def test_early_stopping(mock_train_history):
from nolearn_utils.hooks import EarlyStopping
early_stopper = EarlyStopping(patience=3)
train_history = mock_train_history(valid_loss=[0.7, 0.6, 0.6, 0.6, 0.6],
train_loss=[0.3, 0.3, 0.3, 0.3, 0.3],
epoch=[1, 2, 3, 4, 5])
net = MagicMock()
net.get_all_param = MagicMock()
early_stopper(net, train_history[:1])
early_stopper(net, train_history[:2])
early_stopper(net, train_history[:3])
early_stopper(net, train_history[:4])
with pytest.raises(StopIteration):
early_stopper(net, train_history[:5])
def cascade_model(options):
"""
3D cascade model using Nolearn and Lasagne
Inputs:
- model_options:
- weights_path: path to where weights should be saved
Output:
- nets = list of NeuralNets (CNN1, CNN2)
"""
# model options
channels = len(options['modalities'])
train_split_perc = options['train_split']
num_epochs = options['max_epochs']
max_epochs_patience = options['patience']
# save model to disk to re-use it. Create an experiment folder
# organize experiment
if not os.path.exists(
os.path.join(options['weight_paths'], options['experiment'])):
os.mkdir(os.path.join(options['weight_paths'], options['experiment']))
if not os.path.exists(
os.path.join(options['weight_paths'], options['experiment'],
'nets')):
os.mkdir(
os.path.join(options['weight_paths'], options['experiment'],
'nets'))
# --------------------------------------------------
# first model
# --------------------------------------------------
layer1 = InputLayer(name='in1',
shape=(None, channels) + options['patch_size'])
layer1 = batch_norm(Conv3DLayer(layer1,
name='conv1_1',
num_filters=32,
filter_size=3,
pad='same'),
name='BN1')
layer1 = Pool3DLayer(layer1,
name='avgpool_1',
mode='max',
pool_size=2,
stride=2)
layer1 = batch_norm(Conv3DLayer(layer1,
name='conv2_1',
num_filters=64,
filter_size=3,
pad='same'),
name='BN2')
layer1 = Pool3DLayer(layer1,
name='avgpoo2_1',
mode='max',
pool_size=2,
stride=2)
layer1 = DropoutLayer(layer1, name='l2drop', p=0.5)
layer1 = DenseLayer(layer1, name='d_1', num_units=256)
layer1 = DenseLayer(layer1,
name='out',
num_units=2,
nonlinearity=nonlinearities.softmax)
# save weights
net_model = 'model_1'
net_weights = os.path.join(options['weight_paths'], options['experiment'],
'nets', net_model + '.pkl')
net_history = os.path.join(options['weight_paths'], options['experiment'],
'nets', net_model + '_history.pkl')
net1 = NeuralNet(
layers=layer1,
objective_loss_function=objectives.categorical_crossentropy,
batch_iterator_train=Rotate_batch_Iterator(batch_size=128),
update=updates.adadelta,
on_epoch_finished=[
SaveWeights(net_weights, only_best=True, pickle=False),
SaveTrainingHistory(net_history),
EarlyStopping(patience=max_epochs_patience)
],
verbose=options['net_verbose'],
max_epochs=num_epochs,
train_split=TrainSplit(eval_size=train_split_perc),
)
# --------------------------------------------------
# second model
# --------------------------------------------------
layer2 = InputLayer(name='in2',
shape=(None, channels) + options['patch_size'])
layer2 = batch_norm(Conv3DLayer(layer2,
name='conv1_1',
num_filters=32,
filter_size=3,
pad='same'),
name='BN1')
layer2 = Pool3DLayer(layer2,
name='avgpool_1',
mode='max',
pool_size=2,
stride=2)
layer2 = batch_norm(Conv3DLayer(layer2,
name='conv2_1',
num_filters=64,
filter_size=3,
pad='same'),
name='BN2')
