def call(self, inputs, states, constants):
if not isinstance(constants, (list, tuple)):
keys = values = constants
elif len(constants) == 1:
keys = values = constants[0]
elif len(constants) == 2:
keys, values = constants
else:
raise ValueError(
'constants can either be a list with keys and values or just attention vectors'
)
if not isinstance(states, (list, tuple)):
query = states
else:
query = states[0]
query = self._query_transformation(query)
repeated_query = K.repeat(query, K.shape(keys)[1])
logits = self._attention_logits_dense(K.tanh(repeated_query + keys))
attention_weights = keras.activations.softmax(logits, axis=1)
attention_context = K.sum(attention_weights * values,
axis=1,
keepdims=False)
inputs = inputs + attention_context
return self._cell.call(inputs, states)
def build_generator(self):
"""U-Net Generator"""
def conv2d(layer_input, filters, f_size=4):
"""Layers used during downsampling"""
d = Conv2D(filters, kernel_size=f_size, strides=2,
padding='same')(layer_input)
d = LeakyReLU(alpha=0.2)(d)
d = InstanceNormalization()(d)
return d
def deconv2d(layer_input,
skip_input,
filters,
f_size=4,
dropout_rate=0):
"""Layers used during upsampling"""
u = UpSampling2D(size=2)(layer_input)
u = Conv2D(filters,
kernel_size=f_size,
strides=1,
padding='same',
activation='relu')(u)
if dropout_rate:
u = Dropout(dropout_rate)(u)
u = InstanceNormalization()(u)
u = Concatenate()([u, skip_input])
return u
# Image input
input_img = Input(shape=self.img_shape)
inp_c = Input(shape=(self.class_num, ))
c = Lambda(lambda x: K.repeat(x, self.img_width * self.img_height))(
inp_c)
c = Reshape((self.img_width, self.img_height, self.class_num))(c)
d0 = Concatenate()([input_img, c])
# Downsampling
d1 = conv2d(d0, self.gf)
d2 = conv2d(d1, self.gf * 2)
d3 = conv2d(d2, self.gf * 4)
d4 = conv2d(d3, self.gf * 8)
# Upsampling
u1 = deconv2d(d4, d3, self.gf * 4)
u2 = deconv2d(u1, d2, self.gf * 2)
u3 = deconv2d(u2, d1, self.gf)
u4 = UpSampling2D(size=2)(u3)
output_img = Conv2D(self.img_channel,
kernel_size=4,
strides=1,
padding='same',
activation='tanh')(u4)
return Model([input_img, inp_c], output_img)
def call(self, x, mask=None):
if mask is not None:
# mask (batch, time)
mask = K.cast(mask, K.floatx())
# mask (batch, x_dim, time)
mask = K.repeat(mask, x.shape[-1])
# mask (batch, time, x_dim)
mask = tf.transpose(mask, [0, 2, 1])
x = x * mask
# print(mask)
return K.sum(x, axis=1) / K.sum(mask, axis=1)
def call(self, x, mask=None):
"""1, mask is a bool type tensor, need casting before compute.
2, mask shape in 2 dimension (batch_size, feature_dimension)
"""
if mask is not None:
mask = K.repeat(mask, x.shape[-1])
mask = tf.transpose(mask, [0, 2, 1])
mask = tf.cast(mask, tf.float32)
x = x * mask
return K.sum(x, axis=1) / K.sum(mask, axis=1)
else:
return K.mean(x, axis=1)
def _time_distributed_dense(x,
w,
b=None,
dropout=None,
input_dim=None,
output_dim=None,
timesteps=None,
training=None):
"""Apply `y . w + b` for every temporal slice y of x.
# Arguments
x: input tensor.
w: weight matrix.
b: optional bias vector.
dropout: wether to apply dropout (same dropout mask
for every temporal slice of the input).
input_dim: integer; optional dimensionality of the input.
output_dim: integer; optional dimensionality of the output.
timesteps: integer; optional number of timesteps.
training: training phase tensor or boolean.
# Returns
Output tensor.
