def discriminator(self, inp, cond, stages, t, reuse=False):
alpha_trans = self.alpha_tra
with tf.variable_scope("d_net", reuse=reuse):
x_iden = None
if t:
x_iden = pool(inp, 2)
x_iden = self.from_rgb(x_iden, stages - 2)
x = self.from_rgb(inp, stages - 1)
for i in range(stages - 1, 0, -1):
with tf.variable_scope(self.get_conv_scope_name(i), reuse=reuse):
x = conv2d(x, f=self.get_dnf(i), ks=(3, 3), s=(1, 1), act=lrelu_act())
x = conv2d(x, f=self.get_dnf(i-1), ks=(3, 3), s=(1, 1), act=lrelu_act())
x = pool(x, 2)
if i == stages - 1 and t:
x = tf.multiply(alpha_trans, x) + tf.multiply(tf.subtract(1., alpha_trans), x_iden)
with tf.variable_scope(self.get_conv_scope_name(0), reuse=reuse):
# Real/False branch
cond_compress = fc(cond, units=128, act=lrelu_act())
concat = self.concat_cond4(x, cond_compress)
x_b1 = conv2d(concat, f=self.get_dnf(0), ks=(3, 3), s=(1, 1), act=lrelu_act())
x_b1 = conv2d(x_b1, f=self.get_dnf(0), ks=(4, 4), s=(1, 1), padding='VALID', act=lrelu_act())
output_b1 = fc(x_b1, units=1)
return output_b1
def construct_bottom_block(self, inputs, name):
num_outputs = inputs.shape[self.channel_axis].value
conv1 = ops.conv2d(
inputs, 2*num_outputs, self.conv_size, name+'/conv1')
conv2 = ops.conv2d(
conv1, num_outputs, self.conv_size, name+'/conv2')
return conv2
def discriminator(self, image, is_training, reuse=False):
with tf.variable_scope("discriminator"):
if reuse:
tf.get_variable_scope().reuse_variables()
# [batch,256,256,1] -> [batch,128,128,64]
h0 = lrelu(conv2d(image, self.discriminator_dim,
scope="d_h0_conv"))
# [batch,128,128,64] -> [batch,64,64,64*2]
h1 = lrelu(
batch_norm(conv2d(h0,
self.discriminator_dim * 2,
scope="d_h1_conv"),
is_training,
scope="d_bn_1"))
# [batch,64,64,64*2] -> [batch,32,32,64*4]
h2 = lrelu(
batch_norm(conv2d(h1,
self.discriminator_dim * 4,
scope="d_h2_conv"),
is_training,
scope="d_bn_2"))
# [batch,32,32,64*4] -> [batch,31,31,64*8]
h3 = lrelu(
batch_norm(conv2d(h2,
self.discriminator_dim * 8,
sh=1,
sw=1,
scope="d_h3_conv"),
is_training,
scope="d_bn_3"))
# real or fake binary loss
fc1 = fc(tf.reshape(h3, [self.batch_size, -1]), 1, scope="d_fc1")
return tf.sigmoid(fc1), fc1
def generator(self, z, embed, is_training=True, reuse=False, cond_noise=True):
s = self.output_size
s2, s4, s8, s16 = int(s / 2), int(s / 4), int(s / 8), int(s / 16)
with tf.variable_scope("g_net", reuse=reuse):
# Sample from the multivariate normal distribution of the embeddings
mean, log_sigma = self.generate_conditionals(embed)
net_embed = self.sample_normal_conditional(mean, log_sigma, cond_noise)
# --------------------------------------------------------
# Concatenate the sampled embedding with the z vector
net_input = tf.concat([z, net_embed], 1)
net_h0 = tf.layers.dense(net_input, units=self.gf_dim*8*s16*s16, activation=None,
kernel_initializer=self.w_init)
net_h0 = batch_norm(net_h0, train=is_training, init=self.batch_norm_init, act=None)
net_h0 = tf.reshape(net_h0, [-1, s16, s16, self.gf_dim * 8])
# Residual layer
net = conv2d(net_h0, self.gf_dim * 2, ks=(1, 1), s=(1, 1), padding='valid', init=self.w_init)
net = batch_norm(net, train=is_training, init=self.batch_norm_init, act=tf.nn.relu)
net = conv2d(net, self.gf_dim * 2, ks=(3, 3), s=(1, 1), init=self.w_init)
net = batch_norm(net, train=is_training, init=self.batch_norm_init, act=tf.nn.relu)
net = conv2d(net, self.gf_dim * 8, ks=(3, 3), s=(1, 1), padding='same', init=self.w_init)
net = batch_norm(net, train=is_training, init=self.batch_norm_init, act=None)
net_h1 = tf.add(net_h0, net)
net_h1 = tf.nn.relu(net_h1)
# --------------------------------------------------------
net_h2 = conv2d_transpose(net_h1, self.gf_dim*4, ks=(4, 4), s=(2, 2), init=self.w_init)
net_h2 = conv2d(net_h2, self.gf_dim*4, ks=(3, 3), s=(1, 1), init=self.w_init)
net_h2 = batch_norm(net_h2, train=is_training, init=self.batch_norm_init, act=None)
# --------------------------------------------------------
# Residual layer
net = conv2d(net_h2, self.gf_dim, ks=(1, 1), s=(1, 1), padding='valid', init=self.w_init)
net = batch_norm(net, train=is_training, init=self.batch_norm_init, act=tf.nn.relu)
net = conv2d(net, self.gf_dim, ks=(3, 3), s=(1, 1), init=self.w_init)
net = batch_norm(net, train=is_training, init=self.batch_norm_init, act=tf.nn.relu)