layer2 = Pool3DLayer(layer2,
name='avgpoo2_1',
mode='max',
pool_size=2,
stride=2)
layer2 = DropoutLayer(layer2, name='l2drop', p=0.5)
layer2 = DenseLayer(layer2, name='d_1', num_units=256)
layer2 = DenseLayer(layer2,
name='out',
num_units=2,
nonlinearity=nonlinearities.softmax)
# save weights
net_model = 'model_2'
net_weights2 = os.path.join(options['weight_paths'], options['experiment'],
'nets', net_model + '.pkl')
net_history2 = os.path.join(options['weight_paths'], options['experiment'],
'nets', net_model + '_history.pkl')
net2 = NeuralNet(
layers=layer2,
objective_loss_function=objectives.categorical_crossentropy,
batch_iterator_train=Rotate_batch_Iterator(batch_size=128),
update=updates.adadelta,
on_epoch_finished=[
SaveWeights(net_weights2, only_best=True, pickle=False),
SaveTrainingHistory(net_history2),
EarlyStopping(patience=max_epochs_patience)
],
verbose=options['net_verbose'],
max_epochs=num_epochs,
train_split=TrainSplit(eval_size=train_split_perc),
)
return [net1, net2]
affine_p=1.,
affine_scale_choices=np.linspace(0.5, 1.5, 11),
# affine_shear_choices=np.linspace(-0.5, 0.5, 11),
affine_translation_choices=np.arange(-64, 64, 1),
# affine_rotation_choices=np.arange(0, 360, 1),
adjust_gamma_p=0.5,
adjust_gamma_chocies=np.linspace(0.5, 1.5, 11))
train_iterator = TrainIterator(**train_iterator_kwargs)
test_iterator_kwargs = dict(batch_size=batch_size, )
test_iterator = TestIterator(**test_iterator_kwargs)
save_weights = SaveWeights(model_fname, only_best=True, pickle=False)
save_training_history = SaveTrainingHistory(model_history_fname)
plot_training_history = PlotTrainingHistory(model_graph_fname)
early_stopping = EarlyStopping(patience=100)
conv_kwargs = dict(pad='same',
nonlinearity=nn.nonlinearities.very_leaky_rectify)
pool_kwargs = dict(pool_size=2, )
l = nn.layers.InputLayer(name='in', shape=(None, 3, image_size, image_size))
# 256
l = conv2dbn(l,
name='l1c1',
num_filters=32,
filter_size=(7, 7),
stride=2,
**conv_kwargs)
test_iterator_kwargs = {
'batch_size': batch_size,
'buffer_size': 5,
}
test_iterator = TestIterator(**test_iterator_kwargs)
save_weights = SaveWeights('./examples/mnist/model_weights.pkl',
only_best=True,
pickle=False)
save_training_history = SaveTrainingHistory(
'./examples/mnist/model_history.pkl')
plot_training_history = PlotTrainingHistory(
'./examples/mnist/training_history.png')
early_stopping = EarlyStopping(metrics='valid_accuracy',
patience=100,
verbose=True,
higher_is_better=True)
net = NeuralNet(layers=[
(InputLayer, dict(name='in', shape=(None, 1, image_size, image_size))),
(Conv2DDNNLayer,
dict(name='l1c1', num_filters=32, filter_size=(3, 3), pad='same')),
(Conv2DDNNLayer,
dict(name='l1c2', num_filters=32, filter_size=(3, 3), pad='same')),
(MaxPool2DDNNLayer, dict(name='l1p', pool_size=3, stride=2)),
(Conv2DDNNLayer,
dict(name='l2c1', num_filters=32, filter_size=(3, 3), pad='same')),
(Conv2DDNNLayer,
dict(name='l2c2', num_filters=32, filter_size=(3, 3), pad='same')),
(MaxPool2DDNNLayer, dict(name='l2p', pool_size=3, stride=2)),
(DenseLayer, dict(name='l7', num_units=256)),
def build_model(weights_path, options):
"""
Build the CNN model. Create the Neural Net object and return it back.
Inputs:
- subject name: used to save the net weights accordingly.
- options: several hyper-parameters used to configure the net.
