def build_model(data_tensor, reuse, training, output_shape):
"""Create the hgru from Learning long-range..."""
if isinstance(output_shape, list):
output_shape = output_shape[0]
with tf.variable_scope('cnn', reuse=reuse):
with tf.variable_scope('input', reuse=reuse):
x = tf.layers.conv2d(inputs=data_tensor,
filters=24,
kernel_size=11,
name='l0',
strides=(1, 1),
padding='same',
activation=tf.nn.relu,
trainable=training,
use_bias=True)
layer_hgru = hgru.hGRU('hgru_1',
x_shape=x.get_shape().as_list(),
timesteps=8,
h_ext=15,
strides=[1, 1, 1, 1],
padding='SAME',
aux={
'reuse': False,
'constrain': False
},
train=training)
h2, inh = layer_hgru.build(x)
layers = [[x, inh]]
ds_list = []
bs = data_tensor.get_shape().as_list()[0]
alpha = 5.
beta = 1.
dist = tf.distributions.Beta(alpha, beta)
for idx, acts in enumerate(zip(layers)):
(xs, inhs) = acts[0]
# xs *= mask
# inhs *= tf.abs(1 - mask)
rand = dist.sample([])
sel = tf.cast(tf.round(tf.random_uniform([], minval=1, maxval=bs - 1)),
tf.int32)
xs = tf.stop_gradient(rand * xs + ((1 - rand) * tf.roll(xs, sel, 0)))
xs = tf.reshape(xs, [bs, -1])
inhs = tf.reshape(inhs, [bs, -1])
# Augmentation
rolled_xs = []
for r in range(1, bs - 1):
rolled_xs += [tf.roll(xs, r, 0)]
xs = tf.concat([xs] + rolled_xs, 0)
inhs = tf.tile(inhs, [bs - 1, 1])
# CPC distances
# xs = tf.nn.l2_normalize(xs, -1)
# inhs = tf.nn.l2_normalize(inhs, -1)
ds = tf.reduce_sum((xs * inhs), axis=-1)
ds = tf.reshape(ds, [bs - 1, -1])
ds_list += [ds]
ds_list = tf.transpose(tf.concat(ds_list, -1))
return ds_list, {}
def build_model(data_tensor, reuse, training, output_shape):
"""Create the hgru from Learning long-range..."""
if isinstance(output_shape, list):
output_shape = output_shape[0]
with tf.variable_scope('cnn', reuse=reuse):
with tf.variable_scope('input', reuse=reuse):
o_x = tf.layers.conv2d(inputs=data_tensor,
filters=24,
kernel_size=11,
name='l0',
strides=(1, 1),
padding='same',
activation=tf.nn.relu,
trainable=training,
use_bias=True)
n, h, w, c = o_x.get_shape().as_list()
h = h - 15
w = w - 15
x = tf.image.resize_image_with_crop_or_pad(o_x, h, w)
h = h - 5
w = w - 5
x = tf.image.random_crop(x, [n, h, w, c])
layer_hgru = hgru.hGRU('hgru_1',
x_shape=x.get_shape().as_list(),
timesteps=8,
h_ext=15,
strides=[1, 1, 1, 1],
padding='SAME',
aux={
'reuse': False,
'constrain': False
},
train=training)
h2, inh = layer_hgru.build(x)
layers = [[[o_x, x], inh]]
ds_list = []
nulls = 32
for idx, acts in enumerate(zip(layers)):
(x, inhs) = acts[0]
xs = [x[1]]
x = x[0]
crop_size = o_x.get_shape().as_list()
crop_size[1] = h
crop_size[2] = w
for idx in range(nulls):
xs += [tf.random_crop(o_x, crop_size)]
xs = tf.reshape(tf.concat(xs, 0),
[crop_size[0] * nulls + crop_size[0], -1])
inhs = tf.reshape(inhs, [crop_size[0], -1])
inhs = tf.tile(inhs, [nulls + 1, 1])
# CPC distances
ds = tf.reduce_sum((xs * inhs), axis=-1)
ds = tf.reshape(ds, [crop_size[0], -1])
ds_list += [ds]
ds_list = tf.transpose(tf.concat(ds_list, -1))
return ds_list, {}
def build_model(data_tensor, reuse, training, output_shape):
"""Create the hgru from Learning long-range..."""
if isinstance(output_shape, list):
output_shape = output_shape[0]
alpha = 5.
beta = 1.
