/
pytorch_run.py
executable file
·249 lines (209 loc) · 8.94 KB
/
pytorch_run.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
import numpy as np
from torch.autograd import Variable
import torch.cuda
import _pickle as pickle
import argparse
import os
from tqdm import tqdm
from pytorch_model import ProdLDA
from dataloader import bitermsDataset
from sparseMM import *
parser = argparse.ArgumentParser()
parser.add_argument('-f', '--en1-units', type=int, default=100)
parser.add_argument('-s', '--en2-units', type=int, default=100)
parser.add_argument('-t', '--num-topic', type=int, default=50)
parser.add_argument('-b', '--batch-size', type=int, default=100)
parser.add_argument('-o', '--optimizer', type=str, default='Adam')
parser.add_argument('-r', '--learning-rate', type=float, default=0.002)
parser.add_argument('-m', '--momentum', type=float, default=0.99)
parser.add_argument('-e', '--num-epoch', type=int, default=200)
parser.add_argument('-q', '--init-mult', type=float, default=1.0) # multiplier in initialization of decoder weight
parser.add_argument('-v', '--variance', type=float, default=0.995) # default variance in prior normal
parser.add_argument('--start', action='store_true') # start training at invocation
parser.add_argument('--nogpu', action='store_true') # do not use GPU acceleration
parser.add_argument('-d', '--data-root', type=str, default='data/20news_clean/')
parser.add_argument('-a', '--betavariance', type=float, default=0.04)
args = parser.parse_args()
batch_size=args.batch_size
biterms = []
window_length = 30
mini_doc = 3
# default to use GPU, but have to check if GPU exists
if not args.nogpu:
if torch.cuda.device_count() == 0:
args.nogpu = True
def to_onehot(data, min_length):
return np.bincount(data, minlength=min_length)
def padding(data):
max_len = max([len(doc) for doc in data])
new_data = np.array([np.append(doc,[0] * (max_len - len(doc))) for doc in data]).astype(np.int64)
return new_data
def make_corpus(dataset, vocab):
file_name = args.data_root + "corpus.txt"
import os.path
if os.path.isfile(file_name):
return
with open(file_name, "w") as f:
for i in range(dataset.size):
line = [vocab[x] for x in dataset[i]]
f.write(" ".join(line) + "\n")
def put_dic(biterm, biterms):
if biterm not in biterms:
biterms[biterm] = 1
else:
biterms[biterm] += 1
def make_biterm(data, biterms):
import itertools
print (len(biterms))
for doc in data:
if np.sum(doc) == 0:
continue
doc_len = len(doc)
temp = {}
if doc_len > window_length:
for i in range(doc_len - window_length):
for biterm in itertools.combinations(doc[i:i+window_length], 2):
put_dic(frozenset(biterm), temp)
else:
for biterm in itertools.combinations(doc, 2):
put_dic(frozenset(biterm), temp)
biterms.append(temp)
print (len(biterms))
with open(args.data_root+'biterms.pickle', 'wb') as f:
pickle.dump(biterms, f)
def make_graph(biterms):
merge = {}
for d in biterms:
merge = {**merge, **d}
temp = np.zeros((args.num_input, args.num_input))
for k, v in merge.items():
if len(k) < 2:
a = b = list(k)[0]
else:
a, b = k
temp[a, b] = v
return temp
def make_data():
global data_tr, data_te, tensor_tr, tensor_te, vocab, vocab_size, n_samples_tr
global data_tr_index, data_tr_count, data_te_index, data_te_count, biterms
dataset_tr = args.data_root + 'train.txt.npy'
data_tr = np.load(dataset_tr, encoding='bytes')
vocab = args.data_root + 'vocab.pkl'
vocab = pickle.load(open(vocab,'rb'))
vocab_size=len(vocab)
args.num_input = vocab_size
biterm_file_name = args.data_root+'biterms.pickle'
import os.path
if os.path.isfile(biterm_file_name):
print('loading biterms')
with open(biterm_file_name, 'rb') as f:
biterms = pickle.load(f, encoding='bytes')
else:
print('generating biterm graphs')
make_biterm(data_tr, biterms)
biterms = np.array(biterms)
make_corpus(data_tr, list(zip(*sorted(vocab.items(), key=lambda x:x[1])))[0])
print('Data Loaded')
def make_model():
