def apply_gradients(self, grads_and_vars, global_step=None, name=None):
with tf.init_scope():
self._create_slots([v for (_, v) in grads_and_vars])
accums = []
variables = []
for g, v in grads_and_vars:
accum = self.get_slot(v, 'grad_accum')
variables.append(v)
if isinstance(g, tf.IndexedSlices):
scaled_grad = tf.IndexedSlices(
g.values / self._grad_steps, g.indices, dense_shape=g.dense_shape)
accums.append(accum.assign_add(scaled_grad)) # pytype: disable=attribute-error
else:
accums.append(accum.assign_add(g / self._grad_steps)) # pytype: disable=attribute-error
def _apply_and_zero():
apply_op = self._opt.apply_gradients(list(zip(accums, variables)))
with tf.control_dependencies([apply_op]):
zero_op = [tf.assign(accum, tf.zeros_like(accum)) for accum in accums]
return tf.group(zero_op, tf.assign_add(self._counter, 1))
def _accum():
return tf.group(accums)
accum_step = tf.cond(
tf.equal(tf.mod(global_step, self._grad_steps), self._grad_steps - 1),
_apply_and_zero, _accum)
with tf.control_dependencies([accum_step]):
global_step = tf.assign_add(global_step, 1)
return tf.group(global_step)
def __call__(self, inputs, start_index=None):
dtype = inputs.dtype
inputs_shape = tf.shape(inputs)
batch_size = inputs_shape[0]
length = inputs_shape[1]
channels = inputs_shape[2]
if start_index is None:
start_index = tf.zeros((batch_size, 1), tf.int32)
position = tf.expand_dims(tf.range(length), 0)
position = tf.tile(position, [batch_size, 1]) + start_index
position = tf.cast(position, dtype)
num_timescales = channels // 2
log_timescale_increment = (
math.log(_MAX_TIMESCALE / _MIN_TIMESCALE) /
tf.maximum(tf.cast(num_timescales, dtype) - 1, 1))
inv_timescales = _MIN_TIMESCALE * tf.exp(
tf.cast(tf.range(num_timescales), dtype) *
-log_timescale_increment)
scaled_time = tf.expand_dims(position, 2) * tf.reshape(
inv_timescales, [1, 1, -1])
signal = tf.concat([tf.sin(scaled_time), tf.cos(scaled_time)], axis=2)
signal = tf.pad(signal, [[0, 0], [0, 0], [0, tf.mod(channels, 2)]])
signal = tf.reshape(signal, [-1, length, channels])
return inputs + signal
def _finish(self, update_ops, name_scope):
"""Updates beta_power variables every n batches and incrs counter."""
iter_ = self._get_iter_variable()
beta1_power, beta2_power = self._get_beta_accumulators()
with tf.control_dependencies(update_ops):
with tf.colocate_with(iter_):
def update_beta_op():
update_beta1 = beta1_power.assign(
beta1_power * self._beta1_t,
use_locking=self._use_locking)
update_beta2 = beta2_power.assign(
beta2_power * self._beta2_t,
use_locking=self._use_locking)
return tf.group(update_beta1, update_beta2)
maybe_update_beta = tf.cond(tf.equal(iter_, 0), update_beta_op,
tf.no_op)
with tf.control_dependencies([maybe_update_beta]):
# TODO(cuong): It is suboptimal here because we have to cast twice
# (float to int, and then int to float)
update_iter = iter_.assign(tf.cast(
tf.mod(tf.cast(iter_ + 1.0, tf.int32), self._n_t),
tf.float32),
use_locking=self._use_locking)
return tf.group(*update_ops + [update_iter, maybe_update_beta],
name=name_scope)
def check_integrity_and_batch(*datasets):
"""Checks whether a sequence of frames are from the same video.
Args:
*datasets: datasets each skipping 1 frame from the previous one.
Returns:
batched data and the integrity flag.
"""
not_broken = tf.constant(True)
if "frame_number" in datasets[0]:
frame_numbers = [dataset["frame_number"][0] for dataset in datasets]
not_broken = tf.equal(frame_numbers[-1] - frame_numbers[0],
num_frames - 1)
if self.only_keep_videos_from_0th_frame:
not_broken = tf.logical_and(not_broken, tf.equal(
frame_numbers[0], 0))
if self.avoid_overlapping_frames:
non_overlap = tf.equal(tf.mod(frame_numbers[0], num_frames), 0)
not_broken = tf.logical_and(not_broken, non_overlap)
else:
tf.logging.warning("use_not_breaking_batching is True but "
"no frame_number is in the dataset.")
features = {}
for key in datasets[0].keys():
values = [dataset[key] for dataset in datasets]
batch = tf.stack(values)
features[key] = batch
return features, not_broken
def unwrap(p, discont=np.pi, axis=-1):
"""Unwrap a cyclical phase tensor.
