:
    pass
  input_details = interpreter.get_input_details()
  output_details = interpreter.get_output_details()
  # check the type of the input tensor
  if input_details[0]['dtype'] == np.float32:
    floating_model = True
  # NxHxWxC, H:1, W:2
  height = input_details[0]['shape'][1]
  width = input_details[0]['shape'][2]
  img = Image.open(args.image)
  img = img.resize((width, height))
  # add N dim
  input_data = np.expand_dims(img, axis=0)
  if floating_model:
    input_data = (np.float32(input_data) - args.input_mean) / args.input_std

  interpreter.set_tensor(input_details[0]['index'], input_data)

  start_time = time.time()
  interpreter.invoke()
  stop_time = time.time()

  output_data = interpreter.get_tensor(output_details[0]['index'])
  results = np.squeeze(output_data)
  top_k = results.argsort()[-5:][::-1]
  labels = load_labels(args.label_file)
  for i in top_k:
    if floating_model:
      print('{0:08.6f}'.format(float(results[i]))+":", labels[i])
    else:
      print('{0:08.6f}'.format(float(results[i]/255.0))+":", labels[i])

  print("time: ", stop_time - start_time)

• Inference test

$ python3 label_image.py \
--num_threads 4 \
--image grace_hopper.bmp \
--model_file mobilenet_v1_1.0_224_quant.tflite \
--label_file labels.txt

Operation check 【ObjectDetection】

Sample of MultiThread x4 by Tensorflow Lite + Raspbian Buster (armhf) + RaspberryPi3 [MobileNetV2-SSD / 160ms]

Sample of MultiThread x4 by Tensorflow Lite + Ubuntu18.04 (aarch64) + RaspberryPi3 [MobileNetV2-SSD / 140ms]

Inference test

$ python3 mobilenetv2ssd.py

MobileNetV2-SSD (UINT8) + Corei7 CPU only + USB Camera + 10 Threads + Async

MobileNetV2-SSDLite (UINT8) + RaspberryPi4 CPU only + USB Camera 640x480 + 4 Threads + Sync + Disp 1080p

List of quantized models

https://www.tensorflow.org/lite/guide/hosted_models

Other MobileNetV1 weight files

https://github.com/tensorflow/models/blob/master/research/slim/nets/mobilenet_v1.md

Other MobileNetV2 weight files

https://github.com/tensorflow/models/blob/master/research/slim/nets/mobilenet/README.md

Reference

tflite only python package PINTO0309/Tensorflow-bin#15 Incorrect predictions of Mobilenet_V2 tensorflow/tensorflow#31229 (comment)