Output:
- net: a NeuralNet object
"""
net_model_name = options['experiment']
try:
os.mkdir(os.path.join(weights_path, net_model_name))
except:
pass
net_weights = os.path.join(weights_path, net_model_name,
net_model_name + '.pkl')
net_history = os.path.join(weights_path, net_model_name,
net_model_name + '_history.pkl')
# select hyper-parameters
t_verbose = options['net_verbose']
train_split_perc = options['train_split']
num_epochs = options['max_epochs']
max_epochs_patience = options['patience']
early_stopping = EarlyStopping(patience=max_epochs_patience)
save_weights = SaveWeights(net_weights, only_best=True, pickle=False)
save_training_history = SaveTrainingHistory(net_history)
# build the architecture
ps = options['patch_size'][0]
num_channels = 1
fc_conv = 180
fc_fc = 180
dropout_conv = 0.5
dropout_fc = 0.5
# --------------------------------------------------
# channel_1: axial
# --------------------------------------------------
axial_ch = InputLayer(name='in1', shape=(None, num_channels, ps, ps))
axial_ch = prelu(batch_norm(
Conv2DLayer(axial_ch,
name='axial_ch_conv1',
num_filters=20,
filter_size=3)),
name='axial_ch_prelu1')
axial_ch = prelu(batch_norm(
Conv2DLayer(axial_ch,
name='axial_ch_conv2',
num_filters=20,
filter_size=3)),
name='axial_ch_prelu2')
axial_ch = MaxPool2DLayer(axial_ch, name='axial_max_pool_1', pool_size=2)
axial_ch = prelu(batch_norm(
Conv2DLayer(axial_ch,
name='axial_ch_conv3',
num_filters=40,
filter_size=3)),
name='axial_ch_prelu3')
axial_ch = prelu(batch_norm(
Conv2DLayer(axial_ch,
name='axial_ch_conv4',
num_filters=40,
filter_size=3)),
name='axial_ch_prelu4')
axial_ch = MaxPool2DLayer(axial_ch, name='axial_max_pool_2', pool_size=2)
axial_ch = prelu(batch_norm(
Conv2DLayer(axial_ch,
name='axial_ch_conv5',
num_filters=60,
filter_size=3)),
name='axial_ch_prelu5')
axial_ch = DropoutLayer(axial_ch, name='axial_l1drop', p=dropout_conv)
axial_ch = DenseLayer(axial_ch, name='axial_d1', num_units=fc_conv)
axial_ch = prelu(axial_ch, name='axial_prelu_d1')
# --------------------------------------------------
# channel_1: coronal
# --------------------------------------------------
coronal_ch = InputLayer(name='in2', shape=(None, num_channels, ps, ps))
coronal_ch = prelu(batch_norm(
Conv2DLayer(coronal_ch,
name='coronal_ch_conv1',
num_filters=20,
filter_size=3)),
name='coronal_ch_prelu1')
coronal_ch = prelu(batch_norm(
Conv2DLayer(coronal_ch,
name='coronal_ch_conv2',
num_filters=20,
filter_size=3)),
name='coronal_ch_prelu2')
coronal_ch = MaxPool2DLayer(coronal_ch,
name='coronal_max_pool_1',
pool_size=2)
coronal_ch = prelu(batch_norm(
Conv2DLayer(coronal_ch,
name='coronal_ch_conv3',
num_filters=40,
filter_size=3)),
name='coronal_ch_prelu3')
coronal_ch = prelu(batch_norm(
Conv2DLayer(coronal_ch,
name='coronal_ch_conv4',
num_filters=40,
filter_size=3)),
name='coronal_ch_prelu4')
coronal_ch = MaxPool2DLayer(coronal_ch,
name='coronal_max_pool_2',
pool_size=2)
coronal_ch = prelu(batch_norm(
Conv2DLayer(coronal_ch,
name='coronal_ch_conv5',
num_filters=60,
filter_size=3)),
name='coronal_ch_prelu5')
coronal_ch = DropoutLayer(coronal_ch,
name='coronal_l1drop',
p=dropout_conv)
coronal_ch = DenseLayer(coronal_ch, name='coronal_d1', num_units=fc_conv)
coronal_ch = prelu(coronal_ch, name='coronal_prelu_d1')
# --------------------------------------------------
# channel_1: saggital
# --------------------------------------------------
saggital_ch = InputLayer(name='in3', shape=(None, num_channels, ps, ps))
