PVD/train_generation.py

914 lines
34 KiB
Python
Raw Normal View History

2023-04-11 09:12:58 +00:00
import argparse
import numpy as np
import torch.distributed as dist
2021-10-19 20:54:46 +00:00
import torch.multiprocessing as mp
import torch.nn as nn
import torch.optim as optim
import torch.utils.data
from torch.distributions import Normal
from dataset.shapenet_data_pc import ShapeNet15kPointClouds
2023-04-11 09:12:58 +00:00
from model.pvcnn_generation import PVCNN2Base
2021-10-19 20:54:46 +00:00
from utils.file_utils import *
from utils.visualize import *
2023-04-11 09:12:58 +00:00
"""
2021-10-19 20:54:46 +00:00
some utils
2023-04-11 09:12:58 +00:00
"""
2021-10-19 20:54:46 +00:00
def rotation_matrix(axis, theta):
"""
Return the rotation matrix associated with counterclockwise rotation about
the given axis by theta radians.
"""
axis = np.asarray(axis)
axis = axis / np.sqrt(np.dot(axis, axis))
a = np.cos(theta / 2.0)
b, c, d = -axis * np.sin(theta / 2.0)
aa, bb, cc, dd = a * a, b * b, c * c, d * d
bc, ad, ac, ab, bd, cd = b * c, a * d, a * c, a * b, b * d, c * d
2023-04-11 09:12:58 +00:00
return np.array(
[
[aa + bb - cc - dd, 2 * (bc + ad), 2 * (bd - ac)],
[2 * (bc - ad), aa + cc - bb - dd, 2 * (cd + ab)],
[2 * (bd + ac), 2 * (cd - ab), aa + dd - bb - cc],
]
)
2021-10-19 20:54:46 +00:00
def rotate(vertices, faces):
2023-04-11 09:12:58 +00:00
"""
2021-10-19 20:54:46 +00:00
vertices: [numpoints, 3]
2023-04-11 09:12:58 +00:00
"""
2021-10-19 20:54:46 +00:00
M = rotation_matrix([0, 1, 0], np.pi / 2).transpose()
N = rotation_matrix([1, 0, 0], -np.pi / 4).transpose()
K = rotation_matrix([0, 0, 1], np.pi).transpose()
2023-04-11 09:12:58 +00:00
v, f = vertices[:, [1, 2, 0]].dot(M).dot(N).dot(K), faces[:, [1, 2, 0]]
2021-10-19 20:54:46 +00:00
return v, f
2023-04-11 09:12:58 +00:00
2021-10-19 20:54:46 +00:00
def norm(v, f):
2023-04-11 09:12:58 +00:00
v = (v - v.min()) / (v.max() - v.min()) - 0.5
2021-10-19 20:54:46 +00:00
return v, f
2023-04-11 09:12:58 +00:00
2021-10-19 20:54:46 +00:00
def getGradNorm(net):
2023-04-11 09:12:58 +00:00
pNorm = torch.sqrt(sum(torch.sum(p**2) for p in net.parameters()))
gradNorm = torch.sqrt(sum(torch.sum(p.grad**2) for p in net.parameters()))
2021-10-19 20:54:46 +00:00
return pNorm, gradNorm
def weights_init(m):
"""
xavier initialization
"""
classname = m.__class__.__name__
2023-04-11 09:12:58 +00:00
if classname.find("Conv") != -1 and m.weight is not None:
2021-10-19 20:54:46 +00:00
torch.nn.init.xavier_normal_(m.weight)
2023-04-11 09:12:58 +00:00
elif classname.find("BatchNorm") != -1:
2021-10-19 20:54:46 +00:00
m.weight.data.normal_()
m.bias.data.fill_(0)
2023-04-11 09:12:58 +00:00
"""
2021-10-19 20:54:46 +00:00
models
2023-04-11 09:12:58 +00:00
"""
2021-10-19 20:54:46 +00:00
def normal_kl(mean1, logvar1, mean2, logvar2):
"""
KL divergence between normal distributions parameterized by mean and log-variance.
"""
2023-04-11 09:12:58 +00:00
return 0.5 * (-1.0 + logvar2 - logvar1 + torch.exp(logvar1 - logvar2) + (mean1 - mean2) ** 2 * torch.exp(-logvar2))
2021-10-19 20:54:46 +00:00
def discretized_gaussian_log_likelihood(x, *, means, log_scales):
# Assumes data is integers [0, 1]
assert x.shape == means.shape == log_scales.shape
px0 = Normal(torch.zeros_like(means), torch.ones_like(log_scales))
centered_x = x - means
inv_stdv = torch.exp(-log_scales)
plus_in = inv_stdv * (centered_x + 0.5)
cdf_plus = px0.cdf(plus_in)
2023-04-11 09:12:58 +00:00
min_in = inv_stdv * (centered_x - 0.5)
2021-10-19 20:54:46 +00:00
cdf_min = px0.cdf(min_in)
2023-04-11 09:12:58 +00:00
log_cdf_plus = torch.log(torch.max(cdf_plus, torch.ones_like(cdf_plus) * 1e-12))
log_one_minus_cdf_min = torch.log(torch.max(1.0 - cdf_min, torch.ones_like(cdf_min) * 1e-12))
2021-10-19 20:54:46 +00:00
cdf_delta = cdf_plus - cdf_min
log_probs = torch.where(
2023-04-11 09:12:58 +00:00
x < 0.001,
log_cdf_plus,
torch.where(
x > 0.999, log_one_minus_cdf_min, torch.log(torch.max(cdf_delta, torch.ones_like(cdf_delta) * 1e-12))
),
)
2021-10-19 20:54:46 +00:00
assert log_probs.shape == x.shape
return log_probs
2023-04-11 09:12:58 +00:00
2021-10-19 20:54:46 +00:00
class GaussianDiffusion:
2023-04-11 09:12:58 +00:00
def __init__(self, betas, loss_type, model_mean_type, model_var_type):
2021-10-19 20:54:46 +00:00
