def build_layers(self,
state,
action,
c_names,
units_1,
units_2,
w_i,
b_i,
reg=None):
with tf.variable_scope('conv1'):
conv1 = conv(state, [5, 5, 3, 6], [6], [1, 2, 2, 1], w_i, b_i)
with tf.variable_scope('conv2'):
conv2 = conv(conv1, [3, 3, 6, 12], [12], [1, 2, 2, 1], w_i, b_i)
with tf.variable_scope('flatten'):
flatten = tf.contrib.layers.flatten(conv2)
with tf.variable_scope('dense1'):
dense1 = dense(flatten, units_1, [units_1], w_i, b_i)
with tf.variable_scope('dense2'):
dense2 = dense(dense1, units_2, [units_2], w_i, b_i)
with tf.variable_scope('concat'):
concatenated = tf.concat([dense2, tf.cast(action, tf.float32)], 1)
with tf.variable_scope('dense3'):
dense3 = dense(concatenated, self.atoms, [self.atoms], w_i,
b_i) # 返回
return tf.nn.softmax(dense3)
def _make_discriminator(inputs, data_shape, nplanes=128):
diff_input = inputs['diff_img']
keep_prob = inputs['keep_prob']
g1_weights = tf.get_variable('g1_weights', [5, 5, 3, nplanes])
g1_bias = tf.get_variable('g1_bias', [nplanes])
g1 = tf.nn.relu(utils.conv2d(g1_weights, bias=g1_bias)(diff_input))
g2_weights = tf.get_variable('g2_weights', [5, 5, nplanes, nplanes])
g2_bias = tf.get_variable('g2_bias', [nplanes])
g2 = tf.nn.relu(utils.conv2d(g2_weights, bias=g2_bias)(g1))
flattened = tf.reshape(g2, [-1, data_shape[0] * data_shape[1] * nplanes])
dropout = tf.nn.dropout(flattened, keep_prob)
prob_weights = tf.get_variable(
'prob_weights', [data_shape[0] * data_shape[1] * nplanes, 1])
prob_biases = tf.get_variable('prob_bias', [1])
y_logits = utils.dense(prob_weights, bias=prob_biases)(dropout)
y = tf.nn.sigmoid(y_logits, name='prob_real')
class_weights = tf.get_variable(
'class_weights', [data_shape[0] * data_shape[1] * nplanes, 10])
class_biases = tf.get_variable('class_bias', [1])
class_logits = utils.dense(class_weights, bias=class_biases)(dropout)
class_prob = tf.nn.softmax(class_logits, name='logits_class')
return {
'prob_real': y,
'logits_real': y_logits,
'prob_class': class_prob,
'logits_class': class_logits
}
def __call__(self, x, y=None, sn=False, is_training=True, reuse=False):
with tf.variable_scope(self.name, reuse=reuse):
batch_size = x.get_shape().as_list()[0]
if y is not None:
ydim = y.get_shape().as_list()[-1]
y = tf.reshape(y, [batch_size, 1, 1, ydim])
x = conv_cond_concat(x, y) # [bz, 28, 28, 11]
x = tf.reshape(x, (batch_size, -1))
net = lrelu(dense(x, 512, sn=sn, name='d_fc1'), name='d_l1')
net = lrelu(bn(dense(net, 256, sn=sn, name='d_fc2'),
is_training,
name='d_bn2'),
name='d_l2')
net = lrelu(bn(dense(net, 128, sn=sn, name='d_fc3'),
is_training,
name='d_bn3'),
name='d_l3')
yd = dense(net, 1, sn=sn, name="D_dense")
if self.class_num:
yc = dense(net, self.class_num, sn=sn, name='C_dense')
return yd, net, yc
else:
return yd, net
def forward(x): # input shape (b, d)
h_d0 = tf.expand_dims(x, 1)
h_d1 = ff.dense(h_d0,
V_d1[0] + V_d1[1] * V_d1[2],
activation=tf.nn.leaky_relu) #*V_d1[2]
h_d2 = ff.dense(h_d1, V_d2[0] + V_d2[1] * V_d2[2]) #*V_d2[2]
return tf.reduce_sum(h_d2, 1)
def build_disc0(x, testing=False, reuse=False):
disc0_l1 = x + tf.random_normal(shape=tf.shape(x), stddev=0.05)
disc0_l1 = utils.lrelu(
