File size: 11,834 Bytes
15d238f 5c89480 15d238f 75bf717 15d238f |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 |
import torch.nn as nn
import torch
from torch.optim.lr_scheduler import ReduceLROnPlateau,OneCycleLR,CyclicLR
import pandas as pd
from sklearn.preprocessing import StandardScaler,MinMaxScaler
import matplotlib.pyplot as plt
from torch.distributions import MultivariateNormal, LogNormal,Normal, Chi2
from torch.distributions.distribution import Distribution
from sklearn.metrics import r2_score
import numpy as np
# It's a distribution that is a kernel density estimate of a Gaussian distribution
class GaussianKDE(Distribution):
def __init__(self, X, bw):
"""
X : tensor (n, d)
`n` points with `d` dimensions to which KDE will be fit
bw : numeric
bandwidth for Gaussian kernel
"""
self.X = X
self.bw = bw
self.dims = X.shape[-1]
self.n = X.shape[0]
self.mvn = MultivariateNormal(loc=torch.zeros(self.dims),
scale_tril=torch.eye(self.dims))
def sample(self, num_samples):
"""
We are sampling from a normal distribution with mean equal to the data points in the dataset and
standard deviation equal to the bandwidth
:param num_samples: the number of samples to draw from the KDE
:return: A sample of size num_samples from the KDE.
"""
idxs = (np.random.uniform(0, 1, num_samples) * self.n).astype(int)
norm = Normal(loc=self.X[idxs], scale=self.bw)
return norm.sample()
def score_samples(self, Y, X=None):
"""Returns the kernel density estimates of each point in `Y`.
Parameters
----------
Y : tensor (m, d)
`m` points with `d` dimensions for which the probability density will
be calculated
X : tensor (n, d), optional
`n` points with `d` dimensions to which KDE will be fit. Provided to
allow batch calculations in `log_prob`. By default, `X` is None and
all points used to initialize KernelDensityEstimator are included.
Returns
-------
log_probs : tensor (m)
log probability densities for each of the queried points in `Y`
"""
if X == None:
X = self.X
log_probs = self.mvn.log_prob((X.unsqueeze(1) - Y)).sum(dim=0)
return log_probs
def log_prob(self, Y):
"""Returns the total log probability of one or more points, `Y`, using
a Multivariate Normal kernel fit to `X` and scaled using `bw`.
Parameters
----------
Y : tensor (m, d)
`m` points with `d` dimensions for which the probability density will
be calculated
Returns
-------
log_prob : numeric
total log probability density for the queried points, `Y`
"""
X_chunks = self.X
Y_chunks = Y
self.Y = Y
log_prob = 0
for x in X_chunks:
for y in Y_chunks:
log_prob += self.score_samples(y,x).sum(dim=0)
return log_prob
class Chi2KDE(Distribution):
def __init__(self, X, bw):
"""
X : tensor (n, d)
`n` points with `d` dimensions to which KDE will be fit
bw : numeric
bandwidth for Gaussian kernel
"""
self.X = X
self.bw = bw
self.dims = X.shape[-1]
self.n = X.shape[0]
self.mvn = Chi2(self.dims)
def sample(self, num_samples):
idxs = (np.random.uniform(0, 1, num_samples) * self.n).astype(int)
norm = LogNormal(loc=self.X[idxs], scale=self.bw)
return norm.sample()
def score_samples(self, Y, X=None):
"""Returns the kernel density estimates of each point in `Y`.
Parameters
----------
Y : tensor (m, d)
`m` points with `d` dimensions for which the probability density will
be calculated
X : tensor (n, d), optional
`n` points with `d` dimensions to which KDE will be fit. Provided to
allow batch calculations in `log_prob`. By default, `X` is None and
all points used to initialize KernelDensityEstimator are included.
Returns
-------
log_probs : tensor (m)
log probability densities for each of the queried points in `Y`
"""
if X == None:
X = self.X
log_probs = self.mvn.log_prob(abs(X.unsqueeze(1) - Y)).sum()
return log_probs
def log_prob(self, Y):
"""Returns the total log probability of one or more points, `Y`, using
a Multivariate Normal kernel fit to `X` and scaled using `bw`.
Parameters
----------
Y : tensor (m, d)
`m` points with `d` dimensions for which the probability density will
be calculated
Returns
-------
log_prob : numeric
total log probability density for the queried points, `Y`
"""
X_chunks = self.X
Y_chunks = Y
self.Y = Y
log_prob = 0
for x in X_chunks:
for y in Y_chunks:
log_prob += self.score_samples(y,x).sum(dim=0)
return log_prob
class PlanarFlow(nn.Module):
"""
A single planar flow, computes T(x) and log(det(jac_T)))
"""
def __init__(self, D):
super(PlanarFlow, self).__init__()
self.u = nn.Parameter(torch.Tensor(1, D), requires_grad=True)
self.w = nn.Parameter(torch.Tensor(1, D), requires_grad=True)
self.b = nn.Parameter(torch.Tensor(1), requires_grad=True)
self.h = torch.tanh
self.init_params()
def init_params(self):
self.w.data.uniform_(0.4, 1)
self.b.data.uniform_(0.4, 1)
self.u.data.uniform_(0.4, 1)
def forward(self, z):
linear_term = torch.mm(z, self.w.T) + self.b
return z + self.u * self.h(linear_term)
def h_prime(self, x):
"""
Derivative of tanh
"""
return (1 - self.h(x) ** 2)
def psi(self, z):
inner = torch.mm(z, self.w.T) + self.b
return self.h_prime(inner) * self.w
def log_det(self, z):
inner = 1 + torch.mm(self.psi(z), self.u.T)
return torch.log(torch.abs(inner))
# It's a normalizing flow that takes in a distribution and outputs a distribution.
class NormalizingFlow(nn.Module):
"""
A normalizng flow composed of a sequence of planar flows.
