def train_svm(C=0.1, grid=False):
ds = PascalSegmentation()
svm = LinearSVC(C=C, dual=False, class_weight='auto')
if grid:
data_train = load_pascal("kTrain")
X, y = shuffle(data_train.X, data_train.Y)
# prepare leave-one-label-out by assigning labels to images
image_indicators = np.hstack([np.repeat(i, len(x)) for i, x in
enumerate(X)])
# go down to only 5 "folds"
labels = image_indicators % 5
X, y = np.vstack(X), np.hstack(y)
cv = LeavePLabelOut(labels=labels, p=1)
param_grid = {'C': 10. ** np.arange(-3, 3)}
scorer = Scorer(recall_score, average="macro")
grid_search = GridSearchCV(svm, param_grid=param_grid, cv=cv,
verbose=10, scoring=scorer, n_jobs=-1)
grid_search.fit(X, y)
else:
data_train = load_pascal("train")
X, y = np.vstack(data_train.X), np.hstack(data_train.Y)
svm.fit(X, y)
print(svm.score(X, y))
eval_on_sp(ds, data_train, [svm.predict(x) for x in data_train.X],
print_results=True)
data_val = load_pascal("val")
eval_on_sp(ds, data_val, [svm.predict(x) for x in data_val.X],
print_results=True)
def tuneAndTrainTest(estimatorType,
data,
labels,
numFolds,
patientIds,
roundNum,
estimator,
lossFunction=lossFunction):
folds = LeavePLabelOut(patientIds[0], p=2)
errors = []
counter = 0
for trainIndices, testIndices in folds:
#Tuning
trainData = [data[i] for i in trainIndices]
trainlabels = [labels[i] for i in trainIndices]
testData = [data[i] for i in testIndices]
testLabels = [labels[i] for i in testIndices]
#Training
if roundNum == 0:
estimator.fit(trainData, trainlabels)
else:
estimator = optimizeHyperParams(trainData, trainlabels,
estimatorType)
#Testing
errors.append(lossFunction(estimator, testData, testLabels))
print np.array(errors).mean()
def tuneAndTrain(predictorType,
data,
labels,
patientIds,
numFolds,
lossFunction=lossFunction):
'''
:param data: data matrix, a structured array.
:param predictorType: one of: {'SVM', 'RF'}
:param patientIds: May be ints, strings, etc. denoting which patient every line is taken from.
:return: the mean error of an optimized predictor of type predictorType, and the optimized, trained predictor.
'''
folds = LeavePLabelOut(patientIds, p=2) #
folds = [tup for tup in folds]
errors = []
for trainIndices, testIndices in folds[:numFolds]:
#Tuning
#get actual data and labels for current fold
trainData = data[trainIndices]
trainLabels = labels[trainIndices]
testData = data[testIndices]
testLabels = labels[testIndices]
testNames = patientIds[testIndices]
if np.all(trainLabels == trainLabels[0]):
continue #can't train on elements that are all from the same group
# todo: for testing - skip feature selection
# selectedFeatures = SelectFeatures(trainData, trainLabels)
# selectedTrainData = [trainData[f] for f in selectedFeatures]
# selectedTestData = [testData[f] for f in selectedFeatures]
selectedTrainData = trainData
selectedTestData = testData
# to get the data in a list-of-list format that optimizeHyperParams expects.
selectedTrainData = [list(tup) for tup in selectedTrainData]
selectedTestData = [list(tup) for tup in selectedTestData]
predictor = optimizeHyperParams(selectedTrainData, trainLabels,
predictorType)
#Training
#predictor.fit(selectedFeaturesTrainData, trainLabels) //optimizeHyperParams also trains
#Testing
errors.append(
lossFunction(predictor, selectedTestData, testLabels, testNames))
return np.array(errors).mean()
def twoStepsLoss(estimator, X, y, names):
print names
folds = LeavePLabelOut(names, p=1) #ugly patch - correct syntax?