utils.conv2d(disc0_l1, (3, 3, 3, 96), name='disc0_conv1'))
# 32 x 32 --> 16 x 16
disc0_l2 = utils.batch_normalization(utils.lrelu(
utils.conv2d(disc0_l1, (3, 3, 96, 96),
stride=[1, 2, 2, 1],
name='disc0_conv2')),
name='bn1',
reuse=reuse)
disc0_l2 = tf.nn.dropout(disc0_l2, 0.1 if testing else 1)
# 16 x 16 --> 8x8
disc0_l3 = utils.batch_normalization(utils.lrelu(
utils.conv2d(disc0_l2, (3, 3, 96, 192),
stride=[1, 2, 2, 1],
name='disc0_conv3')),
name='bn2',
reuse=reuse)
# 8x8 --> 8x8
disc0_l4 = utils.batch_normalization(utils.lrelu(
utils.conv2d(disc0_l3, (3, 3, 192, 192), name='disc0_conv4')),
name='bn3',
reuse=reuse)
disc0_l4 = tf.nn.dropout(disc0_l4, 0.1 if testing else 1)
# 8x8 --> 6x6
disc0_l5 = tf.layers.batch_normalization(utils.lrelu(
utils.conv2d(disc0_l4, (3, 3, 192, 192),
padding='VALID',
name='disc0_conv5')),
name='bn4',
reuse=reuse)
disc0_l5 = tf.reshape(disc0_l5, [100, 6, 6, 192])
disc0_shared = utils.lrelu(
utils.network_in_network(disc0_l5,
192,
num_units=192,
name='disc0_shared'))
disc0_shared_flat = tf.reshape(disc0_shared, [-1, 192 * 6 * 6])
disc0_z_recon = utils.dense(disc0_shared_flat,
num_inputs=192 * 6 * 6,
num_units=16,
name='disc0_z_recon')
disc0_shared_pool = tf.reduce_mean(disc0_shared, [1, 2])
disc0_adv = utils.dense(disc0_shared_pool,
num_inputs=192,
num_units=10,
name='disc1_z_adv')
# disc0_adv is the pre-softmax classification output for the discriminator
return disc0_adv, disc0_z_recon
def build_gan(self):
print 'source_Y', self.source_Y.get_shape().as_list()
print 'enX_outputs', self.enX_outputs.get_shape().as_list()
_, D_state_real = self.discriminator(self.source_Y, self.len_Y, reuse=False)
_, D_state_fake = self.discriminator(self.enX_outputs, self.len_X, reuse=True)
self.D_logit_real = dense(D_state_real[-1].c, 1, name="D_dense")
self.D_logit_fake = dense(D_state_fake[-1].c, 1, name="D_dense", reuse=True)
def discriminator(inputs, reuse=False):
with tf.variable_scope('discriminator', reuse=reuse) as d_vs:
layer = dense(inputs, h_dim, activation=tf.nn.elu)
layer = dense(layer, h_dim, activation=tf.nn.elu, batch_residual=True)
layer = dense(layer, h_dim, activation=tf.nn.elu, batch_residual=True)
d_logit = dense(layer, 1, activation=None)
d_prob = tf.nn.sigmoid(d_logit)
d_vars = tf.contrib.framework.get_variables(d_vs)
return d_prob, d_logit, d_vars
def Generator(z, reuse=False):
with tf.variable_scope('generator', reuse=reuse) as g_vs:
layer = dense(z, h_dim, activation=tf.nn.elu)
layer = dense(layer, h_dim, activation=tf.nn.elu, batch_residual=False)
layer = dense(layer, h_dim, activation=tf.nn.elu, batch_residual=False,
dropout=0.9)
g = dense(layer, data_dim, activation=None) # Outputing xy pairs.
g_vars = tf.contrib.framework.get_variables(g_vs)
return g, g_vars
def generator(z, reuse=False):
#inputs = tf.concat(axis=1, values=[z, x])
with tf.variable_scope('generator', reuse=reuse) as g_vs:
layer = dense(z, h_dim, activation=tf.nn.elu)
layer = dense(layer, h_dim, activation=tf.nn.elu, batch_residual=True)
layer = dense(layer, h_dim, activation=tf.nn.elu, batch_residual=True)
g = dense(layer, data_dim, activation=None) # Outputing xy pairs.
