a fork of shap-e for gc
You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.

514 lines
19 KiB

2 years ago
from abc import ABC, abstractmethod
from dataclasses import dataclass
from functools import partial
from typing import Any, Dict, List, Optional, Tuple
import torch
from shap_e.models.nn.utils import sample_pmf
from shap_e.models.volume import Volume, VolumeRange
from shap_e.util.collections import AttrDict
from .model import NeRFModel, Query
def render_rays(
rays: torch.Tensor,
parts: List["RayVolumeIntegral"],
void_model: NeRFModel,
shared: bool = False,
prev_raw_outputs: Optional[List[AttrDict]] = None,
render_with_direction: bool = True,
importance_sampling_options: Optional[Dict[str, Any]] = None,
) -> Tuple["RayVolumeIntegralResults", List["RaySampler"], List[AttrDict]]:
"""
Perform volumetric rendering over a partition of possible t's in the union
of rendering volumes (written below with some abuse of notations)
C(r) := sum(
transmittance(t[i]) *
integrate(
lambda t: density(t) * channels(t) * transmittance(t),
[t[i], t[i + 1]],
)
for i in range(len(parts))
) + transmittance(t[-1]) * void_model(t[-1]).channels
where
1) transmittance(s) := exp(-integrate(density, [t[0], s])) calculates the
probability of light passing through the volume specified by [t[0], s].
(transmittance of 1 means light can pass freely)
2) density and channels are obtained by evaluating the appropriate
part.model at time t.
3) [t[i], t[i + 1]] is defined as the range of t where the ray intersects
(parts[i].volume \\ union(part.volume for part in parts[:i])) at the surface
of the shell (if bounded). If the ray does not intersect, the integral over
this segment is evaluated as 0 and transmittance(t[i + 1]) :=
transmittance(t[i]).
4) The last term is integration to infinity (e.g. [t[-1], math.inf]) that
is evaluated by the void_model (i.e. we consider this space to be empty).
:param rays: [batch_size x ... x 2 x 3] origin and direction.
:param parts: disjoint volume integrals.
:param void_model: use this model to integrate over the empty space
:param shared: All RayVolumeIntegrals are calculated with the same model.
:param prev_raw_outputs: Raw outputs from the previous rendering step
:return: A tuple of
- AttrDict containing the rendered `channels`, `distances`, and the `aux_losses`
- A list of importance samplers for additional fine-grained rendering
- A list of raw output for each interval
"""
if importance_sampling_options is None:
importance_sampling_options = {}
origin, direc = rays[..., 0, :], rays[..., 1, :]
if prev_raw_outputs is None:
prev_raw_outputs = [None] * len(parts)
samplers = []
raw_outputs = []
t0 = None
results = None
for part_i, prev_raw_i in zip(parts, prev_raw_outputs):
# Integrate over [t[i], t[i + 1]]
results_i = part_i.render_rays(
origin,
direc,
t0=t0,
prev_raw=prev_raw_i,
shared=shared,
render_with_direction=render_with_direction,
)
# Create an importance sampler for (optional) fine rendering
samplers.append(
ImportanceRaySampler(
results_i.volume_range, results_i.raw, **importance_sampling_options
)
)
raw_outputs.append(results_i.raw)
# Pass t[i + 1] as the start of integration for the next interval.
t0 = results_i.volume_range.next_t0()
# Combine the results from [t[0], t[i]] and [t[i], t[i+1]]
results = results_i if results is None else results.combine(results_i)
