-
Notifications
You must be signed in to change notification settings - Fork 1
/
icgmm_seq.py
469 lines (373 loc) · 20.5 KB
/
icgmm_seq.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
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
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
import torch
from pydgn.experiment.util import s2c
from torch.distributions import *
from torch.nn import ModuleList, Parameter
from torch_geometric.nn import global_mean_pool, global_add_pool
from torch_scatter import scatter_add, scatter_max
from util import compute_unigram, compute_bigram
class iCGMMSeq(torch.nn.Module):
def __init__(self, dim_node_features, dim_edge_features, dim_target,
readout_class, config):
super(iCGMMSeq, self).__init__()
self.device = None
self.is_first_layer = config['depth'] == 1
self.dim_node_features = dim_node_features
self.dim_target = dim_target
self.readout_class = readout_class
self.depth = config['depth']
self.training = False
self.max_C = config['max_C']
self.return_node_embeddings = False
self.K = dim_node_features
self.Y = dim_target
# HYPER-PRIOR HERE IS GAMMA: see 8.8.2 BRML
self.emission_distr_class = s2c(config['emission_distribution']['class'])
self.emission_distr_prior_params = config['emission_distribution']['prior_params']
# TODO: check this
if 'alpha' in config and 'alpha_prior_params' in config:
raise ValueError('Only one key between alpha or alpha_prior_params can be specified')
if 'alpha' not in config and 'gamma_prior_params' not in config:
raise ValueError('At least one key between alpha or alpha_prior_params must be specified')
if 'alpha' in config:
self.alpha = Parameter(torch.tensor(config['alpha'], dtype=torch.float32), requires_grad=False)
self.alpha_prior_params = None
elif 'alpha_prior_params' in config:
self.alpha = None
self.alpha_prior_params = {k: Parameter(torch.tensor(v, dtype=torch.float32), requires_grad=False)
for k, v in config['alpha_prior_params'].items()}
if 'gamma' in config and 'gamma_prior_params' in config:
raise ValueError('Only one key between gamma or gamma_prior_params can be specified')
if 'gamma' not in config and 'gamma_prior_params' not in config:
raise ValueError('At least one key between gamma or gamma_prior_params must be specified')
if 'gamma' in config:
self.gamma = Parameter(torch.tensor(config['gamma'], dtype=torch.float32), requires_grad=False)
self.gamma_prior_params = None
elif 'gamma_prior_params' in config:
self.gamma = None
self.gamma_prior_params = {k: Parameter(torch.tensor(v, dtype=torch.float32), requires_grad=False)
for k, v in config['gamma_prior_params'].items()}
self.Ccurr = Parameter(torch.tensor(2, dtype=torch.int32),
requires_grad=False) # start with 1/2 states and update during training
self.A = 1 # fixed
self.J = Parameter(torch.tensor(config['J'], dtype=torch.int32), requires_grad=False)
# self.add_self_arc = config['self_arc'] if 'self_arc' in config else False
self.sample_neighboring_macrostate = config['sample_neighboring_macrostate']
self.use_continuous_states = True # fixed
self.unibigram = config['unibigram']
self.aggregation = config['aggregation']
self._init_alpha_gamma()
self.beta = self._init_beta()
self.theta = self._init_theta()
self.init_permanent_accumulators()
def _init_alpha_gamma(self):
if self.alpha_prior_params is not None:
alpha_prior_distr = Gamma(self.alpha_prior_params['a'],
self.alpha_prior_params['b'])
self.alpha = alpha_prior_distr.sample()
if self.gamma_prior_params is not None:
gamma_prior_distr = Gamma(self.gamma_prior_params['a'],
self.gamma_prior_params['b'])
self.gamma = gamma_prior_distr.sample()
def _init_beta(self):
betas = torch.zeros(self.Ccurr + 1)
base_distrib = Beta(1, self.gamma)
# stick breaking process to inizialize beta
betas[0] = 1.
for i in range(0, self.Ccurr):
b = base_distrib.sample()
tmp = betas[i].clone()
betas[i] = b * tmp
betas[i + 1] = (1 - b) * tmp
assert torch.allclose(betas.sum(), torch.tensor(1.)), (betas, betas.sum())
return torch.nn.Parameter(betas, requires_grad=False)
def _init_theta(self):
assert self.Ccurr == 2
theta = ModuleList(
[self.emission_distr_class(self.K, **self.emission_distr_prior_params) for _ in range(self.Ccurr + 1)])
return theta
def init_permanent_accumulators(self):
"""
Accumulators that stay across Gibbs Sampling iterations
(and therefore must be stored/loaded to/from a checkpoint)
TODO tji and assignment matrices should be stored/loaded in checkpoints, but management is not straightforward
"""
self.njk = Parameter(torch.zeros(self.J, self.Ccurr + 1), requires_grad=False)
# counts the table assignments for each group j (restaurant) and mixture component k (dish)
self.tji = [[[0] for _ in range(self.Ccurr)] for _ in range(self.J)]
self.mjk = torch.zeros(self.J, self.Ccurr).to(self.device)