g_vars = tf.contrib.framework.get_variables(g_vs)
return g, g_vars
def __call__(self, x, is_training=True, reuse=False):
with tf.variable_scope(self.name, reuse):
net = lrelu(bn(dense(x, 64, name='c_fc1'),
is_training,
name='c_bn1'),
name='c_l1')
out_logit = dense(net, self.class_num, name='c_l2')
out = tf.nn.softmax(out_logit)
return out_logit, out
def build_ae(self):
self.enX_outputs, _ = self.encoder_X(self.source_X, self.len_X)
self.enX_outputs = tf.reshape(self.enX_outputs, (-1, self.enX_outputs.get_shape().as_list()[-1]))
self.enX_outputs = dense(self.enX_outputs, 3, name='encoder_X_output')
self.enX_outputs = tf.reshape(self.enX_outputs, (-1, self.batch_size, 3))
self.deX_outputs, _ = self.decoder_X(self.enX_outputs, self.len_X)
self.deX_outputs = tf.reshape(self.deX_outputs, (-1, self.deX_outputs.get_shape().as_list()[-1]))
self.deX_outputs = dense(self.deX_outputs, 6, name='decoder_X_output')
self.deX_outputs = tf.reshape(self.deX_outputs, (-1, self.batch_size, 6))
def discriminator_from_params(x, params, isize=28 * 28, n_hid=100):
#bn1 = batch_norm(name='bn1')
#bn2 = batch_norm(name='bn2')
hid = dense(x, n_hid, scope='l1', params=params[:2], normalized=True)
hid = tf.nn.relu(hid)
#hid = tf.tanh(hid)
hid = dense(hid, n_hid, scope='l2', params=params[2:4], normalized=True)
hid = tf.nn.relu(hid)
#hid = tf.tanh(hid)
out = dense(hid, 1, scope='d_out', params=params[4:])
return out
def discriminator(x, z_size, n_hid=500, isize=28 * 28, reuse=False):
#bn1 = batch_norm(name='bn1')
#bn2 = batch_norm(name='bn2')
hid = dense(x, n_hid, scope='l1', reuse=reuse, normalized=True)
hid = tf.nn.relu(hid)
#hid = tf.tanh(hid)
hid = dense(hid, n_hid, scope='l2', reuse=reuse, normalized=True)
#hid = tf.nn.dropout(hid, 0.2)
hid = tf.nn.relu(hid)
#hid = tf.tanh(hid)
out = dense(hid, 1, scope='d_out', reuse=reuse)
return out
def build_disc1(h1, testing=False, reuse=False):
# 16 x 16 --> 8x8
disc1_conv1 = utils.batch_normalization(utils.lrelu(
utils.conv2d(h1, (3, 3, 3, 32),
stride=[1, 2, 2, 1],
name='disc1_conv1')),
name='disc1_bn1',
reuse=reuse)
disc1_conv2 = utils.batch_normalization(utils.lrelu(
utils.conv2d(disc1_conv1, (3, 3, 32, 64), name='disc1_conv2')),
name='disc1_bn2',
reuse=reuse)
# 8x8 --> 8x8
disc1_conv3 = utils.batch_normalization(utils.lrelu(
utils.conv2d(disc1_conv2, (3, 3, 64, 64), name='disc1_conv3')),
name='disc1_bn3',
reuse=reuse)
disc1_conv3 = tf.nn.dropout(disc1_conv3, 0.1 if testing else 1)
# 8x8 --> 6x6
disc1_conv4 = tf.layers.batch_normalization(utils.lrelu(
utils.conv2d(disc1_conv3, (3, 3, 64, 64),
padding='VALID',
name='disc1_conv4')),
name='bn4',
reuse=reuse)
disc1_l5 = tf.reshape(disc1_conv4, [100, 6, 6, 64])
disc1_shared = utils.lrelu(
utils.network_in_network(disc1_l5,
64,
num_units=64,
name='disc1_shared'))
disc1_shared_flat = tf.reshape(disc1_shared, [-1, 64 * 6 * 6])
disc1_z_recon = utils.dense(disc1_shared_flat,
num_inputs=64 * 6 * 6,
num_units=50,
name='disc1_z_recon')
disc1_shared_pool = tf.reduce_mean(disc1_shared, [1, 2])
disc1_adv = utils.dense(disc1_shared_pool,
num_inputs=64,
num_units=1,
name='disc1_z_adv')
# disc1_adv is the pre-sigmoid output of the discriminator
return disc1_adv, disc1_z_recon
def build_gen1(y, z1):
# y is of dimension (batch_size, 8, 8, 3)
gen1_z_embed = utils.batch_normalization(
tf.nn.relu(
utils.dense(z1,
num_inputs=50,
num_units=256,
bias=True,
name='gen1_z_embed')))
y_flatten = tf.reshape(y, (-1, 8 * 8 * 3))
gen1_y_embed = tf.nn.relu(
utils.bias(utils.batch_normalization(
utils.dense(y_flatten,
num_inputs=192,
num_units=512,
bias=False,
name='gen1_y_embed')), (512, ),
name='gen1_y_embed_bias'))
gen1_in = tf.concat([gen1_z_embed, gen1_y_embed], axis=1)
gen1_l1 = tf.transpose(
tf.reshape(
tf.nn.relu(
utils.bias(utils.batch_normalization(
utils.dense(gen1_in,
num_inputs=768,
num_units=1024,
bias=False,
name='gen1_l1')), (1024, ),
name='gen1_l1_bias')), (-1, 64, 4, 4)),
[0, 2, 3, 1])
