Spaces:
Running
Running
File size: 8,779 Bytes
626eca0 |
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 |
import logging
from typing import Dict, List, Optional
import lightning as pl
import torch
from lightning.pytorch.trainer.states import RunningStage
from sklearn.metrics import label_ranking_average_precision_score
from relik.common.log import get_console_logger, get_logger
from relik.retriever.callbacks.base import DEFAULT_STAGES, NLPTemplateCallback
console_logger = get_console_logger()
logger = get_logger(__name__, level=logging.INFO)
class RecallAtKEvaluationCallback(NLPTemplateCallback):
"""
Computes the recall at k for the predictions. Recall at k is computed as the number of
correct predictions in the top k predictions divided by the total number of correct
predictions.
Args:
k (`int`):
The number of predictions to consider.
prefix (`str`, `optional`):
The prefix to add to the metrics.
verbose (`bool`, `optional`, defaults to `False`):
Whether to log the metrics.
prog_bar (`bool`, `optional`, defaults to `True`):
Whether to log the metrics to the progress bar.
"""
def __init__(
self,
k: int = 100,
prefix: Optional[str] = None,
verbose: bool = False,
prog_bar: bool = True,
*args,
**kwargs,
):
super().__init__()
self.k = k
self.prefix = prefix
self.verbose = verbose
self.prog_bar = prog_bar
@torch.no_grad()
def __call__(
self,
trainer: pl.Trainer,
pl_module: pl.LightningModule,
predictions: Dict,
*args,
**kwargs,
) -> dict:
"""
Computes the recall at k for the predictions.
Args:
trainer (:obj:`lightning.trainer.trainer.Trainer`):
The trainer object.
pl_module (:obj:`lightning.core.lightning.LightningModule`):
The lightning module.
predictions (:obj:`Dict`):
The predictions.
Returns:
:obj:`Dict`: The computed metrics.
"""
if self.verbose:
logger.info(f"Computing recall@{self.k}")
# metrics to return
metrics = {}
stage = trainer.state.stage
if stage not in DEFAULT_STAGES:
raise ValueError(
f"Stage {stage} not supported, only `validate` and `test` are supported."
)
for dataloader_idx, samples in predictions.items():
hits, total = 0, 0
for sample in samples:
# compute the recall at k
# cut the predictions to the first k elements
predictions = sample["predictions"][: self.k]
hits += len(set(predictions) & set(sample["gold"]))
total += len(set(sample["gold"]))
# compute the mean recall at k
recall_at_k = hits / total
metrics[f"recall@{self.k}_{dataloader_idx}"] = recall_at_k
metrics[f"recall@{self.k}"] = sum(metrics.values()) / len(metrics)
if self.prefix is not None:
metrics = {f"{self.prefix}_{k}": v for k, v in metrics.items()}
else:
metrics = {f"{stage.value}_{k}": v for k, v in metrics.items()}
pl_module.log_dict(
metrics, on_step=False, on_epoch=True, prog_bar=self.prog_bar
)
if self.verbose:
logger.info(
f"Recall@{self.k} on {stage.value}: {metrics[f'{stage.value}_recall@{self.k}']}"
)
return metrics
class AvgRankingEvaluationCallback(NLPTemplateCallback):
"""
Computes the average ranking of the gold label in the predictions. Average ranking is
computed as the average of the rank of the gold label in the predictions.
Args:
k (`int`):
The number of predictions to consider.
prefix (`str`, `optional`):
The prefix to add to the metrics.
stages (`List[str]`, `optional`):
The stages to compute the metrics on. Defaults to `["validate", "test"]`.
verbose (`bool`, `optional`, defaults to `False`):
Whether to log the metrics.
