def _split_train_val(patient_folders):
"""Splits the patient folders into train and validation splits.
"""
# Construct train and validation splits using default parameters
if paths.SUBMISSION_NR == 1:
print("Using proper validation set")
validation_patients_indices = validation_set.get_cross_validation_indices(
indices=ALL_TRAIN_PATIENT_IDS, validation_index=0)
else:
print("WARNING: no validation set!!")
validation_patients_indices = [1]
train_patients_indices = [
i for i in ALL_TRAIN_PATIENT_IDS
if i not in validation_patients_indices
]
train_patient_folders = [
folder for folder in patient_folders
if not _extract_id_from_path(folder) in validation_patients_indices
]
validation_patient_folders = [
folder for folder in patient_folders
if folder not in train_patient_folders
]
return (train_patient_folders, validation_patient_folders,
validation_patients_indices, train_patients_indices)
def test_default_split(self):
# Create the cross-validation splits
splits = [
validation_set.get_cross_validation_indices(
_TEST_INDICES, i, _TEST_NO_SPLITS, _TEST_SEED)
for i in xrange(_TEST_NO_SPLITS)]
# Load the right ones
with open(_DEFAULT_SPLIT_RESULT_FILE, 'r') as f:
target_splits = pickle.load(f)
# Check for errors
for gen, target in zip(splits, target_splits):
self.assertEqual(gen, target)
def _split_train_val(patient_folders):
"""Splits the patient folders into train and validation splits.
"""
# Construct train and validation splits using default parameters
if paths.SUBMISSION_NR == 1:
print "Using proper validation set"
validation_patients_indices = validation_set.get_cross_validation_indices(
indices=ALL_TRAIN_PATIENT_IDS, validation_index=0)
else:
print "WARNING: no validation set!!"
validation_patients_indices = [1]
train_patients_indices = [i for i in ALL_TRAIN_PATIENT_IDS if i not in validation_patients_indices]
train_patient_folders = [
folder for folder in patient_folders
if not _extract_id_from_path(folder) in validation_patients_indices]
validation_patient_folders = [
folder for folder in patient_folders
if folder not in train_patient_folders]
return (
train_patient_folders, validation_patient_folders,
validation_patients_indices, train_patients_indices)
NUM_TRAIN_PATIENTS = num_patients['train']
NUM_VALID_PATIENTS = num_patients['validation']
NUM_TEST_PATIENTS = num_patients['test']
NUM_PATIENTS = NUM_TRAIN_PATIENTS + NUM_VALID_PATIENTS + NUM_TEST_PATIENTS
##############
# Sunny data #
##############
# This small dataset is loaded into memory
_SUNNY_DATA_PATH = os.path.join(_DATA_FOLDER, "pkl_annotated", "data.pkl")
_sunny_data = _load_file(_SUNNY_DATA_PATH)
num_sunny_images = len(_sunny_data["images"])
_validation_sunny_indices = validation_set.get_cross_validation_indices(
indices=range(num_sunny_images))
_train_sunny_indices = [
i for i in range(num_sunny_images) if i not in _validation_sunny_indices]
sunny_train_images = np.array(_sunny_data["images"])[_train_sunny_indices]
sunny_train_labels = np.array(_sunny_data["labels"])[_train_sunny_indices]
sunny_validation_images = np.array(_sunny_data["images"])[_validation_sunny_indices]
sunny_validation_labels = np.array(_sunny_data["labels"])[_validation_sunny_indices]
###########################
# Data form preprocessing #
###########################
_HOUGH_ROI_PATHS = (
TEMP_FILES_PATH + 'pkl_train_slice2roi.pkl',
import re
from configuration import config
import cPickle as pickle
import utils
from validation_set import get_cross_validation_indices
import random
print "Loading data"
patient_folders = sorted(
glob.glob("/data/dsb15_pkl/pkl_train/*/study/"),
key=lambda folder: int(re.search(r'/(\d+)/', folder).group(1)
)) # glob is non-deterministic!
