def branch_attention(cost_volume_3d, cost_volume_h, cost_volume_v,
cost_volume_45, cost_volume_135):
feature = 4 * 9
k = 9
label = 9
cost1 = convbn(cost_volume_3d, 6, 3, 1, 1)
cost1 = Activation('relu')(cost1)
cost1 = convbn(cost1, 4, 3, 1, 1)
cost1 = Activation('sigmoid')(cost1)
cost_h = Lambda(lambda y: K.repeat_elements(
K.expand_dims(y[:, :, :, :1], 1), 9, 1))(cost1)
cost_h = Lambda(lambda y: K.repeat_elements(y, feature, 4))(cost_h)
cost_v = Lambda(lambda y: K.repeat_elements(
K.expand_dims(y[:, :, :, 1:2], 1), 9, 1))(cost1)
cost_v = Lambda(lambda y: K.repeat_elements(y, feature, 4))(cost_v)
cost_45 = Lambda(lambda y: K.repeat_elements(
K.expand_dims(y[:, :, :, 2:3], 1), 9, 1))(cost1)
cost_45 = Lambda(lambda y: K.repeat_elements(y, feature, 4))(cost_45)
cost_135 = Lambda(lambda y: K.repeat_elements(
K.expand_dims(y[:, :, :, 3:4], 1), 9, 1))(cost1)
cost_135 = Lambda(lambda y: K.repeat_elements(y, feature, 4))(cost_135)
return concatenate([
multiply([cost_h, cost_volume_h]),
multiply([cost_v, cost_volume_v]),
multiply([cost_45, cost_volume_45]),
multiply([cost_135, cost_volume_135])
],
axis=4), cost1
def disparityregression(input):
shape = K.shape(input)
disparity_values = np.linspace(-4, 4, 9)
x = K.constant(disparity_values, shape=[9])
x = K.expand_dims(K.expand_dims(K.expand_dims(x, 0), 0), 0)
x = tf.tile(x, [shape[0], shape[1], shape[2], 1])
out = K.sum(multiply([input, x]), -1)
return out
def channel_attention_free(cost_volume):
x = GlobalAveragePooling3D()(cost_volume)
x = Lambda(
lambda y: K.expand_dims(K.expand_dims(K.expand_dims(y, 1), 1), 1))(x)
x = Conv3D(170, 1, 1, 'same')(x)
x = Activation('relu')(x)
x = Conv3D(81, 1, 1, 'same')(x)
x = Activation('sigmoid')(x)
attention = Lambda(lambda y: K.reshape(y, (K.shape(y)[0], 1, 1, 1, 81)))(x)
x = Lambda(lambda y: K.repeat_elements(y, 4, -1))(attention)
return multiply([x, cost_volume]), attention
def to_3d_135(cost_volume_135):
feature = 4 * 9
channel_135 = GlobalAveragePooling3D(
data_format='channels_last')(cost_volume_135)
channel_135 = Lambda(lambda y: K.expand_dims(
K.expand_dims(K.expand_dims(y, 1), 1), 1))(channel_135)
channel_135 = Conv3D(feature / 2,
1,
1,
'same',
data_format='channels_last')(channel_135)
channel_135 = Activation('relu')(channel_135)
channel_135 = Conv3D(3, 1, 1, 'same',
data_format='channels_last')(channel_135)
channel_135 = Activation('sigmoid')(channel_135)
channel_135 = Lambda(lambda y: K.concatenate([
y[:, :, :, :, 0:1], y[:, :, :, :, 0:1], y[:, :, :, :, 0:1],
y[:, :, :, :, 0:1], y[:, :, :, :, 1:2], y[:, :, :, :, 2:3],
y[:, :, :, :, 2:3], y[:, :, :, :, 2:3], y[:, :, :, :, 2:3]
],
axis=-1))(channel_135)
channel_135 = Lambda(lambda y: K.reshape(y, (K.shape(y)[0], 1, 1, 1, 9)))(
channel_135)
channel_135 = Lambda(lambda y: K.repeat_elements(y, 4, -1))(channel_135)
cv_135_tmp = multiply([channel_135, cost_volume_135])
cv_135_tmp = Conv3D(feature / 2, 1, 1, 'same',
data_format='channels_last')(cv_135_tmp)
cv_135_tmp = Activation('relu')(cv_135_tmp)
cv_135_tmp = Conv3D(3, 1, 1, 'same',
data_format='channels_last')(cv_135_tmp)