folds = [tup for tup in folds]
loss = 0.0
for testIndices in folds[:2]:
print "in fold"
testData = X[testIndices]
testLabels = y[testIndices]
testLabel = (y[testIndices])[0]
sickCount = 0.0
for data in testData:
sickCount = sickCount + (estimator.predict(data)[0])
sickCount = sickCount / len(X)
if sickCount > 0.5: #mostly sick
if testLabel == 0:
loss = loss + 1
else:
if testLabel == 1:
loss = loss + 1
loss = 1 - (loss / 2)
return loss
fmri, stimuli, onsets, conditions = read_data_gauthier(subject)
session_id_onset = np.load('sessions_id_onset.npy')
session_id_onset = [[19.2 / len(onsets[session])] * len(onsets[session])
for session in range(len(onsets))]
betas, reg = de.glm(fmri,
tr,
onsets,
hrf_model=hrf_model,
drift_model='blank',
model=model)
betas = np.vstack(betas)
conditions = np.hstack(conditions)
session_id_onset = np.hstack(session_id_onset)
lplo = LeavePLabelOut(session_id_onset, p=1)
for train_index, test_index in lplo:
# Split into train and test sets
betas_isi = betas[test_index]
conditions_isi = conditions[test_index]
n_points = len(conditions_isi)
if n_points == 12 * 4:
isi = 1.6
elif n_points == 6 * 4:
isi = 3.2
elif n_points == 4 * 4:
isi = 4.8
# Read data
fmri, stimuli, onsets, conditions = read_data_mrt(subject)
session_id_onset = [[session] * len(onsets[session])
for session in range(len(onsets))]
betas, reg = de.glm(fmri, tr, onsets, hrf_model=hrf_model, model=model)
betas = np.vstack(betas)
session_id_onset = np.hstack(session_id_onset)
conditions = np.hstack(conditions)
junk_mask = np.where(conditions != 'ju')
conditions = conditions[junk_mask]
betas = betas[junk_mask]
session_id_onset = session_id_onset[junk_mask]
lplo = LeavePLabelOut(session_id_onset, p=1)
for train_index, test_index in lplo:
# Split into train and test sets
betas_train, betas_test = betas[train_index], betas[test_index]
conditions_train, conditions_test = (conditions[train_index],
conditions[test_index])
# Feature selection
betas_train, betas_test = de.feature_selection(betas_train,
betas_test,
conditions_train,
k=k)
# Fit a logistic regression to score the model
accuracy = de.glm_scoring(betas_train, betas_test, conditions_train,
conditions_test)
print(train, test)
'''
Leave-P-Label-Out
LeavePLabelOut is similar as Leave-One-Label-Out, but removes samples related to P labels for each
training/test set.
'''
from sklearn.cross_validation import LeavePLabelOut
X = [[0., 0.], [1., 1.], [-1., -1.], [2., 2.], [3., 3.], [4., 4.]]
Y = [0, 1, 0, 1, 0, 1]
labels = [1, 1, 2, 2, 3, 3]
lplo = LeavePLabelOut(labels, 2)
print(lplo)
for train, test in lplo:
print(train, test)
'''
Random permutations cross-validation a.k.a. Shuffle & Split¶
ShuffleSplit
The ShuffleSplit iterator will generate a user defined number of independent train / test dataset splits. Samples are first shuffled and then splitted into a pair of train and test sets.
It is possible to control the randomness for reproducibility of the results by explicitly seeding the random_state pseudo random number generator.
