Merge branch 'develop' of https://github.com/PaddlePaddle/DeepSpeech into thchs30_MFA

deepspeech/exps/u2/model.py

@@ -34,9 +34,12 @@ from deepspeech.models.u2 import U2Model
 from deepspeech.training.gradclip import ClipGradByGlobalNormWithLog
 from deepspeech.training.scheduler import WarmupLR
 from deepspeech.training.trainer import Trainer
+from deepspeech.utils import ctc_utils
 from deepspeech.utils import error_rate
 from deepspeech.utils import layer_tools
 from deepspeech.utils import mp_tools
+from deepspeech.utils import text_grid
+from deepspeech.utils import utility
 from deepspeech.utils.log import Log
 
 logger = Log(__name__).getlog()
@@ -278,7 +281,15 @@ class U2Trainer(Trainer):
             shuffle=False,
             drop_last=False,
             collate_fn=SpeechCollator.from_config(config))
-        logger.info("Setup train/valid/test Dataloader!")
+        # return text token id
+        config.collator.keep_transcription_text = False
+        self.align_loader = DataLoader(
+            test_dataset,
+            batch_size=config.decoding.batch_size,
+            shuffle=False,
+            drop_last=False,
+            collate_fn=SpeechCollator.from_config(config))
+        logger.info("Setup train/valid/test/align Dataloader!")
 
     def setup_model(self):
         config = self.config
@@ -353,7 +364,7 @@ class U2Tester(U2Trainer):
                 decoding_chunk_size=-1,  # decoding chunk size. Defaults to -1.
                 # <0: for decoding, use full chunk.
                 # >0: for decoding, use fixed chunk size as set.
-                # 0: used for training, it's prohibited here. 
+                # 0: used for training, it's prohibited here.
                 num_decoding_left_chunks=-1,  # number of left chunks for decoding. Defaults to -1.
                 simulate_streaming=False,  # simulate streaming inference. Defaults to False.
             ))
@@ -498,6 +509,73 @@ class U2Tester(U2Trainer):
         except KeyboardInterrupt:
             sys.exit(-1)
 
+    @paddle.no_grad()
+    def align(self):
+        if self.config.decoding.batch_size > 1:
+            logger.fatal('alignment mode must be running with batch_size == 1')
+            sys.exit(1)
+
+        # xxx.align
+        assert self.args.result_file and self.args.result_file.endswith(
+            '.align')
+
+        self.model.eval()
+        logger.info(f"Align Total Examples: {len(self.align_loader.dataset)}")
+
+        stride_ms = self.align_loader.collate_fn.stride_ms
+        token_dict = self.align_loader.collate_fn.vocab_list
+        with open(self.args.result_file, 'w') as fout:
+            # one example in batch
+            for i, batch in enumerate(self.align_loader):
+                key, feat, feats_length, target, target_length = batch
+
+                # 1. Encoder
+                encoder_out, encoder_mask = self.model._forward_encoder(
+                    feat, feats_length)  # (B, maxlen, encoder_dim)
+                maxlen = encoder_out.size(1)
+                ctc_probs = self.model.ctc.log_softmax(
+                    encoder_out)  # (1, maxlen, vocab_size)
+
+                # 2. alignment
+                ctc_probs = ctc_probs.squeeze(0)
+                target = target.squeeze(0)
+                alignment = ctc_utils.forced_align(ctc_probs, target)
+                logger.info("align ids", key[0], alignment)
+                fout.write('{} {}\n'.format(key[0], alignment))
+
+                # 3. gen praat
+                # segment alignment
+                align_segs = text_grid.segment_alignment(alignment)
+                logger.info("align tokens", key[0], align_segs)
+                # IntervalTier, List["start end token\n"]
+                subsample = utility.get_subsample(self.config)
+                tierformat = text_grid.align_to_tierformat(
+                    align_segs, subsample, token_dict)
+                # write tier
+                align_output_path = os.path.join(
+                    os.path.dirname(self.args.result_file), "align")
+                tier_path = os.path.join(align_output_path, key[0] + ".tier")
+                with open(tier_path, 'w') as f:
+                    f.writelines(tierformat)
+                # write textgrid
+                textgrid_path = os.path.join(align_output_path,
+                                             key[0] + ".TextGrid")
+                second_per_frame = 1. / (1000. /
+                                         stride_ms)  # 25ms window, 10ms stride
+                second_per_example = (
+                    len(alignment) + 1) * subsample * second_per_frame
+                text_grid.generate_textgrid(
+                    maxtime=second_per_example,
+                    intervals=tierformat,
+                    output=textgrid_path)
+
+    def run_align(self):
+        self.resume_or_scratch()
+        try:
+            self.align()
+        except KeyboardInterrupt:
+            sys.exit(-1)
+
     def load_inferspec(self):
         """infer model and input spec.
 

deepspeech/training/trainer.py

@@ -18,8 +18,8 @@ import paddle
 from paddle import distributed as dist
 from tensorboardX import SummaryWriter
 
-from deepspeech.utils import checkpoint
 from deepspeech.utils import mp_tools
+from deepspeech.utils.checkpoint import Checkpoint
 from deepspeech.utils.log import Log
 