self.macrostate_assignments = None
self.xji_z_assignment = None
self.xji_t_assignment = None
self.first_run = True
# Do not delete this!
if self.device: # set by to() method
self.to(self.device)
def _add_beta(self):
# stick breaking process for new state
betas = self.beta
base_distrib = Beta(1, self.gamma)
b = base_distrib.sample()
beta_last = betas[-1].clone()
betas[-1] = b * beta_last
beta_new = (1 - b) * beta_last
betas = torch.cat((betas, beta_new.unsqueeze(0)), dim=0)
self.beta.data = betas
return self.beta
def _add_new_state(self):
# increment state counter
self.Ccurr.data += 1
# add new theta
self.theta.append(self.emission_distr_class(self.K, **self.emission_distr_prior_params))
def to(self, device):
super().to(device)
self.device = device
self.njk.to(device)
self.mjk.to(device)
for t in self.theta:
t.to(device)
return self
def train(self):
self.training = True
for t in self.theta:
t.train()
# print('Training, initializing permanent accumulators')
self.init_permanent_accumulators()
def eval(self):
self.training = False
################################
# WARNING #
################################
# TODO here I am assuming that whenever infer() is called in the training
# engine, the eval() method is also called. This should be avoided somehow
self.first_run = True
for t in self.theta:
t.eval()
def forward(self, data, id_batch):
extra = None
if not self.is_first_layer:
data, extra = data[0], data[1]
return self.sampling(data, extra)
def sampling(self, data, extra=None):
x, y, batch = data.x, data.y, data.batch
prev_stats = None if self.is_first_layer else extra.vo_outs
if prev_stats is not None:
prev_stats.to(self.device)
# assigno to x to groups
if self.macrostate_assignments is None: # self.first_run:
if self.is_first_layer:
# all x to the same group
self.macrostate_assignments = torch.zeros(x.shape[0], dtype=torch.int64).to(x.device)
else:
# sample neighbors macrostates
neighbors_macrostates = prev_stats[:, 0, :-1] # assume only 1 previous layer and no edge states
self.macrostate_assignments = Categorical(neighbors_macrostates).sample()
# compute f(x_u|theta_k) for all k (shape ?x(Ccurr+1))
f_X_theta_log = None
for k in range(self.Ccurr + 1):
log_emission = self.theta[k].get_data_log_likelihood(x) # ?
if f_X_theta_log is None:
f_X_theta_log = log_emission.unsqueeze(1)
else:
f_X_theta_log = torch.cat((f_X_theta_log, log_emission.unsqueeze(1)), dim=1)
if self.training:
if self.xji_z_assignment is None:
################################
# WARNING #
################################
# WE ASSUME FULL BATCH HERE!
# TODO: ALSO, THIS SHOULD BE A PARAMETER TO BE STORED/LOADED!!
# self.xji_z_assignment = (torch.zeros(x.shape[0], dtype=torch.int64) - 1).to(x.device)
self.xji_z_assignment = torch.randint(self.Ccurr.item(), [x.shape[0]], dtype=torch.int64).to(x.device)
idx = torch.stack((self.macrostate_assignments, self.xji_z_assignment), axis=0)
vals = torch.ones(idx.shape[1])
self.njk += torch.sparse_coo_tensor(idx, vals, size=self.njk.shape).coalesce().to_dense()
self.xji_t_assignment = (torch.zeros(x.shape[0], dtype=torch.int64) - 1).to(x.device)
# --------------------------- SAMPLE Z --------------------------- #
posterior_matrix = torch.zeros(x.shape[0], self.Ccurr + 1).to(x.device)
complete_log_likelihood_3 = torch.zeros(1).to(x.device)
complete_log_likelihood_2 = torch.zeros(1).to(x.device)
for u in range(x.shape[0]):
j = self.macrostate_assignments[u]
if self.training:
# Remove ji assignment from the total count
if self.xji_z_assignment[u] != -1:
old_p_zu = self.xji_z_assignment[u]
self.njk[j, old_p_zu] -= 1
else:
old_p_zu = None
unnorm_log_p_zu = (self.alpha * self.beta + self.njk[j]).log() + f_X_theta_log[u]
# log-sum epx trick
max_log_v = torch.max(unnorm_log_p_zu, dim=0, keepdim=True).values
log_zu = (unnorm_log_p_zu - max_log_v).exp().sum(0, keepdim=True).log() + max_log_v
p_zu = (unnorm_log_p_zu - log_zu).exp()
# populate posterior matrix
posterior_matrix[u] += p_zu
if self.Ccurr == self.max_C: # deal with limit case
zu = Categorical(p_zu[:-1]).sample()
else:
zu = Categorical(p_zu).sample()
complete_log_likelihood_2 += f_X_theta_log[u, zu]
complete_log_likelihood_3 += p_zu[zu].log()
if self.training:
if zu == (self.Ccurr):
# add a state
self._add_new_state()
log_emission = self.theta[-1].get_data_log_likelihood(x)
f_X_theta_log = torch.cat((f_X_theta_log, log_emission.unsqueeze(1)), dim=1)
posterior_matrix = torch.cat((posterior_matrix, torch.zeros(x.shape[0], 1).to(x.device)), dim=1)
self.njk.data = torch.cat((self.njk.data, torch.zeros(self.J, 1).to(x.device)), dim=1)
for restaurant in range(self.J): # do not use j as index here
self.tji[restaurant].append([]) # append new list of tables associated with new state
self.mjk = torch.cat((self.mjk, torch.zeros(self.J, 1).to(x.device)), dim=1)