# While integrating out [t[-1], math.inf] is the correct thing to do, this
# erases a lot of useful information. Also, void_model is meant to predict
# the channels at t=math.inf.
# # Add the void background over [t[-1], math.inf] to complete integration.
# results = results.combine(
# RayVolumeIntegralResults(
# output=AttrDict(
# channels=void_model(origin, direc),
# distances=torch.zeros_like(t0),
# aux_losses=AttrDict(),
# ),
# volume_range=VolumeRange(
# t0=t0,
# t1=torch.full_like(t0, math.inf),
# intersected=torch.full_like(results.volume_range.intersected, True),
# ),
# # Void space extends to infinity. It is assumed that no light
# # passes beyond the void.
# transmittance=torch.zeros_like(results_i.transmittance),
# )
# )
results.output.channels = results.output.channels + results.transmittance * void_model(
Query(origin, direc)
)
return results, samplers, raw_outputs
@dataclass
class RayVolumeIntegralResults:
"""
Stores the relevant state and results of
integrate(
lambda t: density(t) * channels(t) * transmittance(t),
[t0, t1],
)
"""
# Rendered output and auxiliary losses
# output.channels has shape [batch_size, *inner_shape, n_channels]
output: AttrDict
"""
Optional values
"""
# Raw values contain the sampled `ts`, `density`, `channels`, etc.
raw: Optional[AttrDict] = None
# Integration
volume_range: Optional[VolumeRange] = None
# If a ray intersects, the transmittance from t0 to t1 (e.g. the
# probability that the ray passes through this volume).
# has shape [batch_size, *inner_shape, 1]
transmittance: Optional[torch.Tensor] = None
def combine(self, cur: "RayVolumeIntegralResults") -> "RayVolumeIntegralResults":
"""
Combines the integration results of `self` over [t0, t1] and
`cur` over [t1, t2] to produce a new set of results over [t0, t2] by
using a similar equation to (4) in NeRF++:
integrate(
lambda t: density(t) * channels(t) * transmittance(t),
[t0, t2]
)
= integrate(
lambda t: density(t) * channels(t) * transmittance(t),
[t0, t1]
) + transmittance(t1) * integrate(
lambda t: density(t) * channels(t) * transmittance(t),
[t1, t2]
)
"""
assert torch.allclose(self.volume_range.next_t0(), cur.volume_range.t0)
def _combine_fn(
prev_val: Optional[torch.Tensor],
cur_val: Optional[torch.Tensor],
*,
prev_transmittance: torch.Tensor,
):
assert prev_val is not None
if cur_val is None:
# cur_output.aux_losses are empty for the void_model.
return prev_val
return prev_val + prev_transmittance * cur_val
output = self.output.combine(
cur.output, combine_fn=partial(_combine_fn, prev_transmittance=self.transmittance)
)
combined = RayVolumeIntegralResults(
output=output,
volume_range=self.volume_range.extend(cur.volume_range),
transmittance=self.transmittance * cur.transmittance,
)
return combined
@dataclass
class RayVolumeIntegral:
model: NeRFModel
volume: Volume
sampler: "RaySampler"
n_samples: int
def render_rays(
self,
origin: torch.Tensor,
direction: torch.Tensor,
t0: Optional[torch.Tensor] = None,
prev_raw: Optional[AttrDict] = None,
shared: bool = False,
render_with_direction: bool = True,
) -> "RayVolumeIntegralResults":
"""
Perform volumetric rendering over the given volume.
:param position: [batch_size, *shape, 3]
:param direction: [batch_size, *shape, 3]
:param t0: Optional [batch_size, *shape, 1]
:param prev_raw: the raw outputs when using multiple levels with this model.
:param shared: means the same model is used for all RayVolumeIntegral's
:param render_with_direction: use the incoming ray direction when querying the model.
:return: RayVolumeIntegralResults
"""
# 1. Intersect the rays with the current volume and sample ts to
# integrate along.
vrange = self.volume.intersect(origin, direction, t0_lower=t0)
ts = self.sampler.sample(vrange.t0, vrange.t1, self.n_samples)
if prev_raw is not None and not shared:
# Append the previous ts now before fprop because previous
# rendering used a different model and we can't reuse the output.
ts = torch.sort(torch.cat([ts, prev_raw.ts], dim=-2), dim=-2).values
# Shape sanity checks
batch_size, *_shape, _t0_dim = vrange.t0.shape
_, *ts_shape, _ts_dim = ts.shape
# 2. Get the points along the ray and query the model
directions = torch.broadcast_to(direction.unsqueeze(-2), [batch_size, *ts_shape, 3])
positions = origin.unsqueeze(-2) + ts * directions
optional_directions = directions if render_with_direction else None
mids = (ts[..., 1:, :] + ts[..., :-1, :]) / 2
raw = self.model(
Query(
position=positions,
direction=optional_directions,
t_min=torch.cat([vrange.t0[..., None, :], mids], dim=-2),
t_max=torch.cat([mids, vrange.t1[..., None, :]], dim=-2),
)
)
raw.ts = ts
if prev_raw is not None and shared:
# We can append the additional queries to previous raw outputs
# before integration
copy = prev_raw.copy()
result = torch.sort(torch.cat([raw.pop("ts"), copy.pop("ts")], dim=-2), dim=-2)
merge_results = partial(self._merge_results, dim=-2, indices=result.indices)
raw = raw.combine(copy, merge_results)
raw.ts = result.values
# 3. Integrate the raw results
output, transmittance = self.integrate_samples(vrange, raw)
