def flow_from_directory(self, directory, *args, **kwargs):
generator = super().flow_from_directory(directory,
class_mode=None,
*args,
**kwargs)
while True:
# Get augmentend image samples
ori = next(generator)
# Get masks for each image sample
#fileName = FILE_DIR + 'resized_COCA_COLA.jpg'
#print (fileName)
mask = np.stack([
random_mask(ori.shape[1], ori.shape[2])
for _ in range(ori.shape[0])
],
axis=0)
# Apply masks to all image sample
masked = deepcopy(ori)
masked[mask == 0] = 1
# Yield ([ori, masl], ori) training batches
# print(masked.shape, ori.shape)
gc.collect()
yield [masked, mask], ori
def flow_from_directory(self, directory, *args, **kwargs):
generator = super().flow_from_directory(directory, class_mode=None, *args, **kwargs)
cnt = 0
# Data augmentation
while True:
# Get augmented image samples
ori = next(generator)
ori_length = ori.shape[0]
# Get masks for each image sample
mask = np.stack([random_mask(ori.shape[1], ori.shape[2]) for _ in range(ori_length)], axis=0)1
# Crop ori, mask and masked images
croped_ori, croped_mask = self.random_crop(ori, mask)
# Apply masks to all image sample
masked = deepcopy(croped_ori)
masked[croped_mask == 0] = 1
# Yield ([ori, masl], ori) training batches
gc.collect()
# Save image
# self.save_img(cnt=cnt, img_size_type='', masked=masked)
# cnt += 1
yield [masked, croped_mask], croped_ori
def flow(self, x, *args, **kwargs):
while True:
# Get augmentend image samples
ori = next(super().flow(x, *args, **kwargs))
# Get masks for each image sample
mask = np.stack([random_mask(ori.shape[1], ori.shape[2]) for _ in range(ori.shape[0])], axis=0)
# Apply masks to all image sample
masked = deepcopy(ori)
masked[mask==0] = 1
# Yield ([ori, masl], ori) training batches
# print(masked.shape, ori.shape)
gc.collect()
yield [masked, mask], ori
def flow_from_directory(self, directory_small, directory_medium,
directory_large, *args, **kwargs):
target_small = (512, 1024)
target_medium = (768, 1536)
target_large = (1536, 3072)
generator_small = super().flow_from_directory(directory_small,
target_size=target_small,
class_mode=None,
*args,
**kwargs)
generator_medium = super().flow_from_directory(
directory_medium,
target_size=target_medium,
class_mode=None,
*args,
**kwargs)
generator_large = super().flow_from_directory(directory_large,
target_size=target_large,
class_mode=None,
*args,
**kwargs)
ratio_small = 1 / 3
ratio_medium = 1 / 3
ratio_large = 1 / 3
cnt = 0
# Data augmentation
while True:
#
# Choice generator
#
img_size_type = np.random.choice(
['small', 'medium', 'large'],
p=[ratio_small, ratio_medium, ratio_large])
if img_size_type == 'small':
generator = generator_small
elif img_size_type == 'medium':
generator = generator_medium
else:
generator = generator_large
# Get augmented image samples
#
ori_img = next(generator) * 255
ori_length = ori_img.shape[0]
# Change color tone for data augmentation using imgaug
seq = iaa.Sequential([
iaa.Add((-20, 20), per_channel=True),
])
ori = seq.augment_images(ori_img) / 255
#
# Create images for training
#
# Crop original images
# croped_ori = self.random_crop(ori)
# croped_ori_length = croped_ori.shape[0]
# Get masks for each image sample
mask = np.stack([
random_mask(ori.shape[1], ori.shape[2])
for _ in range(ori_length)
],
axis=0)
# Crop ori, mask and masked images
croped_ori, croped_mask = self.random_crop(ori=ori, mask=mask)
# Apply masks to all image sample
masked = deepcopy(croped_ori)
masked[croped_mask == 0] = 1
# 白色が50%以上ある場合はやり直し
if self.has_many_white(cnt, masked):
continue
# Yield ([ori, masl], ori) training batches
gc.collect()
#
# Optional
#
# Check mask ratio
# self.has_many_mask(croped_mask[0,:,:,:])
# Save images
# self.save_img(cnt=cnt, img_size_type=img_size_type, mask=mask, masked=masked, ori=ori)
cnt += 1
yield [masked, croped_mask], croped_ori
import cv2
from skimage.io import imsave
# Import modules from libs/ directory
from libs.pconv_model import PConvUnet
from libs.util import random_mask, plot_images
# Load image
img = cv2.imread('./data/building.jpg')
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
img = img / 255
shape = img.shape
print(f"Shape of image is: {shape}")
# Load mask
mask = random_mask(shape[0], shape[1])
# Image + mask
masked_img = deepcopy(img)
masked_img[mask == 0] = 1
model = PConvUnet(weight_filepath='result/logs/')
model.load(r"result/logs/1_weights_2019-02-21-04-59-53.h5", train_bn=False)
# Run prediction quickly
pred = model.scan_predict((img, mask))
# Show result
plot_images([img, masked_img, pred])
imsave('result/test_orginal.png', img)
imsave('result/test_masked.png', masked_img)
# coding: utf-8
# # Mask Generation with OpenCV
# In the paper they generate irregular masks by using occlusion/dis-occlusion between two consecutive frames of videos, as described in [this paper](https://lmb.informatik.uni-freiburg.de/Publications/2010/Bro10e/sundaram_eccv10.pdf).
#
# Instead we'll simply be using OpenCV to generate some irregular masks, which will hopefully perform just as well.
# We've implemented this in the function `random_mask`, which is located in the `util.py` file int he libs directory
import itertools
import matplotlib.pyplot as plt
from libs.util import random_mask
# Plot the results
_, axes = plt.subplots(5, 5, figsize=(20, 20))
axes = list(itertools.chain.from_iterable(axes))
for i in range(len(axes)):
# Generate image
img = random_mask(500, 500)
# Plot image on axis
axes[i].imshow(img * 255)
print("finish")
