def test_stack_augmentations(outdir, datadir):
fname = "/home/leon/uni/bachelor/deepdecoder/test/data/00350.hdf5"
myconfig = config.load(str(datadir.join('augmentation.yml')))
dset = DistributionHDF5Dataset(fname)
batch = next(dset.iter(15**2))
image_save(str(outdir.join('real.png')), tile(batch['real']))
image_save(str(outdir.join('fake.png')), np.clip(tile(batch['fake']), -1, 1))
# for i, name in enumerate(['tag3d']):
for i, name in enumerate(['tag3d', 'tag3d_lighten', 'fake_without_noise', 'fake']):
augment = stack_augmentations(name, myconfig)
xs = augment(batch)
image_save(str(outdir.join('{}_{}_aug_tiled.png'.format(i, name))), tile(xs))
def call(self, samples):
tiles = []
for tag3d, fake in zip(self.preprocess(samples['tag3d']),
self.preprocess(samples['fake'])):
tiles.append(tag3d)
tiles.append(fake)
return tile(tiles, columns_must_be_multiple_of=2)
def call(self, samples):
tiles = []
for tag3d, fake in zip(self.preprocess(samples['tag3d']),
self.preprocess(samples['fake'])):
tiles.append(tag3d)
tiles.append(fake)
return tile(tiles, columns_must_be_multiple_of=2)
def save_samples(data, bits, outfname):
print("bits", bits.shape)
nb_bits = bits.shape[1]
for i in range(len(data)):
s = 64 // nb_bits
for j in range(nb_bits):
data[i, :s, s*j:s*(j+1)] = 2*bits[i, j] - 1
image_save(outfname, np.clip(tile(data), -1, 1))
def save_samples(data, bits, outfname):
print("bits", bits.shape)
nb_bits = bits.shape[1]
for i in range(len(data)):
s = 64 // nb_bits
for j in range(nb_bits):
data[i, :s, s * j:s * (j + 1)] = 2 * bits[i, j] - 1
image_save(outfname, np.clip(tile(data), -1, 1))
def save_real_images(real_hdf5_fname, output_dir, nb_visualise=20**2):
real = next(train_data_generator(real_hdf5_fname, nb_visualise, 10))['data']
print(real.shape)
print(real.max())
print(real.min())
tiled = tile(real)
fname = join(output_dir, 'real_{}.png'.format(nb_visualise))
image_save(fname, tiled[0])
def save_real_images(real_hdf5_fname, output_dir, nb_visualise=20**2):
real = next(train_data_generator(real_hdf5_fname, nb_visualise,
10))['data']
print(real.shape)
print(real.max())
print(real.min())
tiled = tile(real)
fname = join(output_dir, 'real_{}.png'.format(nb_visualise))
image_save(fname, tiled[0])
def visualize_detection_tiles(dset, name, n=20**2):
crown_vis = ResultCrownVisualizer()
imgs_raw = []
imgs_overlayed = []
nb_tags = len(dset['tags'])
n = min(nb_tags, n)
indicies = np.arange(nb_tags)
np.random.shuffle(indicies)
for i in range(n):
p = indicies[i]
position = np.array([dset['tags'][0, 0].shape]) / 2
tag = dset['tags'][p, 0]
bits = (dset['bits'][p:p + 1] + 1) / 2.
overlay = crown_vis(tag, position, np.zeros((1, 3)), bits)[0]
overlayed = crown_vis.add_overlay((tag + 1) / 2, overlay)
imgs_raw.append(tag)
imgs_overlayed.append(2 * overlayed - 1)
tiled_raw = tile([img.swapaxes(0, -1) for img in imgs_raw])
tiled_overlayed = tile([img.swapaxes(0, -1) for img in imgs_overlayed])
image_save(name + '_raw.png', tiled_raw)
image_save(name + '_overlayed.png', tiled_overlayed)
def visualize_detection_tiles(dset, n=12**2):
crown_vis = ResultCrownVisualizer()
imgs = []
indicies = np.arange(len(dset['tags']))
np.random.shuffle(indicies)
for i in range(n):
p = indicies[i]
position = np.array([dset['tags'][0, 0].shape]) / 2 + 140
tag = dset['tags'][p, 0] / 255
overlay = crown_vis(tag, position, np.zeros((1, 3)), dset['bits'][p:p+1])[0]
imgs.append(crown_vis.add_overlay(tag, overlay))
tiled = tile([img.swapaxes(0, -1) for img in imgs])
imsave('visualize_debug_detections.png', tiled.swapaxes(0, -1))
def visualize_detection_tiles(dset, n=12**2):
crown_vis = ResultCrownVisualizer()
imgs = []
indicies = np.arange(len(dset['tags']))
np.random.shuffle(indicies)
for i in range(n):
p = indicies[i]
position = np.array([dset['tags'][0, 0].shape]) / 2 + 140
tag = dset['tags'][p, 0] / 255
overlay = crown_vis(tag, position, np.zeros((1, 3)),
dset['bits'][p:p + 1])[0]
imgs.append(crown_vis.add_overlay(tag, overlay))
tiled = tile([img.swapaxes(0, -1) for img in imgs])
imsave('visualize_debug_detections.png', tiled.swapaxes(0, -1))
def call(self, samples):
real = samples['real']
scores = samples['discriminator_on_real']
idx = np.argsort(scores.flatten())[::-1]
real_sorted = real[idx]
return tile(self.preprocess(real_sorted))
def call(self, samples):
fakes = samples['fake']
scores = samples['discriminator_on_fake']
idx = np.argsort(scores.flatten())[::-1]
fakes_sorted = fakes[idx]
return tile(self.preprocess(fakes_sorted))
def call(self, samples):
real = samples['real']
scores = samples['discriminator_on_real']
idx = np.argsort(scores.flatten())[::-1]
real_sorted = real[idx]
return tile(self.preprocess(real_sorted))
def call(self, samples):
fakes = samples['fake']
scores = samples['discriminator_on_fake']
idx = np.argsort(scores.flatten())[::-1]
fakes_sorted = fakes[idx]
return tile(self.preprocess(fakes_sorted))
def plot_anitaliasing(tag_dist, fname, a, nb_samples=64):
_, masks, depth_map = next(generator(tag_dist, nb_samples, antialiasing=a))
tiled = tile(masks)[0]
imsave(fname.format(a), tiled)
