forked from msabvid/MLMC-MIMC-SGD
-
Notifications
You must be signed in to change notification settings - Fork 0
/
logistic_sgld_cv.py
312 lines (262 loc) · 12.2 KB
/
logistic_sgld_cv.py
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
import sys
import os
from abc import ABC, abstractmethod
import math
import torch
import torch.nn as nn
from torch.utils.data import Dataset, DataLoader
import copy
import argparse
import numpy as np
import math
from sklearn.datasets import make_blobs
import tqdm
import matplotlib.pyplot as plt
import pickle
from lib.data import get_dataset, Dataset_MCMC
from lib.priors import Gaussian, MixtureGaussians
from lib.models import LogisticNets
from lib.hyperparameters import config_priors
class SGLD():
def __init__(self, T, n_steps, data_X, data_Y, prior, device, MAP = None):
self.data_X = data_X
self.data_Y = data_Y
self.dim = data_X.shape[1]-1 # -1 is to subtract column of 1s in the data to account for the intercept
self.T = T
self.n_steps = n_steps
self.device = device
self.data_size = data_X.shape[0]
self.prior = prior
# we estimate the MAP, and the log-likelihood of the dataset using the MAP
if MAP:
self.MAP = LogisticNets(self.dim, N=1)#.to(device=self.device)
self.MAP.load_state_dict(MAP)
self.MAP.to(device=self.device)
else:
self.MAP = self.estimate_MAP(epochs=args.n_steps,batch_size=self.data_size) # object of class LogisticNets
self.grad_loglik_MAP = self.get_grad_loglik_MAP() # tensor of size (1, self.dim+1)
def estimate_MAP(self, epochs, batch_size):
"""
We estimate the MAP using usual SGD with RMSprop
"""
hf = 1e-5#min(self.T/self.n_steps, 0.0001)
epochs = max(epochs, 150000)
print("estimating MAP for control variate")
pbar = tqdm.tqdm(total=epochs)
X = LogisticNets(self.dim, N=1).to(device=self.device)
self.init_weights(X)
sf=self.data_size
for step in range(epochs):
#U = np.random.choice(self.data_size, (1, sf), replace=True)
U = np.arange(self.data_size).reshape(1,batch_size)
X.zero_grad()
drift = self.prior.logprob(X.params) + self.data_size/batch_size * X.loglik(self.data_X, self.data_Y, U)
drift.backward(torch.ones_like(drift))
X.params.data.copy_(X.params.data + hf*(X.params.grad))
pbar.update(1)
if step % 100 == 0:
pbar.write("norm grad log prob={}".format(torch.norm(X.params.grad, p="fro").item()))
if step>100000:
hf = 1e-6
return X
def save_MAP(self, filename):
torch.save(self.MAP.state_dict(), filename)
def get_grad_loglik_MAP(self):
self.MAP.zero_grad()
loss_fn = nn.BCELoss(reduction='sum')
pred = self.MAP(self.data_X.to(self.device))
loss = -loss_fn(pred, self.data_Y.to(self.device)) #!!! I put a minus in front of loss_fn so that we actually compute the log-likelihood! Important for the signs in the Langevin process
loss.backward()
grad_loglik_MAP = copy.deepcopy(self.MAP.params.grad)
return grad_loglik_MAP
def init_weights(self, net):
""" Init weights with prior
"""
#net.params.data.copy_(mu + std * torch.randn_like(net.params))
#net.params.data.copy_(torch.zeros_like(net.params))
N = net.params.shape[0]
try:
net.params.data.copy_(self.MAP.params.data.repeat((N,1)))
except:
net.params.data.copy_(torch.zeros_like(net.params))
def Func(self, nets):
"""Function of X.
