File size: 10,727 Bytes
1f0c7b9 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 |
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch_geometric
from torchdiffeq import odeint
from torch_scatter import scatter_mean, scatter_sum
import torch.distributions as dist
import numpy as np
LOG2PI = np.log(2 * np.pi)
class BFNBase(nn.Module):
# this is a general method which could be used for implement vector field in CNF or
def __init__(self, *args, **kwargs):
super(BFNBase, self).__init__(*args, **kwargs)
# def zero_center_of_mass(self, x_pos, segment_ids):
# size = x_pos.size()
# assert len(size) == 2 # TODO check this
# seg_means = scatter_mean(x_pos, segment_ids, dim=0)
# mean_for_each_segment = seg_means.index_select(0, segment_ids)
# x = x_pos - mean_for_each_segment
# return x
def get_k_params(self, bins):
"""
function to get the k parameters for the discretised variable
"""
# k = torch.ones_like(mu)
# ones_ = torch.ones((mu.size()[1:])).cuda()
# ones_ = ones_.unsqueeze(0)
list_c = []
list_l = []
list_r = []
for k in range(1, int(bins + 1)):
# k = torch.cat([k,torch.ones_like(mu)*(i+1)],dim=1
k_c = (2 * k - 1) / bins - 1
k_l = k_c - 1 / bins
k_r = k_c + 1 / bins
list_c.append(k_c)
list_l.append(k_l)
list_r.append(k_r)
# k_c = torch.cat(list_c,dim=0)
# k_l = torch.cat(list_l,dim=0)
# k_r = torch.cat(list_r,dim=0)
return list_c, list_l, list_l
def discretised_cdf(self, mu, sigma, x):
"""
cdf function for the discretised variable
"""
# in this case we use the discretised cdf for the discretised output function
mu = mu.unsqueeze(1)
sigma = sigma.unsqueeze(1) # B,1,D
f_ = 0.5 * (1 + torch.erf((x - mu) / (sigma * np.sqrt(2))))
flag_upper = torch.ge(x, 1)
flag_lower = torch.le(x, -1)
f_ = torch.where(flag_upper, torch.ones_like(f_), f_)
f_ = torch.where(flag_lower, torch.zeros_like(f_), f_)
return f_
def continuous_var_bayesian_update(self, t, sigma1, x):
"""
x: [N, D]
"""
"""
TODO: rename this function to bayesian flow
"""
# Eq.(77): p_F(θ|x;t) ~ N (μ | γ(t)x, γ(t)(1 − γ(t))I)
gamma = 1 - torch.pow(sigma1, 2 * t) # [B]
mu = gamma * x + torch.randn_like(x) * torch.sqrt(gamma * (1 - gamma))
return mu, gamma
def discrete_var_bayesian_update(self, t, beta1, x, K):
"""
x: [N, K]
"""
# Eq.(182): β(t) = t**2 β(1)
beta = beta1 * (t**2) # (B,)
# Eq.(185): p_F(θ|x;t) = E_{N(y | β(t)(Ke_x−1), β(t)KI)} δ (θ − softmax(y))
# can be sampled by first drawing y ~ N(y | β(t)(Ke_x−1), β(t)KI)
# then setting θ = softmax(y)
one_hot_x = x # (N, K)
mean = beta * (K * one_hot_x - 1)
std = (beta * K).sqrt()
eps = torch.randn_like(mean)
y = mean + std * eps
theta = F.softmax(y, dim=-1)
return theta
def discreteised_var_bayesian_update(self, t, sigma1, x):
"""
x: [N, D]
Note, this is identical to the continuous_var_bayesian_update
"""
gamma = 1 - torch.pow(sigma1, 2 * t)