Args:
p: Phase tensor.
discont: Float, size of the cyclic discontinuity.
axis: Axis of which to unwrap.
Returns:
unwrapped: Unwrapped tensor of same size as input.
"""
dd = diff(p, axis=axis)
ddmod = tf.mod(dd + np.pi, 2.0 * np.pi) - np.pi
idx = tf.logical_and(tf.equal(ddmod, -np.pi), tf.greater(dd, 0))
ddmod = tf.where(idx, tf.ones_like(ddmod) * np.pi, ddmod)
ph_correct = ddmod - dd
idx = tf.less(tf.abs(dd), discont)
ddmod = tf.where(idx, tf.zeros_like(ddmod), dd)
ph_cumsum = tf.cumsum(ph_correct, axis=axis)
shape = p.get_shape().as_list()
shape[axis] = 1
ph_cumsum = tf.concat([tf.zeros(shape, dtype=p.dtype), ph_cumsum],
axis=axis)
unwrapped = p + ph_cumsum
return unwrapped
def get_timing_signal_1d_given_position(channels,
position,
min_timescale=1.0,
max_timescale=1.0e4):
"""Get sinusoids of diff frequencies, with timing position given.
Adapted from add_timing_signal_1d_given_position in
//third_party/py/tensor2tensor/layers/common_attention.py
Args:
channels: scalar, size of timing embeddings to create. The number of
different timescales is equal to channels / 2.
position: a Tensor with shape [batch, seq_len]
min_timescale: a float
max_timescale: a float
Returns:
a Tensor of timing signals [batch, seq_len, channels]
"""
num_timescales = channels // 2
log_timescale_increment = (
math.log(float(max_timescale) / float(min_timescale)) /
(tf.to_float(num_timescales) - 1))
inv_timescales = min_timescale * tf.exp(
tf.to_float(tf.range(num_timescales)) * -log_timescale_increment)
scaled_time = (tf.expand_dims(tf.to_float(position), 2) *
tf.expand_dims(tf.expand_dims(inv_timescales, 0), 0))
signal = tf.concat([tf.sin(scaled_time), tf.cos(scaled_time)], axis=2)
signal = tf.pad(signal, [[0, 0], [0, 0], [0, tf.mod(channels, 2)]])
return signal
def extend_partial_state(self, JCK, r):
"""
Extends partial state by sampling two states to coalesce (Gumbel-max trick to sample without replacement)
JumpChain (JC_K) is a tensor formed from a numpy array of lists of strings, returns a new JumpChain tensor
"""
# Compute combinatorial term
# pdb.set_trace()
q = 1 / ncr(self.N - r, 2)
data = tf.reshape(tf.range((self.N - r) * self.K), (self.K, self.N - r))
data = tf.mod(data, (self.N - r))
data = tf.cast(data, dtype=tf.float32)
# Gumbel-max trick to sample without replacement
z = -tf.math.log(-tf.math.log(tf.random.uniform(tf.shape(data), 0, 1)))
top_values, coalesced_indices = tf.nn.top_k(data + z, 2)
bottom_values, remaining_indices = tf.nn.top_k(tf.negative(data + z), self.N - r - 2)
JC_keep = tf.gather(tf.reshape(JCK, [self.K * (self.N - r)]), remaining_indices)
particles = tf.gather(tf.reshape(JCK, [self.K * (self.N - r)]), coalesced_indices)
particle1 = particles[:, 0]
particle2 = particles[:, 1]
# Form new state
particle_coalesced = particle1 + '+' + particle2
# Form new Jump Chain
JCK = tf.concat([JC_keep, tf.expand_dims(particle_coalesced, axis=1)], axis=1)
#return particle1, particle2, particle_coalesced, coalesced_indices, remaining_indices, q, JCK
return coalesced_indices, remaining_indices, q, JCK
def _link(self, h_clock, x, **kwargs):
# Sanity check
self._check_state(h_clock, (self._state_size, 1))
h, clock = h_clock
# Execute all modules (TODO: too much of brutal force)
h_x = tf.concat([h, x], axis=1, name='h_x')
results = []
for m in self._modules:
assert isinstance(m, Module)
results.append(m.execute(h_x))
result = tf.concat(results, axis=1)
# Bias
bias = None
if self._use_bias: bias = self._get_bias('bias', self._state_size)
# Calculate f(h * Wh + x * Wx + b)
y = self._activation(tf.nn.bias_add(result, bias))
# Calculate and apply mask
mask = tf.cast(tf.equal(tf.mod(clock, self._periods), 0), tf.float32)
# mask = tf.Print(mask, [clock, mask, self._periods], '[Clock, Mask] = ')
updated = tf.multiply(y, mask)