self._add_beta() # append new beta at the end
# ONLY ONE DISH PER TABLE!
# Knowing z, we can compute the conditional table assignments
# assert len(self.tji) == self.J, (len(self.tji), self.J)
# for debug_j in range(len(self.tji)):
# assert len(self.tji[debug_j]) == self.Ccurr.item(), (len(self.tji[debug_j]), self.Ccurr.item())
if len(self.tji[j][zu]) == 0: # no tables yet
self.tji[j][zu].append(1) # create table with one customer
self.mjk[j, zu] += 1 # new table, update count
# assign to table
self.xji_t_assignment[u] = 0
else: # sample table
# Remove ji table assignment from the total count
if self.xji_t_assignment[u] != -1:
assert old_p_zu is not None
# assert len(self.tji[j]) > old_p_zu, (len(self.tji[j]), old_p_zu)
# assert len(self.tji[j]) == self.Ccurr, (len(self.tji[j]), self.Ccurr)
# assert len(self.tji[j][old_p_zu]) > self.xji_t_assignment[u], (len(self.tji[j][old_p_zu]), self.xji_t_assignment[u])
self.tji[j][old_p_zu][self.xji_t_assignment[u]] -= 1
if self.tji[j][old_p_zu][self.xji_t_assignment[u]] == 0:
# the table has become empty, decrease mjk
self.mjk[j, old_p_zu] -= 1
'''
TODO MAKE THIS WORK, but is not necessary to the functioning of iCGMM
# remove the table until we need a new one
del self.tji[j][old_p_zu][self.xji_t_assignment[u]]
# shift table idx only of those nodes with same restaurant and same dish
nodes_zu_j = torch.logical_and((self.xji_z_assignment == old_p_zu), (self.macrostate_assignments == j))
index = self.xji_t_assignment[u]
self.xji_t_assignment[nodes_zu_j][self.xji_t_assignment[nodes_zu_j] > index] -= 1
'''
# use the sampled zu here
unnorm_p_tji = torch.tensor(self.tji[j][zu] + [self.alpha * self.beta[zu]]).to(x.device)
t = Categorical(unnorm_p_tji).sample()
if t == len(self.tji[j][zu]): # add new table to group j for mixture component k
self.tji[j][zu].append(1)
self.mjk[j, zu] += 1 # new table, update count
else:
# Add ji table assignment to the total count
self.tji[j][zu][t] += 1
# assign to table
self.xji_t_assignment[u] = t
self.theta[zu].update_counts(x[u])
self.njk[j, zu] += 1
self.xji_z_assignment[u] = zu
# ASSUMES FULL BATCH
self.first_run = False
# leave this here (do not be tempted to move it above)
complete_log_likelihood = (posterior_matrix * (f_X_theta_log)).sum(1).sum()
# --------------------- HANDLE new/dead states ------------------- #
if self.return_node_embeddings:
posterior_matrix = posterior_matrix[:, :-1]
statistics_batch = self._compute_statistics(posterior_matrix, data, self.device)
node_unigram = compute_unigram(posterior_matrix, self.use_continuous_states)
graph_unigram = self._get_aggregation_fun()(node_unigram, batch)
if self.unibigram:
node_bigram = compute_bigram(posterior_matrix.float(), data.edge_index, batch,
self.use_continuous_states)
graph_bigram = self._get_aggregation_fun()(node_bigram, batch)
node_embeddings_batch = torch.cat((node_unigram, node_bigram), dim=1)
graph_embeddings_batch = torch.cat((graph_unigram, graph_bigram), dim=1)
else:
node_embeddings_batch = node_unigram
graph_embeddings_batch = graph_unigram
embeddings = (None, None, graph_embeddings_batch, statistics_batch, None, None)
else:
embeddings = None
num_nodes = x.shape[0]
return None, embeddings, complete_log_likelihood/num_nodes, \
complete_log_likelihood_2/num_nodes, \
complete_log_likelihood_3/num_nodes, num_nodes
def _update_alpha_gamma(self):
nodes_for_each_group = self.njk.sum(dim=1)
