forked from gwgundersen/rflvm
-
Notifications
You must be signed in to change notification settings - Fork 0
/
fit_model.py
264 lines (232 loc) · 8.5 KB
/
fit_model.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
"""============================================================================
Fit random feature latent variable model.
============================================================================"""
import argparse
from datasets import load_dataset
from logger import (format_number,
Logger)
from models import (BernoulliRFLVM,
GaussianRFLVM,
MultinomialRFLVM,
NegativeBinomialRFLVM,
PoissonRFLVM)
from metrics import (knn_classify,
mean_squared_error,
r_squared)
import numpy as np
from numpy.random import RandomState
from pathlib import Path
import pickle
from time import perf_counter
from visualizer import Visualizer
# -----------------------------------------------------------------------------
def fit_log_plot(args):
"""Fit model, plot visualizations, log metrics.
"""
# Configure logging, dataset, and visualizer.
# -------------------------------------------
p = Path(args.directory)
if not p.exists():
p.mkdir()
log = Logger(directory=args.directory)
log.log(f'Initializing RNG with seed {args.seed}.')
rng = RandomState(args.seed)
ds = load_dataset(rng, args.dataset, args.emissions)
viz = Visualizer(args.directory, ds)
# Set values on `args` so that they are logged.
args.n_burn = int(args.n_iters / 2) # Recommended in Gelman's BDA.
args.dp_prior_obs = ds.latent_dim
args.dp_df = ds.latent_dim + 1
args.marginalize = bool(args.marginalize)
args.log_every = 10
log.log_hline()
log.log_args(args)
# Initialize model.
# -----------------
if args.model == 'bernoulli':
model = BernoulliRFLVM(
rng=rng,
data=ds.Y,
n_burn=args.n_burn,
n_iters=args.n_iters,
latent_dim=ds.latent_dim,
n_clusters=args.n_clusters,
n_rffs=args.n_rffs,
dp_prior_obs=args.dp_prior_obs,
dp_df=args.dp_df
)
elif args.model == 'gaussian':
model = GaussianRFLVM(
rng=rng,
data=ds.Y,
n_burn=args.n_burn,
n_iters=args.n_iters,
latent_dim=ds.latent_dim,
n_clusters=args.n_clusters,
n_rffs=args.n_rffs,
dp_prior_obs=args.dp_prior_obs,
dp_df=args.dp_df,
marginalize=args.marginalize
)
elif args.model == 'poisson':
model = PoissonRFLVM(
rng=rng,
data=ds.Y,
n_burn=args.n_burn,
n_iters=args.n_iters,
latent_dim=ds.latent_dim,
n_clusters=args.n_clusters,
n_rffs=args.n_rffs,
dp_prior_obs=args.dp_prior_obs,
dp_df=args.dp_df
)
elif args.model == 'multinomial':
model = MultinomialRFLVM(
rng=rng,
data=ds.Y,
n_burn=args.n_burn,
n_iters=args.n_iters,
latent_dim=ds.latent_dim,
n_clusters=args.n_clusters,
n_rffs=args.n_rffs,
dp_prior_obs=args.dp_prior_obs,
dp_df=args.dp_df
)
elif args.model == 'negbinom':
if args.dataset == 's-curve' and args.emissions == 'gaussian':
raise NotImplementedError('Sampling `R` requires `Y` to be count '
'data but emissions are Gaussian.')
model = NegativeBinomialRFLVM(
rng=rng,
data=ds.Y,
n_burn=args.n_burn,
n_iters=args.n_iters,
latent_dim=ds.latent_dim,
n_clusters=args.n_clusters,
n_rffs=args.n_rffs,
dp_prior_obs=args.dp_prior_obs,
dp_df=args.dp_df
)
if args.model != args.emissions and args.dataset == 's-curve':
model_name = model.__class__.__name__
log.log_hline()
log.log(f'WARNING: Model is {model_name}, but emissions '
f'are {args.emissions}. Was this intended?')
