-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathprogram_declaration_STEPGAME.py
274 lines (225 loc) · 14.9 KB
/
program_declaration_STEPGAME.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
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
from domiknows.sensor.pytorch.sensors import ReaderSensor, ConcatSensor, FunctionalSensor, JointSensor
from domiknows.sensor.pytorch.learners import ModuleLearner, LSTMLearner
from models import *
from utils import *
from domiknows.sensor.pytorch.relation_sensors import CompositionCandidateSensor
def program_declaration_StepGame(device, *, pmd=False, beta=0.5, sampling=False, sampleSize=1, dropout=False,
constraints=False, spartun=True):
program = None
from graph_stepgame import graph, story, story_contain, question, \
left, right, above, below, lower_left, lower_right, upper_left, upper_right, overlap
story["questions"] = ReaderSensor(keyword="questions")
story["stories"] = ReaderSensor(keyword="stories")
story["relations"] = ReaderSensor(keyword="relation")
story["question_ids"] = ReaderSensor(keyword="question_ids")
story["labels"] = ReaderSensor(keyword="labels")
all_labels = ["left", "right", "above", "below", "lower-left",
"lower-right", "upper-left", "upper-right", "overlap"]
def to_int_list(x):
return torch.LongTensor([int(i) for i in x])
def make_labels(label_list):
labels = label_list.split("@@")
all_labels_list = [[] for _ in range(9)]
for cur_label in labels:
for ind, label in enumerate(all_labels):
all_labels_list[ind].append(1 if label == cur_label else 0)
# label_nums = [0 if label == "Yes" else 1 if label == "No" else 2 for label in labels]
return [to_int_list(labels_list) for labels_list in all_labels_list]
def make_question(questions, stories, relations, q_ids, labels):
all_labels = make_labels(labels)
ids = to_int_list(q_ids.split("@@"))
left_list, right_list, above_list, below_list, lower_left_list, \
lower_right_list, upper_left_list, upper_right_list, over_lap_list = all_labels
return torch.ones(len(questions.split("@@")), 1), questions.split("@@"), stories.split("@@"), \
relations.split("@@"), ids, left_list, right_list, above_list, below_list, lower_left_list, \
lower_right_list, upper_left_list, upper_right_list, over_lap_list
question[story_contain, "question", "story", "relation", "id", "left_label", "right_label",
"above_label", "below_label", "lower_left_label", "lower_right_label", "upper_left_label", "upper_right_label", "overlap_label"] = \
JointSensor(story["questions"], story["stories"], story["relations"],
story["question_ids"], story["labels"], forward=make_question, device=device)
def read_label(_, label):
return label
# question[answer_class] =
# FunctionalSensor(story_contain, "label", forward=read_label, label=True, device=cur_device)
# Replace with all classes
question["input_ids"] = JointSensor(story_contain, 'question', "story",
forward=BERTTokenizer(), device=device)
clf1 = MultipleClassYN_Hidden.from_pretrained('bert-base-uncased', device=device, drp=dropout)
question["hidden_layer"] = ModuleLearner("input_ids", module=clf1, device=device)
question[left] = ModuleLearner("hidden_layer",
module=ClassifyLayer(clf1.hidden_size, device=device, drp=dropout),
device=device)
question[left] = FunctionalSensor(story_contain, "left_label", forward=read_label, label=True, device=device)
question[right] = ModuleLearner("hidden_layer",
module=ClassifyLayer(clf1.hidden_size, device=device, drp=dropout),
device=device)
question[right] = FunctionalSensor(story_contain, "right_label", forward=read_label, label=True, device=device)
question[above] = ModuleLearner("hidden_layer",
module=ClassifyLayer(clf1.hidden_size, device=device, drp=dropout),
device=device)
question[above] = FunctionalSensor(story_contain, "above_label", forward=read_label, label=True, device=device)
question[below] = ModuleLearner("hidden_layer",
module=ClassifyLayer(clf1.hidden_size, device=device, drp=dropout),
device=device)
