This repository contains source code demonstrating how to load a pre-trained and converted TensorFlow Lite model and use it to recognize objects in images.
You also need to install Bazel
in order to build this example code. And be sure you have the Python future
module installed:
pip install future --user
Note that, this repository uses tensorflow lite as bazel external dependency.
Build it for desktop machines (tested on Ubuntu and OS X):
bazel build -c opt --cxxopt="-std=c++14" //:label_image
To run unit tests first pull LFS filesystems
git lfs pull
Run unit tests
bazel test -c opt --cxxopt="-std=c++14" //...
You can use any compatible model, but the following MobileNet v1 model offers a good demonstration of a model trained to recognize 1,000 different objects.
# Get model
curl https://storage.googleapis.com/download.tensorflow.org/models/tflite_11_05_08/mobilenet_v2_1.0_224_quant.tgz | tar xzv -C /tmp
bazel-bin/label_image \
--tflite_model /tmp/mobilenet_v2_1.0_224_quant.tflite \
--labels data/labels.txt \
--image data/grace_hopper.bmp
You should see results like this:
Loaded model /tmp/mobilenet_v2_1.0_224_quant.tflite
resolved reporter
invoked
average time: 68.12 ms
0.860174: 653 653:military uniform
0.0481017: 907 907:Windsor tie
0.00786704: 466 466:bulletproof vest
0.00644932: 514 514:cornet, horn, trumpet, trump
0.00608029: 543 543:drumstick
See the lib/include/argument_parser/cli_options.cpp source code for other command line options, or run following command.
bazel-bin/label_image -h
You should see all the information on command line options as below:
label_image
--count, -c: loop interpreter->Invoke() for certain times
--input_mean, -b: input mean
--input_std, -s: input standard deviation
--image, -i: image_name.bmp
--max_profiling_buffer_entries, -e: maximum profiling buffer entries
--labels, -l: labels for the model
--tflite_model, -m: model_name.tflite
--profiling, -p: [0|1], profiling or not
--num_results, -r: number of results to show
--threads, -t: number of threads
--verbose, -v: [0|1] print more information
--result_directory, -d: directory path
--help, -h: print help
Run with docker images.
# Ubuntu 18.04
docker run -ti -v $(pwd):/workspace registry.gitlab.com/jinay1991/tflite_models
# CentOS 7
docker run -ti -v $(pwd):/workspace registry.gitlab.com/jinay1991/tflite_models/centos
Due to libstdc++ dependencies bazel sometimes doesn't link to appropriate library hence it is advisiable to export variables as given below:
export BAZEL_LINKOPTS=-static-libstdc++
export BAZEL_LINKLIBS=-l%:libstdc++.a
Use devtoolset-8 for compiling source with bazel. (Refer: dockerfile/centos.Dockerfile for more details on dependencies on CentOS 7 or refer below section to configure manually.)
Install dependencies...
sudo yum install -y centos-release-scl
sudo yum install -y devtoolset-8
sudo yum install -y git wget
sudo yum install -y libtool clang-format-6.0
sudo yum install -y epel-release
wget https://copr.fedorainfracloud.org/coprs/vbatts/bazel/repo/epel-7/vbatts-bazel-epel-7.repo && sudo mv vbatts-bazel-epel-7.repo /etc/yum.repos.d
sudo yum install -y bazel
Before running/compiling source, run following commands
scl enable devtoolset-8 bash
export BAZEL_LINKOPTS=-static-libstdc++
export BAZEL_LINKLIBS=-l%:libstdc++.a
To enable profiling, provide command line arg -p 1 or --profiling 1.
As a result, following will be received...
$ bazel run //:label_image -- -p 1
Loaded model data/mobilenet_v2_1.0_224_quant.tflite
Resolved reporter
INFO: Initialized TensorFlow Lite runtime.