gen1_l2 = tf.nn.relu(
utils.bias(utils.batch_normalization(
utils.conv2d_transpose(gen1_l1, (4, 4, 64, 64), (100, 11, 11, 64),
bias=False,
padding='VALID',
stride=(1, 2, 2, 1),
name='gen1_l3')), (64, ),
name='gen1_l3_bias'))
gen1_l3 = tf.sigmoid(
utils.conv2d_transpose(gen1_l2, (6, 6, 3, 64), (100, 16, 16, 3),
padding='VALID',
name='gen1_l4'))
return gen1_l3
def generator(z, n_hid=500, isize=28 * 28, reuse=False, use_bn=False):
bn1 = batch_norm(name='bn1')
bn2 = batch_norm(name='bn2')
hid = dense(z, n_hid, scope='l1', reuse=reuse)
if use_bn:
hid = tf.nn.relu(bn1(hid, train=True))
else:
hid = tf.nn.relu(hid)
hid = dense(hid, n_hid, scope='l2', reuse=reuse)
if use_bn:
hid = tf.nn.relu(bn2(hid, train=True))
else:
hid = tf.nn.relu(hid)
out = tf.nn.sigmoid(dense(hid, isize, scope='g_out', reuse=reuse))
return out
def vgg(inputs, struct, dbb_biases, dbb=False):
def _block(inputs, filters, dbb, dbb_bias):
inputs, mem1 = conv(inputs,
filters,
3,
strides=1,
padding='SAME',
dbb=dbb,
dbb_bias=dbb_bias)
inputs, mem2 = batch_norm(inputs)
#print(inputs)
return inputs, mem1 + mem2
total_mem = 0
inputs, mem1 = _block(inputs, struct[0], dbb, dbb_biases[0])
inputs, mem2 = _block(inputs, struct[1], dbb, dbb_biases[1])
inputs = pool(inputs, 2, strides=2, padding='VALID')
inputs, mem3 = _block(inputs, struct[2], dbb, dbb_biases[2])
inputs, mem4 = _block(inputs, struct[3], dbb, dbb_biases[3])
inputs = pool(inputs, 2, strides=2, padding='VALID')
inputs, mem5 = _block(inputs, struct[4], dbb, dbb_biases[4])
inputs, mem6 = _block(inputs, struct[5], dbb, dbb_biases[5])
inputs, mem7 = _block(inputs, struct[6], dbb, dbb_biases[6])
inputs = pool(inputs, 2, strides=2, padding='VALID')
inputs, mem8 = _block(inputs, struct[7], dbb, dbb_biases[7])
inputs, mem9 = _block(inputs, struct[8], dbb, dbb_biases[8])
inputs, mem10 = _block(inputs, struct[9], dbb, dbb_biases[9])
inputs = pool(inputs, 2, strides=2, padding='VALID')
inputs, mem11 = _block(inputs, struct[10], dbb, dbb_biases[10])
inputs, mem12 = _block(inputs, struct[11], dbb, dbb_biases[11])
inputs, mem13 = _block(inputs, struct[12], dbb, dbb_biases[12])
inputs = pool(inputs, 2, strides=2, padding='VALID')
inputs, mem_flat = flatten(inputs, struct[13], dbb, dbb_biases[13])
inputs, mem14 = dense(inputs, struct[14], dbb=dbb, dbb_bias=dbb_biases[14])
inputs, mem15 = batch_norm(inputs)
inputs, mem16 = dense(inputs, struct[15], dbb=False)
total_mem = mem1 + mem2 + mem3 + mem4 + \
mem5 + mem6 + mem7 + mem8 + \
mem9 + mem10 + mem11 + mem12 + \
mem13 + mem14 + mem15 + mem16 + mem_flat
return inputs, total_mem
def _build_discriminator(inputs, data_shape, nplanes):
diff_input = inputs['diff_img']
keep_prob = inputs['keep_prob']
g1_weights = tf.get_variable('g1_weights', [5, 5, 3, nplanes])
g1_bias = tf.get_variable('g1_bias', [nplanes])
g1 = tf.nn.relu(utils.conv2d(g1_weights, bias=g1_bias)(diff_input))
g2_weights = tf.get_variable('g2_weights', [5, 5, nplanes, nplanes])
g2_bias = tf.get_variable('g2_bias', [nplanes])
g2 = tf.nn.relu(utils.conv2d(g2_weights, bias=g2_bias)(g1))
flattened = tf.reshape(g2, [-1, data_shape[0] * data_shape[1] * nplanes])
dropout = tf.nn.dropout(flattened, keep_prob)
prob_weights = tf.get_variable('prob_weights', [data_shape[0] * data_shape[1] * nplanes, 1])
prob_biases = tf.get_variable('prob_bias', [1])
y_logits = utils.dense(prob_weights, bias=prob_biases)(dropout)
y = tf.nn.sigmoid(y_logits, name='prob_real')
#class_weights = tf.get_variable('class_weights', [data_shape[0] * data_shape[1] * nplanes, 10])
#class_biases = tf.get_variable('class_bias', [1])
#class_logits = utils.dense(class_weights, bias=class_biases)(dropout)
#class_prob = tf.nn.softmax(class_logits, name='logits_class')
return {'prob': y, 'prob_logits': y_logits}#, 'prob_class': class_prob, 'logits_class': class_logits}
def train_output(self, encoder_output, Y, teacher_probs, reuse):
"""Calculate loss and accuracy."""
