def assign_and_sample_proposals(proposed_boxes,
gt_boxes,
gt_classes,
num_samples_per_image=512,
mix_gt_boxes=True,
fg_fraction=0.25,
fg_iou_thresh=0.5,
bg_iou_thresh_hi=0.5,
bg_iou_thresh_lo=0.0):
"""Assigns the proposals with groundtruth classes and performs subsmpling.
Given `proposed_boxes`, `gt_boxes`, and `gt_classes`, the function uses the
following algorithm to generate the final `num_samples_per_image` RoIs.
1. Calculates the IoU between each proposal box and each gt_boxes.
2. Assigns each proposed box with a groundtruth class and box by choosing
the largest IoU overlap.
3. Samples `num_samples_per_image` boxes from all proposed boxes, and
returns box_targets, class_targets, and RoIs.
Args:
proposed_boxes: a tensor of shape of [batch_size, N, 4]. N is the number
of proposals before groundtruth assignment. The last dimension is the
box coordinates w.r.t. the scaled images in [ymin, xmin, ymax, xmax]
format.
gt_boxes: a tensor of shape of [batch_size, MAX_NUM_INSTANCES, 4].
The coordinates of gt_boxes are in the pixel coordinates of the scaled
image. This tensor might have padding of values -1 indicating the invalid
box coordinates.
gt_classes: a tensor with a shape of [batch_size, MAX_NUM_INSTANCES]. This
tensor might have paddings with values of -1 indicating the invalid
classes.
num_samples_per_image: a integer represents RoI minibatch size per image.
mix_gt_boxes: a bool indicating whether to mix the groundtruth boxes before
sampling proposals.
fg_fraction: a float represents the target fraction of RoI minibatch that
is labeled foreground (i.e., class > 0).
fg_iou_thresh: a float represents the IoU overlap threshold for an RoI to be
considered foreground (if >= fg_iou_thresh).
bg_iou_thresh_hi: a float represents the IoU overlap threshold for an RoI to
be considered background (class = 0 if overlap in [LO, HI)).
bg_iou_thresh_lo: a float represents the IoU overlap threshold for an RoI to
be considered background (class = 0 if overlap in [LO, HI)).
Returns:
sampled_rois: a tensor of shape of [batch_size, K, 4], representing the
coordinates of the sampled RoIs, where K is the number of the sampled
RoIs, i.e. K = num_samples_per_image.
sampled_gt_boxes: a tensor of shape of [batch_size, K, 4], storing the
box coordinates of the matched groundtruth boxes of the samples RoIs.
sampled_gt_classes: a tensor of shape of [batch_size, K], storing the
classes of the matched groundtruth boxes of the sampled RoIs.
sampled_gt_indices: a tensor of shape of [batch_size, K], storing the
indices of the sampled groudntruth boxes in the original `gt_boxes`
tensor, i.e. gt_boxes[sampled_gt_indices[:, i]] = sampled_gt_boxes[:, i].
"""
with tf.name_scope('sample_proposals'):
if mix_gt_boxes:
boxes = tf.concat([proposed_boxes, gt_boxes], axis=1)
else:
boxes = proposed_boxes
(matched_gt_boxes, matched_gt_classes, matched_gt_indices, matched_iou,
_) = box_matching(boxes, gt_boxes, gt_classes)
positive_match = tf.greater(matched_iou, fg_iou_thresh)
negative_match = tf.logical_and(
tf.greater_equal(matched_iou, bg_iou_thresh_lo),
tf.less(matched_iou, bg_iou_thresh_hi))
ignored_match = tf.less(matched_iou, 0.0)