new_h = updated + tf.multiply(1.0 - mask, h)
# Return output and new state
new_clock = tf.add(clock, 1)
return new_h, (new_h, new_clock)
def _maybe_apply_grads_and_zero(self, distribution, global_step,
accum_grads, var_list):
cond_return = tf.cond(
tf.equal(tf.mod(global_step, self._grad_steps),
self._grad_steps - 1), lambda: self._apply_and_zero(
distribution, global_step, accum_grads, var_list),
lambda: self._accum(global_step))
return cond_return
def make_hash_fn():
session = tf.Session()
placeholder = tf.placeholder(tf.string, [])
hash_bucket = tf.strings.to_hash_bucket_fast(placeholder, 100000)
result = tf.mod(hash_bucket, 10)
def _hash_fn(x):
return session.run(result, feed_dict={placeholder: x})
return _hash_fn
def get_bmu(self):
square_difference = tf.square(self.input - self.weight)
distance = tf.sqrt(tf.reduce_mean(square_difference, axis=1))
bmu_index = tf.argmin(distance)
bmu_location = tf.to_float(
[tf.div(bmu_index, self.width),
tf.mod(bmu_index, self.width)])
return bmu_location
def false_fn():
"""add mutations."""
mask = tf.cast(
tf.multinomial(tf.log([[1 - mutation_rate, mutation_rate]]),
seq_len), tf.int32)[0]
possible_mutations = tf.random_uniform([seq_len],
minval=1,
maxval=4,
dtype=tf.int32)
x_new = tf.mod(x + mask * possible_mutations, 4)
return x_new
def tf_angle2class(angle):
''' Convert continuous angle to discrete class and residual.
num_class: int scalar, number of classes N
Input:
angle: rad scalar, from 0-2pi (or -pi~pi), class center at
0, 1*(2pi/N), 2*(2pi/N) ... (N-1)*(2pi/N)
Output:
class_id, int, among 0,1,...,N-1
residual_angle: float, a number such that
class*(2pi/N) + residual_angle = angle
'''
twopi = tf.constant(2.0 * np.pi)
angle = tf.mod(angle, twopi)
angle_per_class = twopi / tf.to_float(cfg.model.angles.num_bins)
shifted_angle = tf.mod(angle + angle_per_class / 2.0, twopi)
class_id = tf.to_int32(shifted_angle / angle_per_class)
residual_angle = shifted_angle - (tf.to_float(class_id) * angle_per_class +
angle_per_class / 2.0)
return class_id[:, 0], residual_angle
def get_maximum_index(response):
"""Get the index of the maximum value in the response map"""
response_shape = response.get_shape().as_list()
response_spatial_size = response_shape[-2:] # e.g. [29, 29]
length = response_spatial_size[0] * response_spatial_size[1]
# Get maximum response index (note index starts from zero)
ind_max = tf.argmax(tf.reshape(response, [-1, length]), 1)
ind_row = tf.div(ind_max, response_spatial_size[1])
ind_col = tf.mod(ind_max, response_spatial_size[1])
return ind_row, ind_col
def get_mask(inputs, reverse_mask, data_format='NHWC', dtype=tf.float32):
shape = inputs.get_shape().as_list()
if len(shape) == 2:
N = shape[-1]
range_n = tf.range(N)
odd_ind = tf.mod(range_n, 2)
odd_ind = tf.reshape(odd_ind, [-1, N])
checker = odd_ind
elif len(shape) == 4:
H = shape[2] if data_format == 'NCHW' else shape[1]
W = shape[3] if data_format == 'NCHW' else shape[2]
range_h = tf.range(H)
range_w = tf.range(W)
odd_ind_h = tf.cast(tf.mod(range_h, 2), dtype=tf.bool)
odd_ind_w = tf.cast(tf.mod(range_w, 2), dtype=tf.bool)
odd_h = tf.tile(tf.expand_dims(odd_ind_h, -1), [1, W])
odd_w = tf.tile(tf.expand_dims(odd_ind_w, 0), [H, 1])
checker = tf.logical_xor(odd_h, odd_w)
reshape = [-1, 1, H, W] if data_format == 'NCHW' else [-1, H, W, 1]
checker = tf.reshape(checker, reshape)
else:
raise ValueError(
'Invalid tensor shape. Dimension of the tensor shape must be '
'2 (NxD) or 4 (NxCxHxW or NxHxWxC), got {}.'.format(
inputs.get_shape().as_list()))
if checker.dtype != dtype:
checker = tf.cast(checker, dtype)
if reverse_mask:
checker = 1. - checker
return checker
def kfac_update_ops(self):
"""Sets up the KFAC factor update ops.