cv_135_tmp = Activation('sigmoid')(cv_135_tmp)
attention_135 = Lambda(lambda y: K.concatenate([
y[:, :, :, :, 0:1], y[:, :, :, :, 0:1], y[:, :, :, :, 0:1],
y[:, :, :, :, 0:1], y[:, :, :, :, 1:2], y[:, :, :, :, 2:3],
y[:, :, :, :, 2:3], y[:, :, :, :, 2:3], y[:, :, :, :, 2:3]
],
axis=-1))(cv_135_tmp)
attention_135 = Lambda(lambda y: K.repeat_elements(y, 4, -1))(
attention_135)
cv_135_multi = multiply([attention_135, cost_volume_135])
dres3 = convbn_3d(cv_135_multi, feature, 3, 1)
dres3 = Activation('relu')(dres3)
dres3 = convbn_3d(cv_135_multi, feature / 2, 3, 1)
dres3 = Activation('relu')(dres3)
dres3 = convbn_3d(cv_135_multi, feature / 2, 3, 1)
dres3 = Activation('relu')(dres3)
dres3 = convbn_3d(cv_135_multi, feature / 4, 3, 1)
dres3 = Activation('relu')(dres3)
dres3 = convbn_3d(dres3, 1, 3, 1)
cost3 = Activation('relu')(dres3)
cost3 = Lambda(lambda x: K.permute_dimensions(K.squeeze(x, -1),
(0, 2, 3, 1)))(cost3)
return cost3, cv_135_multi
def channel_attention_mirror(cost_volume):
x = GlobalAveragePooling3D()(cost_volume)
x = Lambda(
lambda y: K.expand_dims(K.expand_dims(K.expand_dims(y, 1), 1), 1))(x)
x = Conv3D(170, 1, 1, 'same')(x)
x = Activation('relu')(x)
x = Conv3D(25, 1, 1, 'same')(x)
x = Activation('sigmoid')(x)
x = Lambda(lambda y: K.reshape(y, (K.shape(y)[0], 5, 5)))(x)
x = Lambda(lambda y: tf.pad(y, [[0, 0], [0, 4], [0, 4]], 'REFLECT'))(x)
attention = Lambda(lambda y: K.reshape(y, (K.shape(y)[0], 1, 1, 1, 81)))(x)
x = Lambda(lambda y: K.repeat_elements(y, 4, -1))(attention)
return multiply([x, cost_volume]), attention
def spatial_attention(cost_volume):
feature = 4 * 9
k = 9
label = 9
dres0 = convbn_3d(cost_volume, feature / 2, 3, 1)
dres0 = Activation('relu')(dres0)
dres0 = convbn_3d(dres0, 1, 3, 1)
cost0 = Activation('relu')(dres0)
cost0 = Lambda(lambda x: K.permute_dimensions(K.squeeze(x, -1),
(0, 2, 3, 1)))(cost0)
cost1 = convbn(cost0, label // 2, (1, k), 1, 1)
cost1 = Activation('relu')(cost1)
cost1 = convbn(cost1, 1, (k, 1), 1, 1)
cost1 = Activation('relu')(cost1)
cost2 = convbn(cost0, label // 2, (k, 1), 1, 1)
cost2 = Activation('relu')(cost2)
cost2 = convbn(cost2, 1, (1, k), 1, 1)
cost2 = Activation('relu')(cost2)
cost = add([cost1, cost2])
cost = Activation('sigmoid')(cost)
cost = Lambda(lambda y: K.repeat_elements(K.expand_dims(y, 1), 9, 1))(cost)
cost = Lambda(lambda y: K.repeat_elements(y, feature, 4))(cost)
return multiply([cost, cost_volume])
def atari_qnet(input_shape, num_actions, net_name, net_size):
net_name = net_name.lower()
# input state
state = Input(shape=input_shape)
# convolutional layers
conv1_32 = Conv2D(32, (8, 8), strides=(4, 4), activation='relu')
conv2_64 = Conv2D(64, (4, 4), strides=(2, 2), activation='relu')
conv3_64 = Conv2D(64, (3, 3), strides=(1, 1), activation='relu')
# if recurrent net then change input shape
if 'drqn' in net_name:
# recurrent net (drqn)
lambda_perm_state = lambda x: K.permute_dimensions(x, [0, 3, 1, 2])
perm_state = Lambda(lambda_perm_state)(state)
dist_state = Lambda(lambda x: K.stack([x], axis=4))(perm_state)