Here is a usage example:
"Random permutation cross-validator
#scores = -np.log10(selector.pvalues_)
#scores /= scores.max()
#print "selected factors:\n"
#support=selector.get_support(indices=True)
#selected= header[selector.get_support(indices=True)]
#for i,item in enumerate(selected):
# print item," ",scores[support[i]]
#print X.shape
#X=selector.transform(X)
#print X.shape
# ridge
lol = LeaveOneLabelOut(offers)
lop = LeavePLabelOut(offers, len(offers_set) - 1)
cvset = lop
rf = RandomForestClassifier(n_estimators=100,
n_jobs=-1,
oob_score=True,
max_depth=5,
min_samples_leaf=5) #0.675
xrf = ExtraTreesClassifier(n_estimators=100, n_jobs=-1) #0.662
dt = DecisionTreeClassifier() #0.559
lr = LogisticRegression(C=0.001) # 0.654
gbr = GradientBoostingRegressor() # 0.691
#gbr1=GradientBoostingRegressor(loss="quantile",alpha=0.6) # 0.542
gbr1 = GradientBoostingRegressor(loss='quantile', alpha=0.5) #0.544
gbc = GradientBoostingClassifier() #0.692
de.design_matrix(len(fmri[session]),
tr,
onsets[session],
conditions[session],
hrf_model=hrf_model,
durations=durations[session])
for session in range(len(fmri))
]
# Stack the BOLD signals and the design matrices
fmri = np.vstack(fmri)
design = np.vstack(design)
stimuli = np.vstack(stimuli)
session_id_fmri = np.hstack(session_id_fmri)
lplo = LeavePLabelOut(session_id_fmri, p=2)
for train_index, test_index in lplo:
# Split into train and test sets
fmri_train, fmri_test = fmri[train_index], fmri[test_index]
design_train, design_test = design[train_index], design[test_index]
stimuli_train, stimuli_test = stimuli[train_index], stimuli[test_index]
# Feature selection
fmri_train, fmri_test = de.feature_selection(
fmri_train, fmri_test, np.argmax(stimuli_train, axis=1))
# Fit a ridge regression to predict the design matrix
prediction_test, prediction_train, score = de.fit_ridge(
fmri_train,
fmri_test,
design_train,
data_train,data_test,target_train,target_test=\
train_test_split(housing.data,housing.target,test_size=0.1,random_state=42)
dtr=tree.DecisionTreeRegressor()
dtr.fit(data_train,target_train)
#score=dtr.score(data_test,target_test)
kf=KFold(data_train.shape[0],n_folds=5)
skf=StratifiedKFold(target_train,n_folds=5)
loo=LeaveOneOut(data_train.shape[0])
lpo=LeavePOut(data_train.shape[0],6)
labels_lolo=[1,1,2,2]
lolo=LeaveOneLabelOut(labels_lolo)
#这个策略在划分样本时，会根据第三方提供的整数型样本类标号进行划分
#每次划分数据集时，去除某个属于某个类裱好的样本作为测试集，剩余的作为训练集
labels_lopo=[1,1,2,2,3,3]
lopo=LeavePLabelOut(labels_lopo,2)
#这个策略每次取得p种类标号的数据作为测试集，其余作为训练集
#注意cross_val_score中的cv参数
cross_score=cross_val_score(dtr,data_train,target_train,cv=skf)
print("交叉验证：")
print(cross_score)
#print(score)
'''
#这种无法使用metrics进行判断，原因在于continous is not supported
pred_test=dtr.predict(data_test)
acc_score=metrics.accuracy_score(target_test,pred_test)
print(acc_score)
'''
n_classes = 12
network = SegNet(in_channels=3, n_classes=n_classes)
network.init_encoder()
network.cuda()
net = skorch.NeuralNet(
module=network,
criterion=torch.nn.CrossEntropyLoss,
train_split=None,
use_cuda=True,
batch_size=10,
)
params = {'lr': [0.01, 0.02], 'max_epochs': [5, 10]}
# if only training
# net.fit(X=X, y=y)
image_indicators = np.hstack([np.repeat(i, len(x)) for i, x in enumerate(X)])
labels = image_indicators % n_classes
#
X, y = np.vstack(X), np.hstack(Y)
cv = LeavePLabelOut(labels=labels, p=1)
gs = GridSearchCV(net, params, scoring='f1', verbose=10, cv=cv, n_jobs=-1)
