def test_squeeze2d(self):
arr = np.reshape(np.arange(0, 1 * 8 * 8 * 1, 1), [1, 8, 8, 1])
x = tf.constant(arr)
with self.session():
squeeze_result = squeeze2d(x).eval()
squeeze_result_old = squeeze2d_old(x).eval()
import pdb
pdb.set_trace()
self.fail()
def split2d(name, z, cfg, objective=0.):
with tf.variable_scope(name):
n_z = ops.int_shape(z)[3]
z1 = z[:, :, :, :n_z // 2]
z2 = z[:, :, :, n_z // 2:]
pz = split2d_prior(z1, cfg)
objective += pz.logp(z2)
z1 = ops.squeeze2d(z1)
eps = pz.get_eps(z2)
return z1, objective, eps
def split2d(z, objective=0., hps=None, name=None):
with tf.variable_scope(name):
n_z = z.get_shape()[3]
z1, z2 = tf.split(z, 2, axis=-1)
pz = split2d_prior(z1, hps=hps)
obj = pz.logp(z2)
objective += obj
z1 = squeeze2d(z1)
eps = pz.get_eps(z2)
return z1, objective, eps
def split2d(z, objective=0., hps=None, name=None):
with tf.variable_scope(name):
n_z = z.get_shape()[3]
z1, z2 = tf.split(z, 2, axis=-1)
pz = split2d_prior(z1, hps=hps)
obj = pz.logp(z2)
objective += obj
z1 = squeeze2d(z1)
eps = pz.get_eps(z2)
return z1, objective, eps
def call(self, inputs, determinant=True):
det_list = []
z_list = []
self.z_shape_list = []
x = inputs
determinant_accumulator = tf.constant(0.)
for l in range(0, self.L-1):
x = ops.squeeze2d(x)
with tf.name_scope("L_%d"%l):
for k in range(0, self.K):
with tf.name_scope("K_%d"%k):
[x, det] = self.blocks[l][k](x, determinant=True)
determinant_accumulator += det
det_list.append(det)
with tf.name_scope("extract_latents"):
num_channels = x.get_shape().as_list()[-1]
self.z_shape_list.append(x[:, :, :, :num_channels//2].get_shape().as_list()[1:])
z_list.append(tf.layers.flatten(x[:, :, :, :num_channels//2]))
x = x[:, :, :, num_channels//2:]
x = ops.squeeze2d(x)
l = max(self.L-1, 0)
with tf.name_scope("L_%d"%l):
for k in range(0, self.K):
with tf.name_scope("K_%d"%k):
[x, det] = self.blocks[l][k](x, determinant=True)
determinant_accumulator += det
det_list.append(det)
self.z_shape_list.append(x.get_shape().as_list()[1:])
z_list.append(tf.layers.flatten(x))
# combine z values
z_list.reverse()
z = tf.concat(z_list, axis=-1, name="combine_multiscale_latents")
# combine
determinants = tf.stack(det_list, name="determinants")
if determinant:
return [z, determinant_accumulator]
return z
def encode(self, inputs, labels, condition=None):
## Dequantization by adding uniform noise
with tf.variable_scope("preprocess"):
self.y = tf.one_hot(labels,
depth=self.num_classes,
dtype=tf.float32)
inputs = tf.cast(inputs, 'float32')
self.height, self.width, self.channels = inputs.get_shape(
).as_list()[1:]
if self.hps.num_bits_x < 8:
inputs = tf.floor(inputs / 2**(8 - self.hps.num_bits_x))
inputs = inputs / self.num_bins - 0.5
inputs = inputs + tf.random_uniform(tf.shape(inputs), 0,
1. / self.num_bins)
objective = tf.zeros(tf.shape(inputs)[0])
objective += -np.log(self.num_bins) * np.prod(
ops.shape(inputs)[1:])
inputs = squeeze2d(inputs)
## Encoder
if self.hps.conditioning and condition is None:
condition = self.y