saggital_ch = prelu(batch_norm(
Conv2DLayer(saggital_ch,
name='saggital_ch_conv1',
num_filters=20,
filter_size=3)),
name='saggital_ch_prelu1')
saggital_ch = prelu(batch_norm(
Conv2DLayer(saggital_ch,
name='saggital_ch_conv2',
num_filters=20,
filter_size=3)),
name='saggital_ch_prelu2')
saggital_ch = MaxPool2DLayer(saggital_ch,
name='saggital_max_pool_1',
pool_size=2)
saggital_ch = prelu(batch_norm(
Conv2DLayer(saggital_ch,
name='saggital_ch_conv3',
num_filters=40,
filter_size=3)),
name='saggital_ch_prelu3')
saggital_ch = prelu(batch_norm(
Conv2DLayer(saggital_ch,
name='saggital_ch_conv4',
num_filters=40,
filter_size=3)),
name='saggital_ch_prelu4')
saggital_ch = MaxPool2DLayer(saggital_ch,
name='saggital_max_pool_2',
pool_size=2)
saggital_ch = prelu(batch_norm(
Conv2DLayer(saggital_ch,
name='saggital_ch_conv5',
num_filters=60,
filter_size=3)),
name='saggital_ch_prelu5')
saggital_ch = DropoutLayer(saggital_ch,
name='saggital_l1drop',
p=dropout_conv)
saggital_ch = DenseLayer(saggital_ch,
name='saggital_d1',
num_units=fc_conv)
saggital_ch = prelu(saggital_ch, name='saggital_prelu_d1')
# FC layer 540
layer = ConcatLayer(name='elem_channels',
incomings=[axial_ch, coronal_ch, saggital_ch])
layer = DropoutLayer(layer, name='f1_drop', p=dropout_fc)
layer = DenseLayer(layer, name='FC1', num_units=540)
layer = prelu(layer, name='prelu_f1')
# concatenate channels 540 + 15
layer = DropoutLayer(layer, name='f2_drop', p=dropout_fc)
atlas_layer = DropoutLayer(InputLayer(name='in4', shape=(None, 15)),
name='Dropout_atlas',
p=.2)
atlas_layer = InputLayer(name='in4', shape=(None, 15))
layer = ConcatLayer(name='elem_channels2', incomings=[layer, atlas_layer])
# FC layer 270
layer = DenseLayer(layer, name='fc_2', num_units=270)
layer = prelu(layer, name='prelu_f2')
# FC output 15 (softmax)
net_layer = DenseLayer(layer,
name='out_layer',
num_units=15,
nonlinearity=softmax)
net = NeuralNet(
layers=net_layer,
objective_loss_function=objectives.categorical_crossentropy,
update=updates.adam,
update_learning_rate=0.001,
on_epoch_finished=[
save_weights,
save_training_history,
early_stopping,
],
verbose=t_verbose,
max_epochs=num_epochs,
train_split=TrainSplit(eval_size=train_split_perc),
)
if options['load_weights'] == 'True':
try:
print " --> loading weights from ", net_weights
net.load_params_from(net_weights)
except:
pass
return net
def main():
c = color_codes()
patch_size = (15, 15, 15)
dir_name = '/home/sergivalverde/w/CNN/images/CH16'
patients = [
f for f in sorted(os.listdir(dir_name))
if os.path.isdir(os.path.join(dir_name, f))
]
names = np.stack([
name for name in
[[
os.path.join(dir_name, patient, 'FLAIR_preprocessed.nii.gz')
for patient in patients
],
[
os.path.join(dir_name, patient, 'DP_preprocessed.nii.gz')
for patient in patients
],
[
os.path.join(dir_name, patient, 'T2_preprocessed.nii.gz')
for patient in patients
],
[
os.path.join(dir_name, patient, 'T1_preprocessed.nii.gz')
for patient in patients
]] if name is not None
],
axis=1)
seed = np.random.randint(np.iinfo(np.int32).max)
''' Here we create an initial net to find conflictive voxels '''
print(c['c'] + '[' + strftime("%H:%M:%S") + '] ' + c['g'] +
'<Running iteration ' + c['b'] + '1>' + c['nc'])
net_name = '/home/sergivalverde/w/CNN/code/CNN1/miccai_challenge2016/deep-challenge2016.init.'
