classes,
others=None,
min_size_given=None,
rng=rng)
print "data.shape"
print data.shape
print "classes.shape"
print classes.shape
print "numpy.unique(classes)"
print numpy.unique(classes)
# partition the data
train_set_x_org, train_set_y_org, valid_set_x_org, valid_set_y_org, test_set_x_org, test_set_y_org = cl.partition_train_valid_test(
data, classes, ratio=(2, 1, 1), rng=rng)
print "valid set shape"
print valid_set_x_org.shape
print valid_set_y_org.shape
# normalization
train_set_x_org, data_min, data_max = cl.normalize_col_scale01(
train_set_x_org, tol=1e-10)
valid_set_x_org, _, _ = cl.normalize_col_scale01(valid_set_x_org,
tol=1e-10,
data_min=data_min,
data_max=data_max)
test_set_x_org, _, _ = cl.normalize_col_scale01(test_set_x_org,
tol=1e-10,
data_min=data_min,
print numpy.unique(classes)
classes_unique,classes=cl.change_class_labels(classes)
print numpy.unique(classes)
# set random state
#numpy.random.seed(1000)
rng=numpy.random.RandomState(2000)
data,classes,others=cl.balance_sample_size(data,classes,others=None,min_size_given=None,rng=rng)
print data.shape
print numpy.unique(classes)
# partition the data
train_set_x_org,train_set_y_org,valid_set_x_org,valid_set_y_org,test_set_x_org,test_set_y_org=cl.partition_train_valid_test(data,classes,ratio=(2,1,1),rng=rng)
# normalization
train_set_x_org,data_min,data_max=cl.normalize_col_scale01(train_set_x_org,tol=1e-10)
valid_set_x_org,_,_=cl.normalize_col_scale01(valid_set_x_org,tol=1e-10,data_min=data_min,data_max=data_max)
test_set_x_org,_,_=cl.normalize_col_scale01(test_set_x_org,tol=1e-10,data_min=data_min,data_max=data_max)
# train
# setting the parameter
pretrain_lr=0.01
finetune_lr=0.1
alpha=0.1
lambda2=1.0
alpha1=0.001
alpha2=0.0
n_hidden=[256,64,16]
print data.shape
print numpy.unique(classes)
kfolds = 10
ind_folds = cl.kfold_cross_validation(classes,
k=kfolds,
shuffle=True,
rng=rng)
for i in range(kfolds):
test_set_x_org = data[ind_folds == i, :]
test_set_y_org = classes[ind_folds == i]
train_set_x_org, train_set_y_org, valid_set_x_org, valid_set_y_org, _, _ = cl.partition_train_valid_test(
data[ind_folds != i, :],
classes[ind_folds != i],
ratio=(3, 1, 0),
rng=rng)
# normalization
train_set_x_org, data_min, data_max = cl.normalize_col_scale01(
train_set_x_org, tol=1e-10)
valid_set_x_org, _, _ = cl.normalize_col_scale01(valid_set_x_org,
tol=1e-10,
data_min=data_min,
data_max=data_max)
test_set_x_org, _, _ = cl.normalize_col_scale01(test_set_x_org,
tol=1e-10,
data_min=data_min,
data_max=data_max)
print data.shape
print numpy.unique(classes)
# normalization
data[:,0:wid]=numpy.absolute(data[:,0:wid]) # change the RNA-seq minus track to positive values
num_feat=len(features)
cl.normalize_matrical_samples(data,num_feat,method="scale01")
# 10-fold CV
kfolds=10
ind_folds=cl.kfold_cross_validation(classes,k=kfolds,shuffle=True,rng=rng)
for i in range(kfolds):
test_set_x_org=data[ind_folds==i,:]
test_set_y_org=classes[ind_folds==i]