# extract features with `TimeDistributed` wrapped convolutional layers
dist_conv1 = TimeDistributed(conv1_32)(dist_state)
dist_conv2 = TimeDistributed(conv2_64)(dist_conv1)
dist_convf = TimeDistributed(conv3_64)(dist_conv2)
feature = TimeDistributed(Flatten())(dist_convf)
elif 'dqn' in net_name:
# fully connected net (dqn)
# extract features with convolutional layers
conv1 = conv1_32(state)
conv2 = conv2_64(conv1)
convf = conv3_64(conv2)
feature = Flatten()(convf)
# network type. Dense for dqn; LSTM or GRU for drqn
if 'lstm' in net_name:
net_type = LSTM
elif 'gru' in net_name:
net_type = GRU
else:
net_type = Dense
# dueling or regular dqn/drqn
if 'dueling' in net_name:
value1 = net_type(net_size, activation='relu')(feature)
adv1 = net_type(net_size, activation='relu')(feature)
value2 = Dense(1)(value1)
adv2 = Dense(num_actions)(adv1)
mean_adv2 = Lambda(lambda x: K.mean(x, axis=1))(adv2)
ones = K.ones([1, num_actions])
lambda_exp = lambda x: K.dot(K.expand_dims(x, axis=1), -ones)
exp_mean_adv2 = Lambda(lambda_exp)(mean_adv2)
sum_adv = add([exp_mean_adv2, adv2])
exp_value2 = Lambda(lambda x: K.dot(x, ones))(value2)
q_value = add([exp_value2, sum_adv])
else:
hid = net_type(net_size, activation='relu')(feature)
q_value = Dense(num_actions)(hid)
# build model
return Model(inputs=state, outputs=q_value)
def channel_attention(cost_volume):
x = GlobalAveragePooling3D()(cost_volume)
x = Lambda(
lambda y: K.expand_dims(K.expand_dims(K.expand_dims(y, 1), 1), 1))(x)
x = Conv3D(170, 1, 1, 'same')(x)
x = Activation('relu')(x)
x = Conv3D(15, 1, 1, 'same')(x) # [B, 1, 1, 1, 15]
x = Activation('sigmoid')(x)
# 15 -> 25
# 0 1 2 3 4
# 5 6 7 8
# 9 10 11
# 12 13
# 14
#
# 0 1 2 3 4
# 1 5 6 7 8
# 2 6 9 10 11
# 3 7 10 12 13
# 4 8 11 13 14
x = Lambda(lambda y: K.concatenate([
y[:, :, :, :, 0:5], y[:, :, :, :, 1:2], y[:, :, :, :, 5:9],
y[:, :, :, :, 2:3], y[:, :, :, :, 6:7], y[:, :, :, :, 9:12],
y[:, :, :, :, 3:4], y[:, :, :, :, 7:8], y[:, :, :, :, 10:11],
y[:, :, :, :, 12:14], y[:, :, :, :, 4:5], y[:, :, :, :, 8:9],
y[:, :, :, :, 11:12], y[:, :, :, :, 13:15]
],
axis=-1))(x)
x = Lambda(lambda y: K.reshape(y, (K.shape(y)[0], 5, 5)))(x)
x = Lambda(lambda y: tf.pad(y, [[0, 0], [0, 4], [0, 4]], 'REFLECT'))(x)
attention = Lambda(lambda y: K.reshape(y, (K.shape(y)[0], 1, 1, 1, 81)))(x)
x = Lambda(lambda y: K.repeat_elements(y, 4, -1))(attention)
return multiply([x, cost_volume]), attention
def _attention_model(self, a, h_prev, Ex_t):
with tf.variable_scope(self.my_scope) as var_scope:
with tf.name_scope(var_scope.original_name_scope):
with tf.variable_scope('AttentionModel'):
CONF = self.C
B = self.ActualBatchSize
L = CONF.L
D = CONF.D
h = h_prev
n = self.output_size
m = CONF.m
self.assertOutputShape(h_prev)
assert K.int_shape(a) == (B, L, D)
assert K.int_shape(h_prev) == (B, n)
assert K.int_shape(Ex_t) == (B, m)
if (CONF.att_model == 'MLP_shared') or (CONF.att_model
== '1x1_conv'):
"""
Here we'll effectively create L MLP stacks all sharing the same weights. Each
stack receives a concatenated vector of a(l) and h as input.