validation_patients_indices = get_cross_validation_indices(indices=range(
1, 501),
validation_index=0)
train_patients_indices = [
i for i in range(1, 501) if i not in validation_patients_indices
]
VALIDATION_REGEX = "|".join(
["(/%d/)" % i for i in validation_patients_indices])
train_patient_folders = [
folder for folder in patient_folders
if re.search(VALIDATION_REGEX, folder) is None
]
validation_patient_folders = [
folder for folder in patient_folders if folder not in train_patient_folders
]
def optimize_expert_weights(expert_predictions,
average_distribution,
mask_matrix=None,
targets=None,
num_cross_validation_masks=2,
num_folds=1,
eps=1e-14,
cutoff=0.01,
do_optimization=True,
expert_weights=None,
optimal_params=None,
special_average=False,
*args, **kwargs):
"""
:param expert_predictions: experts x validation_samples x 600 x
:param mask_matrix: experts x validation_samples x
:param targets: validation_samples x 600 x
:param average_distribution: 600 x
:param eps:
:return:
"""
if expert_weights is not None:
mask_matrix = mask_matrix[expert_weights>cutoff,:] # remove
expert_predictions = expert_predictions[expert_weights>cutoff,:,:] # remove
NUM_EXPERTS = expert_predictions.shape[0]
NUM_FILTER_PARAMETERS = 2
WINDOW_SIZE = 599
# optimizing weights
X = theano.shared(expert_predictions.astype('float32')) # source predictions = (NUM_EXPERTS, NUM_VALIDATIONS, 600)
x_coor = theano.shared(np.linspace(-(WINDOW_SIZE-1)/2, (WINDOW_SIZE-1)/2, num=WINDOW_SIZE, dtype='float32')) # targets = (NUM_VALIDATIONS, 600)
NUM_VALIDATIONS = expert_predictions.shape[1]
ind = theano.shared(np.zeros((NUM_VALIDATIONS,), dtype='int32')) # targets = (NUM_VALIDATIONS, 600)
if optimal_params is None:
params_init = np.concatenate([ np.ones((NUM_EXPERTS,), dtype='float32'),
np.ones((NUM_FILTER_PARAMETERS,), dtype='float32') ])
else:
params_init = optimal_params.astype('float32')
params = theano.shared(params_init.astype('float32'))
#params = T.vector('params', dtype='float32') # expert weights = (NUM_EXPERTS,)
C = 0.0001
if not special_average:
# Create theano expression
# inputs:
W = params[:NUM_EXPERTS]
weights = T.nnet.softmax(W.dimshuffle('x',0)).dimshuffle(1, 0)
preds = X.take(ind, axis=1)
mask = theano.shared(mask_matrix.astype('float32')).take(ind, axis=1)
# expression
masked_weights = mask * weights
tot_masked_weights = T.clip(masked_weights.sum(axis=0), 1e-7, utils.maxfloat)
preds_weighted_masked = preds * masked_weights.dimshuffle(0, 1, 'x')
cumulative_distribution = preds_weighted_masked.sum(axis=0) / tot_masked_weights.dimshuffle(0, 'x')
# loss
l1_loss = weights.sum()
else:
# calculate the weighted average for each of these experts
weights = generate_information_weight_matrix(expert_predictions, average_distribution) # = (NUM_EXPERTS, NUM_VALIDATIONS, 600)
weight_matrix = theano.shared((mask_matrix[:,:,None]*weights).astype('float32'))
pdf = utils.cdf_to_pdf(expert_predictions)
x_log = np.log(pdf)
x_log[pdf<=0] = np.log(eps)
# Compute the mean
X_log = theano.shared(x_log.astype('float32')) # source predictions = (NUM_EXPERTS, NUM_VALIDATIONS, 600)
X_log_i = X_log.take(ind, axis=1)
w_i = weight_matrix.take(ind, axis=1)
W = params[:NUM_EXPERTS]
w_i = w_i * T.nnet.softmax(W.dimshuffle('x',0)).dimshuffle(1, 0, 'x')
#the different predictions, are the experts
geom_av_log = T.sum(X_log_i * w_i, axis=0) / (T.sum(w_i, axis=0) + eps)
geom_av_log = geom_av_log - T.max(geom_av_log,axis=-1).dimshuffle(0,'x') # stabilizes rounding errors?
