def loglik_pos(batch_data, list_type, theta, normalization_params):
# Log-normal distribution
output = dict()
epsilon = 1e-3
# Data outputs
data_mean_log, data_var_log = normalization_params
data_var_log = torch.clamp(data_var_log, epsilon, np.inf)
data, missing_mask = batch_data
data_log = torch.log(1.0 + data)
missing_mask = missing_mask.float()
est_mean, est_var = theta
est_var = torch.clamp(torch.nn.Softplus()(est_var), epsilon, 1.0)
# Affine transformation of the parameters
est_mean = torch.sqrt(data_var_log) * est_mean + data_mean_log
est_var = data_var_log * est_var
# Compute loglik
log_p_x = -0.5 * torch.sum(torch.pow(data_log - est_mean, 2) / est_var, 1) \
- 0.5 * torch.sum(torch.log(2 * np.pi * est_var), 1) - torch.sum(data_log, 1)
log_normal = td.LogNormal(est_mean, torch.sqrt(est_var))
# log_p_x = log_normal.log_prob(data).sum(1) ##
output['log_p_x'] = torch.mul(log_p_x, missing_mask)
output['log_p_x_missing'] = torch.mul(log_p_x, 1.0 - missing_mask)
output['params'] = [est_mean, est_var]
output['samples'] = log_normal.rsample() - 1.0 # -1.0 TODO???
# output['samples'] = torch.clamp(
# torch.exp(td.Normal(est_mean, torch.sqrt(est_var)).rsample()) - 1.0, 0, 1e20)
return output
def gen_data(self):
# sample overall relative abundances of ASVs from a Dirichlet distribution
self.ASV_rel_abundance = tdist.Dirichlet(torch.ones(
self.numASVs)).sample()
# sample spatial embedding of ASVs
self.w = torch.zeros(self.numASVs, self.D)
w_prior = tdist.MultivariateNormal(torch.zeros(self.D),
torch.eye(self.D))
for o in range(0, self.numASVs):
self.w[o, :] = w_prior.sample()
self.data = torch.zeros(self.numParticles, self.numASVs)
num_nonempty = 0
mu_prior = tdist.MultivariateNormal(torch.zeros(self.D),
torch.eye(self.D))
rad_prior = tdist.LogNormal(torch.tensor([self.mu_rad]),
torch.tensor([self.mu_std]))
# replace with neg bin prior
num_reads_prior = tdist.Poisson(
torch.tensor([self.avgNumReadsParticle]))
while (num_nonempty < self.numParticles):
# sample center
mu = mu_prior.sample()
rad = rad_prior.sample()
zr = torch.zeros(1, self.numASVs, dtype=torch.float64)
for o in range(0, self.numASVs):
p = mu - self.w[o, :]
p = torch.pow(p, 2.0) / rad
p = (torch.sum(p)).sqrt()
zr[0, o] = unitboxcar(p, 0.0, 2.0, self.step_approx)
if torch.sum(zr) > 0.95:
particle = Particle(mu, self)
particle.zr = zr
self.particles.append(particle)
# renormalize particle abundances
rn = self.ASV_rel_abundance * zr
rn = rn / torch.sum(rn)
# sample relative abundances for particle
part_rel_abundance = tdist.Dirichlet(rn * self.conc).sample()
# sample number of reads for particle
# (replace w/ neg bin instead of Poisson)
num_reads = num_reads_prior.sample().long().item()
particle.total_reads = num_reads
particle.reads = tdist.Multinomial(
num_reads, probs=part_rel_abundance).sample()
num_nonempty += 1
def decode_x(self, w, z):
params = self.decoder_x(torch.cat((w, z), dim=-1))
px_wz = []
samples = []
for indices in self.likelihood_partition:
data_type = self.likelihood_partition[indices]
params_subset = params[:, indices[0]:(indices[1] + 1)]
