Support paddle 2.x (#538)
* 2.x model
* model test pass
* fix data
* fix soundfile with flac support
* one thread dataloader test pass
* export feasture size
add trainer and utils
add setup model and dataloader
update travis using Bionic dist
* add venv; test under venv
* fix unittest; train and valid
* add train and config
* add config and train script
* fix ctc cuda memcopy error
* fix imports
* fix train valid log
* fix dataset batch shuffle shift start from 1
fix rank_zero_only decreator error
close tensorboard when train over
add decoding config and code
* test process can run
* test with decoding
* test and infer with decoding
* fix infer
* fix ctc loss
lr schedule
sortagrad
logger
* aishell egs
* refactor train
add aishell egs
* fix dataset batch shuffle and add batch sampler log
print model parameter
* fix model and ctc
* sequence_mask make all inputs zeros, which cause grad be zero, this is a bug of LessThanOp
add grad clip by global norm
add model train test notebook
* ctc loss
remove run prefix
using ord value as text id
* using unk when training
compute_loss need text ids
ord id using in test mode, which compute wer/cer
* fix tester
* add lr_deacy
refactor code
* fix tools
* fix ci
add tune
fix gru model bugs
add dataset and model test
* fix decoding
* refactor repo
fix decoding
* fix musan and rir dataset
* refactor io, loss, conv, rnn, gradclip, model, utils
* fix ci and import
* refactor model
add export jit model
* add deploy bin and test it
* rm uselss egs
* add layer tools
* refactor socket server
new model from pretrain
* remve useless
* fix instability loss and grad nan or inf for librispeech training
* fix sampler
* fix libri train.sh
* fix doc
* add license on cpp
* fix doc
* fix libri script
* fix install
* clip 5 wer 7.39, clip 400 wer 7.54, 1.8 clip 400 baseline 7.49
4 years ago
|
|
|
// Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
|
|
|
|
//
|
|
|
|
// Licensed under the Apache License, Version 2.0 (the "COPYING.APACHE2.0");
|
Support paddle 2.x (#538)
* 2.x model
* model test pass
* fix data
* fix soundfile with flac support
* one thread dataloader test pass
* export feasture size
add trainer and utils
add setup model and dataloader
update travis using Bionic dist
* add venv; test under venv
* fix unittest; train and valid
* add train and config
* add config and train script
* fix ctc cuda memcopy error
* fix imports
* fix train valid log
* fix dataset batch shuffle shift start from 1
fix rank_zero_only decreator error
close tensorboard when train over
add decoding config and code
* test process can run
* test with decoding
* test and infer with decoding
* fix infer
* fix ctc loss
lr schedule
sortagrad
logger
* aishell egs
* refactor train
add aishell egs
* fix dataset batch shuffle and add batch sampler log
print model parameter
* fix model and ctc
* sequence_mask make all inputs zeros, which cause grad be zero, this is a bug of LessThanOp
add grad clip by global norm
add model train test notebook
* ctc loss
remove run prefix
using ord value as text id
* using unk when training
compute_loss need text ids
ord id using in test mode, which compute wer/cer
* fix tester
* add lr_deacy
refactor code
* fix tools
* fix ci
add tune
fix gru model bugs
add dataset and model test
* fix decoding
* refactor repo
fix decoding
* fix musan and rir dataset
* refactor io, loss, conv, rnn, gradclip, model, utils
* fix ci and import
* refactor model
add export jit model
* add deploy bin and test it
* rm uselss egs
* add layer tools
* refactor socket server
new model from pretrain
* remve useless
* fix instability loss and grad nan or inf for librispeech training
* fix sampler
* fix libri train.sh
* fix doc
* add license on cpp
* fix doc
* fix libri script
* fix install
* clip 5 wer 7.39, clip 400 wer 7.54, 1.8 clip 400 baseline 7.49
4 years ago
|
|
|
// you may not use this file except in compliance with the License.
|
|
|
|
// You may obtain a copy of the License at
|
|
|
|
//
|
|
|
|
// http://www.apache.org/licenses/LICENSE-2.0
|
|
|
|
//
|
|
|
|
// Unless required by applicable law or agreed to in writing, software
|
|
|
|
// distributed under the License is distributed on an "AS IS" BASIS,
|
|
|
|
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
|
|
// See the License for the specific language governing permissions and
|
|
|
|
// limitations under the License.
