TensorFlow For Poets

Retraining one of Google's CNN image classification models to new categories using Transfer Learning.
This can be an much faster (in a few minutes) than training from scratch (Inception V3 took Google, 2 weeks).

Based on Codelab (updated May 30 2018)

• New Updated Version Tensorflow Hub Example (updated 10 Jul 2018)
• Original from Pete Warden

Contents

Install Tensorflow
Download Flowers
Speedup Training
(Re)Training
Classifying an Image
Training on Your Own Categories
Benchmarks
Compiled
Build from Source

Install Tensorflow:

sudo pip install -U pip sudo pip install tensorflow 

Download Flowers:

mkdir tf_files
cd tf_files
curl -O http://download.tensorflow.org/example_images/flower_photos.tgz
tar xzf flower_photos.tgz

• Retrieving the images

Speedup Training

reduce the number of images by ~70%

ls flower_photos/roses | wc -l
rm flower_photos/*/[3-9]*

also only use 2 flowers e.g. roses and sunflowers

rm flower_photos/daisy/ flower_photos/dandelion/ flower_photos/tulips/ -r

• Precompiled bottlenecks also included as tgz.
• Floydhubrun.shIntel Xeon® · 2 Cores 8 GB RAM
• colabK80 12GBcolab.sh

colab code:

!wget https://raw.githubusercontent.com/EN10/TensorFlowForPoets/master/colab.sh
!bash colab.sh

(Re)Training

Download retrain file:

wget https://raw.githubusercontent.com/EN10/TensorFlow-For-Poets/master/retrain.py

MobileNet 0.5: Faster (< 2m) Less Accurate (Top-1 64%)

python retrain.py \
--bottleneck_dir=tf_files/bottlenecks \
--how_many_training_steps=500 \
--model_dir=tf_files/models/ \
--summaries_dir=tf_files/training_summaries/mobilenet_0.50_224 \
--output_graph=tf_files/retrained_graph.pb \
--output_labels=tf_files/retrained_labels.txt \
--architecture=mobilenet_0.50_224 \
--image_dir=tf_files/flower_photos

mobilenet can be configured from 0.50 to 1.0, more accuracy but slower.

Inception V3: Slower (5 - 20m) More Accurate (Top-1 78%)

python retrain.py \
--bottleneck_dir=tf_files/bottlenecks \
--how_many_training_steps=500 \
--model_dir=tf_files/models/inception_v3 \
--output_graph=tf_files/retrained_graph.pb \
--output_labels=tf_files/retrained_labels.txt \
--image_dir=tf_files/flower_photos

• (Re)training Inception

Classifying an Image

Download label_image file:

wget https://raw.githubusercontent.com/EN10/TensorFlow-For-Poets/master/label_image.py

Predict image label:

python label_image.py \
--graph=tf_files/retrained_graph.pb \
--image=image.jpg

• Label Image

Training on Your Own Categories

retrain.py uses --image_dir as the root folder for training.
Each sub-folder is named after one of your categories and contains only images from that category.

• Training on Your Own Categories

For training to work well, you should gather at least a hundred photos of each kind of object you want to recognize.

• Creating a Set of Training Images

Tool to download images for training:

• Image Batch Downloader

Benchmarks:

rm [3-9]* & 2 Flowers: roses and sunflowers (aka slim)
282 roses & 304 sunfowers i.e. 586 bottlenecks

Model	PAAS	RAM	OS	Tensorflow	CPU	Performance	Notes
MobileNet 0.5	cs50.io	512MB	14.04.5	1.4.1	Not Compiled	1m20s
MobileNet 0.5	cs50.io	1024MB	14.04.5	1.4.1	Not Compiled	1m20s
MobileNet 0.5	colab	1024MB	17.10	1.6.0rc1	Compiled	43s
MobileNet 0.5	colab	12GB	17.10	1.6.0rc1	K80	24s
Inception v3	cs50.io	512MB	14.04.5	1.4.1	Not Compiled	15m
Inception v3	cs50.io	512MB	14.04.5	1.2.1	Compiled	6m30s
Inception v3	floydhub	8GB	16.04.2	1.1.0	Compiled	4m30s	with install
Inception v3	floydhub	8GB	16.04.2	1.1.0	Compiled	4m	with datasets
Inception v3	codenvy.io	3072MB	16.04	1.2.1	Not Compiled	6m45s
Inception v3	codenvy.io	3072MB	16.04	1.2.1	Compiled	3m20s
Inception v3	codenvy.io	2048MB	16.04	1.2.1	Compiled	3m20s
Inception v3	colab	1024MB	17.10	1.6.0rc1	Compiled	3m38s
Inception v3	colab	12GB	17.10	1.6.0rc1	K80	1m23s

Performance = Bottlenecks + Training where Training ~ 1 Min
CPU: TensorFlow binary compiled to use: SSE4.1 SSE4.2 AVX
Precompiled Bottlenecks in /inception_bottlenecks (slim see above)

• Workspace Benchmark

Compiled

Precompiled with FMA, AVX, AVX2, SSE4.1, SSE4.2:

• FloydHub EN10

• FMA, AVX, AVX2, SSE4.1, SSE4.2 Builds Working on codenvy.io not on c9.io

• Precompiled Example for 14.04

wget https://github.com/EN10/BuildTF/raw/771df48529285c69ef760327121e996750b3916e/tensorflow-1.4.0-cp27-none-linux_x86_64.whl
sudo pip install --ignore-installed --upgrade tensorflow-1.4.0-cp27-none-linux_x86_64.whl

OS: Ubuntu 14.04.5 LTS - GCC version 4.8.4 - Python: 2.7.6 - Tensorflow 1.2.1

Build from Source

• Docker Ubuntu
• Install Bazel
• Build from Source
• Add SSE & AVX

Build:

bazel build --config=opt --copt=-msse4.2 --copt=-mavx2 --copt=-mfma //tensorflow/tools/pip_package:build_pip_package