-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy path10_DMN.py
178 lines (173 loc) · 7.42 KB
/
10_DMN.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
# -*- coding: utf-8 -*-
import torch, random, numpy as np
import torch.nn as nn
import torch.optim as optim
import torch.nn.functional as F
from torch.autograd import Variable
from copy import deepcopy
from tqdm import tqdm
USE_CUDA = torch.cuda.is_available()
if USE_CUDA: torch.cuda.set_device(0)
FTensor = torch.cuda.FloatTensor if USE_CUDA else torch.FloatTensor
LTensor = torch.cuda.LongTensor if USE_CUDA else torch.LongTensor
BTensor = torch.cuda.ByteTensor if USE_CUDA else torch.ByteTensor
class Model(nn.Module):
def __init__(self, v_size, h_dim, o_size, w2i, dropout=0.1):
super(Model,self).__init__()
self.h_dim = h_dim
self.embed = nn.Embedding(v_size,h_dim,padding_idx=0)
self.F_gru = nn.GRU(h_dim,h_dim,batch_first=True)
self.Q_gru = nn.GRU(h_dim,h_dim,batch_first=True)
self.gate = nn.Sequential(\
nn.Linear(h_dim*4,h_dim),nn.Tanh(),nn.Linear(h_dim,1),nn.Sigmoid())
self.attn_gcell = nn.GRUCell(h_dim,h_dim)
self.mem_gcell = nn.GRUCell(h_dim,h_dim)
self.ans_gcell = nn.GRUCell(h_dim*2,h_dim)
self.ans_fc = nn.Linear(h_dim,o_size)
self.dropout = nn.Dropout(dropout)
self.w2i = w2i
def init_hidden(self, inputs):
hidden = Variable(torch.zeros(1,inputs.size(0),self.h_dim))
return hidden.cuda() if USE_CUDA else hidden
def init_weight(self):
nn.init.xavier_uniform_(self.embed.state_dict()['weight'])
for name,param in self.F_gru.state_dict().items():
if 'weight' in name: nn.init.xavier_normal_(param)
for name,param in self.Q_gru.state_dict().items():
if 'weight' in name: nn.init.xavier_normal_(param)
for name,param in self.gate.state_dict().items():
if 'weight' in name: nn.init.xavier_normal_(param)
for name,param in self.attn_gcell.state_dict().items():
if 'weight' in name: nn.init.xavier_normal_(param)
for name,param in self.mem_gcell.state_dict().items():
if 'weight' in name: nn.init.xavier_normal_(param)
for name,param in self.ans_gcell.state_dict().items():
if 'weight' in name: nn.init.xavier_normal_(param)
nn.init.xavier_normal_(self.ans_fc.state_dict()['weight'])
self.ans_fc.bias.data.fill_(0)
def forward(self, facts, fms, Qs, qms, n_dec, epi=3, is_training=False):
# Facts
enc_facts = []
for fact,fm in zip(facts,fms):
embed = self.embed(fact)
if is_training: embed = self.dropout(embed)
hidden = self.init_hidden(fact)
outputs,_ = self.F_gru(embed,hidden)
hidden = [o[fm[i].data.tolist().count(0)-1] for i,o in enumerate(outputs)]
enc_facts.append(torch.cat(hidden).view(fact.size(0),-1).unsqueeze(0))
enc_facts = torch.cat(enc_facts)
# Qs
embed = self.embed(Qs)
if is_training: embed = self.dropout(embed)
hidden = self.init_hidden(Qs)
outputs,hidden = self.Q_gru(embed,hidden)
enc_Q = torch.cat([o[qms[i].data.tolist().count(0)-1] for i,o in enumerate(outputs)]).view(Qs.size(0),-1)
# Mem
mem = enc_Q
B = enc_facts.size(0)
T = enc_facts.size(1)
for i in xrange(epi):
hidden = self.init_hidden(enc_facts.transpose(0,1)[0]).squeeze(0)
for t in xrange(T):
z = torch.cat([enc_facts.transpose(0,1)[t]*enc_Q,\
enc_facts.transpose(0,1)[t]*mem,\
torch.abs(enc_facts.transpose(0,1)[t]-enc_Q),\
torch.abs(enc_facts.transpose(0,1)[t]-mem)],1)
g_t = self.gate(z)
hidden = g_t*self.attn_gcell(enc_facts.transpose(0,1)[t],hidden)+(1-g_t)*hidden
mem = self.mem_gcell(hidden,mem)
# Ans
ans_hidden = mem
start = Variable(LTensor([[self.w2i['<s>']]*mem.size(0)])).transpose(0,1)
y_t_1 = self.embed(start).squeeze(1); decodes = []
for t in xrange(n_dec):
ans_hidden = self.ans_gcell(torch.cat([y_t_1,enc_Q],1),ans_hidden)
decodes.append(F.log_softmax(self.ans_fc(ans_hidden),1))
return torch.cat(decodes,1).view(B*n_dec,-1)
class DMN(object):
def __init__(self):
self.b_size = 64
self.lr = 0.001
def train(self):
def load(fn):
data = []; fact = []
for line in open(fn).read().strip().split('\n'):
ind = line.split(' ')[0]
if ind == '1': fact = []
if '?' in line:
q,a,_ = line.split('\t')
q = q.strip().replace('?','').split(' ')[1:]+['?']