|
|
|
|
|
|
|
|
#ifndef CTC_BEAM_SEARCH_DECODER_H_
|
|
|
|
#define CTC_BEAM_SEARCH_DECODER_H_
|
|
|
|
|
|
|
|
#include <string>
|
|
|
|
#include <utility>
|
|
|
|
#include <vector>
|
|
|
|
|
|
|
|
#include "scorer.h"
|
|
|
|
|
|
|
|
/* CTC Beam Search Decoder
|
|
|
|
|
|
|
|
* Parameters:
|
|
|
|
* probs_seq: 2-D vector that each element is a vector of probabilities
|
|
|
|
* over vocabulary of one time step.
|
|
|
|
* vocabulary: A vector of vocabulary.
|
|
|
|
* beam_size: The width of beam search.
|
|
|
|
* cutoff_prob: Cutoff probability for pruning.
|
|
|
|
* cutoff_top_n: Cutoff number for pruning.
|
|
|
|
* ext_scorer: External scorer to evaluate a prefix, which consists of
|
|
|
|
* n-gram language model scoring and word insertion term.
|
|
|
|
* Default null, decoding the input sample without scorer.
|
|
|
|
* Return:
|
|
|
|
* A vector that each element is a pair of score and decoding result,
|
|
|
|
* in desending order.
|
|
|
|
*/
|
|
|
|
std::vector<std::pair<double, std::string>> ctc_beam_search_decoding(
|
|
|
|
const std::vector<std::vector<double>> &probs_seq,
|
|
|
|
const std::vector<std::string> &vocabulary,
|
|
|
|
size_t beam_size,
|
|
|
|
double cutoff_prob = 1.0,
|
|
|
|
size_t cutoff_top_n = 40,
|
|
|
|
Scorer *ext_scorer = nullptr,
|
|
|
|
size_t blank_id = 0);
|
|
|
|
|
|
|
|
|
|
|
|
/* CTC Beam Search Decoder for batch data
|
|
|
|
|
|
|
|
* Parameters:
|
|
|
|
* probs_seq: 3-D vector that each element is a 2-D vector that can be used
|
|
|
|
* by ctc_beam_search_decoder().
|
|
|
|
* vocabulary: A vector of vocabulary.
|
|
|
|
* beam_size: The width of beam search.
|
|
|
|
* num_processes: Number of threads for beam search.
|
|
|
|
* cutoff_prob: Cutoff probability for pruning.
|
|
|
|
* cutoff_top_n: Cutoff number for pruning.
|
|
|
|
* ext_scorer: External scorer to evaluate a prefix, which consists of
|
|
|
|
* n-gram language model scoring and word insertion term.
|
|
|
|
* Default null, decoding the input sample without scorer.
|
|
|
|
* Return:
|
|
|
|
* A 2-D vector that each element is a vector of beam search decoding
|
|
|
|
* result for one audio sample.
