from texture_net import TextureNet
network = TextureNet()
#Loss function
cross_entropy = nn.CrossEntropyLoss() #Softmax function is included
#Optimizer to control step size in gradient descent
optimizer = torch.optim.Adam(network.parameters())
#Transfer model to gpu
if use_gpu:
network = network.cuda()
#Load the data cube and labels
data, data_info = readSEGY(join(dataset_name,'data.segy'))
train_class_imgs, train_coordinates = readLabels(join(dataset_name,'train'), data_info)
val_class_imgs, _ = readLabels(join(dataset_name,'val'), data_info)
#Plot training/validation data with labels
if log_tensorboard:
for class_img in train_class_imgs + val_class_imgs:
logger.log_images(class_img[1] + '_' + str(class_img[2] ), get_slice(data, data_info, class_img[1], class_img[2]), cm='gray')
logger.log_images(class_img[1] + '_' + str(class_img[2]) + '_true_class', class_img[0])
# Training loop
for i in range(2000):
# Get random training batch with augmentation
# This is the bottle-neck for training and could be done more efficient on the GPU...
[batch, labels] = get_random_batch(data, train_coordinates, im_size, batch_size,
from texture_net import TextureNet
network = TextureNet()
#Loss function
cross_entropy = nn.CrossEntropyLoss() #Softmax function is included
#Optimizer to control step size in gradient descent
optimizer = torch.optim.Adam(network.parameters())
#Transfer model to gpu
if use_gpu:
network = network.cuda()
#Load the data cube and labels
data, data_info = readSEGY(dataset_name + '/data.segy')
train_class_imgs, train_coordinates = readLabels(dataset_name + '/train/',
data_info)
val_class_imgs, _ = readLabels(dataset_name + '/val/', data_info)
#Plot training/validation data with labels
if log_tensorboard:
for class_img in train_class_imgs + val_class_imgs:
logger.log_images(class_img[1] + '_' + str(class_img[2]),
get_slice(data, data_info, class_img[1],
class_img[2]),
cm='gray')
logger.log_images(
class_img[1] + '_' + str(class_img[2]) + '_true_class',
class_img[0])
# Training loop
for i in range(2000):
logger.log_images('flipping', batch)
[batch, labels] = get_random_batch(data,
train_coordinates,
65,
32,
random_stretch=.50)
logger.log_images('stretching', batch)
[batch, labels] = get_random_batch(data,
train_coordinates,
65,
32,
random_rot_xy=180)
logger.log_images('rot', batch)
[batch, labels] = get_random_batch(data,
train_coordinates,
65,
32,
random_rot_z=15)
logger.log_images('dip', batch)
train_cls_imgs, train_coordinates = readLabels(join('F3', 'train'),
data_info)
[batch, labels] = get_random_batch(data, train_coordinates, 65, 32)
logger.log_images('salt', batch[:16, :, :, :, :])
logger.log_images('not salt', batch[16:, :, :, :, :])
logger.log_images('data', data[:, :, 50])
random_flip=True)
logger.log_images('flipping', batch)
[batch, labels] = get_random_batch(data,
train_coordinates,
65,
32,
random_stretch=.50)
logger.log_images('stretching', batch)
[batch, labels] = get_random_batch(data,
train_coordinates,
65,
32,
random_rot_xy=180)
logger.log_images('rot', batch)
[batch, labels] = get_random_batch(data,
train_coordinates,
65,
32,
random_rot_z=15)
logger.log_images('dip', batch)
train_cls_imgs, train_coordinates = readLabels('F3/train/', data_info)
[batch, labels] = get_random_batch(data, train_coordinates, 65, 32)
logger.log_images('salt', batch[:16, :, :, :, :])
logger.log_images('not salt', batch[16:, :, :, :, :])
logger.log_images('data', data[:, :, 50])