support RNNLM shallow fusion for LSTM transducer

egs/librispeech/ASR/lstm_transducer_stateless2/decode.py

@@ -115,7 +115,8 @@ from beam_search import (
     greedy_search,
     greedy_search_batch,
     modified_beam_search,
-    modified_beam_search_ngram_rescoring,
+    modified_beam_search_rnnlm_shallow_fusion,
+    
 )
 from librispeech import LibriSpeech
 from train import add_model_arguments, get_params, get_transducer_model
@@ -128,6 +129,7 @@ from icefall.checkpoint import (
     load_checkpoint,
 )
 from icefall.lexicon import Lexicon
+from icefall.rnn_lm.model import RnnLmModel
 from icefall.utils import (
     AttributeDict,
     setup_logger,
@@ -216,7 +218,7 @@ def get_parser():
           - fast_beam_search_nbest
           - fast_beam_search_nbest_oracle
           - fast_beam_search_nbest_LG
-          - modified_beam_search_ngram_rescoring
+          - modified-beam-search_rnnlm_shallow_fusion # for rnn lm shallow fusion
         If you use fast_beam_search_nbest_LG, you have to specify
         `--lang-dir`, which should contain `LG.pt`.
         """,
@@ -307,21 +309,74 @@ def get_parser():
     )
 
     parser.add_argument(
-        "--tokens-ngram",
+        "--rnn-lm-scale",
-        type=int,
+        type=float,
-        default=3,
+        default=0.0,
-        help="""Token Ngram used for rescoring.
+        help="""Used only when --method is modified_beam_search3.
-            Used only when the decoding method is modified_beam_search_ngram_rescoring""",
+        It specifies the path to RNN LM exp dir.
+        """,
     )
 
     parser.add_argument(
-        "--backoff-id",
+        "--rnn-lm-exp-dir",
-        type=int,
+        type=str,
-        default=500,
+        default="rnn_lm/exp",
-        help="""ID of the backoff symbol.
+        help="""Used only when --method is rnn-lm.
-                Used only when the decoding method is modified_beam_search_ngram_rescoring""",
+        It specifies the path to RNN LM exp dir.
+        """,
     )
 
+    parser.add_argument(
+        "--rnn-lm-epoch",
+        type=int,
+        default=7,
+        help="""Used only when --method is rnn-lm.
+        It specifies the checkpoint to use.
+        """,
+    )
+
+    parser.add_argument(
+        "--rnn-lm-avg",
+        type=int,
+        default=2,
+        help="""Used only when --method is rnn-lm.
+        It specifies the number of checkpoints to average.
+        """,
+    )
+
+    parser.add_argument(
+        "--rnn-lm-embedding-dim",
+        type=int,
+        default=2048,
+        help="Embedding dim of the model",
+    )
+
+    parser.add_argument(
+        "--rnn-lm-hidden-dim",
+        type=int,
+        default=2048,
+        help="Hidden dim of the model",
+    )
+
+    parser.add_argument(
+        "--rnn-lm-num-layers",
+        type=int,
+        default=4,
+        help="Number of RNN layers the model",
+    )
+    parser.add_argument(
+        "--rnn-lm-tie-weights",
+        type=str2bool,
+        default=False,
+        help="""True to share the weights between the input embedding layer and the
+        last output linear layer
+        """,
+    )
+    parser.add_argument(
+        "--ilm-scale",
+        type=float,
+        default=-0.1
+    )
     add_model_arguments(parser)
 
     return parser
@@ -336,6 +391,8 @@ def decode_one_batch(
     decoding_graph: Optional[k2.Fsa] = None,
     ngram_lm: Optional[NgramLm] = None,
     ngram_lm_scale: float = 1.0,
+    rnnlm: Optional[RnnLmModel] = None,
+    rnnlm_scale: float = 1.0,
 ) -> Dict[str, List[List[str]]]:
     """Decode one batch and return the result in a dict. The dict has the
     following format:
@@ -469,14 +526,14 @@ def decode_one_batch(
         )
         for hyp in sp.decode(hyp_tokens):
             hyps.append(hyp.split())
-    elif params.decoding_method == "modified_beam_search_ngram_rescoring":
+    elif params.decoding_method == "modified_beam_search_sf_rnnlm":
-        hyp_tokens = modified_beam_search_ngram_rescoring(
+        hyp_tokens = modified_beam_search_sf_rnnlm_batched(
             model=model,
             encoder_out=encoder_out,
             encoder_out_lens=encoder_out_lens,
-            ngram_lm=ngram_lm,
+            sp=sp,
-            ngram_lm_scale=ngram_lm_scale,
+            rnnlm=rnnlm,
-            beam=params.beam_size,
+            rnnlm_scale=rnnlm_scale,
         )
         for hyp in sp.decode(hyp_tokens):
             hyps.append(hyp.split())
@@ -531,7 +588,9 @@ def decode_dataset(
     decoding_graph: Optional[k2.Fsa] = None,
     ngram_lm: Optional[NgramLm] = None,
     ngram_lm_scale: float = 1.0,
-) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    rnnlm: Optional[NgramLm] = None,
+    rnnlm_scale: float = 1.0,
+) -> Dict[str, List[Tuple[List[str], List[str]]]]:
     """Decode dataset.
 
     Args:
@@ -572,6 +631,9 @@ def decode_dataset(
     for batch_idx, batch in enumerate(dl):
         texts = batch["supervisions"]["text"]
         cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+        total_duration = sum([cut.duration for cut in batch["supervisions"]["cut"]])
+        
+        logging.info(f"Decoding {batch_idx}-th batch, batch size is {len(cut_ids)}, total duration is {total_duration}")
 
         hyps_dict = decode_one_batch(
             params=params,
@@ -582,6 +644,8 @@ def decode_dataset(
             batch=batch,
             ngram_lm=ngram_lm,
             ngram_lm_scale=ngram_lm_scale,
+            rnnlm=rnnlm,
+            rnnlm_scale=rnnlm_scale,
         )
 
         for name, hyps in hyps_dict.items():
@@ -607,7 +671,7 @@ def decode_dataset(
 def save_results(
     params: AttributeDict,
     test_set_name: str,
-    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+    results_dict: Dict[str, List[Tuple[List[int], List[int]]]],
 ):
     test_set_wers = dict()
     for key, results in results_dict.items():
@@ -667,7 +731,7 @@ def main():
         "fast_beam_search_nbest_LG",
         "fast_beam_search_nbest_oracle",
         "modified_beam_search",
-        "modified_beam_search_ngram_rescoring",
+        "modified_beam_search_sf_rnnlm",
     )
     params.res_dir = params.exp_dir / params.decoding_method
 