# re-assign negatively matched boxes to the background class.
matched_gt_classes = tf.where(negative_match,
tf.zeros_like(matched_gt_classes),
matched_gt_classes)
matched_gt_indices = tf.where(negative_match,
tf.zeros_like(matched_gt_indices),
matched_gt_indices)
sample_candidates = tf.logical_and(
tf.logical_or(positive_match, negative_match),
tf.logical_not(ignored_match))
sampler = (
balanced_positive_negative_sampler.BalancedPositiveNegativeSampler(
positive_fraction=fg_fraction, is_static=True))
batch_size, _ = sample_candidates.get_shape().as_list()
sampled_indicators = []
for i in range(batch_size):
sampled_indicator = sampler.subsample(sample_candidates[i],
num_samples_per_image,
positive_match[i])
sampled_indicators.append(sampled_indicator)
sampled_indicators = tf.stack(sampled_indicators)
_, sampled_indices = tf.nn.top_k(tf.cast(sampled_indicators,
dtype=tf.int32),
k=num_samples_per_image,
sorted=True)
sampled_indices_shape = tf.shape(sampled_indices)
batch_indices = (
tf.expand_dims(tf.range(sampled_indices_shape[0]), axis=-1) *
tf.ones([1, sampled_indices_shape[-1]], dtype=tf.int32))
gather_nd_indices = tf.stack([batch_indices, sampled_indices], axis=-1)
sampled_rois = tf.gather_nd(boxes, gather_nd_indices)
sampled_gt_boxes = tf.gather_nd(matched_gt_boxes, gather_nd_indices)
sampled_gt_classes = tf.gather_nd(matched_gt_classes,
gather_nd_indices)
sampled_gt_indices = tf.gather_nd(matched_gt_indices,
gather_nd_indices)
return (sampled_rois, sampled_gt_boxes, sampled_gt_classes,
sampled_gt_indices)
def _random_bernoulli(shape, probs, dtype=tf.int64, seed=None, name=None):
"""Returns samples from a Bernoulli distribution."""
with tf1.name_scope(name, "random_bernoulli", [shape, probs]):
probs = tf.convert_to_tensor(value=probs)
random_uniform = tf.random.uniform(shape, dtype=probs.dtype, seed=seed)
return tf.cast(tf.less(random_uniform, probs), dtype)
def one_step(self, current_state, previous_kernel_results, seed=None):
"""Takes one step of the TransitionKernel.
Args:
current_state: `Tensor` or Python `list` of `Tensor`s representing the
current state(s) of the Markov chain(s).
previous_kernel_results: A (possibly nested) `tuple`, `namedtuple` or
`list` of `Tensor`s representing internal calculations made within the
previous call to this function (or as returned by `bootstrap_results`).
seed: Optional, a seed for reproducible sampling.
Returns:
next_state: `Tensor` or Python `list` of `Tensor`s representing the
next state(s) of the Markov chain(s).
kernel_results: A (possibly nested) `tuple`, `namedtuple` or `list` of
`Tensor`s representing internal calculations made within this function.
This inculdes replica states.
"""
# The code below propagates one step states of shape
# [n_replica] + batch_shape + event_shape.
#
# The step is done in three parts:
# 1) Call one_step to transition states via a tempered version of
# self.target_log_prob_fn (see _replica_target_log_prob).
# 2) Permute values in states
# 3) Update state-dependent values, such as log_probs.
#
# We chose to swap states, rather than temperatures, because...
# (i) If swapping temperatures, you *still* have to swap log_probs to
# determine acceptance, as well as states (for kernel results).
# So it's just as difficult to swap temperatures.
# (ii) If swapping temperatures, you have to take care to swap any user-
# supplied temperature related things (like step size).
# A-priori, we don't know what else will need to be swapped!
# (iii)In both cases, the kernel results need to be updated in a non-trivial
# manner....so we either special-case, or use bootstrap.
with tf.name_scope(mcmc_util.make_name(self.name, 'remc', 'one_step')):
# Force a read in case the `inverse_temperatures` is a `tf.Variable`.
inverse_temperatures = tf.convert_to_tensor(
previous_kernel_results.inverse_temperatures,
name='inverse_temperatures')
target_log_prob_for_inner_kernel = _make_replica_target_log_prob_fn(
self.target_log_prob_fn, inverse_temperatures)
# Seed handling complexity is due to users possibly expecting an old-style
# stateful seed to be passed to `self.make_kernel_fn`, and no seed
# expected by `kernel.one_step`.
# In other words:
# - We try `make_kernel_fn` without a seed first; this is the future. The
# kernel will receive a seed later, as part of `one_step`.
# - If the user code doesn't like that (Python complains about a missing
# required argument), we warn and fall back to the previous behavior.
try:
inner_kernel = self.make_kernel_fn( # pylint: disable=not-callable
target_log_prob_for_inner_kernel)
except TypeError as e:
if 'argument' not in str(e):
raise
warnings.warn(
'The `seed` argument to `ReplicaExchangeMC`s `make_kernel_fn` is '
'deprecated. `TransitionKernel` instances now receive seeds via '
'`one_step`.')