train_set_x_org,train_set_y_org,valid_set_x_org,valid_set_y_org,_,_=cl.partition_train_valid_test(data[ind_folds!=i,:],classes[ind_folds!=i],ratio=(3,1,0),rng=rng)
# parameter setting
pretrain_lr=0.04 # 0.1, 0.08, 0.06, 0.04, 0.02, 0.01
finetune_lr=0.04 # 0.1, 0.08, 0.06, 0.04, 0.02, 0.01
alpha=0.1
pretraining_epochs=100 # 20, 50, 100
training_epochs=100 # 50, 100, 200, 500
nkerns=[4,4,8] # [4,4,4], [4,6,8]
batch_size=100 # 100, 200
receptive_fields=((1,4),(1,4),(1,4)) # ((1,4),(1,4),(1,4)), ((1,4),(1,4),(1,2)), ((1,8),(1,4),(1,2))
poolsizes=((1,4),(1,4),(1,2)) # ((1,4),(1,4),(1,4)), ((1,4),(1,4),(1,2)), ((1,6),(1,4),(1,2))
full_hidden_sub=[16] # [8], [16], [32], [64]
full_hidden_all=[32] # [8], [16], [32], [64]
max_num_epoch_change_learning_rate=50 #20, 50, 80 not crucial
max_num_epoch_change_rate=0.8
learning_rate_decay_rate=0.8
filename="/home/yifengli/research/dnashape/result/Data_1000bp.txt";
data=numpy.loadtxt(filename,delimiter='\t',dtype='float16')
filename="/home/yifengli/research/dnashape/result/Classes_1000bp.txt";
classes=numpy.loadtxt(filename,delimiter='\t',dtype=str)
filename="/home/yifengli/research/dnashape/result/Features.txt";
features=numpy.loadtxt(filename,delimiter='\t',dtype=str)
# change class labels
given=["Enhancer","EnhancerFalse"]
data,classes=cl.take_some_classes(data,classes,given)
given={"Enhancer":0,"EnhancerFalse":1}
classes=cl.change_class_labels_to_given(classes,given)
train_set_x_org,train_set_y_org,valid_set_x_org,valid_set_y_org,test_set_x_org,test_set_y_org \
=cl.partition_train_valid_test(data,classes,ratio=(1,1,1))
del data
gc_collect()
rng=numpy.random.RandomState(1000)
numpy.warnings.filterwarnings('ignore')
# train
classifier,training_time=convolutional_mlp.train_model( train_set_x_org=train_set_x_org, train_set_y_org=train_set_y_org,
valid_set_x_org=valid_set_x_org, valid_set_y_org=valid_set_y_org,
n_row_each_sample=4,
learning_rate=0.1, alpha=0.01, n_epochs=1000, rng=rng,
nkerns=[4,4,8],batch_size=500,
receptive_fields=((2,8),(2,8),(2,2)),poolsizes=((1,8),(1,8),(1,2)),full_hidden=8)
# test
test_set_y_pred,test_set_y_pred_prob,test_time=convolutional_mlp.test_model(classifier,test_set_x_org)
def train(self,train_set_x_org=None, train_set_y_org=None,features=None,
num_samplings=100,
randomize_method="random_rescaling",
random_rescaling_alpha=0.5,
random_sampling_portion=0.66,
learning_rate=0.1, alpha=0.01,
lambda1=0.001, lambda2=1.0, alpha1=0.001, alpha2=0.01,
n_hidden=[256,128,16], n_epochs=1000, batch_size=100,
activation_func="relu", rng=numpy.random.RandomState(100),
dfs_select_method="top_num", dfs_threshold=0.001, dfs_top_num=10,
max_num_epoch_change_learning_rate=80,max_num_epoch_change_rate=0.8,learning_rate_decay_rate=0.8):
"""
Train the randomized DFS.
num_samplings: int, number of reruns of DFS.
randomize_method: string, the randomizing method, can be one of {"random_rescaling","random_sampling","random_sampling_and_random_rescaling"}.
