def evaluate(self, loader, verbose=False, write_summary=False, epoch=None):
self.eval()
valid_loss = 0
all_class_probs = []
all_labels = []
with torch.no_grad():
for batch_idx, sample in enumerate(loader):
signal, labels = (
sample["signal"].to(self.device),
sample["labels"].to(self.device).float()
)
outputs = self(signal)
class_logits = outputs["class_logits"].squeeze(-1)
loss = (
focal_loss(
class_logits,
labels,
)
).item()
multiplier = len(labels) / len(loader.dataset)
valid_loss += loss * multiplier
class_probs = torch.sigmoid(class_logits).data.cpu().numpy()
labels = labels.data.cpu().numpy()
all_class_probs.extend(class_probs)
all_labels.extend(labels)
all_class_probs = np.asarray(all_class_probs)
all_labels = np.asarray(all_labels)
metric = compute_inverse_eer(all_labels, all_class_probs)
if write_summary:
self.add_scalar_summaries(
valid_loss,
metric,
writer=self.valid_writer, global_step=self.global_step
)
if verbose:
print("\nValidation loss: {:.4f}".format(valid_loss))
print("Validation metric: {:.4f}".format(metric))
return metric
def train_epoch(self,
train_loader,
epoch,
log_interval,
write_summary=True):
self.train()
print("\n" + " " * 10 +
"****** Epoch {epoch} ******\n".format(epoch=epoch))
history = deque(maxlen=30)
self.optimizer.zero_grad()
accumulated_loss = 0
with tqdm(total=len(train_loader), ncols=80) as pb:
for batch_idx, sample in enumerate(train_loader):
self.global_step += 1
make_step(self.scheduler, step=self.global_step)
signal, labels = (
sample["signal"].to(self.device),
sample["labels"].to(self.device).float(),
)
outputs = self(signal)
class_logits = outputs["class_logits"].squeeze(-1)
loss = (focal_loss(
class_logits,
labels,
)) / self.config.train.accumulation_steps
loss.backward()
accumulated_loss += loss
if batch_idx % self.config.train.accumulation_steps == 0:
self.optimizer.step()
accumulated_loss = 0
self.optimizer.zero_grad()
class_logits = take_first_column(class_logits) # human is 1
labels = take_first_column(labels)
probs = torch.sigmoid(class_logits).data.cpu().numpy()
labels = labels.data.cpu().numpy()
metric = compute_inverse_eer(labels, probs)
history.append(metric)
pb.update()
pb.set_description("Loss: {:.4f}, Metric: {:.4f}".format(
loss.item(), np.mean(history)))
if batch_idx % log_interval == 0:
self.add_scalar_summaries(loss.item(), metric,
self.train_writer,
self.global_step)
if batch_idx == 0:
self.add_image_summaries(signal, self.global_step,
self.train_writer)
val_df_files = [
os.path.join(experiment.predictions,
"val_preds_fold_{}.csv".format(fold))
for fold in range(config.data._n_folds)
]
val_predictions_df = pd.concat([
pd.read_csv(file) for file in val_df_files
]).reset_index(drop=True)
labels = np.asarray([
item["labels"][0]
for item in AntispoofDataset(audio_files=train_df.fname.tolist(),
labels=train_df.labels.values,
transform=None)
])
val_labels_df = pd.DataFrame(labels, columns=["labels"])
val_labels_df["fname"] = train_df.fname
assert set(val_predictions_df.fname) == set(val_labels_df.fname)
val_predictions_df.sort_values(by="fname", inplace=True)
val_labels_df.sort_values(by="fname", inplace=True)
metric = compute_inverse_eer(
val_labels_df.drop("fname", axis=1).values,
val_predictions_df.drop("fname", axis=1).values)
experiment.register_result("metric", metric)
