-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathPreferenceTopic.py
214 lines (196 loc) · 11.5 KB
/
PreferenceTopic.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
from gumbel_softmax import GumbelSoftmax
from tau_scheduler import TauScheduler
import torch
import torch.nn as nn
from tools import Tools
class PriorPreference(nn.Module):
def __init__(self, args, encoder, decoder, hidden_size, n_topic_vocab,
trg_bos_idx, max_seq_len, glo2loc, loc2glo, main_tfr_encoder,
gs: GumbelSoftmax, ts: TauScheduler):
super(PriorPreference, self).__init__()
self.args = args
self.decoder = decoder
self.p_encoder = encoder
self.main_tfr_encoder = main_tfr_encoder
self.n_topic_vocab = n_topic_vocab
self.bos_idx = trg_bos_idx
self.max_seq_len = max_seq_len
self.gs = gs
self.ts = ts
self.glo2loc = glo2loc
self.loc2glo = loc2glo
self.hidden_size = hidden_size
self.gen_proj = nn.Sequential(nn.Linear(self.hidden_size, self.n_topic_vocab))
def forward(self, context, context_len, tp_path, tp_path_len, encoder_embed=None, decoder_embed=None):
bs = context.size(0)
context_mask = Tools.get_mask_via_len(context_len, self.args.context_max_len)
context_hidden = self.main_tfr_encoder(context, context_mask)
tp_path = one_hot_scatter(tp_path, self.n_topic_vocab)
tp_mask = Tools.get_mask_via_len(tp_path_len, self.args.state_num)
tp_hidden = self.p_encoder(tp_path, tp_mask, embed=encoder_embed)
src_hiddens = torch.cat([tp_hidden, context_hidden], 1)
src_mask = torch.cat([tp_mask, context_mask], 2)
seq_gen_gumbel = Tools._generate_init(bs, self.n_topic_vocab, trg_bos_idx=self.bos_idx, training=self.training)
seq_gen_prob = None
seq_gen_prob_raw = None
for _ in range(self.args.preference_num):
dec_output = Tools._single_decode(seq_gen_gumbel.detach(), src_hiddens, src_mask, self.decoder)
single_step_prob, single_step_prob_raw = self.proj(dec_out=dec_output, context=context, src_hidden=src_hiddens, src_mask=src_mask, tp_path=tp_path, embed=decoder_embed)
single_step_gumbel_word = self.gs.forward(single_step_prob, self.ts.step_on(), normed=True)
if self.training:
if seq_gen_prob is not None:
seq_gen_prob = torch.cat([seq_gen_prob, single_step_prob], 1)
seq_gen_prob_raw = torch.cat([seq_gen_prob_raw, single_step_prob_raw], 1)
else:
seq_gen_prob = single_step_prob
seq_gen_prob_raw = single_step_prob_raw
seq_gen_gumbel = torch.cat([seq_gen_gumbel, single_step_gumbel_word], 1)
else:
if seq_gen_prob is not None:
seq_gen_prob = torch.cat([seq_gen_prob, single_step_prob], 1)
else:
seq_gen_prob = single_step_prob
single_step_word = torch.argmax(single_step_prob, -1)
seq_gen_gumbel = torch.cat([seq_gen_gumbel, single_step_word], 1)
if self.training:
return seq_gen_prob, seq_gen_gumbel[:, 1:, :]
else:
return seq_gen_prob, seq_gen_gumbel[:, 1:]
def proj(self, dec_out, context, src_hidden, src_mask, tp_path, embed=None):
gen_logit = self.gen_prob(dec_out, embed)
L_s = dec_out.size(1)
B = context.size(0)
copy_logit = torch.bmm(dec_out, src_hidden.permute(0, 2, 1))
copy_logit = copy_logit.masked_fill((src_mask == 0).expand(-1, L_s, -1), -1e9 if copy_logit.dtype==torch.float32 else -1e4)
logits = torch.cat([gen_logit, copy_logit], -1)