net = NeuralNet(
layers=[
(InputLayer, dict(name='in', shape=(None, 4, 15, 15, 15))),
(Conv3DDNNLayer,
dict(name='conv1_1',
num_filters=32,
filter_size=(5, 5, 5),
pad='same')),
(Pool3DDNNLayer,
dict(name='avgpool_1',
pool_size=2,
stride=2,
mode='average_inc_pad')),
(Conv3DDNNLayer,
dict(name='conv2_1',
num_filters=64,
filter_size=(5, 5, 5),
pad='same')),
(Pool3DDNNLayer,
dict(name='avgpool_2',
pool_size=2,
stride=2,
mode='average_inc_pad')),
(DropoutLayer, dict(name='l2drop', p=0.5)),
(DenseLayer, dict(name='l1', num_units=256)),
(DenseLayer,
dict(name='out', num_units=2,
nonlinearity=nonlinearities.softmax)),
],
objective_loss_function=objectives.categorical_crossentropy,
update=updates.adam,
update_learning_rate=0.0001,
on_epoch_finished=[
SaveWeights(net_name + 'model_weights.pkl',
only_best=True,
pickle=False),
SaveTrainingHistory(net_name + 'model_history.pkl'),
PlotTrainingHistory(net_name + 'training_history.png'),
EarlyStopping(patience=10)
],
verbose=10,
max_epochs=50,
train_split=TrainSplit(eval_size=0.25),
custom_scores=[('dsc', lambda p, t: 2 * np.sum(p * t[:, 1]) / np.sum(
(p + t[:, 1])))],
)
try:
net.load_params_from(net_name + 'model_weights.pkl')
except IOError:
print(c['c'] + '[' + strftime("%H:%M:%S") + '] ' + c['g'] +
'Loading the data for ' + c['b'] + 'iteration 1' + c['nc'])
# Create the data
(x, y, _) = load_patches(dir_name=dir_name,
use_flair=True,
use_pd=True,
use_t2=True,
use_t1=True,
use_gado=False,
flair_name='FLAIR_preprocessed.nii.gz',
pd_name='DP_preprocessed.nii.gz',
t2_name='T2_preprocessed.nii.gz',
t1_name='T1_preprocessed.nii.gz',
gado_name=None,
mask_name='Consensus.nii.gz',
size=patch_size)
print('-- Permuting the data')
np.random.seed(seed)
x_train = np.random.permutation(
np.concatenate(x).astype(dtype=np.float32))
print('-- Permuting the labels')
np.random.seed(seed)
y_train = np.random.permutation(
np.concatenate(y).astype(dtype=np.int32))
y_train = y_train[:, y_train.shape[1] / 2 + 1,
y_train.shape[2] / 2 + 1, y_train.shape[3] / 2 + 1]
print('-- Training vector shape = (' +
','.join([str(length) for length in x_train.shape]) + ')')
print('-- Training labels shape = (' +
','.join([str(length) for length in y_train.shape]) + ')')
print c['c'] + '[' + strftime("%H:%M:%S") + '] ' + c['g'] +\
'Training (' + c['b'] + 'initial' + c['nc'] + c['g'] + ')' + c['nc']
# We try to get the last weights to keep improving the net over and over
net.fit(x_train, y_train)
''' Here we get the seeds '''
print c['c'] + '[' + strftime(
"%H:%M:%S") + '] ' + c['g'] + '<Looking for seeds>' + c['nc']
for patient in names:
output_name = os.path.join('/'.join(patient[0].rsplit('/')[:-1]),
'test.iter1.nii.gz')
try:
load_nii(output_name)
print c['c'] + '[' + strftime("%H:%M:%S") + '] ' \
+ c['g'] + '-- Patient ' + patient[0].rsplit('/')[-2] + ' already done' + c['nc']
except IOError:
print c['c'] + '[' + strftime("%H:%M:%S") + '] '\
+ c['g'] + '-- Testing with patient ' + c['b'] + patient[0].rsplit('/')[-2] + c['nc']
image_nii = load_nii(patient[0])
image = np.zeros_like(image_nii.get_data())
for batch, centers in load_patch_batch(patient, 100000,
patch_size):
y_pred = net.predict_proba(batch)
[x, y, z] = np.stack(centers, axis=1)
image[x, y, z] = y_pred[:, 1]
print c['g'] + '-- Saving image ' + c['b'] + output_name + c['nc']
image_nii.get_data()[:] = image
image_nii.to_filename(output_name)
''' Here we perform the last iteration '''
print c['c'] + '[' + strftime("%H:%M:%S") + '] ' + c[
'g'] + '<Running iteration ' + c['b'] + '2>' + c['nc']
net_name = '/home/sergivalverde/w/CNN/code/CNN1/miccai_challenge2016/deep-challenge2016.final.'
