__author__ = 'adeb'
from shutil import copy2
import inspect
import PIL
import pickle
from spynet.utils.utilities import analyse_classes
from data_brain_parcellation import DatasetBrainParcellation
from network_brain_parcellation import *
from spynet.models.network import *
from spynet.models.neuron_type import *
from spynet.data.dataset import *
from spynet.training.trainer import *
from spynet.training.monitor import *
from spynet.training.parameters_selector import *
from spynet.training.stopping_criterion import *
from spynet.training.cost_function import *
from spynet.training.learning_update import *
from spynet.experiment import Experiment
from spynet.utils.utilities import tile_raster_images
import theano
class ExperimentBrain(Experiment):
"""
Main experiment to train a network on a dataset
"""
def __init__(self, exp_name, data_path):
Experiment.__init__(self, exp_name, data_path)
def copy_file_virtual(self):
copy2(inspect.getfile(inspect.currentframe()), self.path)
def run(self):
###### Create the datasets
# aa = CostNegLLWeighted(np.array([0.9, 0.1]))
# e = theano.function(inputs=[], outputs=aa.test())
# print e()
## Load the data
training_data_path = self.data_path + "train.h5"
ds_training = DatasetBrainParcellation()
ds_training.read(training_data_path)
[ds_training, ds_validation] = ds_training.split_dataset_proportions([0.95, 0.05])
testing_data_path = self.data_path + "test.h5"
ds_testing = DatasetBrainParcellation()
ds_testing.read(testing_data_path)
## Display data sample
# image = PIL.Image.fromarray(tile_raster_images(X=ds_training.inputs[0:50],
# img_shape=(29, 29), tile_shape=(5, 10),
# tile_spacing=(1, 1)))
# image.save(self.path + "filters_corruption_30.png")
## Few stats about the targets
classes, proportion_class = analyse_classes(np.argmax(ds_training.outputs, axis=1), "Training data:")
print classes
## Scale some part of the data
print "Scaling"
s = Scaler([slice(-134, None, None)])
s.compute_parameters(ds_training.inputs)
s.scale(ds_training.inputs)
s.scale(ds_validation.inputs)
s.scale(ds_testing.inputs)
pickle.dump(s, open(self.path + "s.scaler", "wb"))
###### Create the network
net = NetworkUltimateConv()
net.init(33, 29, 5, 134, 135)
print net
###### Configure the trainer
# Cost function
cost_function = CostNegLL(net.ls_params)
# Learning update
learning_rate = 0.05
momentum = 0.5
lr_update = LearningUpdateGDMomentum(learning_rate, momentum)
# Create monitors and add them to the trainer
freq = 1
freq2 = 0.00001
# err_training = MonitorErrorRate(freq, "Train", ds_training)
# err_testing = MonitorErrorRate(freq, "Test", ds_testing)
err_validation = MonitorErrorRate(freq, "Val", ds_validation)
# dice_training = MonitorDiceCoefficient(freq, "Train", ds_training, 135)
dice_testing = MonitorDiceCoefficient(freq, "Test", ds_testing, 135)
# dice_validation = MonitorDiceCoefficient(freq, "Val", ds_validation, 135)
# Create stopping criteria and add them to the trainer
max_epoch = MaxEpoch(300)
early_stopping = EarlyStopping(err_validation, 10, 0.99, 5)
# Create the network selector
params_selector = ParamSelectorBestMonitoredValue(err_validation)
# Create the trainer object
batch_size = 200
t = Trainer(net, cost_function, params_selector, [max_epoch, early_stopping],
lr_update, ds_training, batch_size,
[err_validation, dice_testing])
###### Train the network
t.train()
###### Plot the records
# pred = np.argmax(t.net.predict(ds_testing.inputs, 10000), axis=1)
# d = compute_dice(pred, np.argmax(ds_testing.outputs, axis=1), 134)
# print "Dice test: {}".format(np.mean(d))
# print "Error rate test: {}".format(error_rate(np.argmax(ds_testing.outputs, axis=1), pred))
save_records_plot(self.path, [err_validation], "err", t.n_train_batches, "upper right")
# save_records_plot(self.path, [dice_testing], "dice", t.n_train_batches, "lower right")
###### Save the network
net.save_parameters(self.path + "net.net")
if __name__ == '__main__':
exp_name = "paper_ultimate_conv"
data_path = "./datasets/paper_ultimate_conv/"
exp = ExperimentBrain(exp_name, data_path)
exp.run()