if self.args.scale_prj:
logits *= self.hidden_size ** -0.5
probs = torch.softmax(logits, -1)
gen_prob = probs[:, :, :self.n_topic_vocab]
copy_context_prob = probs[:, :, self.n_topic_vocab + self.args.state_num:]
transfer_context_word = torch.gather(self.glo2loc.unsqueeze(0).expand(B, -1), 1, context)
copy_context_temp = copy_context_prob.new_zeros(B, L_s, self.n_topic_vocab)
copy_context_prob = copy_context_temp.scatter_add(dim=2,
index=transfer_context_word.unsqueeze(1).expand(-1, L_s, -1),
src=copy_context_prob)
copy_tp_prob = probs[:, :, self.n_topic_vocab + self.args.preference_num:self.n_topic_vocab + self.args.preference_num + self.args.state_num]
copy_tp_prob = torch.bmm(copy_tp_prob, tp_path)
probs = gen_prob + copy_tp_prob + copy_context_prob
probs_raw = logits
gen_prob_raw = probs_raw[:, :, :self.n_topic_vocab]
copy_tp_prob_raw = probs_raw[:, :, self.n_topic_vocab + self.args.preference_num:self.n_topic_vocab + self.args.preference_num + self.args.state_num]
copy_tp_prob_raw = torch.bmm(copy_tp_prob_raw, tp_path)
copy_context_prob_raw = probs_raw[:, :, self.n_topic_vocab + self.args.state_num:]
copy_context_temp_raw = copy_context_prob.new_zeros(B, L_s, self.n_topic_vocab)
copy_context_prob_raw = copy_context_temp_raw.scatter_add(dim=2,
index=transfer_context_word.unsqueeze(1).expand(-1,
L_s,
-1),
src=copy_context_prob_raw)
probs_raw = gen_prob_raw + copy_tp_prob_raw + copy_context_prob_raw
return probs, probs_raw
def gen_prob(self, dec_output, embed=None):
if embed is None:
prob = self.gen_proj(dec_output)
else:
assert embed.size(0) == self.gen_proj[0].out_features
prob = dec_output.matmul(embed.T)
return prob
class PosteriorPreference(nn.Module):
def __init__(self, encoder, decoder, main_encoder, hidden_size, n_topic_vocab, glo2loc, loc2glo,
trg_bos_idx, max_seq_len, gs: GumbelSoftmax, ts: TauScheduler):
super(PosteriorPreference, self).__init__()
self.p_encoder = encoder
self.decoder = decoder
self.main_encoder = main_encoder
self.glo2loc = glo2loc
self.loc2glo = loc2glo
self.hidden_size = hidden_size
self.n_topic_vocab = n_topic_vocab
self.trg_bos_idx = trg_bos_idx
self.max_seq_len = max_seq_len
self.gs = gs
self.ts = ts
self.gen_proj = nn.Sequential(nn.Linear(self.hidden_size, self.n_topic_vocab))
def forward(self, context, context_len, ar_gth, ar_gth_len, tp_path, tp_path_len, encoder_embed=None, decoder_embed=None):
bs = context.size(0)
context_mask = Tools.get_mask_via_len(context_len, self.args.context_max_len)
context_hidden = self.main_encoder(context, context_mask)
tp_path = one_hot_scatter(tp_path, self.n_topic_vocab)
tp_mask = Tools.get_mask_via_len(tp_path_len, self.args.state_num)
tp_hidden = self.p_encoder(tp_path, tp_mask, embed=encoder_embed)
if ar_gth is not None and ar_gth_len is not None:
ar_gth_len = [int(length / 2) for length in ar_gth_len]
ar_gth_len = torch.tensor(ar_gth_len).cuda()
ar_gth = ar_gth[:, list(range(1, self.args.action_num, 2))]
ar_gth = one_hot_scatter(ar_gth, self.n_topic_vocab)
ar_mask = Tools.get_mask_via_len(ar_gth_len, int(self.args.action_num / 2))
ar_hidden = self.p_encoder(ar_gth, ar_mask, embed=encoder_embed)
src_hiddens = torch.cat([tp_hidden, context_hidden, ar_hidden], 1)