with tf.variable_scope(self.decoder_scope, reuse=reuse):
logits = dense(encoder_output, self._config.dst_vocab_size, use_bias=False,
name="dst_embedding" if self._config.tie_embedding_and_softmax else "softmax",
reuse=True if self._config.tie_embedding_and_softmax else None) # 2D to 3D
preds = tf.to_int32(tf.argmax(logits, axis=-1))
mask = tf.to_float(tf.not_equal(Y, 0))
# Token-level accuracy
acc = tf.reduce_sum(tf.to_float(tf.equal(preds, Y)) * mask) / tf.reduce_sum(mask)
if not tf.get_variable_scope().reuse:
tf.summary.scalar('accuracy', acc)
if teacher_probs is not None:
# Knowledge distillation
loss = tf.nn.softmax_cross_entropy_with_logits(logits=logits, labels=teacher_probs)
else:
# Smoothed loss
loss = common_layers.smoothing_cross_entropy(logits=logits, labels=Y,
vocab_size=self._config.dst_vocab_size,
confidence=1 - self._config.train.label_smoothing)
loss = tf.reduce_sum(loss * mask) / tf.reduce_sum(mask)
return loss
def test_layers2_7_conv_weight(self):
dnn = self.model(self.config_channels, self.anchors,
len(self.category))
output = dnn(self.image)
state_dict = dnn.state_dict()
name = '.'.join(self.id().split('.')[-1].split('_')[1:])
closure = Closure(name, state_dict, self.model.scope)
closure(output.grad_fn)
self.assertDictEqual(
closure.output, {
'layers2.7.conv.weight': 0,
'layers2.7.bn.weight': 0,
'layers2.7.bn.bias': 0,
'layers2.7.bn.running_mean': 0,
'layers2.7.bn.running_var': 0,
})
self.assertDictEqual(closure.input, {
'layers3.0.conv.weight': 64,
})
d = utils.dense(state_dict[name])
channels = torch.LongTensor(np.argsort(d)[int(len(d) * 0.5):])
prune(closure, channels)
config_channels = model.ConfigChannels(self.config_channels.config,
state_dict)
dnn = self.model(config_channels, self.anchors, len(self.category))
dnn(self.image)
self.assertEqual(len(channels), len(dnn.state_dict()[name]))
def train_output(self, decoder_output, Y, X, reuse):
"""Calculate loss and accuracy."""
with tf.variable_scope(self.decoder_scope, reuse=reuse):
logits_gen = dense(
decoder_output,
self._config.dst_vocab_size,
use_bias=False,
name="dst_embedding"
if self._config.tie_embedding_and_softmax else "softmax",
reuse=True if self._config.tie_embedding_and_softmax else
None) # 2D to 3D
preds_gen = tf.to_int32(tf.argmax(logits_gen, axis=-1))
mask = tf.to_float(tf.not_equal(Y, 0))
acc_gen = tf.reduce_sum(
tf.to_float(tf.equal(preds_gen, Y)) *
mask) / tf.reduce_sum(mask)
# Smoothed loss
loss_gen = common_layers.smoothing_cross_entropy(
logits=logits_gen,
labels=Y,
vocab_size=self._config.dst_vocab_size,
confidence=1 - self._config.train.label_smoothing)
mean_loss_gen = tf.reduce_sum(
loss_gen * mask) / (tf.reduce_sum(mask))
return acc_gen, mean_loss_gen
def test_layers3_0_conv_weight(self):
dnn = self.model(self.config_channels, self.anchors, len(self.category))
output = dnn(self.image)
state_dict = dnn.state_dict()
name = '.'.join(self.id().split('.')[-1].split('_')[1:])
d = utils.dense(state_dict[name])
keep = torch.LongTensor(np.argsort(d)[int(len(d) * 0.5):])
modifier = Modifier(
name, state_dict, dnn,
lambda name, var: var[keep],
lambda name, var, mapper: var[mapper(keep, len(d))],
)
modifier(output.grad_fn)
# check channels
scope = dnn.scope(name)
self.assertEqual(state_dict[name].size(0), len(keep))
self.assertEqual(state_dict[scope + '.bn.weight'].size(0), len(keep))
self.assertEqual(state_dict[scope + '.bn.bias'].size(0), len(keep))
self.assertEqual(state_dict[scope + '.bn.running_mean'].size(0), len(keep))
self.assertEqual(state_dict[scope + '.bn.running_var'].size(0), len(keep))
# check if runnable
config_channels = model.ConfigChannels(self.config_channels.config, state_dict)
dnn = self.model(config_channels, self.anchors, len(self.category))
dnn.load_state_dict(state_dict)
dnn(self.image)
