def compute_metrics(p):
preds = p.predictions[0] if isinstance(p.predictions,
tuple) else p.predictions
preds = paddle.to_tensor(preds)
label = paddle.to_tensor(p.label_ids)
probs = F.softmax(preds, axis=1)
metric = Accuracy()
metric.reset()
result = metric.compute(preds, label)
metric.update(result)
accu = metric.accumulate()
metric.reset()
return {"accuracy": accu}
def evaluate(model, args, mode='test'):
"""evaluate the model"""
model.eval()
metric = Accuracy()
eval_dataloader, processor = load_example(args, mode)
for batch in tqdm(eval_dataloader, total=len(eval_dataloader)):
logits = model(input_ids=batch[0], token_type_ids=batch[1])
labels = batch[2].reshape((
-1,
1,
))
correct = metric.compute(logits, labels)
metric.update(correct)
res = metric.accumulate()
print('Accuracy:', res)
model.train()
return res
class AccuracyAndF1(Metric):
"""
This class encapsulates Accuracy, Precision, Recall and F1 metric logic,
and `accumulate` function returns accuracy, precision, recall and f1.
The overview of all metrics could be seen at the document of `paddle.metric
<https://www.paddlepaddle.org.cn/documentation/docs/zh/develop/api/paddle/metric/Overview_cn.html>`_
for details.
Args:
topk (int or tuple(int), optional):
Number of top elements to look at for computing accuracy.
Defaults to (1,).
pos_label (int, optional): The positive label for calculating precision
and recall.
Defaults to 1.
name (str, optional):
String name of the metric instance. Defaults to 'acc_and_f1'.
Example:
.. code-block::
import paddle
from paddlenlp.metrics import AccuracyAndF1
x = paddle.to_tensor([[0.1, 0.9], [0.5, 0.5], [0.6, 0.4], [0.7, 0.3]])
y = paddle.to_tensor([[1], [0], [1], [1]])
m = AccuracyAndF1()
correct = m.compute(x, y)
m.update(correct)
res = m.accumulate()
print(res) # (0.5, 0.5, 0.3333333333333333, 0.4, 0.45)
"""
def __init__(self,
topk=(1, ),
pos_label=1,
name='acc_and_f1',
*args,
**kwargs):
super(AccuracyAndF1, self).__init__(*args, **kwargs)
self.topk = topk
self.pos_label = pos_label
self._name = name
self.acc = Accuracy(self.topk, *args, **kwargs)
self.precision = Precision(*args, **kwargs)
self.recall = Recall(*args, **kwargs)
self.reset()
def compute(self, pred, label, *args):
"""
Accepts network's output and the labels, and calculates the top-k
(maximum value in topk) indices for accuracy.
Args:
pred (Tensor):
Predicted tensor, and its dtype is float32 or float64, and
has a shape of [batch_size, num_classes].
label (Tensor):
The ground truth tensor, and its dtype is is int64, and has a
shape of [batch_size, 1] or [batch_size, num_classes] in one
hot representation.
Returns:
Tensor: Correct mask, each element indicates whether the prediction
equals to the label. Its' a tensor with a data type of float32 and
has a shape of [batch_size, topk].
"""
self.label = label
self.preds_pos = paddle.nn.functional.softmax(pred)[:, self.pos_label]
return self.acc.compute(pred, label)
def update(self, correct, *args):
"""
Updates the metrics states (accuracy, precision and recall), in order to
calculate accumulated accuracy, precision and recall of all instances.
Args:
correct (Tensor):
Correct mask for calculating accuracy, and it's a tensor with
shape [batch_size, topk] and has a dtype of
float32.
"""
self.acc.update(correct)
self.precision.update(self.preds_pos, self.label)
self.recall.update(self.preds_pos, self.label)
def accumulate(self):
"""
Calculates and returns the accumulated metric.
Returns:
tuple: The accumulated metric. A tuple of shape (acc, precision,
recall, f1, average_of_acc_and_f1)
With the fileds:
- acc (numpy.float64):
The accumulated accuracy.
- precision (numpy.float64):
The accumulated precision.
- recall (numpy.float64):
The accumulated recall.
- f1 (numpy.float64):
The accumulated f1.
- average_of_acc_and_f1 (numpy.float64):
The average of accumulated accuracy and f1.
"""
acc = self.acc.accumulate()
precision = self.precision.accumulate()
recall = self.recall.accumulate()
if precision == 0.0 or recall == 0.0:
f1 = 0.0
else:
# 1/f1 = 1/2 * (1/precision + 1/recall)
f1 = (2 * precision * recall) / (precision + recall)
return (
acc,
precision,
recall,
f1,
(acc + f1) / 2,
)
def reset(self):
"""
Resets all metric states.
"""
self.acc.reset()
self.precision.reset()
self.recall.reset()
self.label = None
self.preds_pos = None
def name(self):
"""
Returns name of the metric instance.
Returns:
str: The name of the metric instance.
"""
return self._name