geom_av = T.exp(geom_av_log)
geom_pdf = geom_av/T.sum(geom_av,axis=-1).dimshuffle(0,'x')
l1_loss = 0
cumulative_distribution = T.cumsum(geom_pdf, axis=-1)
if not do_optimization:
ind.set_value(list(range(NUM_VALIDATIONS)))
f_eval = theano.function([], cumulative_distribution)
cumulative_distribution = f_eval()
return cumulative_distribution[0]
else:
# convert to theano_values (for regularization)
t_valid = theano.shared(targets.astype('float32')) # targets = (NUM_VALIDATIONS, 600)
t_train = theano.shared(targets.astype('float32')) # targets = (NUM_VALIDATIONS, 600)
CRPS_train = T.mean((cumulative_distribution - t_train.take(ind, axis=0))**2) + C * l1_loss
CRPS_valid = T.mean((cumulative_distribution - t_valid.take(ind, axis=0))**2)
iter_optimize = theano.function([], CRPS_train, on_unused_input="ignore", updates=lasagne.updates.adam(CRPS_train, [params], 1.0))
f_val = theano.function([], CRPS_valid)
def optimize_my_params():
for _ in range(40 if special_average else 100): # early stopping
score = iter_optimize()
result = params.get_value()
return result, score
if num_cross_validation_masks==0:
ind.set_value(list(range(NUM_VALIDATIONS)))
params.set_value(params_init)
optimal_params, train_score = optimize_my_params()
final_weights = -1e10 * np.ones(expert_weights.shape,)
final_weights[np.where(expert_weights>cutoff)] = optimal_params[:NUM_EXPERTS]
final_params = np.concatenate(( final_weights, optimal_params[NUM_EXPERTS:]))
return softmax(final_weights), train_score, final_params
else:
final_params = []
final_losses = []
print()
print()
print()
for fold in range(num_folds):
for i_cross_validation in range(num_cross_validation_masks):
print("\r\033[F\033[F\033[Fcross_validation %d/%d"%(fold*num_cross_validation_masks+i_cross_validation+1, num_folds*num_cross_validation_masks))
val_indices = get_cross_validation_indices(list(range(NUM_VALIDATIONS)),
validation_index=i_cross_validation,
number_of_splits=num_cross_validation_masks,
rng_seed=fold,
)
indices = [i for i in range(NUM_VALIDATIONS) if i not in val_indices]
#out, crps, d = scipy.optimize.fmin_l_bfgs_b(f, w_init, fprime=g, pgtol=1e-09, epsilon=1e-08, maxfun=10000)
ind.set_value(indices)
params.set_value(params_init)
result, train_score = optimize_my_params()
final_params.append(result)
ind.set_value(val_indices)
validation_score = f_val()
print(" Current train value: %.6f" % train_score)
print(" Current validation value: %.6f" % validation_score)
final_losses.append(validation_score)
optimal_params = np.mean(final_params, axis=0)
average_loss = np.mean(final_losses)
expert_weights_result = softmax(optimal_params[:NUM_EXPERTS])
filter_param_result = optimal_params[NUM_EXPERTS:NUM_EXPERTS+NUM_FILTER_PARAMETERS]
#print "filter param result:", filter_param_result
return expert_weights_result, average_loss, optimal_params # (NUM_EXPERTS,)
NUM_TRAIN_PATIENTS = num_patients['train']
NUM_VALID_PATIENTS = num_patients['validation']
NUM_TEST_PATIENTS = num_patients['test']
NUM_PATIENTS = NUM_TRAIN_PATIENTS + NUM_VALID_PATIENTS + NUM_TEST_PATIENTS
##############
# Sunny data #
##############
# This small dataset is loaded into memory
_SUNNY_DATA_PATH = os.path.join(_DATA_FOLDER, "pkl_annotated", "data.pkl")
_sunny_data = _load_file(_SUNNY_DATA_PATH)
num_sunny_images = len(_sunny_data["images"])
_validation_sunny_indices = validation_set.get_cross_validation_indices(
indices=range(num_sunny_images))
_train_sunny_indices = [
i for i in range(num_sunny_images) if i not in _validation_sunny_indices
]
sunny_train_images = np.array(_sunny_data["images"])[_train_sunny_indices]
sunny_train_labels = np.array(_sunny_data["labels"])[_train_sunny_indices]
sunny_validation_images = np.array(
_sunny_data["images"])[_validation_sunny_indices]
sunny_validation_labels = np.array(
_sunny_data["labels"])[_validation_sunny_indices]
###########################
# Data form preprocessing #
###########################
import glob
import numpy as np
import re
from configuration import config
import cPickle as pickle
import utils
from validation_set import get_cross_validation_indices
import random
print "Loading data"
patient_folders = sorted(glob.glob("/data/dsb15_pkl/pkl_train/*/study/"), key=lambda folder: int(re.search(r'/(\d+)/', folder).group(1))) # glob is non-deterministic!