|
|
|
|
*/
|
|
|
|
std::vector<std::vector<std::pair<double, std::string>>>
|
|
|
|
ctc_beam_search_decoding_batch(
|
|
|
|
const std::vector<std::vector<std::vector<double>>> &probs_split,
|
|
|
|
const std::vector<std::string> &vocabulary,
|
|
|
|
size_t beam_size,
|
|
|
|
size_t num_processes,
|
|
|
|
double cutoff_prob = 1.0,
|
|
|
|
size_t cutoff_top_n = 40,
|
|
|
|
Scorer *ext_scorer = nullptr,
|
|
|
|
size_t blank_id = 0);
|
|
|
|
|
|
|
|
/**
|
|
|
|
* Store the root and prefixes for decoder
|
|
|
|
*/
|
|
|
|
|
|
|
|
class CtcBeamSearchDecoderStorage {
|
|
|
|
public:
|
|
|
|
PathTrie *root = nullptr;
|
|
|
|
std::vector<PathTrie *> prefixes;
|
|
|
|
|
|
|
|
CtcBeamSearchDecoderStorage() {
|
|
|
|
// init prefixes' root
|
|
|
|
this->root = new PathTrie();
|
|
|
|
this->root->log_prob_b_prev = 0.0;
|
|
|
|
// The score of root is in log scale.Since the prob=1.0, the prob score
|
|
|
|
// in log scale is 0.0
|
|
|
|
this->root->score = root->log_prob_b_prev;
|
|
|
|
// std::vector<PathTrie *> prefixes;
|
|
|
|
this->prefixes.push_back(root);
|
|
|
|
};
|
|
|
|
|
|
|
|
~CtcBeamSearchDecoderStorage() {
|
|
|
|
if (root != nullptr) {
|
|
|
|
delete root;
|
|
|
|
root = nullptr;
|
|
|
|
}
|
|
|
|
};
|
|
|
|
};
|
|
|
|
|
|
|
|
/**
|
|
|
|
* The ctc beam search decoder, support batchsize >= 1
|
|
|
|
*/
|
|
|
|
class CtcBeamSearchDecoderBatch {
|
|
|
|
public:
|
|
|
|
CtcBeamSearchDecoderBatch(const std::vector<std::string> &vocabulary,
|
|
|
|
size_t batch_size,
|
|
|
|
size_t beam_size,
|
|
|
|
size_t num_processes,
|
|
|
|
double cutoff_prob,
|
|
|
|
size_t cutoff_top_n,
|
|
|
|
Scorer *ext_scorer,
|
|
|
|
size_t blank_id);
|
|
|
|
|
|
|
|
~CtcBeamSearchDecoderBatch();
|
|
|
|
void next(const std::vector<std::vector<std::vector<double>>> &probs_split,
|
|
|
|
const std::vector<std::string> &has_value);
|
|
|
|
|
|
|
|
std::vector<std::vector<std::pair<double, std::string>>> decode();
|
|
|
|
|
|
|
|
void reset_state(size_t batch_size,
|
|
|
|
size_t beam_size,
|
|
|
|
size_t num_processes,
|
|
|
|
double cutoff_prob,
|
|
|
|
size_t cutoff_top_n);
|
|
|
|
|
|
|
|
private:
|
|
|
|
std::vector<std::string> vocabulary;
|
|
|
|
size_t batch_size;
|
|
|
|
size_t beam_size;
|
|
|
|
size_t num_processes;
|
|
|
|
double cutoff_prob;
|
|
|
|
size_t cutoff_top_n;
|
|
|
|
Scorer *ext_scorer;
|
|
|
|
size_t blank_id;
|
|
|
|
std::vector<std::unique_ptr<CtcBeamSearchDecoderStorage>>
|
|
|
|
decoder_storage_vector;
|
|
|
|
};
|
|
|
|
|
|
|
|
/**
|
|
|
|
* function for chunk decoding
|
|
|
|
*/
|
|
|
|
void ctc_beam_search_decode_chunk(
|
|
|
|
PathTrie *root,
|
|
|
|
std::vector<PathTrie *> &prefixes,
|
|
|
|
const std::vector<std::vector<double>> &probs_seq,
|
|
|
|
const std::vector<std::string> &vocabulary,
|
|
|
|
size_t beam_size,
|
|
|
|
double cutoff_prob,
|
|
|
|
size_t cutoff_top_n,
|
|
|
|
Scorer *ext_scorer,
|
|
|
|
size_t blank_id);
|
|
|
|
|
|
|
|
std::vector<std::pair<double, std::string>> get_decode_result(
|
|
|
|
std::vector<PathTrie *> &prefixes,
|
|
|
|
const std::vector<std::string> &vocabulary,
|
|
|
|
size_t beam_size,
|
|
|
|
Scorer *ext_scorer);
|
|
|
|
|
|
|
|
/**
|
|
|
|
* free the CtcBeamSearchDecoderStorage
|
|
|
|
*/
|
|
|
|
void free_storage(std::unique_ptr<CtcBeamSearchDecoderStorage> &storage);
|
|
|
|
|
|
|
|
/**
|
|
|
|
* initialize the root
|
|
|
|
*/
|
|
|
|
void ctc_beam_search_decode_chunk_begin(PathTrie *root, Scorer *ext_scorer);
|
|
|
|
|
|
|
|
#endif // CTC_BEAM_SEARCH_DECODER_H_
|