"""
def __init__(
self,
k: int,
prefix: Optional[str] = None,
stages: Optional[List[str]] = None,
verbose: bool = True,
*args,
**kwargs,
):
super().__init__()
self.k = k
self.prefix = prefix
self.verbose = verbose
self.stages = (
[RunningStage(stage) for stage in stages] if stages else DEFAULT_STAGES
)
@torch.no_grad()
def __call__(
self,
trainer: pl.Trainer,
pl_module: pl.LightningModule,
predictions: Dict,
*args,
**kwargs,
) -> dict:
"""
Computes the average ranking of the gold label in the predictions.
Args:
trainer (:obj:`lightning.trainer.trainer.Trainer`):
The trainer object.
pl_module (:obj:`lightning.core.lightning.LightningModule`):
The lightning module.
predictions (:obj:`Dict`):
The predictions.
Returns:
:obj:`Dict`: The computed metrics.
"""
if not predictions:
logger.warning("No predictions to compute the AVG Ranking metrics.")
return {}
if self.verbose:
logger.info(f"Computing AVG Ranking@{self.k}")
# metrics to return
metrics = {}
stage = trainer.state.stage
if stage not in self.stages:
raise ValueError(
f"Stage `{stage}` not supported, only `validate` and `test` are supported."
)
for dataloader_idx, samples in predictions.items():
rankings = []
for sample in samples:
window_candidates = sample["predictions"][: self.k]
window_labels = sample["gold"]
for wl in window_labels:
if wl in window_candidates:
rankings.append(window_candidates.index(wl) + 1)
avg_ranking = sum(rankings) / len(rankings) if len(rankings) > 0 else 0
metrics[f"avg_ranking@{self.k}_{dataloader_idx}"] = avg_ranking
if len(metrics) == 0:
metrics[f"avg_ranking@{self.k}"] = 0
else:
metrics[f"avg_ranking@{self.k}"] = sum(metrics.values()) / len(metrics)
prefix = self.prefix or stage.value
metrics = {
f"{prefix}_{k}": torch.as_tensor(v, dtype=torch.float32)
for k, v in metrics.items()
}
pl_module.log_dict(metrics, on_step=False, on_epoch=True, prog_bar=False)
if self.verbose:
logger.info(
f"AVG Ranking@{self.k} on {prefix}: {metrics[f'{prefix}_avg_ranking@{self.k}']}"
)
return metrics
class LRAPEvaluationCallback(NLPTemplateCallback):
def __init__(
self,
k: int = 100,
prefix: Optional[str] = None,
verbose: bool = False,
prog_bar: bool = True,
*args,
**kwargs,
):
super().__init__()
self.k = k
self.prefix = prefix
self.verbose = verbose
self.prog_bar = prog_bar
@torch.no_grad()
def __call__(
self,
trainer: pl.Trainer,
pl_module: pl.LightningModule,
predictions: Dict,
*args,
**kwargs,
) -> dict:
if self.verbose:
logger.info(f"Computing recall@{self.k}")
# metrics to return
metrics = {}
stage = trainer.state.stage
if stage not in DEFAULT_STAGES:
raise ValueError(
f"Stage {stage} not supported, only `validate` and `test` are supported."
)
for dataloader_idx, samples in predictions.items():
scores = [sample["scores"][: self.k] for sample in samples]
golds = [sample["gold"] for sample in samples]
# compute the mean recall at k
lrap = label_ranking_average_precision_score(golds, scores)
metrics[f"lrap@{self.k}_{dataloader_idx}"] = lrap
metrics[f"lrap@{self.k}"] = sum(metrics.values()) / len(metrics)
prefix = self.prefix or stage.value
metrics = {
f"{prefix}_{k}": torch.as_tensor(v, dtype=torch.float32)
for k, v in metrics.items()
}
pl_module.log_dict(
metrics, on_step=False, on_epoch=True, prog_bar=self.prog_bar
)
if self.verbose:
logger.info(
f"Recall@{self.k} on {stage.value}: {metrics[f'{stage.value}_recall@{self.k}']}"
)
return metrics
|