 __all__ = ["Trainer"]
@@ -139,9 +139,9 @@ class Trainer():
             "epoch": self.epoch,
             "lr": self.optimizer.get_lr()
         })
-        checkpoint.save_parameters(self.checkpoint_dir, self.iteration
-                                   if tag is None else tag, self.model,
-                                   self.optimizer, infos)
+        self.checkpoint.add_checkpoint(self.checkpoint_dir, self.iteration
+                                       if tag is None else tag, self.model,
+                                       self.optimizer, infos)
 
     def resume_or_scratch(self):
         """Resume from latest checkpoint at checkpoints in the output 
@@ -151,7 +151,7 @@ class Trainer():
         resume training.
         """
         scratch = None
-        infos = checkpoint.load_parameters(
+        infos = self.checkpoint.load_latest_parameters(
             self.model,
             self.optimizer,
             checkpoint_dir=self.checkpoint_dir,
@@ -180,7 +180,7 @@ class Trainer():
         from_scratch = self.resume_or_scratch()
         if from_scratch:
             # save init model, i.e. 0 epoch
-            self.save(tag='init')
+            self.save(tag='init', infos=None)
 
         self.lr_scheduler.step(self.iteration)
         if self.parallel:
@@ -263,6 +263,10 @@ class Trainer():
 
         self.checkpoint_dir = checkpoint_dir
 
+        self.checkpoint = Checkpoint(
+            kbest_n=self.config.training.checkpoint.kbest_n,
+            latest_n=self.config.training.checkpoint.latest_n)
+
     @mp_tools.rank_zero_only
     def destory(self):
         """Close visualizer to avoid hanging after training"""

deepspeech/utils/checkpoint.py

@@ -11,9 +11,11 @@
 # 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.
+import glob
 import json
 import os
 import re
+from pathlib import Path
 from typing import Union
 
 import paddle
@@ -25,128 +27,260 @@ from deepspeech.utils.log import Log
 
 logger = Log(__name__).getlog()
 