self.loss_type = loss_type
self.model_mean_type = model_mean_type
self.model_var_type = model_var_type
assert isinstance(betas, np.ndarray)
self.np_betas = betas = betas.astype(np.float64) # computations here in float64 for accuracy
assert (betas > 0).all() and (betas <= 1).all()
2023-04-11 09:12:58 +00:00
(timesteps,) = betas.shape
2021-10-19 20:54:46 +00:00
self.num_timesteps = int(timesteps)
# initialize twice the actual length so we can keep running for eval
# betas = np.concatenate([betas, np.full_like(betas[:int(0.2*len(betas))], betas[-1])])
2023-04-11 09:12:58 +00:00
alphas = 1.0 - betas
2021-10-19 20:54:46 +00:00
alphas_cumprod = torch.from_numpy(np.cumprod(alphas, axis=0)).float()
2023-04-11 09:12:58 +00:00
alphas_cumprod_prev = torch.from_numpy(np.append(1.0, alphas_cumprod[:-1])).float()
2021-10-19 20:54:46 +00:00
self.betas = torch.from_numpy(betas).float()
self.alphas_cumprod = alphas_cumprod.float()
self.alphas_cumprod_prev = alphas_cumprod_prev.float()
# calculations for diffusion q(x_t | x_{t-1}) and others
self.sqrt_alphas_cumprod = torch.sqrt(alphas_cumprod).float()
2023-04-11 09:12:58 +00:00
self.sqrt_one_minus_alphas_cumprod = torch.sqrt(1.0 - alphas_cumprod).float()
self.log_one_minus_alphas_cumprod = torch.log(1.0 - alphas_cumprod).float()
self.sqrt_recip_alphas_cumprod = torch.sqrt(1.0 / alphas_cumprod).float()
self.sqrt_recipm1_alphas_cumprod = torch.sqrt(1.0 / alphas_cumprod - 1).float()
2021-10-19 20:54:46 +00:00
betas = torch.from_numpy(betas).float()
alphas = torch.from_numpy(alphas).float()
# calculations for posterior q(x_{t-1} | x_t, x_0)
2023-04-11 09:12:58 +00:00
posterior_variance = betas * (1.0 - alphas_cumprod_prev) / (1.0 - alphas_cumprod)
2021-10-19 20:54:46 +00:00
# above: equal to 1. / (1. / (1. - alpha_cumprod_tm1) + alpha_t / beta_t)
self.posterior_variance = posterior_variance
# below: log calculation clipped because the posterior variance is 0 at the beginning of the diffusion chain
2023-04-11 09:12:58 +00:00
self.posterior_log_variance_clipped = torch.log(
torch.max(posterior_variance, 1e-20 * torch.ones_like(posterior_variance))
)
self.posterior_mean_coef1 = betas * torch.sqrt(alphas_cumprod_prev) / (1.0 - alphas_cumprod)
self.posterior_mean_coef2 = (1.0 - alphas_cumprod_prev) * torch.sqrt(alphas) / (1.0 - alphas_cumprod)
2021-10-19 20:54:46 +00:00
@staticmethod
def _extract(a, t, x_shape):
"""
Extract some coefficients at specified timesteps,
then reshape to [batch_size, 1, 1, 1, 1, ...] for broadcasting purposes.
"""
2023-04-11 09:12:58 +00:00
(bs,) = t.shape
2021-10-19 20:54:46 +00:00
assert x_shape[0] == bs
out = torch.gather(a, 0, t)
assert out.shape == torch.Size([bs])
return torch.reshape(out, [bs] + ((len(x_shape) - 1) * [1]))
def q_mean_variance(self, x_start, t):
mean = self._extract(self.sqrt_alphas_cumprod.to(x_start.device), t, x_start.shape) * x_start
2023-04-11 09:12:58 +00:00
variance = self._extract(1.0 - self.alphas_cumprod.to(x_start.device), t, x_start.shape)
2021-10-19 20:54:46 +00:00
log_variance = self._extract(self.log_one_minus_alphas_cumprod.to(x_start.device), t, x_start.shape)
return mean, variance, log_variance
def q_sample(self, x_start, t, noise=None):
"""
Diffuse the data (t == 0 means diffused for 1 step)
"""
if noise is None:
noise = torch.randn(x_start.shape, device=x_start.device)
assert noise.shape == x_start.shape
return (
2023-04-11 09:12:58 +00:00
self._extract(self.sqrt_alphas_cumprod.to(x_start.device), t, x_start.shape) * x_start
+ self._extract(self.sqrt_one_minus_alphas_cumprod.to(x_start.device), t, x_start.shape) * noise
2021-10-19 20:54:46 +00:00
)
def q_posterior_mean_variance(self, x_start, x_t, t):
"""
Compute the mean and variance of the diffusion posterior q(x_{t-1} | x_t, x_0)
"""
assert x_start.shape == x_t.shape
posterior_mean = (
2023-04-11 09:12:58 +00:00
self._extract(self.posterior_mean_coef1.to(x_start.device), t, x_t.shape) * x_start
+ self._extract(self.posterior_mean_coef2.to(x_start.device), t, x_t.shape) * x_t
2021-10-19 20:54:46 +00:00
)
posterior_variance = self._extract(self.posterior_variance.to(x_start.device), t, x_t.shape)
2023-04-11 09:12:58 +00:00
posterior_log_variance_clipped = self._extract(
self.posterior_log_variance_clipped.to(x_start.device), t, x_t.shape
)
assert (
