Spaces:
Runtime error
Runtime error
Upload preprocess.py
Browse files- preprocess.py +551 -0
preprocess.py
ADDED
|
@@ -0,0 +1,551 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import time
|
| 2 |
+
import os
|
| 3 |
+
import random
|
| 4 |
+
import json
|
| 5 |
+
import pickle
|
| 6 |
+
import numpy as np
|
| 7 |
+
from tqdm import tqdm
|
| 8 |
+
from termcolor import colored
|
| 9 |
+
from program_translator import ProgramTranslator #
|
| 10 |
+
from config import config
|
| 11 |
+
|
| 12 |
+
|
| 13 |
+
# Print bold tex
|
| 14 |
+
def bold(txt):
|
| 15 |
+
return colored(str(txt), attrs=["bold"])
|
| 16 |
+
|
| 17 |
+
|
| 18 |
+
# Print bold and colored text
|
| 19 |
+
def bcolored(txt, color):
|
| 20 |
+
return colored(str(txt), color, attrs=["bold"])
|
| 21 |
+
|
| 22 |
+
|
| 23 |
+
# Write a line to file
|
| 24 |
+
def writeline(f, line):
|
| 25 |
+
f.write(str(line) + "\n")
|
| 26 |
+
|
| 27 |
+
|
| 28 |
+
# Write a list to file
|
| 29 |
+
def writelist(f, l):
|
| 30 |
+
writeline(f, ",".join(map(str, l)))
|
| 31 |
+
|
| 32 |
+
|
| 33 |
+
# 2d list to numpy
|
| 34 |
+
def vectorize2DList(items, minX=0, minY=0, dtype=np.int):
|
| 35 |
+
maxX = max(len(items), minX)
|
| 36 |
+
maxY = max([len(item) for item in items] + [minY])
|
| 37 |
+
t = np.zeros((maxX, maxY), dtype=dtype)
|
| 38 |
+
tLengths = np.zeros((maxX,), dtype=np.int)
|
| 39 |
+
for i, item in enumerate(items):
|
| 40 |
+
t[i, 0:len(item)] = np.array(item, dtype=dtype)
|
| 41 |
+
tLengths[i] = len(item)
|
| 42 |
+
return t, tLengths
|
| 43 |
+
|
| 44 |
+
|
| 45 |
+
# 3d list to numpy
|
| 46 |
+
def vectorize3DList(items, minX=0, minY=0, minZ=0, dtype=np.int):
|
| 47 |
+
maxX = max(len(items), minX)
|
| 48 |
+
maxY = max([len(item) for item in items] + [minY])
|
| 49 |
+
maxZ = max([len(subitem) for item in items for subitem in item] + [minZ])
|
| 50 |
+
t = np.zeros((maxX, maxY, maxZ), dtype=dtype)
|
| 51 |
+
tLengths = np.zeros((maxX, maxY), dtype=np.int)
|
| 52 |
+
for i, item in enumerate(items):
|
| 53 |
+
for j, subitem in enumerate(item):
|
| 54 |
+
t[i, j, 0:len(subitem)] = np.array(subitem, dtype=dtype)
|
| 55 |
+
tLengths[i, j] = len(subitem)
|
| 56 |
+
return t, tLengths
|
| 57 |
+
|
| 58 |
+
|
| 59 |
+
'''
|
| 60 |
+
Encodes text into integers. Keeps dictionary between string words (symbols)
|
| 61 |
+
and their matching integers. Supports encoding and decoding.
