def train_single_with_warmup(train_dir, test_dir, network=build_rgb_cnn, num_epochs=400, flip=200, input_size=45,
learning_rate=0.005, gray=False, augment=True):
assert (num_epochs > flip)
train_images = load_images(train_dir, is_train=True, permute=False)
training_labels = list([img.label for img in train_images])
classes_set = list(sorted(set(training_labels)))
class_to_index = {key: index for index, key in enumerate(classes_set)}
print('Building network...')
input_var = T.tensor4('inputs')
target_var = T.ivector('targets')
neural_network = network(input_size, input_var)
print_network(neural_network)
train_fn, val_fn, predict_fn = build_network(neural_network, input_var, target_var, learning_rate=learning_rate)
if flip > 0:
# Create validation and training set
(trainset, valset) = split_special(train_images, 2, True)[0] # Use the old validation way
if augment:
print("Augmenting trainset images")
trainset = augmentation(trainset)
print("Augmented to %d images" % len(trainset))
print('Post processing train and validation set')
postprocess(trainset, size=input_size, grayscale=gray)
postprocess(valset, size=input_size, grayscale=gray)
x_train = images_to_vectors(trainset, input_size, num_dimensions=1 if gray else 3)
x_val = images_to_vectors(valset, input_size, num_dimensions=1 if gray else 3)
for img in train_images:
img.disposeImage()
y_train = np.concatenate(np.array([[class_to_index[img.label] for img in trainset]], dtype=np.uint8))
y_val = np.concatenate(np.array([[class_to_index[img.label] for img in valset]], dtype=np.uint8))
print('Training %d iterations as a warmup' % flip)
train(train_fn, val_fn, predict_fn, x_train, y_train, x_val, y_val, flip, show_validation=True)
# Flip or fully train
train_images = load_images(train_dir, is_train=True, permute=False)
train_images = augmentation(train_images)
postprocess(train_images, input_size)
x_train = images_to_vectors(train_images, input_size)
y_train = np.concatenate(np.array([[class_to_index[img.label] for img in train_images]], dtype=np.uint8))
print('Training %d iterations on full set' % (num_epochs - flip))
train(train_fn, val_fn, predict_fn, x_train, y_train,
None if flip <= 0 else x_val, None if flip <= 0 else y_val,
num_epochs if flip <= 0 else num_epochs - flip,
show_validation=False)
# Predict
print('Predicting testset')
test_images = sorted(load_images(test_dir, is_train=False, permute=False), key=lambda x: x.identifier)
postprocess(test_images, size=input_size, grayscale=gray)
x_test = images_to_vectors(test_images, input_size, num_dimensions=1 if gray else 3)
predictions = predict(predict_fn, x_test)
print('Writing to CSV...')
write_csv(test_images, predictions, classes_set, filename='result_warmup.csv')
print('Finished.')
def train_single_with_warmup(train_dir,
test_dir,
network=build_rgb_cnn,
num_epochs=400,
flip=200,
input_size=45,
learning_rate=0.005,
gray=False,
augment=True):
assert (num_epochs > flip)
train_images = load_images(train_dir, is_train=True, permute=False)
training_labels = list([img.label for img in train_images])
classes_set = list(sorted(set(training_labels)))
class_to_index = {key: index for index, key in enumerate(classes_set)}
print('Building network...')