inner_kernel = self.make_kernel_fn( # pylint: disable=not-callable
target_log_prob_for_inner_kernel, self._seed_stream())
# Now that we've constructed the TransitionKernel instance:
# - If we were given a seed, we sanitize it to stateless and pass along
# to `kernel.one_step`. If it doesn't like that, we crash and propagate
# the error. Rationale: The contract is stateless sampling given
# seed, and doing otherwise would not meet it.
# - If not given a seed, we don't pass one along. This avoids breaking
# underlying kernels lacking a `seed` arg on `one_step`.
# TODO(b/159636942): Clean up after 2020-09-20.
if seed is not None:
seed = samplers.sanitize_seed(seed)
inner_seed, swap_seed, logu_seed = samplers.split_seed(
seed, n=3, salt='remc_one_step')
inner_kwargs = dict(seed=inner_seed)
else:
if self._seed_stream.original_seed is not None:
warnings.warn(mcmc_util.SEED_CTOR_ARG_DEPRECATION_MSG)
inner_kwargs = {}
swap_seed, logu_seed = samplers.split_seed(self._seed_stream())
[
pre_swap_replica_states,
pre_swap_replica_results,
] = inner_kernel.one_step(
previous_kernel_results.post_swap_replica_states,
previous_kernel_results.post_swap_replica_results,
**inner_kwargs)
pre_swap_replica_target_log_prob = _get_field(
# These are tempered log probs (have been divided by temperature).
pre_swap_replica_results,
'target_log_prob')
dtype = pre_swap_replica_target_log_prob.dtype
replica_and_batch_shape = ps.shape(
pre_swap_replica_target_log_prob)
batch_shape = replica_and_batch_shape[1:]
replica_and_batch_rank = ps.rank(pre_swap_replica_target_log_prob)
num_replica = ps.size0(inverse_temperatures)
inverse_temperatures = mcmc_util.left_justified_broadcast_to(
inverse_temperatures, replica_and_batch_shape)
# Now that each replica has done one_step, it is time to consider swaps.
# swap.shape = [n_replica], and is a "once only" permutation, meaning it
# is achievable by a sequence of pairwise permutations, where each element
# is moved at most once.
# E.g. if swaps = [1, 0, 2], we will consider swapping temperatures 0 and
# 1, keeping 2 fixed. This exact same swap is considered for *every*
# batch member. Of course some batch members may accept and some reject.
try:
swaps = tf.cast(
self.swap_proposal_fn( # pylint: disable=not-callable
num_replica,
batch_shape=batch_shape,
seed=swap_seed,
step_count=previous_kernel_results.step_count),
dtype=tf.int32)
except TypeError as e:
if 'step_count' not in str(e):
raise
warnings.warn(
'The `swap_proposal_fn` given to ReplicaExchangeMC did not accept '
'the `step_count` argument. Falling back to omitting the '
'argument. This fallback will be removed after 24-Oct-2020.'
)
swaps = tf.cast(
self.swap_proposal_fn( # pylint: disable=not-callable
num_replica,
batch_shape=batch_shape,
seed=swap_seed),
dtype=tf.int32)
null_swaps = mcmc_util.left_justified_expand_dims_like(
tf.range(num_replica, dtype=swaps.dtype), swaps)
swaps = _maybe_embed_swaps_validation(swaps, null_swaps,
self.validate_args)
# Un-temper the log probs. E.g., for replica k, at point x_k, this is
# Log[p(x_k)], and *not* Log[p_x(x_k)] = Log[p(x_k)] * beta_k.
untempered_pre_swap_replica_target_log_prob = (
pre_swap_replica_target_log_prob / inverse_temperatures)
