npy2ckpt

This is a repo to convert deep learning models from .npy format to .ckpt format.

Set environment

Using Anaconda:

conda create -n npy2ckpt python=3.6
source activate npy2ckpt
pip install tensorflow
conda install -c menpo opencv3

Example of use (GoogleNet)

Download GoogleNet model code (.py) and trained variables (.npy) from: http://www.deeplearningmodel.net/ (You can also find there the imagenet-classes.txt file)

Move the files to the models folder.

Change first line of the model code (.py):

#from kaffe.tensorflow import Network
from network import Network

Run the converter code, pointing to the .npy file:

python npy2ckpt_GoogleNet.py models/googlenet.npy

Take a look and how I set the parameters (width, height, channels, input_node_name) in npy2ckpt_GoogleNet.py.

Test the results:

python test_GoogleNet.py

Example of use (OpenPose)

Following this blog post, use the covert.py function in the Caffe2Tensorflow repo and move the output files to the models folder.

Change first line of the model code (.py):

#from kaffe.tensorflow import Network
from network import Network

Run the converter code, pointing to the .npy file:

python npy2ckpt_OpenPoseNet.py models/openposenet.npy

Take a look and how I set the parameters (width, height, channels, input_node_name) in npy2ckpt_OpenPoseNet.py.

Why

Amazing job here to convert models from Caffe to TensorFlow: https://github.com/ethereon/caffe-tensorflow . The thing is that the output is in .npy format, and I'm not very comfortable dealing with that.

Most of the code is borrowed from that repo. I changed some things to update it to Python 3 and TensorFlow 1.