"""
# h.shape = (B,n). Expand it to (B,1,n) and then broadcast to (B,L,n) in order
# to concatenate with feature vectors of 'a' whose shape=(B,L,D)
h = tf.identity(K.tile(K.expand_dims(h, axis=1),
(1, L, 1)),
name='h_t-1')
a = tf.identity(a, name='a')
if CONF.feed_clock_to_att:
assert CONF.build_scanning_RNN, 'Attention model can take Ex_t only in a scanning-LSTM'
# Ex_t.shape = (B,m). Expand it to (B,1,m) and then broadcast to (B,L,m) in order
# to concatenate with feature vectors of 'a' whose shape=(B,L,D)
x = tf.identity(K.tile(K.expand_dims(Ex_t, axis=1),
(1, L, 1)),
name='Ex_t')
# Concatenate a, h nd x. Final shape = (B, L, D+n+m)
att_inp = tf.concat([a, h, x], -1,
name='ai_h_x') # (B, L, D+n+m)
assert K.int_shape(att_inp) == (B, L, D + n + m)
else:
# Concatenate a and h. Final shape = (B, L, D+n)
att_inp = tf.concat([a, h], -1,
name='ai_h') # (B, L, D+n)
assert K.int_shape(att_inp) == (B, L, D + n)
if CONF.att_model == 'MLP_shared':
## For #layers > 1 this implementation will endup being different than the paper's implementation because they only
## Below is how it is implemented in the code released by the authors of the paper
## for i in range(1, CONF.att_a_layers+1):
## if not last_layer:
## a = Dense(CONF['att_a_%d_n'%(i,)], activation=tanh)(a)
## else: # last-layer
## a = AffineTransform(CONF['att_a_%d_n'%(i,)])(a)
## h = AffineTransform(CONF['att_h_%d_n'%(i,)])(h)
## ah = a + K.expand_dims(h, axis=1)
## ah = tanh(ah)
## alpha = Dense(softmax_layer_params, activation=softmax)(ah)
alpha_1_ = tfc.MLPStack(CONF.att_layers)(
att_inp) # (B, L, 1)
assert K.int_shape(alpha_1_) == (B, L, 1
) # (B, L, 1)
alpha_ = K.squeeze(alpha_1_,
axis=2) # output shape = (B, L)
assert K.int_shape(alpha_) == (B, L)
elif CONF.att_model == '1x1_conv':
"""
NOTE: The above model ('MLP_shared') tantamounts to a
1x1 convolution on the Lx1 shaped (L=H.W) convnet features with num_channels=D i.e. an input shape of (H,W,C) or (1,L,D).
Using 'dimctx' kernels of size (1,1) and stride=1 resulting in an output shape of (1,L,dimctx) [or (B, L, 1, dimctx) with the batch dimension included].
This option provides such a convnet layer implementation (which turns out not to be faster than MLP_shared).
"""
att_inp = tf.expand_dims(att_inp, axis=1)
alpha_1_ = tfc.ConvStack(
CONF.att_layers,
(B, 1, L, D + self.output_size))(att_inp)
assert K.int_shape(alpha_1_) == (B, 1, L, 1)
alpha_ = tf.squeeze(alpha_1_, axis=[1,
3]) # (B, L)
assert K.int_shape(alpha_) == (B, L)
elif CONF.att_model == 'MLP_full': # MLP: weights not shared across L
## concatenate a and h_prev and pass them through a MLP. This is different than the theano
## implementation of the paper because we flatten a from (B,L,D) to (B,L*D). Hence each element
## of the L*D vector receives its own weight because the effective weight matrix here would be
## shape (L*D, num_dense_units) as compared to (D, num_dense_units) as in the shared_weights case
## Concatenate a and h. Final shape will be (B, L*D+n)
with tf.variable_scope('a_h'):
a_ = K.batch_flatten(a) # (B, L*D)
a_.set_shape(
(B, L * D)) # Flatten loses shape info
if CONF.build_scanning_RNN and CONF.feed_clock_to_att:
assert CONF.build_scanning_RNN, 'Attention model can take Ex_t only in a scanning-LSTM'
att_inp = tf.concat(
[a_, h, Ex_t], -1,
name="a_h_x") # (B, L*D + n + m)
assert K.int_shape(att_inp) == (
B, L * D + self.output_size +
m), 'shape %s != %s' % (
K.int_shape(att_inp),
(B, L * D + self.output_size + m))
else:
att_inp = tf.concat([a_, h], -1,
name="a_h") # (B, L*D + n)
assert K.int_shape(att_inp) == (
B, L * D +
self.output_size), 'shape %s != %s' % (
K.int_shape(att_inp),
(B, L * D + self.output_size))
alpha_ = tfc.MLPStack(CONF.att_layers)(
att_inp) # (B, L)
assert K.int_shape(alpha_) == (B, L)
else:
raise AttributeError(
'Invalid value of att_model param: %s' %
CONF.att_model)
## Softmax
alpha = tf.identity(tf.nn.softmax(alpha_), name='alpha')
assert K.int_shape(alpha) == (B, L)
## Attention Modulator: Beta
if CONF.build_att_modulator:
beta = tfc.MLPStack(CONF.att_modulator,
self.batch_output_shape)(h_prev)
beta = tf.identity(beta, name='beta')
else:
beta = tf.constant(1., shape=(B, 1), dtype=CONF.dtype)
assert K.int_shape(beta) == (B, 1)
return alpha, beta