|
| 62 |
+
'''
|
| 63 |
+
|
| 64 |
+
|
| 65 |
+
class SymbolDict(object):
|
| 66 |
+
def __init__(self, empty=False):
|
| 67 |
+
self.padding = "<PAD>"
|
| 68 |
+
self.unknown = "<UNK>"
|
| 69 |
+
self.start = "<START>"
|
| 70 |
+
self.end = "<END>"
|
| 71 |
+
|
| 72 |
+
self.invalidSymbols = [self.padding, self.unknown, self.start, self.end]
|
| 73 |
+
|
| 74 |
+
if empty:
|
| 75 |
+
self.sym2id = {}
|
| 76 |
+
self.id2sym = []
|
| 77 |
+
else:
|
| 78 |
+
self.sym2id = {self.padding: 0, self.unknown: 1, self.start: 2, self.end: 3}
|
| 79 |
+
self.id2sym = [self.padding, self.unknown, self.start, self.end]
|
| 80 |
+
self.allSeqs = []
|
| 81 |
+
|
| 82 |
+
def getNumSymbols(self):
|
| 83 |
+
return len(self.sym2id)
|
| 84 |
+
|
| 85 |
+
def isPadding(self, enc):
|
| 86 |
+
return enc == 0
|
| 87 |
+
|
| 88 |
+
def isUnknown(self, enc):
|
| 89 |
+
return enc == 1
|
| 90 |
+
|
| 91 |
+
def isStart(self, enc):
|
| 92 |
+
return enc == 2
|
| 93 |
+
|
| 94 |
+
def isEnd(self, enc):
|
| 95 |
+
return enc == 3
|
| 96 |
+
|
| 97 |
+
def isValid(self, enc):
|
| 98 |
+
return enc < self.getNumSymbols() and enc >= len(self.invalidSymbols)
|
| 99 |
+
|
| 100 |
+
def resetSeqs(self):
|
| 101 |
+
self.allSeqs = []
|
| 102 |
+
|
| 103 |
+
def addSeq(self, seq):
|
| 104 |
+
self.allSeqs += seq
|
| 105 |
+
|
| 106 |
+
# Call to create the words-to-integers vocabulary after (reading word sequences with addSeq).
|
| 107 |
+
def createVocab(self, minCount=0):
|
| 108 |
+
counter = {}
|
| 109 |
+
for symbol in self.allSeqs:
|
| 110 |
+
counter[symbol] = counter.get(symbol, 0) + 1
|
| 111 |
+
for symbol in counter:
|
| 112 |
+
if counter[symbol] > minCount and (symbol not in self.sym2id):
|
| 113 |
+
self.sym2id[symbol] = self.getNumSymbols()
|
| 114 |
+
self.id2sym.append(symbol)
|
| 115 |
+
|
| 116 |
+
# Encodes a symbol. Returns the matching integer.
|
| 117 |
+
def encodeSym(self, symbol):
|
| 118 |
+
if symbol not in self.sym2id:
|
| 119 |
+
symbol = self.unknown
|
| 120 |
+
return self.sym2id[symbol]
|
| 121 |
+
|
| 122 |
+
'''
|
| 123 |
+
Encodes a sequence of symbols.
|
| 124 |
+
Optionally add start, or end symbols.
|
| 125 |
+
Optionally reverse sequence
|
| 126 |
+
'''
|
| 127 |
+
|
| 128 |
+
def encodeSequence(self, decoded, addStart=False, addEnd=False, reverse=False):
|
| 129 |
+
if reverse:
|
| 130 |
+
decoded.reverse()
|
| 131 |
+
if addStart:
|
| 132 |
+
decoded = [self.start] + decoded
|
| 133 |
+
if addEnd:
|
| 134 |
+
decoded = decoded + [self.end]
|
| 135 |
+
encoded = [self.encodeSym(symbol) for symbol in decoded]
|
| 136 |
+
return encoded
|
| 137 |
+
|
| 138 |
+
# Decodes an integer into its symbol
|
| 139 |
+
def decodeId(self, enc):
|
| 140 |
+
return self.id2sym[enc] if enc < self.getNumSymbols() else self.unknown
|
| 141 |
+
|
| 142 |
+
'''
|
| 143 |
+
Decodes a sequence of integers into their symbols.
