en10 / kerasmodels

Keras Models

Keras Models

Models differ based on Accuracy and Performance:

• Keras Models
• Keras Models in Tensorflow
• Other Google Pre-trained Models

Models stored in ~/.keras/models/

Install

Models missing from TensorFlow require installing Keras:
Replace from tensorflow.python.keras... with just from keras...

sudo apt update 
sudo apt install python-dev 
sudo pip install keras

NASnet Error ImportError: No module named nasnet Fix:

sudo pip install git+git://github.com/keras-team/keras.git --upgrade --no-deps

Performance

time python predict.py

TensorFlow binary was not compiled to use: SSE4.1 SSE4.2 AVX

Model	Performance (c9 approx)	Size	Example
Inception V3	10s	92M	predict.py
InceptionResNetV2	50s	215M	irn2.py
NASNetMobile	38s	24M	see below
NASNetLarge	5-12m	344M	nasnet.py

colab

run nasnet.py
in line 1 & 2 add from tensorflow.python.

Performance excluding model download:
cpu - 88s
gpu - 34s

FloydHub

FloydHub increases performance, compiled to use SSE4.1 SSE4.2 AVX
NASNetLarge example run.sh runs < 1m on FloydHub:

floyd run \
--data efcic/datasets/nasnet-large/2:models \
--data efcic/datasets/keras/1:keras \
--env tensorflow-1.4 \
"bash run.sh

FloydHub NASNetLarge Setup:

• mount nasnet_large.h5 model
• mount Keras 2.1.2
• run tensorflow 1.4
• copy model & Keras to local
• install Keras 2.1.2
• run inference