-
Notifications
You must be signed in to change notification settings - Fork 0
/
metrics.py
299 lines (242 loc) · 12.4 KB
/
metrics.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
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
import numpy as np
import numpy as np
import torch
import torchmetrics
from ignite.exceptions import NotComputableError
from ignite.metrics import Precision, Recall, TopKCategoricalAccuracy
from ogb.graphproppred import Evaluator as GraphEvaluator
from ogb.linkproppred import Evaluator as LinkEvaluator
from ogb.nodeproppred import Evaluator as NodeEvaluator
from pytorch_lightning.metrics import F1, AUROC, AveragePrecision, MeanSquaredError, Accuracy
from sklearn.metrics import homogeneity_score, completeness_score, normalized_mutual_info_score, \
adjusted_mutual_info_score
from utils import filter_samples
def clustering_metrics(y_true, y_pred, metrics=["homogeneity", "completeness", "nmi", "ami"]):
results = {}
for metric in metrics:
if "homogeneity" in metric:
results[metric] = homogeneity_score(y_true, y_pred)
elif "completeness" in metric:
results[metric] = completeness_score(y_true, y_pred)
elif "nmi" in metric:
results[metric] = normalized_mutual_info_score(y_true, y_pred, average_method="arithmetic")
elif "ami" in metric:
results[metric] = adjusted_mutual_info_score(y_true, y_pred)
return results
class Metrics(torch.nn.Module):
def __init__(self, prefix, loss_type: str, threshold=0.5, top_k=[1, 5, 10], n_classes: int = None,
multilabel: bool = None, metrics=["precision", "recall", "top_k", "accuracy"]):
super().__init__()
self.loss_type = loss_type.upper()
self.threshold = threshold
self.n_classes = n_classes
self.multilabel = multilabel
self.top_ks = top_k
self.prefix = prefix
self.metrics = {}
for metric in metrics:
if "precision" == metric:
self.metrics[metric] = Precision(average=True, is_multilabel=multilabel)
elif "recall" == metric:
self.metrics[metric] = Recall(average=True, is_multilabel=multilabel)
elif "top_k" in metric:
if n_classes:
top_k = [k for k in top_k if k < n_classes]
if multilabel:
self.metrics[metric] = TopKMultilabelAccuracy(k_s=top_k)
else:
self.metrics[metric] = TopKCategoricalAccuracy(k=max(int(np.log(n_classes)), 1),
output_transform=None)
elif "macro_f1" in metric:
self.metrics[metric] = F1(num_classes=n_classes, average="macro", multilabel=multilabel)
elif "micro_f1" in metric:
self.metrics[metric] = F1(num_classes=n_classes, average="micro", multilabel=multilabel)
elif "mse" == metric:
self.metrics[metric] = MeanSquaredError()
elif "auroc" == metric:
self.metrics[metric] = AUROC(num_classes=n_classes)
elif "avg_precision" in metric:
self.metrics[metric] = AveragePrecision(num_classes=n_classes, )
elif "accuracy" in metric:
self.metrics[metric] = Accuracy(top_k=int(metric.split("@")[-1]) if "@" in metric else None)
elif "ogbn" in metric:
self.metrics[metric] = OGBNodeClfMetrics(NodeEvaluator(metric))
elif "ogbg" in metric:
self.metrics[metric] = OGBNodeClfMetrics(GraphEvaluator(metric))
elif "ogbl" in metric:
self.metrics[metric] = OGBLinkPredMetrics(LinkEvaluator(metric))
else:
print(f"WARNING: metric {metric} doesn't exist")
# Needed to add the PytorchGeometric methods as Modules, so they'll be on the correct CUDA device during training
if isinstance(self.metrics[metric], torchmetrics.metric.Metric):
setattr(self, metric, self.metrics[metric])
self.reset_metrics()
