File size: 15,656 Bytes
d758c99 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 |
import ast
import collections
import collections.abc
import enum
import itertools
import json
import os
import operator
import re
import copy
import random
import asdl
import attr
import pyrsistent
import entmax
import torch
import torch.nn.functional as F
from seq2struct.utils import vocab
from seq2struct.models.nl2code import decoder
@attr.s
class TreeState:
node = attr.ib()
parent_field_type = attr.ib()
class TreeTraversal:
class Handler:
handlers = {}
@classmethod
def register_handler(cls, func_type):
if func_type in cls.handlers:
raise RuntimeError(f"{func_type} handler is already registered")
def inner_func(func):
cls.handlers[func_type] = func.__name__
return func
return inner_func
@attr.s(frozen=True)
class QueueItem:
item_id = attr.ib()
state = attr.ib()
node_type = attr.ib()
parent_action_emb = attr.ib()
parent_h = attr.ib()
parent_field_name = attr.ib()
def to_str(self):
return "<state: {}, node_type: {}, parent_field_name: {}>".format(
self.state, self.node_type, self.parent_field_name
)
class State(enum.Enum):
SUM_TYPE_INQUIRE = 0
SUM_TYPE_APPLY = 1
CHILDREN_INQUIRE = 2
CHILDREN_APPLY = 3
LIST_LENGTH_INQUIRE = 4
LIST_LENGTH_APPLY = 5
GEN_TOKEN = 6
POINTER_INQUIRE = 7
POINTER_APPLY = 8
NODE_FINISHED = 9
def __init__(self, model, desc_enc):
if model is None:
return
self.model = model
self.desc_enc = desc_enc
model.state_update.set_dropout_masks(batch_size=1)
self.recurrent_state = decoder.lstm_init(
model._device, None, self.model.recurrent_size, 1
)
self.prev_action_emb = model.zero_rule_emb
root_type = model.preproc.grammar.root_type
if root_type in model.preproc.ast_wrapper.sum_types:
initial_state = TreeTraversal.State.SUM_TYPE_INQUIRE
else:
initial_state = TreeTraversal.State.CHILDREN_INQUIRE
self.queue = pyrsistent.pvector()
self.cur_item = TreeTraversal.QueueItem(
item_id=0,
state=initial_state,
node_type=root_type,
parent_action_emb=self.model.zero_rule_emb,
parent_h=self.model.zero_recurrent_emb,
parent_field_name=None,
)
self.next_item_id = 1
self.update_prev_action_emb = TreeTraversal._update_prev_action_emb_apply_rule
def clone(self):
other = self.__class__(None, None)
other.model = self.model
other.desc_enc = self.desc_enc
other.recurrent_state = self.recurrent_state
other.prev_action_emb = self.prev_action_emb
other.queue = self.queue
other.cur_item = self.cur_item
other.next_item_id = self.next_item_id
other.actions = self.actions
other.update_prev_action_emb = self.update_prev_action_emb
return other
def step(self, last_choice, extra_choice_info=None, attention_offset=None):
while True:
self.update_using_last_choice(
last_choice, extra_choice_info, attention_offset
)
handler_name = TreeTraversal.Handler.handlers[self.cur_item.state]
handler = getattr(self, handler_name)
choices, continued = handler(last_choice)
if continued:
last_choice = choices
continue
else:
return choices
def update_using_last_choice(
self, last_choice, extra_choice_info, attention_offset
):
if last_choice is None:
return
if self.model.visualize_flag:
print("cur_item.state", self.cur_item.state)
if (
self.cur_item.state == TreeTraversal.State.SUM_TYPE_APPLY
or self.cur_item.state == TreeTraversal.State.CHILDREN_APPLY
or self.cur_item.state == TreeTraversal.State.LIST_LENGTH_APPLY
):
print("last choice", self.model.preproc.all_rules[last_choice])
else:
print("last choice", last_choice)
self.update_prev_action_emb(self, last_choice, extra_choice_info)
@classmethod
def _update_prev_action_emb_apply_rule(cls, self, last_choice, extra_choice_info):
rule_idx = self.model._tensor([last_choice])
self.prev_action_emb = self.model.rule_embedding(rule_idx)
@classmethod
def _update_prev_action_emb_gen_token(cls, self, last_choice, extra_choice_info):
# token_idx shape: batch (=1), LongTensor
token_idx = self.model._index(self.model.terminal_vocab, last_choice)
