def sample_pdf(bins, weights, N_samples, det=False):
"""Sample additional points for training fine network
Args:
bins: int. Height in pixels.
weights: int. Width in pixels.
N_samples: float. Focal length of pinhole camera.
det
Returns:
samples: array of shape [batch_size, 3]. Depth samples for fine network
"""
weights += 1e-5
pdf = weights / F.sum(weights, axis=-1, keepdims=True)
cdf = F.cumsum(pdf, axis=-1)
# if isinstance(pdf, nn.Variable):
# cdf = nn.Variable.from_numpy_array(tf.math.cumsum(pdf.d, axis=-1))
# else:
# cdf = nn.Variable.from_numpy_array(tf.math.cumsum(pdf.data, axis=-1)).data
cdf = F.concatenate(F.constant(0, cdf[..., :1].shape), cdf, axis=-1)
if det:
u = F.arange(0., 1., 1 / N_samples)
u = F.broadcast(u[None, :], cdf.shape[:-1] + (N_samples, ))
u = u.data if isinstance(cdf, nn.NdArray) else u
else:
u = F.rand(shape=cdf.shape[:-1] + (N_samples, ))
indices = F.searchsorted(cdf, u, right=True)
# if isinstance(cdf, nn.Variable):
# indices = nn.Variable.from_numpy_array(
# tf.searchsorted(cdf.d, u.d, side='right').numpy())
# else:
# indices = nn.Variable.from_numpy_array(
# tf.searchsorted(cdf.data, u.data, side='right').numpy())
below = F.maximum_scalar(indices - 1, 0)
above = F.minimum_scalar(indices, cdf.shape[-1] - 1)
indices_g = F.stack(below, above, axis=below.ndim)
cdf_g = F.gather(cdf,
indices_g,
axis=-1,
batch_dims=len(indices_g.shape) - 2)
bins_g = F.gather(bins,
indices_g,
axis=-1,
batch_dims=len(indices_g.shape) - 2)
denom = (cdf_g[..., 1] - cdf_g[..., 0])
denom = F.where(F.less_scalar(denom, 1e-5), F.constant(1, denom.shape),
denom)
t = (u - cdf_g[..., 0]) / denom
samples = bins_g[..., 0] + t * (bins_g[..., 1] - bins_g[..., 0])
return samples
def ray_march(self, camloc, raydir, t0, t1, N, n_chunks, t_argmin=False):
# Points computation
BR, _ = t0.shape
t0 = F.reshape(t0, (BR, 1, 1))
t1 = F.reshape(t1, (BR, 1, 1))
camloc = F.reshape(camloc, (BR, 1, 3))
raydir = F.reshape(raydir, (BR, 1, 3))
step = (t1 - t0) / (N - 1)
intervals = F.reshape(F.arange(0, N), (1, N, 1))
ts = t0 + step * intervals
points = camloc + ts * raydir
points = F.reshape(points, (BR * N, 3))
# SDF computation
sdf_points = []
batch = (BR * N) // n_chunks
for r in range(0, BR * N, batch):
sdf_points.append(self.sdf(points[r:r + batch, :]))
sdf_points = F.reshape(F.concatenate(*sdf_points, axis=0), (BR, N, 1)) if n_chunks != 1 else \
F.reshape(sdf_points[0], (BR, N, 1))
# t_argmin computation
if t_argmin:
idx_min = F.min(sdf_points, axis=1, keepdims=True, only_index=True)
t_argmin = F.reshape(F.gather(ts, idx_min, axis=1, batch_dims=1),
(BR, 1))
return t_argmin
# Intersection check
points = F.reshape(points, (BR, N, 3))
sdf_pos = F.greater_equal_scalar(sdf_points[:, :-1, :], 0)
sdf_neg = F.less_equal_scalar(sdf_points[:, 1:, :], 0)
mask_hit = sdf_pos * sdf_neg
decreasing_consts = F.reshape(F.arange(N, 1, -1), (1, N - 1, 1))
vals = mask_hit * decreasing_consts
idx_max = F.max(vals, axis=1, only_index=True)
points = points[:, :-1, :]
x_hit = F.gather(points, idx_max, axis=1, batch_dims=1)
x_hit = F.reshape(x_hit, (BR, 3))
mask_hit = F.greater_scalar(F.sum(mask_hit, axis=1), 0)
mask_hit = F.reshape(mask_hit, (BR, 1))
x_hit_rm0 = x_hit
step = F.reshape(step, (BR, 1))
raydir = F.reshape(raydir, (BR, 3))
x_hit_rm1 = x_hit_rm0 + step * raydir
return x_hit_rm0, x_hit_rm1, mask_hit
def generate_z_anchor(self):
z_anchor_list = []
for _ in range(2):
z_anchor_var = F.gather(
self.init_z_var, combination(
self.n_train, self.batch_size)) + F.randn(
sigma=self.anch_std,
shape=(self.batch_size, self.latent_dim))
z_anchor_list.append(z_anchor_var)
return z_anchor_list
def _extract(a, t, x_shape=None):
"""
Extract some coefficients at specified timesteps.
If x_shape is not None,
reshape the extracted one to [batch_size, 1, 1, 1, 1, ...] corrensponding to x_shape for broadcasting purposes.
"""
B, = t.shape
out = F.gather(a, t, axis=0)
if x_shape is None:
return out
assert x_shape[0] == B
assert out.shape == (B, )
out_shape = np.array(list(x_shape))
out_shape[1:] = 1
return F.reshape(out, out_shape)
def pack_padded_sequence(padded_sequence,
lengths,
batch_first=False,
enforce_sorted=True):
r"""Pack a padded variable-length sequences.