def update_metrics(self, y_hat: torch.Tensor, y: torch.Tensor, weights=None):
"""
:param y_pred:
:param y_true:
:param weights:
"""
y_pred = y_hat.detach()
y_true = y.detach()
y_pred, y_true = filter_samples(y_pred, y_true, weights)
# Apply softmax/sigmoid activation if needed
if "LOGITS" in self.loss_type or "FOCAL" in self.loss_type:
if "SOFTMAX" in self.loss_type:
y_pred = torch.softmax(y_pred, dim=1)
else:
y_pred = torch.sigmoid(y_pred)
elif "NEGATIVE_LOG_LIKELIHOOD" == self.loss_type or "SOFTMAX_CROSS_ENTROPY" in self.loss_type:
y_pred = torch.softmax(y_pred, dim=1)
for metric in self.metrics:
# torchmetrics metrics
if isinstance(self.metrics[metric], torchmetrics.metric.Metric):
self.metrics[metric].update(y_pred, y_true)
# Torch ignite metrics
elif "precision" in metric or "recall" in metric or "accuracy" in metric:
if not self.multilabel and y_true.dim() == 1:
self.metrics[metric].update((self.hot_encode(y_pred.argmax(1, keepdim=False), type_as=y_true),
self.hot_encode(y_true, type_as=y_pred)))
else:
self.metrics[metric].update(((y_pred > self.threshold).type_as(y_true), y_true))
# Torch ignite metrics
elif metric == "top_k":
self.metrics[metric].update((y_pred, y_true))
# OGB metrics
elif "ogb" in metric:
if metric in ["ogbl-ddi", "ogbl-collab"]:
y_true = y_true[:, 0]
elif "ogbg-mol" in metric:
# print(tensor_sizes({"y_pred": y_pred, "y_true": y_true}))
pass
self.metrics[metric].update((y_pred, y_true))
else:
raise Exception(f"Metric {metric} has problem at .update()")
def hot_encode(self, labels, type_as):
if labels.dim() == 2:
return labels
elif labels.dim() == 1:
labels = torch.eye(self.n_classes)[labels].type_as(type_as)
return labels
def compute_metrics(self):
logs = {}
for metric in self.metrics:
try:
if "ogb" in metric:
logs.update(self.metrics[metric].compute(prefix=self.prefix))
elif metric == "top_k" and isinstance(self.metrics[metric], TopKMultilabelAccuracy):
logs.update(self.metrics[metric].compute(prefix=self.prefix))
elif metric == "top_k" and isinstance(self.metrics[metric], TopKCategoricalAccuracy):
metric_name = (metric if self.prefix is None else \
self.prefix + metric) + f"@{self.metrics[metric]._k}"
logs[metric_name] = self.metrics[metric].compute()
else:
metric_name = metric if self.prefix is None else self.prefix + metric
logs[metric_name] = self.metrics[metric].compute()
except Exception as e:
print(f"Had problem with metric {metric}, {str(e)}\r")
# Needed for Precision(average=False) metrics
logs = {k: v.mean() if isinstance(v, torch.Tensor) and v.numel() > 1 else v for k, v in logs.items()}
return logs
def reset_metrics(self):
for metric in self.metrics:
self.metrics[metric].reset()
class OGBNodeClfMetrics(torchmetrics.Metric):
def __init__(self, evaluator, compute_on_step: bool = True, dist_sync_on_step: bool = False,
process_group=None, dist_sync_fn=None):
super().__init__(compute_on_step, dist_sync_on_step, process_group, dist_sync_fn)
self.evaluator = evaluator
self.y_pred = []
self.y_true = []
def reset(self):
self.y_pred = []
self.y_true = []
def update(self, y_pred, y_true):
if isinstance(self.evaluator, (NodeEvaluator, GraphEvaluator)):
assert y_pred.dim() == 2
if y_true.dim() == 1 or y_true.size(1) == 1:
y_pred = y_pred.argmax(axis=1)
if y_pred.dim() <= 1:
y_pred = y_pred.unsqueeze(-1)
if y_true.dim() <= 1:
y_true = y_true.unsqueeze(-1)
self.y_true.append(y_true)
self.y_pred.append(y_pred)