ar_mask[ar_mask] = False
src_mask = torch.cat([tp_mask, context_mask, ar_mask], 2)
else:
src_hiddens = torch.cat([tp_hidden, context_hidden], 1)
src_mask = torch.cat([tp_mask, context_mask], 2)
seq_gen_gumbel = Tools._generate_init(bs, self.n_topic_vocab, trg_bos_idx=self.trg_bos_idx, training=self.training)
seq_gen_prob = None
for _ in range(self.args.preference_num):
dec_output = Tools._single_decode(seq_gen_gumbel.detach(), src_hiddens, src_mask, self.decoder)
single_step_prob = self.proj(dec_out=dec_output, src_hidden=src_hiddens, src_mask=src_mask, context=context, tp=tp_path, embed=decoder_embed, no_action=ar_gth is None)
if self.training:
single_step_gumbel_word = self.gs.forward(single_step_prob, self.ts.step_on(), normed=True)
if seq_gen_prob is not None:
seq_gen_prob = torch.cat([seq_gen_prob, single_step_prob], 1)
else:
seq_gen_prob = single_step_prob
seq_gen_gumbel = torch.cat([seq_gen_gumbel, single_step_gumbel_word], 1)
else:
if seq_gen_prob is not None:
seq_gen_prob = torch.cat([seq_gen_prob, single_step_prob], 1)
else:
seq_gen_prob = single_step_prob
single_step_word = torch.argmax(single_step_prob, -1)
seq_gen_gumbel = torch.cat([seq_gen_gumbel, single_step_word], 1)
if self.training:
return seq_gen_prob, seq_gen_gumbel[:, 1:, :]
else:
return seq_gen_prob, seq_gen_gumbel[:, 1:]
def proj(self, dec_out, src_hidden, src_mask, context, tp, embed=None, no_action=False):
B, L_s = dec_out.size(0), dec_out.size(1)
gen_logit = self.gen_prob(dec_out, embed)
if no_action:
hidden_no_At = src_hidden
mask_no_At = src_mask
else:
hidden_no_At = src_hidden[:, :-self.args.action_num // 2, :]
mask_no_At = src_mask[:, :, :-self.args.action_num // 2]
copy_logit = torch.bmm(dec_out, hidden_no_At.permute(0, 2, 1))
copy_logit = copy_logit.masked_fill((mask_no_At == 0).expand(-1, L_s, -1), -1e9 if copy_logit.dtype==torch.float32 else -1e4)
logits = torch.cat([gen_logit, copy_logit], -1)
if self.args.scale_prj:
logits *= self.hidden_size ** -0.5
probs = torch.softmax(logits, -1)
gen_prob = probs[:, :, :self.n_topic_vocab]
copy_tp_path_prob = probs[:, :, self.n_topic_vocab:self.n_topic_vocab + self.args.state_num]
copy_tp_path_prob = torch.bmm(copy_tp_path_prob, tp)
copy_context_prob = probs[:, :, self.n_topic_vocab + self.args.state_num:]
transfer_context_word = torch.gather(self.glo2loc.unsqueeze(0).expand(B, -1), 1, context)
copy_context_temp = copy_context_prob.new_zeros(B, L_s, self.n_topic_vocab)
copy_context_prob = copy_context_temp.scatter_add(dim=2,
index=transfer_context_word.unsqueeze(1).expand(-1, L_s, -1),
src=copy_context_prob)
probs = gen_prob + copy_tp_path_prob + copy_context_prob
return probs
def gen_prob(self, dec_output, embed=None):
if embed is None:
prob = self.gen_proj(dec_output)
else:
assert embed.size(0) == self.gen_proj[0].out_features
prob = dec_output.matmul(embed.T)
return prob
def one_hot_scatter(indice, num_classes, dtype=torch.float):
indice_shape = list(indice.shape)
placeholder = torch.zeros(*(indice_shape + [num_classes]), device=indice.device, dtype=dtype)
v = 1 if dtype == torch.long else 1.0
placeholder.scatter_(-1, indice.unsqueeze(-1), v)
return placeholder