def _build_generator(inputs, data_shape, nplanes):
class_cond = inputs['class_cond']
base_img = inputs['base_img']
noise = inputs['noise']
class_weights = tf.get_variable('class_weights',
[10, data_shape[0] * data_shape[1]])
# We don't actually need a bias here as we are learning a bitplane per class anyways
class_vec = utils.dense(class_weights)(class_cond)
class_plane = tf.nn.relu(
tf.reshape(class_vec, [-1, data_shape[0], data_shape[1], 1]))
# Reshape the noise
noise_plane = tf.reshape(noise, [-1, data_shape[0], data_shape[1], 1])
# Now concatenate the tensors
stacked_input = tf.concat([base_img, noise_plane, class_plane], axis=3)
g1_weights = tf.get_variable('g1_weights',
[7, 7, data_shape[2] + 2, nplanes])
g1_bias = tf.get_variable('g1_bias', [nplanes])
g1 = tf.nn.relu(utils.conv2d(g1_weights, bias=g1_bias)(stacked_input))
g2_weights = tf.get_variable('g2_weights', [7, 7, nplanes, nplanes])
g2_bias = tf.get_variable('g2_bias', [nplanes])
g2 = tf.nn.relu(utils.conv2d(g2_weights, bias=g2_bias)(g1))
g3_weights = tf.get_variable('g3_weights', [5, 5, nplanes, data_shape[2]])
g3_bias = tf.get_variable('g3_bias', data_shape[2])
g3 = tf.nn.tanh(
utils.conv2d(g3_weights, bias=g3_bias, name='diff_img')(g2))
return {'diff_img': g3}
def train_output(self, decoder_output, Y, reuse, decoder_scope):
"""Calculate loss and accuracy."""
with tf.variable_scope(decoder_scope, reuse=reuse):
if self._config.is_lsoftmax is None:
self._config.is_lsoftmax = False
if not self._config.is_lsoftmax:
logits = dense(decoder_output, self._config.dst_vocab_size, use_bias=False,
name="dst_embedding" if self._config.tie_embedding_and_softmax else "softmax",
reuse=True if self._config.tie_embedding_and_softmax else None)
else:
with tf.variable_scope("dst_embedding" if self._config.tie_embedding_and_softmax else "softmax",
"dense", reuse=reuse):
input_size = decoder_output.get_shape().as_list()[-1]
inputs_shape = tf.unstack(tf.shape(decoder_output))
decoder_output_tmp = tf.reshape(decoder_output, [-1, input_size])
Y_tmp = tf.reshape(Y, [-1])
with tf.variable_scope(tf.get_variable_scope(),
reuse=True if self._config.tie_embedding_and_softmax else None):
weights = tf.get_variable("kernel", [self._config.dst_vocab_size, input_size])
weights = tf.transpose(weights)
logits = lsoftmax(decoder_output_tmp, weights, Y_tmp)
logits = tf.reshape(logits, inputs_shape[:-1] + [self._config.dst_vocab_size])
preds = tf.to_int32(tf.argmax(logits, axis=-1))
mask = tf.to_float(tf.not_equal(Y, 0))
acc = tf.reduce_sum(tf.to_float(tf.equal(preds, Y)) * mask) / tf.reduce_sum(mask)
# Smoothed loss
loss = common_layers.smoothing_cross_entropy(logits=logits, labels=Y,
vocab_size=self._config.dst_vocab_size,
confidence=1 - self._config.train.label_smoothing)
mean_loss = tf.reduce_sum(loss * mask) / (tf.reduce_sum(mask))
return acc, mean_loss
def __init__(self, ob_dim, ac_dim): #pylint: disable=W0613
X = tf.placeholder(tf.float32,
shape=[None, ob_dim * 2 + ac_dim * 2 + 2
]) # batch of observations
vtarg_n = tf.placeholder(tf.float32, shape=[None], name='vtarg')
wd_dict = {}
h1 = tf.nn.elu(
dense(X,
64,
"h1",
weight_init=U.normc_initializer(1.0),
bias_init=0,
weight_loss_dict=wd_dict))
h2 = tf.nn.elu(
dense(h1,
64,
"h2",
weight_init=U.normc_initializer(1.0),
bias_init=0,
weight_loss_dict=wd_dict))
vpred_n = dense(h2,
1,
"hfinal",
weight_init=U.normc_initializer(1.0),
bias_init=0,
weight_loss_dict=wd_dict)[:, 0]
sample_vpred_n = vpred_n + tf.random_normal(tf.shape(vpred_n))
wd_loss = tf.get_collection("vf_losses", None)
loss = U.mean(tf.square(vpred_n - vtarg_n)) + tf.add_n(wd_loss)
loss_sampled = U.mean(
tf.square(vpred_n - tf.stop_gradient(sample_vpred_n)))
self._predict = U.function([X], vpred_n)
optim = kfac.KfacOptimizer(learning_rate=0.001, cold_lr=0.001*(1-0.9), momentum=0.9, \
