How to assign a float weight to every training sample?

[lqniunjunlper]

Is the following right?
input_types: add a dense_vector(1)
network config: add a data_layer weight=data_layer( , size=1)
cost layer: coeff=weight?

[lqniunjunlper]

Or using scaling_layer?

[lqniunjunlper]

hi @reyoung
After adding a weight layer, i got the following error:
So strange, input order is not right!

[reyoung]

@NIULQfromNJU In dataprovider, we calculate the input order by reference order of data_layer in TrainerConfig. But user also can yield a dictionary in dataprovider in latest code. See http://www.paddlepaddle.org/doc_cn/ui/data_provider/pydataprovider2.html#id1 's 显式指定返回的数据对应关系.

Sample weight should be added to cost layer. It is ok to add ascaling_layer to cost. And we should add a weight parameter to cost layer later.

The sample config should be

feas = data_layer(name='feas', size=100)

hidden = fc_layer(input=feas, size=100)

predict = fc_layer(input=hidden, size=10, act=SoftmaxActivation())

cost = classification_cost(input=predict, label=data_layer(name='label', size=10))

cost_after_scale = scaling_layer(input=cost, weight=data_layer(name='weight', size=1))

outputs(cost_after_scale)

and the data provider could be

def process(settings, filename):
     with open(filename) as f:
           ...
           yield { 'feas': feas, 'label': label, 'weight': weight}

[lqniunjunlper]

hi @reyoung
you mean that previous paddle (version 0.8.0b0) does not support this explicit way but the latest version?
Another question: which input_types can be used to represent the float weight? Is the following right?

initializer

input_types=[dense_vector(1)]

process

weight = []
weight.append(1.0)
yield weight

config

weight=data_layer(name='weight', size=1)

thx~

[reyoung]

@NIULQfromNJU It seems that the usage above should be ok. And explicit provide data is supported since the version 0.8.0b1 .