net = conv2d(net, self.gf_dim*4, ks=(3, 3), s=(1, 1), init=self.w_init)
net = batch_norm(net, train=is_training, init=self.batch_norm_init, act=None)
net_h3 = tf.add(net_h2, net)
net_h3 = tf.nn.relu(net_h3)
# --------------------------------------------------------
net_h4 = conv2d_transpose(net_h3, self.gf_dim*2, ks=(4, 4), s=(2, 2), init=self.w_init)
net_h4 = conv2d(net_h4, self.gf_dim*2, ks=(3, 3), s=(1, 1), init=self.w_init)
net_h4 = batch_norm(net_h4, train=is_training, init=self.batch_norm_init, act=tf.nn.relu)
net_h5 = conv2d_transpose(net_h4, self.gf_dim, ks=(4, 4), s=(2, 2), init=self.w_init)
net_h5 = conv2d(net_h5, self.gf_dim, ks=(3, 3), s=(1, 1), init=self.w_init)
net_h5 = batch_norm(net_h5, train=is_training, init=self.batch_norm_init, act=tf.nn.relu)
net_logits = conv2d_transpose(net_h5, self.image_dims[-1], ks=(4, 4), s=(2, 2), init=self.w_init)
net_logits = conv2d(net_logits, self.image_dims[-1], ks=(3, 3), s=(1, 1), init=self.w_init)
net_output = tf.nn.tanh(net_logits)
return net_output, mean, log_sigma
def generator_encode_image(self, image, is_training=True):
net_h0 = conv2d(image, self.gf_dim, ks=(3, 3), s=(1, 1), act=tf.nn.relu)
net_h1 = conv2d(net_h0, self.gf_dim * 2, ks=(4, 4), s=(2, 2))
net_h1 = batch_norm(net_h1, train=is_training, init=self.batch_norm_init, act=tf.nn.relu)
output_tensor = conv2d(net_h1, self.gf_dim * 4, ks=(4, 4), s=(2, 2))
output_tensor = batch_norm(output_tensor, train=is_training, init=self.batch_norm_init, act=tf.nn.relu)
return output_tensor
def generator_residual_layer(self, input_layer, is_training=True):
net_h0 = input_layer
net_h1 = conv2d(net_h0, self.gf_dim * 4, ks=(4, 4), s=(1, 1))
net_h1 = batch_norm(net_h1, train=is_training, init=self.batch_norm_init, act=tf.nn.relu)
net_h2 = conv2d(net_h1, self.gf_dim * 4, ks=(4, 4), s=(1, 1))
net_h2 = batch_norm(net_h2, train=is_training, init=self.batch_norm_init)
return tf.nn.relu(tf.add(net_h0, net_h2))
def construct_down_block(self, inputs, name, down_outputs, first=False):
num_outputs = self.conf.start_channel_num if first else 2 * \
inputs.shape[self.channel_axis].value
conv1 = ops.conv2d(
inputs, num_outputs, self.conv_size, name+'/conv1')
conv2 = ops.conv2d(
conv1, num_outputs, self.conv_size, name+'/conv2',)
down_outputs.append(conv2)
pool = ops.pool2d(
conv2, self.pool_size, name+'/pool')
return pool
def _make_descriminator(self, input, phase_train):
conv1 = ops.batch_norm(ops.conv2d(input, self.df_dim,
name='d_h0_conv'),
name='d_bn0',
phase_train=phase_train)
h0 = ops.lrelu(conv1)
h1 = ops.lrelu(
ops.batch_norm(ops.conv2d(h0, self.df_dim * 2, name='d_h1_conv'),
name='d_bn1',
phase_train=phase_train))
#h2 = ops.lrelu(ops.batch_norm(ops.conv2d(h1, self.df_dim*4, name='d_h2_conv'), name='d_bn2'))
#h3 = ops.lrelu(ops.batch_norm(ops.conv2d(h2, self.df_dim*8, name='d_h3_conv'), name='d_bn3'))
h2 = ops.lrelu(tf.reshape(h1, [self.batch_size, -1]), 1, 'd_h1_lin')
return h2
def __init__(self,
num_actions,
state,
action=None,
target=None,
learning_rate=None,
scope='DQN'):
# State - Input state to pass through the network
# Action - Action for which the Q value should be predicted (only required for training)
# Target - Target Q value (only required for training)
self.input = state
self.action = action
self.target = target
self.num_actions = num_actions
self.scope = scope
if learning_rate is not None:
self.optimizer = tf.train.RMSPropOptimizer(learning_rate,
momentum=0.95,
epsilon=0.01)
with tf.variable_scope(self.scope):
with tf.variable_scope('input_layers'):
self.input_float = tf.to_float(self.input)
self.input_norm = tf.divide(self.input_float, 255.0)
self.conv1 = conv2d(self.input_norm,
8,
32,
4,
tf.nn.relu,
scope='conv1')
self.conv2 = conv2d(self.conv1,
4,
64,
2,
tf.nn.relu,
scope='conv2')
self.conv3 = conv2d(self.conv2,
3,
64,
1,
tf.nn.relu,
scope='conv3')
self.flatten = flatten(self.conv3, scope='flatten')
self.dense = dense(self.flatten, 512, tf.nn.relu, scope='dense')
self.output = dense(self.dense, self.num_actions, scope='output')
self.network_params = tf.trainable_variables(scope=self.scope)
def generator(self,
z_var,
cond_inp,
stages,
t,
reuse=False,
cond_noise=True):
alpha_trans = self.alpha_tra