|
| 144 |
+
If delim is given, joins the symbols using delim,
|
| 145 |
+
Optionally reverse the resulted sequence
|
| 146 |
+
'''
|
| 147 |
+
|
| 148 |
+
def decodeSequence(self, encoded, delim=None, reverse=False, stopAtInvalid=True):
|
| 149 |
+
length = 0
|
| 150 |
+
for i in range(len(encoded)):
|
| 151 |
+
if not self.isValid(encoded[i]) and stopAtInvalid:
|
| 152 |
+
break
|
| 153 |
+
length += 1
|
| 154 |
+
encoded = encoded[:length]
|
| 155 |
+
|
| 156 |
+
decoded = [self.decodeId(enc) for enc in encoded]
|
| 157 |
+
if reverse:
|
| 158 |
+
decoded.reverse()
|
| 159 |
+
|
| 160 |
+
if delim is not None:
|
| 161 |
+
return delim.join(decoded)
|
| 162 |
+
|
| 163 |
+
return decoded
|
| 164 |
+
|
| 165 |
+
|
| 166 |
+
'''
|
| 167 |
+
Preprocesses a given dataset into numpy arrays.
|
| 168 |
+
By calling preprocess, the class:
|
| 169 |
+
1. Reads the input data files into dictionary.
|
| 170 |
+
2. Saves the results jsons in files and loads them instead of parsing input if files exist/
|
| 171 |
+
3. Initializes word embeddings to random / GloVe.
|
| 172 |
+
4. Optionally filters data according to given filters.
|
| 173 |
+
5. Encodes and vectorize the data into numpy arrays.
|
| 174 |
+
6. Buckets the data according to the instances length.
|
| 175 |
+
'''
|
| 176 |
+
|
| 177 |
+
|
| 178 |
+
class Preprocesser(object):
|
| 179 |
+
def __init__(self):
|
| 180 |
+
self.questionDict = SymbolDict()
|
| 181 |
+
self.answerDict = SymbolDict(empty=True)
|
| 182 |
+
self.qaDict = SymbolDict()
|
| 183 |
+
|
| 184 |
+
self.specificDatasetDicts = None
|
| 185 |
+
|
| 186 |
+
self.programDict = SymbolDict()
|
| 187 |
+
self.programTranslator = ProgramTranslator(self.programDict, 2)
|
| 188 |
+
|
| 189 |
+
'''
|
| 190 |
+
Tokenizes string into list of symbols.
|
| 191 |
+
|
| 192 |
+
Args:
|
| 193 |
+
text: raw string to tokenize.
|
| 194 |
+
ignorePuncts: punctuation to ignore
|
| 195 |
+
keptPunct: punctuation to keep (as symbol)
|
| 196 |
+
endPunct: punctuation to remove if appears at the end
|
| 197 |
+
delim: delimiter between symbols
|
| 198 |
+
clean: True to replace text in string
|
| 199 |
+
replacelistPre: dictionary of replacement to perform on the text before tokanization
|
| 200 |
+
replacelistPost: dictionary of replacement to perform on the text after tokanization
|
| 201 |
+
'''
|
| 202 |
+
# sentence tokenizer
|
| 203 |
+
allPunct = ["?", "!", "\\", "/", ")", "(", ".", ",", ";", ":"]
|
| 204 |
+
|
| 205 |
+
def tokenize(self, text, ignoredPuncts=["?", "!", "\\", "/", ")", "("],
|
| 206 |
+
keptPuncts=[".", ",", ";", ":"], endPunct=[">", "<", ":"], delim=" ",
|
| 207 |
+
clean=False, replacelistPre=dict(), replacelistPost=dict()):
|
| 208 |
+
|
| 209 |
+
if clean:
|
| 210 |
+
for word in replacelistPre:
|
| 211 |
+
origText = text
|
| 212 |
+
text = text.replace(word, replacelistPre[word])
|
| 213 |
+
if (origText != text):
|
| 214 |
+
print(origText)
|
| 215 |
+