Recall we want to approximate E(F(X)) where X is a random vector
Parameters
----------
nets : LogisticNets
"""
with torch.no_grad():
F = torch.norm(nets.params, p=2, dim=1)**2
return F.cpu().numpy()
def solve(self, N, sf):
"""Solve N stochastic langevin processes with subsample size sf
It returns E(F(X)) with X sampled from the posterior
Parameters
----------
N: int
Number of Langevin Processes to solve
sf: int
Subsample size
"""
hf = 5e-6#0.0001#0.1/self.data_size#1e-8#1/self.data_size#self.T/self.n_steps
#sigma = 1/math.sqrt(self.data_size)
sigma = math.sqrt(2)
sum1, sum2 = np.zeros(self.n_steps), np.zeros(self.n_steps) #first and second order moments
var = np.zeros_like(sum1)
print("Solving SGLD...")
n_steps = min(100000, self.n_steps)
for N1 in range(0,N,5000):
N2 = min(5000, N-N1) # we will do batches of paths of size N2
X = LogisticNets(self.dim, N2).to(device=self.device)
self.init_weights(X)
# Euler scheme on Stochastic Langevin process
pbar = tqdm.tqdm(total=n_steps)
for step in range(n_steps):
dW = math.sqrt(hf) * torch.randn_like(X.params)
U = np.random.choice(self.data_size, (N2, sf), replace=True)
#U = np.array([np.random.choice(self.data_size, (sf), replace=False) for i in range(N2)])
X.zero_grad()
drift_langevin = self.prior.logprob(X.params) + self.data_size/sf * X.loglik(self.data_X, self.data_Y,U)
drift_langevin.backward(torch.ones_like(drift_langevin))
X.params.data.copy_(X.params.data + hf*(X.params.grad) + sigma*dW)
sum1[step] += np.sum(self.Func(X)) # first moment
sum2[step] += np.sum(self.Func(X)**2) # second moment
var[step] = np.var(self.Func(X))
if step%100==0:
pbar.update(100)
return sum1/N, sum2/N, var
def solve_with_cv(self, N, sf):
"""Solve N stochastic langevin processes with subsample size sf
It returns E(F(X)) with X sampled from the posterior
Parameters
----------
N: int
Number of Langevin Processes to solve
sf: int
Subsample size
"""
hf = 5e-6#0.0001 #0.1/self.data_size#1e-8#1/self.data_size#self.T/self.n_steps
#sigma = 1/math.sqrt(self.data_size)
sigma = math.sqrt(2)
sum1, sum2 = np.zeros(self.n_steps), np.zeros(self.n_steps) #first and second order moments
var = np.zeros_like(sum1)
print("Solving SGLD with CV...")
n_steps = min(self.n_steps, 100000)
for N1 in range(0,N,5000):
N2 = min(5000, N-N1) # we will do batches of paths of size N2
X = LogisticNets(self.dim, N2).to(device=self.device)
self.init_weights(X)
# we extend MAP and grad_loklik_MAP to run several processes forward at the same time
grad_loglik_MAP = self.grad_loglik_MAP.repeat((N2,1))
MAP = copy.deepcopy(self.MAP)
MAP.params.data = self.MAP.params.data.repeat((N2,1))
# Euler scheme on Stochastic Langevin process with cv
pbar = tqdm.tqdm(total=n_steps)
for step in range(n_steps):
dW = math.sqrt(hf) * torch.randn_like(X.params)
U = np.random.choice(self.data_size, (N2, sf), replace=True)
#U = np.array([np.random.choice(self.data_size, (sf), replace=False) for i in range(N2)])
X.zero_grad()
MAP.zero_grad()
X.forward_backward_pass(self.data_X, self.data_Y, U)
MAP.forward_backward_pass(self.data_X, self.data_Y, U)
params_updated = X.params.data + hf * (self.prior.grad_logprob(X.params.data) +
grad_loglik_MAP + self.data_size/sf*(X.params.grad - MAP.params.grad)) + sigma*dW
X.params.data.copy_(params_updated)
sum1[step] += np.sum(self.Func(X)) # first order moment
sum2[step] += np.sum(self.Func(X)**2) # second order moment
var[step] = np.var(self.Func(X))
if step%100==0:
pbar.update(100)
return sum1/N, sum2/N, var
def make_plots(path_results,results):
"""
Make plots and write results
Parameters
----------
path: str
path where plot and data should be saved
results: List(Dict)