Built TfLite Interpreter
invoked
average time: 20.885 ms. (i.e. 47.8813 images/second)
============================== Run Order ==============================
[node type] [start] [first] [avg ms] [%] [cdf%] [mem KB] [times called] [Name]
CONV_2D 0.000 1.504 1.504 7.205% 7.205% 0.000 1 [MobilenetV2/Conv/Relu6]
DEPTHWISE_CONV_2D 1.504 0.689 0.689 3.301% 10.506% 0.000 1 [MobilenetV2/expanded_conv/depthwise/Relu6]
CONV_2D 2.194 0.464 0.464 2.223% 12.729% 0.000 1 [MobilenetV2/expanded_conv/project/add_fold]
CONV_2D 2.658 2.111 2.111 10.114% 22.843% 0.000 1 [MobilenetV2/expanded_conv_1/expand/Relu6]
DEPTHWISE_CONV_2D 4.769 0.440 0.440 2.108% 24.951% 0.000 1 [MobilenetV2/expanded_conv_1/depthwise/Relu6]
CONV_2D 5.210 0.175 0.175 0.838% 25.789% 0.000 1 [MobilenetV2/expanded_conv_1/project/add_fold]
CONV_2D 5.386 0.870 0.870 4.168% 29.957% 0.000 1 [MobilenetV2/expanded_conv_2/expand/Relu6]
DEPTHWISE_CONV_2D 6.257 0.826 0.826 3.957% 33.915% 0.000 1 [MobilenetV2/expanded_conv_2/depthwise/Relu6]
CONV_2D 7.083 0.300 0.300 1.437% 35.352% 0.000 1 [MobilenetV2/expanded_conv_2/project/add_fold]
ADD 7.383 0.901 0.901 4.317% 39.668% 0.000 1 [MobilenetV2/expanded_conv_2/add]
CONV_2D 8.284 0.590 0.590 2.827% 42.495% 0.000 1 [MobilenetV2/expanded_conv_3/expand/Relu6]
DEPTHWISE_CONV_2D 8.874 0.145 0.145 0.695% 43.190% 0.000 1 [MobilenetV2/expanded_conv_3/depthwise/Relu6]
CONV_2D 9.019 0.077 0.077 0.369% 43.559% 0.000 1 [MobilenetV2/expanded_conv_3/project/add_fold]
CONV_2D 9.096 0.198 0.198 0.949% 44.507% 0.000 1 [MobilenetV2/expanded_conv_4/expand/Relu6]
DEPTHWISE_CONV_2D 9.294 0.189 0.189 0.905% 45.413% 0.000 1 [MobilenetV2/expanded_conv_4/depthwise/Relu6]
CONV_2D 9.483 0.096 0.096 0.460% 45.873% 0.000 1 [MobilenetV2/expanded_conv_4/project/add_fold]
ADD 9.579 0.293 0.293 1.404% 47.276% 0.000 1 [MobilenetV2/expanded_conv_4/add]
CONV_2D 9.872 0.198 0.198 0.949% 48.225% 0.000 1 [MobilenetV2/expanded_conv_5/expand/Relu6]
DEPTHWISE_CONV_2D 10.070 0.210 0.210 1.006% 49.231% 0.000 1 [MobilenetV2/expanded_conv_5/depthwise/Relu6]
CONV_2D 10.280 0.141 0.141 0.676% 49.907% 0.000 1 [MobilenetV2/expanded_conv_5/project/add_fold]
ADD 10.421 0.393 0.393 1.883% 51.789% 0.000 1 [MobilenetV2/expanded_conv_5/add]
CONV_2D 10.814 0.415 0.415 1.988% 53.778% 0.000 1 [MobilenetV2/expanded_conv_6/expand/Relu6]
DEPTHWISE_CONV_2D 11.230 0.096 0.096 0.460% 54.238% 0.000 1 [MobilenetV2/expanded_conv_6/depthwise/Relu6]
CONV_2D 11.326 0.122 0.122 0.584% 54.822% 0.000 1 [MobilenetV2/expanded_conv_6/project/add_fold]
CONV_2D 11.448 0.222 0.222 1.064% 55.886% 0.000 1 [MobilenetV2/expanded_conv_7/expand/Relu6]
DEPTHWISE_CONV_2D 11.670 0.181 0.181 0.867% 56.753% 0.000 1 [MobilenetV2/expanded_conv_7/depthwise/Relu6]
