def _attention(self, direct, cur_token, prev, to_apply, to_apply_proj):
with layer.mixed(size=cur_token.size,
bias_attr=Attr.Param(direct + '.bp', initial_std=0.),
act=Act.Linear()) as proj:
proj += layer.full_matrix_projection(input=cur_token,
param_attr=Attr.Param(direct +
'.wp'))
proj += layer.full_matrix_projection(input=prev,
param_attr=Attr.Param(direct +
'.wr'))
expanded = layer.expand(input=proj, expand_as=to_apply)
att_context = layer.addto(input=[expanded, to_apply_proj],
act=Act.Tanh(),
bias_attr=False)
att_weights = layer.fc(input=att_context,
param_attr=Attr.Param(direct + '.w'),
bias_attr=Attr.Param(direct + '.b',
initial_std=0.),
act=Act.SequenceSoftmax(),
size=1)
scaled = layer.scaling(input=to_apply, weight=att_weights)
applied = layer.pooling(input=scaled,
pooling_type=paddle.pooling.Sum())
return applied
weight = layer.data(name='weight', type=data_type.dense_vector(1))
combine_weight = layer.data(name='weight_combine',
type=data_type.dense_vector(10))
score = layer.data(name='score', type=data_type.dense_vector(1))
hidden = layer.fc(input=pixel,
size=100,
act=activation.Sigmoid(),
param_attr=attr.Param(name='hidden'))
inference = layer.fc(input=hidden, size=10, act=activation.Softmax())
conv = layer.img_conv(input=pixel,
filter_size=1,
filter_size_y=1,
num_channels=8,
num_filters=16,
act=activation.Linear())
class ImageLayerTest(unittest.TestCase):
def test_conv_layer(self):
conv_shift = layer.conv_shift(a=pixel, b=score)
print layer.parse_network(conv, conv_shift)
def test_pooling_layer(self):
maxpool = layer.img_pool(input=conv,
pool_size=2,
num_channels=16,
padding=1,
pool_type=pooling.Max())
spp = layer.spp(input=conv,
pyramid_height=2,