print(text)
|
| 216 |
+
print("")
|
| 217 |
+
|
| 218 |
+
for punct in endPunct:
|
| 219 |
+
if text[-1] == punct:
|
| 220 |
+
print(text)
|
| 221 |
+
text = text[:-1]
|
| 222 |
+
print(text)
|
| 223 |
+
print("")
|
| 224 |
+
|
| 225 |
+
for punct in keptPuncts:
|
| 226 |
+
text = text.replace(punct, delim + punct + delim)
|
| 227 |
+
|
| 228 |
+
for punct in ignoredPuncts:
|
| 229 |
+
text = text.replace(punct, "")
|
| 230 |
+
|
| 231 |
+
ret = text.lower().split(delim)
|
| 232 |
+
|
| 233 |
+
if clean:
|
| 234 |
+
origRet = ret
|
| 235 |
+
ret = [replacelistPost.get(word, word) for word in ret]
|
| 236 |
+
if origRet != ret:
|
| 237 |
+
print(origRet)
|
| 238 |
+
print(ret)
|
| 239 |
+
|
| 240 |
+
ret = [t for t in ret if t != ""]
|
| 241 |
+
return ret
|
| 242 |
+
|
| 243 |
+
# Read class' generated files.
|
| 244 |
+
# files interface
|
| 245 |
+
def readFiles(self, instancesFilename):
|
| 246 |
+
with open(instancesFilename, "r") as inFile:
|
| 247 |
+
instances = json.load(inFile)
|
| 248 |
+
|
| 249 |
+
with open(config.questionDictFile(), "rb") as inFile:
|
| 250 |
+
self.questionDict = pickle.load(inFile)
|
| 251 |
+
|
| 252 |
+
with open(config.answerDictFile(), "rb") as inFile:
|
| 253 |
+
self.answerDict = pickle.load(inFile)
|
| 254 |
+
|
| 255 |
+
with open(config.qaDictFile(), "rb") as inFile:
|
| 256 |
+
self.qaDict = pickle.load(inFile)
|
| 257 |
+
|
| 258 |
+
return instances
|
| 259 |
+
|
| 260 |
+
'''
|
| 261 |
+
Generate class' files. Save json representation of instances and
|
| 262 |
+
symbols-to-integers dictionaries.
|
| 263 |
+
'''
|
| 264 |
+
|
| 265 |
+
def writeFiles(self, instances, instancesFilename):
|
| 266 |
+
with open(instancesFilename, "w") as outFile:
|
| 267 |
+
json.dump(instances, outFile)
|
| 268 |
+
|
| 269 |
+
with open(config.questionDictFile(), "wb") as outFile:
|
| 270 |
+
pickle.dump(self.questionDict, outFile)
|
| 271 |
+
|
| 272 |
+
with open(config.answerDictFile(), "wb") as outFile:
|
| 273 |
+
pickle.dump(self.answerDict, outFile)
|
| 274 |
+
|
| 275 |
+
with open(config.qaDictFile(), "wb") as outFile:
|
| 276 |
+
pickle.dump(self.qaDict, outFile)
|
| 277 |
+
|
| 278 |
+
# Write prediction json to file and optionally a one-answer-per-line output file
|
| 279 |
+
def writePreds(self, res, tier, suffix=""):
|
| 280 |
+
if res is None:
|
| 281 |
+
return
|
| 282 |
+
preds = res["preds"]
|
| 283 |
+
sortedPreds = sorted(preds, key=lambda instance: instance["index"])
|
| 284 |
+
with open(config.predsFile(tier + suffix), "w") as outFile:
|
| 285 |
+
outFile.write(json.dumps(sortedPreds))
|
| 286 |
+
with open(config.answersFile(tier + suffix), "w") as outFile:
|
| 287 |
+
for instance in sortedPreds:
|
| 288 |
+
writeline(outFile, instance["prediction"])
|
| 289 |
+
|
| 290 |
+
def readPDF(self, instancesFilename):
|
| 291 |
+
instances = []
|
| 292 |
+
|
| 293 |
+
if os.path.exists(instancesFilename):
|
| 294 |
+
