def _multihead_attention_layer(self,
layer_idx,
query,
memory=None,
mask=None):
if memory is None:
memory = query
# Linear project to d_model dimension: [batch, q_size/k_size, d_model]
Q = ne.fully_conn(query,
self.att_weights["W_att_Q_l{}_0".format(layer_idx)],
self.att_biases["b_att_Q_l{}_0".format(layer_idx)])
Q = ne.leaky_relu(Q, self.leaky_ratio[layer_idx])
K = ne.fully_conn(memory,
self.att_weights["W_att_K_l{}_0".format(layer_idx)],
self.att_biases["b_att_K_l{}_0".format(layer_idx)])
K = ne.leaky_relu(K, self.leaky_ratio[layer_idx])
V = ne.fully_conn(memory,
self.att_weights["W_att_V_l{}_0".format(layer_idx)],
self.att_biases["b_att_V_l{}_0".format(layer_idx)])
V = ne.leaky_relu(V, self.leaky_ratio[layer_idx])
# Split the matrix to multiple heads and then concatenate to have a larger
# batch size: [h*batch, q_size/k_size, d_model/num_heads]
Q_split = tf.concat(tf.split(Q, self.num_att_header, axis=2), axis=0)
K_split = tf.concat(tf.split(K, self.num_att_header, axis=2), axis=0)
V_split = tf.concat(tf.split(V, self.num_att_header, axis=2), axis=0)
if mask != None:
mask = tf.tile(mask, [self.num_att_header, 1, 1])
# Apply scaled dot product attention
d = self.feature_size // self.num_att_header
assert d == Q_split.shape[-1] == K_split.shape[-1] == V_split.shape[-1]
out = tf.matmul(Q_split,
tf.transpose(K_split,
[0, 2, 1])) # [h*batch, q_size, k_size]
out = out / tf.sqrt(tf.cast(d, tf.float32)) # scaled by sqrt(d_k)
if mask is not None:
# masking out (0.0) => setting to -inf.
out = tf.multiply(out, mask) + (1.0 - mask) * (-1e10)
out = ne.softmax(out) # [h * batch, q_size, k_size]
out = ne.dropout(out, self.drop_rate[layer_idx], self.is_training)
out = tf.matmul(out, V_split) # [h * batch, q_size, d_model]
# Merge the multi-head back to the original shape
out = tf.concat(tf.split(out, self.num_att_header, axis=0),
axis=2) # [bs, q_size, d_model]
return out
def in_layer(self, inputs, W_name="W_in_", b_name="b_in_"):
net = inputs
# fc
h, w, c = net.get_shape().as_list()[1:]
assert h * w * c == self.in_state
net = tf.reshape(net, [-1, self.in_state])
for layer_id in range(len(self.in_fc_states)):
weight_name = "{}{}".format(W_name, layer_id)
bias_name = "{}{}".format(b_name, layer_id)
curr_weight = self.in_weight[weight_name]
curr_bias = self.in_bias[bias_name]
# batch normalization
if self.in_norm == "BATCH":
net = ne.batch_norm(net, self.is_training)
elif self.in_norm == "LAYER":
net = ne.layer_norm(net, self.is_training)
#net = ne.leaky_brelu(net, self.conv_leaky_ratio[layer_id], self.layer_low_bound, self.output_up_bound) # Nonlinear act
net = ne.leaky_relu(net, self.in_leaky_ratio)
net = ne.fully_conn(net, curr_weight, curr_bias)
out_channel = self.in_fc_states[-1] // h // w
assert h * w * out_channel == self.in_fc_states[-1]
net = tf.reshape(net, [-1, h, w, out_channel])
net = tf.identity(net, name='in_output')
#import pdb; pdb.set_trace()
return net
def out_layer(self, inputs, label=None, W_name="W_out_", b_name="b_out_"):
net = inputs
h, w, c = net.get_shape().as_list()[1:]
assert h*w*c == self.out_state
net = tf.reshape(net, [-1, self.out_state])
if self.use_class_label:
net = tf.concat([net, label], -1)
for layer_id in range(len(self.out_fc_states)):
weight_name = "{}{}".format(W_name, layer_id)
bias_name = "{}{}".format(b_name, layer_id)
curr_weight = self.out_weight[weight_name]
curr_bias = self.out_bias[bias_name]
net = ne.fully_conn(net, curr_weight, curr_bias)
if self.out_norm == "BATCH":
net = ne.batch_norm(net, self.is_training)
elif self.out_norm == "LAYER":
net = ne.layer_norm(net, self.is_training)
#net = ne.leaky_brelu(net, self.conv_leaky_ratio[layer_id], self.layer_low_bound, self.output_up_bound) # Nonlinear act
net = ne.leaky_relu(net, self.out_leaky_ratio)
