def face_parse(image, label):
conv = dict(kernel=3, padding='SAME')
deconv = dict(kernel=3, padding='SAME', bias=True)
subpixel = dict(kernel=3, factor=2, padding='SAME')
with tf.default_args(conv=conv, deconv=deconv, subpixel=subpixel):
net = image
net = net.conv(16).bn().relu().conv(16).bn().relu().maxpool()
net = net.conv(32).bn().relu().conv(32).bn().relu().maxpool()
net = net.conv(64).bn().relu().conv(64).bn().relu().maxpool()
net = net.conv(64).bn().relu().conv(64).bn().relu().maxpool()
net = net.conv(64).bn().relu()
net = net.deconv(64).bn().relu()
net = net.subpixel().deconv(64).bn().relu().deconv(64).bn().relu()
net = net.subpixel().deconv(64).bn().relu().deconv(64).bn().relu()
net = net.subpixel().deconv(32).bn().relu().deconv(32).bn().relu()
net = net.subpixel().deconv(16).bn().relu().deconv(11, bias=True)
prob = net.softmax()
summary_parse(prob)
losses = tf.softmax_cross_entropy(net, label, name='losses')
loss = losses.mean()
return tf.dic(losses=losses, loss=loss, outputs=[prob])
def dataset_train(batch, shape=None, folder=None):
import os
from scipy.io import loadmat
# load .mat file
folder = folder or _folder()
fpath = os.path.join(folder, 'train_32x32.mat')
data = loadmat(fpath)
x = data['X'] #
y = data['y'] # 1~10 ?
x = x.transpose((3, 0, 1, 2))
y[y == 10] = 0
n = x.shape[0] # total data
# convert to tensors
x = tf.convert_to_tensor(x) # uint8 (73257, 32, 32, 3)
y = tf.convert_to_tensor(y) # uint8 (73257, 1)
# random select
i = tf.random_uniform((batch, ), minval=0, maxval=n, dtype=tf.int32)
img = tf.gather(x, i).to_float() / 255.
label = tf.gather(y, i).to_int32()
if shape is not None:
img = tf.image.resize_images(img, shape)
return tf.dic(image=img, label=label)
def trainset(batch, size=(256, 256), threads=8, sizedown=None, removebg=False):
import sflow.tf as tf
# qtrain : from train
with tf.name_scope('helen_trainset'):
# fid = tf.feed.read_line(_trainfile(), shuffle=True)
line = tf.feed.read_line(_trainfile(), shuffle=True)
_, fid = tf.decode_csv(line, [[0], ['']], field_delim=',')
trimed = tf.string_split([fid], delimiter=' ') # remove space
# now trimed.values[0] has fid
fid = trimed.values[0] # ex) '100032540_1'
qtrain = feed_img_label(fid, size, shuffle=True, threads=threads)
image, label = qtrain.dequeue_many(batch)
image = image.to_float() / 255.
image = tf.identity(image, name='image')
label = label.to_float() / 255.
if sizedown is not None:
image = image.sizedown(sizedown)
label = label.sizedown(sizedown)
label = normalize_label(label)
if removebg is True:
# remove background image
bg = 1. - label[:, :, :, :1]
image = image * bg
return tf.dic(image=image, label=label)
def testset(batch, size=(256, 256), threads=8, sizedown=None, removebg=False):
import sflow.tf as tf
with tf.name_scope('helen_testset'):
# qtest : from testing.txt format , seperated
line = tf.feed.read_line(_testfile())
_, fid = tf.decode_csv(line, [[0], ['']], field_delim=',')
trimed = tf.string_split([fid], delimiter=' ') # remove space
# now trimed.values[0] has fid
fid = trimed.values[0] # ex) '100032540_1'
qtest = feed_img_label(fid,
size,
shuffle=False,
rotate=False,
threads=threads)
data = qtest.dequeue_many(batch)
image = data[0].to_float() / 255.
label = data[1].to_float() / 255.
if sizedown is not None:
image = image.sizedown(sizedown)
label = label.sizedown(sizedown)
label = normalize_label(label)
if removebg is True:
# remove background image
bg = 1. - label[:, :, :, :1]
image = image * bg
return tf.dic(image=image, label=label)
def dataset_test(batch, shuffle=True, size=(28, 28), folder=None):
if not shuffle:
raise NotImplementedError
_, (img, label) = _load_data(size, folder)
with tf.name_scope('dataset_mnist_test'):
img = tf.constant(img)
label = tf.constant(label)
ind = tf.random_uniform((batch, ),
minval=0,
maxval=img.dims[0],
dtype=tf.int32)
x = tf.gather(img, ind)
y = tf.gather(label, ind)
if x.ndim == 3:
x = x.expand_dims(3)
if x.dtype != tf.float32:
x = x.to_float() / 255.
