-
Notifications
You must be signed in to change notification settings - Fork 88
/
main.py
255 lines (213 loc) · 8.58 KB
/
main.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
#!/usr/bin/env python2
# check profiler
if not isinstance(__builtins__, dict) or "profile" not in __builtins__:
__builtins__.__dict__["profile"] = lambda x: x
from misc import util
from misc.indices import QUESTION_INDEX, ANSWER_INDEX, MODULE_INDEX, MODULE_TYPE_INDEX, \
NULL, NULL_ID, UNK_ID
from misc.visualizer import visualizer
import models
from models.nmn import MLPFindModule, MultiplicativeFindModule
import tasks
import apollocaffe
import argparse
import json
import logging.config
import random
import numpy as np
import yaml
def main():
config = configure()
task = tasks.load_task(config)
model = models.build_model(config.model, config.opt)
for i_epoch in range(config.opt.iters):
train_loss, train_acc, _ = \
do_iter(task.train, model, config, train=True)
val_loss, val_acc, val_predictions = \
do_iter(task.val, model, config, vis=True)
test_loss, test_acc, test_predictions = \
do_iter(task.test, model, config)
logging.info(
"%5d | %8.3f %8.3f %8.3f | %8.3f %8.3f %8.3f",
i_epoch,
train_loss, val_loss, test_loss,
train_acc, val_acc, test_acc)
with open("logs/val_predictions_%d.json" % i_epoch, "w") as pred_f:
print >>pred_f, json.dumps(val_predictions)
#with open("logs/test_predictions_%d.json" % i_epoch, "w") as pred_f:
# print >>pred_f, json.dumps(test_predictions)
def configure():
apollocaffe.set_random_seed(0)
np.random.seed(0)
random.seed(0)
arg_parser = argparse.ArgumentParser()
arg_parser.add_argument(
"-c", "--config", dest="config", required=True,
help="model configuration file")
arg_parser.add_argument(
"-l", "--log-config", dest="log_config", default="config/log.yml",
help="log configuration file")
args = arg_parser.parse_args()
config_name = args.config.split("/")[-1].split(".")[0]
with open(args.log_config) as log_config_f:
log_filename = "logs/%s.log" % config_name
log_config = yaml.load(log_config_f)
log_config["handlers"]["fileHandler"]["filename"] = log_filename
logging.config.dictConfig(log_config)
with open(args.config) as config_f:
config = util.Struct(**yaml.load(config_f))
assert not hasattr(config, "name")
config.name = config_name
return config
def do_iter(task_set, model, config, train=False, vis=False):
loss = 0.0
acc = 0.0
predictions = []
n_batches = 0
# sort first to guarantee deterministic behavior with fixed seed
data = list(sorted(task_set.data))
np.random.shuffle(data)
if vis:
visualizer.begin(config.name, 100)
for batch_start in range(0, len(data), config.opt.batch_size):
batch_end = batch_start + config.opt.batch_size
batch_data = data[batch_start:batch_end]
batch_loss, batch_acc, batch_preds = do_batch(
batch_data, model, config, train, vis)
loss += batch_loss
acc += batch_acc
predictions += batch_preds
n_batches += 1
if vis:
visualize(batch_data, model)
if vis:
visualizer.end()
if n_batches == 0:
return 0, 0, dict()
assert len(predictions) == len(data)
loss /= n_batches
acc /= n_batches
return loss, acc, predictions
def do_batch(data, model, config, train, vis):
predictions = forward(data, model, config, train, vis)
answer_loss = backward(data, model, config, train, vis)
acc = compute_acc(predictions, data, config)
return answer_loss, acc, predictions
# TODO this is ugly and belongs somewhere else
def featurize_layouts(datum, max_layouts):
# TODO pre-fill module type index
layout_reprs = np.zeros((max_layouts, len(MODULE_INDEX) + 7))
for i_layout in range(len(datum.layouts)):
layout = datum.layouts[i_layout]
labels = util.flatten(layout.labels)
modules = util.flatten(layout.modules)
for i_mod in range(len(modules)):
if isinstance(modules[i_mod], MLPFindModule) or isinstance(modules[i_mod], MultiplicativeFindModule):