# 4. Clean up results that do not intersect with the volume.
transmittance = torch.where(
vrange.intersected, transmittance, torch.ones_like(transmittance)
)
def _mask_fn(_key: str, tensor: torch.Tensor):
return torch.where(vrange.intersected, tensor, torch.zeros_like(tensor))
def _is_tensor(_key: str, value: Any):
return isinstance(value, torch.Tensor)
output = output.map(map_fn=_mask_fn, should_map=_is_tensor)
return RayVolumeIntegralResults(
output=output,
raw=raw,
volume_range=vrange,
transmittance=transmittance,
)
def integrate_samples(
self,
volume_range: VolumeRange,
raw: AttrDict,
) -> Tuple[AttrDict, torch.Tensor]:
"""
Integrate the raw.channels along with other aux_losses and values to
produce the final output dictionary containing rendered `channels`,
estimated `distances` and `aux_losses`.
:param volume_range: Specifies the integral range [t0, t1]
:param raw: Contains a dict of function evaluations at ts. Should have
density: torch.Tensor [batch_size, *shape, n_samples, 1]
channels: torch.Tensor [batch_size, *shape, n_samples, n_channels]
aux_losses: {key: torch.Tensor [batch_size, *shape, n_samples, 1] for each key}
no_weight_grad_aux_losses: an optional set of losses for which the weights
should be detached before integration.
after the call, integrate_samples populates some intermediate calculations
for later use like
weights: torch.Tensor [batch_size, *shape, n_samples, 1] (density *
transmittance)[i] weight for each rgb output at [..., i, :].
:returns: a tuple of (
a dictionary of rendered outputs and aux_losses,
transmittance of this volume,
)
"""
# 1. Calculate the weights
_, _, dt = volume_range.partition(raw.ts)
ddensity = raw.density * dt
mass = torch.cumsum(ddensity, dim=-2)
transmittance = torch.exp(-mass[..., -1, :])
alphas = 1.0 - torch.exp(-ddensity)
Ts = torch.exp(torch.cat([torch.zeros_like(mass[..., :1, :]), -mass[..., :-1, :]], dim=-2))
# This is the probability of light hitting and reflecting off of
# something at depth [..., i, :].
weights = alphas * Ts
# 2. Integrate all results
def _integrate(key: str, samples: torch.Tensor, weights: torch.Tensor):
if key == "density":
# Omit integrating the density, because we don't need it
return None
return torch.sum(samples * weights, dim=-2)
def _is_tensor(_key: str, value: Any):
return isinstance(value, torch.Tensor)
if raw.no_weight_grad_aux_losses:
extra_aux_losses = raw.no_weight_grad_aux_losses.map(
partial(_integrate, weights=weights.detach()), should_map=_is_tensor
)
else:
extra_aux_losses = {}
output = raw.map(partial(_integrate, weights=weights), should_map=_is_tensor)
if "no_weight_grad_aux_losses" in output:
del output["no_weight_grad_aux_losses"]
output.aux_losses.update(extra_aux_losses)
# Integrating the ts yields the distance away from the origin; rename the variable.
output.distances = output.ts
del output["ts"]
del output["density"]
assert output.distances.shape == (*output.channels.shape[:-1], 1)
assert output.channels.shape[:-1] == raw.channels.shape[:-2]
assert output.channels.shape[-1] == raw.channels.shape[-1]
# 3. Reduce loss
def _reduce_loss(_key: str, loss: torch.Tensor):
return loss.view(loss.shape[0], -1).sum(dim=-1)
# 4. Store other useful calculations
raw.weights = weights
output.aux_losses = output.aux_losses.map(_reduce_loss)
return output, transmittance
def _merge_results(
self, a: Optional[torch.Tensor], b: torch.Tensor, dim: int, indices: torch.Tensor
):
"""
:param a: [..., n_a, ...]. The other dictionary containing the b's may
contain extra tensors from earlier calculations, so a can be None.
:param b: [..., n_b, ...]
:param dim: dimension to merge
:param indices: how the merged results should be sorted at the end
:return: a concatted and sorted tensor of size [..., n_a + n_b, ...]