input_var = T.tensor4('inputs')
target_var = T.ivector('targets')
neural_network = network(input_size, input_var)
print_network(neural_network)
train_fn, val_fn, predict_fn = build_network(neural_network,
input_var,
target_var,
learning_rate=learning_rate)
if flip > 0:
# Create validation and training set
(trainset,
valset) = split_special(train_images, 2,
True)[0] # Use the old validation way
if augment:
print("Augmenting trainset images")
trainset = augmentation(trainset)
print("Augmented to %d images" % len(trainset))
print('Post processing train and validation set')
postprocess(trainset, size=input_size, grayscale=gray)
postprocess(valset, size=input_size, grayscale=gray)
x_train = images_to_vectors(trainset,
input_size,
num_dimensions=1 if gray else 3)
x_val = images_to_vectors(valset,
input_size,
num_dimensions=1 if gray else 3)
for img in train_images:
img.disposeImage()
y_train = np.concatenate(
np.array([[class_to_index[img.label] for img in trainset]],
dtype=np.uint8))
y_val = np.concatenate(
np.array([[class_to_index[img.label] for img in valset]],
dtype=np.uint8))
print('Training %d iterations as a warmup' % flip)
train(train_fn,
val_fn,
predict_fn,
x_train,
y_train,
x_val,
y_val,
flip,
show_validation=True)
# Flip or fully train
train_images = load_images(train_dir, is_train=True, permute=False)
train_images = augmentation(train_images)
postprocess(train_images, input_size)
x_train = images_to_vectors(train_images, input_size)
y_train = np.concatenate(
np.array([[class_to_index[img.label] for img in train_images]],
dtype=np.uint8))
print('Training %d iterations on full set' % (num_epochs - flip))
train(train_fn,
val_fn,
predict_fn,
x_train,
y_train,
None if flip <= 0 else x_val,
None if flip <= 0 else y_val,
num_epochs if flip <= 0 else num_epochs - flip,
show_validation=False)
# Predict
print('Predicting testset')
test_images = sorted(load_images(test_dir, is_train=False, permute=False),
key=lambda x: x.identifier)
postprocess(test_images, size=input_size, grayscale=gray)
x_test = images_to_vectors(test_images,
input_size,
num_dimensions=1 if gray else 3)
predictions = predict(predict_fn, x_test)
print('Writing to CSV...')
write_csv(test_images,
predictions,
classes_set,
filename='result_warmup.csv')
print('Finished.')
def cross_validate_neural(train_dir, networks, weights, epochs, input_sizes, learning_rates, grays, num_folds=5, augment=True):
print('Cross-validation using %d folds' % num_folds)
if isinstance(weights, tuple) and len(weights) > 1:
print('Using weights %s' % str.join(',', [str(x) for x in weights]))
train_images = load_images(train_dir, is_train=True, permute=False)
training_labels = list([img.label for img in train_images])
classes_set = list(sorted(set(training_labels)))
class_to_index = {key: index for index, key in enumerate(classes_set)}
if not isinstance(networks, tuple):
networks = (networks)
if not isinstance(weights, tuple):
weights = (weights)
if not isinstance(epochs, tuple):
epochs = (epochs)
if not isinstance(input_sizes, tuple):
input_sizes = (input_sizes)
if not isinstance(learning_rates, tuple):
learning_rates = (learning_rates)
if not isinstance(grays, tuple):
grays = (grays)
# Create validation and training set
val_losses = []
train_losses = []
val_accs = []
for trainset, valset in split_special(train_images, num_folds, num_folds == 2):
# For each network in ensemble (if available)
predictions = []
print('Evaluating fold... (%d models in ensemble)' % len(weights))
if augment:
trainset = augmentation(trainset)
print("Augmented to %d images" % len(trainset))
# Convert classes to uin8 indices
y_train = np.concatenate(np.array([[class_to_index[img.label] for img in trainset]], dtype=np.uint8))
y_val = np.concatenate(np.array([[class_to_index[img.label] for img in valset]], dtype=np.uint8))
for input_size, network, num_epochs, gray, learning_rate in zip(input_sizes, networks, epochs,
grays, learning_rates):
# Postprocess images
postprocess(trainset, size=input_size, grayscale=gray)
postprocess(valset, size=input_size, grayscale=gray)
# Get their numpy vectors to feed to lasagne
x_train = images_to_vectors(trainset, input_size, num_dimensions=1 if gray else 3)
x_val = images_to_vectors(valset, input_size, num_dimensions=1 if gray else 3)
# Prepare theano tensors for classes and inputs
input_var = T.tensor4('inputs')
target_var = T.ivector('targets')
neural_network = network(input_size, input_var)