y = y.to_int32()
return tf.dic(image=x, label=y)
def trainset_old(batch,
size=(256, 256),
threads=8,
sizedown=None,
removebg=False):
import sflow.tf as tf
# qtrain : from train
with tf.name_scope('helen_trainset'):
fid = tf.feed.read_line(_trainfile(), shuffle=True)
qtrain = feed_img_label(fid, size, shuffle=True, threads=threads)
image, label = qtrain.dequeue_many(batch)
image = image.to_float() / 255.
image = tf.identity(image, name='image')
label = label.to_float() / 255.
if sizedown is not None:
image = image.sizedown(sizedown)
label = label.sizedown(sizedown)
label = normalize_label(label)
if removebg is True:
# remove background image
bg = 1. - label[:, :, :, :1]
image = image * bg
return tf.dic(image=image, label=label, batch=batch)
def dataset_emotion(batch, threads=8, shuffle=None, capacity=10, folder=None):
"""
:param batch:
:param threads:
:param shuffle:
:param capacity:
:param folder:
:return: dict(image, label)
"""
with tf.name_scope('fer2013.emotion'):
# face image 48x48x1
img = tf.placeholder(tf.float32, shape=(48, 48, 1), name='image')
# emotion label
label = tf.placeholder(tf.int32, shape=(), name='emotion')
placeholders = [img, label]
q = tf.feed.gen_producer(placeholders,
gen_random_face(folder),
capacity=capacity,
threads=threads,
shuffle=shuffle)
d = q.dequeue_many(batch)
return tf.dic(image=d[0], label=d[1])
def attr_dataset(batch, attr, value, size=None, crop=None,
threads=8, shuffle=None,
partition='train', folder=None, fliplr=False):
"""
get data set given attribute data
:param batch:
:param attr:
:param value:
:param size:
:param crop:
:param threads:
:param shuffle:
:param partition:
:param folder:
:param fliplr:
:return:
"""
folder = folder or datasetfolder()
shuffle = shuffle or (partition == 'train')
attr = attr.lower()
files = _pair_list_file(attr, folder=folder, partition=partition)
if not (os.path.exists(files[0]) and os.path.exists(files[1])):
_prepare_attr_pair_list(attr, folder=folder, partition=partition)
if value is False:
filelist = files[0]
else:
filelist = files[1]
imgfolder = os.path.join(folder, 'Img/img_align_celeba/')
imgfolder = tf.constant(imgfolder)
with tf.name_scope(None, 'celeba.attr.{0}.{1}'.format(attr, value)):
line = tf.feed.read_line(filelist)
line = line.print('line', first_n=10)
fields = tf.decode_csv(line, [['']] + [[-1]]*40, field_delim=' ')
fname = fields[0]
fname = imgfolder + fname
img = tf.feed.read_image(fname, channels=3)
attrs = tf.stack(fields[1:])
attrs = tf.equal(attrs, 1).to_float()
capacity = 512
shapes = [(218, 178, 3), (40,)]
q = tf.feed.queue_producer([img, attrs], capacity, shapes=shapes,
threads=threads, shuffle=shuffle)
img, attrs = q.dequeue_many(batch)
img = img.to_float()/255.
crop = crop or (178, 178)
if crop is not None:
img = tf.img.crop_center(crop, img)
if size is not None:
img = tf.image.resize_images(img, size)
if fliplr:
img = tf.img.rand_fliplr(img, p=0.5)
return tf.dic(image=img, label=attrs, batch=batch)
def dataset_train(batch, **kwargs):
"""
sketch collection to face photo collection
:param batch:
:param kwargs:
:return:
"""
a = dataset_trainA(batch, **kwargs)
b = dataset_trainB(batch, **kwargs)
return tf.dic(A=a, B=b, batch=batch)
def attribute_dataset(batch,
size=None,
crop=None,
threads=8,
shuffle=None,
partition='train',
folder=None,
fliplr=False):
folder = folder or datasetfolder()
# assert size is None # not implemented size
attrfile = 'list_attr_{0}.txt'.format(partition)
shuffle = shuffle or (partition == 'train')
trainfile = os.path.join(folder, 'Anno', attrfile)
imgfolder = os.path.join(folder, 'Img/img_align_celeba/')
# imgfolder = os.path.join(folder, 'process/resized/')
imgfolder = tf.constant(imgfolder)
with tf.name_scope('celeba_attribute'):
line = tf.feed.read_line(trainfile)
line = line.print('line', first_n=10)
fields = tf.decode_csv(line, [['']] + [[-1]] * 40, field_delim=' ')
fname = fields[0]
fname = imgfolder + fname
img = tf.feed.read_image(fname, channels=3)
attrs = tf.stack(fields[1:])
attrs = tf.equal(attrs, 1).to_float()
capacity = 512
shapes = [(218, 178, 3), (40, )]
q = tf.feed.queue_producer([img, attrs],
capacity,
shapes=shapes,
threads=threads,
shuffle=shuffle)
img, attrs = q.dequeue_many(batch)
# img = tf.img.rand_fliplr(img, p=0.5)
img = img.to_float() / 255.