gs.fit(X=X, y=Y)
print(gs.best_score_, gs.best_params_)
design = [
de.design_matrix(len(fmri[session]),
tr,
onsets[session],
conditions[session],
hrf_model=hrf_model,
drift_model='blank') for session in range(len(fmri))
]
# Stack the BOLD signals and the design matrices
fmri = np.vstack(fmri)
design = np.vstack(design)
stimuli = np.vstack(stimuli)
session_id_fmri = np.hstack(session_id_fmri)
lplo = LeavePLabelOut(session_id_fmri, p=1)
for train_index, test_index in lplo:
# Split into train and test sets
_, fmri_isi = fmri[train_index], fmri[test_index]
_, design_isi = design[train_index], design[test_index]
_, stimuli_isi = stimuli[train_index], stimuli[test_index]
n_points = np.sum(stimuli_isi[:, 1:])
if n_points == 12 * 4:
isi = 1.6
logistic_window = 4
delay = 1
elif n_points == 6 * 4:
isi = 3.2
logistic_window = 4
def fit(self, X_task, y, sub_ids):
DEBUG_FLAG = True
# self.max_epochs = 333
self.batch_size = 100
n_input = X_task.shape[1] # sklearn-like structure
n_output = n_input
rng = np.random.RandomState(42)
self.input_taskdata = T.matrix(dtype='float32', name='input_taskdata')
self.params_from_last_iters = []
index = T.iscalar(name='index')
# prepare data for theano computation
if not DEBUG_FLAG:
X_train_s = theano.shared(value=np.float32(X_task),
name='X_train_s')
y_train_s = theano.shared(value=np.int32(y), name='y_train_s')
else:
# from sklearn.cross_validation import StratifiedShuffleSplit
# folder = StratifiedShuffleSplit(y, n_iter=1, test_size=0.20)
# new_trains, inds_val = iter(folder).next()
from sklearn.cross_validation import LeavePLabelOut
folder = LeavePLabelOut(sub_ids, p=1)
new_trains, inds_val = iter(folder).next()
# valid_subs = np.array([1107, 1109, 1110, 1114, # healthy
# 2105, 2106, 2113, 2125, # depression
# 3280, 3279, 3276, 3275]) # sz
# inds_val = np.in1d(sub_ids, valid_subs)
# new_trains = np.logical_not(inds_val)
print('Data points in train set: %i' % np.sum(new_trains))
print('Data points in validation set: %i' % np.sum(inds_val))
print('Data features: %i' % n_input)
X_train, X_val = X_task[new_trains], X_task[inds_val]
y_train, y_val = y[new_trains], y[inds_val]
X_train_s = theano.shared(value=np.float32(X_train),
name='X_train_s',
borrow=False)
y_train_s = theano.shared(value=np.int32(y_train),
name='y_train_s',
borrow=False)
# X_val_s = theano.shared(value=np.float32(X_val),
# name='X_train_s', borrow=False)
# y_val_s = theano.shared(value=np.int32(y_val),
# name='y_cal_s', borrow=False)
self.dbg_epochs_ = list()
self.dbg_acc_train_ = list()
self.dbg_acc_val_ = list()
self.dbg_ae_cost_ = list()
self.dbg_lr_cost_ = list()
self.dbg_ae_nonimprovesteps = list()
self.dbg_acc_other_ds_ = list()
self.dbg_combined_cost_ = list()
self.dbg_prfs_ = list()
self.dbg_prfs_other_ds_ = list()
train_samples = len(X_train)
# V -> supervised / logistic regression
# W -> unsupervised / auto-encoder
# computational graph: auto-encoder
W0_vals = rng.randn(n_input, self.n_hidden).astype(
np.float32) * self.gain1
self.W0s = theano.shared(W0_vals)
# self.W1s = self.W0s.T # tied
W1_vals = rng.randn(self.n_hidden, n_input).astype(
np.float32) * self.gain1
self.W1s = theano.shared(W1_vals)
bW0_vals = np.zeros(self.n_hidden).astype(np.float32)
self.bW0s = theano.shared(value=bW0_vals, name='bW0')
bW1_vals = np.zeros(n_output).astype(np.float32)
self.bW1s = theano.shared(value=bW1_vals, name='bW1')
