1Copyright 2017 The Android Open Source Project 2 3Licensed under the Apache License, Version 2.0 (the "License"); 4you may not use this file except in compliance with the License. 5You may obtain a copy of the License at 6 7 http://www.apache.org/licenses/LICENSE-2.0 8 9Unless required by applicable law or agreed to in writing, software 10distributed under the License is distributed on an "AS IS" BASIS, 11WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12See the License for the specific language governing permissions and 13limitations under the License. 14------------------------------------------------------------------ 15 16This directory contains models data for the Android Neural Networks API benchmarks. 17 18Included models: 19 20------------------------------------------------------------------ 21- mobilenet_v1_(0.25_128|0.5_160|0.75_192|1.0_224).tflite 22MobileNet tensorflow lite model based on: 23"MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications" 24https://arxiv.org/abs/1704.04861 25Apache License, Version 2.0 26 27Downloaded from 28http://download.tensorflow.org/models/mobilenet_v1_2018_08_02/mobilenet_v1_${variant}.tgz 29on Oct 5 2018 and converted using ToT toco. 30Golden output generated with ToT tensorflow (Linux, CPU). 31 32------------------------------------------------------------------ 33- mobilenet_v1_(0.25_128|0.5_160|0.75_192|1.0_224)_quant.tflite 348bit quantized MobileNet tensorflow lite model based on: 35"MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications" 36https://arxiv.org/abs/1704.04861 37Apache License, Version 2.0 38 39Downloaded from 40http://download.tensorflow.org/models/mobilenet_v1_2018_08_02/mobilenet_v1_${variant}_quant.tgz 41on Oct 5 2018. 42Golden output generated with ToT tflite (Linux, CPU). 43 44------------------------------------------------------------------ 45- mobilenet_v2_(0.35_128|0.5_160|0.75_192|1.0_224).tflite 46MobileNet v2 tensorflow lite model based on: 47"MobileNetV2: Inverted Residuals and Linear Bottlenecks" 48https://arxiv.org/abs/1801.04381 49Apache License, Version 2.0 50 51Downloaded from 52https://storage.googleapis.com/mobilenet_v2/checkpoints/mobilenet_v2_${variant}.tgz 53on Oct 16 2018 and converted using ToT toco. 54Golden output generated with ToT tensorflow (Linux, CPU). 55 56------------------------------------------------------------------ 57- mobilenet_v2_1.0_224_quant.tflite 588bit quantized MobileNet v2 tensorflow lite model based on: 59"MobileNetV2: Inverted Residuals and Linear Bottlenecks" 60https://arxiv.org/abs/1801.04381 61Apache License, Version 2.0 62 63Downloaded from 64http://download.tensorflow.org/models/tflite_11_05_08/mobilenet_v2_1.0_224_quant.tgz 65on Oct 30 2018. 66Golden output generated with ToT tflite (Linux, CPU). 67 68------------------------------------------------------------------ 69- ssd_mobilenet_v1_coco_float.tflite 70Float version of MobileNet SSD tensorflow model based on: 71"Speed/accuracy trade-offs for modern convolutional object detectors." 72https://arxiv.org/abs/1611.10012 73Apache License, Version 2.0 74 75Generated from 76http://download.tensorflow.org/models/object_detection/ssd_mobilenet_v1_coco_2018_01_28.tar.gz 77on Sep 24 2018. 78See also: https://github.com/tensorflow/models/tree/master/research/object_detection 79Golden output generated with ToT tflite (Linux, x86_64 CPU). 80 81------------------------------------------------------------------ 82- ssd_mobilenet_v1_coco_quantized.tflite 838bit quantized MobileNet SSD tensorflow lite model based on: 84"Speed/accuracy trade-offs for modern convolutional object detectors." 85https://arxiv.org/abs/1611.10012 86Apache License, Version 2.0 87 88Generated from 89http://download.tensorflow.org/models/object_detection/ssd_mobilenet_v1_quantized_300x300_coco14_sync_2018_07_18.tar.gz 90on Sep 19 2018. 91See also: https://github.com/tensorflow/models/tree/master/research/object_detection 92Golden output generated with ToT tflite (Linux, CPU). 93 94------------------------------------------------------------------ 95- tts_float.tflite 96TTS tensorflow lite model based on: 97"Fast, Compact, and High Quality LSTM-RNN Based Statistical Parametric Speech Synthesizers for 98Mobile Devices" 99https://ai.google/research/pubs/pub45379 100Apache License, Version 2.0 101 102Note that the tensorflow lite model is the acoustic model in the paper. It is used because it is 103much heavier than the duration model. 104------------------------------------------------------------------ 105- asr_float.tflite 106ASR tensorflow lite model based on the ASR acoustic model in: 107"Personalized Speech recognition on mobile devices" 108https://arxiv.org/abs/1603.03185 109Apache License, Version 2.0 110 111------------------------------------------------------------------ 112Input files: 113------------------------------------------------------------------ 114- ssd_mobilenet_v1_coco_*/tarmac.input 115Photo of airport tarmac by krtaylor@google.com, Apache License, Version 2.0 116- cup_(128|160|192|224).input 117Photo of cup by pszczepaniak@google.com, Apache License, Version 2.0 118- banana_(128|160|192|224).input 119Photo of banana by pszczepaniak@google.com, Apache License, Version 2.0 120- tts_float/arctic_*.input 121Linguistic features and durations generated from text sentences from the CMU Arctic set 122(http://www.festvox.org/cmu_arctic/cmuarctic.data), Apache License, Version 2.0 123- asr_float/*.input 124Acoustic features generated from audio files from the LibriSpeech dataset 125(http://www.openslr.org/12/), Creative Commons Attribution 4.0 International License 126------------------------------------------------------------------ 127 128TODO(pszczepaniak): Provide at least 5 inputs outputs for each model 129 130