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
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
def _u_step(self, h_cur, u):
s = self._step_basic(h_cur, u)
with layer.mixed(size=1, bias_attr=False,
act=Act.SequenceSoftmax()) as h_weights:
h_weights += layer.identity_projection(s)
applied_weights = layer.scaling(input=u, weight=h_weights)
u_ctx = layer.pooling(input=applied_weights,
pooling_type=paddle.pooling.Sum())
return u_ctx
def sum_weighted_loss(loss, weight, size=1):
"""Loss has input batch_size x image_size, weight has input batch_size x weight
( i * w ) / sum(W)
The output is normalized weighted loss
"""
weighted_loss = pd.mixed(
size=size, input=[pd.dotmul_operator(a=loss, b=weight, scale=1.0)])
weight_fac = pd.sum_cost(input=weight)
weight_fac = util_layers.math_op(input=weight_fac, act=pd.activation.Inv())
weighted_loss = pd.scaling(input=loss, weight=weight_fac)
weighted_loss = pd.sum_cost(input=weighted_loss)
return weighted_loss
def test_math_layer(self):
addto = layer.addto(input=[pixel, pixel])
linear_comb = layer.linear_comb(weights=weight, vectors=hidden, size=10)
interpolation = layer.interpolation(
input=[hidden, hidden], weight=score)
bilinear = layer.bilinear_interp(input=conv, out_size_x=4, out_size_y=4)
power = layer.power(input=pixel, weight=score)
scaling = layer.scaling(input=pixel, weight=score)
slope = layer.slope_intercept(input=pixel)
tensor = layer.tensor(a=pixel, b=pixel, size=1000)
cos_sim = layer.cos_sim(a=pixel, b=pixel)
trans = layer.trans(input=tensor)
print layer.parse_network(addto, linear_comb, interpolation, power,
scaling, slope, tensor, cos_sim, trans)
def ns_ele_l2_cost(input,
label,
weight,
height,
width,
num_channel=None,
interp='nearest'):
assert interp in image_resize_func.keys()
# make sure all the input label and weight have the same size
input = pd.bilinear_interp(input=input,
out_size_x=width,
out_size_y=height)
label = image_resize_func[interp](input=label,
out_size_x=width,
out_size_y=height)
weight = image_resize_func[interp](input=weight,
out_size_x=width,
out_size_y=height)
# reshape the orignal layer
# input has shape c x h x w change to h x w x c
input_ts = pd.transpose(input=input,
trans_order=[1, 2, 0],
height=height,
width=width)
input_rs = pd.resize(input=input_ts, size=num_channel, height=1, width=1)
label_ts = pd.transpose(input=label,
trans_order=[1, 2, 0],
height=height,
width=width)
label_rs = pd.resize(input=label_ts, size=num_channel, height=1, width=1)
weight_rs = pd.resize(input=weight, size=1, height=1, width=1)
cost_rs = pd.mse_cost(input=input_rs, label=label_rs)
sqrt_l2_cost = util_layers.math_op(input=cost_rs, act=pd.activation.Sqrt())
sqrt_l2_cost = pd.mixed(
size=1,
input=[pd.dotmul_operator(a=sqrt_l2_cost, b=weight_rs, scale=1.0)])
sqrt_l2_cost = pd.resize(input=sqrt_l2_cost,
size=height * width,
height=height,
width=width)
weight_fac = pd.sum_cost(input=weight)
weight_fac = util_layers.math_op(input=weight_fac, act=pd.activation.Inv())
sqrt_l2_cost = pd.scaling(input=sqrt_l2_cost, weight=weight_fac)
cost = pd.sum_cost(input=sqrt_l2_cost)
return cost
def _attention_flow(self, h, u):
bs = layer.recurrent_group(input=[h, layer.StaticInput(u)],
step=self._h_step,
reverse=False)
b_weights = layer.mixed(act=Act.SequenceSoftmax(),
bias_attr=False,
input=layer.identity_projection(bs))
h_step_scaled = layer.scaling(input=h, weight=b_weights)
h_step = layer.pooling(input=h_step_scaled,
pooling_type=paddle.pooling.Sum())
h_expr = layer.expand(input=h_step, expand_as=h)
u_expr = layer.recurrent_group(input=[h, layer.StaticInput(u)],
step=self._u_step,
reverse=False)
g = self._beta(h, u_expr, h_expr)
return g
def test_math_layer(self):
addto = layer.addto(input=[pixel, pixel])
linear_comb = layer.linear_comb(
weights=combine_weight, vectors=hidden, size=10)
interpolation = layer.interpolation(
input=[hidden, hidden], weight=score)
bilinear = layer.bilinear_interp(input=conv, out_size_x=4, out_size_y=4)
power = layer.power(input=pixel, weight=score)
scaling = layer.scaling(input=pixel, weight=score)
slope = layer.slope_intercept(input=pixel)
tensor = layer.tensor(a=pixel, b=pixel, size=1000)
cos_sim = layer.cos_sim(a=pixel, b=pixel)
trans = layer.trans(input=tensor)
print layer.parse_network([
addto, linear_comb, interpolation, power, scaling, slope, tensor,
cos_sim, trans
])
def ele_norm_cost(input,
label,
weight,
height=None,
width=None,
num_channel=None,
cost_type='l1'):
if height > 1 and width > 1:
input = pd.bilinear_interp(input=input,
out_size_x=width,
out_size_y=height)
label = pd.bilinear_interp(input=label,
out_size_x=width,
out_size_y=height)
if weight:
weight = pd.nearest_interp(input=weight,
out_size_x=width,
out_size_y=height)
size = height * width * num_channel
if weight:
input = pd.mixed(
size=size,
input=[pd.dotmul_operator(a=input, b=weight, scale=1.0)])
label = pd.mixed(
size=size,
input=[pd.dotmul_operator(a=label, b=weight, scale=1.0)])
cost = cost_func[cost_type](input=input, label=label)
fac = pd.sum_cost(input=weight)
fac = util_layers.math_op(input=fac, act=pd.activation.Inv())
cost = pd.scaling(input=cost, weight=fac)
cost = pd.sum_cost(input=cost)
else:
cost = cost_func[cost_type](input=input, label=label)
fac = 1.0 / float(height * width)
cost = pd.slope_intercept(input=cost, slope=fac, intercept=0.0)
cost = pd.sum_cost(input=cost)
return cost