"""
if not input_dim:
input_dim = K.shape(x)[2]
if not timesteps:
timesteps = K.shape(x)[1]
if not output_dim:
output_dim = K.int_shape(w)[1]
if dropout is not None and 0. < dropout < 1.:
# apply the same dropout pattern at every timestep
ones = K.ones_like(K.reshape(x[:, 0, :], (-1, input_dim)))
dropout_matrix = K.dropout(ones, dropout)
expanded_dropout_matrix = K.repeat(dropout_matrix, timesteps)
x = K.in_train_phase(x * expanded_dropout_matrix, x, training=training)
# collapse time dimension and batch dimension together
x = K.reshape(x, (-1, input_dim))
x = K.dot(x, w)
if b is not None:
x = K.bias_add(x, b)
# reshape to 3D tensor
if K.backend() == 'tensorflow':
x = K.reshape(x, K.stack([-1, timesteps, output_dim]))
x.set_shape([None, None, output_dim])
else:
x = K.reshape(x, (-1, timesteps, output_dim))
return x
def call(self, x, mask=None):
if mask is not None:
# mask (batch, time)
mask = K.cast(mask, K.floatx())
if K.ndim(x) != K.ndim(mask):
mask = K.repeat(mask, x.shape[-1])
mask = tf.transpose(mask, [0, 2, 1])
x = x * mask
if K.ndim(x) == 2:
x = K.expand_dims(x)
return K.sum(x, axis=self.axis)
else:
if K.ndim(x) == 2:
x = K.expand_dims(x)
return K.sum(x, axis=self.axis)
def select_best_leaf(self, y_pred):
if self.N > self.num_leaves:
# if there are more leaf nodes than total nodes in the hierarchy (should always be the case,
# but allowed to work either way) then pad with a zero for each non-leaf node in the taxonomy
y_pred = self._pad(y_pred)
# propagate the probabilities (algo 1)
propagated_probabilities = K.transpose(
K.dot(self.A, K.transpose(y_pred)))
# grab the mask vector for root and repeat it <batch size> times
root = K.repeat(self.root, K.shape(y_pred)[0])
# reshape into (<batch size>, N)
predictions = K.reshape(root, (K.shape(y_pred)[0], ))
# each branch will walk futher out toward leaf nodes (and loops on leaf nodes)
for _ in range(self.depth):
predictions = self._branch(propagated_probabilities, predictions)
return predictions
def call(self, x, mask=None):
'''mask是上一层的'''
'''# using 'mask' you can access the mask passed from the previous layer'''
# x [batch_size, seq_len, embedding_size]
if self.supports_masking:
# mask [batch_size, seq_len]
if mask is None:
# 先判断是否非零，然后执行OR运算，计算每个序列的有效长度
mask = K.any(K.not_equal(x, 0), -1) # [batch_size, seq_len]
mask = K.cast(mask, K.floatx())
return K.sum(x, axis=1) / K.sum(mask, axis=1, keepdims=True)
if mask is not None:
mask = K.cast(mask, K.floatx())
# [batch_size, embedding_size, seq_len]
mask = K.repeat(mask, x.shape[-1].value)
# [batch_size, seq_len, embedding_size]
mask = tf.transpose(mask, [0, 2, 1])
x = x * mask
return K.sum(x, axis=1) / K.sum(mask, axis=1)
def test_sequence_example_into_input_layer(self):
examples = [_make_sequence_example().SerializeToString()] * 100
ctx_cols, seq_cols = self._build_feature_columns()
def _parse_example(example):
ctx, seq = parsing_ops.parse_single_sequence_example(
example,
context_features=fc.make_parse_example_spec_v2(ctx_cols),
sequence_features=fc.make_parse_example_spec_v2(seq_cols))
ctx.update(seq)
return ctx
ds = dataset_ops.Dataset.from_tensor_slices(examples)
ds = ds.map(_parse_example)
ds = ds.batch(20)
# Test on a single batch
features = dataset_ops.make_one_shot_iterator(ds).get_next()
# Tile the context features across the sequence features
sequence_input_layer = ksfc.SequenceFeatures(seq_cols)
seq_input, _ = sequence_input_layer(features)
dense_input_layer = dense_features.DenseFeatures(ctx_cols)
ctx_input = dense_input_layer(features)