-__all__ = ["load_parameters", "save_parameters"]
-
-
-def _load_latest_checkpoint(checkpoint_dir: str) -> int:
-    """Get the iteration number corresponding to the latest saved checkpoint.
-    Args:
-        checkpoint_dir (str): the directory where checkpoint is saved.
-    Returns:
-        int: the latest iteration number. -1 for no checkpoint to load.
-    """
-    checkpoint_record = os.path.join(checkpoint_dir, "checkpoint")
-    if not os.path.isfile(checkpoint_record):
-        return -1
-
-    # Fetch the latest checkpoint index.
-    with open(checkpoint_record, "rt") as handle:
-        latest_checkpoint = handle.readlines()[-1].strip()
-        iteration = int(latest_checkpoint.split(":")[-1])
-    return iteration
-
-
-def _save_record(checkpoint_dir: str, iteration: int):
-    """Save the iteration number of the latest model to be checkpoint record.
-    Args:
-        checkpoint_dir (str): the directory where checkpoint is saved.
-        iteration (int): the latest iteration number.
-    Returns:
-        None
-    """
-    checkpoint_record = os.path.join(checkpoint_dir, "checkpoint")
-    # Update the latest checkpoint index.
-    with open(checkpoint_record, "a+") as handle:
-        handle.write("model_checkpoint_path:{}\n".format(iteration))
-
-
-def load_parameters(model,
-                    optimizer=None,
-                    checkpoint_dir=None,
-                    checkpoint_path=None):
-    """Load a specific model checkpoint from disk. 
-    Args:
-        model (Layer): model to load parameters.
-        optimizer (Optimizer, optional): optimizer to load states if needed.
-            Defaults to None.
-        checkpoint_dir (str, optional): the directory where checkpoint is saved.
-        checkpoint_path (str, optional): if specified, load the checkpoint
-            stored in the checkpoint_path(prefix) and the argument 'checkpoint_dir' will 
-            be ignored. Defaults to None. 
-    Returns:
-        configs (dict): epoch or step, lr and other meta info should be saved.
-    """
-    configs = {}
-
-    if checkpoint_path is not None:
-        tag = os.path.basename(checkpoint_path).split(":")[-1]
-    elif checkpoint_dir is not None:
-        iteration = _load_latest_checkpoint(checkpoint_dir)
-        if iteration == -1:
-            return configs
-        checkpoint_path = os.path.join(checkpoint_dir, "{}".format(iteration))
-    else:
-        raise ValueError(
-            "At least one of 'checkpoint_dir' and 'checkpoint_path' should be specified!"
-        )
-
-    rank = dist.get_rank()
-
-    params_path = checkpoint_path + ".pdparams"
-    model_dict = paddle.load(params_path)
-    model.set_state_dict(model_dict)
-    logger.info("Rank {}: loaded model from {}".format(rank, params_path))
-
-    optimizer_path = checkpoint_path + ".pdopt"
-    if optimizer and os.path.isfile(optimizer_path):
-        optimizer_dict = paddle.load(optimizer_path)
-        optimizer.set_state_dict(optimizer_dict)
-        logger.info("Rank {}: loaded optimizer state from {}".format(
-            rank, optimizer_path))
-
-    info_path = re.sub('.pdparams$', '.json', params_path)
-    if os.path.exists(info_path):
-        with open(info_path, 'r') as fin:
-            configs = json.load(fin)
-    return configs
-
-
-@mp_tools.rank_zero_only
-def save_parameters(checkpoint_dir: str,
-                    tag_or_iteration: Union[int, str],
-                    model: paddle.nn.Layer,
-                    optimizer: Optimizer=None,
-                    infos: dict=None):
-    """Checkpoint the latest trained model parameters.
-    Args:
-        checkpoint_dir (str): the directory where checkpoint is saved.
-        tag_or_iteration (int or str): the latest iteration(step or epoch) number.
-        model (Layer): model to be checkpointed.
-        optimizer (Optimizer, optional): optimizer to be checkpointed.
-            Defaults to None.
-        infos (dict or None): any info you want to save.
-    Returns:
-        None
-    """
-    checkpoint_path = os.path.join(checkpoint_dir,
-                                   "{}".format(tag_or_iteration))
-
-    model_dict = model.state_dict()
-    params_path = checkpoint_path + ".pdparams"
-    paddle.save(model_dict, params_path)
-    logger.info("Saved model to {}".format(params_path))
-
-    if optimizer:
-        opt_dict = optimizer.state_dict()
+__all__ = ["Checkpoint"]
+
+
+class Checkpoint(object):
+    def __init__(self, kbest_n: int=5, latest_n: int=1):
+        self.best_records: Mapping[Path, float] = {}
+        self.latest_records = []
+        self.kbest_n = kbest_n
+        self.latest_n = latest_n
+        self._save_all = (kbest_n == -1)
+
+    def add_checkpoint(self,
+                       checkpoint_dir,
+                       tag_or_iteration,
+                       model,
+                       optimizer,
+                       infos,
+                       metric_type="val_loss"):
+        if (metric_type not in infos.keys()):
+            self._save_parameters(checkpoint_dir, tag_or_iteration, model,
+                                  optimizer, infos)
+            return
+
+        #save best
+        if self._should_save_best(infos[metric_type]):
+            self._save_best_checkpoint_and_update(
+                infos[metric_type], checkpoint_dir, tag_or_iteration, model,
+                optimizer, infos)
+        #save latest
+        self._save_latest_checkpoint_and_update(
+            checkpoint_dir, tag_or_iteration, model, optimizer, infos)
+
+        if isinstance(tag_or_iteration, int):
+            self._save_checkpoint_record(checkpoint_dir, tag_or_iteration)
+
+    def load_latest_parameters(self,
+                               model,
+                               optimizer=None,
+                               checkpoint_dir=None,
+                               checkpoint_path=None):
+        """Load a last model checkpoint from disk. 
+        Args:
+            model (Layer): model to load parameters.
+            optimizer (Optimizer, optional): optimizer to load states if needed.
+                Defaults to None.
+            checkpoint_dir (str, optional): the directory where checkpoint is saved.
+            checkpoint_path (str, optional): if specified, load the checkpoint
+                stored in the checkpoint_path(prefix) and the argument 'checkpoint_dir' will 
+                be ignored. Defaults to None. 
+        Returns:
+            configs (dict): epoch or step, lr and other meta info should be saved.
+        """
+        return self._load_parameters(model, optimizer, checkpoint_dir,
+                                     checkpoint_path, "checkpoint_latest")
+
+    def load_best_parameters(self,
+                             model,
+                             optimizer=None,
+                             checkpoint_dir=None,
+                             checkpoint_path=None):
+        """Load a last model checkpoint from disk. 
+        Args:
+            model (Layer): model to load parameters.
+            optimizer (Optimizer, optional): optimizer to load states if needed.
+                Defaults to None.
+            checkpoint_dir (str, optional): the directory where checkpoint is saved.
+            checkpoint_path (str, optional): if specified, load the checkpoint
+                stored in the checkpoint_path(prefix) and the argument 'checkpoint_dir' will 
+                be ignored. Defaults to None. 
+        Returns:
+            configs (dict): epoch or step, lr and other meta info should be saved.
+        """
+        return self._load_parameters(model, optimizer, checkpoint_dir,
+                                     checkpoint_path, "checkpoint_best")
+
+    def _should_save_best(self, metric: float) -> bool:
+        if not self._best_full():
+            return True
+
+        # already full
+        worst_record_path = max(self.best_records, key=self.best_records.get)
+        # worst_record_path = max(self.best_records.iteritems(), key=operator.itemgetter(1))[0]
+        worst_metric = self.best_records[worst_record_path]
+        return metric < worst_metric
+
+    def _best_full(self):
+        return (not self._save_all) and len(self.best_records) == self.kbest_n
+
+    def _latest_full(self):
+        return len(self.latest_records) == self.latest_n
+
+    def _save_best_checkpoint_and_update(self, metric, checkpoint_dir,
+                                         tag_or_iteration, model, optimizer,
+                                         infos):
+        # remove the worst
+        if self._best_full():
+            worst_record_path = max(self.best_records,
+                                    key=self.best_records.get)
+            self.best_records.pop(worst_record_path)
+            if (worst_record_path not in self.latest_records):
+                logger.info(
+                    "remove the worst checkpoint: {}".format(worst_record_path))
+                self._del_checkpoint(checkpoint_dir, worst_record_path)
+
+        # add the new one
+        self._save_parameters(checkpoint_dir, tag_or_iteration, model,
+                              optimizer, infos)
+        self.best_records[tag_or_iteration] = metric
+
+    def _save_latest_checkpoint_and_update(
+            self, checkpoint_dir, tag_or_iteration, model, optimizer, infos):
+        # remove the old
+        if self._latest_full():
+            to_del_fn = self.latest_records.pop(0)
+            if (to_del_fn not in self.best_records.keys()):
+                logger.info(
+                    "remove the latest checkpoint: {}".format(to_del_fn))
+                self._del_checkpoint(checkpoint_dir, to_del_fn)
+        self.latest_records.append(tag_or_iteration)
+
+        self._save_parameters(checkpoint_dir, tag_or_iteration, model,
+                              optimizer, infos)
+
+    def _del_checkpoint(self, checkpoint_dir, tag_or_iteration):
+        checkpoint_path = os.path.join(checkpoint_dir,
+                                       "{}".format(tag_or_iteration))
+        for filename in glob.glob(checkpoint_path + ".*"):
+            os.remove(filename)
+            logger.info("delete file: {}".format(filename))
+
+    def _load_checkpoint_idx(self, checkpoint_record: str) -> int:
+        """Get the iteration number corresponding to the latest saved checkpoint.
+        Args:
+            checkpoint_path (str): the saved path of checkpoint.
+        Returns:
+            int: the latest iteration number. -1 for no checkpoint to load.
+        """
+        if not os.path.isfile(checkpoint_record):
+            return -1
+
+        # Fetch the latest checkpoint index.
+        with open(checkpoint_record, "rt") as handle:
+            latest_checkpoint = handle.readlines()[-1].strip()
+            iteration = int(latest_checkpoint.split(":")[-1])
+        return iteration
+
+    def _save_checkpoint_record(self, checkpoint_dir: str, iteration: int):
+        """Save the iteration number of the latest model to be checkpoint record.
+        Args:
+            checkpoint_dir (str): the directory where checkpoint is saved.
+            iteration (int): the latest iteration number.
+        Returns:
+            None
+        """
+        checkpoint_record_latest = os.path.join(checkpoint_dir,
+                                                "checkpoint_latest")
+        checkpoint_record_best = os.path.join(checkpoint_dir, "checkpoint_best")
+
+        with open(checkpoint_record_best, "w") as handle:
+            for i in self.best_records.keys():
+                handle.write("model_checkpoint_path:{}\n".format(i))
+        with open(checkpoint_record_latest, "w") as handle:
+            for i in self.latest_records:
+                handle.write("model_checkpoint_path:{}\n".format(i))
+
+    def _load_parameters(self,
+                         model,
+                         optimizer=None,
+                         checkpoint_dir=None,
+                         checkpoint_path=None,
+                         checkpoint_file=None):
+        """Load a last model checkpoint from disk. 
+        Args:
+            model (Layer): model to load parameters.
+            optimizer (Optimizer, optional): optimizer to load states if needed.
+                Defaults to None.
+            checkpoint_dir (str, optional): the directory where checkpoint is saved.
+            checkpoint_path (str, optional): if specified, load the checkpoint
+                stored in the checkpoint_path(prefix) and the argument 'checkpoint_dir' will 
+                be ignored. Defaults to None. 
+            checkpoint_file "checkpoint_latest" or "checkpoint_best"
+        Returns:
+            configs (dict): epoch or step, lr and other meta info should be saved.
+        """
+        configs = {}
+
+        if checkpoint_path is not None:
+            tag = os.path.basename(checkpoint_path).split(":")[-1]
+        elif checkpoint_dir is not None and checkpoint_file is not None:
+            checkpoint_record = os.path.join(checkpoint_dir, checkpoint_file)
+            iteration = self._load_checkpoint_idx(checkpoint_record)
+            if iteration == -1:
+                return configs
+            checkpoint_path = os.path.join(checkpoint_dir,
+                                           "{}".format(iteration))
+        else:
+            raise ValueError(
+                "At least one of 'checkpoint_dir' and 'checkpoint_file' and 'checkpoint_path' should be specified!"
+            )
+
+        rank = dist.get_rank()
+
+        params_path = checkpoint_path + ".pdparams"
+        model_dict = paddle.load(params_path)
+        model.set_state_dict(model_dict)
+        logger.info("Rank {}: loaded model from {}".format(rank, params_path))
+
         optimizer_path = checkpoint_path + ".pdopt"
-        paddle.save(opt_dict, optimizer_path)
-        logger.info("Saved optimzier state to {}".format(optimizer_path))
+        if optimizer and os.path.isfile(optimizer_path):
+            optimizer_dict = paddle.load(optimizer_path)
+            optimizer.set_state_dict(optimizer_dict)
+            logger.info("Rank {}: loaded optimizer state from {}".format(
+                rank, optimizer_path))
+
+        info_path = re.sub('.pdparams$', '.json', params_path)
+        if os.path.exists(info_path):
+            with open(info_path, 'r') as fin:
+                configs = json.load(fin)
+        return configs
+
+    @mp_tools.rank_zero_only
+    def _save_parameters(self,
+                         checkpoint_dir: str,
+                         tag_or_iteration: Union[int, str],
+                         model: paddle.nn.Layer,
+                         optimizer: Optimizer=None,
+                         infos: dict=None):
+        """Checkpoint the latest trained model parameters.
+        Args:
+            checkpoint_dir (str): the directory where checkpoint is saved.
+            tag_or_iteration (int or str): the latest iteration(step or epoch) number.
+            model (Layer): model to be checkpointed.
+            optimizer (Optimizer, optional): optimizer to be checkpointed.
+                Defaults to None.
+            infos (dict or None): any info you want to save.
+        Returns:
+            None
+        """
+        checkpoint_path = os.path.join(checkpoint_dir,
+                                       "{}".format(tag_or_iteration))
+
+        model_dict = model.state_dict()
+        params_path = checkpoint_path + ".pdparams"
+        paddle.save(model_dict, params_path)
+        logger.info("Saved model to {}".format(params_path))
 