@@ -692,7 +756,12 @@ def main():
     else:
         params.suffix += f"-context-{params.context_size}"
         params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
-    params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}"
+    
+    if "rnnlm" in params.decoding_method:
+        params.suffix += f"-rnnlm-lm-scale-{params.rnn_lm_scale}"
+    
+    if "ILME" in params.decoding_method:
+        params.suffix += f"-ILME-scale={params.ilm_scale}"
 
     if params.use_averaged_model:
         params.suffix += "-use-averaged-model"
@@ -806,14 +875,28 @@ def main():
     model.to(device)
     model.eval()
 
-    lm_filename = f"{params.tokens_ngram}gram.fst.txt"
+    # only load rnnlm if used
-    logging.info(f"lm filename: {lm_filename}")
+    if "rnnlm" in params.decoding_method:
-    ngram_lm = NgramLm(
+        rnn_lm_scale = params.rnn_lm_scale
-        str(params.lang_dir / lm_filename),
+
-        backoff_id=params.backoff_id,
+        rnn_lm_model = RnnLmModel(
-        is_binary=False,
+            vocab_size=params.vocab_size,
-    )
+            embedding_dim=params.rnn_lm_embedding_dim,
-    logging.info(f"num states: {ngram_lm.lm.num_states}")
+            hidden_dim=params.rnn_lm_hidden_dim,
+            num_layers=params.rnn_lm_num_layers,
+            tie_weights=params.rnn_lm_tie_weights,
+        )
+        assert params.rnn_lm_avg == 1
+
+        load_checkpoint(
+            f"{params.rnn_lm_exp_dir}/epoch-{params.rnn_lm_epoch}.pt",
+            rnn_lm_model,
+        )
+        rnn_lm_model.to(device)
+        rnn_lm_model.eval()
+    
+    else:
+        rnn_lm_model = None
 
     if "fast_beam_search" in params.decoding_method:
         if params.decoding_method == "fast_beam_search_nbest_LG":
@@ -861,6 +944,8 @@ def main():
             decoding_graph=decoding_graph,
             ngram_lm=ngram_lm,
             ngram_lm_scale=params.ngram_lm_scale,
+            rnnlm=rnn_lm_model,
+            rnnlm_scale=rnn_lm_scale,
         )
 
         save_results(

egs/librispeech/ASR/pruned_transducer_stateless2/beam_search.py

@@ -16,7 +16,7 @@
 
 import warnings
 from dataclasses import dataclass
-from typing import Dict, List, Optional, Union
+from typing import Dict, List, Optional
 
 import k2
 import sentencepiece as spm
@@ -25,13 +25,8 @@ from model import Transducer
 
 from icefall import NgramLm, NgramLmStateCost
 from icefall.decode import Nbest, one_best_decoding
-from icefall.utils import (
+from icefall.rnn_lm.model import RnnLmModel
-    DecodingResults,
+from icefall.utils import add_eos, add_sos, get_texts
-    add_eos,
-    add_sos,
-    get_texts,
-    get_texts_with_timestamp,
-)
 
 
 def fast_beam_search_one_best(
@@ -43,8 +38,7 @@ def fast_beam_search_one_best(
     max_states: int,
     max_contexts: int,
     temperature: float = 1.0,
-    return_timestamps: bool = False,
+) -> List[List[int]]:
-) -> Union[List[List[int]], DecodingResults]:
     """It limits the maximum number of symbols per frame to 1.
 
     A lattice is first obtained using fast beam search, and then
@@ -68,12 +62,8 @@ def fast_beam_search_one_best(
         Max contexts pre stream per frame.
       temperature:
         Softmax temperature.
-      return_timestamps:
-        Whether to return timestamps.
     Returns:
-      If return_timestamps is False, return the decoded result.
+      Return the decoded result.
-      Else, return a DecodingResults object containing
-      decoded result and corresponding timestamps.
     """
     lattice = fast_beam_search(
         model=model,
@@ -87,11 +77,8 @@ def fast_beam_search_one_best(
     )
 
     best_path = one_best_decoding(lattice)
-
+    hyps = get_texts(best_path)
-    if not return_timestamps:
+    return hyps
-        return get_texts(best_path)
-    else:
-        return get_texts_with_timestamp(best_path)
 
 
 def fast_beam_search_nbest_LG(
@@ -106,8 +93,7 @@ def fast_beam_search_nbest_LG(
     nbest_scale: float = 0.5,
     use_double_scores: bool = True,
     temperature: float = 1.0,
-    return_timestamps: bool = False,
+) -> List[List[int]]:
-) -> Union[List[List[int]], DecodingResults]:
     """It limits the maximum number of symbols per frame to 1.
 
     The process to get the results is:
@@ -144,12 +130,8 @@ def fast_beam_search_nbest_LG(
         single precision.
       temperature:
         Softmax temperature.
-      return_timestamps:
-        Whether to return timestamps.
     Returns:
-      If return_timestamps is False, return the decoded result.
+      Return the decoded result.
-      Else, return a DecodingResults object containing
-      decoded result and corresponding timestamps.
     """
     lattice = fast_beam_search(
         model=model,
@@ -214,10 +196,9 @@ def fast_beam_search_nbest_LG(
     best_hyp_indexes = ragged_tot_scores.argmax()
     best_path = k2.index_fsa(nbest.fsa, best_hyp_indexes)
 
-    if not return_timestamps:
+    hyps = get_texts(best_path)
-        return get_texts(best_path)
+
-    else:
+    return hyps
-        return get_texts_with_timestamp(best_path)
 
 
 def fast_beam_search_nbest(
@@ -232,8 +213,7 @@ def fast_beam_search_nbest(
     nbest_scale: float = 0.5,
     use_double_scores: bool = True,
     temperature: float = 1.0,
-    return_timestamps: bool = False,
+) -> List[List[int]]:
-) -> Union[List[List[int]], DecodingResults]:
     """It limits the maximum number of symbols per frame to 1.
 