with tf.variable_scope('g_net', reuse=reuse):
with tf.variable_scope(self.get_conv_scope_name(0), reuse=reuse):
mean_lr, log_sigma_lr = self.generate_conditionals(cond_inp)
cond = self.sample_normal_conditional(mean_lr, log_sigma_lr,
cond_noise)
# import pdb
# pdb.set_trace()
x = tf.concat([z_var, cond], axis=1)
x = fc(x, units=4 * 4 * self.get_nf(0))
x = layer_norm(x)
x = tf.reshape(x, [-1, 4, 4, self.get_nf(0)])
x = conv2d(x, f=self.get_nf(0), ks=(3, 3), s=(1, 1))
x = layer_norm(x, act=tf.nn.relu)
x = conv2d(x, f=self.get_nf(0), ks=(3, 3), s=(1, 1))
x = layer_norm(x, act=tf.nn.relu)
x_iden = None
for i in range(1, stages):
if (i == stages - 1) and t:
x_iden = self.to_rgb(x, stages - 2)
x_iden = upscale(x_iden, 2)
with tf.variable_scope(self.get_conv_scope_name(i),
reuse=reuse):
x = upscale(x, 2)
x = conv2d(x, f=self.get_nf(i), ks=(3, 3), s=(1, 1))
x = layer_norm(x, act=tf.nn.relu)
x = conv2d(x, f=self.get_nf(i), ks=(3, 3), s=(1, 1))
x = layer_norm(x, act=tf.nn.relu)
x = self.to_rgb(x, stages - 1)
if t:
x = tf.multiply(tf.subtract(1., alpha_trans),
x_iden) + tf.multiply(alpha_trans, x)
return x, mean_lr, log_sigma_lr
def encoder(self, images, is_training, reuse=False):
with tf.variable_scope("generator"):
if reuse:
tf.get_variable_scope().reuse_variables()
encode_layers = dict()
def encode_layer(x, output_filters, layer):
act = lrelu(x)
conv = conv2d(act,
output_filters=output_filters,
scope="g_e%d_conv" % layer)
enc = batch_norm(conv, is_training, scope="g_e%d_bn" % layer)
encode_layers["e%d" % layer] = enc
return enc
e1 = conv2d(images, self.generator_dim, scope="g_e1_conv")
encode_layers["e1"] = e1
e2 = encode_layer(e1, self.generator_dim * 2, 2)
e3 = encode_layer(e2, self.generator_dim * 4, 3)
e4 = encode_layer(e3, self.generator_dim * 8, 4)
e5 = encode_layer(e4, self.generator_dim * 8, 5)
e6 = encode_layer(e5, self.generator_dim * 8, 6)
e7 = encode_layer(e6, self.generator_dim * 8, 7)
e8 = encode_layer(e7, self.generator_dim * 8, 8)
return e8, encode_layers
def generator(self, image, embed, is_training=True, reuse=False, sampler=False):
s = 64
s2, s4, s8, s16 = int(s / 2), int(s / 4), int(s / 8), int(s / 16)
with tf.variable_scope("stageII_g_net", reuse=reuse):
encoded_img = self.generator_encode_image(image, is_training=is_training)
# Sample from the multivariate normal distribution of the embeddings
mean, log_sigma = self.generate_conditionals(embed)
net_embed = self.sample_normal_conditional(mean, log_sigma)
# --------------------------------------------------------
# Concatenate the encoded image and the embeddings
net_embed = tf.expand_dims(tf.expand_dims(net_embed, 1), 1)
net_embed = tf.tile(net_embed, [1, s4, s4, 1])
imgenc_embed = tf.concat([encoded_img, net_embed], 3)
pre_res = conv2d(imgenc_embed, self.gf_dim * 4, ks=(3, 3), s=(1, 1))
pre_res = batch_norm(pre_res, train=is_training, init=self.batch_norm_init, act=tf.nn.relu)
r_block1 = self.generator_residual_layer(pre_res, is_training=is_training)
r_block2 = self.generator_residual_layer(r_block1, is_training=is_training)
r_block3 = self.generator_residual_layer(r_block2, is_training=is_training)
r_block4 = self.generator_residual_layer(r_block3, is_training=is_training)
return self.generator_upsample(r_block4, is_training=is_training), mean, log_sigma
def from_rgb(self, x, stage):
with tf.variable_scope(self.get_rgb_name(stage)):
return conv2d(x,
f=self.get_dnf(stage),
ks=(1, 1),
s=(1, 1),
act=lrelu_act())
def build_down_block(self, inputs, name, first=False, last=False):
if first:
num_outputs = self.conf.start_channel_num
elif last:
num_outputs = inputs.shape[self.channel_axis].value
else:
num_outputs = 2 * inputs.shape[self.channel_axis].value
conv1 = ops.conv2d(inputs, num_outputs, self.conv_size,
name + '/conv1')
conv2 = ops.conv2d(
conv1,
num_outputs,
self.conv_size,
name + '/conv2',
)
pool = ops.pool2d(conv2, self.pool_size, name + '/pool')
return pool
def generator_upsample(self, input_layer, is_training=True):
net_h0 = conv2d_transpose(input_layer,
self.gf_dim * 2,
ks=(4, 4),
init=self.w_init)
net_h0 = conv2d(net_h0, self.gf_dim * 2, ks=(3, 3), s=(1, 1))
net_h0 = batch_norm(net_h0,
train=is_training,
init=self.batch_norm_init,
act=tf.nn.relu)
net_h1 = conv2d_transpose(net_h0,