-    info_path = re.sub('.pdparams$', '.json', params_path)
-    infos = {} if infos is None else infos
-    with open(info_path, 'w') as fout:
-        data = json.dumps(infos)
-        fout.write(data)
+        if optimizer:
+            opt_dict = optimizer.state_dict()
+            optimizer_path = checkpoint_path + ".pdopt"
+            paddle.save(opt_dict, optimizer_path)
+            logger.info("Saved optimzier state to {}".format(optimizer_path))
 
-    if isinstance(tag_or_iteration, int):
-        _save_record(checkpoint_dir, tag_or_iteration)
+        info_path = re.sub('.pdparams$', '.json', params_path)
+        infos = {} if infos is None else infos
+        with open(info_path, 'w') as fout:
+            data = json.dumps(infos)
+            fout.write(data)

deepspeech/utils/ctc_utils.py

@@ -38,21 +38,23 @@ def remove_duplicates_and_blank(hyp: List[int], blank_id=0) -> List[int]:
     new_hyp: List[int] = []
     cur = 0
     while cur < len(hyp):
+        # add non-blank into new_hyp
         if hyp[cur] != blank_id:
             new_hyp.append(hyp[cur])
+        # skip repeat label
         prev = cur
         while cur < len(hyp) and hyp[cur] == hyp[prev]:
             cur += 1
     return new_hyp
 
 
-def insert_blank(label: np.ndarray, blank_id: int=0):
+def insert_blank(label: np.ndarray, blank_id: int=0) -> np.ndarray:
     """Insert blank token between every two label token.
 
     "abcdefg" -> "-a-b-c-d-e-f-g-"
 
     Args:
-        label ([np.ndarray]): label ids, (L).
+        label ([np.ndarray]): label ids, List[int], (L).
         blank_id (int, optional): blank id. Defaults to 0.
 
     Returns:
@@ -61,13 +63,13 @@ def insert_blank(label: np.ndarray, blank_id: int=0):
     label = np.expand_dims(label, 1)  #[L, 1]
     blanks = np.zeros((label.shape[0], 1), dtype=np.int64) + blank_id
     label = np.concatenate([blanks, label], axis=1)  #[L, 2]
-    label = label.reshape(-1)  #[2L]
-    label = np.append(label, label[0])  #[2L + 1]
+    label = label.reshape(-1)  #[2L], -l-l-l
+    label = np.append(label, label[0])  #[2L + 1], -l-l-l-
     return label
 
 
 def forced_align(ctc_probs: paddle.Tensor, y: paddle.Tensor,
-                 blank_id=0) -> list:
+                 blank_id=0) -> List[int]:
     """ctc forced alignment.
 
     https://distill.pub/2017/ctc/
@@ -77,23 +79,25 @@ def forced_align(ctc_probs: paddle.Tensor, y: paddle.Tensor,
         y (paddle.Tensor): label id sequence tensor, 1d tensor (L)
         blank_id (int): blank symbol index
     Returns:
-        paddle.Tensor: best alignment result, (T).
+        List[int]: best alignment result, (T).
     """
-    y_insert_blank = insert_blank(y, blank_id)
+    y_insert_blank = insert_blank(y, blank_id)  #(2L+1)
 
     log_alpha = paddle.zeros(
         (ctc_probs.size(0), len(y_insert_blank)))  #(T, 2L+1)
     log_alpha = log_alpha - float('inf')  # log of zero
+    # TODO(Hui Zhang): zeros not support paddle.int16
     state_path = (paddle.zeros(
-        (ctc_probs.size(0), len(y_insert_blank)), dtype=paddle.int16) - 1
-                  )  # state path
+        (ctc_probs.size(0), len(y_insert_blank)), dtype=paddle.int32) - 1
+                  )  # state path, Tuple((T, 2L+1))
 
     # init start state
-    log_alpha[0, 0] = ctc_probs[0][y_insert_blank[0]]  # Sb
-    log_alpha[0, 1] = ctc_probs[0][y_insert_blank[1]]  # Snb
+    # TODO(Hui Zhang): VarBase.__getitem__() not support np.int64
+    log_alpha[0, 0] = ctc_probs[0][int(y_insert_blank[0])]  # State-b, Sb
+    log_alpha[0, 1] = ctc_probs[0][int(y_insert_blank[1])]  # State-nb, Snb
 
