def knockoff_stats(X, y, met, epoch, loss='mean_squared_error', verb=2):
p = X.shape[1] // 2
# implement DeepPINK
dp = Sequential()
dp.add(PairwiseConnected(input_shape=(2 * p, )))
dp.add(
Dense(p,
activation=met,
kernel_regularizer=keras.regularizers.l1(0.01)))
dp.add(
Dense(1,
activation=None,
kernel_regularizer=keras.regularizers.l1(0.01)))
dp.compile(loss=loss, optimizer=keras.optimizers.Adam())
dp.fit(X, y, epochs=epoch, batch_size=32, verbose=verb)
# calculate knockoff statistics W_j
weights = dp.get_weights()
# w = weights[1] @ weights[3]
w = np.matmul(weights[1], weights[3])
w = w.reshape(p, )
z = weights[0][:p]
z_tilde = weights[0][p:]
W = (w * z)**2 - (w * z_tilde)**2
return W
def knockoff_stats(X,
y,
met,
epoch,
l1,
lr,
loss='mean_squared_error',
verb=2):
p = X.shape[1] // 2
# implement DeepPINK
es = EarlyStopping(monitor='val_loss', patience=30, verbose=2)
dp = Sequential()
dp.add(PairwiseConnected(input_shape=(2 * p, )))
dp.add(
Dense(p,
activation=met,
kernel_regularizer=keras.regularizers.l1(l1=l1)))
dp.add(Dense(1, activation=None))
dp.compile(loss=loss, optimizer=keras.optimizers.Adam(learning_rate=lr))
dp.fit(X,
y,
epochs=epoch,
batch_size=32,
verbose=verb,
validation_split=0.1,
callbacks=[es])
# calculate knockoff statistics W_j
weights = dp.get_weights()
w = weights[1] @ weights[3]
w = w.reshape(p, )
z = weights[0][:p]
z_tilde = weights[0][p:]
W = (w * z)**2 - (w * z_tilde)**2
return W
random.seed(58 * i)
######### load data #########
ind_train = np.random.choice(np.arange(n), n_train, False)
ind_test = list(set(np.arange(n)) - set(ind_train))
Xnew_train = Xnew[ind_train, :]
Xnew_test = Xnew[ind_test, :]
y_train = y[ind_train]
y_test = y[ind_test]
indmat[:, i] = ind_train
#############################
dp = Sequential()
dp.add(PairwiseConnected(input_shape=(2 * d, )))
dp.add(
Dense(d,
activation=dnn_met,
kernel_regularizer=keras.regularizers.l1_l2(l1=0.001, l2=0.001)))
dp.add(Dropout(0.4)) # play with this number, such as 0.4, 0.6, 0.7
dp.add(
Dense(1,
activation='relu',
kernel_regularizer=keras.regularizers.l1_l2(l1=0.001, l2=0.001)))
dp.compile(loss=dnn_loss, optimizer=keras.optimizers.Adam())
dp.fit(Xnew_train,
y_train,
epochs=dnn_epoch,
batch_size=bs,
verbose=dnn_verb)