instances = self.readFiles(instancesFilename)
|
| 295 |
+
|
| 296 |
+
return instances
|
| 297 |
+
|
| 298 |
+
def readData(self, datasetFilename, instancesFilename, train):
|
| 299 |
+
# data extraction
|
| 300 |
+
datasetReader = {
|
| 301 |
+
"PDF": self.readPDF
|
| 302 |
+
}
|
| 303 |
+
|
| 304 |
+
return datasetReader[config.dataset](datasetFilename, instancesFilename, train)
|
| 305 |
+
|
| 306 |
+
def vectorizeData(self, data):
|
| 307 |
+
# if "SHARED" tie symbol representations in questions and answers
|
| 308 |
+
if config.ansEmbMod == "SHARED":
|
| 309 |
+
qDict = self.qaDict
|
| 310 |
+
else:
|
| 311 |
+
qDict = self.questionDict
|
| 312 |
+
|
| 313 |
+
encodedQuestion = [qDict.encodeSequence(d["questionSeq"]) for d in data]
|
| 314 |
+
question, questionL = vectorize2DList(encodedQuestion)
|
| 315 |
+
|
| 316 |
+
# pass the whole instances? if heavy then not good
|
| 317 |
+
imageId = [d["imageId"] for d in data]
|
| 318 |
+
instance = data
|
| 319 |
+
|
| 320 |
+
return {"question": question,
|
| 321 |
+
"questionLength": questionL,
|
| 322 |
+
"imageId": imageId
|
| 323 |
+
}
|
| 324 |
+
|
| 325 |
+
# Separates data based on a field length
|
| 326 |
+
def lseparator(self, key, lims):
|
| 327 |
+
maxI = len(lims)
|
| 328 |
+
|
| 329 |
+
def separatorFn(x):
|
| 330 |
+
v = x[key]
|
| 331 |
+
for i, lim in enumerate(lims):
|
| 332 |
+
if len(v) < lim:
|
| 333 |
+
return i
|
| 334 |
+
return maxI
|
| 335 |
+
|
| 336 |
+
return {"separate": separatorFn, "groupsNum": maxI + 1}
|
| 337 |
+
|
| 338 |
+
# Buckets data to groups using a separator
|
| 339 |
+
def bucket(self, instances, separator):
|
| 340 |
+
buckets = [[] for i in range(separator["groupsNum"])]
|
| 341 |
+
for instance in instances:
|
| 342 |
+
bucketI = separator["separate"](instance)
|
| 343 |
+
buckets[bucketI].append(instance)
|
| 344 |
+
return [bucket for bucket in buckets if len(bucket) > 0]
|
| 345 |
+
|
| 346 |
+
# Re-buckets bucket list given a seperator
|
| 347 |
+
def rebucket(self, buckets, separator):
|
| 348 |
+
res = []
|
| 349 |
+
for bucket in buckets:
|
| 350 |
+
res += self.bucket(bucket, separator)
|
| 351 |
+
return res
|
| 352 |
+
|
| 353 |
+
# Buckets data based on question / program length
|
| 354 |
+
def bucketData(self, data, noBucket=False):
|
| 355 |
+
if noBucket:
|
| 356 |
+
buckets = [data]
|
| 357 |
+
else:
|
| 358 |
+
if config.noBucket:
|
| 359 |
+
buckets = [data]
|
| 360 |
+
elif config.noRebucket:
|
| 361 |
+
questionSep = self.lseparator("questionSeq", config.questionLims)
|
| 362 |
+
buckets = self.bucket(data, questionSep)
|
| 363 |
+
else:
|
| 364 |
+
programSep = self.lseparator("programSeq", config.programLims)
|
| 365 |
+
questionSep = self.lseparator("questionSeq", config.questionLims)
|
| 366 |
+
buckets = self.bucket(data, programSep)
|
| 367 |
+
buckets = self.rebucket(buckets, questionSep)
|
| 368 |
+
return buckets
|
| 369 |
+
|
| 370 |
+
'''
|
| 371 |
+
Prepares data:
