-
Notifications
You must be signed in to change notification settings - Fork 1
/
predict.py
198 lines (167 loc) · 6.19 KB
/
predict.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
import json
import os
import pickle
from collections import defaultdict
from operator import itemgetter
import numpy as np
import pytorch_lightning as pl
import torch
from absl import app, flags
from numpy import average
from pytorch_lightning import Trainer, seed_everything
from pytorch_lightning.callbacks import ModelCheckpoint
from data.data import SequenceDataLoader
from model import LSTMRegressor
from utils import backoff_model, create_matrix
flags.DEFINE_boolean('debug', False, '')
flags.DEFINE_integer('epochs', 9, '')
flags.DEFINE_integer('batch_size', 1, '')
flags.DEFINE_float('lr', 1e-4 , '')
flags.DEFINE_float('momentum', .9, '')
flags.DEFINE_float('dropout', .3, '')
# flags.DEFINE_string('dataset', 'SequenceDataLoader', '')
# flags.DEFINE_string('model', 'bert-base-uncased', '')
flags.DEFINE_integer('seq_length', 32, '')
flags.DEFINE_integer('embedding_size', 512, '')
flags.DEFINE_integer('lstm_size', 128, '')
flags.DEFINE_integer('hidden_size', 128, '')
flags.DEFINE_integer('layers', 2, '')
FLAGS = flags.FLAGS
def replace_sos_eos(word):
if word == '<s>':
word = '<sos>'
elif word == '</s>':
word = '<eos>'
return word
def prob_method():
data_module = SequenceDataLoader(FLAGS.batch_size)
data_module.prepare_data('data')
vocab = data_module.vocab
lookup = vocab.get_itos()
vocab_size = len(vocab)
padding_idx = vocab['<pad>']
# Intialise model with trained parameters
model = LSTMRegressor(vocab_size,
FLAGS.embedding_size,
FLAGS.lstm_size,
FLAGS.hidden_size,
FLAGS.seq_length,
padding_idx,
FLAGS.batch_size,
FLAGS.layers,
FLAGS.dropout,
FLAGS.lr)
print(model)
checkpoint = torch.load('callback_logs/new_data/version1/epoch=8-step=4850.ckpt')
model.load_state_dict(checkpoint['state_dict'])
# Set model/data loader to eval mode
data_module.setup('test')
model.eval()
torch.set_grad_enabled(False)
device = model.device
# Initialise
ngram_dict = defaultdict(list)
lookup = vocab.get_itos()
reverse_lookup = vocab.get_stoi()
# backoff LM
bigram = create_matrix('data/new_lm.arpa', reverse_lookup)
ngram_save = defaultdict(list)
# Calculate n-gram probabilities
for count, batch in enumerate(data_module.test_dataloader()):
print(count*512)
# import pdb; pdb.set_trace()
pred = model.test_step(batch)
x, y, x_len, y_len = batch
x = x.view(x.shape[0]*x.shape[1], -1, 1).numpy()
x = np.repeat(x, repeats=len(lookup), axis=1)
pred_indices = pred.indices.view(pred.indices.shape[1]*pred.indices.shape[0], pred.indices.shape[2], 1).numpy()
pred_prob = pred.values.view(pred.values.shape[1]*pred.values.shape[0], pred.values.shape[2], 1).numpy()
# bigram
indices = np.dstack((x, pred_indices)).reshape(-1,2)
prob = pred_prob.reshape(-1,1)
result = list(zip(indices, prob))
with open(f'data/output/ngram_prob_{str(count)}', 'wb') as outfile:
pickle.dump(result, outfile)
return ngram_dict
def modified_prob_method():
data_module = SequenceDataLoader(FLAGS.batch_size)
data_module.prepare_data('data')
vocab = data_module.vocab
lookup = vocab.get_itos()
reverse_lookup = vocab.get_stoi()
vocab_size = len(vocab)
padding_idx = vocab['<pad>']
# Intialise model with trained parameters
model = LSTMRegressor(vocab_size,
FLAGS.embedding_size,
FLAGS.lstm_size,
FLAGS.hidden_size,
FLAGS.seq_length,
padding_idx,
FLAGS.batch_size,
FLAGS.layers,
FLAGS.dropout,
FLAGS.lr)
print(model)
checkpoint = torch.load('callback_logs/new_data/version1/epoch=8-step=4850.ckpt')
model.load_state_dict(checkpoint['state_dict'])
# Set model/data loader to eval mode
data_module.setup('test')
model.eval()
torch.set_grad_enabled(False)
device = model.device
# initialise
bigram = defaultdict(list)
trigram = defaultdict(list)
# backoff LM
lm = backoff_model(os.path.join('data', 'new_lm.arpa'))
unigram = lm.ngrams[1][()]
unigram['<sos'] = unigram['<s>']
unigram['<eos'] = unigram['</s>']
unigram.pop('<s>')
unigram.pop('</s>')
lm_vocab = list(unigram.keys())
# bigram
for word_pred, val in lm.ngrams[2].items():
word = replace_sos_eos(word_pred[0])
if word in reverse_lookup:
pred = model.test_step(
(torch.tensor([[reverse_lookup[word]]]),
None,
torch.tensor([1], dtype=torch.int32),
None)
)
for prob, word_next in zip(pred.values[0], pred.indices[0]):
word_next = replace_sos_eos(lookup[word_next])
if word_next in val:
bigram[((word,), word_next)].append(prob)
bigram_average = {}
for k,v in bigram.items():
bigram_average[k] = average(v)
with open('data/output/bigram', 'wb') as f:
pickle.dump(bigram_average, f)
# trigram
for word_pred, val in lm.ngrams[3].items():
word = (replace_sos_eos(word_pred[0]), replace_sos_eos(word_pred[1]))
if (word[0] in reverse_lookup) and (word[1] in reverse_lookup):
pred = model.test_step(
(torch.tensor([[reverse_lookup[i] for i in word]]),
None,
torch.tensor([2], dtype=torch.int32),
None)
)
for prob, word_next in zip(pred.values[-1], pred.indices[-1]):
word_next = replace_sos_eos(lookup[word_next])
if word_next in val:
trigram[((word), word_next)].append(prob)
trigram_average = {}
for k,v in trigram.items():
trigram_average[k] = average(v)
with open('data/output/trigram', 'wb') as f:
pickle.dump(trigram_average, f)
def main(_):
seed_everything(42, workers=True)
# prob_method()
modified_prob_method()
if __name__ == '__main__':
app.run(main)