Impact of increasing the pyramid sizeĭreaming is performed on multiple image resolutions stacked 'vertically' (we call this an image pyramid). Left: ResNet50-ImageNet (we can see more animal features) Right: ResNet50-Places365 (human built stuff, etc.). If we keep every other parameter the same but we swap the pretrained weights we get these: The 1st and 3rd were created using VGG 16 (ImageNet) and the middle one using ResNet50 pretrained on Places 365. Optimizing deeper Layers = Amplify high-level featuresīy using deeper network layers you'll get higher level patterns (eyes, snouts, animal heads): Here the first 2 images came from ResNet50 and the last one came from GoogLeNet (both pretrained on ImageNet). Here are some examples that you can create using this code! Optimizing shallower layers = Amplify low-level featuresīy using shallower layers of neural networks you'll get lower level patterns (edges, circles, colors, etc.) as the output: This repo is an attempt of making the cleanest DeepDream repo that I'm aware of + it's written in PyTorch! ❤️ Static Image Examples Most of the original Deep Dream repos were written in Caffe and the ones written in PyTorch are usually really hard to read and understand. Who would have said that neural networks had this creativity hidden inside? 🎨 Why yet another Deep Dream repo? So from an input image like the one on the left after 'dreaming' we get the image on the right: In a nutshell the algorithm maximizes the activations of chosen network layers by doing a gradient ascent. Note: it's pretty large, ~10 MBs, so it may take a couple of attempts to load it in the browser here on GitHub. I strongly suggest you start with the Jupyter notebook that I've created! It's got a full support for the command line usage and a Jupyter Notebook!Īnd it will give you the power to create these weird, psychedelic-looking images: This repo contains a PyTorch implementation of the Deep Dream algorithm ( 🔗 blog by Mordvintstev et al.).

Deep dream google download