     The process to get the results is:
@@ -270,12 +250,8 @@ def fast_beam_search_nbest(
         single precision.
       temperature:
         Softmax temperature.
-      return_timestamps:
-        Whether to return timestamps.
     Returns:
-      If return_timestamps is False, return the decoded result.
+      Return the decoded result.
-      Else, return a DecodingResults object containing
-      decoded result and corresponding timestamps.
     """
     lattice = fast_beam_search(
         model=model,
@@ -304,10 +280,9 @@ def fast_beam_search_nbest(
 
     best_path = k2.index_fsa(nbest.fsa, max_indexes)
 
-    if not return_timestamps:
+    hyps = get_texts(best_path)
-        return get_texts(best_path)
+
-    else:
+    return hyps
-        return get_texts_with_timestamp(best_path)
 
 
 def fast_beam_search_nbest_oracle(
@@ -323,8 +298,7 @@ def fast_beam_search_nbest_oracle(
     use_double_scores: bool = True,
     nbest_scale: float = 0.5,
     temperature: float = 1.0,
-    return_timestamps: bool = False,
+) -> List[List[int]]:
-) -> Union[List[List[int]], DecodingResults]:
     """It limits the maximum number of symbols per frame to 1.
 
     A lattice is first obtained using fast beam search, and then
@@ -365,12 +339,8 @@ def fast_beam_search_nbest_oracle(
         yields more unique paths.
       temperature:
         Softmax temperature.
-      return_timestamps:
-        Whether to return timestamps.
     Returns:
-      If return_timestamps is False, return the decoded result.
+      Return the decoded result.
-      Else, return a DecodingResults object containing
-      decoded result and corresponding timestamps.
     """
     lattice = fast_beam_search(
         model=model,
@@ -409,10 +379,8 @@ def fast_beam_search_nbest_oracle(
 
     best_path = k2.index_fsa(nbest.fsa, max_indexes)
 
-    if not return_timestamps:
+    hyps = get_texts(best_path)
-        return get_texts(best_path)
+    return hyps
-    else:
-        return get_texts_with_timestamp(best_path)
 
 
 def fast_beam_search(
@@ -502,11 +470,8 @@ def fast_beam_search(
 
 
 def greedy_search(
-    model: Transducer,
+    model: Transducer, encoder_out: torch.Tensor, max_sym_per_frame: int
-    encoder_out: torch.Tensor,
+) -> List[int]:
-    max_sym_per_frame: int,
-    return_timestamps: bool = False,
-) -> Union[List[int], DecodingResults]:
     """Greedy search for a single utterance.
     Args:
       model:
@@ -516,12 +481,8 @@ def greedy_search(
       max_sym_per_frame:
         Maximum number of symbols per frame. If it is set to 0, the WER
         would be 100%.
-      return_timestamps:
-        Whether to return timestamps.
     Returns:
-      If return_timestamps is False, return the decoded result.
+      Return the decoded result.
-      Else, return a DecodingResults object containing
-      decoded result and corresponding timestamps.
     """
     assert encoder_out.ndim == 3
 
@@ -547,10 +508,6 @@ def greedy_search(
     t = 0
     hyp = [blank_id] * context_size
 
-    # timestamp[i] is the frame index after subsampling
-    # on which hyp[i] is decoded
-    timestamp = []
-
     # Maximum symbols per utterance.
     max_sym_per_utt = 1000
 
@@ -577,7 +534,6 @@ def greedy_search(
         y = logits.argmax().item()
         if y not in (blank_id, unk_id):
             hyp.append(y)
-            timestamp.append(t)
             decoder_input = torch.tensor(
                 [hyp[-context_size:]], device=device
             ).reshape(1, context_size)
@@ -592,21 +548,14 @@ def greedy_search(
             t += 1
     hyp = hyp[context_size:]  # remove blanks
 
-    if not return_timestamps:
+    return hyp
-        return hyp
-    else:
-        return DecodingResults(
-            tokens=[hyp],
-            timestamps=[timestamp],
-        )
 
 
 def greedy_search_batch(
     model: Transducer,
     encoder_out: torch.Tensor,
     encoder_out_lens: torch.Tensor,
-    return_timestamps: bool = False,
+) -> List[List[int]]:
-) -> Union[List[List[int]], DecodingResults]:
     """Greedy search in batch mode. It hardcodes --max-sym-per-frame=1.
     Args:
       model:
@@ -616,12 +565,9 @@ def greedy_search_batch(
       encoder_out_lens:
         A 1-D tensor of shape (N,), containing number of valid frames in
         encoder_out before padding.
-      return_timestamps:
-        Whether to return timestamps.
     Returns:
-      If return_timestamps is False, return the decoded result.
+      Return a list-of-list of token IDs containing the decoded results.
-      Else, return a DecodingResults object containing
+      len(ans) equals to encoder_out.size(0).
-      decoded result and corresponding timestamps.
     """
     assert encoder_out.ndim == 3
     assert encoder_out.size(0) >= 1, encoder_out.size(0)
@@ -646,10 +592,6 @@ def greedy_search_batch(
 
     hyps = [[blank_id] * context_size for _ in range(N)]
 
-    # timestamp[n][i] is the frame index after subsampling
-    # on which hyp[n][i] is decoded
-    timestamps = [[] for _ in range(N)]
-
     decoder_input = torch.tensor(
         hyps,
         device=device,
@@ -663,7 +605,7 @@ def greedy_search_batch(
     encoder_out = model.joiner.encoder_proj(packed_encoder_out.data)
 
     offset = 0
-    for (t, batch_size) in enumerate(batch_size_list):
+    for batch_size in batch_size_list:
         start = offset
         end = offset + batch_size
         current_encoder_out = encoder_out.data[start:end]
@@ -685,7 +627,6 @@ def greedy_search_batch(
         for i, v in enumerate(y):
             if v not in (blank_id, unk_id):
                 hyps[i].append(v)
-                timestamps[i].append(t)
                 emitted = True
         if emitted:
             # update decoder output
@@ -700,19 +641,11 @@ def greedy_search_batch(
 
     sorted_ans = [h[context_size:] for h in hyps]
     ans = []
-    ans_timestamps = []
     unsorted_indices = packed_encoder_out.unsorted_indices.tolist()
     for i in range(N):
         ans.append(sorted_ans[unsorted_indices[i]])
-        ans_timestamps.append(timestamps[unsorted_indices[i]])
 