out_channel_size = self.out_fc_states[-1]//h//w
assert h*w*out_channel_size == self.out_fc_states[-1]
net = tf.reshape(net, [-1, h, w, out_channel_size])
#net = ne.max_pool_2x2(net) # Pooling
net = tf.identity(net, name='out_output')
#import pdb; pdb.set_trace()
return net
def _func(net, layer_id, postfix="", act_func="leaky"):
weight_name = "{}{}{}".format(W_name, layer_id, postfix)
bias_name = "{}{}{}".format(b_name, layer_id, postfix)
curr_weight = self.enfc_weights[weight_name]
curr_bias = self.enfc_biases[bias_name]
net = ne.fully_conn(net, weights=curr_weight, biases=curr_bias)
# batch normalization
if self.use_norm == "BATCH":
net = ne.batch_norm(net, self.is_training, axis=1)
elif self.use_norm == "LAYER":
net = ne.layer_norm(net, self.is_training)
#net = ne.leaky_brelu(net, self.enfc_leaky_ratio[layer_id], self.enfc_low_bound[layer_id], self.enfc_up_bound[layer_id]) # Nonlinear act
if act_func == "leaky":
net = ne.leaky_relu(net, self.enfc_leaky_ratio[layer_id])
elif act_func == "soft":
net = tf.nn.softplus(net)
#net = ne.drop_out(net, self.enfc_drop_rate[layer_id], self.is_training)
return net
def _fc_layers(self, inputs, weights_dict, biases_dict, fc_leaky_ratio,
fc_drop_rate, num_fc, W_name, b_name):
net = inputs
for layer_id in range(num_fc):
weight_name = "{}{}".format(W_name, layer_id)
bias_name = "{}{}".format(b_name, layer_id)
curr_weight = weights_dict[weight_name]
curr_bias = biases_dict[bias_name]
net = ne.fully_conn(net, weights=curr_weight, biases=curr_bias)
# batch normalization
if self.use_norm == "BATCH":
net = ne.batch_norm(net, self.is_training, axis=-1)
#net = ne.leaky_brelu(net, self.enfc_leaky_ratio[layer_id], self.enfc_low_bound[layer_id], self.enfc_up_bound[layer_id]) # Nonlinear act
net = ne.leaky_relu(net, fc_leaky_ratio[layer_id])
net = ne.drop_out(net, fc_drop_rate[layer_id], self.is_training)
#net = ne.elu(net)
net = tf.identity(net, name='output')
return net
def defc_layers(self, inputs, W_name="W_defc", b_name="b_defc"):
net = inputs
for layer_id in range(self.num_enfc):
weight_name = "{}{}".format(W_name, layer_id)
bias_name = "{}{}".format(b_name, layer_id)
curr_weight = self.defc_weights[weight_name]
curr_bias = self.defc_biases[bias_name]
net = ne.fully_conn(net, weights=curr_weight, biases=curr_bias)
# batch normalization
if self.use_batch_norm:
net = ne.batch_norm(net, self.is_training, axis=1)
#net = ne.leaky_brelu(net, self.defc_leaky_ratio[layer_id], self.layer_low_bound, self.layer_up_bound) # Nonlinear act
net = ne.leaky_relu(net, self.defc_leaky_ratio[layer_id])
net = ne.drop_out(net, self.defc_drop_rate[layer_id],
self.is_training)
#net = ne.elu(net)
net = tf.identity(net, name='output')
net = tf.reshape(net, [-1] + self.decv_in_shape)
return net
def enfc_layers(self, inputs, W_name="W_enfc", b_name="b_enfc"):
net = tf.reshape(inputs, [
-1, self.conv_out_shape[0] * self.conv_out_shape[1] *
self.conv_out_shape[2]
])
for layer_id in range(self.num_enfc):
weight_name = "{}{}".format(W_name, layer_id)
bias_name = "{}{}".format(b_name, layer_id)
curr_weight = self.enfc_weights[weight_name]
curr_bias = self.enfc_biases[bias_name]
net = ne.fully_conn(net, weights=curr_weight, biases=curr_bias)
# batch normalization
if self.use_norm == "BATCH":
net = ne.batch_norm(net, self.is_training, axis=1)
elif self.use_norm == "LAYER":
net = ne.layer_norm(net, self.is_training)
#net = ne.leaky_brelu(net, self.enfc_leaky_ratio[layer_id], self.enfc_low_bound[layer_id], self.enfc_up_bound[layer_id]) # Nonlinear act
net = ne.leaky_relu(net, self.enfc_leaky_ratio[layer_id])
net = ne.drop_out(net, self.enfc_drop_rate[layer_id],
self.is_training)
#net = ne.elu(net)
net = tf.identity(net, name='output')
return net