def predict_to_file(filename, path_to_model):
model = load_model(path_to_model)
# Data
h5_file_location = os.path.join('C:\\Users\Jeftha\stack\Rommel\ISMI\data',
'prostatex-test.hdf5')
h5_file = h5py.File(h5_file_location, 'r')
x, _, attr = get_train_data(h5_file, ['ADC'])
x = np.expand_dims(x, axis=-1)
windowed = []
for lesion in x:
windowed.append(apply_window(lesion, (500, 1100)))
x = np.asarray(windowed)
predictions = model.predict(x, verbose=1)
with open(filename, 'w') as f:
f.write('proxid,clinsig\n')
for i in range(len(predictions)):
patient_id = attr[i]['patient_id']
fid = attr[i]['fid']
line = "ProstateX-%s-%s,%f\n" % (patient_id, fid, predictions[i])
f.write(line)
print(line)
def get_model(configuration='baseline'):
AUGMENTATION_CONFIGURATION = configuration
## Model
model = Sequential()
model.add(
Conv2D(32, (3, 3),
kernel_initializer='he_normal',
bias_initializer=RandomNormal(mean=0.1),
input_shape=(16, 16, 1)))
model.add(LeakyReLU())
model.add(BatchNormalization())
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(
Conv2D(64, (2, 2),
kernel_initializer='he_normal',
bias_initializer=RandomNormal(mean=0.1)))
model.add(LeakyReLU())
model.add(BatchNormalization())
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(
Conv2D(64, (2, 2),
kernel_initializer='he_normal',
bias_initializer=RandomNormal(mean=0.1)))
model.add(LeakyReLU())
model.add(BatchNormalization())
model.add(
Conv2D(64, (2, 2),
kernel_initializer='he_normal',
bias_initializer=RandomNormal(mean=0.1)))
model.add(LeakyReLU())
model.add(BatchNormalization())
model.add(
Conv2D(64, (1, 1),
kernel_initializer='he_normal',
bias_initializer=RandomNormal(mean=0.1)))
model.add(LeakyReLU())
model.add(BatchNormalization())
model.add(Conv2D(1, (1, 1), activation='sigmoid'))
model.add(Flatten())
# For a binary classification problem
sgd = SGD(lr=0.0005, momentum=0.9)
model.compile(optimizer=sgd,
loss='binary_crossentropy',
metrics=['accuracy'])
## Data
h5_file_location = os.path.join('C:\\Users\Jeftha\stack\Rommel\ISMI\data',
'prostatex-train.hdf5')
h5_file = h5py.File(h5_file_location, 'r')
train_data_list, train_labels_list, attr = get_train_data(h5_file, ['ADC'])
windowed = []
for lesion in train_data_list:
windowed.append(apply_window(lesion, (500, 1100)))
train_data_list = np.asarray(windowed)
train_data, val_data, train_labels, val_labels = train_test_split(
train_data_list, train_labels_list, attr, test_size=0.33)
train_data = np.expand_dims(train_data, axis=-1)
val_data = np.expand_dims(val_data, axis=-1)
train_labels = np.expand_dims(train_labels, axis=-1)
val_labels = np.expand_dims(val_labels, axis=-1)
## Stuff for training
generator = get_generator(configuration=AUGMENTATION_CONFIGURATION)
train_generator = generator.flow(
train_data,
train_labels) #, save_to_dir="/nfs/home4/schellev/augmented_images")
batch_size = 128
steps_per_epoch = len(train_labels_list) // batch_size
auc_history = AucHistory(train_data,
train_labels,
val_data,
val_labels,
output_graph_name=AUGMENTATION_CONFIGURATION)
model.fit_generator(train_generator,
steps_per_epoch,
epochs=15000,
verbose=2,
callbacks=[auc_history],
max_q_size=50,
workers=8)
return model
def predict_to_file(filename, path_to_model):
model = load_model(path_to_model)
model.summary()
# Data
if IS_TRAIN:
file = 'prostatex-train.hdf5'
else:
file = 'prostatex-test.hdf5'
h5_file_location = os.path.join('/scratch-shared/ISMI/prostatex', file)
h5_file = h5py.File(h5_file_location, 'r')
x, y, attr = get_train_data(h5_file, ['ADC'])
x = np.expand_dims(x, axis=-1)
windowed = []
for lesion in x:
windowed.append(apply_window(lesion, (500, 1100)))
x = np.asarray(windowed)
print(attr[0])
activations = get_activations(model, -3, x)
print('length of predictions:', len(activations))
print('length of prediction[0]:', len(activations[0]))
print(activations[0].shape)
activations = activations[0]
with open(filename, 'w') as f:
header_items = ['proxid']
if IS_TRAIN:
header_items.append('clinsig')
for index in range(activations.shape[-1]):
header_items.append('feature' + str(index).zfill(2))
attr_features = ['Age', 'Zone']
header_items.extend(attr_features)
header = ','.join(header_items) + '\n'
f.write(header)
for i in range(activations.shape[0]):
patient_id = attr[i]['patient_id']
fid = attr[i]['fid']
line = ["ProstateX-%s-%s" % (patient_id, fid)]
if IS_TRAIN:
line.append('%f' % float(y[i]))
features = ['%f' % feature for feature in activations[i, 0, 0, :]]
line.extend(features)
attribute_values = []
for attribute in attr_features:
attribute_values.append(str(attr[i][attribute], 'utf-8'))
line.extend(attribute_values)
line = ','.join(line) + '\n'
f.write(line)
print(line)