net = NeuralNet(
layers=[
(InputLayer, dict(name='in', shape=(None, 4, 15, 15, 15))),
(Conv3DDNNLayer,
dict(name='conv1_1',
num_filters=32,
filter_size=(5, 5, 5),
pad='same')),
(Pool3DDNNLayer,
dict(name='avgpool_1',
pool_size=2,
stride=2,
mode='average_inc_pad')),
(Conv3DDNNLayer,
dict(name='conv2_1',
num_filters=64,
filter_size=(5, 5, 5),
pad='same')),
(Pool3DDNNLayer,
dict(name='avgpool_2',
pool_size=2,
stride=2,
mode='average_inc_pad')),
(DropoutLayer, dict(name='l2drop', p=0.5)),
(DenseLayer, dict(name='l1', num_units=256)),
(DenseLayer,
dict(name='out', num_units=2,
nonlinearity=nonlinearities.softmax)),
],
objective_loss_function=objectives.categorical_crossentropy,
update=updates.adam,
update_learning_rate=0.0001,
on_epoch_finished=[
SaveWeights(net_name + 'model_weights.pkl',
only_best=True,
pickle=False),
SaveTrainingHistory(net_name + 'model_history.pkl'),
PlotTrainingHistory(net_name + 'training_history.png'),
],
batch_iterator_train=BatchIterator(batch_size=4096),
verbose=10,
max_epochs=2000,
train_split=TrainSplit(eval_size=0.25),
custom_scores=[('dsc', lambda p, t: 2 * np.sum(p * t[:, 1]) / np.sum(
(p + t[:, 1])))],
)
try:
net.load_params_from(net_name + 'model_weights.pkl')
except IOError:
pass
print c['c'] + '[' + strftime("%H:%M:%S") + '] '\
+ c['g'] + 'Loading the data for ' + c['b'] + 'iteration 2' + c['nc']
(x, y,
names) = load_patches(dir_name='/home/sergivalverde/w/CNN/images/CH16',
use_flair=True,
use_pd=True,
use_t2=True,
use_t1=True,
use_gado=False,
flair_name='FLAIR_preprocessed.nii.gz',
pd_name='DP_preprocessed.nii.gz',
t2_name='T2_preprocessed.nii.gz',
gado_name=None,
t1_name='T1_preprocessed.nii.gz',
mask_name='Consensus.nii.gz',
size=patch_size,
roi_name='test.iter1.nii.gz')
print '-- Permuting the data'
np.random.seed(seed)
x_train = np.random.permutation(np.concatenate(x).astype(dtype=np.float32))
print '-- Permuting the labels'
np.random.seed(seed)
y_train = np.random.permutation(np.concatenate(y).astype(dtype=np.int32))
y_train = y_train[:, y_train.shape[1] / 2 + 1, y_train.shape[2] / 2 + 1,
y_train.shape[3] / 2 + 1]
print '-- Training vector shape = (' + ','.join(
[str(length) for length in x_train.shape]) + ')'
print '-- Training labels shape = (' + ','.join(
[str(length) for length in y_train.shape]) + ')'
print c['c'] + '[' + strftime("%H:%M:%S") + '] '\
+ c['g'] + 'Training (' + c['b'] + 'final' + c['nc'] + c['g'] + ')' + c['nc']
net.fit(x_train, y_train)