-    if not return_timestamps:
+    return ans
-        return ans
-    else:
-        return DecodingResults(
-            tokens=ans,
-            timestamps=ans_timestamps,
-        )
 
 
 @dataclass
@@ -725,11 +658,9 @@ class Hypothesis:
     # It contains only one entry.
     log_prob: torch.Tensor
 
-    # timestamp[i] is the frame index after subsampling
-    # on which ys[i] is decoded
-    timestamp: List[int]
-
     state_cost: Optional[NgramLmStateCost] = None
+    state: Optional = None
+    lm_score: Optional=None
 
     @property
     def key(self) -> str:
@@ -878,8 +809,7 @@ def modified_beam_search(
     encoder_out_lens: torch.Tensor,
     beam: int = 4,
     temperature: float = 1.0,
-    return_timestamps: bool = False,
+) -> List[List[int]]:
-) -> Union[List[List[int]], DecodingResults]:
     """Beam search in batch mode with --max-sym-per-frame=1 being hardcoded.
 
     Args:
@@ -894,12 +824,9 @@ def modified_beam_search(
         Number of active paths during the beam search.
       temperature:
         Softmax temperature.
-      return_timestamps:
-        Whether to return timestamps.
     Returns:
-      If return_timestamps is False, return the decoded result.
+      Return a list-of-list of token IDs. ans[i] is the decoding results
-      Else, return a DecodingResults object containing
+      for the i-th utterance.
-      decoded result and corresponding timestamps.
     """
     assert encoder_out.ndim == 3, encoder_out.shape
     assert encoder_out.size(0) >= 1, encoder_out.size(0)
@@ -927,7 +854,6 @@ def modified_beam_search(
             Hypothesis(
                 ys=[blank_id] * context_size,
                 log_prob=torch.zeros(1, dtype=torch.float32, device=device),
-                timestamp=[],
             )
         )
 
@@ -935,7 +861,7 @@ def modified_beam_search(
 
     offset = 0
     finalized_B = []
-    for (t, batch_size) in enumerate(batch_size_list):
+    for batch_size in batch_size_list:
         start = offset
         end = offset + batch_size
         current_encoder_out = encoder_out.data[start:end]
@@ -1013,44 +939,30 @@ def modified_beam_search(
 
                 new_ys = hyp.ys[:]
                 new_token = topk_token_indexes[k]
-                new_timestamp = hyp.timestamp[:]
                 if new_token not in (blank_id, unk_id):
                     new_ys.append(new_token)
-                    new_timestamp.append(t)
 
                 new_log_prob = topk_log_probs[k]
-                new_hyp = Hypothesis(
+                new_hyp = Hypothesis(ys=new_ys, log_prob=new_log_prob)
-                    ys=new_ys, log_prob=new_log_prob, timestamp=new_timestamp
-                )
                 B[i].add(new_hyp)
 
     B = B + finalized_B
     best_hyps = [b.get_most_probable(length_norm=True) for b in B]
 
     sorted_ans = [h.ys[context_size:] for h in best_hyps]
-    sorted_timestamps = [h.timestamp for h in best_hyps]
     ans = []
-    ans_timestamps = []
     unsorted_indices = packed_encoder_out.unsorted_indices.tolist()
     for i in range(N):
         ans.append(sorted_ans[unsorted_indices[i]])
-        ans_timestamps.append(sorted_timestamps[unsorted_indices[i]])
 
-    if not return_timestamps:
+    return ans
-        return ans
-    else:
-        return DecodingResults(
-            tokens=ans,
-            timestamps=ans_timestamps,
-        )
 
 
 def _deprecated_modified_beam_search(
     model: Transducer,
     encoder_out: torch.Tensor,
     beam: int = 4,
-    return_timestamps: bool = False,
+) -> List[int]:
-) -> Union[List[int], DecodingResults]:
     """It limits the maximum number of symbols per frame to 1.
 
     It decodes only one utterance at a time. We keep it only for reference.
@@ -1065,13 +977,8 @@ def _deprecated_modified_beam_search(
         A tensor of shape (N, T, C) from the encoder. Support only N==1 for now.
       beam:
         Beam size.
-      return_timestamps:
-        Whether to return timestamps.
-
     Returns:
-      If return_timestamps is False, return the decoded result.
+      Return the decoded result.
-      Else, return a DecodingResults object containing
-      decoded result and corresponding timestamps.
     """
 
     assert encoder_out.ndim == 3
@@ -1091,7 +998,6 @@ def _deprecated_modified_beam_search(
         Hypothesis(
             ys=[blank_id] * context_size,
             log_prob=torch.zeros(1, dtype=torch.float32, device=device),
-            timestamp=[],
         )
     )
     encoder_out = model.joiner.encoder_proj(encoder_out)
@@ -1150,24 +1056,17 @@ def _deprecated_modified_beam_search(
         for i in range(len(topk_hyp_indexes)):
             hyp = A[topk_hyp_indexes[i]]
             new_ys = hyp.ys[:]
-            new_timestamp = hyp.timestamp[:]
             new_token = topk_token_indexes[i]
             if new_token not in (blank_id, unk_id):
                 new_ys.append(new_token)
-                new_timestamp.append(t)
             new_log_prob = topk_log_probs[i]
-            new_hyp = Hypothesis(
+            new_hyp = Hypothesis(ys=new_ys, log_prob=new_log_prob)
-                ys=new_ys, log_prob=new_log_prob, timestamp=new_timestamp
-            )
             B.add(new_hyp)
 
     best_hyp = B.get_most_probable(length_norm=True)
     ys = best_hyp.ys[context_size:]  # [context_size:] to remove blanks
 