encoding = (self.input_taskdata.dot(self.W0s) + self.bW0s).dot(
self.W1s) + self.bW1s
self.ae_loss = T.sum((self.input_taskdata - encoding)**2, axis=1)
self.ae_cost = (T.mean(self.ae_loss) / n_input)
# params1 = [self.W0s, self.bW0s, self.bW1s]
# gparams1 = [T.grad(cost=self.ae_cost, wrt=param1) for param1 in params1]
#
# lr = self.learning_rate
# updates = self.RMSprop(cost=self.ae_cost, params=params1,
# lr=self.learning_rate)
# f_train_ae = theano.function(
# [index],
# [self.ae_cost],
# givens=givens_ae,
# updates=updates)
# computation graph: logistic regression
clf_n_output = len(np.unique(y))
print('SSFLogreg: Fitting %i classes' % clf_n_output)
my_y = T.ivector(name='y')
bV0_vals = np.zeros(self.n_hidden).astype(np.float32)
self.bV0 = theano.shared(value=bV0_vals, name='bV0')
bV1_vals = np.zeros(clf_n_output).astype(np.float32)
self.bV1 = theano.shared(value=bV1_vals, name='bV1')
# V0_vals = rng.randn(n_input, self.n_hidden).astype(np.float32) * self.gain1
# self.V0s = theano.shared(V0_vals)
V1_vals = rng.randn(self.n_hidden, clf_n_output).astype(
np.float32) * self.gain1
self.V1s = theano.shared(V1_vals)
self.p_y_given_x = T.nnet.softmax(
# T.dot(T.dot(self.input_taskdata, self.V0s) + self.bV0, self.V1s) + self.bV1
T.dot(T.dot(self.input_taskdata, self.W0s) + self.bV0, self.V1s) +
self.bV1)
self.lr_cost = -T.mean(
T.log(self.p_y_given_x)[T.arange(my_y.shape[0]), my_y])
self.lr_cost = (
self.lr_cost + T.mean(abs(self.W0s)) * self.penalty_l1 +
# T.mean(abs(self.V0s)) * self.penalty_l1 +
T.mean(abs(self.bV0)) * self.penalty_l1 +
T.mean(abs(self.V1s)) * self.penalty_l1 +
T.mean(abs(self.bV1)) * self.penalty_l1 + T.mean(
(self.W0s**np.float32(2))) * self.penalty_l2 +
# T.mean((self.V0s ** 2)) * self.penalty_l2 +
T.mean((self.bV0**np.float32(2))) * self.penalty_l2 + T.mean(
(self.V1s**np.float32(2))) * self.penalty_l2 + T.mean(
(self.bV1**np.float32(2))) * self.penalty_l2)
self.y_pred = T.argmax(self.p_y_given_x, axis=1)
# params2 = [self.V0s, self.bV0, self.V1s, self.bV1]
# params2 = [self.W0s, self.bV0, self.V1s, self.bV1]
# updates2 = self.RMSprop(cost=self.lr_cost, params=params2,
# lr=self.learning_rate)
# f_train_lr = theano.function(
# [index],
# [self.lr_cost],
# givens=givens_lr,
# updates=updates2)
self.covar_cost = T.dot(self.W0s.T, self.W0s) - T.eye(
self.W0s.shape[1])
self.covar_cost = T.sum(T.sum(self.covar_cost**2, axis=1),
axis=0) # Frobenius
# combined loss for AE and LR
combined_params = [
self.W0s,
self.bW0s,
self.bW1s,
self.W1s,
# self.V0s, self.V1s, self.bV0, self.bV1]
self.V1s,
self.bV0,
self.bV1
]
self.combined_cost = (
(np.float32(1) - self.lambda_param) * self.ae_cost +
self.lambda_param * self.lr_cost + self.gamma * self.covar_cost)
combined_updates = self.RMSprop2(cost=self.combined_cost,
params=combined_params,
lr=self.learning_rate)
givens_combined = {
self.input_taskdata:
X_train_s[index * self.batch_size:(index + 1) * self.batch_size],
my_y:
y_train_s[index * self.batch_size:(index + 1) * self.batch_size]
}
f_train_combined = theano.function(
[index],
# [self.combined_cost, self.ae_cost, self.lr_cost, self.lr_cost],
[self.combined_cost, self.ae_cost, self.lr_cost, self.covar_cost],
givens=givens_combined,
updates=combined_updates,
allow_input_downcast=False)
# optimization loop
start_time = time.time()
ae_last_cost = np.inf
lr_last_cost = np.inf
no_improve_steps = 0
acc_train, acc_val = 0., 0.
