afcarl / deepdream-mxnet

An implementation of deepdream for mxnet

deepdream-mxnet

An implementation of deepdream for mxnet

Installation

0. Recommended Create a new Virtualenvironment
1. Install MXNet in your venv (compile it by yourself, or use pip (pip install mxnet)
2. Install all other requirements with pip install -r requirements.txt
3. Profit

Usage

Assume you are having the following files:

• Inception-BN-0126.params (trained model)
• Inception-BN-symbol.json (network definition)

1. You will need to rename Inception-BN-symbol.json to orig-Inception-BN-symbol.json
2. Choose the layer you want to visualize by looking at the names of the Convolutional layers in the symbol definition file
3. Remember how many channels this layer has.
4. Set all required values in the config file (config.json)
   1. input_shape the shape of the input images in the form num_channels, height, width
   2. batch_size the batch size to use while dreaming
   3. scale_n the number of downscale steps for the laplacian gradient normalization (2^scale_n should be smaller than your image size)
   4. num_steps number of optimization steps per octave
   5. num_octaves number of times the image shall be increased in size
   6. octave_scale how much the size should increase
   7. step_size step size for applying the gradient on the input iamge
   8. max_tile_size max size of each tile in pixels for saving GPU memory
   9. mean RGB mean values that should be subtracted from the input image (not mandatory)
5. Start the visualization with: python deepdream.py <model-prefix> <epoch> <layer_name> <layer_id> -g <gpu_to_use> --all -c config.json --folder <place to save resulting images>, with our example data: python deepdream.py Inception-BN 0126 conv_4d_double_3x3_1 140 --all -c config.json -g 0 --folder images/inception
6. Profit again!

Questions?

Feel free to open an issue.

Improvements?

I'm happy to review your Pull Request!