-    if not return_timestamps:
+    return ys
-        return ys
-    else:
-        return DecodingResults(tokens=[ys], timestamps=[best_hyp.timestamp])
 
 
 def beam_search(
@@ -1175,8 +1074,7 @@ def beam_search(
     encoder_out: torch.Tensor,
     beam: int = 4,
     temperature: float = 1.0,
-    return_timestamps: bool = False,
+) -> List[int]:
-) -> Union[List[int], DecodingResults]:
     """
     It implements Algorithm 1 in https://arxiv.org/pdf/1211.3711.pdf
 
@@ -1191,13 +1089,8 @@ def beam_search(
         Beam size.
       temperature:
         Softmax temperature.
-      return_timestamps:
-        Whether to return timestamps.
-
     Returns:
-      If return_timestamps is False, return the decoded result.
+      Return the decoded result.
-      Else, return a DecodingResults object containing
-      decoded result and corresponding timestamps.
     """
     assert encoder_out.ndim == 3
 
@@ -1224,7 +1117,7 @@ def beam_search(
     t = 0
 
     B = HypothesisList()
-    B.add(Hypothesis(ys=[blank_id] * context_size, log_prob=0.0, timestamp=[]))
+    B.add(Hypothesis(ys=[blank_id] * context_size, log_prob=0.0))
 
     max_sym_per_utt = 20000
 
@@ -1285,13 +1178,7 @@ def beam_search(
             new_y_star_log_prob = y_star.log_prob + skip_log_prob
 
             # ys[:] returns a copy of ys
-            B.add(
+            B.add(Hypothesis(ys=y_star.ys[:], log_prob=new_y_star_log_prob))
-                Hypothesis(
-                    ys=y_star.ys[:],
-                    log_prob=new_y_star_log_prob,
-                    timestamp=y_star.timestamp[:],
-                )
-            )
 
             # Second, process other non-blank labels
             values, indices = log_prob.topk(beam + 1)
@@ -1300,14 +1187,7 @@ def beam_search(
                     continue
                 new_ys = y_star.ys + [i]
                 new_log_prob = y_star.log_prob + v
-                new_timestamp = y_star.timestamp + [t]
+                A.add(Hypothesis(ys=new_ys, log_prob=new_log_prob))
-                A.add(
-                    Hypothesis(
-                        ys=new_ys,
-                        log_prob=new_log_prob,
-                        timestamp=new_timestamp,
-                    )
-                )
 
             # Check whether B contains more than "beam" elements more probable
             # than the most probable in A
@@ -1323,11 +1203,7 @@ def beam_search(
 
     best_hyp = B.get_most_probable(length_norm=True)
     ys = best_hyp.ys[context_size:]  # [context_size:] to remove blanks
-
+    return ys
-    if not return_timestamps:
-        return ys
-    else:
-        return DecodingResults(tokens=[ys], timestamps=[best_hyp.timestamp])
 
 
 def fast_beam_search_with_nbest_rescoring(
@@ -1347,8 +1223,7 @@ def fast_beam_search_with_nbest_rescoring(
     use_double_scores: bool = True,
     nbest_scale: float = 0.5,
     temperature: float = 1.0,
-    return_timestamps: bool = False,
+) -> Dict[str, List[List[int]]]:
-) -> Dict[str, Union[List[List[int]], DecodingResults]]:
     """It limits the maximum number of symbols per frame to 1.
     A lattice is first obtained using fast beam search, num_path are selected
     and rescored using a given language model. The shortest path within the
@@ -1390,13 +1265,10 @@ def fast_beam_search_with_nbest_rescoring(
         yields more unique paths.
       temperature:
         Softmax temperature.
-      return_timestamps:
-        Whether to return timestamps.
     Returns:
       Return the decoded result in a dict, where the key has the form
-      'ngram_lm_scale_xx' and the value is the decoded results
+      'ngram_lm_scale_xx' and the value is the decoded results. `xx` is the
-      optionally with timestamps. `xx` is the ngram LM scale value
+      ngram LM scale value used during decoding, i.e., 0.1.
-      used during decoding, i.e., 0.1.
     """
     lattice = fast_beam_search(
         model=model,
@@ -1474,18 +1346,16 @@ def fast_beam_search_with_nbest_rescoring(
         log_semiring=False,
     )
 
-    ans: Dict[str, Union[List[List[int]], DecodingResults]] = {}
+    ans: Dict[str, List[List[int]]] = {}
     for s in ngram_lm_scale_list:
         key = f"ngram_lm_scale_{s}"
         tot_scores = am_scores.values + s * ngram_lm_scores
         ragged_tot_scores = k2.RaggedTensor(nbest.shape, tot_scores)
         max_indexes = ragged_tot_scores.argmax()
         best_path = k2.index_fsa(nbest.fsa, max_indexes)
+        hyps = get_texts(best_path)
 
-        if not return_timestamps:
+        ans[key] = hyps
-            ans[key] = get_texts(best_path)
-        else:
-            ans[key] = get_texts_with_timestamp(best_path)
 
     return ans
 
@@ -1509,8 +1379,7 @@ def fast_beam_search_with_nbest_rnn_rescoring(
     use_double_scores: bool = True,
     nbest_scale: float = 0.5,
     temperature: float = 1.0,
-    return_timestamps: bool = False,
+) -> Dict[str, List[List[int]]]:
-) -> Dict[str, Union[List[List[int]], DecodingResults]]:
     """It limits the maximum number of symbols per frame to 1.
     A lattice is first obtained using fast beam search, num_path are selected
     and rescored using a given language model and a rnn-lm.
@@ -1556,13 +1425,10 @@ def fast_beam_search_with_nbest_rnn_rescoring(
         yields more unique paths.
       temperature:
         Softmax temperature.
-      return_timestamps:
-        Whether to return timestamps.
     Returns:
       Return the decoded result in a dict, where the key has the form
-      'ngram_lm_scale_xx' and the value is the decoded results
+      'ngram_lm_scale_xx' and the value is the decoded results. `xx` is the
-      optionally with timestamps. `xx` is the ngram LM scale value
+      ngram LM scale value used during decoding, i.e., 0.1.
-      used during decoding, i.e., 0.1.
     """
     lattice = fast_beam_search(
         model=model,
@@ -1674,44 +1540,150 @@ def fast_beam_search_with_nbest_rnn_rescoring(
             ragged_tot_scores = k2.RaggedTensor(nbest.shape, tot_scores)
             max_indexes = ragged_tot_scores.argmax()
             best_path = k2.index_fsa(nbest.fsa, max_indexes)
+            hyps = get_texts(best_path)
 