self.gf_dim,
ks=(4, 4),
init=self.w_init)
net_h1 = conv2d(net_h1, self.gf_dim, ks=(3, 3), s=(1, 1))
net_h1 = batch_norm(net_h1,
train=is_training,
init=self.batch_norm_init,
act=tf.nn.relu)
net_h2 = conv2d_transpose(net_h1,
self.gf_dim // 2,
ks=(4, 4),
init=self.w_init)
net_h2 = conv2d(net_h2, self.gf_dim // 2, ks=(3, 3), s=(1, 1))
net_h2 = batch_norm(net_h2,
train=is_training,
init=self.batch_norm_init,
act=tf.nn.relu)
net_h3 = conv2d_transpose(net_h2,
self.gf_dim // 4,
ks=(4, 4),
init=self.w_init)
net_h3 = conv2d(net_h3, self.gf_dim // 4, ks=(3, 3), s=(1, 1))
net_h3 = batch_norm(net_h3,
train=is_training,
init=self.batch_norm_init,
act=tf.nn.relu)
return conv2d(net_h3,
self.image_dims[-1],
ks=(3, 3),
s=(1, 1),
act=tf.nn.tanh)
def encode_layer(x, output_filters, layer):
act = lrelu(x)
conv = conv2d(act,
output_filters=output_filters,
scope="g_e%d_conv" % layer)
enc = batch_norm(conv, is_training, scope="g_e%d_bn" % layer)
encode_layers["e%d" % layer] = enc
return enc
def inference(self, images):
cur_out_num = self.conf.ch_num
outs = ops.conv2d(
images, cur_out_num, (3, 3), 'conv_s', train=self.conf.is_train,
stride=2, act_fn=None, data_format=self.conf.data_format)
cur_out_num *= 2
cur_outs = ops.dw_block( # 112 * 112 * 64
outs, cur_out_num, 1, 'conv_1_0', self.conf.keep_r,
self.conf.is_train, data_format=self.conf.data_format)
outs = tf.concat([outs, cur_outs], axis=1, name='add0')
cur_out_num *= 2
outs = ops.dw_block( # 56 * 56 * 128
outs, cur_out_num, 2, 'conv_1_1', self.conf.keep_r,
self.conf.is_train, data_format=self.conf.data_format)
cur_outs = ops.dw_block( # 56 * 56 * 128
outs, cur_out_num, 1, 'conv_1_2', self.conf.keep_r,
self.conf.is_train, data_format=self.conf.data_format)
outs = tf.concat([outs, cur_outs], axis=1, name='add1')
#outs = tf.add(outs, cur_outs, name='add1')
cur_out_num *= 2
outs = ops.dw_block( # 28 * 28 * 256
outs, cur_out_num, 2, 'conv_1_3', self.conf.keep_r,
self.conf.is_train, data_format=self.conf.data_format)
#cur_out_num *= 2
cur_outs = ops.dw_block( # 28 * 28 * 256
outs, cur_out_num, 1, 'conv_1_4', self.conf.keep_r,
self.conf.is_train, data_format=self.conf.data_format)
outs = tf.concat([outs, cur_outs], axis=1, name='add2')
cur_out_num *= 2
outs = ops.dw_block( # 14 * 14 * 512
outs, cur_out_num, 2, 'conv_1_5', self.conf.keep_r,
self.conf.is_train, data_format=self.conf.data_format)
cur_outs = ops.simple_group_block( # 14 * 14 * 512
outs, self.conf.block_num, self.conf.keep_r, self.conf.is_train,
'conv_2_1', self.conf.data_format, self.conf.group_num)
outs = tf.add(outs, cur_outs, name='add21')
outs = self.get_block_func()( # 14 * 14 * 512
outs, self.conf.block_num, self.conf.keep_r, self.conf.is_train,
'conv_2_2', self.conf.data_format, self.conf.group_num)
#outs = tf.add(outs, cur_outs, name='add22')
cur_outs = self.get_block_func()( # 14 * 14 * 512
outs, self.conf.block_num, self.conf.keep_r, self.conf.is_train,
'conv_2_3', self.conf.data_format, self.conf.group_num)
outs = tf.add(outs, cur_outs, name='add23')
cur_out_num *= 2
outs = ops.dw_block( # 7 * 7 * 1024
outs, cur_out_num, 2, 'conv_3_0', self.conf.keep_r,
self.conf.is_train, data_format=self.conf.data_format)
outs = ops.dw_block( # 7 * 7 * 1024
outs, cur_out_num, 1, 'conv_3_1', self.conf.keep_r,
self.conf.is_train, self.conf.use_rev_conv,
self.conf.rev_kernel_size,
#act_fn=None,
data_format=self.conf.data_format)
outs = self.get_out_func()(
outs, 'out', self.conf.class_num, self.conf.is_train,
data_format=self.conf.data_format)
return outs
def construct_up_block(self, inputs, down_inputs, name, final=False):
num_outputs = inputs.shape[self.channel_axis].value
conv1 = self.deconv_func()(inputs, num_outputs, self.conv_size,
name + '/conv1')
conv1 = tf.concat([conv1, down_inputs],
self.channel_axis,
name=name + '/concat')
conv2 = self.conv_func()(conv1, num_outputs, self.conv_size,
name + '/conv2')
num_outputs = self.conf.class_num if final else num_outputs / 2
conv3 = ops.conv2d(conv2, num_outputs, self.conv_size, name + '/conv3')
return conv3
def encode_layer(x, output_filters, layer, keep_rate=1.0):
# act = lrelu(x)
enc = tf.nn.relu(x)
enc = tf.nn.dropout(enc, keep_rate)
enc = conv2d(enc,
output_filters=output_filters,
scope="g_e%d_conv" % layer)