global model
net_arch = args # en1_units, en2_units, num_topic, num_input
net_arch.num_input = vocab_size
model = ProdLDA(net_arch)
if not args.nogpu:
model = model.cuda()
def make_optimizer():
global optimizer
if args.optimizer == 'Adam':
optimizer = torch.optim.Adadelta(model.params, lr=1, rho=0.99)
elif args.optimizer == 'SGD':
optimizer = torch.optim.SGD(model.parameters(), args.learning_rate, momentum=args.momentum)
else:
assert False, 'Unknown optimizer {}'.format(args.optimizer)
def train(dataloader):
for epoch in range(args.num_epoch):
loss_epoch = 0.0
model.train() # switch to training mode
#total_batch = int(n_samples_tr / batch_size)
sparsity = False
count = 0
b_count = 0
GCNsInputList = []
target_input = []
losses = []
data_size = len(dataloader)
for biterm in dataloader:
mm = SparseMM.apply
biterm = torch.FloatTensor(biterm).float().cuda()
target_input.append(Variable(biterm))
sparse_biterms = to_sparse(biterm.float().cuda())
ones = torch.cuda.FloatTensor(biterm.shape[0]).fill_(1).unsqueeze(-1)
indices = to_sparse(mm(sparse_biterms, ones))._indices()
values = torch.cuda.FloatTensor(indices.size()[1]).fill_(1)
adj_mask = torch.cuda.sparse.FloatTensor(indices, values, (sparse_biterms.size()[0], 1))
eye = sparse_ones(biterm.size()[0]).cuda()
adj = (sparse_biterms + eye).coalesce()
degree_matrix = mm(adj, ones)
degree_matrix = torch.pow(degree_matrix, -0.5)
degree_matrix = degree_matrix * adj_mask.to_dense()
degrees = sparse_diag(degree_matrix.squeeze(-1)).coalesce()
adj = mm(adj, degrees.to_dense())
adj = mm(degrees, adj)
indices = (sparse_biterms + eye).coalesce()._indices()
values = adj[tuple(indices[i] for i in range(indices.shape[0]))]
adj = torch.cuda.sparse.FloatTensor(indices, values, sparse_biterms.size())
GCNsInputList.append( (Variable(sparse_biterms),
Variable(adj)) )
b_count += 1
if b_count % batch_size != 0:
continue
if b_count > data_size:
break
_, loss = model(GCNsInputList, None, compute_loss=True, l1=False, target=target_input)
# optimize
optimizer.zero_grad() # clear previous gradients
loss.backward() # backprop
#torch.nn.utils.clip_grad_norm(model.params, 5)
optimizer.step() # update parameters
# report
loss_epoch += loss.data[0] # add loss to loss_epoch
GCNsInputList = []
target_input = []
if count % 10 == 0:
print('Epoch {}, count {}, loss={}'.format(epoch, count, loss_epoch / (count+1)))
count += 1
emb = torch.nn.functional.softmax(model.decoder.weight, 0).data.cpu().numpy().T
print_top_words(emb, list(zip(*sorted(vocab.items(), key=lambda x: x[1])))[0])
def print_top_words(beta, feature_names, n_top_words=10):
print( '---------------Printing the Topics------------------')
with open('topic_interpretability/data/topics_20news.txt', 'w') as f:
for i in range(len(beta)):
print(" ".join([feature_names[j+1]
for j in beta[i][1:].argsort()[:-n_top_words - 1:-1]]))
f.write(" ".join([feature_names[j+1]
for j in beta[i][1:].argsort()[:-n_top_words - 1:-1]]) + '\n')
print( '---------------End of Topics------------------')
def sample(data,biterms):
rng = np.random.RandomState(10)
data_new = []
biterms_new = []
for i in tqdm(range(len(data))):
ixs = rng.randint(data.shape[0], size=mini_doc)
sample_biterms = make_graph(biterms[ixs])
sample_datas = data[ixs].sum(0)
data_new.append(sample_datas)
biterms_new.append(sample_biterms)
return np.array(data_new), biterms_new
def save_checkpoint(model, path):
torch.save(model, os.path.join(path, 'model.pt'))
if __name__=='__main__' and args.start:
make_data()
make_model()
make_optimizer()
data_new = []
print (len(biterms))
dataset = bitermsDataset(np.array(biterms), args.num_input, mini_doc, data=None)
print ('start training')
train(dataset)
emb = torch.nn.functional.softmax(model.decoder.weight, 0).data.cpu().numpy().T
print_top_words(emb, list(zip(*sorted(vocab.items(), key=lambda x:x[1])))[0])
save_checkpoint(model, 'saved_models')