Returns:
An op that when run will run the update ops at their update frequencies.
"""
# This if-statement is a trick/hack to maintain compatibility with
# CrossShardOptimizer or other optimizers that might not call our
# custom minimize() method (that would otherwise always make the variables).
if not self._made_vars_already:
(cov_update_thunks,
inv_update_thunks) = self.make_vars_and_create_op_thunks()
logging.warning(
"It looks like apply_gradients() was called before "
"minimze() was called. This is not recommended, and you "
"should avoid using optimizer wrappers like "
"CrossShardOptimizer with K-FAC that try to bypass the "
"minimize() method. The burn-in feature won't work when "
"the class is used this way, for example. And K-FAC does "
"its own cross-relica syncronization.")
else:
(_, cov_update_thunks, _,
inv_update_thunks) = self.create_ops_and_vars_thunks()
should_do_cov_updates = tf.equal(
tf.mod(self.counter, self._cov_update_every), 0)
maybe_cov_updates = utils.smart_cond(
should_do_cov_updates,
lambda: tf.group(*(thunk() for thunk in cov_update_thunks)),
tf.no_op)
maybe_pre_update_adapt_damping = self.maybe_pre_update_adapt_damping()
with tf.control_dependencies(
[maybe_cov_updates, maybe_pre_update_adapt_damping]):
should_do_inv_updates = tf.equal(
tf.mod(self.counter, self._invert_every), 0)
maybe_inv_updates = utils.smart_cond(
should_do_inv_updates,
lambda: tf.group(*(thunk() for thunk in inv_update_thunks)),
tf.no_op)
return maybe_inv_updates
def tf_class2angle(pred_cls, residual, to_label_format=True):
''' Inverse function to angle2class.
If to_label_format, adjust angle to the range as in labels.
'''
tf_pi = tf.constant(np.pi, dtype=tf.float32)
angle_per_class = 2.0 * tf_pi / tf.to_float(cfg.model.angles.num_bins)
angle_center = tf.to_float(pred_cls) * angle_per_class
angle = angle_center + residual
if to_label_format:
angle = tf.mod(angle + tf_pi,
2.0 * tf_pi) - tf_pi # Transfer to range -pi,pi
return angle
def noise_from_step_num():
"""Quantization noise equal to (phi * (step_num + 1)) mod 1.0.
Not using random_uniform here due to a problem on TPU in that random seeds
are not respected, which may cause the parameters on different replicas
to go out-of-sync.
Returns:
a float32 scalar
"""
step = tf.to_int32(tf.train.get_or_create_global_step()) + 1
phi = ((5 ** 0.5) - 1) / 2
# Naive computation tf.mod(phi * step, 1.0) in float32 would be disastrous
# due to loss of precision when the step number gets large.
# Computation in doubles does not work on TPU, so we use this complicated
# alternative computation which does not suffer from these roundoff errors.
ret = 0.0
for i in range(30):
ret += (((phi * (2 ** i)) % 1.0) # double-precision computation in python
* to_float(tf.mod(step // (2 ** i), 2)))
return tf.mod(ret, 1.0)
def tf_popvec(y):
"""Population vector read-out in tensorflow."""