validation_patients_indices = get_cross_validation_indices(indices=range(1,501), validation_index=0)
train_patients_indices = [i for i in range(1,501) if i not in validation_patients_indices]
VALIDATION_REGEX = "|".join(["(/%d/)"%i for i in validation_patients_indices])
train_patient_folders = [folder for folder in patient_folders if re.search(VALIDATION_REGEX, folder) is None]
validation_patient_folders = [folder for folder in patient_folders if folder not in train_patient_folders]
import os
import os.path
def copy(from_folder, to_folder):
command = "cp -r %s %s/."%(from_folder, to_folder)
print command
os.system(command)
for folder in train_patient_folders:
def optimize_expert_weights(expert_predictions,
average_distribution,
mask_matrix=None,
targets=None,
num_cross_validation_masks=2,
num_folds=1,
eps=1e-14,
cutoff=0.01,
do_optimization=True,
expert_weights=None,
optimal_params=None,
special_average=False,
*args, **kwargs):
"""
:param expert_predictions: experts x validation_samples x 600 x
:param mask_matrix: experts x validation_samples x
:param targets: validation_samples x 600 x
:param average_distribution: 600 x
:param eps:
:return:
"""
if expert_weights is not None:
mask_matrix = mask_matrix[expert_weights>cutoff,:] # remove
expert_predictions = expert_predictions[expert_weights>cutoff,:,:] # remove
NUM_EXPERTS = expert_predictions.shape[0]
NUM_FILTER_PARAMETERS = 2
WINDOW_SIZE = 599
# optimizing weights
X = theano.shared(expert_predictions.astype('float32')) # source predictions = (NUM_EXPERTS, NUM_VALIDATIONS, 600)
x_coor = theano.shared(np.linspace(-(WINDOW_SIZE-1)/2, (WINDOW_SIZE-1)/2, num=WINDOW_SIZE, dtype='float32')) # targets = (NUM_VALIDATIONS, 600)
NUM_VALIDATIONS = expert_predictions.shape[1]
ind = theano.shared(np.zeros((NUM_VALIDATIONS,), dtype='int32')) # targets = (NUM_VALIDATIONS, 600)
if optimal_params is None:
params_init = np.concatenate([ np.ones((NUM_EXPERTS,), dtype='float32'),
np.ones((NUM_FILTER_PARAMETERS,), dtype='float32') ])
else:
params_init = optimal_params.astype('float32')
params = theano.shared(params_init.astype('float32'))
#params = T.vector('params', dtype='float32') # expert weights = (NUM_EXPERTS,)
C = 0.0001
if not special_average:
# Create theano expression
# inputs:
W = params[:NUM_EXPERTS]
weights = T.nnet.softmax(W.dimshuffle('x',0)).dimshuffle(1, 0)
preds = X.take(ind, axis=1)
mask = theano.shared(mask_matrix.astype('float32')).take(ind, axis=1)
# expression
masked_weights = mask * weights
tot_masked_weights = T.clip(masked_weights.sum(axis=0), 1e-7, utils.maxfloat)
preds_weighted_masked = preds * masked_weights.dimshuffle(0, 1, 'x')
cumulative_distribution = preds_weighted_masked.sum(axis=0) / tot_masked_weights.dimshuffle(0, 'x')
# loss
l1_loss = weights.sum()
else:
# calculate the weighted average for each of these experts
weights = generate_information_weight_matrix(expert_predictions, average_distribution) # = (NUM_EXPERTS, NUM_VALIDATIONS, 600)
weight_matrix = theano.shared((mask_matrix[:,:,None]*weights).astype('float32'))
pdf = utils.cdf_to_pdf(expert_predictions)
x_log = np.log(pdf)
x_log[pdf<=0] = np.log(eps)
# Compute the mean
X_log = theano.shared(x_log.astype('float32')) # source predictions = (NUM_EXPERTS, NUM_VALIDATIONS, 600)
X_log_i = X_log.take(ind, axis=1)
w_i = weight_matrix.take(ind, axis=1)
W = params[:NUM_EXPERTS]
w_i = w_i * T.nnet.softmax(W.dimshuffle('x',0)).dimshuffle(1, 0, 'x')
#the different predictions, are the experts
geom_av_log = T.sum(X_log_i * w_i, axis=0) / (T.sum(w_i, axis=0) + eps)
geom_av_log = geom_av_log - T.max(geom_av_log,axis=-1).dimshuffle(0,'x') # stabilizes rounding errors?