posterior_mean.shape[0]
== posterior_variance.shape[0]
== posterior_log_variance_clipped.shape[0]
== x_start.shape[0]
)
2021-10-19 20:54:46 +00:00
return posterior_mean, posterior_variance, posterior_log_variance_clipped
def p_mean_variance(self, denoise_fn, data, t, clip_denoised: bool, return_pred_xstart: bool):
model_output = denoise_fn(data, t)
2023-04-11 09:12:58 +00:00
if self.model_var_type in ["fixedsmall", "fixedlarge"]:
2021-10-19 20:54:46 +00:00
# below: only log_variance is used in the KL computations
model_variance, model_log_variance = {
# for fixedlarge, we set the initial (log-)variance like so to get a better decoder log likelihood
2023-04-11 09:12:58 +00:00
"fixedlarge": (
self.betas.to(data.device),
torch.log(torch.cat([self.posterior_variance[1:2], self.betas[1:]])).to(data.device),
),
"fixedsmall": (
self.posterior_variance.to(data.device),
self.posterior_log_variance_clipped.to(data.device),
),
2021-10-19 20:54:46 +00:00
}[self.model_var_type]
model_variance = self._extract(model_variance, t, data.shape) * torch.ones_like(data)
model_log_variance = self._extract(model_log_variance, t, data.shape) * torch.ones_like(data)
else:
raise NotImplementedError(self.model_var_type)
2023-04-11 09:12:58 +00:00
if self.model_mean_type == "eps":
2021-10-19 20:54:46 +00:00
x_recon = self._predict_xstart_from_eps(data, t=t, eps=model_output)
if clip_denoised:
2023-04-11 09:12:58 +00:00
x_recon = torch.clamp(x_recon, -0.5, 0.5)
2021-10-19 20:54:46 +00:00
model_mean, _, _ = self.q_posterior_mean_variance(x_start=x_recon, x_t=data, t=t)
else:
raise NotImplementedError(self.loss_type)
assert model_mean.shape == x_recon.shape == data.shape
assert model_variance.shape == model_log_variance.shape == data.shape
if return_pred_xstart:
return model_mean, model_variance, model_log_variance, x_recon
else:
return model_mean, model_variance, model_log_variance
def _predict_xstart_from_eps(self, x_t, t, eps):
assert x_t.shape == eps.shape
return (
2023-04-11 09:12:58 +00:00
self._extract(self.sqrt_recip_alphas_cumprod.to(x_t.device), t, x_t.shape) * x_t
- self._extract(self.sqrt_recipm1_alphas_cumprod.to(x_t.device), t, x_t.shape) * eps
2021-10-19 20:54:46 +00:00
)
2023-04-11 09:12:58 +00:00
""" samples """
2021-10-19 20:54:46 +00:00
def p_sample(self, denoise_fn, data, t, noise_fn, clip_denoised=False, return_pred_xstart=False):
"""
Sample from the model
"""
2023-04-11 09:12:58 +00:00
model_mean, _, model_log_variance, pred_xstart = self.p_mean_variance(
denoise_fn, data=data, t=t, clip_denoised=clip_denoised, return_pred_xstart=True
)
2021-10-19 20:54:46 +00:00
noise = noise_fn(size=data.shape, dtype=data.dtype, device=data.device)
assert noise.shape == data.shape
# no noise when t == 0
nonzero_mask = torch.reshape(1 - (t == 0).float(), [data.shape[0]] + [1] * (len(data.shape) - 1))
sample = model_mean + nonzero_mask * torch.exp(0.5 * model_log_variance) * noise
assert sample.shape == pred_xstart.shape
return (sample, pred_xstart) if return_pred_xstart else sample
2023-04-11 09:12:58 +00:00
def p_sample_loop(self, denoise_fn, shape, device, noise_fn=torch.randn, clip_denoised=True, keep_running=False):
2021-10-19 20:54:46 +00:00
"""
Generate samples
keep_running: True if we run 2 x num_timesteps, False if we just run num_timesteps
"""
assert isinstance(shape, (tuple, list))
img_t = noise_fn(size=shape, dtype=torch.float, device=device)
for t in reversed(range(0, self.num_timesteps if not keep_running else len(self.betas))):
t_ = torch.empty(shape[0], dtype=torch.int64, device=device).fill_(t)
2023-04-11 09:12:58 +00:00
img_t = self.p_sample(
denoise_fn=denoise_fn,
data=img_t,
t=t_,
noise_fn=noise_fn,
clip_denoised=clip_denoised,
return_pred_xstart=False,
)
2021-10-19 20:54:46 +00:00
assert img_t.shape == shape
return img_t
2023-04-11 09:12:58 +00:00
def p_sample_loop_trajectory(
self, denoise_fn, shape, device, freq, noise_fn=torch.randn, clip_denoised=True, keep_running=False
):
2021-10-19 20:54:46 +00:00
"""
Generate samples, returning intermediate images
Useful for visualizing how denoised images evolve over time
Args:
repeat_noise_steps (int): Number of denoising timesteps in which the same noise
is used across the batch. If >= 0, the initial noise is the same for all batch elemements.