num_units = y.get_shape().as_list()[-1]
pref = np.arange(0, 2 * np.pi, 2 * np.pi / num_units) # preferences
cos_pref = np.cos(pref)
sin_pref = np.sin(pref)
temp_sum = tf.reduce_sum(y, axis=-1)
temp_cos = tf.reduce_sum(y * cos_pref, axis=-1) / temp_sum
temp_sin = tf.reduce_sum(y * sin_pref, axis=-1) / temp_sum
loc = tf.atan2(temp_sin, temp_cos)
return tf.mod(loc, 2 * np.pi)
def test(self, restore_model, save_dir, is_normalize_img=True):
from test_datasets import BasicDataset
dataset = BasicDataset(data_list_file=self.dataset_config['data_list_file'], img_dir=self.dataset_config['img_dir'], is_normalize_img=is_normalize_img)
save_name_list = dataset.data_list[:, -1]
iterator = dataset.create_one_shot_iterator(dataset.data_list, num_parallel_calls=self.num_input_threads)
batch_img0, batch_img1, batch_img2 = iterator.get_next()
img_shape = tf.shape(batch_img0)
h = img_shape[1]
w = img_shape[2]
new_h = tf.where(tf.equal(tf.mod(h, 64), 0), h, (tf.to_int32(tf.floor(h / 64) + 1)) * 64)
new_w = tf.where(tf.equal(tf.mod(w, 64), 0), w, (tf.to_int32(tf.floor(w / 64) + 1)) * 64)
batch_img0 = tf.image.resize_images(batch_img0, [new_h, new_w], method=1, align_corners=True)
batch_img1 = tf.image.resize_images(batch_img1, [new_h, new_w], method=1, align_corners=True)
batch_img2 = tf.image.resize_images(batch_img2, [new_h, new_w], method=1, align_corners=True)
flow_fw, flow_bw = pyramid_processing(batch_img0, batch_img1, batch_img2, train=False, trainable=False, is_scale=True)
flow_fw['full_res'] = flow_resize(flow_fw['full_res'], [h, w], method=1)
flow_bw['full_res'] = flow_resize(flow_bw['full_res'], [h, w], method=1)
flow_fw_color = flow_to_color(flow_fw['full_res'], mask=None, max_flow=256)
flow_bw_color = flow_to_color(flow_bw['full_res'], mask=None, max_flow=256)
restore_vars = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES)
saver = tf.train.Saver(var_list=restore_vars)
sess = tf.Session()
sess.run(tf.global_variables_initializer())
sess.run(iterator.initializer)
saver.restore(sess, restore_model)
if not os.path.exists(save_dir):
os.makedirs(save_dir)
for i in range(dataset.data_num):
np_flow_fw, np_flow_bw, np_flow_fw_color, np_flow_bw_color = sess.run([flow_fw['full_res'], flow_bw['full_res'], flow_fw_color, flow_bw_color])
misc.imsave('%s/flow_fw_color_%s.png' % (save_dir, save_name_list[i]), np_flow_fw_color[0])
misc.imsave('%s/flow_bw_color_%s.png' % (save_dir, save_name_list[i]), np_flow_bw_color[0])
write_flo('%s/flow_fw_%s.flo' % (save_dir, save_name_list[i]), np_flow_fw[0])
write_flo('%s/flow_bw_%s.flo' % (save_dir, save_name_list[i]), np_flow_bw[0])
print('Finish %d/%d' % (i+1, dataset.data_num))
def apply_gradients(self, grads_and_vars, global_step=None, name=None):
counter = tf.get_variable(shape=[],
initializer=tf.zeros_initializer,
name="counter")
accums = []
update_op = []
variables = []
for (grad, param) in grads_and_vars:
if grad is None or param is None:
continue
if self._grad_clipping is not None:
grad_clipping = self._steps * self._grad_clipping
grad = tf.clip_by_value(grad, -grad_clipping, grad_clipping)
variables.append(param)
param_name = self._get_variable_name(param.name)
accum = tf.get_variable(name=param_name + "/accum",
shape=param.shape.as_list(),
dtype=tf.float32,
trainable=False,
initializer=tf.zeros_initializer())
accums.append(accum)
if isinstance(grad, tf.IndexedSlices):