# batch norm is important for ae, or aw would output nothing!!!
enc = batch_norm(enc, is_training, scope="g_e%d_bn" % layer)
encode_layers["e%d" % layer] = enc
return enc
def __init__(self, sess, img_height, img_width, c_dim, a_dim, reuse=False):
self.sess = sess
with tf.variable_scope("policy", reuse=reuse):
#ob_ph: (mb_size, img_height, img_width, c_dim)
self.ob_ph = tf.placeholder(tf.uint8,
[None, img_height, img_width, c_dim],
name="observation")
ob_normalized = tf.cast(self.ob_ph, tf.float32) / 255.0
#conv1: (mb_size, img_height1, img_width1, 32)
h = ops.conv2d(ob_normalized, 32, 8, 8, 4, 4, name="conv1")
h = tf.nn.relu(h)
#conv2: (mb_size, img_height2, img_width2, 64)
h = ops.conv2d(h, 64, 4, 4, 2, 2, name="conv2")
h = tf.nn.relu(h)
#conv3: (mb_size, img_height3, img_width3, 64)
h = ops.conv2d(h, 64, 3, 3, 1, 1, name="conv3")
h = tf.nn.relu(h)
#fc: (mb_size, 512)
h = ops.fc(tf.reshape(h,
[-1, h.shape[1] * h.shape[2] * h.shape[3]]),
512,
name="fc1")
h = tf.nn.relu(h)
#pi: (mb_size, a_dim)
#value: (mb_size, 1)
pi = ops.fc(h, a_dim, name="fc_pi")
value = ops.fc(h, 1, name="fc_value")
#value: (mb_size)
#action: (mb_size)
self.value = value[:, 0]
self.cat_dist = tf.distributions.Categorical(pi)
self.action = self.cat_dist.sample(1)[0]
self.pi = pi
def __init__(self, sess, img_height, img_width, c_dim, a_dim, name="policy", reuse=False):
self.sess = sess
with tf.variable_scope(name, reuse=reuse):
#ob_ph: (mb_size, s_dim)
self.ob_ph = tf.placeholder(tf.uint8, [None, img_height, img_width, c_dim], name="observation")
ob_normalized = tf.cast(self.ob_ph, tf.float32) / 255.0
#conv1: (mb_size, img_height1, img_width1, 32)
h = ops.conv2d(ob_normalized, 32, 8, 8, 4, 4, name="conv1")
h = tf.nn.relu(h)
#conv2: (mb_size, img_height2, img_width2, 64)
h = ops.conv2d(h, 64, 4, 4, 2, 2, name="conv2")
h = tf.nn.relu(h)
#conv3: (mb_size, img_height3, img_width3, 64)
h = ops.conv2d(h, 64, 3, 3, 1, 1, name="conv3")
h = tf.nn.relu(h)
#fc: (mb_size, 512)
h = ops.fc(tf.reshape(h, [-1, h.shape[1]*h.shape[2]*h.shape[3]]), 512, name="fc1")
h = tf.nn.relu(h)
with tf.variable_scope("actor", reuse=reuse):
#fc_logits: (mb_size, a_dim)
logits = ops.fc(h, a_dim, name="a_fc_logits")
with tf.variable_scope("critic", reuse=reuse):
#value: (mb_size, 1)
value = ops.fc(h, 1, name="c_fc_value")
#value: (mb_size)
#action: (mb_size)
#neg_logprob: (mb_size)
self.value = value[:, 0]
self.distrib = distribs.CategoricalDistrib(logits)
self.action = self.distrib.sample()
self.neg_logprob = self.distrib.neg_logp(self.action)
def build_down_block(self,
inputs,
name,
layer_index=0,
first=False,
last=False):
if first:
num_outputs = self.conf.start_channel_num
elif last:
num_outputs = inputs.shape[self.channel_axis].value
else:
num_outputs = 2 * inputs.shape[self.channel_axis].value
print("drop out keep probabilities on layer %d: " % (layer_index),
self.conf.drop_outs[layer_index])
conv1 = ops.conv2d(inputs, num_outputs, self.conv_size,
name + '/conv1')
conv1 = ops.dropout(conv1, self.conf.drop_outs[layer_index][0],
name + '/dropout1')
conv2 = ops.conv2d(conv1, num_outputs, self.conv_size, name + '/conv2')
conv2 = ops.dropout(conv2, self.conf.drop_outs[layer_index][1],
name + '/dropout2')
if layer_index > 1:
conv3 = ops.conv2d(conv2, num_outputs, self.conv_size,
name + '/conv3')
conv3 = ops.dropout(conv3, self.conf.drop_outs[layer_index][2],
name + '/dropout3')
# conv4 = ops.conv2d(conv3, num_outputs, self.conv_size, name+'/conv4',self.regularizer)
# conv4 = ops.dropout(conv4,self.conf.drop_outs[layer_index][3],name+'/dropout4')
pool = ops.pool2d(conv3, self.pool_size, name + '/pool')
pool = ops.dropout(pool, self.conf.drop_outs[layer_index][3],
name + '/dropout_pool')
else:
pool = ops.pool2d(conv2, self.pool_size, name + '/pool')
pool = ops.dropout(pool, self.conf.drop_outs[layer_index][2],
name + '/dropout_pool')
return pool
def discriminator(self, inputs, embed, is_training=True, reuse=False):
s16 = self.output_size / 16
lrelu = lambda l: tf.nn.leaky_relu(l, 0.2)
with tf.variable_scope("d_net", reuse=reuse):
net_ho = conv2d(inputs, self.df_dim, ks=(4, 4), s=(2, 2), act=lrelu, init=self.w_init)
net_h1 = conv2d(net_ho, self.df_dim * 2, ks=(4, 4), s=(2, 2), init=self.w_init)
net_h1 = batch_norm(net_h1, train=is_training, init=self.batch_norm_init, act=lrelu)
net_h2 = conv2d(net_h1, self.df_dim * 4, ks=(4, 4), s=(2, 2), init=self.w_init)
net_h2 = batch_norm(net_h2, train=is_training, init=self.batch_norm_init, act=lrelu)
net_h3 = conv2d(net_h2, self.df_dim * 8, ks=(4, 4), s=(2, 2), init=self.w_init)
net_h3 = batch_norm(net_h3, train=is_training, init=self.batch_norm_init)
# --------------------------------------------------------
# Residual layer
net = conv2d(net_h3, self.df_dim * 2, ks=(1, 1), s=(1, 1), padding='valid', init=self.w_init)
net = batch_norm(net, train=is_training, init=self.batch_norm_init, act=lrelu)
net = conv2d(net, self.df_dim * 2, ks=(3, 3), s=(1, 1), init=self.w_init)
net = batch_norm(net, train=is_training, init=self.batch_norm_init, act=lrelu)
net = conv2d(net, self.df_dim * 8, ks=(3, 3), s=(1, 1), init=self.w_init)
net = batch_norm(net, train=is_training, init=self.batch_norm_init)
net_h4 = tf.add(net_h3, net)
net_h4 = tf.nn.leaky_relu(net_h4, 0.2)
# --------------------------------------------------------
# Compress embeddings
net_embed = tf.layers.dense(embed, units=self.compressed_embed_dim, activation=lrelu)
# Append embeddings in depth
net_embed = tf.expand_dims(tf.expand_dims(net_embed, 1), 1)
net_embed = tf.tile(net_embed, [1, 4, 4, 1])
net_h4_concat = tf.concat([net_h4, net_embed], 3)
net_h4 = conv2d(net_h4_concat, self.df_dim * 8, ks=(1, 1), s=(1, 1), padding='valid', init=self.w_init)
net_h4 = batch_norm(net_h4, train=is_training, init=self.batch_norm_init, act=lrelu)
net_logits = conv2d(net_h4, 1, ks=(s16, s16), s=(s16, s16), padding='valid', init=self.w_init)
return tf.nn.sigmoid(net_logits), net_logits
def encoder(self, images, is_training, reuse=False):
"""
Encoder network
:param images:
:param is_training:
:param reuse:
:return:
"""
with tf.variable_scope("generator"):
if reuse:
tf.get_variable_scope().reuse_variables()
encode_layers = dict()
def encode_layer(x, output_filters, layer, keep_rate=1.0):
# act = lrelu(x)
enc = tf.nn.relu(x)
enc = tf.nn.dropout(enc, keep_rate)
enc = conv2d(enc,
output_filters=output_filters,
scope="g_e%d_conv" % layer)
