def create_forward_flow(self):
"""
Implements the forward propagation for the model:
CONV2D -> RELU -> MAXPOOL -> CONV2D -> RELU -> MAXPOOL -> FLATTEN -> FULLYCONNECTED
Note that for simplicity and grading purposes, we'll hard-code some values
such as the stride and kernel (filter) sizes.
Normally, functions should take these values as function parameters.
Arguments:
X -- input dataset placeholder, of shape (input size, number of examples)
parameters -- python dictionary containing your parameters "W1", "W2"
the shapes are given in initialize_parameters
Returns:
Z3 -- the output of the last LINEAR unit
"""
X_input = Input(self.input_dims)
X = X_input.copy()
for i, d_params in enumerate(self.conv_layer_params):
X = Conv2D(filters=d_params['n'],
kernel_size=d_params['kernel_size'],
strides=d_params.get('strides', (1, 1)),
padding=d_params.get('padding', 'SAME'),
name='conv{}'.format(i))(X)
X = BatchNormalization(axis=3, name='bn{}'.format(X))(X)
X = Activation(d_params.get('activation', 'relu'))(X)
if not d_params.get('has_pool', True):
continue
if d_params.get('pool_op', 'max') == 'max':
X = MaxPooling2D(pool_size=d_params.get('pool_size', (4, 4)),
strides=d_params.get('strides_pool', (1, 1)),
padding=d_params.get('padding_pool', 'same'),
name='pool{}'.format(i))(X)
else:
raise NotImplementedError
# Flatten result of convolution layers
X = Flatten()(X)
# Add hidden layers of fully connected part
for i, d_params in enumerate(self.fc_layer_params):
X = Dense(d_params['nhn'],
activation=d_params.get('activation', 'relu'),
name='fc{}'.format(i))(X)
# Add output layer
X = Dense(self.fc_layer_params[-1]['ny'],
activation=self.fc_layer_params[-1]['activation'],
name='output_layer')(X)
model = Model(inputs=X_input, outputs=X, name=self.name)
return X_input, X, model