mu = gamma * x + torch.randn_like(x) * torch.sqrt(gamma * (1 - gamma))
return mu, gamma
def ctime4continuous_loss(self, t, sigma1, x_pred, x, segment_ids=None):
# Eq.(101): L∞(x) = −ln(σ1) * E_{t∼U (0,1), p_F(θ|x;t)} [|x − x_hat(θ,t)|**2 / (σ_1**2)**t]
if segment_ids is not None:
loss = scatter_mean(
torch.pow(sigma1, -2 * t.view(-1))
* ((x_pred - x).view(x.shape[0], -1).abs().pow(2).sum(dim=1)),
segment_ids,
dim=0,
)
else:
loss = torch.pow(sigma1, -2 * t.view(-1)) * (x_pred - x).view(
x.shape[0], -1
).abs().pow(2).sum(dim=1)
return -torch.log(sigma1) * loss
def dtime4continuous_loss(self, i, N, sigma1, x_pred, x, segment_ids=None):
# TODO not debuged yet
weight = N * (1 - torch.pow(sigma1, 2 / N)) / (2 * torch.pow(sigma1, 2 * i / N))
# print(x_pred.shape, x.shape , i.shape,weight.shape)
# print(segment_ids)
if segment_ids is not None:
loss = scatter_mean(
weight.view(-1) * ((x_pred - x) ** 2).sum(-1), segment_ids, dim=0
)
else:
loss = (
N
* (1 - torch.pow(sigma1, 2 / N))
/ (2 * torch.pow(sigma1, 2 * i / N))
* (x_pred - x).view(x.shape[0], -1).abs().pow(2).sum(dim=1)
)
# print(loss.shape)
return loss
def ctime4discrete_loss(self, t, beta1, one_hot_x, p_0, K, segment_ids=None):
# Eq.(205): L∞(x) = Kβ(1) E_{t∼U (0,1), p_F (θ|x,t)} [t|e_x − e_hat(θ, t)|**2,
# where e_hat(θ, t) = (\sum_k p_O^(1) (k | θ; t)e_k, ..., \sum_k p_O^(D) (k | θ; t)e_k)
e_x = one_hot_x # [N, K]
e_hat = p_0 # (N, K)
assert e_x.size() == e_hat.size()
if segment_ids is not None:
L_infinity = scatter_mean(
K * beta1 * t.view(-1) * ((e_x - e_hat) ** 2).sum(dim=-1),
segment_ids,
dim=0,
)
else:
L_infinity = K * beta1 * t.view(-1) * ((e_x - e_hat) ** 2).sum(dim=-1)
return L_infinity
def dtime4discrete_loss_prob(
self, i, N, beta1, one_hot_x, p_0, K, n_samples=200, segment_ids=None
):
# this is based on the official implementation of BFN.
# import pdb
# pdb.set_trace()
target_x = one_hot_x # [D, K]
e_hat = p_0 # (D, K)
alpha = beta1 * (2 * i - 1) / N**2 # [D]
alpha = alpha.view(-1, 1) # [D, 1]
classes = torch.arange(K, device=target_x.device).long().unsqueeze(0) # [ 1, K]
e_x = F.one_hot(classes.long(), K) #[1,K, K]
# print(e_x.shape)
receiver_components = dist.Independent(
dist.Normal(
alpha.unsqueeze(-1) * ((K * e_x) - 1), # [D K, K]
(K * alpha.unsqueeze(-1)) ** 0.5, # [D, 1, 1]
),
1,
) # [D,T, K, K]
receiver_mix_distribution = dist.Categorical(probs=e_hat) # [D, K]
receiver_dist = dist.MixtureSameFamily(
receiver_mix_distribution, receiver_components
) # [D, K]
sender_dist = dist.Independent( dist.Normal(
alpha* ((K * target_x) - 1), ((K * alpha) ** 0.5)
),1) # [D, K]
y = sender_dist.sample(torch.Size([n_samples]))
loss = N * (sender_dist.log_prob(y) - receiver_dist.log_prob(y)).mean(0).mean(
-1, keepdims=True
)