"""
assert isinstance(shape, (tuple, list))
2023-04-11 09:12:58 +00:00
total_steps = self.num_timesteps if not keep_running else len(self.betas)
2021-10-19 20:54:46 +00:00
img_t = noise_fn(size=shape, dtype=torch.float, device=device)
imgs = [img_t]
2023-04-11 09:12:58 +00:00
for t in reversed(range(0, total_steps)):
2021-10-19 20:54:46 +00:00
t_ = torch.empty(shape[0], dtype=torch.int64, device=device).fill_(t)
2023-04-11 09:12:58 +00:00
img_t = self.p_sample(
denoise_fn=denoise_fn,
data=img_t,
t=t_,
noise_fn=noise_fn,
clip_denoised=clip_denoised,
return_pred_xstart=False,
)
if t % freq == 0 or t == total_steps - 1:
2021-10-19 20:54:46 +00:00
imgs.append(img_t)
assert imgs[-1].shape == shape
return imgs
2023-04-11 09:12:58 +00:00
"""losses"""
2021-10-19 20:54:46 +00:00
def _vb_terms_bpd(self, denoise_fn, data_start, data_t, t, clip_denoised: bool, return_pred_xstart: bool):
true_mean, _, true_log_variance_clipped = self.q_posterior_mean_variance(x_start=data_start, x_t=data_t, t=t)
model_mean, _, model_log_variance, pred_xstart = self.p_mean_variance(
2023-04-11 09:12:58 +00:00
denoise_fn, data=data_t, t=t, clip_denoised=clip_denoised, return_pred_xstart=True
)
2021-10-19 20:54:46 +00:00
kl = normal_kl(true_mean, true_log_variance_clipped, model_mean, model_log_variance)
2023-04-11 09:12:58 +00:00
kl = kl.mean(dim=list(range(1, len(data_start.shape)))) / np.log(2.0)
2021-10-19 20:54:46 +00:00
return (kl, pred_xstart) if return_pred_xstart else kl
def p_losses(self, denoise_fn, data_start, t, noise=None):
"""
Training loss calculation
"""
B, D, N = data_start.shape
assert t.shape == torch.Size([B])
if noise is None:
noise = torch.randn(data_start.shape, dtype=data_start.dtype, device=data_start.device)
assert noise.shape == data_start.shape and noise.dtype == data_start.dtype
data_t = self.q_sample(x_start=data_start, t=t, noise=noise)
2023-04-11 09:12:58 +00:00
if self.loss_type == "mse":
2021-10-19 20:54:46 +00:00
# predict the noise instead of x_start. seems to be weighted naturally like SNR
eps_recon = denoise_fn(data_t, t)
assert data_t.shape == data_start.shape
assert eps_recon.shape == torch.Size([B, D, N])
assert eps_recon.shape == data_start.shape
2023-04-11 09:12:58 +00:00
losses = ((noise - eps_recon) ** 2).mean(dim=list(range(1, len(data_start.shape))))
elif self.loss_type == "kl":
2021-10-19 20:54:46 +00:00
losses = self._vb_terms_bpd(
2023-04-11 09:12:58 +00:00
denoise_fn=denoise_fn,
data_start=data_start,
data_t=data_t,
t=t,
clip_denoised=False,
return_pred_xstart=False,
)
2021-10-19 20:54:46 +00:00
else:
raise NotImplementedError(self.loss_type)
assert losses.shape == torch.Size([B])
return losses
2023-04-11 09:12:58 +00:00
"""debug"""
2021-10-19 20:54:46 +00:00
def _prior_bpd(self, x_start):
with torch.no_grad():
B, T = x_start.shape[0], self.num_timesteps
2023-04-11 09:12:58 +00:00
t_ = torch.empty(B, dtype=torch.int64, device=x_start.device).fill_(T - 1)
2021-10-19 20:54:46 +00:00
qt_mean, _, qt_log_variance = self.q_mean_variance(x_start, t=t_)
2023-04-11 09:12:58 +00:00
kl_prior = normal_kl(
mean1=qt_mean,
logvar1=qt_log_variance,
mean2=torch.tensor([0.0]).to(qt_mean),
logvar2=torch.tensor([0.0]).to(qt_log_variance),
)
2021-10-19 20:54:46 +00:00
assert kl_prior.shape == x_start.shape
2023-04-11 09:12:58 +00:00
return kl_prior.mean(dim=list(range(1, len(kl_prior.shape)))) / np.log(2.0)
2021-10-19 20:54:46 +00:00
def calc_bpd_loop(self, denoise_fn, x_start, clip_denoised=True):
with torch.no_grad():
B, T = x_start.shape[0], self.num_timesteps
2023-04-11 09:12:58 +00:00
vals_bt_, mse_bt_ = torch.zeros([B, T], device=x_start.device), torch.zeros([B, T], device=x_start.device)
2021-10-19 20:54:46 +00:00
for t in reversed(range(T)):
t_b = torch.empty(B, dtype=torch.int64, device=x_start.device).fill_(t)
# Calculate VLB term at the current timestep
new_vals_b, pred_xstart = self._vb_terms_bpd(
2023-04-11 09:12:58 +00:00
denoise_fn,
data_start=x_start,
data_t=self.q_sample(x_start=x_start, t=t_b),
t=t_b,
clip_denoised=clip_denoised,
return_pred_xstart=True,
)
2021-10-19 20:54:46 +00:00
# MSE for progressive prediction loss
assert pred_xstart.shape == x_start.shape
2023-04-11 09:12:58 +00:00
new_mse_b = ((pred_xstart - x_start) ** 2).mean(dim=list(range(1, len(x_start.shape))))
assert new_vals_b.shape == new_mse_b.shape == torch.Size([B])
2021-10-19 20:54:46 +00:00
# Insert the calculated term into the tensor of all terms
2023-04-11 09:12:58 +00:00
mask_bt = t_b[:, None] == torch.arange(T, device=t_b.device)[None, :].float()
2021-10-19 20:54:46 +00:00