-            if not return_timestamps:
+            ans[key] = hyps
-                ans[key] = get_texts(best_path)
-            else:
-                ans[key] = get_texts_with_timestamp(best_path)
 
     return ans
 
+def modified_beam_search_sf_rnnlm(
+    model: Transducer,
+    encoder_out: torch.Tensor,
+    sp,
+    rnnlm: RnnLmModel,
+    rnnlm_scale: float,
+    beam: int = 4,
+):
+    encoder_out = model.joiner.encoder_proj(encoder_out)
+    lm_scale = rnnlm_scale
 
-def modified_beam_search_ngram_rescoring(
+    assert rnnlm is not None
+    assert encoder_out.ndim == 2, encoder_out.shape
+    rnnlm.clean_cache()
+    blank_id = model.decoder.blank_id
+    sos_id = sp.piece_to_id("<sos/eos>")
+    eos_id = sp.piece_to_id("<sos/eos>")
+    unk_id = getattr(model, "unk_id", blank_id)
+    context_size = model.decoder.context_size
+    device = next(model.parameters()).device
+
+    B = HypothesisList()
+    B.add(
+        Hypothesis(
+            ys=[blank_id] * context_size,
+            log_prob=torch.zeros(1, dtype=torch.float32, device=device),
+        )
+    )
+    
+    T = encoder_out.shape[0]
+    for t in range(T):
+        current_encoder_out = encoder_out[t : t + 1]
+        A = list(B)
+        B = HypothesisList()
+
+        ys_log_probs = torch.cat(
+            [hyp.log_prob.reshape(1, 1) for hyp in A]
+        )  # (num_hyps, 1)
+
+        decoder_input = torch.tensor(
+            [hyp.ys[-context_size:] for hyp in A],
+            device=device,
+            dtype=torch.int64,
+        )  # (num_hyps, context_size)
+        decoder_out = model.decoder(decoder_input, need_pad=False).squeeze(1)
+        decoder_out = model.joiner.decoder_proj(decoder_out)
+
+        # decoder_out is of shape (num_hyps, joiner_dim)
+        current_encoder_out = current_encoder_out.repeat(len(A), 1)
+        # current_encoder_out is of shape (num_hyps, encoder_out_dim)
+        logits = model.joiner(
+            current_encoder_out,
+            decoder_out,
+            project_input=False,
+        )  # (num_hyps,  vocab_size)
+        log_probs = logits.log_softmax(dim=-1)  # (num_hyps, vocab_size)
+        log_probs.add_(ys_log_probs)
+
+        vocab_size = log_probs.size(-1)
+        log_probs = log_probs.reshape(-1)
+        topk_log_probs, topk_indexes = log_probs.topk(
+            beam
+        )  # get topk tokens and scores
+
+        with warnings.catch_warnings():
+            warnings.simplefilter("ignore")
+            topk_hyp_indexes = (topk_indexes // vocab_size).tolist()
+            topk_token_indexes = (topk_indexes % vocab_size).tolist()
+
+            for k in range(len(topk_hyp_indexes)):
+                hyp_idx = topk_hyp_indexes[k]
+                hyp = A[hyp_idx]  # get hyp
+                new_ys = hyp.ys[:]
+                state = "ys=" + "+".join(list(map(str, new_ys)))
+                tokens = k2.RaggedTensor([new_ys[context_size:]])
+
+                lm_score = rnnlm.predict(
+                    tokens, state, sos_id, eos_id, blank_id
+                )  # get rnnlm score
+
+                hyp_log_prob = topk_log_probs[k]  # get score of current hyp
+                new_token = topk_token_indexes[k]  # get token
+                if new_token not in (blank_id, unk_id):
+                    new_ys.append(new_token)
+                    # state_cost = hyp.state_cost.forward_one_step(new_token)
+                    hyp_log_prob += (
+                        lm_score[new_token] * lm_scale
+                    )  # add the lm score
+                else:
+                    new_ys = new_ys
+                new_log_prob = hyp_log_prob
+
+                new_hyp = Hypothesis(
+                    ys=new_ys,
+                    log_prob=new_log_prob,
+                )
+                B.add(new_hyp)
+
+    best_hyp = B.get_most_probable(length_norm=True)
+    return best_hyp.ys[context_size:]
+  
+def modified_beam_search_rnnlm_shallow_fusion(
     model: Transducer,
     encoder_out: torch.Tensor,
     encoder_out_lens: torch.Tensor,
-    ngram_lm: NgramLm,
+    sp: spm.SentencePieceProcessor,
-    ngram_lm_scale: float,
+    rnnlm: RnnLmModel,
+    rnnlm_scale: float,
     beam: int = 4,
-    temperature: float = 1.0,
 ) -> List[List[int]]:
-    """Beam search in batch mode with --max-sym-per-frame=1 being hardcoded.
+    """Modified_beam_search + RNNLM shallow fusion
 
     Args:
-      model:
+        model (Transducer):
-        The transducer model.
+            The transducer model
-      encoder_out:
+        encoder_out (torch.Tensor): 
-        Output from the encoder. Its shape is (N, T, C).
+            Encoder output in (N,T,C)
-      encoder_out_lens:
+        encoder_out_lens (torch.Tensor): 
-        A 1-D tensor of shape (N,), containing number of valid frames in
+            A 1-D tensor of shape (N,), containing the number of 
-        encoder_out before padding.
+            valid frames in encoder_out before padding.
-      beam:
+        sp: 
-        Number of active paths during the beam search.
+            Sentence piece generator.
-      temperature:
+        rnnlm (RnnLmModel): 
-        Softmax temperature.
+            RNNLM
+        rnnlm_scale (float): 
+            scale of RNNLM in shallow fusion
+        beam (int, optional): 
+            Beam size. Defaults to 4.
+
     Returns:
       Return a list-of-list of token IDs. ans[i] is the decoding results
       for the i-th utterance.
     """
     assert encoder_out.ndim == 3, encoder_out.shape
     assert encoder_out.size(0) >= 1, encoder_out.size(0)
+    assert rnnlm is not None
+    lm_scale = rnnlm_scale
+    vocab_size = rnnlm.vocab_size
     
     packed_encoder_out = torch.nn.utils.rnn.pack_padded_sequence(
         input=encoder_out,
@@ -1721,26 +1693,33 @@ def modified_beam_search_ngram_rescoring(
     )
 
     blank_id = model.decoder.blank_id
+    sos_id = sp.piece_to_id("<sos/eos>")
+    eos_id = sp.piece_to_id("<sos/eos>")
     unk_id = getattr(model, "unk_id", blank_id)
     context_size = model.decoder.context_size
     device = next(model.parameters()).device
-    lm_scale = ngram_lm_scale
     
     batch_size_list = packed_encoder_out.batch_sizes.tolist()
     N = encoder_out.size(0) 
     assert torch.all(encoder_out_lens > 0), encoder_out_lens
     assert N == batch_size_list[0], (N, batch_size_list)
 
