Input optimization with keras models. This can be used for feature visualization, neural style transfer,adversarial examples, and other stuff.
This project provides 3 main things:
- A way of inserting layers into an already trained keras models**
- New layers which facilitate feature visualization and other types on input optimization.
- A way of using keras optimizers to optimize your input tensors**
Update (2019/01/03) : I've updated the code to work with weight normalization optimizers (pulled straight from https://github.com/openai/weightnorm/blob/master/keras/weightnorm.py). All you need to do is load the optimizer from kinopt.weightnorm as you would any keras optimizer and it should work.
** Note: Most of this project was done awhile back and used kind of hack-y ways to accomplish some of these goals. Since then, I've figured out how to do some of these with mostly standard Keras/Tensorflow machinery. Both the old and the new ways are explained in featvis_long_tutorial.py
With this, you will be able to compare the feature visualization of any keras model see this article for more info. For example, you can have different models draw bees:
The code works for model trained in keras (any backend), however to run this code you'll need to use the tensorflow backend. Only tested with python 3.5-3.6
Standard installation, git clone this repo
git clone https://github.com/PhilippeNguyen/kinopt.git
then pip install
pip install -e ./kinopt
See featvis_example.py for a quick script showing how to do feature visualization.
Explanation of some of the script arguments:
- --model: path to a saved keras model. Models are saved with the model.save function. Doesn't work if you only save the weights using the Model.save_weights function. Note that you'll want to have the pointwise activations (relu/softmax/...) separated from the linear transforms (Dense/Convs/...), my script at /other_scripts/separate_activations.py should do this for you. Keras gives you access to a bunch of classification models, see keras_applications (comes installed with keras). Every one of these models should work, though I couldn't try NASNetLarge since it was too big to download.
- --preprocess_mode: the preprocessing function to use. See keras_applications.imagenet_utils.preprocess_input for the different types of modes available. If you're using a keras_application model, then they'll be using one of the modes, and you'll probably need to dig around the source code on github to see what it is.
- --layer_identifier: Either the layer name or the layer index in Model.get_config()['layers']
- --neuron_index: This is the channel index of the tensor corresponding to the layer identified above
- --output: name/path of the output png to save the optimized input image.
See featvis_long_tutorial.py for long explanation of what's going on, as well as alternate methods of doing input optimization in keras so you won't have to deal with my janky code. Note that optimizing in the fourier space (as mentioned here) is implemented, but not shown in the feat vis exmaples. I also add a random roll transformation which is not mentioned in the article.
With slight modifications you can do some dumb stuff like activating as many neurons as you can.
You can also do neural style transfer, see neural_style_example.py