# Since `swaps` is its own inverse permutation we automatically know the
# swap counterpart: range(num_replica). We use this idea to compute the
# acceptance in a vectorized manner at the cost of wasting roughly half
# our computation. Although we could use `unique` to solve this problem,
# we expect the cost of `unique` to be higher than the dozens of wasted
# arithmetic calculations. Worse, it'd mean we need dynamic sized Tensors
# (eg, using `tf.where(bool)`) and so we wouldn't be able to XLA compile.
# Note: diffs would normally be "proposed - current" however energy is
# flipped since `energy == -log_prob`.
energy_diff = (untempered_pre_swap_replica_target_log_prob -
mcmc_util.index_remapping_gather(
untempered_pre_swap_replica_target_log_prob,
swaps,
name='gather_swap_tlp'))
swapped_inverse_temperatures = mcmc_util.index_remapping_gather(
inverse_temperatures, swaps, name='gather_swap_temps')
inverse_temp_diff = swapped_inverse_temperatures - inverse_temperatures
# If i and j are swapping, log_accept_ratio[] i and j are equal.
log_accept_ratio = (energy_diff *
mcmc_util.left_justified_expand_dims_to(
inverse_temp_diff, replica_and_batch_rank))
log_accept_ratio = tf.where(tf.math.is_finite(log_accept_ratio),
log_accept_ratio,
tf.constant(-np.inf, dtype=dtype))
# Produce Log[Uniform] draws that are identical at swapped indices.
log_uniform = tf.math.log(
samplers.uniform(shape=replica_and_batch_shape,
dtype=dtype,
seed=logu_seed))
anchor_swaps = tf.minimum(swaps, null_swaps)
log_uniform = mcmc_util.index_remapping_gather(
log_uniform, anchor_swaps)
is_swap_accepted_mask = tf.less(log_uniform,
log_accept_ratio,
name='is_swap_accepted_mask')
def _swap_tensor(x):
return mcmc_util.choose(
is_swap_accepted_mask,
mcmc_util.index_remapping_gather(x, swaps), x)
post_swap_replica_states = [
_swap_tensor(s) for s in pre_swap_replica_states
]
expanded_null_swaps = mcmc_util.left_justified_broadcast_to(
null_swaps, replica_and_batch_shape)
is_swap_proposed = _compute_swap_notmatrix(
# Broadcast both so they have shape [num_replica] + batch_shape.
# This (i) makes them have same shape as is_swap_accepted, and
# (ii) keeps shape consistent if someday swaps has a batch shape.
expanded_null_swaps,
mcmc_util.left_justified_broadcast_to(swaps,
replica_and_batch_shape))
# To get is_swap_accepted in ordered position, we use
# _compute_swap_notmatrix on current and next replica positions.
post_swap_replica_position = _swap_tensor(expanded_null_swaps)
is_swap_accepted = _compute_swap_notmatrix(
post_swap_replica_position, expanded_null_swaps)
if self._state_includes_replicas:
post_swap_states = post_swap_replica_states
else:
post_swap_states = [s[0] for s in post_swap_replica_states]
post_swap_replica_results = _make_post_swap_replica_results(
pre_swap_replica_results, inverse_temperatures,
swapped_inverse_temperatures, is_swap_accepted_mask,
_swap_tensor)
if mcmc_util.is_list_like(current_state):
# We *always* canonicalize the states in the kernel results.
states = post_swap_states
else:
states = post_swap_states[0]
post_swap_kernel_results = ReplicaExchangeMCKernelResults(
post_swap_replica_states=post_swap_replica_states,
pre_swap_replica_results=pre_swap_replica_results,
post_swap_replica_results=post_swap_replica_results,
is_swap_proposed=is_swap_proposed,
is_swap_accepted=is_swap_accepted,
is_swap_proposed_adjacent=_sub_diag(is_swap_proposed),
is_swap_accepted_adjacent=_sub_diag(is_swap_accepted),
# Store the original pkr.inverse_temperatures in case its a
# `tf.Variable`.
inverse_temperatures=previous_kernel_results.
inverse_temperatures,
swaps=swaps,
step_count=previous_kernel_results.step_count + 1,
seed=samplers.zeros_seed() if seed is None else seed,
)
return states, post_swap_kernel_results
def _sample_n(self, n, seed=None):
probs = self._probs_parameter_no_checks()
new_shape = tf.concat([[n], tf.shape(probs)], 0)
uniform = tf.random.uniform(new_shape, seed=seed, dtype=probs.dtype)
sample = tf.less(uniform, probs)
return tf.cast(sample, self.dtype)
def filter_fn(source):
return tf.less(tf.shape(source)[0], max_len + 1)