-    for t in range(1, ctc_probs.size(0)):
-        for s in range(len(y_insert_blank)):
+    for t in range(1, ctc_probs.size(0)):  # T
+        for s in range(len(y_insert_blank)):  # 2L+1
             if y_insert_blank[s] == blank_id or s < 2 or y_insert_blank[
                     s] == y_insert_blank[s - 2]:
                 candidates = paddle.to_tensor(
@@ -106,11 +110,13 @@ def forced_align(ctc_probs: paddle.Tensor, y: paddle.Tensor,
                     log_alpha[t - 1, s - 2],
                 ])
                 prev_state = [s, s - 1, s - 2]
-            log_alpha[t, s] = paddle.max(candidates) + ctc_probs[t][
-                y_insert_blank[s]]
+            # TODO(Hui Zhang): VarBase.__getitem__() not support np.int64
+            log_alpha[t, s] = paddle.max(candidates) + ctc_probs[t][int(
+                y_insert_blank[s])]
             state_path[t, s] = prev_state[paddle.argmax(candidates)]
 
-    state_seq = -1 * paddle.ones((ctc_probs.size(0), 1), dtype=paddle.int16)
+    # TODO(Hui Zhang): zeros not support paddle.int16
+    state_seq = -1 * paddle.ones((ctc_probs.size(0), 1), dtype=paddle.int32)
 
     candidates = paddle.to_tensor([
         log_alpha[-1, len(y_insert_blank) - 1],  # Sb

deepspeech/utils/text_grid.py

@@ -0,0 +1,127 @@
+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# 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.
+from typing import Dict
+from typing import List
+from typing import Text
+
+import textgrid
+
+
+def segment_alignment(alignment: List[int], blank_id=0) -> List[List[int]]:
+    """segment ctc alignment ids by continuous blank and repeat label.
+
+    Args:
+        alignment (List[int]): ctc alignment id sequence. 
+            e.g. [0, 0, 0, 1, 1, 1, 2, 0, 0, 3]
+        blank_id (int, optional): blank id. Defaults to 0.
+
+    Returns:
+        List[List[int]]: token align, segment aligment id sequence. 
+            e.g. [[0, 0, 0, 1, 1, 1], [2], [0, 0, 3]]
+    """
+    # convert alignment to a praat format, which is a doing phonetics
+    # by computer and helps analyzing alignment
+    align_segs = []
+    # get frames level duration for each token
+    start = 0
+    end = 0
+    while end < len(alignment):
+        while end < len(alignment) and alignment[end] == blank_id:  # blank
+            end += 1
+        if end == len(alignment):
+            align_segs[-1].extend(alignment[start:])
+            break
+        end += 1
+        while end < len(alignment) and alignment[end - 1] == alignment[
+                end]:  # repeat label
+            end += 1
+        align_segs.append(alignment[start:end])
+        start = end
+    return align_segs
+
+
+def align_to_tierformat(align_segs: List[List[int]],
+                        subsample: int,
+                        token_dict: Dict[int, Text],
+                        blank_id=0) -> List[Text]:
+    """Generate textgrid.Interval format from alignment segmentations.
+
+    Args:
+        align_segs (List[List[int]]): segmented ctc alignment ids.
+        subsample (int): 25ms frame_length, 10ms hop_length, 1/subsample
+        token_dict (Dict[int, Text]): int -> str map.
+
+    Returns:
+        List[Text]: list of textgrid.Interval text, str(start, end, text).
+    """
+    hop_length = 10  # ms
+    second_ms = 1000  # ms
+    frame_per_second = second_ms / hop_length  # 25ms frame_length, 10ms hop_length
+    second_per_frame = 1.0 / frame_per_second
+
+    begin = 0
+    duration = 0
+    tierformat = []
+
+    for idx, tokens in enumerate(align_segs):
+        token_len = len(tokens)
+        token = tokens[-1]
+        # time duration in second
+        duration = token_len * subsample * second_per_frame
+        if idx < len(align_segs) - 1:
+            print(f"{begin:.2f} {begin + duration:.2f} {token_dict[token]}")
+            tierformat.append(
+                f"{begin:.2f} {begin + duration:.2f} {token_dict[token]}\n")
+        else:
+            for i in tokens:
+                if i != blank_id:
+                    token = i
+                    break
+            print(f"{begin:.2f} {begin + duration:.2f} {token_dict[token]}")
+            tierformat.append(
+                f"{begin:.2f} {begin + duration:.2f} {token_dict[token]}\n")
+        begin = begin + duration
+
+    return tierformat
+
+
+def generate_textgrid(maxtime: float,
+                      intervals: List[Text],
+                      output: Text,
+                      name: Text='ali') -> None:
+    """Create alignment textgrid file.
+
+    Args:
+        maxtime (float): audio duartion.
+        intervals (List[Text]): ctc output alignment. e.g. "start-time end-time word" per item.
+        output (Text): textgrid filepath.
+        name (Text, optional): tier or layer name. Defaults to 'ali'.
+    """
+    # Download Praat: https://www.fon.hum.uva.nl/praat/
+    avg_interval = maxtime / (len(intervals) + 1)
+    print(f"average second/token: {avg_interval}")
+    margin = 0.0001
+
+    tg = textgrid.TextGrid(maxTime=maxtime)
+    tier = textgrid.IntervalTier(name=name, maxTime=maxtime)
+
+    i = 0
+    for dur in intervals:
+        s, e, text = dur.split()
+        tier.add(minTime=float(s) + margin, maxTime=float(e), mark=text)
+
+    tg.append(tier)
+
+    tg.write(output)
+    print("successfully generator textgrid {}.".format(output))