This method packs a padded variable-length sequences.
:math:`T` is the max length over the lengths of sequences.
:math:`B` is the batch size equal to the length of the sequences.
:math:`*` is the remaining dimensions including none.
.. note::
This function **must** be used the dynamic computation mode.
Example:
.. code-block:: python
import numpy as np
import nnabla as nn
import nnabla.functions as F
import nnabla.utils.rnn as rnn_utils
nn.set_auto_forward(True)
l2v = lambda ldata: nn.Variable.from_numpy_array(np.asarray(ldata))
a = l2v([1, 1, 1, 1])
b = l2v([2, 2, 2])
c = l2v([2, 2, 2])
d = l2v([3, 3])
e = l2v([3, 3])
sequences = [a, b, c, d, e]
lengths = l2v([seq.shape[0] for seq in sequences])
padded_sequence = rnn_utils.pad_sequence(sequences)
print(padded_sequence.d)
packed_sequence = rnn_utils.pack_padded_sequence(padded_sequence, lengths)
print(packed_sequence.data.d)
print(packed_sequence.batch_sizes.d)
Args:
padded_sequence (:obj:`nnabla.Variable`): Padded sequence of (:math:`T \times B \times *`)
or (:math:`B \times T \times *`) shape.
lengths (:obj:`nnabla.Variable`): Sequence length for each batch and always resides in CPU.
batch_first (bool): `padded_sequence` is of (:math:`T`, :math:`B`, :math:`*`) shape if False,
otherwise (:math:`B`, :math:`T`, :math:`*`).
enforce_sorted (bool): Sequences are sorted by the length in a decreasing order if True. Default is True.
Returns:
:obj:`PackedSequence`
"""
if enforce_sorted:
sorted_indices = None
unsorted_indices = None
else:
# TODO: replace cuda context when the bug fix of the sort
with nn.context_scope(nn.Context()):
lengths, sorted_indices = F.sort(lengths,
axis=0,
reverse=True,
with_index=True)
B = sorted_indices.shape[0]
unsorted_indices = F.scatter_nd(F.arange(0, B),
sorted_indices.reshape((1, B)),
shape=(B, ))
axis = 0 if batch_first else 1
padded_sequence = F.gather(padded_sequence, sorted_indices, axis)
packed_sequence, batch_sizes = F.pack_padded_sequence(
padded_sequence, lengths, batch_first)
packed_sequence0 = PackedSequence()
packed_sequence0.data = packed_sequence
packed_sequence0.batch_sizes = batch_sizes
packed_sequence0.sorted_indices = sorted_indices
packed_sequence0.unsorted_indices = unsorted_indices
return packed_sequence0
def pad_packed_sequence(sequence,
batch_first=False,
padding_value=0.0,
total_length=None):
"""Pad packed sequence.
This method unpacks the packed sequqnce and pad it, the inverse operation of :func:`pack_padded_sequence`.
:math:`T_i` is the length of the :math:`i`-th Variable in the sequences.
:math:`B` is the batch size equal to the length of the sequences.
:math:`T` is the max of :math:`T_i` for all :math:`i`.
:math:`*` is the remaining dimensions including none.
.. note::
This function **must** be used the dynamic computation mode.
Example:
.. code-block:: python
import numpy as np
import nnabla as nn
import nnabla.functions as F
import nnabla.utils.rnn as rnn_utils
nn.set_auto_forward(True)
l2v = lambda ldata: nn.Variable.from_numpy_array(np.asarray(ldata))
a = l2v([3, 3])
b = l2v([2, 2, 2])
c = l2v([2, 2, 2])
d = l2v([1, 1, 1, 1])
e = l2v([3, 3])
sequences = [a, b, c, d, e]
packed_sequence = rnn_utils.pack_sequence(sequences, enforce_sorted=False)
print(packed_sequence.data.d)
print(packed_sequence.batch_sizes.d)
padded_sequence, lengths = rnn_utils.pad_packed_sequence(packed_sequence)
print(padded_sequence.d)
print(lengths.d)
Args:
sequence (:obj:`PackedSequence`): PackedSequence.
batch_first (bool): If False, output is of (:math:`T`, :math:`B`, :math:`*`) shape,
otherwise (:math:`B`, :math:`T`, :math:`*`).
padding_value (float): Padding value.
total_length (int): If not None, the outputs are padded up to the `total_length`.
If the `total_length` is less than the max length in the `sequences`,
the error is thrown.
This is normally used in the distributed training to align with
the longest sequence in a distributed system.
Returns:
:obj:`nnabla.Variable` of (:math:`T`, :math:`B`, :math:`*`) or (:math:`B`, :math:`T`, :math:`*`) shape
"""
packed_sequence = sequence.data
batch_sizes = sequence.batch_sizes
sorted_indices = sequence.sorted_indices
unsorted_indices = sequence.unsorted_indices
T = batch_sizes.shape[0]
if total_length is not None:
if total_length < T:
raise ValueError(
"`total length ({})` must be greater than or equal to the maximum length ({})."
.format(total_length, T))
padded_sequence, lengths = F.pad_packed_sequence(packed_sequence,
batch_sizes, batch_first,
padding_value,
total_length)
if unsorted_indices is not None:
axis = 0 if batch_first else 1
padded_sequence = F.gather(padded_sequence, unsorted_indices, axis)
lengths = lengths[unsorted_indices]
return padded_sequence, lengths