# action_emb shape: batch (=1) x emb_size
self.prev_action_emb = self.model.terminal_embedding(token_idx)
@classmethod
def _update_prev_action_emb_pointer(cls, self, last_choice, extra_choice_info):
# TODO batching
self.prev_action_emb = self.model.pointer_action_emb_proj[
self.cur_item.node_type
](self.desc_enc.pointer_memories[self.cur_item.node_type][:, last_choice])
def pop(self):
if self.queue:
self.cur_item = self.queue[-1]
self.queue = self.queue.delete(-1)
return True
return False
@Handler.register_handler(State.SUM_TYPE_INQUIRE)
def process_sum_inquire(self, last_choice):
# 1. ApplyRule, like expr -> Call
# a. Ask which one to choose
output, self.recurrent_state, rule_logits = self.model.apply_rule(
self.cur_item.node_type,
self.recurrent_state,
self.prev_action_emb,
self.cur_item.parent_h,
self.cur_item.parent_action_emb,
self.desc_enc,
)
self.cur_item = attr.evolve(
self.cur_item, state=TreeTraversal.State.SUM_TYPE_APPLY, parent_h=output
)
self.update_prev_action_emb = (
TreeTraversal._update_prev_action_emb_apply_rule
)
choices = self.rule_choice(self.cur_item.node_type, rule_logits)
return choices, False
@Handler.register_handler(State.SUM_TYPE_APPLY)
def process_sum_apply(self, last_choice):
# b. Add action, prepare for #2
sum_type, singular_type = self.model.preproc.all_rules[last_choice]
assert sum_type == self.cur_item.node_type
self.cur_item = attr.evolve(
self.cur_item,
node_type=singular_type,
parent_action_emb=self.prev_action_emb,
state=TreeTraversal.State.CHILDREN_INQUIRE,
)
return None, True
@Handler.register_handler(State.CHILDREN_INQUIRE)
def process_children_inquire(self, last_choice):
# 2. ApplyRule, like Call -> expr[func] expr*[args] keyword*[keywords]
# Check if we have no children
type_info = self.model.ast_wrapper.singular_types[
self.cur_item.node_type
]
if not type_info.fields:
if self.pop():
last_choice = None
return last_choice, True
else:
return None, False
# a. Ask about presence
output, self.recurrent_state, rule_logits = self.model.apply_rule(
self.cur_item.node_type,
self.recurrent_state,
self.prev_action_emb,
self.cur_item.parent_h,
self.cur_item.parent_action_emb,
self.desc_enc,
)
self.cur_item = attr.evolve(
self.cur_item, state=TreeTraversal.State.CHILDREN_APPLY, parent_h=output
)
self.update_prev_action_emb = (
TreeTraversal._update_prev_action_emb_apply_rule
)
choices = self.rule_choice(self.cur_item.node_type, rule_logits)
return choices, False
@Handler.register_handler(State.CHILDREN_APPLY)
def process_children_apply(self, last_choice):
# b. Create the children
node_type, children_presence = self.model.preproc.all_rules[last_choice]
assert node_type == self.cur_item.node_type
self.queue = self.queue.append(
TreeTraversal.QueueItem(
item_id=self.cur_item.item_id,
state=TreeTraversal.State.NODE_FINISHED,
node_type=None,
parent_action_emb=None,
parent_h=None,
parent_field_name=None,
)
)
for field_info, present in reversed(
list(
zip(
self.model.ast_wrapper.singular_types[node_type].fields,
children_presence,
)
)
):
if not present:
continue
# seq field: LIST_LENGTH_INQUIRE x
# sum type: SUM_TYPE_INQUIRE x
# product type:
# no children: not possible
# children: CHILDREN_INQUIRE
# constructor type: not possible x
# builtin type: GEN_TOKEN x
child_type = field_type = field_info.type
if field_info.seq:
child_state = TreeTraversal.State.LIST_LENGTH_INQUIRE
elif field_type in self.model.ast_wrapper.sum_types:
child_state = TreeTraversal.State.SUM_TYPE_INQUIRE
elif field_type in self.model.ast_wrapper.product_types:
assert self.model.ast_wrapper.product_types[field_type].fields
child_state = TreeTraversal.State.CHILDREN_INQUIRE
elif field_type in self.model.preproc.grammar.pointers:
child_state = TreeTraversal.State.POINTER_INQUIRE
elif field_type in self.model.ast_wrapper.primitive_types:
child_state = TreeTraversal.State.GEN_TOKEN
child_type = present
else:
raise ValueError(
"Unable to handle field type {}".format(field_type)
)
self.queue = self.queue.append(
TreeTraversal.QueueItem(