deepspeech/utils/utility.py

@@ -79,3 +79,22 @@ def log_add(args: List[int]) -> float:
     a_max = max(args)
     lsp = math.log(sum(math.exp(a - a_max) for a in args))
     return a_max + lsp
+
+
+def get_subsample(config):
+    """Subsample rate from config.
+
+    Args:
+        config (yacs.config.CfgNode): yaml config
+
+    Returns:
+        int: subsample rate.
+    """
+    input_layer = config["model"]["encoder_conf"]["input_layer"]
+    assert input_layer in ["conv2d", "conv2d6", "conv2d8"]
+    if input_layer == "conv2d":
+        return 4
+    elif input_layer == "conv2d6":
+        return 6
+    elif input_layer == "conv2d8":
+        return 8

examples/aishell/s0/conf/deepspeech2.yaml

@@ -48,6 +48,9 @@ training:
   weight_decay: 1e-06
   global_grad_clip: 3.0
   log_interval: 100
+  checkpoint:
+    kbest_n: 50
+    latest_n: 5
   
 decoding:
   batch_size: 128

examples/aishell/s1/conf/chunk_conformer.yaml

@@ -93,6 +93,9 @@ training:
     warmup_steps: 25000
     lr_decay: 1.0
   log_interval: 100
+  checkpoint:
+    kbest_n: 50
+    latest_n: 5
 
 
 decoding:

examples/aishell/s1/conf/conformer.yaml

@@ -88,6 +88,9 @@ training:
     warmup_steps: 25000
     lr_decay: 1.0
   log_interval: 100
+  checkpoint:
+    kbest_n: 50
+    latest_n: 5
 
 
 decoding:

examples/aishell/s1/local/align.sh

@@ -0,0 +1,43 @@
+#! /usr/bin/env bash
+
+if [ $# != 2 ];then
+    echo "usage: ${0} config_path ckpt_path_prefix"
+    exit -1
+fi
+
+ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
+echo "using $ngpu gpus..."
+
+device=gpu
+if [ ngpu == 0 ];then
+    device=cpu
+fi
+config_path=$1
+ckpt_prefix=$2
+
+ckpt_name=$(basename ${ckpt_prefxi})
+
+mkdir -p exp
+
+
+
+batch_size=1
+output_dir=${ckpt_prefix}
+mkdir -p ${output_dir}
+
+# align dump in `result_file`
+# .tier, .TextGrid dump in `dir of result_file`
+python3 -u ${BIN_DIR}/alignment.py \
+--device ${device} \
+--nproc 1 \
+--config ${config_path} \
+--result_file ${output_dir}/${type}.align \
+--checkpoint_path ${ckpt_prefix} \
+--opts decoding.batch_size ${batch_size}
+
+if [ $? -ne 0 ]; then
+    echo "Failed in ctc alignment!"
+    exit 1
+fi
+
+exit 0

examples/aishell/s1/run.sh

@@ -30,10 +30,15 @@ fi
 
 if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
     # test ckpt avg_n
-    CUDA_VISIBLE_DEVICES=4 ./local/test.sh ${conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} || exit -1
+    CUDA_VISIBLE_DEVICES=0 ./local/test.sh ${conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} || exit -1
 fi
 
 if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then
+    # ctc alignment of test data
+    CUDA_VISIBLE_DEVICES=0 ./local/align.sh ${conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} || exit -1
+fi
+
+if [ ${stage} -le 5 ] && [ ${stop_stage} -ge 5 ]; then
     # export ckpt avg_n
     CUDA_VISIBLE_DEVICES= ./local/export.sh ${conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} exp/${ckpt}/checkpoints/${avg_ckpt}.jit
 fi

examples/librispeech/s0/conf/deepspeech2.yaml

@@ -48,6 +48,9 @@ training:
   weight_decay: 1e-06
   global_grad_clip: 5.0
   log_interval: 100
+  checkpoint:
+    kbest_n: 50
+    latest_n: 5
 
 decoding:
   batch_size: 128

examples/librispeech/s1/conf/chunk_confermer.yaml

@@ -93,6 +93,9 @@ training:
     warmup_steps: 25000
     lr_decay: 1.0
   log_interval: 100
+  checkpoint:
+    kbest_n: 50
+    latest_n: 5
 
 
 decoding:

examples/librispeech/s1/conf/chunk_transformer.yaml

@@ -86,6 +86,9 @@ training:
     warmup_steps: 25000
     lr_decay: 1.0
   log_interval: 100
+  checkpoint:
+    kbest_n: 50
+    latest_n: 5
 
 
 decoding:

examples/librispeech/s1/conf/conformer.yaml

@@ -89,6 +89,9 @@ training:
     warmup_steps: 25000
     lr_decay: 1.0
   log_interval: 100
+  checkpoint:
+    kbest_n: 50
+    latest_n: 5
 
 
 decoding:

examples/librispeech/s1/conf/transformer.yaml

@@ -84,6 +84,9 @@ training:
     warmup_steps: 25000
     lr_decay: 1.0
   log_interval: 100
+  checkpoint:
+    kbest_n: 50
+    latest_n: 5
 
 
 decoding:

examples/librispeech/s1/local/align.sh

@@ -0,0 +1,43 @@
+#! /usr/bin/env bash
+
+if [ $# != 2 ];then
+    echo "usage: ${0} config_path ckpt_path_prefix"
+    exit -1
+fi
+
+ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
+echo "using $ngpu gpus..."
+
+device=gpu
+if [ ngpu == 0 ];then
+    device=cpu
+fi
+config_path=$1
+ckpt_prefix=$2
+
+ckpt_name=$(basename ${ckpt_prefxi})
+
+mkdir -p exp
+
+
+
+batch_size=1
+output_dir=${ckpt_prefix}
+mkdir -p ${output_dir}
+
+# align dump in `result_file`
+# .tier, .TextGrid dump in `dir of result_file`
+python3 -u ${BIN_DIR}/alignment.py \
+--device ${device} \
+--nproc 1 \
+--config ${config_path} \
+--result_file ${output_dir}/${type}.align \
+--checkpoint_path ${ckpt_prefix} \
+--opts decoding.batch_size ${batch_size}
+
+if [ $? -ne 0 ]; then
+    echo "Failed in ctc alignment!"
+    exit 1
+fi
+
+exit 0

examples/librispeech/s1/run.sh

@@ -33,6 +33,11 @@ if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
 fi
 
 if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then
+    # ctc alignment of test data
+    CUDA_VISIBLE_DEVICES=0 ./local/align.sh ${conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} || exit -1
+fi
+
+if [ ${stage} -le 5 ] && [ ${stop_stage} -ge 5 ]; then
     # export ckpt avg_n
     CUDA_VISIBLE_DEVICES= ./local/export.sh ${conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} exp/${ckpt}/checkpoints/${avg_ckpt}.jit
 fi

examples/tiny/s0/conf/deepspeech2.yaml

@@ -43,12 +43,16 @@ model:
   share_rnn_weights: True 
 
 training:
-  n_epoch: 24
+  n_epoch: 10
   lr: 1e-5 
   lr_decay: 1.0 
   weight_decay: 1e-06
   global_grad_clip: 5.0
   log_interval: 1
+  checkpoint:
+    kbest_n: 3
+    latest_n: 2
+
 
 decoding:
   batch_size: 128

examples/tiny/s1/conf/chunk_confermer.yaml

@@ -91,6 +91,9 @@ training:
     warmup_steps: 25000
     lr_decay: 1.0
   log_interval: 1
+  checkpoint:
+    kbest_n: 10
+    latest_n: 1
 
 
 decoding:

examples/tiny/s1/conf/chunk_transformer.yaml

@@ -84,6 +84,9 @@ training:
     warmup_steps: 25000
     lr_decay: 1.0
   log_interval: 1
+  checkpoint:
+    kbest_n: 10
+    latest_n: 1
 
 
 decoding:

examples/tiny/s1/conf/conformer.yaml

@@ -87,6 +87,9 @@ training:
     warmup_steps: 25000
     lr_decay: 1.0
   log_interval: 1
+  checkpoint:
+    kbest_n: 10
+    latest_n: 1
 
 
 decoding:

examples/tiny/s1/conf/transformer.yaml

@@ -84,6 +84,9 @@ training:
     warmup_steps: 25000
     lr_decay: 1.0
   log_interval: 1
+  checkpoint:
+    kbest_n: 10
+    latest_n: 1
 
 
 decoding:

examples/tiny/s1/local/align.sh

@@ -0,0 +1,43 @@
+#! /usr/bin/env bash
+
+if [ $# != 2 ];then
+    echo "usage: ${0} config_path ckpt_path_prefix"
+    exit -1
+fi
+
+ngpu=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
+echo "using $ngpu gpus..."
+
+device=gpu
+if [ ngpu == 0 ];then
+    device=cpu
+fi
+config_path=$1
+ckpt_prefix=$2
+
+ckpt_name=$(basename ${ckpt_prefxi})
+
+mkdir -p exp
+
+
+
+batch_size=1
+output_dir=${ckpt_prefix}
+mkdir -p ${output_dir}
+
+# align dump in `result_file`
+# .tier, .TextGrid dump in `dir of result_file`
+python3 -u ${BIN_DIR}/alignment.py \
+--device ${device} \
+--nproc 1 \
+--config ${config_path} \
+--result_file ${output_dir}/${type}.align \
+--checkpoint_path ${ckpt_prefix} \
+--opts decoding.batch_size ${batch_size}
+
+if [ $? -ne 0 ]; then
+    echo "Failed in ctc alignment!"
+    exit 1
+fi
+
+exit 0

examples/tiny/s1/run.sh

@@ -34,6 +34,12 @@ if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
 fi
 
 if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then
+    # ctc alignment of test data
+    CUDA_VISIBLE_DEVICES=0 ./local/align.sh ${conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} || exit -1
+fi
+
+if [ ${stage} -le 5 ] && [ ${stop_stage} -ge 5 ]; then
     # export ckpt avg_n
-    CUDA_VISIBLE_DEVICES=  ./local/export.sh ${conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} exp/${ckpt}/checkpoints/${avg_ckpt}.jit
+    CUDA_VISIBLE_DEVICES= ./local/export.sh ${conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} exp/${ckpt}/checkpoints/${avg_ckpt}.jit
 fi
+

tools/Makefile

@@ -38,4 +38,4 @@ soxbindings.done:
 mfa.done:
 	test -d montreal-forced-aligner || wget https://github.com/MontrealCorpusTools/Montreal-Forced-Aligner/releases/download/v1.0.1/montreal-forced-aligner_linux.tar.gz
 	tar xvf montreal-forced-aligner_linux.tar.gz
-	touch mfa.done
+	touch mfa.done