|
| 372 |
+
1. Filters data according to above arguments.
|
| 373 |
+
2. Takes only a subset of the data based on config.trainedNum / config.testedNum
|
| 374 |
+
3. Buckets data according to question / program length
|
| 375 |
+
4. Vectorizes data into numpy arrays
|
| 376 |
+
'''
|
| 377 |
+
|
| 378 |
+
def prepareData(self, data, train, filterKey=None, noBucket=False):
|
| 379 |
+
filterDefault = {"maxQLength": 0, "maxPLength": 0, "onlyChain": False, "filterOp": 0}
|
| 380 |
+
|
| 381 |
+
filterTrain = {"maxQLength": config.tMaxQ, "maxPLength": config.tMaxP,
|
| 382 |
+
"onlyChain": config.tOnlyChain, "filterOp": config.tFilterOp}
|
| 383 |
+
|
| 384 |
+
filterVal = {"maxQLength": config.vMaxQ, "maxPLength": config.vMaxP,
|
| 385 |
+
"onlyChain": config.vOnlyChain, "filterOp": config.vFilterOp}
|
| 386 |
+
|
| 387 |
+
filters = {"train": filterTrain, "evalTrain": filterTrain,
|
| 388 |
+
"val": filterVal, "test": filterDefault}
|
| 389 |
+
|
| 390 |
+
if filterKey is None:
|
| 391 |
+
fltr = filterDefault
|
| 392 |
+
else:
|
| 393 |
+
fltr = filters[filterKey]
|
| 394 |
+
|
| 395 |
+
# split data when finetuning on validation set
|
| 396 |
+
if config.trainExtra and config.extraVal and (config.finetuneNum > 0):
|
| 397 |
+
if train:
|
| 398 |
+
data = data[:config.finetuneNum]
|
| 399 |
+
else:
|
| 400 |
+
data = data[config.finetuneNum:]
|
| 401 |
+
|
| 402 |
+
typeFilter = config.typeFilters[fltr["filterOp"]]
|
| 403 |
+
# filter specific settings
|
| 404 |
+
if fltr["onlyChain"]:
|
| 405 |
+
data = [d for d in data if all((len(inputNum) < 2) for inputNum in d["programInputs"])]
|
| 406 |
+
if fltr["maxQLength"] > 0:
|
| 407 |
+
data = [d for d in data if len(d["questionSeq"]) <= fltr["maxQLength"]]
|
| 408 |
+
if fltr["maxPLength"] > 0:
|
| 409 |
+
data = [d for d in data if len(d["programSeq"]) <= fltr["maxPLength"]]
|
| 410 |
+
if len(typeFilter) > 0:
|
| 411 |
+
data = [d for d in data if d["programSeq"][-1] not in typeFilter]
|
| 412 |
+
|
| 413 |
+
# run on subset of the data. If 0 then use all data
|
| 414 |
+
num = config.trainedNum if train else config.testedNum
|
| 415 |
+
# retainVal = True to retain same clevr_sample of validation across runs
|
| 416 |
+
if (not train) and (not config.retainVal):
|
| 417 |
+
random.shuffle(data)
|
| 418 |
+
if num > 0:
|
| 419 |
+
data = data[:num]
|
| 420 |
+
# set number to match dataset size
|
| 421 |
+
if train:
|
| 422 |
+
config.trainedNum = len(data)
|
| 423 |
+
else:
|
| 424 |
+
config.testedNum = len(data)
|
| 425 |
+
|
| 426 |
+
# bucket
|
| 427 |
+
buckets = self.bucketData(data, noBucket=noBucket)
|
| 428 |
+
|
| 429 |
+
# vectorize
|
| 430 |
+
return [self.vectorizeData(bucket) for bucket in buckets]
