def forward(self, X):
hidden_size = self.hidden_size
input_dim = self.input_dim
num_classes = self.num_classes
#--@TODO slim.stack
# Verbose way:
#x = slim.fully_connected(x, 32, scope='fc/fc_1')
#x = slim.fully_connected(x, 64, scope='fc/fc_2')
#x = slim.fully_connected(x, 128, scope='fc/fc_3')
## Equivalent, TF-Slim way using slim.stack:
#slim.stack(x, slim.fully_connected, [32, 64, 128], scope='fc')
#--------so no local variable, make it get_var or class member var so we can share
if isinstance(X, tf.Tensor):
w_h = melt.get_weights('w_h', [input_dim, hidden_size])
else:
with tf.device('/cpu:0'):
w_h = melt.get_weights('w_h', [input_dim, hidden_size])
b_h = melt.get_bias('b_h', [hidden_size])
w_o = melt.get_weights('w_o', [hidden_size, num_classes])
b_o = melt.get_bias('b_o', [num_classes])
py_x = melt.mlp_forward(X,
w_h,
b_h,
w_o,
b_o,
activation=self.activation)
return py_x
def mlp_nobias(x, hidden_size, output_size, activation=tf.nn.relu, scope=None):
scope = 'mlp_nobias' if scope is None else scope
with tf.variable_scope(scope):
input_dim = utils.last_dimension(x.get_shape(), min_rank=2)
if isinstance(x, tf.Tensor):
w_h = melt.get_weights('w_h', [input_dim, hidden_size])
else:
with tf.device('/cpu:0'):
w_h = melt.get_weights('w_h', [input_dim, hidden_size])
w_o = melt.get_weights('w_o', [hidden_size, output_size])
return melt.mlp_forward_nobias(x, w_h, w_o, activation)
def mlp(x, hidden_size, output_size, activation=tf.nn.relu, scope=None):
scope = 'mlp' if scope is None else scope
with tf.variable_scope(scope):
hidden = fully_connected(x, hidden_size, activation)
w_o = melt.get_weights('w_o', [hidden_size, output_size])
b_o = melt.get_bias('b_o', [output_size])
return tf.nn.xw_plus_b(hidden, w_o, b_o)
def fully_connected(x, output_size, activation=tf.nn.relu, scope=None):
#@TODO -1 or last dim ? NotImplementedError("Negative indices are currently unsupported")
#input_dim = tf.shape(x)[-1]
#@TODO how is slim.fully_connected get inputdim and use..
#below will now work int() argument must be a string or a number, not 'Tensor' [input_dim, output_size])
#input_dim = tf.shape(x)[1]
#check contrib\layers\python\layers\layers.py
scope = 'fc' if scope is None else scope
with tf.variable_scope(scope):
input_dim = utils.last_dimension(x.get_shape(), min_rank=2)
if isinstance(x, tf.Tensor):
w_h = melt.get_weights('w_h', [input_dim, output_size])
else:
with tf.device('/cpu:0'):
w_h = melt.get_weights('w_h', [input_dim, output_size])
b_h = melt.get_bias('b_h', [output_size])
return activation(melt.matmul(x, w_h) + b_h)