# with tf.variable_scope("cond_preprocess"):
# condition = tf.layers.dense(condition, units=10, use_bias=False)
z, objective, eps = codec(inputs,
cond=condition,
objective=objective,
hps=self.hps,
reverse=False)
## Prior
with tf.variable_scope("prior"):
self.hps.top_shape = z.get_shape().as_list()[1:]
logp, sample, get_eps = prior(self.y, self.hps)
obj = logp(z)
eps.append(get_eps(z))
objective += obj
self.objective = -objective
# Class label predict with latent representation
if self.hps.ycond:
z_y = tf.reduce_mean(z, axis=[1, 2])
self.logits = linear_zeros(z_y,
self.num_classes,
name="classifier")
return eps
def _f_loss(self, x, y):
with tf.variable_scope('model', reuse=tf.AUTO_REUSE):
y_onehot = tf.cast(tf.one_hot(y, self.cfg.n_y, 1, 0), 'float32')
# Discrete -> Continuous
objective = tf.zeros_like(x, dtype='float32')[:, 0, 0, 0]
z = x # + tf.random_uniform(tf.shape(x), 0, 1./self.cfg.n_bins)
objective += - np.log(self.cfg.n_bins) * \
np.prod(ops.int_shape(z)[1:])
# Encode
z = ops.squeeze2d(z, 2) # > 16x16x12
z, objective, eps = self.encoder(z, objective)
# Prior
self.cfg.top_shape = ops.int_shape(z)[1:]
logp, _, _ = prior("prior", y_onehot, self.cfg)
objective += logp(z)
# Generative loss
nobj = - objective
bits_x = nobj / (np.log(2.) * int(x.get_shape()[1]) * int(
x.get_shape()[2]) * int(x.get_shape()[3])) # bits per subpixel
# Predictive loss
if self.cfg.weight_y > 0 and self.cfg.ycond:
assert(False)
# Classification loss
h_y = tf.reduce_mean(z, axis=[1, 2])
y_logits = ops.dense(
"classifier", h_y, self.cfg.n_y, has_bn=False)
bits_y = tf.nn.softmax_cross_entropy_with_logits_v2(
labels=y_onehot, logits=y_logits) / np.log(2.)
# Classification accuracy
y_predicted = tf.argmax(y_logits, 1, output_type=tf.int32)
classification_error = 1 - \
tf.cast(tf.equal(y_predicted, y), tf.float32)
else:
bits_y = tf.zeros_like(bits_x)
classification_error = tf.ones_like(bits_x)
return bits_x, bits_y, classification_error, eps
def split2d(z, objective=0., hps=None, name=None):
'''
:param z: inputs，通过Squeeze处理之后变为z
:param objective：未知
:param hps：参数列表
:param name:
:return:
'''
with tf.variable_scope(name):
# 2019.11.6
n_z = z.get_shape()[3]
# n_z为inputs的通道数
z1, z2 = tf.split(z, 2, axis=-1)
pz = split2d_prior(z1, hps=hps)
obj = pz.logp(z2)
objective += obj
z1 = squeeze2d(z1)
eps = pz.get_eps(z2)
return z1, objective, eps
def f_encode(self, x, y):
with tf.variable_scope('model', reuse=tf.AUTO_REUSE):
y_onehot = tf.cast(tf.one_hot(y, self.cfg.n_y, 1, 0), 'float32')
# Discrete -> Continuous
objective = tf.zeros_like(x, dtype='float32')[:, 0, 0, 0]
z = x + tf.random_uniform(tf.shape(x), 0, 1. / self.cfg.n_bins)
objective += - np.log(self.cfg.n_bins) * \
np.prod(ops.int_shape(z)[1:])
# Encode
z = ops.squeeze2d(z, 2) # > 16x16x12
z, objective, eps = self.encoder(z, objective)
# Prior
self.cfg.top_shape = ops.int_shape(z)[1:]
logp, _, _eps = prior("prior", y_onehot, self.cfg)
objective += logp(z)
eps.append(_eps(z))
return eps
def encode(self, inputs, labels, condition=None):