vals_bt_ = vals_bt_ * (~mask_bt) + new_vals_b[:, None] * mask_bt
mse_bt_ = mse_bt_ * (~mask_bt) + new_mse_b[:, None] * mask_bt
assert mask_bt.shape == vals_bt_.shape == vals_bt_.shape == torch.Size([B, T])
prior_bpd_b = self._prior_bpd(x_start)
total_bpd_b = vals_bt_.sum(dim=1) + prior_bpd_b
2023-04-11 09:12:58 +00:00
assert vals_bt_.shape == mse_bt_.shape == torch.Size(
[B, T]
) and total_bpd_b.shape == prior_bpd_b.shape == torch.Size([B])
2021-10-19 20:54:46 +00:00
return total_bpd_b.mean(), vals_bt_.mean(), prior_bpd_b.mean(), mse_bt_.mean()
class PVCNN2(PVCNN2Base):
sa_blocks = [
((32, 2, 32), (1024, 0.1, 32, (32, 64))),
((64, 3, 16), (256, 0.2, 32, (64, 128))),
((128, 3, 8), (64, 0.4, 32, (128, 256))),
(None, (16, 0.8, 32, (256, 256, 512))),
]
fp_blocks = [
((256, 256), (256, 3, 8)),
((256, 256), (256, 3, 8)),
((256, 128), (128, 2, 16)),
((128, 128, 64), (64, 2, 32)),
]
2023-04-11 09:12:58 +00:00
def __init__(
self,
num_classes,
embed_dim,
use_att,
dropout,
extra_feature_channels=3,
width_multiplier=1,
voxel_resolution_multiplier=1,
):
2021-10-19 20:54:46 +00:00
super().__init__(
2023-04-11 09:12:58 +00:00
num_classes=num_classes,
embed_dim=embed_dim,
use_att=use_att,
dropout=dropout,
extra_feature_channels=extra_feature_channels,
width_multiplier=width_multiplier,
voxel_resolution_multiplier=voxel_resolution_multiplier,
2021-10-19 20:54:46 +00:00
)
class Model(nn.Module):
2023-04-11 09:12:58 +00:00
def __init__(self, args, betas, loss_type: str, model_mean_type: str, model_var_type: str):
2021-10-19 20:54:46 +00:00
super(Model, self).__init__()
self.diffusion = GaussianDiffusion(betas, loss_type, model_mean_type, model_var_type)
2023-04-11 09:12:58 +00:00
self.model = PVCNN2(
num_classes=args.nc,
embed_dim=args.embed_dim,
use_att=args.attention,
dropout=args.dropout,
extra_feature_channels=0,
)
2021-10-19 20:54:46 +00:00
def prior_kl(self, x0):
return self.diffusion._prior_bpd(x0)
def all_kl(self, x0, clip_denoised=True):
2023-04-11 09:12:58 +00:00
total_bpd_b, vals_bt, prior_bpd_b, mse_bt = self.diffusion.calc_bpd_loop(self._denoise, x0, clip_denoised)
2021-10-19 20:54:46 +00:00
2023-04-11 09:12:58 +00:00
return {"total_bpd_b": total_bpd_b, "terms_bpd": vals_bt, "prior_bpd_b": prior_bpd_b, "mse_bt": mse_bt}
2021-10-19 20:54:46 +00:00
def _denoise(self, data, t):
2023-04-11 09:12:58 +00:00
B, D, N = data.shape
2021-10-19 20:54:46 +00:00
assert data.dtype == torch.float
assert t.shape == torch.Size([B]) and t.dtype == torch.int64
out = self.model(data, t)
assert out.shape == torch.Size([B, D, N])
return out
def get_loss_iter(self, data, noises=None):
B, D, N = data.shape
t = torch.randint(0, self.diffusion.num_timesteps, size=(B,), device=data.device)
if noises is not None:
2023-04-11 09:12:58 +00:00
noises[t != 0] = torch.randn((t != 0).sum(), *noises.shape[1:]).to(noises)
2021-10-19 20:54:46 +00:00
2023-04-11 09:12:58 +00:00
losses = self.diffusion.p_losses(denoise_fn=self._denoise, data_start=data, t=t, noise=noises)
2021-10-19 20:54:46 +00:00
assert losses.shape == t.shape == torch.Size([B])
return losses
2023-04-11 09:12:58 +00:00
def gen_samples(self, shape, device, noise_fn=torch.randn, clip_denoised=True, keep_running=False):
return self.diffusion.p_sample_loop(
self._denoise,
shape=shape,
device=device,
noise_fn=noise_fn,
clip_denoised=clip_denoised,
keep_running=keep_running,
)
2021-10-19 20:54:46 +00:00
2023-04-11 09:12:58 +00:00
def gen_sample_traj(self, shape, device, freq, noise_fn=torch.randn, clip_denoised=True, keep_running=False):
return self.diffusion.p_sample_loop_trajectory(
self._denoise,
shape=shape,
device=device,
noise_fn=noise_fn,
freq=freq,
clip_denoised=clip_denoised,
keep_running=keep_running,
)
2021-10-19 20:54:46 +00:00
def train(self):
self.model.train()
def eval(self):
self.model.eval()
def multi_gpu_wrapper(self, f):
self.model = f(self.model)
def get_betas(schedule_type, b_start, b_end, time_num):
2023-04-11 09:12:58 +00:00
if schedule_type == "linear":
2021-10-19 20:54:46 +00:00
betas = np.linspace(b_start, b_end, time_num)
2023-04-11 09:12:58 +00:00
elif schedule_type == "warm0.1":
2021-10-19 20:54:46 +00:00
betas = b_end * np.ones(time_num, dtype=np.float64)
warmup_time = int(time_num * 0.1)
betas[:warmup_time] = np.linspace(b_start, b_end, warmup_time, dtype=np.float64)
2023-04-11 09:12:58 +00:00
elif schedule_type == "warm0.2":
2021-10-19 20:54:46 +00:00
betas = b_end * np.ones(time_num, dtype=np.float64)
warmup_time = int(time_num * 0.2)
betas[:warmup_time] = np.linspace(b_start, b_end, warmup_time, dtype=np.float64)
2023-04-11 09:12:58 +00:00
