def fully_connected_flow(self, x, DEBUG=True):
for _ in range(self.fully_connected_flow_layers):
temp = self.fully_connected_res_flow(x, DEBUG=True)
x = temp
x = Dense(1, activation='sigmoid')(x)
if DEBUG:
print(x.get_shape().as_list())
return x
def fully_connected_flow(self, x, outputsize):
print(" shape after all conv layers ", x.get_shape().as_list())
for _ in range(self.fully_connected_flow_layers):
temp = self.fully_connected_res_flow(x)
x = temp
print(" fully residual block at ", _, " ",
x.get_shape().as_list())
x = Dense(outputsize, activation='linear')(x)
print(" final shape ", x.get_shape().as_list())
return x
def mnist_mlp_model(input):
# Using Keras model API with Flatten results in split ngraph at Flatten() or Reshape() op.
# Use tf.reshape instead
known_shape = input.get_shape()[1:]
size = np.prod(known_shape)
print('size', size)
y = tf.reshape(input, [-1, size])
# y = Flatten()(input)
y = Dense(input_shape=[1, 784], units=30, use_bias=True)(y)
y = PolyAct()(y)
y = Dense(units=10, use_bias=True, name="output")(y)
known_shape = y.get_shape()[1:]
size = np.prod(known_shape)
print('size', size)
return y
def build_decoder(input_shape,
latent_dim,
hidden_dim,
filters,
kernels,
conv_activation=None,
dense_activation=None):
"""
Return decoder as model
input_shape: shape of the input data
latent_dim : dimension of the latent variable
hidden_dim : dimension of the dense hidden layer
filters: list of the sizes of the filters used for this model
list of the size of the kernels used for each filter of this model
conv_activation: type of activation layer used after the convolutional layers
dense_activation: type of activation layer used after the dense layers
"""
input_layer = Input(shape=(latent_dim, ))
h = Dense(hidden_dim, activation=dense_activation)(input_layer)
h = PReLU()(h)
w = int(np.ceil(input_shape[0] / 2**(len(filters))))
h = Dense(w * w * filters[-1], activation=dense_activation)(h)
h = PReLU()(h)
h = Reshape((w, w, filters[-1]))(h)
for i in range(len(filters) - 1, -1, -1):
h = Conv2DTranspose(filters[i], (kernels[i], kernels[i]),
activation=conv_activation,
padding='same',
strides=(2, 2))(h)
h = PReLU()(h)
h = Conv2DTranspose(filters[i], (kernels[i], kernels[i]),
activation=conv_activation,
padding='same')(h)
h = PReLU()(h)
h = Conv2D(input_shape[-1], (3, 3), activation='sigmoid',
padding='same')(h)
cropping = int(h.get_shape()[1] - input_shape[0])
if cropping > 0:
print('in cropping')
if cropping % 2 == 0:
h = Cropping2D(cropping / 2)(h)
else:
h = Cropping2D(((cropping // 2, cropping // 2 + 1),
(cropping // 2, cropping // 2 + 1)))(h)
return Model(input_layer, h)
print(l_cov1.get_shape())
l_pool2 = MaxPooling1D(5)(l_cov2)
print(l_pool2.get_shape())
l_cov3 = Conv1D(128, 5, activation='relu')(l_pool2)
print(l_cov3.get_shape())
l_pool3 = MaxPooling1D(35)(l_cov3) # global max pooling
print(l_pool3.get_shape())
l_flat = Flatten()(l_pool3)
print(l_flat.get_shape())
l_dense = Dense(128, activation='relu')(l_flat)
print(l_dense.get_shape())
preds = Dense(len(macronum), activation='softmax')(l_dense)
print(preds.get_shape())
# #### Here we attempt to train the neural network, but the kernel dies every time we use the keras.fit() to train it.
# It says that it will take too much memory. There are similar cases on google, and people advised to not use Sagemaker to run it and use our personal computers instead.
# In[17]:
model = Model(sequence_input, preds)
model.compile(loss='categorical_crossentropy',
optimizer='rmsprop',
metrics=['acc'])