ctx_input = backend.repeat(ctx_input, array_ops.shape(seq_input)[1])
concatenated_input = merge.concatenate([seq_input, ctx_input])
rnn_layer = recurrent.RNN(recurrent.SimpleRNNCell(10))
output = rnn_layer(concatenated_input)
with self.cached_session() as sess:
sess.run(variables.global_variables_initializer())
features_r = sess.run(features)
self.assertAllEqual(features_r['int_list'].dense_shape, [20, 3, 6])
output_r = sess.run(output)
self.assertAllEqual(output_r.shape, [20, 10])
def call(self, inputs): return K.repeat(inputs, self.n)
def call(self, inputs): return K.repeat(inputs, self.n)
def call(self, inputs, states, training=None):
# dropout matrices for input units
dp_mask = self._dropout_mask
# dropout matrices for recurrent units
rec_dp_mask = self._recurrent_dropout_mask
h_tm1 = states[0] # previous memory state
c_tm1 = states[1] # previous carry state
# alignment model
h_att = K.repeat(h_tm1, self.timestep_dim)
att = _time_distributed_dense(inputs,
self.attention_weights,
self.attention_bias,
input_dim=self.input_dim,
output_dim=self.units,
timesteps=self.timestep_dim)
attention_ = self.attention_activation(
K.dot(h_att, self.attention_recurrent_weights) + att) # energy
attention_ = K.squeeze(K.dot(attention_,
self.attention_recurrent_bias),
2) # energy
alpha = K.exp(attention_)
if dp_mask is not None:
alpha *= dp_mask[0]
alpha /= K.sum(alpha, axis=1, keepdims=True)
alpha_r = K.repeat(alpha, self.input_dim)
alpha_r = K.permute_dimensions(alpha_r, (0, 2, 1))
# make context vector (soft attention after Bahdanau et al.)
z_hat = inputs * alpha_r
context_sequence = z_hat
z_hat = K.sum(z_hat, axis=1)
if self.implementation == 1:
if 0 < self.dropout < 1.:
inputs_i = inputs * dp_mask[0]
inputs_f = inputs * dp_mask[1]
inputs_c = inputs * dp_mask[2]
inputs_o = inputs * dp_mask[3]
else:
inputs_i = inputs
inputs_f = inputs
inputs_c = inputs
inputs_o = inputs
x_i = K.dot(inputs_i, self.kernel_i)
x_f = K.dot(inputs_f, self.kernel_f)
x_c = K.dot(inputs_c, self.kernel_c)
x_o = K.dot(inputs_o, self.kernel_o)
if self.use_bias:
x_i = K.bias_add(x_i, self.bias_i)
x_f = K.bias_add(x_f, self.bias_f)
x_c = K.bias_add(x_c, self.bias_c)
x_o = K.bias_add(x_o, self.bias_o)
if 0 < self.recurrent_dropout < 1.:
h_tm1_i = h_tm1 * rec_dp_mask[0]
h_tm1_f = h_tm1 * rec_dp_mask[1]
h_tm1_c = h_tm1 * rec_dp_mask[2]
h_tm1_o = h_tm1 * rec_dp_mask[3]
else:
h_tm1_i = h_tm1
h_tm1_f = h_tm1
h_tm1_c = h_tm1
h_tm1_o = h_tm1
i = self.recurrent_activation(
x_i + K.dot(h_tm1_i, self.recurrent_kernel_i) +
K.dot(z_hat, self.attention_i))
f = self.recurrent_activation(
x_f + K.dot(h_tm1_f, self.recurrent_kernel_f) +
K.dot(z_hat, self.attention_f))
c = f * c_tm1 + i * self.activation(
x_c + K.dot(h_tm1_c, self.recurrent_kernel_c) +
K.dot(z_hat, self.attention_c))
o = self.recurrent_activation(
x_o + K.dot(h_tm1_o, self.recurrent_kernel_o) +
K.dot(z_hat, self.attention_o))
else:
if 0. < self.dropout < 1.:
inputs *= dp_mask[0]
z = K.dot(inputs, self.kernel)
if 0. < self.recurrent_dropout < 1.:
h_tm1 *= rec_dp_mask[0]
z += K.dot(h_tm1, self.recurrent_kernel)
z += K.dot(z_hat, self.attention_kernel)
if self.use_bias:
z = K.bias_add(z, self.bias)
z0 = z[:, :self.units]
z1 = z[:, self.units:2 * self.units]
z2 = z[:, 2 * self.units:3 * self.units]
z3 = z[:, 3 * self.units:]
i = self.recurrent_activation(z0)
f = self.recurrent_activation(z1)
c = f * c_tm1 + i * self.activation(z2)
o = self.recurrent_activation(z3)
h = o * self.activation(c)
if 0 < self.dropout + self.recurrent_dropout:
if training is None:
h._uses_learning_phase = True
return h, [h, c]