scaled_grad = tf.IndexedSlices(grad.values / self._steps,
grad.indices,
dense_shape=grad.dense_shape)
update_op.append(accum.assign_add(scaled_grad))
else:
update_op.append(accum.assign_add(grad / self._steps))
def _apply_and_zero():
with tf.control_dependencies(update_op):
apply_op = self._opt.apply_gradients(
list(zip(accums, variables)), global_step, name)
with tf.control_dependencies([apply_op]):
zero_op = [
tf.assign(accum, tf.zeros_like(accum))
for accum in accums + [counter]
]
return tf.group(zero_op)
def _accum():
return tf.group(update_op)
# Control that the counter has been incremented already
with tf.control_dependencies([counter.assign_add(1)]):
return tf.cond(tf.equal(tf.mod(counter, self._steps), 0),
_apply_and_zero, _accum)
def getBMU(self):
#Best Matching Unit
#Eucledian distance
square_distance = tf.square(self.input - self.weight)
distance = tf.sqrt(tf.reduce_sum(square_distance, axis=1))
#Get BMU index
bmu_index = tf.argmin(distance)
#Get the position
bmu_position = tf.to_float(
[tf.div(bmu_index, self.width),
tf.mod(bmu_index, self.width)])
return bmu_position
def _data_augmentation(self, image_blur, image_sharp):
rot = tf.random_uniform(shape=[1], minval=0, maxval=3,
dtype=tf.int32)[0]
flip_rl = tf.random_uniform(shape=[1],
minval=0,
maxval=3,
dtype=tf.int32)[0]
flip_updown = tf.random_uniform(shape=[1],
minval=0,
maxval=3,
dtype=tf.int32)[0]
image_blur = tf.image.rot90(image_blur, rot)
image_sharp = tf.image.rot90(image_sharp, rot)
rl = tf.equal(tf.mod(flip_rl, 2), 0)
ud = tf.equal(tf.mod(flip_updown, 2), 0)
image_blur = tf.cond(
rl,
true_fn=lambda: tf.image.flip_left_right(image_blur),
false_fn=lambda: (image_blur))
image_sharp = tf.cond(
rl,
true_fn=lambda: tf.image.flip_left_right(image_sharp),
false_fn=lambda: (image_sharp))
image_blur = tf.cond(ud,
true_fn=lambda: tf.image.flip_up_down(image_blur),
false_fn=lambda: (image_blur))
image_sharp = tf.cond(
ud,
true_fn=lambda: tf.image.flip_up_down(image_sharp),
false_fn=lambda: (image_sharp))
return image_blur, image_sharp
def tf_class2angle2(pred_cls, residuals, to_label_format=True):
''' Inverse function to angle2class.
If to_label_format, adjust angle to the range as in labels.
'''
residual = tf.gather_nd(
residuals,
tf.transpose(
tf.stack([tf.range(pred_cls.shape[0], dtype=tf.int64), pred_cls])))
tf_pi = tf.constant(np.pi, dtype=tf.float32)
angle_per_class = 2.0 * tf_pi / tf.to_float(cfg.model.angles.num_bins)
angle_center = tf.to_float(pred_cls) * angle_per_class
angle = angle_center + residual
if to_label_format:
angle = tf.mod(angle + tf_pi,
2.0 * tf_pi) - tf_pi # Transfer to range -pi,pi
return angle
def add_timing_signal_1d_given_position(x,
position,
min_timescale=1.0,
max_timescale=1.0e4):
channels = tf.shape(x)[2]
num_timescales = channels // 2
log_timescale_increment = (
math.log(float(max_timescale) / float(min_timescale)) /
(tf.to_float(num_timescales) - 1))
inv_timescales = min_timescale * tf.exp(
tf.to_float(tf.range(num_timescales)) * -log_timescale_increment)
scaled_time = (tf.expand_dims(tf.to_float(position), 2) *
tf.expand_dims(tf.expand_dims(inv_timescales, 0), 0))
signal = tf.concat([tf.sin(scaled_time), tf.cos(scaled_time)], axis=2)
signal = tf.pad(signal, [[0, 0], [0, 0], [0, tf.mod(channels, 2)]])
return x + signal
def _update_history(self, hist_idx, hist_size, zs, Hzs, s_hists, y_hists):
ops = []
for z, Hz, s_hist, y_hist in zip(zs, Hzs, s_hists, y_hists):