|
| 431 |
+
|
| 432 |
+
# Prepares all the tiers of a dataset. See prepareData method for further details.
|
| 433 |
+
def prepareDataset(self, dataset, noBucket=False):
|
| 434 |
+
if dataset is None:
|
| 435 |
+
return None
|
| 436 |
+
|
| 437 |
+
for tier in dataset:
|
| 438 |
+
if dataset[tier] is not None:
|
| 439 |
+
dataset[tier]["data"] = self.prepareData(dataset[tier]["instances"],
|
| 440 |
+
train=dataset[tier]["train"], filterKey=tier,
|
| 441 |
+
noBucket=noBucket)
|
| 442 |
+
|
| 443 |
+
for tier in dataset:
|
| 444 |
+
if dataset[tier] is not None:
|
| 445 |
+
del dataset[tier]["instances"]
|
| 446 |
+
|
| 447 |
+
return dataset
|
| 448 |
+
|
| 449 |
+
# Initializes word embeddings to random uniform / random normal / GloVe.
|
| 450 |
+
def initializeWordEmbeddings(self, wordsDict=None, noPadding=False):
|
| 451 |
+
# default dictionary to use for embeddings
|
| 452 |
+
if wordsDict is None:
|
| 453 |
+
wordsDict = self.questionDict
|
| 454 |
+
|
| 455 |
+
# uniform initialization
|
| 456 |
+
if config.wrdEmbUniform:
|
| 457 |
+
lowInit = -1.0 * config.wrdEmbScale
|
| 458 |
+
highInit = 1.0 * config.wrdEmbScale
|
| 459 |
+
embeddings = np.random.uniform(low=lowInit, high=highInit,
|
| 460 |
+
size=(wordsDict.getNumSymbols(), config.wrdEmbDim))
|
| 461 |
+
# normal initialization
|
| 462 |
+
else:
|
| 463 |
+
embeddings = config.wrdEmbScale * np.random.randn(wordsDict.getNumSymbols(),
|
| 464 |
+
config.wrdEmbDim)
|
| 465 |
+
|
| 466 |
+
# if wrdEmbRandom = False, use GloVE
|
| 467 |
+
counter = 0
|
| 468 |
+
if (not config.wrdEmbRandom):
|
| 469 |
+
with open(config.wordVectorsFile, 'r') as inFile:
|
| 470 |
+
for line in inFile:
|
| 471 |
+
line = line.strip().split()
|
| 472 |
+
word = line[0].lower()
|
| 473 |
+
vector = [float(x) for x in line[1:]]
|
| 474 |
+
index = wordsDict.sym2id.get(word)
|
| 475 |
+
if index is not None:
|
| 476 |
+
embeddings[index] = vector
|
| 477 |
+
counter += 1
|
| 478 |
+
|
| 479 |
+
print(counter)
|
| 480 |
+
print(self.questionDict.sym2id)
|
| 481 |
+
print(len(self.questionDict.sym2id))
|
| 482 |
+
print(self.answerDict.sym2id)
|
| 483 |
+
print(len(self.answerDict.sym2id))
|
| 484 |
+
print(self.qaDict.sym2id)
|
| 485 |
+
print(len(self.qaDict.sym2id))
|
| 486 |
+
|
| 487 |
+
if noPadding:
|
| 488 |
+
return embeddings # no embedding for padding symbol
|
| 489 |
+
else:
|
| 490 |
+
return embeddings[1:]
|
| 491 |
+
|
| 492 |
+
'''
|
| 493 |
+
Initializes words embeddings for question words and optionally for answer words
|
| 494 |
+
(when config.ansEmbMod == "BOTH"). If config.ansEmbMod == "SHARED", tie embeddings for
|
| 495 |
+
question and answer same symbols.