tot_number_of_tables = self.mjk.sum()
if self.alpha_prior_params is not None:
# sampling schema for alpha (from "Hierarchical Dirichlet Processes" by Jordan)
w_aux = torch.zeros(self.J)
s_aux = torch.zeros(self.J)
for j in range(self.J):
# Eq. (48)
w_aux_distr = Beta(self.alpha + 1, nodes_for_each_group[j])
w_aux[j] = w_aux_distr.sample()
# Eq. (49)
s_aux_distr = Bernoulli(nodes_for_each_group[j] / (nodes_for_each_group[j] + self.alpha))
s_aux[j] = s_aux_distr.sample()
# Eq. (47)
alpha_distr = Gamma(self.alpha_prior_params['a'] + tot_number_of_tables - torch.sum(s_aux),
self.alpha_prior_params['b'] - torch.sum(torch.log(w_aux)))
self.alpha = alpha_distr.sample()
if self.gamma_prior_params is not None:
# sampling schema for gamma (from "THE STICKY HDP-HMM" by Fox et al)
# Eq. (D.8)
eta_distr = Beta(self.gamma + 1,
tot_number_of_tables)
eta = eta_distr.sample()
zeta_distr = Bernoulli(tot_number_of_tables / (tot_number_of_tables + self.gamma))
zeta = zeta_distr.sample()
gamma_distr = Gamma(self.gamma_prior_params['a'] + self.Ccurr - zeta,
self.gamma_prior_params['b'] - torch.log(eta))
self.gamma = gamma_distr.sample()
def update(self):
# Check whether we need to add/remove states
states_to_remove = [k for k in range(self.Ccurr) if self.njk.sum(dim=0)[k] == 0]
# Q: How is this handled in other libraries?
for index in sorted(states_to_remove, reverse=True):
del self.theta[index]
self.njk.data = torch.cat((self.njk.data[:, :index], self.njk.data[:, index + 1:]), dim=1)
# reduce index value of state assignments that come after the
# removed index (remember, we are iterating in reverse order)
self.xji_z_assignment[self.xji_z_assignment > index] -= 1
for j in range(0, self.J):
del self.tji[j][index]
self.mjk.data = torch.cat((self.mjk.data[:, :index],
self.mjk.data[:, index + 1:]),
dim=1)
self.Ccurr.data -= 1
unnorm_p_beta = torch.cat((self.mjk.sum(dim=0),
torch.tensor([self.gamma])))
p_beta = unnorm_p_beta / unnorm_p_beta.sum()
self.beta.data = Dirichlet(p_beta).sample()
# update thetas, including resampling the categorical for the new state
for k in range(self.Ccurr + 1):
self.theta[k].update_parameters()
self._update_alpha_gamma()
def stopping_criterion(self, depth, max_layers, train_loss, train_score, val_loss, val_score,
dict_per_layer, layer_config, logger=None):
return depth == max_layers
def _compute_statistics(self, posteriors, data, device):
statistics = torch.full((posteriors.shape[0], self.A, posteriors.shape[1] + 1), 0., dtype=torch.float32).to(
device)
srcs, dsts = data.edge_index
assert self.A == 1
sparse_adj_matr = torch.sparse_coo_tensor(data.edge_index, \
torch.ones(data.edge_index.shape[1], dtype=posteriors.dtype).to(
device), \
torch.Size([posteriors.shape[0],
posteriors.shape[0]])).to(device).transpose(0, 1)
statistics[:, 0, :-1] = torch.sparse.mm(sparse_adj_matr, posteriors)
# Deal with nodes with degree 0: add a single fake neighbor with uniform posterior
degrees = statistics[:, :, :-1].sum(dim=[1, 2]).floor()
statistics[degrees == 0., :, :] = 1. / self.Ccurr.float()
return statistics
def _compute_sizes(self, batch, device):
return scatter_add(torch.ones(len(batch), dtype=torch.int).to(device), batch)
def _compute_max_ariety(self, degrees, batch):
return scatter_max(degrees, batch)
def _get_aggregation_fun(self):
if self.aggregation == 'mean':
aggregate = global_mean_pool
elif self.aggregation == 'sum':
aggregate = global_add_pool
return aggregate