elif schedule_type == "warm0.5":
2021-10-19 20:54:46 +00:00
betas = b_end * np.ones(time_num, dtype=np.float64)
warmup_time = int(time_num * 0.5)
betas[:warmup_time] = np.linspace(b_start, b_end, warmup_time, dtype=np.float64)
else:
raise NotImplementedError(schedule_type)
return betas
2023-04-11 09:12:58 +00:00
def get_dataset(dataroot, npoints, category):
tr_dataset = ShapeNet15kPointClouds(
root_dir=dataroot,
categories=[category],
split="train",
2021-10-19 20:54:46 +00:00
tr_sample_size=npoints,
te_sample_size=npoints,
2023-04-11 09:12:58 +00:00
scale=1.0,
2021-10-19 20:54:46 +00:00
normalize_per_shape=False,
normalize_std_per_axis=False,
2023-04-11 09:12:58 +00:00
random_subsample=True,
)
te_dataset = ShapeNet15kPointClouds(
root_dir=dataroot,
categories=[category],
split="val",
2021-10-19 20:54:46 +00:00
tr_sample_size=npoints,
te_sample_size=npoints,
2023-04-11 09:12:58 +00:00
scale=1.0,
2021-10-19 20:54:46 +00:00
normalize_per_shape=False,
normalize_std_per_axis=False,
all_points_mean=tr_dataset.all_points_mean,
all_points_std=tr_dataset.all_points_std,
)
return tr_dataset, te_dataset
def get_dataloader(opt, train_dataset, test_dataset=None):
2023-04-11 09:12:58 +00:00
if opt.distribution_type == "multi":
2021-10-19 20:54:46 +00:00
train_sampler = torch.utils.data.distributed.DistributedSampler(
2023-04-11 09:12:58 +00:00
train_dataset, num_replicas=opt.world_size, rank=opt.rank
2021-10-19 20:54:46 +00:00
)
if test_dataset is not None:
test_sampler = torch.utils.data.distributed.DistributedSampler(
2023-04-11 09:12:58 +00:00
test_dataset, num_replicas=opt.world_size, rank=opt.rank
2021-10-19 20:54:46 +00:00
)
else:
test_sampler = None
else:
train_sampler = None
test_sampler = None
2023-04-11 09:12:58 +00:00
train_dataloader = torch.utils.data.DataLoader(
train_dataset,
batch_size=opt.bs,
sampler=train_sampler,
shuffle=train_sampler is None,
num_workers=int(opt.workers),
drop_last=True,
)
2021-10-19 20:54:46 +00:00
if test_dataset is not None:
2023-04-11 09:12:58 +00:00
test_dataloader = torch.utils.data.DataLoader(
train_dataset,
batch_size=opt.bs,
sampler=test_sampler,
shuffle=False,
num_workers=int(opt.workers),
drop_last=False,
)
2021-10-19 20:54:46 +00:00
else:
test_dataloader = None
return train_dataloader, test_dataloader, train_sampler, test_sampler
def train(gpu, opt, output_dir, noises_init):
set_seed(opt)
logger = setup_logging(output_dir)
2023-04-11 09:12:58 +00:00
if opt.distribution_type == "multi":
should_diag = gpu == 0
2021-10-19 20:54:46 +00:00
else:
should_diag = True
if should_diag:
2023-04-11 09:12:58 +00:00
(outf_syn,) = setup_output_subdirs(output_dir, "syn")
2021-10-19 20:54:46 +00:00
2023-04-11 09:12:58 +00:00
if opt.distribution_type == "multi":
2021-10-19 20:54:46 +00:00
if opt.dist_url == "env://" and opt.rank == -1:
opt.rank = int(os.environ["RANK"])
2023-04-11 09:12:58 +00:00
base_rank = opt.rank * opt.ngpus_per_node
2021-10-19 20:54:46 +00:00
opt.rank = base_rank + gpu
2023-04-11 09:12:58 +00:00
dist.init_process_group(
backend=opt.dist_backend, init_method=opt.dist_url, world_size=opt.world_size, rank=opt.rank
)
2021-10-19 20:54:46 +00:00
opt.bs = int(opt.bs / opt.ngpus_per_node)
opt.workers = 0
2023-04-11 09:12:58 +00:00
opt.saveIter = int(opt.saveIter / opt.ngpus_per_node)
2021-10-19 20:54:46 +00:00
opt.diagIter = int(opt.diagIter / opt.ngpus_per_node)
opt.vizIter = int(opt.vizIter / opt.ngpus_per_node)
2023-04-11 09:12:58 +00:00
""" data """
2021-10-19 20:54:46 +00:00
train_dataset, _ = get_dataset(opt.dataroot, opt.npoints, opt.category)
dataloader, _, train_sampler, _ = get_dataloader(opt, train_dataset, None)
2023-04-11 09:12:58 +00:00
"""
2021-10-19 20:54:46 +00:00
create networks
2023-04-11 09:12:58 +00:00
"""
2021-10-19 20:54:46 +00:00
betas = get_betas(opt.schedule_type, opt.beta_start, opt.beta_end, opt.time_num)
model = Model(opt, betas, opt.loss_type, opt.model_mean_type, opt.model_var_type)
2023-04-11 09:12:58 +00:00
if opt.distribution_type == "multi": # Multiple processes, single GPU per process
2021-10-19 20:54:46 +00:00
def _transform_(m):
2023-04-11 09:12:58 +00:00
return nn.parallel.DistributedDataParallel(m, device_ids=[gpu], output_device=gpu)
2021-10-19 20:54:46 +00:00
torch.cuda.set_device(gpu)
model.cuda(gpu)
model.multi_gpu_wrapper(_transform_)
2023-04-11 09:12:58 +00:00
elif opt.distribution_type == "single":
2021-10-19 20:54:46 +00:00
def _transform_(m):
return nn.parallel.DataParallel(m)
2023-04-11 09:12:58 +00:00
2021-10-19 20:54:46 +00:00
model = model.cuda()
model.multi_gpu_wrapper(_transform_)
