It'll be interesting to see what imagery people are able to generate using the described technique. If you post images to Google+, Facebook, or Twitter, be sure to tag them with #deepdream so other researchers can check them out too.
This notebook is designed to have as few dependencies as possible:
# imports and basic notebook setup from cStringIO import StringIO import numpy as np import scipy.ndimage as nd import PIL.Image from IPython.display import clear_output, Image, display from google.protobuf import text_format import caffe # If your GPU supports CUDA and Caffe was built with CUDA support, # uncomment the following to run Caffe operations on the GPU. # caffe.set_mode_gpu() # caffe.set_device(0) # select GPU device if multiple devices exist def showarray(a, fmt='jpeg'): a = np.uint8(np.clip(a, 0, 255)) f = StringIO() PIL.Image.fromarray(a).save(f, fmt) display(Image(data=f.getvalue()))
model_path = '../caffe/models/bvlc_googlenet/' # substitute your path here net_fn = model_path + 'deploy.prototxt' param_fn = model_path + 'bvlc_googlenet.caffemodel' # Patching model to be able to compute gradients. # Note that you can also manually add "force_backward: true" line to "deploy.prototxt". model = caffe.io.caffe_pb2.NetParameter() text_format.Merge(open(net_fn).read(), model) model.force_backward = True open('tmp.prototxt', 'w').write(str(model)) net = caffe.Classifier('tmp.prototxt', param_fn, mean = np.float32([104.0, 116.0, 122.0]), # ImageNet mean, training set dependent channel_swap = (2,1,0)) # the reference model has channels in BGR order instead of RGB # a couple of utility functions for converting to and from Caffe's input image layout def preprocess(net, img): return np.float32(np.rollaxis(img, 2)[::-1]) - net.transformer.mean['data'] def deprocess(net, img): return np.dstack((img + net.transformer.mean['data'])[::-1])
Making the "dream" images is very simple. Essentially it is just a gradient ascent process that tries to maximize the L2 norm of activations of a particular DNN layer. Here are a few simple tricks that we found useful for getting good images:
First we implement a basic gradient ascent step function, applying the first two tricks:
def objective_L2(dst): dst.diff[:] = dst.data def make_step(net, step_size=1.5, end='inception_4c/output', jitter=32, clip=True, objective=objective_L2): '''Basic gradient ascent step.''' src = net.blobs['data'] # input image is stored in Net's 'data' blob dst = net.blobs[end] ox, oy = np.random.randint(-jitter, jitter+1, 2) src.data = np.roll(np.roll(src.data, ox, -1), oy, -2) # apply jitter shift net.forward(end=end) objective(dst) # specify the optimization objective net.backward(start=end) g = src.diff # apply normalized ascent step to the input image src.data[:] += step_size/np.abs(g).mean() * g src.data = np.roll(np.roll(src.data, -ox, -1), -oy, -2) # unshift image if clip: bias = net.transformer.mean['data'] src.data[:] = np.clip(src.data, -bias, 255-bias)
Next we implement an ascent through different scales. We call these scales "octaves".
def deepdream(net, base_img, iter_n=10, octave_n=4, octave_scale=1.4, end='inception_4c/output', clip=True, **step_params): # prepare base images for all octaves octaves = [preprocess(net, base_img)] for i in xrange(octave_n-1): octaves.append(nd.zoom(octaves[-1], (1, 1.0/octave_scale,1.0/octave_scale), order=1)) src = net.blobs['data'] detail = np.zeros_like(octaves[-1]) # allocate image for network-produced details for octave, octave_base in enumerate(octaves[::-1]): h, w = octave_base.shape[-2:] if octave > 0: # upscale details from the previous octave h1, w1 = detail.shape[-2:] detail = nd.zoom(detail, (1, 1.0*h/h1,1.0*w/w1), order=1) src.reshape(1,3,h,w) # resize the network's input image size src.data = octave_base+detail for i in xrange(iter_n): make_step(net, end=end, clip=clip, **step_params) # visualization vis = deprocess(net, src.data) if not clip: # adjust image contrast if clipping is disabled vis = vis*(255.0/np.percentile(vis, 99.98)) showarray(vis) print octave, i, end, vis.shape clear_output(wait=True) # extract details produced on the current octave detail = src.data-octave_base # returning the resulting image return deprocess(net, src.data)
Now we are ready to let the neural network reveal its dreams! Let's take a cloud image as a starting point:
img = np.float32(PIL.Image.open('sky1024px.jpg')) showarray(img)
Running the next code cell starts the detail generation process. You may see how new patterns start to form, iteration by iteration, octave by octave.
3 9 inception_4c/output (575, 1024, 3)
The complexity of the details generated depends on which layer's activations we try to maximize. Higher layers produce complex features, while lower ones enhance edges and textures, giving the image an impressionist feeling:
_=deepdream(net, img, end='inception_3b/5x5_reduce')
3 9 inception_3b/5x5_reduce (575, 1024, 3)
We encourage readers to experiment with layer selection to see how it affects the results. Execute the next code cell to see the list of different layers. You can modify the
make_step function to make it follow some different objective, say to select a subset of activations to maximize, or to maximize multiple layers at once. There is a huge design space to explore!
What if we feed the
deepdream function its own output, after applying a little zoom to it? It turns out that this leads to an endless stream of impressions of the things that the network saw during training. Some patterns fire more often than others, suggestive of basins of attraction.
We will start the process from the same sky image as above, but after some iteration the original image becomes irrelevant; even random noise can be used as the starting point.
!mkdir frames frame = img frame_i = 0
h, w = frame.shape[:2] s = 0.05 # scale coefficient for i in xrange(100): frame = deepdream(net, frame) PIL.Image.fromarray(np.uint8(frame)).save("frames/%04d.jpg"%frame_i) frame = nd.affine_transform(frame, [1-s,1-s,1], [h*s/2,w*s/2,0], order=1) frame_i += 1
Be careful running the code above, it can bring you into very strange realms!
The image detail generation method described above tends to produce some patterns more often the others. One easy way to improve the generated image diversity is to tweak the optimization objective. Here we show just one of many ways to do that. Let's use one more input image. We'd call it a "guide".
guide = np.float32(PIL.Image.open('flowers.jpg')) showarray(guide)
Note that the neural network we use was trained on images downscaled to 224x224 size. So high resolution images might have to be downscaled, so that the network could pick up their features. The image we use here is already small enough.
Now we pick some target layer and extract guide image features.
end = 'inception_3b/output' h, w = guide.shape[:2] src, dst = net.blobs['data'], net.blobs[end] src.reshape(1,3,h,w) src.data = preprocess(net, guide) net.forward(end=end) guide_features = dst.data.copy()
Instead of maximizing the L2-norm of current image activations, we try to maximize the dot-products between activations of current image, and their best matching correspondences from the guide image.
def objective_guide(dst): x = dst.data.copy() y = guide_features ch = x.shape x = x.reshape(ch,-1) y = y.reshape(ch,-1) A = x.T.dot(y) # compute the matrix of dot-products with guide features dst.diff.reshape(ch,-1)[:] = y[:,A.argmax(1)] # select ones that match best _=deepdream(net, img, end=end, objective=objective_guide)
3 9 inception_3b/output (575, 1024, 3)
This way we can affect the style of generated images without using a different training set.