def save_img(self):
if self.img4save is None:
return
lines = [self.img4save]
if self.learn_from_unlabel:
lines.append(self.semi_img4save)
tile_img = None
for l in lines:
for i, sect in enumerate(l):
#l[0] = (l[0] + 1) * 0.5
if sect is None:
l[i] = np.zeros_like(l[0])
continue
if i != 0:
sect = onehot2label(sect)
l[i] = np.transpose(sect, (1, 2, 0))
l = np.concatenate(l, axis=1)
if tile_img is None:
tile_img = l
else:
tile_img = np.concatenate((tile_img, l), axis=0)
out = np.uint8(tile_img * 255)
out = Image.fromarray(out)
out_dir_name = self.opt.out_dir + '/out_img'
os.makedirs(out_dir_name, exist_ok=True)
out.save(out_dir_name + '/' + str(self.num_saved_img) + '.png')
self.num_saved_img += 1
def save_img(self):
out_dir = self.opt.out_dir + '/img/'
os.makedirs(out_dir, exist_ok=True)
train = self.img4save
valid1 = self.evaluate()
valid2 = self.evaluate()
tile_img = None
for seed, real, fake in [train, valid1, valid2]:
seed = onehot2label(seed, skip_bg=True, dtype='float32')
tp = (1, 2, 0)
seed = np.transpose(seed, tp)
real = np.transpose(real, tp)
fake = np.transpose(fake, tp)
real = (real + 1) / 2
fake = (fake + 1) / 2
row_img = np.concatenate((seed, real, fake), axis=1)
if tile_img is None:
tile_img = row_img
else:
tile_img = np.concatenate((tile_img, row_img), axis=0)
img_array = np.uint8(tile_img * 255)
img = Image.fromarray(img_array)
img.save(out_dir + 'tile_img-' + str(self.num_saved_img) + '.png')
self.num_saved_img += 1
def main():
out_dir = 'predict_to'
in_dir = 'predict_from'
batch_size = 1
gen_npz = 'pretrained/gen.npz'
opt = get_options()
gen = ResNetDeepLab(opt)
gen.to_gpu(0)
chainer.serializers.load_npz(gen_npz, gen)
gen.to_cpu()
num = 0
ksize = 5
os.makedirs(out_dir, exist_ok=True)
files = glob(in_dir + '/*.*')
for filename in files:
print(filename)
img_array = np.array(Image.open(filename), dtype='float32')
img_array = img_array.transpose((2, 0, 1)) / 255
x = chainer.Variable(img_array[np.newaxis, :3, :, :])
out = gen(x)
onehot = out.array[0]
x = x.array[0]
out = onehot2label(onehot)
bg_onehot = np.argmax(onehot, axis=0)
bg_onehot = bg_onehot == 4
bg_threshold = 0.5
bg_num = x * bg_onehot
bg_num = bg_num > bg_threshold
bg_num = np.sum(bg_num, axis=0)
bg_num = np.sum(bg_num == 3)
bg_ratio = bg_num / np.sum(bg_onehot)
if bg_ratio < 0.6:
print('bg is black')
continue
out = np.transpose(out * 255, (1, 2, 0)).astype('uint8')
#out = remove_noise(out, ksize=ksize)
#exist eye ?
if not is_exist_color(out, [255, 0, 0], threshold_num=32):
print('not exist eye')
continue
#exist face ?
if not is_exist_color(out, [0, 255, 0], threshold_num=100):
print('not exist face')
continue
#exist hair ?
if not is_exist_color(out, [0, 0, 255], threshold_num=100):
print('not exist hair')
continue
x = np.transpose(x * 255, (1, 2, 0)).astype('uint8')
out_img = np.concatenate((x, out), axis=1)
img = Image.fromarray(out_img)
path = out_dir + '/' + str(num) + '.png'
img.save(path)
num += 1
def main():
out_predict_dir = 'out'
device = 0
gen_npz = 'trained/gen_snapshot_epoch-900.npz'
opt = get_options()
opt.spade_ch = 32
opt.ngf = 64
opt.ndf = 64
gen = SPADEGenerator(opt)
gen.to_gpu(device)
chainer.serializers.load_npz(gen_npz, gen)
os.makedirs(out_predict_dir, exist_ok=True)
out_dir = out_predict_dir + '/predicted'
os.makedirs(out_dir, exist_ok=True)
num = 0
dir_path = 'datasets/resnet-large_hc'
files = glob(dir_path + '/*.png')
random.shuffle(files)
for img_path in files:
if not os.path.exists(img_path):
continue
img = Image.open(img_path)
if img == None:
continue
print(img_path)
img_array = np.array(img).astype('float32') / 255
img_array = np.transpose(img_array, (2, 0, 1))
t_array = img_array[:3, :, :256]
x_array = img_array[:3, :, 256:512]
c_array = img_array[:3, :, 512:]
#to onehot
x_array = label2onehot(x_array,
threshold=0.4,
skip_bg=True,
dtype='float32').astype('float32')
c_array = c_array * x_array[2]
#cast 16bit -> 32bit (cannot use tensor core)
x = Variable(cuda.to_gpu(x_array[np.newaxis, :, :, :]))
c = Variable(cuda.to_gpu(c_array[np.newaxis, :, :, :]))
out = gen([x, c])[0]
out = cp.asnumpy(out.array[0])
out = (out + 1) / 2
x = cp.asnumpy(x.array[0])
x = onehot2label(x, skip_bg=True, dtype='float32')
out = np.transpose(out * 255, (1, 2, 0)).astype('uint8')
x = np.transpose(x * 255, (1, 2, 0)).astype('uint8')
y = np.transpose(t_array * 255, (1, 2, 0)).astype('uint8')
out_img = np.concatenate((x, y, out), axis=1)
img = Image.fromarray(out_img)
path = out_dir + '/' + str(num) + '.png'
img.save(path)
num += 1