File size: 10,865 Bytes
c6e3750 |
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 |
from collections import defaultdict, Counter
import copy, pprint, re
import itertools
from tqdm import tqdm
import scoring_metrics as metrics
import text_processing_utils as text_utils
class ScoreAccumulator(object):
def __init__(
self,
measures=["diversity"],
default_matching="",
require_prompt=False,
require_target=True,
verbose=True,
):
"""
Score accumulation object for measuring model inference performance
on a dataset of songs.
Capable of measuring:
Lexical diversity : "diversity"
Meter consistency : "meter"
Syllabic consistency : "syllable"
Parameters
----------
measures : list, optional
list of measures to calculate for each song, by default ["diversity"]
default_matching : str
string defining the default stanza or line matching for comparison
require_prompt : bool, optional
whether to require the model prompt for each song, by default False
require_target : bool, optional
whether to requre a target response for each song, by default True
"""
self.measures = measures
self.default_matching = default_matching
self.require_prompt = require_prompt
self.require_target = require_target
self.verbose = verbose
self.__total_score = Counter({measure: 0 for measure in self.measures})
self.__total_target_score = Counter({measure: 0 for measure in self.measures})
self._database = {}
self.measure_functions = {
"diversity": metrics.measure_lex_div,
"meter": metrics.measure_meter,
"syllable": metrics.measure_syllable,
}
supported_meas = {"diversity", "meter", "syllable"}
for measure in self.measures:
if measure not in {"diversity", "meter", "syllable"}:
raise NotImplementedError(f"Only measures {supported_meas} supported.")
def get_total_pred_score(self, measure):
"""
Return current score totals of model predictions for a given
measure
Parameters
----------
measure : str
measure to get score for
Returns
-------
float
average (by song) cumulative score
"""
measure_score = self.__total_score[measure]
return measure_score / len(self._database)
def get_total_target_score(self, measure):
"""
Return current score total of target responses for a given
measure
Parameters
----------
measure : str
measure to get score for
Returns
-------
float
average (by song) cumulative score
"""
measure_score = self.__total_target_score[measure]
return measure_score / len(self._database)
def get_ids(self):
return list(self._database.keys())
def get_prompts(self, ids=[]):
if ids:
return [self._database[id]["prompt"] for id in ids]
return [self._database[songid]["prompt"] for songid in self.get_ids]
def get_preds(self, ids=[]):
if ids:
return [self._database[id]["model_response"] for id in ids]
return [self._database[songid]["model_response"] for songid in self.get_ids]
def get_targets(self, ids=[]):
if ids:
return [self._database[id]["target_response"] for id in ids]
return [self._database[songid]["target_response"] for songid in self.get_ids]
def get_pred_scores(self, ids=[]):
if ids:
return [self._database[id]["pred_scores"] for id in ids]
return [self._database[songid]["pred_scores"] for songid in self.get_ids]
def get_target_scores(self, ids=[]):
if ids:
return [self._database[id]["target_scores"] for id in ids]
return [self._database[songid]["target_scores"] for songid in self.get_ids]
def score_all_songs(self, song_database):
"""
Scores all songs in a song database (a list of song dictionaries) and adds
them to accumulators database.
Parameters
----------
song_database : list
list of song dictionaries containing {id, prompt, model_response, and
target_response} keys
"""
# TODO: add safety checks
for song_dict in tqdm(song_database, position=0):
self.score_song(song_dict)
def score_song(self, song_dict):
"""
Adds song to accumulator and scores song.