## Dequantization by adding uniform noise
with tf.variable_scope("preprocess"):
self.y = tf.one_hot(labels, depth=self.num_classes, dtype=tf.float32)
inputs = tf.cast(inputs, 'float32')
self.height, self.width, self.channels = inputs.get_shape().as_list()[1:]
if self.hps.num_bits_x < 8:
inputs = tf.floor(inputs/2**(8-self.hps.num_bits_x))
inputs = inputs / self.num_bins - 0.5
inputs = inputs + tf.random_uniform(tf.shape(inputs), 0, 1./self.num_bins)
objective = tf.zeros(tf.shape(inputs)[0])
objective += -np.log(self.num_bins) * np.prod(ops.shape(inputs)[1:])
inputs = squeeze2d(inputs)
## Encoder
if self.hps.conditioning and condition is None:
condition = self.y
# with tf.variable_scope("cond_preprocess"):
# condition = tf.layers.dense(condition, units=10, use_bias=False)
z, objective, eps = codec(inputs, cond=condition, objective=objective, hps=self.hps, reverse=False)
## Prior
with tf.variable_scope("prior"):
self.hps.top_shape = z.get_shape().as_list()[1:]
logp, sample, get_eps = prior(self.y, self.hps)
obj = logp(z)
eps.append(get_eps(z))
objective += obj
self.objective = -objective
# Class label predict with latent representation
if self.hps.ycond:
z_y = tf.reduce_mean(z, axis=[1, 2])
self.logits = linear_zeros(z_y, self.num_classes, name="classifier")
return eps
def encode(self, inputs, labels, condition=None):
# line268
# Dequantization by adding uniform noise 加入均匀噪声来反量化
with tf.variable_scope("preprocess"):
# 采用One-hot编码
self.y = tf.one_hot(labels,
depth=self.num_classes,
dtype=tf.float32)
inputs = tf.cast(inputs, 'float32')
# tf.cast()数据类型转换
self.height, self.width, self.channels = inputs.get_shape(
).as_list()[1:]
# num_bits、num_bin：？？？
if self.hps.num_bits_x < 8:
inputs = tf.floor(inputs / 2**(8 - self.hps.num_bits_x))
inputs = inputs / self.num_bins - 0.5
# 输入加入随机正态分布
inputs = inputs + tf.random_uniform(tf.shape(inputs), 0,
1. / self.num_bins)
# objective：？？？
objective = tf.zeros(tf.shape(inputs)[0])
# np.prod：计算数组中所有元素的乘积
objective += -np.log(self.num_bins) * np.prod(
ops.shape(inputs)[1:])
# 使用Squeezing操作（在进入Squeezing之前的操作未知）
print("before inter squeeze2d, input.shape=" + inputs.shape())
inputs = squeeze2d(inputs)
# inputs的shape为[图片数, 高度, 宽度, 通道数]
# Encoder 编码
# 下面作用未知
if self.hps.conditioning and condition is None:
condition = self.y
# with tf.variable_scope("cond_preprocess"):
# condition = tf.layers.dense(condition, units=10, use_bias=False)
print("before inter model.codec, inputs.shape=" + inputs.shape())
z, objective, eps = codec(inputs,
cond=condition,
objective=objective,
hps=self.hps,
reverse=False)
# line 11
# Prior 先验
with tf.variable_scope("prior"):
self.hps.top_shape = z.get_shape().as_list()[1:]
logp, sample, get_eps = prior(self.y, self.hps)
obj = logp(z)
eps.append(get_eps(z))
objective += obj
self.objective = -objective
# Class label predict with latent representation：潜变量标签预测
if self.hps.ycond:
z_y = tf.reduce_mean(z, axis=[1, 2])
self.logits = linear_zeros(z_y,
self.num_classes,
name="classifier")
return eps