elif gpu is not None:
torch.cuda.set_device(gpu)
model = model.cuda(gpu)
else:
2023-04-11 09:12:58 +00:00
raise ValueError("distribution_type = multi | single | None")
2021-10-19 20:54:46 +00:00
if should_diag:
logger.info(opt)
2023-04-11 09:12:58 +00:00
optimizer = optim.Adam(model.parameters(), lr=opt.lr, weight_decay=opt.decay, betas=(opt.beta1, 0.999))
2021-10-19 20:54:46 +00:00
lr_scheduler = optim.lr_scheduler.ExponentialLR(optimizer, opt.lr_gamma)
2023-04-11 09:12:58 +00:00
if opt.model != "":
2021-10-19 20:54:46 +00:00
ckpt = torch.load(opt.model)
2023-04-11 09:12:58 +00:00
model.load_state_dict(ckpt["model_state"])
optimizer.load_state_dict(ckpt["optimizer_state"])
2021-10-19 20:54:46 +00:00
2023-04-11 09:12:58 +00:00
if opt.model != "":
start_epoch = torch.load(opt.model)["epoch"] + 1
2021-10-19 20:54:46 +00:00
else:
start_epoch = 0
def new_x_chain(x, num_chain):
return torch.randn(num_chain, *x.shape[1:], device=x.device)
for epoch in range(start_epoch, opt.niter):
2023-04-11 09:12:58 +00:00
if opt.distribution_type == "multi":
2021-10-19 20:54:46 +00:00
train_sampler.set_epoch(epoch)
lr_scheduler.step(epoch)
for i, data in enumerate(dataloader):
2023-04-11 09:12:58 +00:00
x = data["train_points"].transpose(1, 2)
noises_batch = noises_init[data["idx"]].transpose(1, 2)
2021-10-19 20:54:46 +00:00
2023-04-11 09:12:58 +00:00
"""
2021-10-19 20:54:46 +00:00
train diffusion
2023-04-11 09:12:58 +00:00
"""
2021-10-19 20:54:46 +00:00
2023-04-11 09:12:58 +00:00
if opt.distribution_type == "multi" or (opt.distribution_type is None and gpu is not None):
2021-10-19 20:54:46 +00:00
x = x.cuda(gpu)
noises_batch = noises_batch.cuda(gpu)
2023-04-11 09:12:58 +00:00
elif opt.distribution_type == "single":
2021-10-19 20:54:46 +00:00
x = x.cuda()
noises_batch = noises_batch.cuda()
loss = model.get_loss_iter(x, noises_batch).mean()
optimizer.zero_grad()
loss.backward()
netpNorm, netgradNorm = getGradNorm(model)
if opt.grad_clip is not None:
torch.nn.utils.clip_grad_norm_(model.parameters(), opt.grad_clip)
optimizer.step()
if i % opt.print_freq == 0 and should_diag:
2023-04-11 09:12:58 +00:00
logger.info(
"[{:>3d}/{:>3d}][{:>3d}/{:>3d}] loss: {:>10.4f}, "
"netpNorm: {:>10.2f}, netgradNorm: {:>10.2f} ".format(
epoch,
opt.niter,
i,
len(dataloader),
loss.item(),
netpNorm,
netgradNorm,
)
)
2021-10-19 20:54:46 +00:00
if (epoch + 1) % opt.diagIter == 0 and should_diag:
2023-04-11 09:12:58 +00:00
logger.info("Diagnosis:")
2021-10-19 20:54:46 +00:00
x_range = [x.min().item(), x.max().item()]
kl_stats = model.all_kl(x)
2023-04-11 09:12:58 +00:00
logger.info(
" [{:>3d}/{:>3d}] "
"x_range: [{:>10.4f}, {:>10.4f}], "
"total_bpd_b: {:>10.4f}, "
"terms_bpd: {:>10.4f}, "
"prior_bpd_b: {:>10.4f} "
"mse_bt: {:>10.4f} ".format(
epoch,
opt.niter,
*x_range,
kl_stats["total_bpd_b"].item(),
kl_stats["terms_bpd"].item(),
kl_stats["prior_bpd_b"].item(),
kl_stats["mse_bt"].item(),
)
)
2021-10-19 20:54:46 +00:00
if (epoch + 1) % opt.vizIter == 0 and should_diag:
2023-04-11 09:12:58 +00:00
logger.info("Generation: eval")
2021-10-19 20:54:46 +00:00
model.eval()
with torch.no_grad():
x_gen_eval = model.gen_samples(new_x_chain(x, 25).shape, x.device, clip_denoised=False)
x_gen_list = model.gen_sample_traj(new_x_chain(x, 1).shape, x.device, freq=40, clip_denoised=False)
x_gen_all = torch.cat(x_gen_list, dim=0)
gen_stats = [x_gen_eval.mean(), x_gen_eval.std()]
gen_eval_range = [x_gen_eval.min().item(), x_gen_eval.max().item()]
2023-04-11 09:12:58 +00:00
logger.info(
" [{:>3d}/{:>3d}] "
"eval_gen_range: [{:>10.4f}, {:>10.4f}] "
"eval_gen_stats: [mean={:>10.4f}, std={:>10.4f}] ".format(
epoch,
opt.niter,
*gen_eval_range,
*gen_stats,
)
)
visualize_pointcloud_batch(
"%s/epoch_%03d_samples_eval.png" % (outf_syn, epoch), x_gen_eval.transpose(1, 2), None, None, None
)
2021-10-19 20:54:46 +00:00
2023-04-11 09:12:58 +00:00
visualize_pointcloud_batch(
"%s/epoch_%03d_samples_eval_all.png" % (outf_syn, epoch), x_gen_all.transpose(1, 2), None, None, None
)
2021-10-19 20:54:46 +00:00
2023-04-11 09:12:58 +00:00
visualize_pointcloud_batch("%s/epoch_%03d_x.png" % (outf_syn, epoch), x.transpose(1, 2), None, None, None)
2021-10-19 20:54:46 +00:00
2023-04-11 09:12:58 +00:00
logger.info("Generation: train")
2021-10-19 20:54:46 +00:00
model.train()
if (epoch + 1) % opt.saveIter == 0:
if should_diag:
save_dict = {
2023-04-11 09:12:58 +00:00
"epoch": epoch,
"model_state": model.state_dict(),
"optimizer_state": optimizer.state_dict(),
2021-10-19 20:54:46 +00:00
}
2023-04-11 09:12:58 +00:00
torch.save(save_dict, "%s/epoch_%d.pth" % (output_dir, epoch))
