Create model with own images

[Manfred Pauli]

Hello,

I want to train a Caffe model with my own image set. I've split them it in a test and training set, created labels and stored them in a lmdb (following https://github.com/BVLC/caffe/issues/550). 

Can I now use simply the "bvlc_googlenet/solver.prototxt" (or quick_solver.prototxt) by modifying the "num_output: 1000" to "num_output: <my number of categories/labels>" in the train_val.prototxt? Will this work?

Or do I have to dive deep in creating my own train_val.prototxt? 

Thank you

[Nguyen Tuan Hung]

I have trained my own model based on cifar10_train_test.prototxt for my image database with high accuracy.

Hence, I think bvlc_googlenet/solver.prototxt could work well for you. 

Of course, you should modify the num_output to be equal to your number of categories.

Then, you can try to modify a little bit in bvlc_googlenet model to improve the accuracy on your database.

--