paddle.infer output multiple outputs

[pengwangucla]

For paddle api v2,

The infer is only for a single output:

gen_result=paddle.infer(
    output_layer=beam_gen,
    parameters=parameters,
    input=gen_data, 
    field=['id', 'value'])

Even for multiple fields, what should I do if I want to simultaneously obtain multiple outputs from layer1 and layer2, is it possible to upgrade the function as

gen_result=paddle.infer(
    output_layer=[beam_gen1, beam_gen2],
    parameters=parameters,
    input=gen_data, 
    field=['id', 'value'])

Also set the gen results to be a list.

@reyoung @luotao1

[lcy-seso]

Hi @pengwangucla the last large modification of C++ code of beam search is made by me. There are some restrictions on getting the output of a layer in recurrent_layer_group.

In the old version (v1), this can be done by directly put the layer you need in "output of the recurrent layer group", just like you do in the above example. I am not sure whether it works in V2 version. I will check it.

[lcy-seso]

I check the C++ codes. I am still not sure whether the v2 infer function can return multiple outputs. But, currently, in the C++ codes, except the first output of beam search (the generated sequence), the other output can be any layer defined in the step function, but only the top1 result during beam search is returned.

[lcy-seso]

@luotao1 has told me, v2 can print multiple outputs.

I find you use beam search as an example. Beam search is implemented in recurrent layer group, that works very differently from others layers in PaddlePaddle.

I guess you do not want outputs of multiples layers in recurrent layer group, but just want to get outputs of more than one layers of a network not implemented by recurrent layer group?

[pengwangucla]

I guess you do not want outputs of multiples layers in recurrent layer group, but just want to get outputs of more than one layers of a network not implemented by recurrent layer group?

I just use beam search as an example, I just need to output different intermedia output inside a forward network.

So is the usage the same as I wrote ? just put a list for output_layer ?

[luotao1]

@pengwangucla You usage as follows is correct.

gen_result=paddle.infer(
    output_layer=[beam_gen1, beam_gen2],
    parameters=parameters,
    input=gen_data, 
    field=['id', 'value'])

You can run print gen_result to see the result. And the final print order of result is:
beam_gen1.id, beam_gen2.id, beam_gen1.value, beam_gen2.value

I will add the usage of multiple outputs in the annotation of paddle.infer.