def compute(self, prefix=None):
if isinstance(self.evaluator, NodeEvaluator):
output = self.evaluator.eval({"y_pred": torch.cat(self.y_pred, dim=0),
"y_true": torch.cat(self.y_true, dim=0)})
elif isinstance(self.evaluator, LinkEvaluator):
y_pred_pos = torch.cat(self.y_pred, dim=0).squeeze(-1)
y_pred_neg = torch.cat(self.y_true, dim=0)
output = self.evaluator.eval({"y_pred_pos": y_pred_pos,
"y_pred_neg": y_pred_neg})
output = {k.strip("_list"): v.mean().item() for k, v in output.items()}
elif isinstance(self.evaluator, GraphEvaluator):
input_shape = {"y_true": torch.cat(self.y_pred, dim=0),
"y_pred": torch.cat(self.y_true, dim=0)}
output = self.evaluator.eval(input_shape)
else:
raise Exception(f"implement eval for {self.evaluator}")
if prefix is None:
return {f"{k}": v for k, v in output.items()}
else:
return {f"{prefix}{k}": v for k, v in output.items()}
class OGBLinkPredMetrics(torchmetrics.Metric):
def __init__(self, evaluator: LinkEvaluator, compute_on_step: bool = True, dist_sync_on_step: bool = False,
process_group=None, dist_sync_fn=None):
super().__init__(compute_on_step, dist_sync_on_step, process_group, dist_sync_fn)
self.evaluator = evaluator
self.outputs = {}
def reset(self):
self.outputs = {}
def update(self, e_pred_pos, e_pred_neg):
if e_pred_pos.dim() > 1:
e_pred_pos = e_pred_pos.squeeze(-1)
# if e_pred_neg.dim() <= 1:
# e_pred_neg = e_pred_neg.unsqueeze(-1)
# print("e_pred_pos", e_pred_pos.shape)
# print("e_pred_neg", e_pred_neg.shape)
output = self.evaluator.eval({"y_pred_pos": e_pred_pos,
"y_pred_neg": e_pred_neg})
for k, v in output.items():
if isinstance(v, float):
score = torch.tensor([v])
self.outputs.setdefault(k.strip("_list"), []).append(score)
else:
self.outputs.setdefault(k.strip("_list"), []).append(v.mean())
def compute(self, prefix=None):
output = {k: torch.stack(v, dim=0).mean().item() for k, v in self.outputs.items()}
if prefix is None:
return {f"{k}": v for k, v in output.items()}
else:
return {f"{prefix}{k}": v for k, v in output.items()}
class TopKMultilabelAccuracy(torchmetrics.Metric):
"""
Calculates the top-k categorical accuracy.
- `update` must receive output of the form `(y_pred, y)` or `{'y_pred': y_pred, 'y': y}` Tensors of size (batch_size, n_classes).
"""
def __init__(self, k_s=[5, 10, 50, 100, 200], compute_on_step: bool = True, dist_sync_on_step: bool = False,
process_group=None, dist_sync_fn=None):
super().__init__(compute_on_step, dist_sync_on_step, process_group, dist_sync_fn)
self.k_s = k_s
def reset(self):
self._num_correct = {k: 0 for k in self.k_s}
self._num_examples = 0
def update(self, y_pred, y_true):
batch_size, n_classes = y_true.size()
_, top_indices = y_pred.topk(k=max(self.k_s), dim=1, largest=True, sorted=True)
for k in self.k_s:
y_true_select = torch.gather(y_true, 1, top_indices[:, :k])
corrects_in_k = y_true_select.sum(1) * 1.0 / k
corrects_in_k = corrects_in_k.sum(0) # sum across all samples to get # of true positives
self._num_correct[k] += corrects_in_k.item()
self._num_examples += batch_size
def compute(self, prefix=None) -> dict:
if self._num_examples == 0:
raise NotComputableError("TopKCategoricalAccuracy must have at"
"least one example before it can be computed.")
if prefix is None:
return {f"top_k@{k}": self._num_correct[k] / self._num_examples for k in self.k_s}
else:
return {f"{prefix}top_k@{k}": self._num_correct[k] / self._num_examples for k in self.k_s}