each element of args is a dictionary with the following (key,value) pairs
- "moment1":E(F(X))
- "moment2":E(F(X)**2)
- "label":str
"""
if not os.path.exists(path_results):
os.path.makedirs(path_results)
fig, ax = plt.subplots()
for d in results:
n_steps = range(len(d["moment1"]))
var = d["moment2"] - d["moment1"]**2
ax.plot(n_steps, d["moment1"], label=d["label"])
ax.fill_between(n_steps, d["moment1"]-np.sqrt(var), d["moment1"]+np.sqrt(var), alpha=0.5)
ax.legend()
fig.savefig(os.path.join(path_results,"sgld_cv.pdf"))
return 0
def set_seed(seed):
np.random.seed(seed)
torch.manual_seed(seed)
if __name__ == '__main__':
#CONFIGURATION
parser = argparse.ArgumentParser()
parser.add_argument('--N', type=int, default=10000, help='number samples from posterior to calculate E(F(X))')
parser.add_argument('--device', default='cpu', help='device')
parser.add_argument('--dim', type=int, default=2, help='dimension of data if data is synthetic')
parser.add_argument('--data_size', type=int, default=512, help="dataset size is data is synthetic")
parser.add_argument('--subsample_size', type=int, default=32, help="subsample size")
parser.add_argument('--T', type=int, default=10, help='horizon time')
parser.add_argument('--n_steps', type=int, default=10000, help='number of steps in time discretisation')
parser.add_argument('--seed', type=int, default=1, help='seed')
parser.add_argument('--type_data', type=str, default="synthetic")
parser.add_argument('--prior', type=str, default="Gaussian", help="type of prior")
args = parser.parse_args()
if args.device=='cpu' or (not torch.cuda.is_available()):
device='cpu'
else:
device = 'cuda:'+str(args.device)
# Target Logistic regression, and synthetic data
data_X, data_Y = get_dataset(m=args.data_size, d=args.dim,
type_regression="logistic", type_data=args.type_data, data_dir="./data/")
data_X = data_X.to(device=device)
data_Y = data_Y.to(device=device)
# path numerical results
dim = data_X.shape[1]
data_size = data_X.shape[0]
path_results = "./numerical_results/sgld_cv/logistic/{}_d{}_m{}_s{}".format(args.type_data, dim, data_size, args.subsample_size)
if not os.path.exists(path_results):
os.makedirs(path_results)
# load the mode
path_mode = "./numerical_results/sgld_cv/logistic/mode_{}_d{}_m{}".format(args.type_data, dim, data_size)
if not os.path.exists(path_mode):
os.makedirs(path_mode)
try:
MAP = torch.load(os.path.join(path_mode, "MAP.pth.tar"), map_location="cpu") # state_dict
except:
MAP = None
# prior configuration
PRIORS = {"Gaussian":Gaussian, "MixtureGaussians":MixtureGaussians}
CONFIG_PRIORS = config_priors(dim, device)
prior = PRIORS[args.prior](**CONFIG_PRIORS[args.prior])
# SGLD object
set_seed(args.seed)
sgld = SGLD(T=args.T,
n_steps=args.n_steps,
data_X=data_X,
data_Y=data_Y,
device=device,
prior=prior,
MAP=MAP)
# 1. We calculate E(F(X)) and E(F(X)**2) for stochastic Langevin process
sgld_1, sgld_2, sgld_var = sgld.solve(N=args.N, sf=args.subsample_size)
# 2. We calculate E(F(X)) and E(F(X)**2) for stochastic Langevin process with control variate
sgld_cv_1, sgld_cv_2, sgld_cv_var = sgld.solve_with_cv(N=args.N, sf=args.subsample_size)
# Plots and results
results = [dict(moment1=sgld_1, moment2=sgld_2, var=sgld_var, label="$E(F(X))$ - sgld"),
dict(moment1=sgld_cv_1, moment2=sgld_cv_2, var=sgld_cv_var, label="$E(F(X))$ - sgld_cv")]
print("sgld_var={:.4f}, sgld_cv_var={:.4f}".format(sgld_var[-1], sgld_cv_var[-1]))
print("sgld_1={:.4f}, sgld_cv_1={:.4f}".format(sgld_1[-1], sgld_cv_1[-1]))
#make_plots(path_results, results)
with open(os.path.join(path_results, "sgld_results.pickle"), "wb") as f:
pickle.dump(results, f)
sgld.save_MAP(os.path.join(path_mode, "MAP.pth.tar"))