dannysperling/mobilecnn

Mobile CNN

This project represents a method of running convolutional neural networks on an Android device. Much of the code in this project is based on caffe-android-demo by sh1r0, to whom many thanks are given.

##Project The project is an android application built using Android-Studio. Importing the project into Android-Studio should build the application using gradle and allow for on-device debugging with connected devices.

##Models The models used for this project were in caffe. In order to get mobile caffe built into the libraris found in the /jni folder, steps were taken following another project by sh1r0 in conjunction with the first: caffe-android-lib. Instead of using the online models, the weights for the smaller trained model are compressed in models/weights.7z, along with the required model definition for deployment. These two files must be copied onto the sdcard of the android device to get the project working.

Also present in the models folder are the .prototxt files used to train the network. mobilecnn.prototxt was used in conjunction with mobilecnn_solver.prototxt to originally train the network off of the AlexNet network. This network should be initialized in caffe using the weights from the trained bvlc_alexnet model. Once training has been satisfactorily completed, weights from the resulting network can be used to initialize training on mobilecnn_small.prototxt and mobilecnn_solver_small.prototxt to connect and fine tune the network.