2021-10-19 20:54:46 +00:00
2023-04-11 09:12:58 +00:00
if opt.distribution_type == "multi":
2021-10-19 20:54:46 +00:00
dist.barrier()
2023-04-11 09:12:58 +00:00
map_location = {"cuda:%d" % 0: "cuda:%d" % gpu}
2021-10-19 20:54:46 +00:00
model.load_state_dict(
2023-04-11 09:12:58 +00:00
torch.load("%s/epoch_%d.pth" % (output_dir, epoch), map_location=map_location)["model_state"]
)
2021-10-19 20:54:46 +00:00
dist.destroy_process_group()
2023-04-11 09:12:58 +00:00
2021-10-19 20:54:46 +00:00
def main():
opt = parse_args()
2023-04-11 09:12:58 +00:00
if opt.category == "airplane":
2021-10-19 20:54:46 +00:00
opt.beta_start = 1e-5
opt.beta_end = 0.008
2023-04-11 09:12:58 +00:00
opt.schedule_type = "warm0.1"
2021-10-19 20:54:46 +00:00
exp_id = os.path.splitext(os.path.basename(__file__))[0]
dir_id = os.path.dirname(__file__)
output_dir = get_output_dir(dir_id, exp_id)
copy_source(__file__, output_dir)
2023-04-11 09:12:58 +00:00
""" workaround """
2021-10-19 20:54:46 +00:00
train_dataset, _ = get_dataset(opt.dataroot, opt.npoints, opt.category)
noises_init = torch.randn(len(train_dataset), opt.npoints, opt.nc)
if opt.dist_url == "env://" and opt.world_size == -1:
opt.world_size = int(os.environ["WORLD_SIZE"])
2023-04-11 09:12:58 +00:00
if opt.distribution_type == "multi":
2021-10-19 20:54:46 +00:00
opt.ngpus_per_node = torch.cuda.device_count()
opt.world_size = opt.ngpus_per_node * opt.world_size
mp.spawn(train, nprocs=opt.ngpus_per_node, args=(opt, output_dir, noises_init))
else:
train(opt.gpu, opt, output_dir, noises_init)
def parse_args():
parser = argparse.ArgumentParser()
2023-04-11 09:12:58 +00:00
parser.add_argument("--dataroot", default="ShapeNetCore.v2.PC15k/")
parser.add_argument("--category", default="chair")
parser.add_argument("--bs", type=int, default=16, help="input batch size")
parser.add_argument("--workers", type=int, default=16, help="workers")
parser.add_argument("--niter", type=int, default=10000, help="number of epochs to train for")
parser.add_argument("--nc", default=3)
parser.add_argument("--npoints", default=2048)
"""model"""
parser.add_argument("--beta_start", default=0.0001)
parser.add_argument("--beta_end", default=0.02)
parser.add_argument("--schedule_type", default="linear")
parser.add_argument("--time_num", default=1000)
# params
parser.add_argument("--attention", default=True)
parser.add_argument("--dropout", default=0.1)
parser.add_argument("--embed_dim", type=int, default=64)
parser.add_argument("--loss_type", default="mse")
parser.add_argument("--model_mean_type", default="eps")
parser.add_argument("--model_var_type", default="fixedsmall")
parser.add_argument("--lr", type=float, default=2e-4, help="learning rate for E, default=0.0002")
parser.add_argument("--beta1", type=float, default=0.5, help="beta1 for adam. default=0.5")
parser.add_argument("--decay", type=float, default=0, help="weight decay for EBM")
parser.add_argument("--grad_clip", type=float, default=None, help="weight decay for EBM")
parser.add_argument("--lr_gamma", type=float, default=0.998, help="lr decay for EBM")
parser.add_argument("--model", default="", help="path to model (to continue training)")
"""distributed"""
parser.add_argument("--world_size", default=1, type=int, help="Number of distributed nodes.")
parser.add_argument(
"--dist_url", default="tcp://127.0.0.1:9991", type=str, help="url used to set up distributed training"
)
parser.add_argument("--dist_backend", default="nccl", type=str, help="distributed backend")
parser.add_argument(
"--distribution_type",
default="single",
choices=["multi", "single", None],
help="Use multi-processing distributed training to launch "
"N processes per node, which has N GPUs. This is the "
"fastest way to use PyTorch for either single node or "
"multi node data parallel training",
)
parser.add_argument("--rank", default=0, type=int, help="node rank for distributed training")
parser.add_argument("--gpu", default=None, type=int, help="GPU id to use. None means using all available GPUs.")
"""eval"""
parser.add_argument("--saveIter", default=100, help="unit: epoch")
parser.add_argument("--diagIter", default=50, help="unit: epoch")
parser.add_argument("--vizIter", default=50, help="unit: epoch")
parser.add_argument("--print_freq", default=50, help="unit: iter")
2021-10-19 20:54:46 +00:00
2023-04-11 09:12:58 +00:00
parser.add_argument("--manualSeed", default=42, type=int, help="random seed")
2021-10-19 20:54:46 +00:00
opt = parser.parse_args()
return opt
2023-04-11 09:12:58 +00:00
if __name__ == "__main__":
2021-10-19 20:54:46 +00:00
main()