layout_reprs[i_layout, labels[i_mod]] += 1
mt = MODULE_TYPE_INDEX.index(modules[i_mod])
layout_reprs[i_layout, len(MODULE_INDEX) + mt] += 1
return layout_reprs
def forward(data, model, config, train, vis):
model.reset()
# load batch data
max_len = max(len(d.question) for d in data)
max_layouts = max(len(d.layouts) for d in data)
channels, size, trailing = data[0].load_features().shape
assert trailing == 1
has_rel_features = data[0].load_rel_features() is not None
if has_rel_features:
rel_channels, size_1, size_2 = data[0].load_rel_features().shape
assert size_1 == size_2 == size
questions = np.ones((config.opt.batch_size, max_len)) * NULL_ID
features = np.zeros((config.opt.batch_size, channels, size, 1))
if has_rel_features:
rel_features = np.zeros((config.opt.batch_size, rel_channels, size, size))
else:
rel_features = None
layout_reprs = np.zeros(
(config.opt.batch_size, max_layouts, len(MODULE_INDEX) + 7))
for i, datum in enumerate(data):
questions[i, max_len-len(datum.question):] = datum.question
features[i, ...] = datum.load_features()
if has_rel_features:
rel_features[i, ...] = datum.load_rel_features()
layout_reprs[i, ...] = featurize_layouts(datum, max_layouts)
layouts = [d.layouts for d in data]
# apply model
model.forward(
layouts, layout_reprs, questions, features, rel_features,
dropout=(train and config.opt.dropout), deterministic=not train)
# extract predictions
if config.opt.multiclass:
pred_words = []
for i in range(model.prediction_data.shape[0]):
preds = model.prediction_data[i, :]
chosen = np.where(preds > 0.5)[0]
pred_words.append(set(ANSWER_INDEX.get(w) for w in chosen))
else:
pred_ids = np.argmax(model.prediction_data, axis=1)
pred_words = [ANSWER_INDEX.get(w) for w in pred_ids]
predictions = list()
for i in range(len(data)):
qid = data[i].id
answer = pred_words[i]
predictions.append({"question_id": qid, "answer": answer})
return predictions
def backward(data, model, config, train, vis):
n_answers = len(data[0].answers)
loss = 0
for i in range(n_answers):
if config.opt.multiclass:
output_i = np.zeros((config.opt.batch_size, len(ANSWER_INDEX)))
for i_datum, datum in enumerate(data):
for answer in datum.answers[i]:
output_i[i_datum, answer] = 1
else:
output_i = UNK_ID * np.ones(config.opt.batch_size)
output_i[:len(data)] = \
np.asarray([d.answers[i] for d in data])
loss += model.loss(output_i, multiclass=config.opt.multiclass)
if train:
model.train()
return loss
def visualize(batch_data, model):
i_datum = 0
#mod_layout_choice = model.module_layout_choices[i_datum]
#print model.apollo_net.blobs.keys()
#att_blob_name = "Find_%d_softmax" % (mod_layout_choice * 100 + 1)
#
datum = batch_data[i_datum]
question = " ".join([QUESTION_INDEX.get(w) for w in datum.question[1:-1]]),
preds = model.prediction_data[i_datum,:]
top = np.argsort(preds)[-5:]
top_answers = reversed([ANSWER_INDEX.get(p) for p in top])
#att_data = model.apollo_net.blobs[att_blob_name].data[i_datum,...]
#att_data = att_data.reshape((14, 14))
att_data = np.zeros((14, 14))
chosen_parse = datum.parses[model.layout_ids[i_datum]]
fields = [
question,
str(chosen_parse),
"<img src='../../%s'>" % datum.image_path,
att_data,
", ".join(top_answers),
", ".join([ANSWER_INDEX.get(a) for a in datum.answers])
]
visualizer.show(fields)
def compute_acc(predictions, data, config):
score = 0.0
for prediction, datum in zip(predictions, data):
pred_answer = prediction["answer"]
if config.opt.multiclass:
answers = [set(ANSWER_INDEX.get(aa) for aa in a) for a in datum.answers]
else:
answers = [ANSWER_INDEX.get(a) for a in datum.answers]
matching_answers = [a for a in answers if a == pred_answer]
if len(answers) == 1:
score += len(matching_answers)
else:
score += min(len(matching_answers) / 3.0, 1.0)
score /= len(data)
return score
if __name__ == "__main__":
main()