if data_type == 'real':
cov_diag = self.likelihood_params['lik_var'] * torch.ones_like(
params_subset).to(self.device)
dist = D.Normal(loc=params_subset, scale=cov_diag.sqrt())
elif data_type == 'categorical':
dist = D.OneHotCategorical(logits=params_subset)
elif data_type == 'binary':
dist = D.Bernoulli(logits=params_subset)
elif data_type == 'positive':
lognormal_var = self.likelihood_params[
'lik_var_lognormal'] * torch.ones_like(params_subset).to(
self.device)
dist = D.LogNormal(loc=params_subset,
scale=lognormal_var.sqrt())
elif data_type == 'count':
positive_params_subset = F.softplus(params_subset)
dist = D.Poisson(rate=positive_params_subset)
elif data_type == 'binomial':
num_trials = self.likelihood_params['binomial_num_trials']
dist = D.Binomial(total_count=num_trials, logits=params_subset)
elif data_type == 'ordinal':
h = params_subset[:, 0:1]
thetas = torch.cumsum(F.softplus(params_subset[:, 1:]), axis=1)
prob_lessthans = torch.sigmoid(thetas - h)
probs = torch.cat((prob_lessthans, torch.ones(len(prob_lessthans), 1)), axis=1) - \
torch.cat((torch.zeros(len(prob_lessthans), 1), prob_lessthans), axis=1)
dist = D.OneHotCategorical(probs=probs)
else:
raise NotImplementedError
samples.append(dist.sample())
px_wz.append(dist)
sample_x = torch.cat(samples, axis=1)
return params, sample_x, px_wz
def dist(
self,
batch: dict[str, Union[torch.Tensor, list[torch.Tensor]]],
) -> distributions.Distribution:
"""Возвращает распределение доходности."""
logits, mean, std = self(batch)
try:
weights_dist = distributions.Categorical(logits=logits)
except ValueError:
raise GradientsError(
f"Ошибка при обновлении градиентов: NaN in Categorical distribution"
)
comp_dist = distributions.LogNormal(mean, std)
return distributions.MixtureSameFamily(weights_dist, comp_dist)
# sample embedding of OTUs
w = torch.zeros(O, D)
w_prior = tdist.MultivariateNormal(torch.zeros(D), torch.eye(D))
for o in range(0, O):
w[o, :] = w_prior.sample()
# hyperparameters for particle radius
#eta_1 = 1.0
#eta_2 = 2.0
#rad_prior = tdist.Gamma(torch.tensor([eta_1]), torch.tensor([eta_2]))
#rad = rad_prior.sample()*4.0
#rad = rad * rad
mu_rad = numpy.log(1.0)
mu_std = 1.0
rad_prior = tdist.LogNormal(torch.tensor([mu_rad]), torch.tensor([mu_std]))
#rad = rad_prior.sample()
rad = torch.tensor([1.0])
# sample particle center
mu_prior = tdist.MultivariateNormal(torch.zeros(D), torch.eye(D))
mu = mu_prior.sample()
print('mu=', mu)
print('rad=', rad)
# sample indicator as to whether OTUs occur in the particle
#z = torch.zeros(O,1,dtype=torch.int64)
zr = torch.zeros(O, 1, dtype=torch.float64)
# annealing parameter for unit step approximation
def lognormal_log_pdf(x, mu, logvar):
scale = 0.5 * torch.exp(logvar)
p_dist = dist.LogNormal(mu, scale)
logprob = p_dist.log_prob(x)
return torch.sum(logprob, dim=1)
from matplotlib import pyplot as plt
import torch
import torch.distributions as ds
# Sample synthetic data from a heavitailed distribution for the sizes
# Sample random ids from 1 to 1000 for the packet senders
N_SAMPLES = 100000000
N_IDS = 10000
