mtyka / deepdream_highres

Deepdream in Tensorflow for highres images

Highres Deepdream

The size/resolution limitation of the original deepdream Caffe implementation made deep dreaming large photos tricky. However, the offical Google Deepdream example that comes with Tensorflow by Alex Mordvintsev, doesnt suffer from those limitations. However it only came as a IPython notebook, so here's a little commandline app based on Alex's notebook for just applying deepdream to large images in bulk.

Install Requirements:

• Python 2.7
• TensorFlow (>=r0.7)
• NumPy

Download pretrained Inception V5 from here: https://storage.googleapis.com/download.tensorflow.org/models/inception5h.zip

Run deepdream like this: python deepdream.py --input mymegapixelimage.jpg --output output.jpg --layer import/mixed5a_5x5_pre_relu --frames 7 --octaves 10 --iterations 10

Or create a zoom-in series like so:

python deepdream.py --input mystart.jpg --output output.jpg --layer import/mixed4d_3x3 --frames 100 --frame_scale 1.4 --frame_crop --octaves 4 --iterations 10

For help & options: python deepdream.py --help

deepdream_highres's Issues

How fast is it supposed to run on GPU?

Hi, I am just trying to find out whether my setup is messed up or if everything is working as intended.

I am running Ubuntu 16.04, CUDA 9.0 and cudNN 7.04 on Tensorflow 1.5.

Both CPU and GPU are utilized, I am seeing around 10% GPU usage and 100% CPU. A 720p image with your endless zoom parameters takes around 45 seconds.

Everything that has to do with layers of the convnet is reported as running on CPU:
I tensorflow/core/common_runtime/placer.cc:874] import/conv2d0_w: (Const)/job:localhost/replica:0/task:0/device:CPU:0

Everything else is reported as running on GPU.

Is that just how it is or can I run everything on GPU? Cheers

It seems that thats is in 2.x

Tensorflow has dropped support for python 2.x a bit back all fork and see if i can share something :)