CONV_2D 11.851 0.214 0.214 1.025% 57.778% 0.000 1 [MobilenetV2/expanded_conv_7/project/add_fold]
ADD 12.065 0.177 0.177 0.848% 58.626% 0.000 1 [MobilenetV2/expanded_conv_7/add]
CONV_2D 12.242 0.238 0.238 1.140% 59.766% 0.000 1 [MobilenetV2/expanded_conv_8/expand/Relu6]
DEPTHWISE_CONV_2D 12.480 0.187 0.187 0.896% 60.662% 0.000 1 [MobilenetV2/expanded_conv_8/depthwise/Relu6]
CONV_2D 12.667 0.186 0.186 0.891% 61.553% 0.000 1 [MobilenetV2/expanded_conv_8/project/add_fold]
ADD 12.853 0.183 0.183 0.877% 62.430% 0.000 1 [MobilenetV2/expanded_conv_8/add]
CONV_2D 13.037 0.162 0.162 0.776% 63.206% 0.000 1 [MobilenetV2/expanded_conv_9/expand/Relu6]
DEPTHWISE_CONV_2D 13.199 0.130 0.130 0.623% 63.829% 0.000 1 [MobilenetV2/expanded_conv_9/depthwise/Relu6]
CONV_2D 13.329 0.130 0.130 0.623% 64.452% 0.000 1 [MobilenetV2/expanded_conv_9/project/add_fold]
ADD 13.459 0.185 0.185 0.886% 65.338% 0.000 1 [MobilenetV2/expanded_conv_9/add]
CONV_2D 13.644 0.174 0.174 0.834% 66.172% 0.000 1 [MobilenetV2/expanded_conv_10/expand/Relu6]
DEPTHWISE_CONV_2D 13.818 0.144 0.144 0.690% 66.861% 0.000 1 [MobilenetV2/expanded_conv_10/depthwise/Relu6]
CONV_2D 13.962 0.266 0.266 1.274% 68.136% 0.000 1 [MobilenetV2/expanded_conv_10/project/add_fold]
CONV_2D 14.228 0.446 0.446 2.137% 70.273% 0.000 1 [MobilenetV2/expanded_conv_11/expand/Relu6]
DEPTHWISE_CONV_2D 14.675 0.312 0.312 1.495% 71.767% 0.000 1 [MobilenetV2/expanded_conv_11/depthwise/Relu6]
CONV_2D 14.987 0.512 0.512 2.453% 74.220% 0.000 1 [MobilenetV2/expanded_conv_11/project/add_fold]
ADD 15.500 0.330 0.330 1.581% 75.801% 0.000 1 [MobilenetV2/expanded_conv_11/add]
CONV_2D 15.831 0.457 0.457 2.189% 77.991% 0.000 1 [MobilenetV2/expanded_conv_12/expand/Relu6]
DEPTHWISE_CONV_2D 16.289 0.290 0.290 1.389% 79.380% 0.000 1 [MobilenetV2/expanded_conv_12/depthwise/Relu6]
CONV_2D 16.579 0.474 0.474 2.271% 81.651% 0.000 1 [MobilenetV2/expanded_conv_12/project/add_fold]
ADD 17.053 0.252 0.252 1.207% 82.858% 0.000 1 [MobilenetV2/expanded_conv_12/add]
CONV_2D 17.305 0.398 0.398 1.907% 84.765% 0.000 1 [MobilenetV2/expanded_conv_13/expand/Relu6]
DEPTHWISE_CONV_2D 17.703 0.070 0.070 0.335% 85.100% 0.000 1 [MobilenetV2/expanded_conv_13/depthwise/Relu6]
CONV_2D 17.773 0.171 0.171 0.819% 85.920% 0.000 1 [MobilenetV2/expanded_conv_13/project/add_fold]
CONV_2D 17.944 0.279 0.279 1.337% 87.256% 0.000 1 [MobilenetV2/expanded_conv_14/expand/Relu6]
DEPTHWISE_CONV_2D 18.223 0.113 0.113 0.541% 87.798% 0.000 1 [MobilenetV2/expanded_conv_14/depthwise/Relu6]
CONV_2D 18.337 0.286 0.286 1.370% 89.168% 0.000 1 [MobilenetV2/expanded_conv_14/project/add_fold]
ADD 18.623 0.128 0.128 0.613% 89.781% 0.000 1 [MobilenetV2/expanded_conv_14/add]