train_fn, val_fn, predict_fn = build_network(neural_network, input_var, target_var)
training_loss, val_loss, val_acc = train(train_fn, val_fn, predict_fn, x_train, y_train, x_val, y_val,
num_epochs)
if len(weights) <= 1:
val_losses.append(val_loss)
train_losses.append(training_loss)
val_accs.append(val_acc)
else:
weighted_logloss = predict(predict_fn, x_val)
print('Weighted logloss of fold: %f' % weighted_logloss)
predictions.append(weighted_logloss)
# Calculate weighted logloss for this fold
if len(weights) > 1:
val_losses.append(logloss(predictions, weights, y_val, num_classes=81))
# Process metrics
if len(weights) <= 1:
mean_val_loss = np.mean(val_losses)
mean_train_loss = np.mean(train_losses)
mean_accuracy = np.mean(val_accs)
std_val_loss = np.std(val_losses)
std_val_acc = np.std(val_accs)
std_train_loss = np.std(train_losses)
print('Mean validation loss: %f (std: %f)' % (mean_val_loss, std_val_loss))
print('Mean training loss: %f (std: %f)' % (mean_train_loss, std_train_loss))
print('Mean training accuracy: %f (std %f)' % (mean_accuracy, std_val_acc))
else:
print('Weighted logloss: %f (std %f)' %(np.mean(val_losses), np.std(val_losses)))
def cross_validate_neural(train_dir,
networks,
weights,
epochs,
input_sizes,
learning_rates,
grays,
num_folds=5,
augment=True):
print('Cross-validation using %d folds' % num_folds)
if isinstance(weights, tuple) and len(weights) > 1:
print('Using weights %s' % str.join(',', [str(x) for x in weights]))
train_images = load_images(train_dir, is_train=True, permute=False)
training_labels = list([img.label for img in train_images])
classes_set = list(sorted(set(training_labels)))
class_to_index = {key: index for index, key in enumerate(classes_set)}
if not isinstance(networks, tuple):
networks = (networks)
if not isinstance(weights, tuple):
weights = (weights)
if not isinstance(epochs, tuple):
epochs = (epochs)
if not isinstance(input_sizes, tuple):
input_sizes = (input_sizes)
if not isinstance(learning_rates, tuple):
learning_rates = (learning_rates)
if not isinstance(grays, tuple):
grays = (grays)
# Create validation and training set
val_losses = []
train_losses = []
val_accs = []
for trainset, valset in split_special(train_images, num_folds,
num_folds == 2):
# For each network in ensemble (if available)
predictions = []
print('Evaluating fold... (%d models in ensemble)' % len(weights))
if augment:
trainset = augmentation(trainset)
print("Augmented to %d images" % len(trainset))
# Convert classes to uin8 indices
y_train = np.concatenate(
np.array([[class_to_index[img.label] for img in trainset]],
dtype=np.uint8))
y_val = np.concatenate(
np.array([[class_to_index[img.label] for img in valset]],
dtype=np.uint8))
for input_size, network, num_epochs, gray, learning_rate in zip(
input_sizes, networks, epochs, grays, learning_rates):
# Postprocess images
postprocess(trainset, size=input_size, grayscale=gray)
postprocess(valset, size=input_size, grayscale=gray)
# Get their numpy vectors to feed to lasagne
x_train = images_to_vectors(trainset,
input_size,
num_dimensions=1 if gray else 3)
x_val = images_to_vectors(valset,
input_size,
num_dimensions=1 if gray else 3)
# Prepare theano tensors for classes and inputs
input_var = T.tensor4('inputs')
target_var = T.ivector('targets')
neural_network = network(input_size, input_var)
train_fn, val_fn, predict_fn = build_network(
neural_network, input_var, target_var)
training_loss, val_loss, val_acc = train(train_fn, val_fn,
predict_fn, x_train,
y_train, x_val, y_val,
num_epochs)
if len(weights) <= 1:
val_losses.append(val_loss)
train_losses.append(training_loss)
val_accs.append(val_acc)
else:
weighted_logloss = predict(predict_fn, x_val)
print('Weighted logloss of fold: %f' % weighted_logloss)
predictions.append(weighted_logloss)
# Calculate weighted logloss for this fold
if len(weights) > 1:
val_losses.append(
logloss(predictions, weights, y_val, num_classes=81))
# Process metrics
if len(weights) <= 1:
mean_val_loss = np.mean(val_losses)
mean_train_loss = np.mean(train_losses)
mean_accuracy = np.mean(val_accs)
std_val_loss = np.std(val_losses)
std_val_acc = np.std(val_accs)
std_train_loss = np.std(train_losses)
print('Mean validation loss: %f (std: %f)' %
(mean_val_loss, std_val_loss))
print('Mean training loss: %f (std: %f)' %
(mean_train_loss, std_train_loss))
print('Mean training accuracy: %f (std %f)' %
(mean_accuracy, std_val_acc))
else:
print('Weighted logloss: %f (std %f)' %
(np.mean(val_losses), np.std(val_losses)))