Song dicts are formatted as follows:
{
id: unique id number,
prompt: prompt string, #if required
model_response: response from model,
target_response: ground truth text #if required
}
Parameters
----------
song_dict : dict
dict containing model predictions and targets for song
"""
song_dict = copy.deepcopy(song_dict)
try:
id = song_dict["id"]
pred = song_dict["model_response"]
if self.require_prompt:
song_dict["prompt"]
if self.require_target:
target = song_dict["target_response"]
except Exception as e:
raise AssertionError(f"Provided song dict formatted incorrectly: {e}")
assert id not in self._database, f"Multiple entries found for unique id: {id}"
scores = {
"pred_scores": {measure: 0 for measure in self.measures},
"target_scores": {measure: 0 for measure in self.measures},
}
pred_stanzas, pred_matching = self.get_stanzas(pred, get_matching=True)
target_stanzas, target_matching = self.get_stanzas(target, get_matching=True)
for measure in (
tqdm(self.measures, position=1, leave=0) if self.verbose else self.measures
):
scores["pred_scores"][measure] = self.measure_stanzas(
pred_stanzas,
self.measure_functions[measure],
matching=pred_matching if measure != "diversity" else "",
)
scores["target_scores"][measure] = self.measure_stanzas(
target_stanzas,
self.measure_functions[measure],
matching=target_matching if measure != "diversity" else "",
)
self.__total_score += Counter(scores["pred_scores"])
self.__total_target_score += Counter(scores["target_scores"])
song_dict.pop("id")
song_dict.update(scores)
self._database[id] = song_dict
def get_stanzas(self, song_text, get_matching=False):
"""
Given the raw text of a song, gets the stanza paragraphs according to
their in-text label as {hook, verse, chorus, bridge, intro, outro}
Parameters
----------
song_text : str
raw song text
get_matching : bool, optional,
whether to return a matching
Returns
-------
list : str
raw text, stripped of their stanza identifiers, split into stanzas
"""
stanza_pairs = text_utils.get_stanzas(song_text)
matching_map = {}
matching_str = ""
match_idx = 0 # NOTE: using ints prohibits > 10 keywords per song
stanzas = []
for i, (kword, stanza) in enumerate(stanza_pairs):
assert match_idx < 10, "Stanza keywords are too complex, please simplify"
stanzas.append(
[n for n in stanza.split("\n") if len(re.sub(r"\s+", "", n)) > 0]
)
if i == 0:
matching_map[kword] = str(match_idx)
matching_str = str(match_idx)
match_idx += 1
elif kword in matching_map:
matching_str += "-" + matching_map[kword]
else:
matching_map[kword] = str(match_idx)
matching_str += "-" + matching_map[kword]
match_idx += 1
if get_matching:
return stanzas, matching_str
else:
return stanzas
def build_matches(self, stanzas, matching=""):
"""
Builds matchings between stanzas given a matching string. If no matching
string provided, matchings are just the stanzas themselves.
Parameters
----------
stanzas : list
list of input strings (stanzas) for different stanzas in a song
matching : str, optional
string defining the stanza or line matching for comparison, by default ""
Returns
-------
list
list of "matched" stanza pairings
"""
if matching == "":
return [(s,) for s in stanzas]
matching = matching.split("-")
assert len(matching) == len(stanzas), "Incorrect matching to stanzas"
matches_dict = defaultdict(lambda: [])
pairings = []
# map stanzas to their matching char
for i, p in enumerate(stanzas):
matches_dict[matching[i]].append(p)
# build pairings
for matches in matches_dict.values():
for p1, p2 in itertools.combinations(matches, 2):
pairings.append([p1, p2])
return pairings
def measure_stanzas(self, stanzas, measure_function, matching=""):
"""
Computes and returns a measure of a list of stanzas (string segments made up of numerous
lines). If measure is a consistency score, a comparison stanza is required.
Comparison metrics will be computed along the cartesian product of stanzas with matching
keys.
For example, given:
stanzas = ["line1", "line2", "line3", "line4"]
matching = "a-b-a-b"
-> Comparisons computed between [("line1", "line3"), ("line2", "line4")]
Parameters
----------
stanzas : list
list of input strings (stanzas) for different stanzas in a song
measure_function : function
callable function that takes in a pairing and returns a float
matching : str
string defining the stanza or line matching for comparison
Returns
-------
float
averaged (across measurings) measure score
"""
pairings = self.build_matches(stanzas, matching)
if len(pairings) == 0:
# NOTE: dangerous, since if no pairings the song structure could be fine..
return 0
score = 0
for pairing in pairings:
score += measure_function(*pairing)
final_score = score / len(pairings)
return final_score
|