+    # get initial lm score and lm state by scoring the "sos" token
+    sos_token = torch.tensor([[sos_id]]).to(torch.int64).to(device)
+    init_score, init_states = rnnlm.score_token(sos_token)
+    
     B = [HypothesisList() for _ in range(N)]
     for i in range(N):
         B[i].add(
             Hypothesis(
                 ys=[blank_id] * context_size,
                 log_prob=torch.zeros(1, dtype=torch.float32, device=device),
-                state_cost=NgramLmStateCost(ngram_lm),
+                state=init_states,
+                lm_score=init_score.reshape(-1)
             )
         )
 
+    rnnlm.clean_cache()
     encoder_out = model.joiner.encoder_proj(packed_encoder_out.data)
     
     offset = 0
@@ -1748,7 +1727,7 @@ def modified_beam_search_ngram_rescoring(
     for batch_size in batch_size_list:
         start = offset
         end = offset + batch_size
-        current_encoder_out = encoder_out.data[start:end]
+        current_encoder_out = encoder_out.data[start:end] # get batch
         current_encoder_out = current_encoder_out.unsqueeze(1).unsqueeze(1)
         # current_encoder_out's shape is (batch_size, 1, 1, encoder_out_dim)
         offset = end
@@ -1762,12 +1741,8 @@ def modified_beam_search_ngram_rescoring(
         B = [HypothesisList() for _ in range(batch_size)]
         
         ys_log_probs = torch.cat(
-            [
+            [hyp.log_prob.reshape(1, 1) for hyps in A for hyp in hyps]
-                hyp.log_prob.reshape(1, 1) + hyp.state_cost.lm_score * lm_scale
+        )
-                for hyps in A
-                for hyp in hyps
-            ]
-        )  # (num_hyps, 1)
         
         decoder_input = torch.tensor(
             [hyp.ys[-context_size:] for hyps in A for hyp in hyps],
@@ -1777,10 +1752,7 @@ def modified_beam_search_ngram_rescoring(
         
         decoder_out = model.decoder(decoder_input, need_pad=False).unsqueeze(1)
         decoder_out = model.joiner.decoder_proj(decoder_out)
-        # decoder_out is of shape (num_hyps, 1, 1, joiner_dim)
         
-        # Note: For torch 1.7.1 and below, it requires a torch.int64 tensor
-        # as index, so we use `to(torch.int64)` below.
         current_encoder_out = torch.index_select(
             current_encoder_out,
             dim=0,
@@ -1795,12 +1767,14 @@ def modified_beam_search_ngram_rescoring(
         
         logits = logits.squeeze(1).squeeze(1)  # (num_hyps, vocab_size)
 
-        log_probs = (logits / temperature).log_softmax(
+        log_probs = logits.log_softmax(
             dim=-1
         )  # (num_hyps, vocab_size)
 
         log_probs.add_(ys_log_probs)
+        
         vocab_size = log_probs.size(-1)
+
         log_probs = log_probs.reshape(-1)
         
         row_splits = hyps_shape.row_splits(1) * vocab_size
@@ -1811,6 +1785,38 @@ def modified_beam_search_ngram_rescoring(
             shape=log_probs_shape, value=log_probs
         )
         
+
+        # for all hyps with a non-blank new token, score it
+        token_list = []
+        hs = []
+        cs = []
+        for i in range(batch_size):
+            topk_log_probs, topk_indexes = ragged_log_probs[i].topk(beam)
+
+            with warnings.catch_warnings():
+                warnings.simplefilter("ignore")
+                topk_hyp_indexes = (topk_indexes // vocab_size).tolist()
+                topk_token_indexes = (topk_indexes % vocab_size).tolist()
+            for k in range(len(topk_hyp_indexes)):
+                hyp_idx = topk_hyp_indexes[k]
+                hyp = A[i][hyp_idx]
+                
+                new_token = topk_token_indexes[k]
+                if new_token not in (blank_id, unk_id):
+
+                    assert new_token != 0, new_token
+                    token_list.append([new_token])
+                    hs.append(hyp.state[0])
+                    cs.append(hyp.state[1])
+        # forward RNNLM to get new states and scores
+        if len(token_list) != 0:
+            tokens_to_score = torch.tensor(token_list).to(torch.int64).to(device).reshape(-1,1)
+            
+            hs = torch.cat(hs, dim=1).to(device)
+            cs = torch.cat(cs, dim=1).to(device)
+            scores, lm_states = rnnlm.score_token(tokens_to_score, (hs,cs))
+        
+        count = 0 # index, used to locate score and lm states
         for i in range(batch_size):
             topk_log_probs, topk_indexes = ragged_log_probs[i].topk(beam)
 
@@ -1823,21 +1829,29 @@ def modified_beam_search_ngram_rescoring(
                 hyp_idx = topk_hyp_indexes[k]
                 hyp = A[i][hyp_idx]
 
-                new_ys = hyp.ys[:]
+                ys = hyp.ys[:]
+                
+                lm_score = hyp.lm_score
+                state = hyp.state
+                
+                hyp_log_prob = topk_log_probs[k]  # get score of current hyp
                 new_token = topk_token_indexes[k]
                 if new_token not in (blank_id, unk_id):
-                    new_ys.append(new_token)
-                    state_cost = hyp.state_cost.forward_one_step(new_token)
-                else:
-                    state_cost = hyp.state_cost
                     