"""
if randomize_method=="random_rescaling":
train_set_x,train_set_y,valid_set_x,valid_set_y,_,_=cl.partition_train_valid_test(train_set_x_org, train_set_y_org,ratio=(3,1,0), rng=rng)
self.num_samplings=num_samplings
self.randomize_method=randomize_method
self.random_rescaling_alpha=random_rescaling_alpha
self.random_sampling_portion=random_sampling_portion
self.learning_rate=learning_rate
self.alpha=alpha
self.lambda1=lambda1
self.lambda2=lambda2
self.alpha1=alpha1
self.alpha2=alpha2
self.n_hidden=n_hidden
self.n_epochs=n_epochs
self.batch_size=batch_size
self.activation_func="relu"
self.max_num_epoch_change_learning_rate=max_num_epoch_change_learning_rate
self.max_num_epoch_change_rate=max_num_epoch_change_rate
self.learning_rate_decay_rate=learning_rate_decay_rate
self.features=features
self.classifiers=[]
self.feature_counts=numpy.zeros(shape=(self.n_in,),dtype=int)
self.feature_weights=numpy.zeros(shape=(self.n_in,self.num_samplings),dtype=float)
self.training_time=0
self.classes_unique=numpy.unique(train_set_y_org)
self.rescale_Ws=[]
for ns in range(self.num_samplings):
print "The {0}-th run of randomized DFS...".format(ns)
rng_ns=numpy.random.RandomState(ns)
# generate a subsample of data points
if randomize_method=="random_sampling" or randomize_method=="random_sampling_and_random_rescaling":
# sample data
train_set_x,train_set_y,ind_train,_=cl.sampling(train_set_x_org,train_set_y_org,others=None,portion=random_sampling_portion,max_size_given=None,rng=rng_ns)
valid_set_x=numpy.delete(train_set_x_org,ind_train,axis=0)
valid_set_y=numpy.delete(train_set_y_org,ind_train)
# reorder the sampled data
train_set_x,train_set_y,_=cl.sort_classes(train_set_x,train_set_y)
if randomize_method=="random_rescaling" or randomize_method=="random_sampling_and_random_rescaling":
rescale_ws=rng_ns.uniform(low=random_rescaling_alpha,high=1.0,size=(1,self.num_features))
train_set_x=train_set_x*rescale_ws ### may multiplify, not division, leave it for future!!!!!!!!!!
valid_set_x=valid_set_x*rescale_ws
self.rescale_Ws.extend([rescale_ws])
# run DFS
classifier,training_time=deep_feat_select_mlp.train_model(train_set_x_org=train_set_x, train_set_y_org=train_set_y,
valid_set_x_org=valid_set_x, valid_set_y_org=valid_set_y,
learning_rate=learning_rate, alpha=alpha, lambda1=lambda1, lambda2=lambda2,
alpha1=alpha1, alpha2=alpha2, n_hidden=n_hidden,
n_epochs=n_epochs, batch_size=batch_size, activation_func=activation_func, rng=rng_ns,
max_num_epoch_change_learning_rate=max_num_epoch_change_learning_rate,
max_num_epoch_change_rate=max_num_epoch_change_rate,
learning_rate_decay_rate=learning_rate_decay_rate)
features_selected,logind_selected,weights_selected,weights=deep_feat_select_mlp.select_features(classifier,features,select_method=dfs_select_method,threshold=dfs_threshold,top_num=dfs_top_num)
print weights
self.classifiers.append(classifier)
self.feature_counts=self.feature_counts+numpy.array(logind_selected,dtype=int)
print self.feature_counts
self.feature_weights[:,ns]=weights
self.training_time=self.training_time+training_time
# final clean up
self.feature_importance=self.feature_counts/self.num_samplings
print self.feature_importance
return self.feature_importance,self.feature_weights,self.training_time
filename = "/home/yifengli/research/dnashape/result/Data_1000bp.txt"
data = numpy.loadtxt(filename, delimiter='\t', dtype='float16')
filename = "/home/yifengli/research/dnashape/result/Classes_1000bp.txt"
classes = numpy.loadtxt(filename, delimiter='\t', dtype=str)
filename = "/home/yifengli/research/dnashape/result/Features.txt"
features = numpy.loadtxt(filename, delimiter='\t', dtype=str)
# change class labels
given = ["Enhancer", "EnhancerFalse"]
data, classes = cl.take_some_classes(data, classes, given)
given = {"Enhancer": 0, "EnhancerFalse": 1}
classes = cl.change_class_labels_to_given(classes, given)
train_set_x_org,train_set_y_org,valid_set_x_org,valid_set_y_org,test_set_x_org,test_set_y_org \
=cl.partition_train_valid_test(data,classes,ratio=(1,1,1))
del data
gc_collect()
rng = numpy.random.RandomState(1000)
numpy.warnings.filterwarnings('ignore')
# train
classifier, training_time = convolutional_mlp.train_model(
train_set_x_org=train_set_x_org,
train_set_y_org=train_set_y_org,
valid_set_x_org=valid_set_x_org,
valid_set_y_org=valid_set_y_org,
n_row_each_sample=4,
learning_rate=0.1,
alpha=0.01,
n_epochs=1000,
