def test_model(test_loader, net):
net.eval()
device = params['device']
batch_size = params['batch_size']
test_loss = 0
test_acc = 0
test_iou = {}
with torch.no_grad():
for batch_index, (img, target) in enumerate(test_loader):
img, target = img.to(device), target.to(device)
if model_version == 'deeplab':
output = net(img)['out']
else:
output = net(img)
target = target.long()
loss = criterion(output, target).item()
test_loss += loss
pred = aux.get_predicted_image(output)
output, target, pred = output.detach().cpu(), target.detach(
).cpu(), pred.detach().cpu()
# compute number of correct predictions in the batch
test_accuracy = metrics.calculate_accuracy(output, target)
test_acc += test_accuracy
iou_inds = metrics.calculate_iou(pred, target)
for key in iou_inds:
if key not in test_iou:
test_iou[key] = iou_inds[key]
else:
test_iou[key] += iou_inds[key]
test_loss = test_loss / (len(test_loader.dataset) / batch_size)
test_acc = 100 * (test_acc / (len(test_loader.dataset) / batch_size))
test_iou = metrics.convert_batched_iou(
test_iou, (len(test_loader.dataset) / batch_size))
mIoU = metrics.get_mIoU(test_iou)
mIoU_desc = metrics.miou_to_string(test_iou)
return test_loss, test_acc, mIoU, mIoU_desc
def val_one_epoch(val_loader, net):
net.eval()
device = params['device']
batch_size = params['batch_size']
val_loss = 0
val_acc = 0
val_iou = {}
pred = 0
with torch.no_grad():
for batch_index, (img, target) in enumerate(val_loader):
img, target = img.to(device), target.to(device)
output = net(img)
target = target.long()
loss = criterion(output, target).item()
val_loss += loss
pred = aux.get_predicted_image(output)
# Desvinculamos el valor de los nuevos targets y los pasamos a CPU para calcular las métricas
output, target, pred = output.detach().cpu(), target.detach(
).cpu(), pred.detach().cpu()
# compute number of correct predictions in the batch
val_accuracy = metrics.calculate_accuracy(output, target)
val_acc += val_accuracy
iou_inds = metrics.calculate_iou(pred, target)
for key in iou_inds:
if key not in val_iou:
val_iou[key] = iou_inds[key]
else:
val_iou[key] += iou_inds[key]
#print('Batch index: {}, loss: {}, accuracy: {:.2f}%'.format(batch_index, loss, val_accuracy * 100))
# Average acc across all correct predictions batches now
val_loss = val_loss / (len(val_loader.dataset) / batch_size)
val_acc = 100 * (val_acc / (len(val_loader.dataset) / batch_size))
val_iou = metrics.convert_batched_iou(
val_iou, (len(val_loader.dataset) / batch_size))
mIoU = metrics.get_mIoU(val_iou)
#print('\nValidation set: Average loss: {:.4f}, Accuracy: {:.0f}%, mIoU: {:.4f}\n'.format(val_loss, val_acc, mIoU))
mIoU_desc = metrics.miou_to_string(val_iou)
return val_loss, val_acc, mIoU, mIoU_desc
def train_one_epoch(train_loader, net, optimizer, criterion, hparams):
# Activate the train=True flag inside the model
net.train()
device = hparams['device']
batch_size = hparams['batch_size']
train_loss, train_accs = 0, 0
train_iou = {}
times_per_step_iteration = []
times_per_metric_iteration = []
times_per_iteration = []
for batch_index, (img, target) in enumerate(train_loader):
#Arrancamos temporizador general
start_total.record()
img, target = img.to(device), target.to(device)
optimizer.zero_grad()
# Arrancamos temporizador para inferencia
start.record()
output = net(img)
target = target.long()
loss = criterion(output, target)
loss.backward()
optimizer.step()
pred = aux.get_predicted_image(output)
#Paramos temporizador de inferencia
end.record()
torch.cuda.synchronize()
times_per_step_iteration.append(start.elapsed_time(end))
# Accuracy
#Arrancamos temporizador para métricas
start.record()
# Desvinculamos el valor de los nuevos targets y los pasamos a CPU para calcular las métricas
output, target, pred = output.detach().cpu(), target.detach().cpu(
), pred.detach().cpu()
train_loss += loss.item()
# Devuelve values, indices. Los indices son el nº de feature map (clase) en la que se encuentra el valor más alto en el pixel
train_accuracy = metrics.calculate_accuracy(output,
target) #, predicted
train_accs += train_accuracy
iou_inds = metrics.calculate_iou(pred, target)
for key in iou_inds:
if key not in train_iou:
train_iou[key] = iou_inds[key]
else:
train_iou[key] += iou_inds[key]
#Paramos temporizador para métricas
end.record()
torch.cuda.synchronize()
times_per_metric_iteration.append(start.elapsed_time(end))
#Paramos temporizador general
end_total.record()
torch.cuda.synchronize()
times_per_iteration.append(start_total.elapsed_time(end))
avg_time_taken = sum(times_per_iteration) / len(
times_per_iteration)
avg_time_step_taken = sum(times_per_step_iteration) / len(
times_per_step_iteration)
avg_time_metrics_taken = sum(times_per_metric_iteration) / len(
times_per_metric_iteration)
print('Average Time spent total: {:.02f}s'.format(avg_time_taken *
1e-3))
print('Average Time spent by steps: {:.02f}s'.format(
avg_time_step_taken * 1e-3))
print('Average Time spent by metrics: {:.02f}s'.format(
avg_time_metrics_taken * 1e-3))
print('Average Time spent by data load: {:.02f}s'.format(
avg_time_taken * 1e-3 - avg_time_step_taken * 1e-3 -
avg_time_metrics_taken * 1e-3))
train_loss = train_loss / (len(train_loader.dataset) / batch_size)
train_accs = 100 * (train_accs /
(len(train_loader.dataset) / batch_size))
train_iou = metrics.convert_batched_iou(
train_iou, (len(train_loader.dataset) / batch_size))
mIoU = metrics.get_mIoU(train_iou)
mIoU_desc = metrics.miou_to_string(train_iou)
return train_loss, train_accs, mIoU, mIoU_desc