|
| 496 |
+
'''
|
| 497 |
+
|
| 498 |
+
def initializeQAEmbeddings(self):
|
| 499 |
+
# use same embeddings for questions and answers
|
| 500 |
+
if config.ansEmbMod == "SHARED":
|
| 501 |
+
qaEmbeddings = self.initializeWordEmbeddings(self.qaDict)
|
| 502 |
+
ansMap = np.array([self.qaDict.sym2id[sym] for sym in self.answerDict.id2sym])
|
| 503 |
+
embeddings = {"qa": qaEmbeddings, "ansMap": ansMap}
|
| 504 |
+
# use different embeddings for questions and answers
|
| 505 |
+
else:
|
| 506 |
+
qEmbeddings = self.initializeWordEmbeddings(self.questionDict)
|
| 507 |
+
aEmbeddings = None
|
| 508 |
+
if config.ansEmbMod == "BOTH":
|
| 509 |
+
aEmbeddings = self.initializeWordEmbeddings(self.answerDict, noPadding=True)
|
| 510 |
+
embeddings = {"q": qEmbeddings, "a": aEmbeddings}
|
| 511 |
+
return embeddings
|
| 512 |
+
|
| 513 |
+
'''
|
| 514 |
+
Preprocesses a given dataset into numpy arrays:
|
| 515 |
+
1. Reads the input data files into dictionary.
|
| 516 |
+
2. Saves the results jsons in files and loads them instead of parsing input if files exist/
|
| 517 |
+
3. Initializes word embeddings to random / GloVe.
|
| 518 |
+
4. Optionally filters data according to given filters.
|
| 519 |
+
5. Encodes and vectorize the data into numpy arrays.
|
| 520 |
+
5. Buckets the data according to the instances length.
|
| 521 |
+
'''
|
| 522 |
+
|
| 523 |
+
def preprocessData(self, question, debug=False):
|
| 524 |
+
# Read data into json and symbols' dictionaries
|
| 525 |
+
print(bold("Loading data..."))
|
| 526 |
+
start = time.time()
|
| 527 |
+
with open(config.questionDictFile(), "rb") as inFile:
|
| 528 |
+
self.questionDict = pickle.load(inFile)
|
| 529 |
+
with open(config.qaDictFile(), "rb") as inFile:
|
| 530 |
+
self.qaDict = pickle.load(inFile)
|
| 531 |
+
with open(config.answerDictFile(), "rb") as inFile:
|
| 532 |
+
self.answerDict = pickle.load(inFile)
|
| 533 |
+
question = question.replace('?', '').replace(', ', '').lower().split()
|
| 534 |
+
encodedQuestion = self.questionDict.encodeSequence(question)
|
| 535 |
+
data = {'question': np.array([encodedQuestion]), 'questionLength': np.array([len(encodedQuestion)])}
|
| 536 |
+
print("took {:.2f} seconds".format(time.time() - start))
|
| 537 |
+
|
| 538 |
+
# Initialize word embeddings (random / glove)
|
| 539 |
+
print(bold("Loading word vectors..."))
|
| 540 |
+
start = time.time()
|
| 541 |
+
embeddings = self.initializeQAEmbeddings()
|
| 542 |
+
print("took {:.2f} seconds".format(time.time() - start))
|
| 543 |
+
|
| 544 |
+
answer = 'yes' # DUMMY_ANSWER
|
| 545 |
+
self.answerDict.addSeq([answer])
|
| 546 |
+
self.qaDict.addSeq([answer])
|
| 547 |
+
|
| 548 |
+
config.questionWordsNum = self.questionDict.getNumSymbols()
|
| 549 |
+
config.answerWordsNum = self.answerDict.getNumSymbols()
|
| 550 |
+
|
| 551 |
+
return data, embeddings, self.answerDict
|