clip_kl=0.3, epsilon=0.1, stats_decay=0.95, \
async=1, kfac_update=2, cold_iter=50, \
weight_decay_dict=wd_dict, max_grad_norm=None)
vf_var_list = []
for var in tf.trainable_variables():
if "vf" in var.name:
vf_var_list.append(var)
update_op, self.q_runner = optim.minimize(loss,
loss_sampled,
var_list=vf_var_list)
self.do_update = U.function([X, vtarg_n], update_op) #pylint: disable=E1101
U.initialize() # Initialize uninitialized TF variables
def __call__(self, z, y=None, is_training=True, reuse=False):
with tf.variable_scope(self.name, reuse=reuse):
batch_size = z.get_shape().as_list()[0]
if y is not None:
z = tf.concat([z, y], 1)
net = tf.nn.relu(
bn(dense(z, 128, name='g_fc1'), is_training, name='g_bn1'))
net = tf.nn.relu(
bn(dense(net, 256, name='g_fc2'), is_training, name='g_bn2'))
net = tf.nn.relu(
bn(dense(net, 512, name='g_fc3'), is_training, name='g_bn3'))
net = tf.nn.relu(
bn(dense(net, 1024, name='g_fc4'), is_training, name='g_bn4'))
net = tf.nn.sigmoid(dense(net, 784, name='g_fc5'))
out = tf.reshape(net, (batch_size, 28, 28, 1))
return out
def encoder_impl(self, encoder_input, is_training):
attention_dropout_rate = self._config.attention_dropout_rate if is_training else 0.0
residual_dropout_rate = self._config.residual_dropout_rate if is_training else 0.0
# Mask
encoder_padding = tf.equal(
tf.reduce_sum(tf.abs(encoder_input), axis=-1), 0.0)
encoder_output = dense(encoder_input,
self._config.hidden_units,
activation=tf.identity,
use_bias=True,
name="src_change")
encoder_output = tf.contrib.layers.layer_norm(encoder_output,
center=True,
scale=True,
trainable=True)
# Add positional signal
encoder_output = common_attention.add_timing_signal_1d(encoder_output)
# Dropout
encoder_output = tf.layers.dropout(encoder_output,
rate=residual_dropout_rate,
training=is_training)
# Blocks
for i in range(self._config.num_blocks_enc):
with tf.variable_scope("block_{}".format(i)):
# Multihead Attention
encoder_output = residual(
encoder_output,
multihead_attention(
query_antecedent=encoder_output,
memory_antecedent=None,
bias=common_attention.attention_bias_ignore_padding(
encoder_padding),
total_key_depth=self._config.hidden_units,
total_value_depth=self._config.hidden_units,
output_depth=self._config.hidden_units,
num_heads=self._config.num_heads,
dropout_rate=attention_dropout_rate,
name='encoder_self_attention',
summaries=True),
dropout_rate=residual_dropout_rate)
# Feed Forward
encoder_output = residual(
encoder_output,
ff_hidden(inputs=encoder_output,
hidden_size=4 * self._config.hidden_units,
output_size=self._config.hidden_units,
activation=self._ff_activation),
dropout_rate=residual_dropout_rate)
# Mask padding part to zeros.
encoder_output *= tf.expand_dims(1.0 - tf.to_float(encoder_padding),
axis=-1)
return encoder_output
def __call__(self, x, y=None, sn=False, is_training=True, reuse=False):
with tf.variable_scope(self.name, reuse=reuse):
batch_size = x.get_shape().as_list()[0]
if y is not None:
ydim = y.get_shape().as_list()[-1]
y = tf.reshape(y, [batch_size, 1, 1, ydim])
x = conv_cond_concat(x, y) # [bz, 28, 28, 11]
# [bz, 14, 14, 64]
net = lrelu(conv2d(x,
64,
4,
4,
2,
2,
sn=sn,
padding="SAME",
name='d_conv1'),
name='d_l1')
# [bz, 7, 7, 128]
net = lrelu(bn(conv2d(net,
128,
4,
4,
2,
2,
sn=sn,
padding="SAME",
name='d_conv2'),
is_training,
name='d_bn2'),
name='d_l2')
net = tf.reshape(net, [batch_size, 7 * 7 * 128])
# [bz, 1024]
net = lrelu(bn(dense(net, 1024, sn=sn, name='d_fc3'),
is_training,
name='d_bn3'),
name='d_l3')
# [bz, 1]
yd = dense(net, 1, sn=sn, name='D_dense')
if self.class_num:
yc = dense(net, self.class_num, sn=sn, name='C_dense')
return yd, net, yc
else:
return yd, net
def test_output(self, decoder_output, reuse):
"""During test, we only need the last prediction at each time."""