for i_epoch in range(self.max_epochs):
if i_epoch == 1:
epoch_dur = time.time() - start_time
total_mins = (epoch_dur * self.max_epochs) / 60
hs, mins = divmod(total_mins, 60)
print("Max estimated duration: %i hours and %i minutes" %
(hs, mins))
# AE
n_batches = train_samples // self.batch_size
for i in range(n_batches):
# lr_cur_cost = f_train_lr(i)[0]
# ae_cur_cost = lr_cur_cost
combined_cost, ae_cur_cost, lr_cur_cost, covar_cost = f_train_combined(
i)
# evaluate epoch cost
if ae_last_cost - ae_cur_cost < 0.1:
no_improve_steps += 1
else:
ae_last_cost = ae_cur_cost
no_improve_steps = 0
# logistic
lr_last_cost = lr_cur_cost
acc_train = self.score(X_train, y_train)
acc_val, prfs_val = self.score(X_val, y_val, return_prfs=True)
print(
'E:%i, ae_cost:%.4f, lr_cost:%.4f, covar_cost:%.4f, train_score:%.2f, vald_score:%.2f, ae_badsteps:%i'
% (i_epoch + 1, ae_cur_cost, lr_cur_cost, covar_cost,
acc_train, acc_val, no_improve_steps))
# if (i_epoch % 10 == 0):
self.dbg_ae_cost_.append(ae_cur_cost)
self.dbg_lr_cost_.append(lr_cur_cost)
self.dbg_combined_cost_.append(combined_cost)
self.dbg_epochs_.append(i_epoch + 1)
self.dbg_ae_nonimprovesteps.append(no_improve_steps)
self.dbg_acc_train_.append(acc_train)
self.dbg_acc_val_.append(acc_val)
self.dbg_prfs_.append(prfs_val)
# save paramters from last 100 iterations
# if i_epoch > (self.max_epochs - 100):
# print('Param pool!')
param_pool = self.get_param_pool()
self.params_from_last_iters.append(param_pool)
total_mins = (time.time() - start_time) / 60
hs, mins = divmod(total_mins, 60)
print("Final duration: %i hours and %i minutes" % (hs, mins))
return self
def main():
n_samples, step = 40, 20
load_data = LoadHAR(add_pitch=False,
add_roll=False,
add_filter=False,
n_samples=n_samples,
step=step,
normalize='channels',
comp_magnitude=False,
simple_labels=True,
common_labels=False)
conf = ModelConfiguration()
conf.load_datasets([load_data.uci_hapt], label_limit=6)
user_idx = -1
user = None # 'UCI HAPT10'
if user is not None:
train_idx = conf.users != user
test_idx = conf.users == user
conf.cv = ((train_idx, test_idx), )
print('Testing user: %s' % user)
else:
# Cross validate on users
conf.cv = LeavePLabelOut(conf.users, p=1)
# Divide into K folds balanced on labels
# conf.cv = StratifiedKFold(conf.users, n_folds=10)
# And shuffle
# conf.cv = StratifiedShuffleSplit(np.argmax(conf.y, axis=1), n_iter=1, test_size=0.1, random_state=None)
# Pure shuffle
# conf.cv = ShuffleSplit(conf.y.shape[0], n_iter=2, test_size=0.1)
for train_index, test_index in conf.cv:
conf.user = user
model = tconvRNN(n_in=(n_samples, conf.n_features),
n_filters=[64, 64, 64, 64],
filter_sizes=[5] * 4,
pool_sizes=[0] * 4,
n_hidden=[128, 128],
conv_dropout=0.3,
rnn_in_dropout=0.0,
rnn_hid_dropout=0.0,
output_dropout=0.5,
n_out=conf.n_classes,
trans_func=leaky_rectify,
out_func=softmax,
stats=conf.stats)
if len(conf.cv) > 1:
user_idx += 1
if len(conf.cv) == len(conf.user_names):
conf.user = conf.user_names[user_idx]
else:
conf.user = conf.name + ' K_%d' % user_idx
# Generate root path and edit
root_path = model.get_root_path()
model.root_path = "%s_cv_%s_%s" % (root_path, conf.d, conf.user)
paths.path_exists(model.root_path)
rmdir(root_path)
scriptpath = path.realpath(__file__)
filename = path.basename(scriptpath)
print(scriptpath, model.root_path, filename)
shutil.copy(scriptpath, model.root_path + '/' + filename)
train = TrainModel(model=model,
anneal_lr=0.9,
anneal_lr_freq=1,
output_freq=1,
pickle_f_custom_freq=100,
f_custom_eval=None)
train.pickle = False
train.write_to_logger(
"Using StratifiedShuffleSplit with n_iter=1, test_size=0.1, random_state=None"
)
conf.run(train_index,
test_index,
lr=0.003,
n_epochs=500,
model=model,
train=train,
load_data=load_data,
batch_size=100)