# batch norm is important for ae, or aw would output nothing!!!
enc = batch_norm(enc, is_training, scope="g_e%d_bn" % layer)
encode_layers["e%d" % layer] = enc
return enc
e1 = conv2d(images, self.network_dim,
scope="g_e1_env") # 128 x 128
encode_layers["e1"] = e1
e2 = encode_layer(e1, self.network_dim * 2, 2) # 64 x 64
e3 = encode_layer(e2, self.network_dim * 4, 3) # 32 x 32
e4 = encode_layer(e3, self.network_dim * 8, 4) # 16 x 16
e5 = encode_layer(e4, self.network_dim * 8, 5) # 8 x 8
e6 = encode_layer(e5, self.network_dim * 8, 6) # 4 x 4
e7 = encode_layer(e6, self.network_dim * 8, 7) # 2 x 2
e8 = encode_layer(e7, self.network_dim * 8, 8) # 1 x 1
return e8, encode_layers
def __call__(self, speech_features, is_training, reuse=False):
with tf.variable_scope(self.scope, reuse=reuse):
if reuse == True:
tf.get_variable_scope().reuse_variables()
batch_norm = BatchNorm(epsilon=1e-5, momentum=0.9)
speech_features = batch_norm(speech_features,
reuse=reuse,
is_training=is_training)
speech_feature_shape = speech_features.get_shape().as_list()
speech_features_reshaped = tf.reshape(tensor=speech_features,
shape=[
-1,
speech_feature_shape[1],
1,
speech_feature_shape[2]
])
'''
condition_shape = condition.get_shape().as_list()
condition_reshaped = tf.reshape(tensor=condition, shape=[-1, condition_shape[1]])
if self._condition_speech_features:
#Condition input speech feature windows
speech_feature_condition = tf.transpose(tf.reshape(tensor=condition_reshaped, shape=[-1, condition_shape[1], 1, 1]), perm=[0,2,3,1])
speech_feature_condition = tf.tile(speech_feature_condition, [1, speech_feature_shape[1], 1, 1])
speech_features_reshaped = tf.concat((speech_features_reshaped, speech_feature_condition), axis=-1, name='conditioning_speech_features')
'''
factor = self._speech_encoder_size_factor
with tf.name_scope('conv1_time'):
'''
conv1_time = tf.nn.relu(conv2d(inputs=speech_features_reshaped,
n_filters=int(32*factor),
k_h=3, k_w=1,
stride_h=2, stride_w=1,
activation=tf.identity,
scope='conv1'))
'''
conv1_time = conv2d(inputs=speech_features_reshaped,
n_filters=int(32 * factor),
k_h=3,
k_w=1,
stride_h=2,
stride_w=1,
activation=tf.identity,
scope='conv1')
conv1_time = tf.nn.relu(
tf.contrib.layers.batch_norm(conv1_time,
decay=0.9,
updates_collections=None,
epsilon=1e-5,
center=True,
scale=True,
is_training=is_training,
reuse=reuse,
scope='BN1'))
with tf.name_scope('conv2_time'):
'''
conv2_time = tf.nn.relu(conv2d(inputs=conv1_time,
n_filters=int(32*factor),
k_h=3, k_w=1,
stride_h=2, stride_w=1,
activation=tf.identity,
scope='conv2'))
'''
conv2_time = conv2d(inputs=conv1_time,
n_filters=int(32 * factor),
k_h=3,
k_w=1,
stride_h=2,
stride_w=1,
activation=tf.identity,
scope='conv2')
conv2_time = tf.nn.relu(
tf.contrib.layers.batch_norm(conv2_time,
decay=0.9,
updates_collections=None,
epsilon=1e-5,
center=True,
scale=True,
is_training=is_training,
reuse=reuse,
scope='BN2'))
with tf.name_scope('conv3_time'):
'''
conv3_time = tf.nn.relu(conv2d(inputs=conv2_time,
n_filters=int(64*factor),
k_h=3, k_w=1,
stride_h=2, stride_w=1,
activation=tf.identity,
scope='conv3'))
'''
conv3_time = conv2d(inputs=conv2_time,
n_filters=int(64 * factor),
k_h=3,
k_w=1,
stride_h=2,
stride_w=1,
activation=tf.identity,
scope='conv3')
conv3_time = tf.nn.relu(
tf.contrib.layers.batch_norm(conv3_time,
decay=0.9,
updates_collections=None,
epsilon=1e-5,
center=True,
scale=True,
is_training=is_training,
reuse=reuse,
scope='BN3'))
with tf.name_scope('conv4_time'):
conv4_time = tf.nn.relu(
conv2d(inputs=conv3_time,
n_filters=int(64 * factor),
k_h=3,
k_w=1,
stride_h=2,
stride_w=1,
activation=tf.identity,
scope='conv4'))
previous_shape = conv4_time.get_shape().as_list()
time_conv_flattened = tf.reshape(conv4_time, [
-1, previous_shape[1] * previous_shape[2] * previous_shape[3]
])
#Condition audio encoding on speaker style
with tf.name_scope('concat_audio_embedding'):
#concatenated = tf.concat((time_conv_flattened, condition_reshaped), axis=1, name='conditioning_audio_embedding')
concatenated = time_conv_flattened
units_in = concatenated.get_shape().as_list()[1]
with tf.name_scope('fc1'):
fc1 = tf.nn.relu(
fc_layer(concatenated,
num_units_in=units_in,
num_units_out=128,
scope='fc1'))
fc1 = tf.layers.dropout(fc1, rate=0.2, training=is_training)
with tf.name_scope('fc2'):
fc2 = tf.nn.relu(
fc_layer(fc1,
num_units_in=128,
num_units_out=self._speech_encoding_dim,
scope='fc2'))
fc2 = tf.layers.dropout(fc2, rate=0.2, training=is_training)
return fc2
def generator(input_, angles, reuse=False):
""" Generator.