item_id=self.next_item_id,
state=child_state,
node_type=child_type,
parent_action_emb=self.prev_action_emb,
parent_h=self.cur_item.parent_h,
parent_field_name=field_info.name,
)
)
self.next_item_id += 1
advanced = self.pop()
assert advanced
last_choice = None
return last_choice, True
@Handler.register_handler(State.LIST_LENGTH_INQUIRE)
def process_list_length_inquire(self, last_choice):
list_type = self.cur_item.node_type + "*"
output, self.recurrent_state, rule_logits = self.model.apply_rule(
list_type,
self.recurrent_state,
self.prev_action_emb,
self.cur_item.parent_h,
self.cur_item.parent_action_emb,
self.desc_enc,
)
self.cur_item = attr.evolve(
self.cur_item, state=TreeTraversal.State.LIST_LENGTH_APPLY, parent_h=output
)
self.update_prev_action_emb = (
TreeTraversal._update_prev_action_emb_apply_rule
)
choices = self.rule_choice(list_type, rule_logits)
return choices, False
@Handler.register_handler(State.LIST_LENGTH_APPLY)
def process_list_length_apply(self, last_choice):
list_type, num_children = self.model.preproc.all_rules[last_choice]
elem_type = self.cur_item.node_type
assert list_type == elem_type + "*"
child_node_type = elem_type
if elem_type in self.model.ast_wrapper.sum_types:
child_state = TreeTraversal.State.SUM_TYPE_INQUIRE
if self.model.preproc.use_seq_elem_rules:
child_node_type = elem_type + "_seq_elem"
elif elem_type in self.model.ast_wrapper.product_types:
child_state = TreeTraversal.State.CHILDREN_INQUIRE
elif elem_type == "identifier":
child_state = TreeTraversal.State.GEN_TOKEN
child_node_type = "str"
elif elem_type in self.model.ast_wrapper.primitive_types:
# TODO: Fix this
raise ValueError("sequential builtin types not supported")
else:
raise ValueError(
"Unable to handle seq field type {}".format(elem_type)
)
for i in range(num_children):
self.queue = self.queue.append(
TreeTraversal.QueueItem(
item_id=self.next_item_id,
state=child_state,
node_type=child_node_type,
parent_action_emb=self.prev_action_emb,
parent_h=self.cur_item.parent_h,
parent_field_name=self.cur_item.parent_field_name,
)
)
self.next_item_id += 1
advanced = self.pop()
assert advanced
last_choice = None
return last_choice, True
@Handler.register_handler(State.GEN_TOKEN)
def process_gen_token(self, last_choice):
if last_choice == vocab.EOS:
if self.pop():
last_choice = None
return last_choice, True
else:
return None, False
self.recurrent_state, output, gen_logodds = self.model.gen_token(
self.cur_item.node_type,
self.recurrent_state,
self.prev_action_emb,
self.cur_item.parent_h,
self.cur_item.parent_action_emb,
self.desc_enc,
)
self.update_prev_action_emb = (
TreeTraversal._update_prev_action_emb_gen_token
)
choices = self.token_choice(output, gen_logodds)
return choices, False
@Handler.register_handler(State.POINTER_INQUIRE)
def process_pointer_inquire(self, last_choice):
# a. Ask which one to choose
output, self.recurrent_state, logits, attention_logits = self.model.compute_pointer_with_align(
self.cur_item.node_type,
self.recurrent_state,
self.prev_action_emb,
self.cur_item.parent_h,
self.cur_item.parent_action_emb,
self.desc_enc,
)
self.cur_item = attr.evolve(
self.cur_item, state=TreeTraversal.State.POINTER_APPLY, parent_h=output
)
self.update_prev_action_emb = (
TreeTraversal._update_prev_action_emb_pointer
)
choices = self.pointer_choice(
self.cur_item.node_type, logits, attention_logits
)
return choices, False
@Handler.register_handler(State.POINTER_APPLY)
def process_pointer_apply(self, last_choice):
if self.pop():
last_choice = None
return last_choice, True
else:
return None, False
@Handler.register_handler(State.NODE_FINISHED)
def process_node_finished(self, last_choice):
if self.pop():
last_choice = None
return last_choice, True
else:
return None, False
def rule_choice(self, node_type, rule_logits):
raise NotImplementedError
def token_choice(self, output, gen_logodds):
raise NotImplementedError
def pointer_choice(self, node_type, logits, attention_logits):
raise NotImplementedError
|