CONV_2D 18.751 0.211 0.211 1.011% 90.792% 0.000 1 [MobilenetV2/expanded_conv_15/expand/Relu6]
DEPTHWISE_CONV_2D 18.962 0.104 0.104 0.498% 91.290% 0.000 1 [MobilenetV2/expanded_conv_15/depthwise/Relu6]
CONV_2D 19.066 0.274 0.274 1.313% 92.603% 0.000 1 [MobilenetV2/expanded_conv_15/project/add_fold]
ADD 19.340 0.088 0.088 0.422% 93.024% 0.000 1 [MobilenetV2/expanded_conv_15/add]
CONV_2D 19.428 0.272 0.272 1.303% 94.328% 0.000 1 [MobilenetV2/expanded_conv_16/expand/Relu6]
DEPTHWISE_CONV_2D 19.700 0.103 0.103 0.493% 94.821% 0.000 1 [MobilenetV2/expanded_conv_16/depthwise/Relu6]
CONV_2D 19.803 0.480 0.480 2.300% 97.121% 0.000 1 [MobilenetV2/expanded_conv_16/project/add_fold]
CONV_2D 20.283 0.372 0.372 1.782% 98.903% 0.000 1 [MobilenetV2/Conv_1/Relu6]
AVERAGE_POOL_2D 20.655 0.019 0.019 0.091% 98.994% 0.000 1 [MobilenetV2/Logits/AvgPool]
CONV_2D 20.674 0.210 0.210 1.006% 100.000% 0.000 1 [MobilenetV2/Logits/Conv2d_1c_1x1/BiasAdd]
RESHAPE 20.884 0.000 0.000 0.000% 100.000% 0.000 1 [output]
============================== Top by Computation Time ==============================
[node type] [start] [first] [avg ms] [%] [cdf%] [mem KB] [times called] [Name]
CONV_2D 2.658 2.111 2.111 10.114% 10.114% 0.000 1 [MobilenetV2/expanded_conv_1/expand/Relu6]
CONV_2D 0.000 1.504 1.504 7.205% 17.319% 0.000 1 [MobilenetV2/Conv/Relu6]
ADD 7.383 0.901 0.901 4.317% 21.636% 0.000 1 [MobilenetV2/expanded_conv_2/add]
CONV_2D 5.386 0.870 0.870 4.168% 25.804% 0.000 1 [MobilenetV2/expanded_conv_2/expand/Relu6]
DEPTHWISE_CONV_2D 6.257 0.826 0.826 3.957% 29.761% 0.000 1 [MobilenetV2/expanded_conv_2/depthwise/Relu6]
DEPTHWISE_CONV_2D 1.504 0.689 0.689 3.301% 33.062% 0.000 1 [MobilenetV2/expanded_conv/depthwise/Relu6]
CONV_2D 8.284 0.590 0.590 2.827% 35.888% 0.000 1 [MobilenetV2/expanded_conv_3/expand/Relu6]
CONV_2D 14.987 0.512 0.512 2.453% 38.341% 0.000 1 [MobilenetV2/expanded_conv_11/project/add_fold]
CONV_2D 19.803 0.480 0.480 2.300% 40.641% 0.000 1 [MobilenetV2/expanded_conv_16/project/add_fold]
CONV_2D 16.579 0.474 0.474 2.271% 42.912% 0.000 1 [MobilenetV2/expanded_conv_12/project/add_fold]
Number of nodes executed: 65
============================== Summary by node type ==============================
[Node type] [count] [avg ms] [avg %] [cdf %] [mem KB] [times called]
CONV_2D 36 13.695 65.611% 65.611% 0.000 36
DEPTHWISE_CONV_2D 17 4.229 20.261% 85.872% 0.000 17
ADD 10 2.930 14.037% 99.909% 0.000 10
AVERAGE_POOL_2D 1 0.019 0.091% 100.000% 0.000 1
RESHAPE 1 0.000 0.000% 100.000% 0.000 1
Timings (microseconds): count=1 curr=20873
Memory (bytes): count=0
65 nodes observed
0.705882: 653:military uniform
0.541176: 835:suit, suit of clothes
0.533333: 753:racket, racquet
0.505882: 907:Windsor tie
Retriving 173 tensors...