Deep Dream Travel allows you to turn a single image into a video, diving through various layers of a neural network.
This effect is achieved by recursively feeding the image into the network and regularly switching the targeted layer.
A higher resolution example can be found on YouTube.
Create and activate a new Anaconda environment
conda create --name DeepDreamTravel
conda activate DeepDreamTravel
Install required dependencies
pip install -r requirements.txt
conda install psutil
Install Caffe depending on your operating system
conda install caffe <-- Linux
conda install caffe -c willyd <-- Windows
python deepdreamtravel.py INPUT_FILE OUTPUT_VIDEO
or view all options
python deepdreamtravel.py -h
View example.py for an example of running DeepDreamTravel inside your application.
Available arguments are listed bellow.
from deepdreamtravel import DeepDreamTravel
dreamer = DeepDreamTravel('protxt', 'caffemodel') # View example.py for an example
dreamer.generate(input_image="noise.jpg",
node_switch=10, # How often should layers switch
resize_coeff=0.05, # Zoom amount in %
show_iter=100, # The image will be shown every X iterations
offset=[0,0,0,0], # Left, Top, Right, Bottom zoom offset
temp_dir="tmp", # Temp directory to store frames
start_iter=0, # If generation was interrupted, enter the next iteration here
start_index=0, # Which layer index should be used first. Also used when interruption occured
start_offset=0, # In-case there are invalid layers, enter the offset given by terminal output when interruption happened.
max_iteration=False, # If false, all layers will be "explored". Else set maximum number of iterations
octaves=4, # Higher number leads to bigger ("better") visuals. Takes significantly more time to generate! Lower number in-case of errors with small images.
iter_n=10, # Number of iterations per octave.
octave_scale=1.4, # Resize amount per octave. Higher number leads to higher dream states.
output_video="DeepDreamTravel.avi", # Output location and name. Don't forget to end with .avi!
fps=30, # Frames per second
delete_temp=True, # Clean up generated images after video was generated?
max_memory=3 # Maximum amount of memory to use in GB
)Note: The examples were generated with Places205-GoogLeNet. You can find other interesting models here.
If you are having issues with the installation, Nerdy Rodent was kind enough to create a tutorial.
Additionally, support, questions, and requests are provided here
Pull requests are welcome. For major changes, please open an issue first to discuss what you would like to change.