design/sequence decoder #4905
design/sequence decoder #4905
Conversation
doc/design/ops/sequence_decoder.md
Outdated
| ## Beam Search | ||
| Beam search is a heuristic search algorithm that explores a graph by expanding the most promising node in a limited set. | ||
|
|
||
| It is the core component of Sequence Decoder. |
There was a problem hiding this comment.
Sequence Decoder => *sequence decoder*
doc/design/ops/sequence_decoder.md
Outdated
|
|
||
| It is the core component of Sequence Decoder. | ||
|
|
||
| In the original implementation of `RecurrentGradientMachine`, the beam search is a method in RNN, |
There was a problem hiding this comment.
Are there more than one implementations of RecrurentGradientMachine? What is the other implementations besides the "original" one?
doc/design/ops/sequence_decoder.md
Outdated
| @@ -0,0 +1,436 @@ | |||
| # A LoD-based Sequence Decoder | |||
| In tasks such as machine translation and image to text, | |||
| a **sequence decoder** is necessary to generate sequences. | |||
There was a problem hiding this comment.
a [sequence decoder](https://github.com/PaddlePaddle/book/blob/develop/08.machine_translation/README.md) is necessary
doc/design/ops/sequence_decoder.md
Outdated
| @@ -0,0 +1,436 @@ | |||
| # A LoD-based Sequence Decoder | |||
There was a problem hiding this comment.
Design: Sequence Decoder Generating LoDTensors
| For example, the RNN sates, candidates IDs and probabilities of beam search can be represented as `LoDTensors`; | ||
| the selected candidate's IDs in each time step can be stored in a `TensorArray`, and `Packed` to the sentences translated. | ||
|
|
||
| ## Changing LoD's absolute offset to relative offsets |
There was a problem hiding this comment.
This is not part of the design. This is an issue. I copy-n-pasted the content into issue #4945
| ``` | ||
|
|
||
| the first level represents that there are two sequences, | ||
| their offsets in the second-level LoD is `[0, 3)` and `[3, 5)`. |
There was a problem hiding this comment.
Is here [3, 5) or [3, 6), I think it is the latter one.
| the first level represents that there are two sequences, | ||
| their offsets in the second-level LoD is `[0, 3)` and `[3, 5)`. | ||
|
|
||
| The second level is the same with the relative offset example because the lower level is a tensor. |
There was a problem hiding this comment.
the same as or the same to？
| # for example | ||
| # decoder_mem.lod is | ||
| # [[0 1 3], | ||
| # [0 1 3 6]] |
There was a problem hiding this comment.
Why is decoder_mem here a sequence? Usually, the memory of an RNN is its hidden state in the last time step. Is memory here a TensorArray? and we decide to memorize hidden states in all the previous time step? If so, is this design compatible to the memory in dynamic RNN?
But I guess, maybe this memory has something different?
| # its tensor content is [a1 a2 a3 a4 a5] | ||
| # which means there are 2 sentences to translate | ||
| # - the first sentence has 1 translation prefixes, the offsets are [0, 1) | ||
| # - the second sentence has 2 translation prefixes, the offsets are [1, 3) and [3, 6) |
There was a problem hiding this comment.
Does here means memory need to memorize the entire unfinished prefixes? (This is required for beam search.)
| # the following has 2, 3, 2, 3 candidates | ||
| # the encoder_ctx_expanded's content will be | ||
| # [a1 a1 a2 a2 a3 a3 a3 a4 a4 a5 a5 a5] | ||
| encoder_ctx_expanded = pd.lod_expand(encoder_ctx, target_word) |
There was a problem hiding this comment.
The name target word is confusing. In the generation, we do not have target word.
| # which means there are 2 sentences to translate | ||
| # - the first sentence has 1 translation prefixes, the offsets are [0, 1) | ||
| # - the second sentence has 2 translation prefixes, the offsets are [1, 3) and [3, 6) | ||
| # the target_word.lod is |
There was a problem hiding this comment.
The name target word for generation is really confusing. For the generation, we only have source words to be translated into target words.
doc/design/ops/sequence_decoder.md
Outdated
| bias=None, | ||
| act=pd.activation.Softmax()) | ||
| # topk_scores, a tensor, [None, k] | ||
| topk_scores, topk_ids = pd.top_k(scores) |
There was a problem hiding this comment.
This top_k is special. It is not only a simple "selecting top k items", but also "selecting the top k from a distribution to form a new batch". How do we handle this?
doc/design/ops/sequence_decoder.md
Outdated
| # selected_ids is the selected candidates that will be append to the translation | ||
| # selected_scores is the scores of the selected candidates | ||
| # generated_scores is the score of the translations(with candidates appended) | ||
| selected_ids, selected_scores, generated_scores = decoder.beam_search( |
There was a problem hiding this comment.
First, I want to leave some of my questions here for our discussion.
The difficulties for dynamic beam search through some preliminary operators lies in (maybe not limited to):
- how to loop.
- maybe this is done by dynamic RNN currently, but the
loopis a generally used operation, should it be independent of RNN?
- maybe this is done by dynamic RNN currently, but the
- how to stop the loop (the condition operation?).
- Samples in a batch may hit different conditions, as a result, the batch size is dynamically changing.
- Maybe all branches of a condition will be executed? I am not sure what is the design of the current condition operator, and do we decide to use it in beam search?
- How to construct the beam?
- Dynamic expansion to form a larger batch but have to track the entire unfinished prefixes.
- How the track the unfinished prefixes, and who to track this? It seems that in the current design the decoder memory tracks this?
- How to shrink the beam?
…into design/sequence_decoder
An LoD-based Sequence Decoder (Beam Search)
refactor the
beamSearchinRecurrentGradientMachine