def on_batch_end(self, state):
import torchelie.utils as tu
cm = state['cm']
pred = tu.as_multiclass_shape(state['pred']).argmax(1)
true = state['batch'][1]
for p, t in zip(pred, true):
cm[p][t] += 1
def analyze(self,
batch: torch.Tensor,
pred: torch.Tensor,
true: torch.Tensor,
pred_label: Optional[torch.Tensor] = None,
paths: Optional[List[str]] = None) -> None:
pred = as_multiclass_shape(pred, as_probs=True)
for_label = pred.gather(1, true.unsqueeze(1))
if pred_label is None:
pred_label = pred.argmax(dim=1)
best_pred = pred.gather(1, pred_label.unsqueeze(1))
if paths is None:
paths = [None] * len(batch)
this_data = list(
zip(batch, for_label, true, pred_label == true, paths, pred_label,
best_pred))
self.best += this_data
self.best.sort(key=lambda x: -x[1])
self.best = self.best[:self.topk]
self.worst += this_data
self.worst.sort(key=lambda x: x[1])
self.worst = self.worst[:self.topk]
self.confused += this_data
self.confused.sort(key=lambda x: abs(self.center_value - x[1]))
self.confused = self.confused[:self.topk]
def on_batch_end(self, state):
pred, y = state['pred'], state['batch'][1]
pred = tu.as_multiclass_shape(pred)
batch_correct = pred.argmax(1).eq(y).float()
if isinstance(self.avg, RunningAvg):
self.avg.log(batch_correct.sum(), pred.shape[0])
else:
self.avg.log(batch_correct.mean())
if self.post_each_batch:
state['metrics']['acc'] = self.avg.get()
def analyze(self, batch, pred, true, pred_label=None, paths=None):
pred = as_multiclass_shape(pred, as_probs=True)
for_label = pred.gather(1, true.unsqueeze(1))
if pred_label is None:
pred_label = pred.argmax(dim=1)
if paths is None:
paths = [None] * len(batch)
this_data = list(zip(batch, for_label, true, pred_label == true,
paths, pred_label))
self.best += this_data
self.best.sort(key=lambda x: -x[1])
self.best = self.best[:self.topk]
self.worst += this_data
self.worst.sort(key=lambda x: x[1])
self.worst = self.worst[:self.topk]
self.confused += this_data
self.confused.sort(key=lambda x: abs(self.center_value - x[1]))
self.confused = self.confused[:self.topk]
