caffe2tf

This Python library (compatible with Tensorflow version 0.12) enables the use of Caffe prototxt file for creating the equivalent neural network in TensorFlow. While experimenting with different network architectures and hyperparameters, it is convenient to keep track of portable, language independent files encoding this information, rather than making copies of Python files. caffe2tf is different from, but complimentary to tf-slim and the wonderful library caffe-tensorflow; these libraries allow creation of a network via abstractions within Python code (additionally, caffe-tensorflow converts Caffe model parameter files to NumPy pickles).

caffe2tf is a work in progress, and currently supports the following layers:

• Input
• 2D Convolution
• Pooling (avg/max)
• ReLU
• Dropout
• Inner Product
• Softmax
• Sigmoid
• Reshape
• Concat
• Slice

I am currently adding support for layers needed for my experiments, but if there is interest from the community, I will be happy to add more!

Support planned for:

• 3D Convolution
• Batch Normalization
• Deconvolution

For loading pre-trained parameters, the code currently relies on the NumPy pickle created by caffe-tensorflow for the corresponding network parameter file. Note that parsing the original Caffe file via Protobuf is painfully slow, which is why it is best to rely on the converted file. Note that this feature is untested. I will update this file once correctness is verified.

Additionally, the code includes a modified copy of caffe.proto (hope this isn't a copyright infringement; someone, please let me know!), which has the unimplemented features commented out. This will be updated as support is added for additional layers.

Dependencies

Developed in Python 2.7. Requires the following external Python modules:

• Google Protobuf (can be installed via pip)
• Tensorflow (refer to the installation instructions page)
• caffe-tensorflow (only for using pre-trained weights from Caffe model file; not needed for successful compilation)

Installation

Download caffe.proto, and compile using protoc, which should be included in Google Protobuf.

protoc --python_out=. caffe.proto

This will create caffe_pb2.py in the current directory; move it inside the caffe2tf directory, along with the rest of the source code. That's it!

Usage

from caffe2tf.network import Network
import tensorflow as tf
# Create network.
net = Network("the_proto_file.prototxt")
# Load variables (requires active Tensorflow session)
with tf.Session() as sess:
    net.load("the_model_pickle_file", sess)

To access the output of a particular layer, use the top value as defined in the prototxt file for the model. For example:

# Example for VGG-16.
output = net.layer_output("conv_1_1")

Finally

I am excited to hear your comments, suggestions, and criticisms -- be it regarding feature requests, bug reports, coding style, or the library architecture! Feel free to start an Issues thread or get in touch directly via email (ID on my GitHub profile).