def old_target_for_fno(fno):
layer = 'mixed5a_3x3_bottleneck_pre_relu'
channel_1 = 10
channel_2 = 3
if (fno % 240) < 120:
rt = (fno % 240) / 120.0
rit = 1.0 - rt
target = tf.reduce_mean(rit * tfi.T(layer)[:, :, :, channel_1] +
rt * tfi.T(layer)[:, :, :, channel_2])
else:
target = tf.reduce_mean(tfi.T(layer)[:, :, :, channel_2])
return target
def add_texture_ring( current_img, pattern_step, offset, size, layer, channel ):
if pattern_step >= offset and pattern_step < offset + size:
ring_id = pattern_step - offset
ring_img = ring_masks[ring_id]
target = tfi.T(layer)[:,:,:,channel]
textured = tfi.render_deepdream( target, current_img, iter_n=3, step=1.5, octave_n=4, octave_scale=1.5 )
return tfi.masked_mix( current_img, textured, ring_img )
else:
return current_img
def target_for_fno(fno):
return tf.reduce_mean(tfi.T('mixed5a_3x3_bottleneck_pre_relu')[:, :, :, 3])
import tfi
import os
import numpy as np
import PIL.Image
import tensorflow as tf
in_name = 'images/nfrac_1400x840.jpg'
out_name = 'images/start_frame_1400x840.jpeg'
img0 = PIL.Image.open(in_name)
img0 = np.float32(img0)
tfi.reset_graph_and_session()
target = tf.square(tfi.T('mixed4c'))
print('Rendering {}'.format(out_name))
test_img = tfi.render_deepdream(target,
img0,
iter_n=20,
step=0.75,
octave_n=4,
octave_scale=1.5)
tfi.savejpeg(test_img, out_name)
end_frame = 5445
end_colours = np.float32(PIL.Image.open('images/stage03_end_colours.jpeg'))
credit_img = np.float32(PIL.Image.open('images/credits.jpeg'))
complete_fade_img = credit_img * 0
# Stage 3b - rapid zoom in and fade to credits
for frame in range(start_frame, end_frame):
fno = frame + 1
print('Stage 03b - frame {}'.format(fno))
section_id = ((fno - start_frame) // 30) + 2
layer_1 = targets[section_id * 2]
channel_1 = targets[section_id * 2 + 1]
target = tf.reduce_mean(tfi.T(layer_1)[:, :, :, channel_1])
delta_rot = 0.1
delta_zoom = 1.05
total_rot += delta_rot
total_zoom *= delta_zoom
current_img = tfi.affine_zoom(current_img, delta_zoom, delta_rot)
r = (fno - start_frame) / (end_frame - start_frame)
mix_amount = 0.99 * (1 - r) + 0.96 * r
current_img = tfi.mix_images(current_img, end_colours, mix_amount)
current_img = tfi.render_deepdream(target,
current_img,
iter_n=2,
for frame in range(nframes):
fno = end_frame - 1 - frame
section_id = (fno // channel_step)
prev_layer = targets[(section_id - 1) * 2]
prev_channel = targets[(section_id - 1) * 2 + 1]
layer = targets[section_id * 2]
channel = targets[section_id * 2 + 1]
print('Rendering frame {}, using layer {}, channel {}'.format(
fno, layer, channel))
if channel > 1000:
target = tf.square(tfi.T(layer))
else:
target = tfi.T(layer)[:, :, :, channel]
# Mixed target for first half of each channel_step
if (fno % 240) < 120:
r = (fno % 240) / 120.0
ri = 1.0 - r
if prev_layer == layer:
# If the layers match, then shape matches, and we can get a simpler combination
target = tf.reduce_mean(ri *
tfi.T(prev_layer)[:, :, :, prev_channel] +
r * tfi.T(layer)[:, :, :, channel])
else:
target = ri * tf.reduce_mean(
tfi.T(prev_layer)[:, :, :, prev_channel]) + r * tf.reduce_mean(
print('Stage 01 - overlap, frame {}'.format(fno))
section_id = 7
layer_1 = targets[ section_id * 2 + 2]
channel_1 = targets[ section_id * 2 + 3]
# Lattice with gems, which is what we expect to merge with in stage 2
layer_2 = 'mixed5a_3x3_bottleneck_pre_relu'
channel_2 = 3
if ( fno % 240 ) < 120:
r = (fno - frames)/120.0
ri = 1.0 - r
if layer_1 == layer_2:
# If the layers match, then shape matches, and we can get a simpler combination
target = tf.reduce_mean( ri * tfi.T(layer_1)[:,:,:,channel_1] + r * tfi.T(layer_2)[:,:,:,channel_2] )
else:
target = ri * tf.reduce_mean( tfi.T(layer_1)[:,:,:,channel_1] ) + r * tf.reduce_mean( tfi.T(layer_2)[:,:,:,channel_2] )
else:
target = tf.reduce_mean( tfi.T(layer_2)[:,:,:,channel_2] )
delta_rot = transition_rot
delta_zoom = transition_zoom
if ( fno % 240 ) > 230:
delta_rot *= ( 240 - ( fno % 240 ) ) / 10.0
if fno > 980 and ( fno % 240 ) < 10:
delta_rot *= ( fno % 240 ) / 10.0
total_rot += delta_rot
test_layers = [
'mixed4a_3x3_bottleneck_pre_relu', 'mixed4b_3x3_bottleneck_pre_relu',
'head1_bottleneck_pre_relu', 'mixed4c_3x3_bottleneck_pre_relu',
'mixed5b_3x3_bottleneck_pre_relu', 'mixed4e_3x3_bottleneck_pre_relu'
]
iterations = 30
source_img = 'images/example.jpg'
img0 = PIL.Image.open(source_img)
img0 = np.float32(img0)
if not os.path.exists('explore_layers'):
os.makedirs('explore_layers')
for layer in test_layers:
num_channels = tfi.T(layer).get_shape()[3]
directory = 'explore_layers/{}'.format(layer)
if not os.path.exists(directory):
os.makedirs(directory)
print('Rendering {}, all {} channels squared'.format(layer, num_channels))
test_img = tfi.render_deepdream(tf.square(tfi.T(layer)),
img0,
iter_n=iterations,
step=2.0,
octave_n=4,
octave_scale=1.5)
tfi.savejpeg(test_img, ('{}/all_channels_squared.jpeg'.format(directory)))
for channel in range(0, num_channels):
print('Rendering {}, channel {}'.format(layer, channel))