Parameters
----------
input_: tensor, input images.
angles: tensor, target gaze direction.
reuse: bool, reuse the net if True.
Returns
-------
x: tensor, generated image.
"""
channel = 64
style_dim = angles.get_shape().as_list()[-1]
angles_reshaped = tf.reshape(angles, [-1, 1, 1, style_dim])
angles_tiled = tf.tile(angles_reshaped,
[1, tf.shape(input_)[1],
tf.shape(input_)[2], 1])
x = tf.concat([input_, angles_tiled], axis=3)
with tf.compat.v1.variable_scope('generator', reuse=reuse):
# input layer
x = conv2d(x,
channel,
d_h=1,
d_w=1,
scope='conv2d_input',
use_bias=False,
pad=3,
conv_filters_dim=7)
x = instance_norm(x, scope='in_input')
x = relu(x)
# encoder
for i in range(2):
x = conv2d(x,
2 * channel,
d_h=2,
d_w=2,
scope='conv2d_%d' % i,
use_bias=False,
pad=1,
conv_filters_dim=4)
x = instance_norm(x, scope='in_conv_%d' % i)
x = relu(x)
channel = 2 * channel
# bottleneck
for i in range(6):
x_a = conv2d(x,
channel,
conv_filters_dim=3,
d_h=1,
d_w=1,
pad=1,
use_bias=False,
scope='conv_res_a_%d' % i)
x_a = instance_norm(x_a, 'in_res_a_%d' % i)
x_a = relu(x_a)
x_b = conv2d(x_a,
channel,
conv_filters_dim=3,
d_h=1,
d_w=1,
pad=1,
use_bias=False,
scope='conv_res_b_%d' % i)
x_b = instance_norm(x_b, 'in_res_b_%d' % i)
x = x + x_b
# decoder
for i in range(2):
x = deconv2d(x,
int(channel / 2),
conv_filters_dim=4,
d_h=2,
d_w=2,
use_bias=False,
scope='deconv_%d' % i)
x = instance_norm(x, scope='in_decon_%d' % i)
x = relu(x)
channel = int(channel / 2)
x = conv2d(x,
3,
conv_filters_dim=7,
d_h=1,
d_w=1,
pad=3,
use_bias=False,
scope='output')
x = tanh(x)
return x
def discriminator(params, x_init, reuse=False):
""" Discriminator.
Parameters
----------
params: dict.
x_init: input tensor.
reuse: bool, reuse the net if True.
Returns
-------
x_gan: tensor, outputs for adversarial training.
x_reg: tensor, outputs for gaze estimation.
"""
layers = 5
channel = 64
image_size = params.image_size
with tf.compat.v1.variable_scope('discriminator', reuse=reuse):
# 64 3 -> 32 64 -> 16 128 -> 8 256 -> 4 512 -> 2 1024
x = conv2d(x_init,
channel,
conv_filters_dim=4,
d_h=2,
d_w=2,
scope='conv_0',
pad=1,
use_bias=True)
x = lrelu(x)
for i in range(1, layers):
x = conv2d(x,
channel * 2,
conv_filters_dim=4,
d_h=2,
d_w=2,
scope='conv_%d' % i,
pad=1,
use_bias=True)
x = lrelu(x)
channel = channel * 2
filter_size = int(image_size / 2**layers)
x_gan = conv2d(x,
1,
conv_filters_dim=filter_size,
d_h=1,
d_w=1,
pad=1,
scope='conv_logit_gan',
use_bias=False)
x_reg = conv2d(x,
2,
conv_filters_dim=filter_size,
d_h=1,
d_w=1,
pad=0,
scope='conv_logit_reg',
use_bias=False)
x_reg = tf.reshape(x_reg, [-1, 2])
return x_gan, x_reg
def logits(self, x_onehot):
batch_size = self.batch_size
seq_len = self.seq_len
vocab_size = self.vocab_size
dis_emb_dim = self.dis_emb_dim
num_rep = self.num_rep
sn = self.sn
# get the embedding dimension for each presentation
emb_dim_single = int(dis_emb_dim / num_rep)
assert isinstance(emb_dim_single, int) and emb_dim_single > 0
filter_sizes = [2, 3, 4, 5]
num_filters = [300, 300, 300, 300]
dropout_keep_prob = 0.75
d_embeddings = tf.get_variable('d_emb', shape=[vocab_size, dis_emb_dim],
initializer=create_linear_initializer(vocab_size))
input_x_re = tf.reshape(x_onehot, [-1, vocab_size])
emb_x_re = tf.matmul(input_x_re, d_embeddings)
# batch_size x seq_len x dis_emb_dim
emb_x = tf.reshape(emb_x_re, [batch_size, seq_len, dis_emb_dim])
# batch_size x seq_len x dis_emb_dim x 1
emb_x_expanded = tf.expand_dims(emb_x, -1)
# print('shape of emb_x_expanded: {}'.format(
# emb_x_expanded.get_shape().as_list()))
# Create a convolution + maxpool layer for each filter size
pooled_outputs = []
for filter_size, num_filter in zip(filter_sizes, num_filters):
conv = conv2d(emb_x_expanded, num_filter, k_h=filter_size, k_w=emb_dim_single,
d_h=1, d_w=emb_dim_single, sn=sn, stddev=None, padding='VALID',
scope="conv-%s" % filter_size) # batch_size x (seq_len-k_h+1) x num_rep x num_filter
out = tf.nn.relu(conv, name="relu_new")
pooled = tf.nn.max_pool(out, ksize=[1, seq_len - filter_size + 1, 1, 1],
strides=[1, 1, 1, 1], padding='VALID',
name="pool_new") # batch_size x 1 x num_rep x num_filter
pooled_outputs.append(pooled)
# Combine all the pooled features
num_filters_total = sum(num_filters)
# batch_size x 1 x num_rep x num_filters_total
h_pool = tf.concat(pooled_outputs, 3)