"""
def __init__(self, D, n_flows=2):
"""
The function takes in two arguments, D and n_flows. D is the dimension of the data, and n_flows
is the number of flows. The function then creates a list of PlanarFlow objects, where the number
of PlanarFlow objects is equal to n_flows
:param D: the dimensionality of the data
:param n_flows: number of flows to use, defaults to 2 (optional)
"""
super(NormalizingFlow, self).__init__()
self.flows = nn.ModuleList(
[PlanarFlow(D) for _ in range(n_flows)])
def sample(self, base_samples):
"""
Transform samples from a simple base distribution
by passing them through a sequence of Planar flows.
"""
samples = base_samples
for flow in self.flows:
samples = flow(samples)
return samples
def forward(self, x):
"""
Computes and returns the sum of log_det_jacobians
and the transformed samples T(x).
"""
sum_log_det = 0
transformed_sample = x
for i in range(len(self.flows)):
log_det_i = (self.flows[i].log_det(transformed_sample))
sum_log_det += log_det_i
transformed_sample = self.flows[i](transformed_sample)
return transformed_sample, sum_log_det
def random_normal_samples(n, dim=2):
return torch.zeros(n, dim).normal_(mean=0, std=1.5)
class nflow():
def __init__(self,dim=2,latent=16,batchsize:int=1,dataset=None):
"""
The function __init__ initializes the class NormalizingFlowModel with the parameters dim,
latent, batchsize, and datasetPath
:param dim: The dimension of the data, defaults to 2 (optional)
:param latent: The number of latent variables in the model, defaults to 16 (optional)
:param batchsize: The number of samples to generate at a time, defaults to 1
:type batchsize: int (optional)
:param datasetPath: The path to the dataset, defaults to data/dataset.csv
:type datasetPath: str (optional)
"""
self.dim = dim
self.batchsize = batchsize
self.model = NormalizingFlow(dim, latent)
self.dataset = dataset
def compile(self,optim:torch.optim=torch.optim.Adam,distribution:str='GaussianKDE',lr:float=0.00015,bw:float=0.1,wd=0.0015):
"""
It takes in a dataset, a model, and a distribution, and returns a compiled model
:param optim: the optimizer to use
:type optim: torch.optim
:param distribution: the type of distribution to use, defaults to GaussianKDE
:type distribution: str (optional)
:param lr: learning rate
:type lr: float
:param bw: bandwidth for the KDE
:type bw: float
"""
if wd:
self.opt = optim(
params=self.model.parameters(),
lr=lr,
weight_decay = wd
# momentum=0.9
# momentum=0.1
)
else:
self.opt = optim(
params=self.model.parameters(),
lr=lr,
# momentum=0.9
# momentum=0.1
)
self.scaler = StandardScaler()
self.scaler_mm = MinMaxScaler(feature_range=(0,1))
df = pd.read_csv(self.dataset)
df = df.iloc[:,1:]
if 'Chi2' in distribution:
self.scaled=self.scaler_mm.fit_transform(df)
else: self.scaled = self.scaler.fit_transform(df)
self.density = globals()[distribution](X=torch.tensor(self.scaled, dtype=torch.float32), bw=bw)
# self.dl = torch.utils.data.DataLoader(scaled,batchsize=self.batchsize)
self.scheduler = ReduceLROnPlateau(self.opt, patience=10000)
self.losses = []
def train(self,iters:int=1000):
"""
> We sample from a normal distribution, pass the samples through the model, and then calculate
the loss
:param iters: number of iterations to train for, defaults to 1000
:type iters: int (optional)
"""
for idx in range(iters):
if idx % 100 == 0:
print("Iteration {}".format(idx))
samples = torch.autograd.Variable(random_normal_samples(self.batchsize,self.dim))
z_k, sum_log_det = self.model(samples)
log_p_x = self.density.log_prob(z_k)
# Reverse KL since we can evaluate target density but can't sample
loss = (-sum_log_det - (log_p_x)).mean()/self.density.n
self.opt.zero_grad()
loss.backward()
self.opt.step()
self.scheduler.step(loss)
self.losses.append(loss.item())
if idx % 100 == 0:
print("Loss {}".format(loss.item()))
yield idx,loss.item()
def performance(self):
"""
The function takes the model and the scaled data as inputs, samples from the model, and then
prints the r2 score of the samples and the scaled data.
"""
samples = ((self.model.sample(torch.tensor(self.scaled).float())).detach().numpy())
print('r2', r2_score(self.scaled,samples))
|