# loss = (
# (sender_dist.log_prob(y) - receiver_dist.log_prob(y))
# .mean(0)
# .flatten(start_dim=1)
# .mean(1, keepdims=True)
# )
# #
return loss.mean()
def dtime4discrete_loss(self, i, N, beta1, one_hot_x, p_0, K, segment_ids=None):
# i in {1,n}
# Algorithm 7 in BFN
e_x = one_hot_x # [D, K]
e_hat = p_0 # (D, K)
assert e_x.size() == e_hat.size()
alpha = beta1 * (2 * i - 1) / N**2 # [D]
# print(alpha.shape)
mean_ = alpha * (K * e_x - 1) # [D, K]
std_ = torch.sqrt(alpha * K) # [D,1] TODO check shape
eps = torch.randn_like(mean_) # [D,K,]
y_ = mean_ + std_ * eps
# modify this line:
matrix_ek = torch.eye(K, K).unsqueeze(0).to(e_x.device)
matrix_ek.repeat(alpha.size(0), 1, 1) # [D,K,K]
mean_matrix = alpha.unsqueeze(-1) * (K * matrix_ek - 1) # [D,K,K]
std_matrix = torch.sqrt(alpha * K).unsqueeze(-1) #
likelihood = (
torch.exp(
-((y_.unsqueeze(1).repeat(1, K, 1) - mean_matrix) ** 2)
/ (2 * std_matrix**2)
)
/ (std_matrix * np.sqrt(2 * np.pi))
).prod(
-1
) # [D,K]
if segment_ids is not None:
L_N = -scatter_mean(
torch.log((likelihood * e_hat).sum(dim=-1)), segment_ids, dim=0
)
else:
L_N = -torch.log((likelihood * e_hat).sum(dim=-1)) # [D]
# print(L_N.shape)
#
return N * L_N
def dtime4discrete_loss_gjj(self, i, N, beta1, one_hot_x, p_0, K, segment_ids=None):
# i in {1,n}
# Algorithm 7 in BFN
e_x = one_hot_x # [D, K]
e_hat = p_0 # (D, K)
assert e_x.size() == e_hat.size()
alpha = beta1 * (2 * i - 1) / N**2 # [D]
# print(alpha.shape)
mean_ = alpha * (K * e_x - 1) # [D, K]
std_ = torch.sqrt(alpha * K) # [D,1] TODO check shape
eps = torch.randn_like(mean_) # [D,K,]
y_ = mean_ + std_ * eps # [D, K]
# modify this line:
matrix_ek = torch.eye(K, K).to(e_x.device) # [K, K]
mean_matrix = K * matrix_ek - 1 # [K, K]
std_matrix = torch.sqrt(alpha * K).unsqueeze(-1) #
_log_gaussians = ( # [D, K]
(-0.5 * LOG2PI - torch.log(std_matrix))
- (y_.unsqueeze(1) - mean_matrix) ** 2 / (2 * std_matrix**2)
).sum(-1)
_inner_log_likelihood = torch.log(
torch.sum(e_hat * torch.exp(_log_gaussians), dim=-1)
) # (D,)
_inner_log_likelihood = torch.log(e_hat) + _log_gaussians # [D, K]
log_likelihood = torch.logsumexp(_inner_log_likelihood, dim=-1) # [D]
if segment_ids is not None:
L_N = -scatter_mean(log_likelihood, segment_ids, dim=0)
else:
L_N = -log_likelihood.sum(dim=-1) # [D]
# print(L_N.shape)
#
return N * L_N
def ctime4discreteised_loss(self, t, sigma1, x_pred, x, segment_ids=None):
if segment_ids is not None:
loss = scatter_sum(
(x_pred - x).view(x.shape[0], -1).abs().pow(2), segment_ids, dim=0
)
else:
raise NotImplementedError
loss = (x_pred - x).view(x.shape[0], -1).abs().pow(2).sum(dim=1)
return -torch.log(sigma1) * loss * torch.pow(sigma1, -2 * t.view(-1))
def interdependency_modeling(self):
raise NotImplementedError
def forward(self):
raise NotImplementedError
def loss_one_step(self):
raise NotImplementedError
def sample(self):
raise NotImplementedError
|