# print('shape of h_pool: {}'.format(h_pool.get_shape().as_list()))
h_pool_flat = tf.reshape(h_pool, [-1, num_filters_total])
# Add highway
# (batch_size*num_rep) x num_filters_total
h_highway = highway(h_pool_flat, h_pool_flat.get_shape()[1], 1, 0)
# Add dropout
h_drop = tf.nn.dropout(h_highway, dropout_keep_prob, name='dropout_new')
# fc
fc_out = linear(h_drop, output_size=100,
use_bias=True, sn=sn, scope='fc_new')
logits = linear(fc_out, output_size=1,
use_bias=True, sn=sn, scope='logits_new')
logits = tf.squeeze(logits, -1) # batch_size*num_rep
return logits
def forward_pass(self,
state_in,
reshape=True,
sigmoid_out=False,
reuse=None):
self.state_in = state_in
shape_in = self.state_in.get_shape().as_list()
# Get number of input channels for weight/bias init
channels_in = shape_in[-1]
with tf.variable_scope(self.scope, reuse=reuse):
if reshape:
# Reshape [batch_size, traj_len, H, W, C] into [batch_size*traj_len, H, W, C]
self.state_in = tf.reshape(
self.state_in, [-1, shape_in[2], shape_in[3], shape_in[4]])
self.conv1 = conv2d(
self.state_in,
self.num_filters,
self.kernels[0],
self.strides[0],
kernel_init=tf.random_uniform_initializer((-1.0 / tf.sqrt(
float(channels_in * self.kernels[0] * self.kernels[0]))), (
1.0 / tf.sqrt(
float(channels_in * self.kernels[0] *
self.kernels[0])))),
bias_init=tf.random_uniform_initializer((-1.0 / tf.sqrt(
float(channels_in * self.kernels[0] * self.kernels[0]))), (
1.0 / tf.sqrt(
float(channels_in * self.kernels[0] *
self.kernels[0])))),
scope='conv1')
self.conv1 = lrelu(self.conv1, self.lrelu_alpha, scope='conv1')
self.conv2 = conv2d(
self.conv1,
self.num_filters,
self.kernels[1],
self.strides[1],
kernel_init=tf.random_uniform_initializer((-1.0 / tf.sqrt(
float(self.num_filters * self.kernels[1] *
self.kernels[1]))), (1.0 / tf.sqrt(
float(self.num_filters * self.kernels[1] *
self.kernels[1])))),
bias_init=tf.random_uniform_initializer((-1.0 / tf.sqrt(
float(self.num_filters * self.kernels[1] *
self.kernels[1]))), (1.0 / tf.sqrt(
float(self.num_filters * self.kernels[1] *
self.kernels[1])))),
scope='conv2')
self.conv2 = lrelu(self.conv2, self.lrelu_alpha, scope='conv2')
self.conv3 = conv2d(
self.conv2,
self.num_filters,
self.kernels[2],
self.strides[2],
kernel_init=tf.random_uniform_initializer((-1.0 / tf.sqrt(
float(self.num_filters * self.kernels[2] *
self.kernels[2]))), (1.0 / tf.sqrt(
float(self.num_filters * self.kernels[2] *
self.kernels[2])))),
bias_init=tf.random_uniform_initializer((-1.0 / tf.sqrt(
float(self.num_filters * self.kernels[2] *
self.kernels[2]))), (1.0 / tf.sqrt(
float(self.num_filters * self.kernels[2] *
self.kernels[2])))),
scope='conv3')
self.conv3 = lrelu(self.conv3, self.lrelu_alpha, scope='conv3')
self.conv4 = conv2d(
self.conv3,
self.num_filters,
self.kernels[3],
self.strides[3],
kernel_init=tf.random_uniform_initializer((-1.0 / tf.sqrt(
float(self.num_filters * self.kernels[3] *
self.kernels[3]))), (1.0 / tf.sqrt(
float(self.num_filters * self.kernels[3] *
self.kernels[3])))),
bias_init=tf.random_uniform_initializer((-1.0 / tf.sqrt(
float(self.num_filters * self.kernels[3] *
self.kernels[3]))), (1.0 / tf.sqrt(
float(self.num_filters * self.kernels[3] *
self.kernels[3])))),
scope='conv4')
self.conv4 = lrelu(self.conv4, self.lrelu_alpha, scope='conv4')
self.flatten = flatten(self.conv4)
self.dense = dense(self.flatten,
self.dense_size,
kernel_init=tf.random_uniform_initializer(
(-1.0 / tf.sqrt(float(self.num_filters))),
(1.0 / tf.sqrt(float(self.num_filters)))),
bias_init=tf.random_uniform_initializer(
(-1.0 / tf.sqrt(float(self.num_filters))),
(1.0 / tf.sqrt(float(self.num_filters)))))
self.output = dense(self.dense,
1,
kernel_init=tf.random_uniform_initializer(
(-1.0 / tf.sqrt(float(self.dense_size))),
(1.0 / tf.sqrt(float(self.dense_size)))),
bias_init=tf.random_uniform_initializer(
(-1.0 / tf.sqrt(float(self.dense_size))),
(1.0 / tf.sqrt(float(self.dense_size)))),
scope='output')
if sigmoid_out:
self.output = tf.nn.sigmoid(self.output)
if reshape:
# Reshape 1d reward output [batch_size*traj_len] into batches [batch_size, traj_len]
self.output = tf.reshape(self.output, [-1, shape_in[1]])
self.network_params = tf.trainable_variables(scope=self.scope)
return self.output
def to_rgb(self, x, stage):
with tf.variable_scope(self.get_rgb_name(stage)):
x = conv2d(x, f=9, ks=(2, 2), s=(1, 1), act=tf.nn.relu)
x = conv2d(x, f=3, ks=(1, 1), s=(1, 1))
return x