def get_classifier(input_data, num_classes):
"""Get VGG classifier layers as fc layers."""
flatten = sym.flatten(data=input_data, name="flatten")
fc6 = sym.dense(data=flatten, units=4096, name="fc6")
relu6 = sym.relu(data=fc6, name="relu6")
drop6 = sym.dropout(data=relu6, rate=0.5, name="drop6")
fc7 = sym.dense(data=drop6, units=4096, name="fc7")
relu7 = sym.relu(data=fc7, name="relu7")
drop7 = sym.dropout(data=relu7, rate=0.5, name="drop7")
fc8 = sym.dense(data=drop7, units=num_classes, name="fc8")
return fc8
def lstm_gate(op, U, b, x, num_units):
"""
op - nonlinearity operation
x - input tensor of shape (1,a)
U - weight matrix of shape (a,b)
b - bias (1,b)
return tensor of shape (1,b)
"""
return op(sym.dense(x, U, b, units=num_units))
def lstm_and_dense_layer(num_timesteps: int, num_inputs: int, num_classes: int, num_hidden: int,
dense_init=np.random.normal):
"""
Use `model` with `num_hidden` LSTM units, translate them `num_classes` classes using a dense layer.
"""
W = Variable("W", init=dense_init(size=[num_hidden, num_classes]))
b = Variable("b", init=dense_init(size=[1, num_classes]))
X, outputs = lstm_layer(num_timesteps, num_inputs, num_units=num_hidden)
cls = sym.dense(outputs[-1], W, b, units=num_classes)
return X, cls
def _get_model(dshape):
data = sym.Variable('data', shape=dshape)
fc1 = sym.dense(data, units=dshape[-1]*2, use_bias=True)
left, right = sym.split(fc1, indices_or_sections=2, axis=1)
return sym.Group(((left + 1), (right - 1)))
def _get_model(dshape):
data = sym.Variable('data', shape=dshape)
fc1 = sym.dense(data, units=dshape[-1] * 2, use_bias=True)
left, right = sym.split(fc1, indices_or_sections=2, axis=1)
return sym.Group(((left + 1), (right - 1)))