bs = data_tensor.get_shape().as_list()[0]
dist = tf.distributions.Beta(alpha, beta)
rand = dist.sample([])
sel = tf.cast(tf.round(tf.random_uniform([], minval=1, maxval=bs - 1)),
tf.int32)
with tf.variable_scope('cnn', reuse=reuse):
with tf.variable_scope('input', reuse=reuse):
o_x = tf.layers.conv2d(inputs=data_tensor,
filters=24,
kernel_size=11,
name='l0',
strides=(1, 1),
padding='same',
activation=tf.nn.relu,
trainable=training,
use_bias=True)
x = rand * o_x + ((1 - rand) * tf.roll(o_x, sel, 0))
layer_hgru = hgru.hGRU('hgru_1',
x_shape=x.get_shape().as_list(),
timesteps=8,
h_ext=15,
strides=[1, 1, 1, 1],
padding='SAME',
aux={
'reuse': False,
'constrain': False
},
train=training)
h2, inh = layer_hgru.build(x)
with tf.variable_scope('ro', reuse=reuse):
h2 = normalization.batch(bottom=h2,
renorm=False,
name='hgru_bn',
training=training)
activity = conv.readout_layer(activity=h2,
reuse=reuse,
training=training,
output_shape=output_shape)
extra_activities = {'activity': h2}
return activity, extra_activities
def build_model(data_tensor, reuse, training, output_shape):
"""Create the hgru from Learning long-range..."""
if isinstance(output_shape, list):
output_shape = output_shape[0]
with tf.variable_scope('cnn', reuse=reuse):
with tf.variable_scope('input', reuse=reuse):
x = tf.layers.conv2d(
inputs=data_tensor,
filters=24,
kernel_size=11,
name='l0',
strides=(1, 1),
padding='same',
activation=tf.nn.relu,
trainable=training,
use_bias=True)
layer_hgru = hgru.hGRU(
'hgru_1',
x_shape=x.get_shape().as_list(),
timesteps=8,
h_ext=15,
strides=[1, 1, 1, 1],
padding='SAME',
aux={'reuse': False, 'constrain': False},
train=training)
h2, inh = layer_hgru.build(x)
# h2 = normalization.batch(
# bottom=h2,
# renorm=True,
# name='hgru_bn',
# training=training)
# inh = normalization.batch(
# bottom=inh,
# renorm=True,
# name='hgru_inh',
# training=training)
layers = [[x, tf.nn.relu(inh)]]
ds_list = []
bs = data_tensor.get_shape().as_list()[0]
for idx, acts in enumerate(zip(layers)):
(xs, inhs) = acts[0]
xs = tf.reshape(xs, [bs, -1])
inhs = tf.reshape(inhs, [bs, -1])
# Augmentation
# roll_choices = tf.range(
# bs - 1,
# dtype=tf.int32)
# samples = tf.multinomial(
# tf.log([tf.ones_like(tf.cast(roll_choices, tf.float32))]), 1)
# rand_roll = roll_choices[tf.cast(samples[0][0], tf.int32)]
# rolled_xs = tf.roll(xs, rand_roll, 0)
# xs = tf.concat([xs, rolled_xs], 0)
# inhs = tf.concat([inhs, inhs], 0)
rolled_xs = []
for r in range(1, bs - 1):
rolled_xs += [tf.roll(xs, r, 0)]
xs = tf.concat([xs] + rolled_xs, 0)
inhs = tf.tile(inhs, [bs - 1, 1])
# CPC distances
# denom_xs = tf.sqrt(
# tf.reduce_sum(tf.matmul(xs, xs, transpose_b=True), axis=-1))
# denom_inhs = tf.sqrt(
# tf.reduce_sum(tf.matmul(inhs, inhs, transpose_b=True), axis=-1))
# num = tf.reduce_sum(xs * inhs, axis=-1)
# ds = num / (denom_xs * denom_inhs)
xs = tf.nn.l2_normalize(xs, 0)
inhs = tf.nn.l2_normalize(inhs, 0)
# xs = tf.nn.l2_normalize(xs, -1)
# inhs = tf.nn.l2_normalize(inhs, -1)
# ds = 1 - tf.losses.cosine_distance(
# xs,
# inhs,
# axis=-1,
# reduction=tf.losses.Reduction.NONE)
ds = tf.reduce_sum(xs * inhs, axis=-1)
ds = tf.reshape(ds, [bs - 1, -1])
# ds = tf.nn.softmax(ds, 0)
ds_list += [ds]
ds_list = tf.transpose(tf.concat(ds_list, -1))
return ds_list, {}