if size is not None:
img = tf.image.resize_images(img, size)
if crop is not None:
img = tf.img.crop_center(crop, img)
if fliplr:
img = tf.img.rand_fliplr(img, p=0.5)
return tf.dic(image=img, label=attrs, batch=batch)
def attr_pair(batch, attr, size=None, crop=None, threads=8,
shuffle=None, partition='train',
folder=None, fliplr=False):
size = size or (128, 128)
data0 = attr_dataset(batch, attr, value=False, size=size, crop=crop, threads=threads,
shuffle=shuffle, partition=partition,
folder=folder, fliplr=fliplr)
data1 = attr_dataset(batch, attr, value=True, size=size, crop=crop, threads=threads,
shuffle=shuffle, partition=partition,
folder=folder, fliplr=fliplr)
return tf.dic(x0=data0.image, x1=data1.image)
def face_parse2(image, label):
conv = dict(kernel=3, padding='SAME')
deconv = dict(kernel=3, padding='SAME', bias=True)
# subpixel = dict(kernel=3, factor=2, padding='SAME')
maxpool_where = dict(kernel=2)
unpool_where = dict(kernel=2)
with tf.default_args(conv=conv,
deconv=deconv,
maxpool_where=maxpool_where,
unpool_where=unpool_where):
net = image
wheres = []
net, where = net.conv(16).bn().relu().conv(
16).bn().relu().maxpool_where()
wheres.append(where)
net, where = net.conv(32).bn().relu().conv(
32).bn().relu().maxpool_where()
wheres.append(where)
net, where = net.conv(64).bn().relu().conv(
64).bn().relu().maxpool_where()
wheres.append(where)
net, where = net.conv(64).bn().relu().conv(
64).bn().relu().maxpool_where()
wheres.append(where)
net = net.conv(64).bn().relu()
net = net.deconv(64).bn().relu()
net = net.unpool_where(
wheres.pop()).deconv(64).bn().relu().deconv(64).bn().relu()
net = net.unpool_where(
wheres.pop()).deconv(64).bn().relu().deconv(32).bn().relu()
net = net.unpool_where(
wheres.pop()).deconv(32).bn().relu().deconv(16).bn().relu()
net = net.unpool_where(wheres.pop()).deconv(16).bn().relu().deconv(
11, bias=True)
net = tf.summary_activation(net, name='logits')
prob = net.softmax()
summary_parse(prob)
# losses = tf.nn.sigmoid_cross_entropy_with_logits(net, label)
losses = tf.softmax_cross_entropy(net, label, name='losses')
loss = losses.mean()
return tf.dic(loss=loss, logits=net, label=label, image=image)
def attribute(batch, size=None, threads=8):
"""
add comment size, fields
:param batch:
:param size:
:param threads:
:return:
"""
return tf.dic(
train=attribute_dataset(batch,
size=size,
threads=threads,
partition='train'),
valid=attribute_dataset(batch,
size=size,
threads=threads,
partition='valid'),
)
def dataset_train(batch, size=(28, 28), folder=None):
(img, label), _ = _load_data(size, folder)
with tf.name_scope('dataset_mnist'):
# train.shape == (60000, 28, 28)
img = tf.constant(img)
label = tf.constant(label)
ind = tf.random_uniform((batch, ),
minval=0,
maxval=img.dims[0],
dtype=tf.int32)
x = tf.gather(img, ind)
y = tf.gather(label, ind)
if x.ndim == 3:
x = x.expand_dims(3)
if x.dtype != tf.float32:
x = x.to_float() / 255.
y = y.to_int32()
return tf.dic(image=x, label=y)
def dataset_train(batch, **kwargs):
a = dataset_pair(batch, 'train', shuffle=True, **kwargs)
b = dataset_pair(batch, 'train', shuffle=True, **kwargs)
return tf.dic(A=a.A, B=b.B, batch=batch)
def dataset_train(batch, **kwargs):
a = dataset_trainA(batch, **kwargs)
b = dataset_trainB(batch, **kwargs)
return tf.dic(A=a, B=b, batch=batch)
def dataset_pair_val(batch, **kwargs):
data = dataset_reader_AB(batch, _folder(), 'val', **kwargs).split(2,
axis=2)
return tf.dic(A=data[0], B=data[1], batch=batch)
def dataset_valid(batch, **kwargs):
a = dataset_valA(batch, **kwargs)
b = dataset_valB(batch, **kwargs)
return tf.dic(A=a, B=b, batch=batch)
def dataset_test(batch, **kwargs):
a = dataset_testA(batch, **kwargs)
b = dataset_testB(batch, **kwargs)
return tf.dic(A=a, B=b, batch=batch)
def dataset_pair(batch, partition, **kwargs):
shape = (256, 512, 3)
data = dataset_reader(batch, _folder(), partition, shape=shape, **kwargs)
a, b = tf.split(data, 2, axis=2)
return tf.dic(A=a, B=b, batch=batch)