question[below] = FunctionalSensor(story_contain, "below_label", forward=read_label, label=True, device=device)
question[lower_left] = ModuleLearner("hidden_layer",
module=ClassifyLayer(clf1.hidden_size, device=device, drp=dropout),
device=device)
question[lower_left] = FunctionalSensor(story_contain, "lower_left_label", forward=read_label, label=True,
device=device)
question[lower_right] = ModuleLearner("hidden_layer",
module=ClassifyLayer(clf1.hidden_size, device=device, drp=dropout),
device=device)
question[lower_right] = FunctionalSensor(story_contain, "lower_right_label", forward=read_label, label=True,
device=device)
question[upper_left] = ModuleLearner("hidden_layer",
module=ClassifyLayer(clf1.hidden_size, device=device, drp=dropout),
device=device)
question[upper_left] = FunctionalSensor(story_contain, "upper_left_label", forward=read_label, label=True,
device=device)
question[upper_right] = ModuleLearner("hidden_layer",
module=ClassifyLayer(clf1.hidden_size, device=device, drp=dropout),
device=device)
question[upper_right] = FunctionalSensor(story_contain, "upper_right_label", forward=read_label, label=True,
device=device)
question[overlap] = ModuleLearner("hidden_layer",
module=ClassifyLayer(clf1.hidden_size, device=device, drp=dropout),
device=device)
question[overlap] = FunctionalSensor(story_contain, "overlap_label", forward=read_label, label=True, device=device)
poi_list = [question, left, right, above, below, lower_left, lower_right, upper_left, upper_right,
overlap]
from domiknows.program.metric import PRF1Tracker, PRF1Tracker, DatanodeCMMetric, MacroAverageTracker, ValueTracker
from domiknows.program.loss import NBCrossEntropyLoss, BCEWithLogitsIMLoss, BCEFocalLoss
from domiknows.program import LearningBasedProgram, SolverPOIProgram
from domiknows.program.lossprogram import SampleLossProgram, PrimalDualProgram
from domiknows.program.model.pytorch import model_helper, PoiModel, SolverModel
infer_list = ['local/argmax'] # ['ILP', 'local/argmax']
if pmd:
program = PrimalDualProgram(graph, SolverModel, poi=poi_list,
inferTypes=infer_list,
loss=MacroAverageTracker(NBCrossEntropyLoss()),
beta=beta,
metric={'argmax': PRF1Tracker(DatanodeCMMetric('local/argmax'))},
device=device)
elif sampling:
program = SampleLossProgram(graph, SolverModel, poi=poi_list,
inferTypes=infer_list,
loss=MacroAverageTracker(NBCrossEntropyLoss()),
metric={'argmax': PRF1Tracker(DatanodeCMMetric('local/argmax'))},
sample=True,
sampleSize=sampleSize,
sampleGlobalLoss=False,
beta=1,
device=device)
else:
program = SolverPOIProgram(graph,
poi=poi_list,
inferTypes=infer_list,
loss=MacroAverageTracker(NBCrossEntropyLoss()),
metric={'argmax': PRF1Tracker(DatanodeCMMetric('local/argmax'))},
device=device)
return program
def program_declaration_StepGame_T5(device, *, pmd=False, beta=0.5, sampling=False, sampleSize=1, dropout=False,
constraints=False, spartun=True):
from graph_stepgame import graph, story, story_contain, question, \
left, right, above, below, lower_left, lower_right, upper_left, upper_right, overlap, output_for_loss
story["questions"] = ReaderSensor(keyword="questions")
story["stories"] = ReaderSensor(keyword="stories")
story["relations"] = ReaderSensor(keyword="relation")
story["question_ids"] = ReaderSensor(keyword="question_ids")
story["labels"] = ReaderSensor(keyword="labels")
all_labels = ["left", "right", "above", "below", "lower-left",
"lower-right", "upper-left", "upper-right", "overlap"]
map_label_index = {text: i for i, text in enumerate(all_labels)}
def to_int_list(x):
return torch.LongTensor([int(i) for i in x])
def to_float_list(x):
return torch.Tensor([float(i) for i in x])
def make_labels(label_list):
labels = label_list.split("@@")
# label_nums = [0 if label == "Yes" else 1 if label == "No" else 2 for label in labels]
return labels
def make_question(questions, stories, relations, q_ids, labels):
text_label = make_labels(labels)
ids = to_int_list(q_ids.split("@@"))
return torch.ones(len(questions.split("@@")), 1), questions.split("@@"), stories.split("@@"), relations.split(