"""
if a is None:
return None
merged = torch.cat([a, b], dim=dim)
return torch.gather(merged, dim=dim, index=torch.broadcast_to(indices, merged.shape))
class RaySampler(ABC):
@abstractmethod
def sample(self, t0: torch.Tensor, t1: torch.Tensor, n_samples: int) -> torch.Tensor:
"""
:param t0: start time has shape [batch_size, *shape, 1]
:param t1: finish time has shape [batch_size, *shape, 1]
:param n_samples: number of ts to sample
:return: sampled ts of shape [batch_size, *shape, n_samples, 1]
"""
class StratifiedRaySampler(RaySampler):
"""
Instead of fixed intervals, a sample is drawn uniformly at random from each
interval.
"""
def __init__(self, depth_mode: str = "linear"):
"""
:param depth_mode: linear samples ts linearly in depth. harmonic ensures
closer points are sampled more densely.
"""
self.depth_mode = depth_mode
assert self.depth_mode in ("linear", "geometric", "harmonic")
def sample(
self,
t0: torch.Tensor,
t1: torch.Tensor,
n_samples: int,
epsilon: float = 1e-3,
) -> torch.Tensor:
"""
:param t0: start time has shape [batch_size, *shape, 1]
:param t1: finish time has shape [batch_size, *shape, 1]
:param n_samples: number of ts to sample
:return: sampled ts of shape [batch_size, *shape, n_samples, 1]
"""
ones = [1] * (len(t0.shape) - 1)
ts = torch.linspace(0, 1, n_samples).view(*ones, n_samples).to(t0.dtype).to(t0.device)
if self.depth_mode == "linear":
ts = t0 * (1.0 - ts) + t1 * ts
elif self.depth_mode == "geometric":
ts = (t0.clamp(epsilon).log() * (1.0 - ts) + t1.clamp(epsilon).log() * ts).exp()
elif self.depth_mode == "harmonic":
# The original NeRF recommends this interpolation scheme for
# spherical scenes, but there could be some weird edge cases when
# the observer crosses from the inner to outer volume.
ts = 1.0 / (1.0 / t0.clamp(epsilon) * (1.0 - ts) + 1.0 / t1.clamp(epsilon) * ts)
mids = 0.5 * (ts[..., 1:] + ts[..., :-1])
upper = torch.cat([mids, t1], dim=-1)
lower = torch.cat([t0, mids], dim=-1)
t_rand = torch.rand_like(ts)
ts = lower + (upper - lower) * t_rand
return ts.unsqueeze(-1)
class ImportanceRaySampler(RaySampler):
"""
Given the initial estimate of densities, this samples more from
regions/bins expected to have objects.
"""
def __init__(
self, volume_range: VolumeRange, raw: AttrDict, blur_pool: bool = False, alpha: float = 1e-5
):
"""
:param volume_range: the range in which a ray intersects the given volume.
:param raw: dictionary of raw outputs from the NeRF models of shape
[batch_size, *shape, n_coarse_samples, 1]. Should at least contain
:param ts: earlier samples from the coarse rendering step
:param weights: discretized version of density * transmittance
:param blur_pool: if true, use 2-tap max + 2-tap blur filter from mip-NeRF.
:param alpha: small value to add to weights.
"""
self.volume_range = volume_range
self.ts = raw.ts.clone().detach()
self.weights = raw.weights.clone().detach()
self.blur_pool = blur_pool
self.alpha = alpha
@torch.no_grad()
def sample(self, t0: torch.Tensor, t1: torch.Tensor, n_samples: int) -> torch.Tensor:
"""
:param t0: start time has shape [batch_size, *shape, 1]
:param t1: finish time has shape [batch_size, *shape, 1]
:param n_samples: number of ts to sample
:return: sampled ts of shape [batch_size, *shape, n_samples, 1]
"""
lower, upper, _ = self.volume_range.partition(self.ts)
batch_size, *shape, n_coarse_samples, _ = self.ts.shape
weights = self.weights
if self.blur_pool:
padded = torch.cat([weights[..., :1, :], weights, weights[..., -1:, :]], dim=-2)
maxes = torch.maximum(padded[..., :-1, :], padded[..., 1:, :])
weights = 0.5 * (maxes[..., :-1, :] + maxes[..., 1:, :])
weights = weights + self.alpha
pmf = weights / weights.sum(dim=-2, keepdim=True)
inds = sample_pmf(pmf, n_samples)
assert inds.shape == (batch_size, *shape, n_samples, 1)
assert (inds >= 0).all() and (inds < n_coarse_samples).all()
t_rand = torch.rand(inds.shape, device=inds.device)
lower_ = torch.gather(lower, -2, inds)
upper_ = torch.gather(upper, -2, inds)
ts = lower_ + (upper_ - lower_) * t_rand
ts = torch.sort(ts, dim=-2).values
return ts