def forward_propagation(self, x):
"""
Runs a forward propagation of the model for a batch of examples, activating
all layers besides the top one.
The function should return the linear input to the top layer, i.e.,
non-normalized scores, with higher scores corresponding to greater
probability of an image belonging a particular class.
Inputs:
x: batch with shape [batch_size, 1, 28, 28]
Output:
A Tensor with shape [batch_size, 10] containing predicted scores (logits)
"""
# normalize data to improve performance
x = x / 127.5 - 1.0
# flatten the input image data to a 2d matrix in the shape [N, d]
x = nn.reshape(x, [x.shape[0], -1], inplace=True)
# define the structure of the network to get the logits result from x
#######################################################################
"*** YOUR CODE HERE ***"
Y1 = nn.matmul(x, self.param["w"]) + self.param["b"]
Y1_out = nn.relu(Y1)
Y2 = nn.matmul(Y1_out, self.param["w1"]) + self.param["b1"]
Y2_out = nn.relu(Y2)
logits = nn.matmul(Y2_out, self.param["w2"]) + self.param["b2"]
#######################################################################
return logits
def forward_propagation(self, x):
"""
Runs a forward propagation of the model for a batch of examples, activating
all layers besides the top one.
The function should return the linear input to the top layer, i.e.,
non-normalized scores, with higher scores corresponding to greater
probability of an image belonging a particular class.
Inputs:
x: batch with shape [batch_size, 1, 28, 28]
Output:
A Tensor with shape [batch_size, 10] containing predicted scores (logits)
"""
# normalize data to improve performance
x = x / 127.5 - 1.0
# flatten the input image data to a 2d matrix in the shape [N, d]
x = nn.reshape(x, [x.shape[0], -1], inplace=True)
# define the structure of the network to get the logits result from x
#######################################################################
"*** YOUR CODE HERE ***"
logits = None
#print(self.param['w1'])
#print(self.param['b1'])
internal_hidden_layer_z1 = nn.relu(
nn.matmul(x, self.param['w1']) + self.param['b1'])
internal_hidden_layer_z2 = nn.sigmoid(
nn.matmul(internal_hidden_layer_z1, self.param['w2']) +
self.param['b2'])
internal_hidden_layer_z3 = nn.prelu(
nn.matmul(internal_hidden_layer_z2, self.param['w3']) +
self.param['b3'], 0.0001)
logits = nn.matmul(internal_hidden_layer_z3,
self.param['w4']) + self.param['b4']
#######################################################################
return logits