def predict(self, dataset):
weights_original = self.copy_param()
metrics = Metrics()
while dataset.next_rel():
metrics.reset_task_metrics(dataset.cur_rel)
for idx_iter in range(self.hyp.num_test_inner_iter):
for idx_data in range(self.hyp.num_shot):
dscp1, dscp2, dscpn, dscpr = dataset.get_sup_inner_data()
for idx_iter in range(self.hyp.num_test_inner_iter):
self.__inner_train_step(dscp1, dscp2, dscpn, dscpr,
False)
batch_dscp_cand, dscpr = dataset.get_qur_task_cand_rel()
cands_emb = []
if self.memory:
self.qur_dscpr_emb, self.qur_head_memory_dscpr_emb, self.qur_tail_memory_dscpr_emb = self.encoder.encode_rel(
dscpr)
cands_emb = [
self.encoder(bucket, self.qur_tail_memory_dscpr_emb)
for bucket in batch_dscp_cand
]
else:
self.qur_dscpr_emb = self.encoder(dscpr)
cands_emb = [
self.encoder(bucket) for bucket in batch_dscp_cand
]
self.qur_cands_emb = T.cat(cands_emb, 0)
while 1:
qur_task = dataset.get_qur_inner_data()
if qur_task is None:
break
sim_cands = self.__predict_triplet(*qur_task)
sim_cands = sim_cands.detach(
).data # sim_cands[0]: y_true sim, sim_cands[1:]: y_cands sim
sim_cands = sim_cands.cpu().numpy()
metrics.add(sim_cands)
self.set_param(weights_original)
if self.memory:
self.qur_dscpr_emb = self.qur_cands_emb = self.qur_head_memory_dscpr_emb = self.qur_tail_memory_dscpr_emb = None
else:
self.qur_dscpr_emb = self.qur_cands_emb = None
metrics.log_task_metric()
metrics.log_overall_metric()
with torch.no_grad():
for batch_idx, (images, masks) in enumerate(validation_loader):
images = images.cuda()
masks = masks.cuda()
preds = net(images)
loss = loss_fn(preds, masks)
test_loss += loss.item()
preds = preds.argmax(dim=1)
preds = preds.view(-1).cpu().data.numpy()
masks = masks.view(-1).cpu().data.numpy()
metrics.add(preds, masks)
n_iter = (epoch - 1) * iter_per_epoch + batch_idx + 1
miou = metrics.iou()
precision = metrics.precision()
recall = metrics.recall()
metrics.clear()
utils.visualize_scalar(
writer,
'Test/mIOU',
miou,
epoch,
)
utils.visualize_scalar(