"@@"), ids, text_label
question[story_contain, "question", "story", "relation", "id", "text_labels"] = \
JointSensor(story["questions"], story["stories"], story["relations"],
story["question_ids"], story["labels"], forward=make_question, device=device)
T5_model = T5WithLora("google/flan-t5-base", device=device, adapter=True)
# defined loss based on the model
LossT5 = T5LossFunction(T5_model=T5_model)
t5_outTokenizer = T5TokenizerOutput('google/flan-t5-base')
t5_inTokenizer = T5TokenizerInput('google/flan-t5-base')
question[output_for_loss] = JointSensor(story_contain, 'question', "story",
forward=t5_inTokenizer, device=device)
question["input_ids"] = JointSensor(story_contain, 'question', "story", True,
forward=t5_inTokenizer, device=device)
question[output_for_loss] = FunctionalSensor(story_contain,
'text_labels',
forward=t5_outTokenizer,
label=True,
device=device)
all_answers = [left, right, above, below, lower_left, lower_right, upper_left, upper_right, overlap]
question["output_encoder"] = ModuleLearner(story_contain, "input_ids", module=T5_model, device=device)
question["output_decoder"] = FunctionalSensor(story_contain, "output_encoder",
forward=T5TokenizerDecoder('google/flan-t5-base'), device=device)
def read_decoder(_, decoder_list):
text_label = [[0] * 15 for _ in range(len(decoder_list))]
for ind, text_decode in enumerate(decoder_list):
text_decode = text_decode.replace("and", "")
all_relations = text_decode.strip().split(", ")
for relation in all_relations:
relation = relation.strip()
if relation not in map_label_index:
continue
text_label[ind][map_label_index[relation]] = 1
list_tensor = [to_float_list(labels_list) for labels_list in text_label]
return torch.stack(list_tensor)
def read_label(_, relation_list, index):
label = relation_list[:, index].reshape((-1, 1))
label = torch.concat((torch.ones_like(label) - label, label), dim=-1)
return label
question["output_relations"] = FunctionalSensor(story_contain, "output_decoder", forward=read_decoder,
device=device)
question[left] = FunctionalSensor(story_contain, "output_relations", 0, forward=read_label, device=device)
question[right] = FunctionalSensor(story_contain, "output_relations", 1, forward=read_label, device=device)
question[above] = FunctionalSensor(story_contain, "output_relations", 2, forward=read_label, device=device)
question[below] = FunctionalSensor(story_contain, "output_relations", 3, forward=read_label, device=device)
question[lower_left] = FunctionalSensor(story_contain, "output_relations", 4, forward=read_label, device=device)
question[lower_right] = FunctionalSensor(story_contain, "output_relations", 5, forward=read_label, device=device)
question[upper_left] = FunctionalSensor(story_contain, "output_relations", 6, forward=read_label, device=device)
question[upper_right] = FunctionalSensor(story_contain, "output_relations", 7, forward=read_label, device=device)
question[overlap] = FunctionalSensor(story_contain, "output_relations", 8, forward=read_label, device=device)
poi_list = [question, left, right, above, below, lower_left,
lower_right, upper_left, upper_right, overlap, output_for_loss]
from domiknows.program.metric import PRF1Tracker, DatanodeCMMetric, ValueTracker
from domiknows.program import SolverPOIProgram
from domiknows.program.lossprogram import SampleLossProgram, PrimalDualProgram
from domiknows.program.model.pytorch import SolverModel
infer_list = ['local/argmax'] # ['ILP', 'local/argmax']
if pmd:
print("Using PMD program")
program = PrimalDualProgram(graph, SolverModel, poi=poi_list,
inferTypes=infer_list,
loss=ValueTracker(LossT5),
beta=beta,
device=device)
elif sampling:
program = SampleLossProgram(graph, SolverModel, poi=poi_list,
inferTypes=infer_list,
loss=ValueTracker(LossT5),
sample=True,
sampleSize=sampleSize,
sampleGlobalLoss=False,
beta=1,
device=device)
else:
print("Using Base program")
program = SolverPOIProgram(graph,
poi=poi_list,
inferTypes=infer_list,
loss=ValueTracker(LossT5),
device=device)
return program