-                # We only keep AM scores in new_hyp.log_prob
+                    ys.append(new_token)
-                new_log_prob = (
+                    hyp_log_prob += (
-                    topk_log_probs[k] - hyp.state_cost.lm_score * lm_scale
+                        lm_score[new_token] * lm_scale
-                )
+                    )  # add the lm score
+                    
+                    lm_score = scores[count]
+                    state = (lm_states[0][:, count, :].unsqueeze(1), lm_states[1][:, count, :].unsqueeze(1))
+                    count += 1
                 
                 new_hyp = Hypothesis(
-                    ys=new_ys, log_prob=new_log_prob, state_cost=state_cost
+                    ys=ys, 
+                    log_prob=hyp_log_prob,
+                    state=state,
+                    lm_score=lm_score
                 )
                 B[i].add(new_hyp) 
 

icefall/rnn_lm/model.py

@@ -18,8 +18,9 @@ import logging
 
 import torch
 import torch.nn.functional as F
+import k2
 
-from icefall.utils import make_pad_mask
+from icefall.utils import add_eos, add_sos, make_pad_mask
 
 
 class RnnLmModel(torch.nn.Module):
@@ -72,6 +73,8 @@ class RnnLmModel(torch.nn.Module):
         else:
             logging.info("Not tying weights")
 
+        self.cache = {}
+
     def forward(
         self, x: torch.Tensor, y: torch.Tensor, lengths: torch.Tensor
     ) -> torch.Tensor:
@@ -118,3 +121,124 @@ class RnnLmModel(torch.nn.Module):
         nll_loss = nll_loss.reshape(batch_size, -1)
 
         return nll_loss
+        
+    def get_init_states(self, sos):
+        p = next(self.parameters())
+
+    def predict_batch(self, tokens, token_lens, sos_id, eos_id, blank_id):
+        device = next(self.parameters()).device
+        batch_size = len(token_lens)
+
+        sos_tokens = add_sos(tokens, sos_id)
+        tokens_eos = add_eos(tokens, eos_id)
+        sos_tokens_row_splits = sos_tokens.shape.row_splits(1)
+
+        sentence_lengths = (
+            sos_tokens_row_splits[1:] - sos_tokens_row_splits[:-1]
+        )
+
+        x_tokens = sos_tokens.pad(mode="constant", padding_value=blank_id)
+        y_tokens = tokens_eos.pad(mode="constant", padding_value=blank_id)
+
+        x_tokens = x_tokens.to(torch.int64).to(device)
+        y_tokens = y_tokens.to(torch.int64).to(device)
+        sentence_lengths = sentence_lengths.to(torch.int64).to(device)
+
+        embedding = self.input_embedding(x_tokens)
+
+        # Note: We use batch_first==True
+        rnn_out, states = self.rnn(embedding)
+        logits = self.output_linear(rnn_out)
+        mask = torch.zeros(logits.shape).bool().to(device)
+        for i in range(batch_size):
+            mask[i, token_lens[i], :] = True
+        logits = logits[mask].reshape(batch_size, -1)
+        
+        return logits[:,:].log_softmax(-1), states
+    
+    def clean_cache(self):
+        self.cache = {}
+        
+    def score_token(self, tokens: torch.Tensor, state=None):
+        device = next(self.parameters()).device
+        batch_size = tokens.size(0)
+        if state:
+            h,c = state
+        else:
+            h = torch.zeros(self.rnn.num_layers, batch_size, self.rnn.input_size).to(device)
+            c = torch.zeros(self.rnn.num_layers, batch_size, self.rnn.input_size).to(device)
+            
+        embedding = self.input_embedding(tokens)
+        rnn_out, states = self.rnn(embedding, (h,c))
+        logits = self.output_linear(rnn_out)
+        
+        return logits[:,0].log_softmax(-1), states
+
+    def forward_with_state(self, tokens, token_lens, sos_id, eos_id, blank_id, state=None):
+        batch_size = len(token_lens)
+        if state:
+            h,c = state
+        else:
+            h = torch.zeros(self.rnn.num_layers, batch_size, self.rnn.input_size)
+            c = torch.zeros(self.rnn.num_layers, batch_size, self.rnn.input_size)
+        
+        device = next(self.parameters()).device
+
+        sos_tokens = add_sos(tokens, sos_id)
+        tokens_eos = add_eos(tokens, eos_id)
+        sos_tokens_row_splits = sos_tokens.shape.row_splits(1)
+
+        sentence_lengths = (
+            sos_tokens_row_splits[1:] - sos_tokens_row_splits[:-1]
+        )
+
+        x_tokens = sos_tokens.pad(mode="constant", padding_value=blank_id)
+        y_tokens = tokens_eos.pad(mode="constant", padding_value=blank_id)
+
+        x_tokens = x_tokens.to(torch.int64).to(device)
+        y_tokens = y_tokens.to(torch.int64).to(device)
+        sentence_lengths = sentence_lengths.to(torch.int64).to(device)
+
+        embedding = self.input_embedding(x_tokens)
+
+        # Note: We use batch_first==True
+        rnn_out, states = self.rnn(embedding, (h,c))
+        logits = self.output_linear(rnn_out)
+
+        return logits, states
+      
+if __name__=="__main__":
+    LM = RnnLmModel(500, 2048, 2048, 3, True)
+    h0 = torch.zeros(3, 1, 2048)
+    c0 = torch.zeros(3, 1, 2048)
+    seq = [[0,1,2,3]]
+    seq_lens = [len(s) for s in seq]
+    tokens = k2.RaggedTensor(seq)
+    output1, state = LM.forward_with_state(
+        tokens,
+        seq_lens,
+        1,
+        1,
+        0,
+        state=(h0,c0)
+    )
+    seq = [[0,1,2,3,4]]
+    seq_lens = [len(s) for s in seq]
+    tokens = k2.RaggedTensor(seq)
+    output2, _ = LM.forward_with_state(
+        tokens,
+        seq_lens,
+        1,
+        1,
+        0,
+        state=(h0,c0)
+    )
+    
+    seq = [[4]]
+    seq_lens = [len(s) for s in seq]
+    output3 = LM.score_token(seq, seq_lens, state)
+    
+    print("Finished")
+    
+    
+