with tf.variable_scope("decoder", reuse=reuse):
last_logits = dense(decoder_output[:,-1], self._config.dst_vocab_size, use_bias=False,
name="dst_embedding" if self._config.tie_embedding_and_softmax else "softmax",
reuse=True if self._config.tie_embedding_and_softmax else None)
last_preds = tf.to_int32(tf.argmax(last_logits, axis=-1))
z = tf.nn.log_softmax(last_logits)
last_k_scores, last_k_preds = tf.nn.top_k(z, k=self._config.test.beam_size, sorted=False)
last_k_preds = tf.to_int32(last_k_preds)
return last_preds, last_k_preds, last_k_scores
def test_loss(self, decoder_output, Y, reuse):
"""This function help users to compute PPL during test."""
with tf.variable_scope("decoder", reuse=reuse):
logits = dense(decoder_output, self._config.dst_vocab_size, use_bias=False,
name="dst_embedding" if self._config.tie_embedding_and_softmax else "softmax",
reuse=True if self._config.tie_embedding_and_softmax else None)
mask = tf.to_float(tf.not_equal(Y, 0))
labels = tf.one_hot(Y, depth=self._config.dst_vocab_size)
loss = tf.nn.softmax_cross_entropy_with_logits(logits=logits, labels=labels)
loss_sum = tf.reduce_sum(loss * mask)
return loss_sum
def lenet_conv(inputs, struct, dbb=False):
total_mem = 0
if dbb:
inputs, mem1 = conv(inputs, struct[0], 5, strides=1, padding='VALID', \
dbb=dbb, dbb_bias=13)
inputs = pool(inputs, 2, strides=2, padding='VALID')
inputs, mem2 = conv(inputs, struct[1], 5, strides=1, padding='VALID', \
dbb=dbb, dbb_bias=25)
inputs = pool(inputs, 2, strides=2, padding='VALID')
inputs, mem3 = flatten(inputs, struct[2], dbb=dbb, dbb_bias=156)
inputs, mem4 = dense(inputs, struct[3], dbb=dbb, dbb_bias=54)
inputs, mem5 = dense(inputs, 10, dbb=False)
else:
inputs, mem1 = conv(inputs,
struct[0],
5,
strides=1,
padding='VALID',
dbb=dbb)
inputs = pool(inputs, 2, strides=2, padding='VALID')
inputs, mem2 = conv(inputs,
struct[1],
5,
strides=1,
padding='VALID',
dbb=dbb)
inputs = pool(inputs, 2, strides=2, padding='VALID')
inputs, mem3 = flatten(inputs, struct[2], dbb=dbb)
inputs, mem4 = dense(inputs, struct[3], dbb=dbb)
inputs, mem5 = dense(inputs, 10, dbb=False)
total_mem = mem1 + mem2 + mem3 + mem4 + mem5
return inputs, total_mem
def _build_autoencoder(inputs, data_shape, num_planes, latent_dim):
base_img = inputs['base_img']
class_cond = inputs['class_cond']
# We don't actually need a bias here as we are learning a bitplane per class anyways
class_weights = tf.get_variable('class_weights', [10, data_shape[0] * data_shape[1]])
class_vec = utils.dense(class_weights)(class_cond)
class_plane = tf.nn.relu(tf.reshape(class_vec, [-1, data_shape[0], data_shape[1], 1]))
# Now concatenate the tensors
stacked_input = tf.concat([base_img, class_plane], axis=3)
c1_weights = tf.get_variable('d1_weights', [7, 7, data_shape[2] + 1, num_planes])
c1_bias = tf.get_variable('d1_bias', [num_planes])
c1 = utils.conv2d(c1_weights, bias=c1_bias)(stacked_input)
c2_weights = tf.get_variable('d2_weights', [7, 7, num_planes, num_planes])
c2_bias = tf.get_variable('d2_bias', [num_planes])
c2 = utils.conv2d(c2_weights, bias=c2_bias)(c1)
c2_dropout = tf.nn.dropout(c2, inputs['keep_prob'])
c3_weights = tf.get_variable('d3_weights', [7, 7, num_planes, num_planes])
c3_bias = tf.get_variable('d3_bias', [num_planes])
c3 = utils.conv2d(c3_weights, bias=c3_bias)(c2_dropout)
c4_weights = tf.get_variable('d4_weights', [7, 7, num_planes, num_planes])
c4_bias = tf.get_variable('d4_bias', [num_planes])
c4 = utils.conv2d(c4_weights, bias=c4_bias)(c3)
c4_dropout = tf.nn.dropout(c4, inputs['keep_prob'])
c5_weights = tf.get_variable('d5_weights', [7, 7, num_planes, 3])
c5_bias = tf.get_variable('d5_bias', [3])
c5 = utils.conv2d(c5_weights, bias=c5_bias)(c4_dropout)
diff_img = c5
return {'diff_img': diff_img}