FNAME = 'trace/generated_data_1e5_1e9'
size_dist = ds.LogNormal(torch.tensor([0.0]), torch.tensor([2.0]))
id_dist = ds.Uniform(torch.tensor([0]).float(), torch.tensor([N_IDS]).float())
sizes = []
ids = []
for index in range(N_SAMPLES):
if index % 1000000 == 0:
print(index)
packet_size = int(size_dist.sample().long().numpy()[0] + 1)
packet_id = int(id_dist.sample().long().numpy()[0])
sizes.append(packet_size)
ids.append(packet_id)
arr = [str(a) + " " + str(b) + "\n" for a, b in zip(sizes, ids)]
with open(FNAME, 'w') as f:
f.write("".join(arr))
""" from math import comb import matplotlib.pyplot as plt import numpy as np import pandas as pd import torch import torch.nn as nn import torch.nn.functional as F import torch.distributions as D import torch.optim as optim from torch.utils.data import Dataset, DataLoader from tqdm import trange m = D.LogNormal(torch.tensor([0.0]), torch.tensor([1.0])) m.log_prob(torch.tensor([2.])) # GLOBAL SEED = 42 VERBOSITY = 0 # VAE NUM_EPOCHS_TUNE = 200 NUM_EPOCHS = 500 HIDDEN_DIM = 64 Z_PRIOR_VAR = 0.5**2 X_POST_VAR = 0.1**2
def sample_logp(self, N, I, noise=None):
max_tau = self.tau_count
burnin = self.burnin
logp0 = 0.0
T = len(I)
P = self.P.exp()
N_0 = self.N_0.exp()
s_d = self.s_d.exp()
s_p = self.s_p.exp()
delta = self.delta.exp()
p_tau = td.Categorical(logits=self.tau_logits)
tau = p_tau.sample([N])
logp0 = logp0 + p_tau.log_prob(tau)
tau = tau + 1
#tau = self.tau_logits.softmax(0)
p_et = td.Gamma(1 / s_p**2, 1 / s_p**2)
et = p_et.sample([N, T+burnin])
logp0 = logp0 + p_et.log_prob(et).sum(1)
p_epsilont = td.Gamma(1 / s_d**2, 1 / s_d**2)
epsilont = p_epsilont.sample([N, T+burnin])
logp0 = logp0 + p_epsilont.log_prob(epsilont).sum(-1)
#z = self.z_0.exp().expand(N)
zs = torch.zeros(N, max_tau+burnin+T, dtype=P.dtype, device=P.device)
zs[torch.arange(N),:max_tau] = self.z_0.sigmoid()
z = zs[torch.arange(N),max_tau-1]
for t in torch.arange(max_tau, T+burnin+max_tau):
ztmtau = zs[torch.arange(N), t-tau]
# print("----")
# print(zs[:,:t+1])
# print(tau_t)
# print(ztmtau)
# tstart = max(0, t+1-len(tau))
# ztmtau = zs[torch.arange(N), tstart:t+1] @ tau[:min(t,len(tau))+1]
z = P * ztmtau * (-ztmtau / N_0).exp() * et[:,t-max_tau] + z * (-delta * epsilont[:,t-max_tau]).exp()
pz = td.LogNormal(z.log(), self.rand_std.exp())
z = pz.sample([])
logp0 = logp0 + pz.log_prob(z)
zs[:,t] = z
xs = zs[:, -T:].detach()
'''
if noise is None:
noise_dist = td.Normal(xs, self.noise_std.exp())
else:
noise_dist = td.Normal(xs, max(noise, 1e-6))
xs = noise_dist.sample([])
xs = xs.reshape(N, -1)
nat = Likelihoods.gaussian_nat(noise_dist)
norm = Likelihoods.gaussian_norm(noise_dist)
return (tau, et, epsilont, zs), xs, norm - logp0, nat
'''
return (tau, et, epsilont, zs), xs, -logp0, torch.zeros(N,T*2,dtype=xs.dtype,device=xs.device)
def entropy(self):
return dists.LogNormal(self.loc, self.scale).entropy()
def sample(self, batch_size):
return dists.LogNormal(self.loc, self.scale).rsample((batch_size, ))
def log_prob(self, value):
return dists.LogNormal(self.loc, self.scale).log_prob(value).sum(-1)