a = a.split()+['</s>']
data.append([deepcopy(fact),q,a])
else:
fact.append(line.replace('.','').split(' ')[1:]+['</s>'])
return data
flat = lambda a:[c for b in a for c in b]
data_tr = load('data/10/qa5_three-arg-relations_train.txt')
fact,q,a = zip(*data_tr)
w2i = {w:i+4 for i,w in enumerate(set(flat(flat(fact))+flat(q)+flat(a)))}
w2i.update({'<PAD>':0,'<UNK>':1,'<s>':2,'</s>':3})
idx = lambda x:Variable(LTensor([w2i[w] if w in w2i else w2i['<UNK>'] for w in x])).view(1,-1)
for i in xrange(len(data_tr)):
for j,fact in enumerate(data_tr[i][0]):
data_tr[i][0][j] = idx(fact)
data_tr[i][1] = idx(data_tr[i][1])
data_tr[i][2] = idx(data_tr[i][2])
model = Model(len(w2i)+1,80,len(w2i)+1,w2i)
model.init_weight()
if USE_CUDA: model = model.cuda()
loss_func = nn.CrossEntropyLoss(ignore_index=0)
opt = optim.Adam(model.parameters(),lr=self.lr)
# training
def pad(batch):
fact,q,a = zip(*batch)
max_f = max([len(f) for f in fact])
max_l = max([x.size(1) for x in flat(fact)])
max_q = max([x.size(1) for x in q])
max_a = max([x.size(1) for x in a])
facts,fm,Qs,As = [],[],[],[]
for i in xrange(len(batch)):
facts.append(torch.cat([torch.cat(\
[fact[i][j],Variable(LTensor([w2i['<PAD>']]*(max_l-fact[i][j].size(1)))).view(1,-1)],1)\
if fact[i][j].size(1)<max_l else fact[i][j] for j in xrange(len(fact[i]))]+\
[Variable(LTensor([w2i['<PAD>']]*max_l)).view(1,-1) for _ in xrange(max_f-len(fact[i]))]))
fm.append(torch.cat([Variable(BTensor(tuple(map(lambda s:s==0,t.data))),volatile=False) for t in facts[-1]]).view(facts[-1].size(0),-1))
Qs.append(torch.cat([q[i],Variable(LTensor([w2i['<PAD>']]*(max_q-q[i].size(1)))).view(1,-1)],1) if q[i].size(1)<max_q else q[i])
As.append(torch.cat([a[i],Variable(LTensor([w2i['<PAD>']]*(max_a-a[i].size(1)))).view(1,-1)],1) if a[i].size(1)<max_a else a[i])
Qs,As = torch.cat(Qs),torch.cat(As)
qm = torch.cat([Variable(BTensor(tuple(map(lambda s:s==0,t.data))),volatile=False) for t in Qs]).view(Qs.size(0),-1)
return facts, fm, Qs, qm, As
for epoch in xrange(50):
random.shuffle(data_tr); losses = []
for i in tqdm(xrange(0,len(data_tr),self.b_size)):
facts,fm,Qs,Qm,As = pad(data_tr[i:i+self.b_size])
model.zero_grad()
pred = model(facts,fm,Qs,Qm,As.size(1),3,True)
loss = loss_func(pred,As.view(-1))
loss.backward(); opt.step()
losses.append(loss.data.tolist())
if np.mean(losses) < 0.01: break
print np.mean(losses); losses = []
# testing
def pad_fact(fact):
x_max = max([s.size(1) for s in fact])
fact = torch.cat([torch.cat([fact[i],Variable(LTensor([w2i['<PAD>']]*(x_max-fact[i].size(1)))).view(1,-1)],1)\
if fact[i].size(1)<x_max else fact[i] for i in xrange(len(fact))])
fm = torch.cat([Variable(BTensor(tuple(map(lambda s:s==0,t.data))),volatile=False) for t in fact]).view(fact.size(0),-1)
return fact, fm
data_te = load('data/10/qa5_three-arg-relations_test.txt')
for i in xrange(len(data_te)):
for j,fact in enumerate(data_te[i][0]):
data_te[i][0][j] = idx(fact)
data_te[i][1] = idx(data_te[i][1])
data_te[i][2] = idx(data_te[i][2])
accu = 0
for fact,Q,A in data_te:
fact,fm = pad_fact(fact)
qm = Variable(BTensor([0]*Q.size(1)),volatile=False).unsqueeze(0)
A = A.squeeze(0)
p = model([fact],[fm],Q,qm,A.size(0),3)
if p.max(1)[1].data.tolist() == A.data.tolist(): accu += 1
print 100.*accu/len(data_te)
if __name__ == '__main__':
DMN().train()