def validation(validation_data_loader: torch.utils.data.DataLoader,
model: SupervisedModel, local_rank: int) -> tuple:
"""
:param validation_data_loader: Validation data loader
:param model: ResNet based classifier.
:param local_rank: local rank.
:return: validation loss, the number of corrected samples, and the size of samples on a local
"""
model.eval()
sum_loss = torch.tensor([0.]).to(local_rank)
num_corrects = torch.tensor([0.]).to(local_rank)
with torch.no_grad():
for data, targets in validation_data_loader:
data, targets = data.to(local_rank), targets.to(local_rank)
unnormalized_features = model(data)
loss = torch.nn.functional.cross_entropy(unnormalized_features,
targets,
reduction="sum")
predicted = torch.max(unnormalized_features.data, 1)[1]
sum_loss += loss.item()
num_corrects += (predicted == targets).sum()
return sum_loss, num_corrects
batch_size=64,
num_batches=31000)
# Create object to local contrast normalize a batch.
# Note: Every batch must be normalized before use.
normer = util.Normer3(filter_size=5, num_channels=1)
module_list = [normer]
preprocessor = util.Preprocessor(module_list)
print('Training Model')
for x_batch, y_batch in train_iterator:
#x_batch = preprocessor.run(x_batch)
x_batch = (x_batch - mean) / std
# loop over batch
for i in range(len(x_batch)):
# hide patch for an image
x_batch[i] = hide_patch(x_batch[i])
monitor.start()
log_prob, accuracy = model.train(x_batch, y_batch)
monitor.stop(1 - accuracy) # monitor takes error instead of accuracy
if monitor.test:
monitor.start()
x_val_batch, y_val_batch = val_iterator.next()
#x_val_batch = preprocessor.run(x_val_batch)
x_val_batch = (x_val_batch - mean) / std
val_accuracy = model.eval(x_val_batch, y_val_batch)
monitor.stop_test(1 - val_accuracy)
X_test = numpy.float32(X_test)
X_test /= 255.0
X_test *= 2.0
train_dataset = supervised_dataset.SupervisedDataset(X_train, y_train)
val_dataset = supervised_dataset.SupervisedDataset(X_val, y_val)
train_iterator = train_dataset.iterator(
mode='random_uniform', batch_size=64, num_batches=31000)
val_iterator = val_dataset.iterator(
mode='random_uniform', batch_size=64, num_batches=31000)
# Create object to local contrast normalize a batch.
# Note: Every batch must be normalized before use.
normer = util.Normer3(filter_size=5, num_channels=1)
module_list = [normer]
preprocessor = util.Preprocessor(module_list)
print('Training Model')
for x_batch, y_batch in train_iterator:
x_batch = preprocessor.run(x_batch)
monitor.start()
log_prob, accuracy = model.train(x_batch, y_batch)
monitor.stop(1-accuracy)
if monitor.test:
monitor.start()
x_val_batch, y_val_batch = val_iterator.next()
x_val_batch = preprocessor.run(x_val_batch)
val_accuracy = model.eval(x_val_batch, y_val_batch)
monitor.stop_test(1-val_accuracy)
y_test = test_data_container.y
train_dataset = supervised_dataset.SupervisedDataset(X_train, y_train)
test_dataset = supervised_dataset.SupervisedDataset(X_test, y_test)
train_iterator = train_dataset.iterator(mode='random_uniform',
batch_size=64,
num_batches=31000)
test_iterator = test_dataset.iterator(mode='random_uniform',
batch_size=64,
num_batches=31000)
# Create object to local contrast normalize a batch.
# Note: Every batch must be normalized before use.
normer = util.Normer3(filter_size=5, num_channels=1)
module_list = [normer]
preprocessor = util.Preprocessor(module_list)
print('Training Model')
for x_batch, y_batch in train_iterator:
x_batch = preprocessor.run(x_batch)
monitor.start()
log_prob, accuracy = model.train(x_batch, y_batch)
monitor.stop(1 - accuracy)
if monitor.test:
monitor.start()
x_test_batch, y_test_batch = test_iterator.next()
x_test_batch = preprocessor.run(x_test_batch)
test_accuracy = model.eval(x_test_batch, y_test_batch)
monitor.stop_test(1 - test_accuracy)
X_test /= 255.0
X_test *= 2.0
y_test = test_data_container.y
train_dataset = supervised_dataset.SupervisedDataset(X_train, y_train)
test_dataset = supervised_dataset.SupervisedDataset(X_test, y_test)
train_iterator = train_dataset.iterator(
mode='random_uniform', batch_size=64, num_batches=31000)
test_iterator = test_dataset.iterator(
mode='random_uniform', batch_size=64, num_batches=31000)
# Create object to local contrast normalize a batch.
# Note: Every batch must be normalized before use.
normer = util.Normer3(filter_size=5, num_channels=1)
module_list = [normer]
preprocessor = util.Preprocessor(module_list)
print('Training Model')
for x_batch, y_batch in train_iterator:
x_batch = preprocessor.run(x_batch)
monitor.start()
log_prob, accuracy = model.train(x_batch, y_batch)
monitor.stop(1-accuracy)
if monitor.test:
monitor.start()
x_test_batch, y_test_batch = test_iterator.next()
x_test_batch = preprocessor.run(x_test_batch)
test_accuracy = model.eval(x_test_batch, y_test_batch)
monitor.stop_test(1-test_accuracy)