# Visualize the initial value of `X`.
viz.plot_X_init(model.X)
# Fit model.
# ----------
s_start = perf_counter()
for t in range(args.n_iters):
s = perf_counter()
model.step()
e = perf_counter() - s
if t == model.n_burn:
log.log_hline()
log.log(f'Burn in complete on iter = {t}. Now plotting using mean '
f'of `X` samples after burn in.')
if (t % args.log_every == 0) or (t == args.n_iters - 1):
assert(model.t-1 == t)
plot_and_print(t, rng, log, viz, ds, model, e)
elapsed_time = (perf_counter() - s_start) / 3600
log.log_hline()
log.log(f'Finished job in {format_number(elapsed_time)} (hrs).')
log.log_hline()
# -----------------------------------------------------------------------------
def plot_and_print(t, rng, log, viz, ds, model, elapsed_time):
"""Utility function for plotting images and printing logs.
"""
# Generate model predictions.
# ---------------------------
Y_pred, F_pred, K_pred = model.predict(model.X, return_latent=True)
# Plot visualizations.
# --------------------
viz.plot_iteration(t, Y_pred, F_pred, K_pred, model.X)
log.log_hline()
log.log(t)
# Log metrics.
# ------------
mse_Y = mean_squared_error(Y_pred, ds.Y)
log.log_pair('MSE Y', mse_Y)
if ds.has_true_F:
mse_F = mean_squared_error(F_pred, ds.F)
log.log_pair('MSE F', mse_F)
if ds.has_true_K:
mse_K = mean_squared_error(K_pred, ds.K)
log.log_pair('MSE K', mse_K)
if ds.has_true_X:
r2_X = r_squared(model.X, ds.X)
log.log_pair('R2 X', r2_X)
if ds.is_categorical:
knn_acc = knn_classify(model.X, ds.labels, rng)
log.log_pair('KNN acc', knn_acc)
# Log parameters.
# ---------------
log.log_pair('DPMM LL', model.calc_dpgmm_ll())
log.log_pair('K', model.Z_count.tolist())
log.log_pair('alpha', model.alpha)
n_mh_iters = (model.t + 1) * model.M
log.log_pair('W MH acc', model.mh_accept / n_mh_iters)
if hasattr(model, 'R'):
log.log_pair('R median', np.median(model.R))
# Record time.
# ------------
log.log_pair('time', elapsed_time)
# Flush and save state.
# ---------------------
params = model.get_params()
fpath = f'{args.directory}/{args.model}_rflvm.pickle'
pickle.dump(params, open(fpath, 'wb'))
# -----------------------------------------------------------------------------
if __name__ == '__main__':
EMISSIONS = ['bernoulli', 'gaussian', 'multinomial', 'negbinom', 'poisson']
p = argparse.ArgumentParser()
p.add_argument('--directory',
help='Experimental directory.',
required=False,
default='experiments')
p.add_argument('--model',
help='Model to fit.',
required=False,
default='gaussian',
choices=EMISSIONS)
p.add_argument('--seed',
help='Random seed.',
required=False,
default=0,
type=int)
p.add_argument('--dataset',
help='Experimental dataset.',
type=str,
default='s-curve',
choices=['bridges', 'congress', 's-curve'])
p.add_argument('--emissions',
help='Emissions used S-curve dataset.',
required=False,
type=str,
default='gaussian',
choices=EMISSIONS)
p.add_argument('--n_iters',
help='Number of iterations for the Gibbs sampler.',
required=False,
type=int,
default=2000)
p.add_argument('--n_rffs',
help='Number of random Fourier features.',
required=False,
type=int,
default=100)
p.add_argument('--marginalize',
help='Whether or not to marginalize out `beta` in the '
'Gaussian model.',
type=int,
default=1)
p.add_argument('--n_clusters',
help='Number of initial clusters for `W`.',
required=False,
type=int,
default=1)
# Parse and validate script arguments.
# ------------------------------------
args = p.parse_args()
fit_log_plot(args)