def setUp(self):
"""
Creating true multi-layer perceptron with one hidden layer
"""
np.random.seed(1337)
batch_size = 16
input_size = 20
hidden_size = 40
output_size = 5
self.x_val = np.random.randn(batch_size, input_size)
self.w1_val = np.random.randn(input_size, hidden_size)
self.w2_val = np.random.randn(hidden_size, output_size)
self.tf_x = tf.constant(self.x_val)
self.tf_w1 = tf.constant(self.w1_val)
self.tf_w2 = tf.constant(self.w2_val)
self.tf_h = tf.nn.sigmoid(self.tf_x @ self.tf_w1)
self.tf_o = tf.nn.sigmoid(self.tf_h @ self.tf_w2)
self.my_x = ad.Variable(self.x_val, name="x_val")
self.my_w1 = ad.Variable(self.w1_val, name="w1_val")
self.my_w2 = ad.Variable(self.w2_val, name="w2_val")
self.var_h = ad.Sigmoid(self.my_x @ self.my_w1)
self.var_o = ad.Sigmoid(self.var_h @ self.my_w2)
self.my_graph = self.var_o
self.tf_graph = self.tf_o
def setUp(self):
"""
Graph looks like this:
x_val w_val
\ /
MatMul
|
Sigmoid
x_val.shape = (2, 3)
w_val.shape = (3, 5)
MatMul.shape = (2, 5)
Sigmoid.shape = (2, 5)
"""
np.random.seed(1337)
self.x_val = np.random.randn(2, 3)
self.w_val = np.random.randn(3, 5)
self.b_val = np.random.randn(5)
self.tf_x = tf.constant(self.x_val, dtype=tf.float64)
self.tf_w = tf.constant(self.w_val, dtype=tf.float64)
self.tf_b = tf.constant(self.b_val, dtype=tf.float64)
self.tf_mul = self.tf_x @ self.tf_w + self.tf_b
self.tf_graph = tf.nn.sigmoid(self.tf_mul)
self.my_x = ad.Variable(self.x_val, name="x_val")
self.my_w = ad.Variable(self.w_val, name="w_val")
self.my_b = ad.Variable(self.b_val, name="b_val")
self.var_mul = self.my_x @ self.my_w + self.my_b
self.my_graph = ad.Sigmoid(self.var_mul)
def output_layer(self, S_Old, X):
S = S_Old
val_z = ad.MatMul(X, self._Uz) + ad.MatMul(S, self._Wz) + self._bz
Z = ad.Sigmoid(val_z)
#print("Z",Z())
val_g = ad.MatMul(X, self._Ug) + ad.MatMul(S, self._Wg) + self._bg
G = ad.Sigmoid(val_g)
#print("G",G())
val_r = ad.MatMul(X, self._Ur) + ad.MatMul(S, self._Wr) + self._br
R = ad.Sigmoid(val_r)
#print("R",R())
val_h = ad.MatMul(X, self._Uh) + ad.MatMul(S * R, self._Wh) + self._bh
H = ad.Sigmoid(val_h)
#print("H",H())
S_New = ((ad.Variable(np.ones_like(G.eval())) - G) * H) + (Z * S)
#print("Snew",S_New())
#val = (-G * ((ad.Variable(np.ones_like(G.eval()))- G ) * H) * (self._Ug+self._Wg*self._Wg*temp)) + (((ad.Variable(np.ones_like(G.eval()))- G ) * H*(ad.Variable(np.ones_like(H.eval()))- H ))*(self._Uz+self._Wh*self._Wg*R*temp)+ (self._Wg*S*(ad.Variable(np.ones_like(R.eval()))- R ) * R*(self._Ug+self._Wg*self._Wg*temp))) +(Z*self._Wg*temp) + (S*(self._Ug+self._Wg*self._Wg*temp))
#print(val())
return S_New
def output(self, X):
S = ad.Sigmoid(ad.MatMul(X, self._W) + self._B)
#print("S:",S())
S1 = self.layer1.output_layer(S, X)
S_list = []
S_list.append(S1)
for i in range(self.number_of_layers):
S_list.append(self.layers[i].output_layer(S_list[i], X))
S_final = S_list[-1]
#print(S_final.shape)
#print(self.Wf.shape)
#print(self.Bf.shape)
val = ad.MatMul(S_final, self._Wf) + self._Bf
#print("The output:",val())
return val
def test_sigmoid(self):
my_graph = ad.Sigmoid(self.my_w0)
tf_graph = tf.nn.sigmoid(self.tf_w0)
wrt_vars = [self.my_w0]
tf_vars = [self.tf_w0]
utils.custom_test(self, my_graph, wrt_vars, tf_graph, tf_vars)