geom_av = T.exp(geom_av_log)
geom_pdf = geom_av/T.sum(geom_av,axis=-1).dimshuffle(0,'x')
l1_loss = 0
cumulative_distribution = T.cumsum(geom_pdf, axis=-1)
if not do_optimization:
ind.set_value(range(NUM_VALIDATIONS))
f_eval = theano.function([], cumulative_distribution)
cumulative_distribution = f_eval()
return cumulative_distribution[0]
else:
# convert to theano_values (for regularization)
t_valid = theano.shared(targets.astype('float32')) # targets = (NUM_VALIDATIONS, 600)
t_train = theano.shared(targets.astype('float32')) # targets = (NUM_VALIDATIONS, 600)
CRPS_train = T.mean((cumulative_distribution - t_train.take(ind, axis=0))**2) + C * l1_loss
CRPS_valid = T.mean((cumulative_distribution - t_valid.take(ind, axis=0))**2)
iter_optimize = theano.function([], CRPS_train, on_unused_input="ignore", updates=lasagne.updates.adam(CRPS_train, [params], 1.0))
f_val = theano.function([], CRPS_valid)
def optimize_my_params():
for _ in xrange(40 if special_average else 100): # early stopping
score = iter_optimize()
result = params.get_value()
return result, score
if num_cross_validation_masks==0:
ind.set_value(range(NUM_VALIDATIONS))
params.set_value(params_init)
optimal_params, train_score = optimize_my_params()
final_weights = -1e10 * np.ones(expert_weights.shape,)
final_weights[np.where(expert_weights>cutoff)] = optimal_params[:NUM_EXPERTS]
final_params = np.concatenate(( final_weights, optimal_params[NUM_EXPERTS:]))
return softmax(final_weights), train_score, final_params
else:
final_params = []
final_losses = []
print
print
print
for fold in xrange(num_folds):
for i_cross_validation in xrange(num_cross_validation_masks):
print "\r\033[F\033[F\033[Fcross_validation %d/%d"%(fold*num_cross_validation_masks+i_cross_validation+1, num_folds*num_cross_validation_masks)
val_indices = get_cross_validation_indices(range(NUM_VALIDATIONS),
validation_index=i_cross_validation,
number_of_splits=num_cross_validation_masks,
rng_seed=fold,
)
indices = [i for i in range(NUM_VALIDATIONS) if i not in val_indices]
#out, crps, d = scipy.optimize.fmin_l_bfgs_b(f, w_init, fprime=g, pgtol=1e-09, epsilon=1e-08, maxfun=10000)
ind.set_value(indices)
params.set_value(params_init)
result, train_score = optimize_my_params()
final_params.append(result)
ind.set_value(val_indices)
validation_score = f_val()
print " Current train value: %.6f" % train_score
print " Current validation value: %.6f" % validation_score
final_losses.append(validation_score)
optimal_params = np.mean(final_params, axis=0)
average_loss = np.mean(final_losses)
expert_weights_result = softmax(optimal_params[:NUM_EXPERTS])
filter_param_result = optimal_params[NUM_EXPERTS:NUM_EXPERTS+NUM_FILTER_PARAMETERS]
#print "filter param result:", filter_param_result
return expert_weights_result, average_loss, optimal_params # (NUM_EXPERTS,)
