def cal_hidden_keras(self, test, layernum):
if layernum == 0:
acx = test
else:
acx = get_activations_single_layer(self.model, np.array(test),
self.layerName(layernum - 1))
units = int(
int(self.model.layers[layernum].trainable_weights[0].shape[1]) / 4)
if len(acx.shape) < len(test.shape):
acx = np.array([acx])
inp = keras.layers.Input(batch_shape=(None, acx.shape[1],
acx.shape[2]),
name="input")
rnn, s, c = keras.layers.LSTM(units,
return_sequences=True,
stateful=False,
return_state=True,
name="RNN")(inp)
states = keras.models.Model(inputs=[inp], outputs=[s, c, rnn])
for layer in states.layers:
if layer.name == "RNN":
layer.set_weights(self.model.layers[layernum].get_weights())
h_t_keras, c_t_keras, rnn = states.predict(acx)
return rnn
def displayInfo(self, test):
model = self._code_repr_model
print("method name: %s" % (test))
conf = get_activations_single_layer(model, np.array([test]), self.layerName(-1))
print("current confidence: %.2f\n" % (conf))
if conf >= 0.5:
return (1, conf)
else:
return (-1, 1 - conf)
def displayInfo(self, test):
model = self.model
text = self.fromIDToText(test)
print("review content: %s" % (text))
conf = get_activations_single_layer(model, np.array([test]),
self.layerName(-1))
print("current confidence: %.2f\n" % (conf))
if conf >= 0.5:
return (1, conf)
else:
return (-1, 1 - conf)
def cal_hidden_state(self, test, layernum):
if layernum == 0:
acx = test
else:
acx = get_activations_single_layer(self.model, np.array([test]),
self.layerName(layernum - 1))
units = int(
int(self.model.layers[layernum].trainable_weights[0].shape[1]) / 4)
# print("No units: ", units)
# get weight
W = self.model.layers[layernum].get_weights()[0]
U = self.model.layers[layernum].get_weights()[1]
b = self.model.layers[layernum].get_weights()[2]
W_i = W[:, :units]
W_f = W[:, units:units * 2]
W_c = W[:, units * 2:units * 3]
W_o = W[:, units * 3:]
U_i = U[:, :units]
U_f = U[:, units:units * 2]
U_c = U[:, units * 2:units * 3]
U_o = U[:, units * 3:]
b_i = b[:units]
b_f = b[units:units * 2]
b_c = b[units * 2:units * 3]
b_o = b[units * 3:]
# calculate the hidden state value
h_t = np.zeros((self.imagesize, units))
c_t = np.zeros((self.imagesize, units))
f_t = np.zeros((self.imagesize, units))
h_t0 = np.zeros((1, units))
c_t0 = np.zeros((1, units))
for i in range(0, self.imagesize):
f_gate = hard_sigmoid(
np.dot(acx[i, :], W_f) + np.dot(h_t0, U_f) + b_f)
i_gate = hard_sigmoid(
np.dot(acx[i, :], W_i) + np.dot(h_t0, U_i) + b_i)
o_gate = hard_sigmoid(
np.dot(acx[i, :], W_o) + np.dot(h_t0, U_o) + b_o)
new_C = np.tanh(np.dot(acx[i, :], W_c) + np.dot(h_t0, U_c) + b_c)
c_t0 = f_gate * c_t0 + i_gate * new_C
h_t0 = o_gate * np.tanh(c_t0)
c_t[i, :] = c_t0
h_t[i, :] = h_t0
f_t[i, :] = f_gate
return [h_t, c_t, f_t]
def cal_hidden_state(self, x_mn, x_api, x_token):
print(self.layerName(6))
acx = get_activations_single_layer(self._code_repr_model, x_mn, x_api, x_token, self.layerName(2))
units = int(int(self._code_repr_model.layers[6].trainable_weights[0].shape[1]) / 4)
# print("No units: ", units)
# lstm_layer = model.layers[1]
W = self._code_repr_model.layers[6].get_weights()[0]
U = self._code_repr_model.layers[6].get_weights()[1]
b = self._code_repr_model.layers[6].get_weights()[2]
W_i = W[:, :units]
W_f = W[:, units: units * 2]
W_c = W[:, units * 2: units * 3]
W_o = W[:, units * 3:]
U_i = U[:, :units]
U_f = U[:, units: units * 2]
U_c = U[:, units * 2: units * 3]
U_o = U[:, units * 3:]
b_i = b[:units]
b_f = b[units: units * 2]
b_c = b[units * 2: units * 3]
b_o = b[units * 3:]
# calculate the hidden state value
h_t = np.zeros((6, units))
c_t = np.zeros((6, units))
f_t = np.zeros((6, units))
h_t0 = np.zeros((1, units))
c_t0 = np.zeros((1, units))
for i in range(0, 6):
f_gate = hard_sigmoid(np.dot(acx[i, :], W_f) + np.dot(h_t0, U_f) + b_f)
i_gate = hard_sigmoid(np.dot(acx[i, :], W_i) + np.dot(h_t0, U_i) + b_i)
o_gate = hard_sigmoid(np.dot(acx[i, :], W_o) + np.dot(h_t0, U_o) + b_o)
new_C = np.tanh(np.dot(acx[i, :], W_c) + np.dot(h_t0, U_c) + b_c)
c_t0 = f_gate * c_t0 + i_gate * new_C
h_t0 = o_gate * np.tanh(c_t0)
c_t[i, :] = c_t0
h_t[i, :] = h_t0
f_t[i, :] = f_gate
return h_t, c_t, f_t
def get_lstm_state(self, test):
return get_activations_single_layer(self.lstm_state_model,
np.array([test]), 'lstm_1')