# Use the Hessian or the Gauss-Newton approximation instead of a simple
# difference of gradients.
ops.append(tf.scatter_update(s_hist, hist_idx, z))
ops.append(tf.scatter_update(y_hist, hist_idx, Hz))
hist_idx_new = tf.mod(hist_idx + 1, self._conf['hist'])
hist_size_new = tf.minimum(hist_size + 1, self._conf['hist'])
with tf.control_dependencies(ops):
ops.append(tf.assign(hist_idx, hist_idx_new))
ops.append(tf.assign(hist_size, hist_size_new))
return tf.group(ops)
def get_timing_signal_1d(length,
channels,
min_timescale=1.0,
max_timescale=1.0e4):
position = tf.to_float(tf.range(length))
num_timescales = channels // 2
log_timescale_increment = (
math.log(float(max_timescale) / float(min_timescale)) /
(tf.to_float(num_timescales) - 1))
inv_timescales = min_timescale * tf.exp(
tf.to_float(tf.range(num_timescales)) * -log_timescale_increment)
scaled_time = tf.expand_dims(position, 1) * tf.expand_dims(
inv_timescales, 0)
signal = tf.concat([tf.sin(scaled_time), tf.cos(scaled_time)], axis=1)
signal = tf.pad(signal, [[0, 0], [0, tf.mod(channels, 2)]])
signal = tf.reshape(signal, [1, length, channels])
return signal
def preprocess_example(self, example, mode, hparams):
if self.has_inputs:
# Stack encoded score components depthwise as inputs.
inputs = []
for name, _ in self.score_encoders():
inputs.append(tf.expand_dims(example[name], axis=1))
del example[name]
example['inputs'] = tf.stack(inputs, axis=2)
if self.random_crop_in_train and mode == tf_estimator.ModeKeys.TRAIN:
# Take a random crop of the training example.
assert not self.has_inputs
max_offset = tf.maximum(
tf.shape(example['targets'])[0] -
hparams.max_target_seq_length, 0)
offset = tf.cond(
max_offset > 0, lambda: tf.random_uniform(
[], maxval=max_offset, dtype=tf.int32), lambda: 0)
example['targets'] = (
example['targets'][offset:offset +
hparams.max_target_seq_length])
return example
elif self.split_in_eval and mode == tf_estimator.ModeKeys.EVAL:
# Split the example into non-overlapping segments.
assert not self.has_inputs
length = tf.shape(example['targets'])[0]
extra_length = tf.mod(length, hparams.max_target_seq_length)
examples = {
'targets':
tf.reshape(example['targets'][:length - extra_length],
[-1, hparams.max_target_seq_length, 1, 1])
}
extra_example = {
'targets':
tf.reshape(example['targets'][-extra_length:], [1, -1, 1, 1])
}
dataset = tf.data.Dataset.from_tensor_slices(examples)
extra_dataset = tf.data.Dataset.from_tensor_slices(extra_example)
return dataset.concatenate(extra_dataset)
else:
# If not cropping or splitting, do standard preprocessing.
return super(Score2PerfProblem,
self).preprocess_example(example, mode, hparams)
def _compute_top_k(x):
"""Compute top-k values for each row in a batch."""
cls_outputs_per_sample, box_outputs_per_sample = x
cls_outputs_per_sample_reshape = tf.reshape(cls_outputs_per_sample,
[-1])
_, cls_topk_indices = tf.nn.top_k(cls_outputs_per_sample_reshape,
k=anchors.MAX_DETECTION_POINTS)
# Gets top-k class and box scores.
indices = tf.div(cls_topk_indices, params['num_classes'])
classes = tf.mod(cls_topk_indices, params['num_classes'])
cls_indices = tf.stack([indices, classes], axis=1)
cls_outputs_after_topk = tf.gather_nd(cls_outputs_per_sample,
cls_indices)
box_outputs_after_topk = tf.gather_nd(box_outputs_per_sample,
tf.expand_dims(indices, 1))
return [
indices, classes, cls_outputs_after_topk, box_outputs_after_topk
]
def get_background_lr(global_step, steps_per_phase):
"""A periodic linear-warmp background learning rate schedule."""
truncated_global_step = tf.mod(global_step, steps_per_phase)
if steps_per_phase > 1:
background = tf.minimum(
tf.train.polynomial_decay(1.0,
truncated_global_step,
steps_per_phase,
end_learning_rate=0.9,
power=0.5),
tf.train.polynomial_decay(0.0,
truncated_global_step,
steps_per_phase,
end_learning_rate=10.0,
power=1.0))
else:
background = 1.0
return background