From fb0da9f849a5d3bd473dcdbaf6197c6a5ff24a57 Mon Sep 17 00:00:00 2001
From: 游雁 <zhifu.gzf@alibaba-inc.com>
Date: 星期二, 07 五月 2024 15:53:26 +0800
Subject: [PATCH] decoding key
---
funasr/models/sense_voice/whisper_lib/decoding.py | 203 ++++++++++++++++++++------------------------------
1 files changed, 82 insertions(+), 121 deletions(-)
diff --git a/funasr/models/sense_voice/whisper_lib/decoding.py b/funasr/models/sense_voice/whisper_lib/decoding.py
index 62be3bc..609d6a6 100644
--- a/funasr/models/sense_voice/whisper_lib/decoding.py
+++ b/funasr/models/sense_voice/whisper_lib/decoding.py
@@ -19,7 +19,11 @@
@torch.no_grad()
def detect_language(
- model: "Whisper", mel: Tensor, tokenizer: Tokenizer = None, initial_prompt = None, x = None,
+ model: "Whisper",
+ mel: Tensor,
+ tokenizer: Tokenizer = None,
+ initial_prompt=None,
+ x=None,
) -> Tuple[Tensor, List[dict]]:
"""
Detect the spoken language in the audio, and return them as list of strings, along with the ids
@@ -34,16 +38,9 @@
list of dictionaries containing the probability distribution over all languages.
"""
if tokenizer is None:
- tokenizer = get_tokenizer(
- model.is_multilingual, num_languages=model.num_languages
- )
- if (
- tokenizer.language is None
- or tokenizer.language_token not in tokenizer.sot_sequence
- ):
- raise ValueError(
- "This model doesn't have language tokens so it can't perform lang id"
- )
+ tokenizer = get_tokenizer(model.is_multilingual, num_languages=model.num_languages)
+ if tokenizer.language is None or tokenizer.language_token not in tokenizer.sot_sequence:
+ raise ValueError("This model doesn't have language tokens so it can't perform lang id")
single = mel.ndim == 2
if single:
@@ -59,17 +56,21 @@
# FIX(funasr): sense vocie
# x = torch.tensor([[tokenizer.sot]] * n_audio).to(mel.device) # [n_audio, 1]
if x is None:
- x = torch.tensor([tokenizer.encode(initial_prompt, allowed_special="all")] * n_audio).to(mel.device) # [n_audio, 1]
+ x = torch.tensor([tokenizer.encode(initial_prompt, allowed_special="all")] * n_audio).to(
+ mel.device
+ ) # [n_audio, 1]
else:
x = x.to(mel.device)
+ # FIX(funasr): sense vocie
+ # logits = model.logits(x[:, :-1], mel)[:, -1]
+ logits = model.logits(x[:, :], mel)[:, -1]
- logits = model.logits(x[:,:-1], mel)[:, -1]
# collect detected languages; suppress all non-language tokens
mask = torch.ones(logits.shape[-1], dtype=torch.bool)
mask[list(tokenizer.all_language_tokens)] = False
mask[tokenizer.no_speech] = False
-
+
logits[:, mask] = -np.inf
language_tokens = logits.argmax(dim=-1)
language_token_probs = logits.softmax(dim=-1).cpu()
@@ -77,7 +78,10 @@
language_probs = [
{
c: language_token_probs[i, j].item()
- for j, c in zip(list(tokenizer.all_language_tokens) + [tokenizer.no_speech], list(tokenizer.all_language_codes) + ["nospeech"])
+ for j, c in zip(
+ list(tokenizer.all_language_tokens) + [tokenizer.no_speech],
+ list(tokenizer.all_language_codes) + ["nospeech"],
+ )
}
for i in range(n_audio)
]
@@ -119,14 +123,16 @@
suppress_blank: bool = True # this will suppress blank outputs
gain_event: bool = False # this will suppress blank outputs
- gain_tokens_bg: Optional[Union[str, List[int]]] = "<|Applause|><|Laughter|>"
- gain_tokens_ed: Optional[Union[str, List[int]]] = "<|/Applause|><|/Laughter|>"
- gain_tokens_score: List[float] = field(default_factory=lambda: [25.0, 5.0]) #[25, 5]
+ gain_tokens_bg: Optional[Union[str, List[int]]] = "<|Speech|><|BGM|><|Applause|><|Laughter|>"
+ gain_tokens_ed: Optional[Union[str, List[int]]] = (
+ "<|/Speech|><|/BGM|><|/Applause|><|/Laughter|>"
+ )
+ gain_tokens_score: List[float] = field(default_factory=lambda: [1, 1, 25.0, 5.0]) # [25, 5]
use_emo_threshold: bool = False # this will suppress blank outputs
emo_unk_token: Optional[Union[str, List[int]]] = "<|SPECIAL_TOKEN_1|>"
emo_target_tokens: Optional[Union[str, List[int]]] = "<|HAPPY|><|SAD|><|ANGRY|>"
- emo_target_threshold: List[float] = field(default_factory=lambda: [0.1, 0.1, 0.1]) #[25, 5]
+ emo_target_threshold: List[float] = field(default_factory=lambda: [0.1, 0.1, 0.1]) # [25, 5]
# timestamp sampling options
without_timestamps: bool = False # use <|notimestamps|> to sample text tokens only
@@ -203,9 +209,7 @@
class SequenceRanker:
- def rank(
- self, tokens: List[List[Tensor]], sum_logprobs: List[List[float]]
- ) -> List[int]:
+ def rank(self, tokens: List[List[Tensor]], sum_logprobs: List[List[float]]) -> List[int]:
"""
Given a list of groups of samples and their cumulative log probabilities,
return the indices of the samples in each group to select as the final result
@@ -243,9 +247,7 @@
def reset(self):
"""Initialize any stateful variables for decoding a new sequence"""
- def update(
- self, tokens: Tensor, logits: Tensor, sum_logprobs: Tensor
- ) -> Tuple[Tensor, bool]:
+ def update(self, tokens: Tensor, logits: Tensor, sum_logprobs: Tensor) -> Tuple[Tensor, bool]:
"""Specify how to select the next token, based on the current trace and logits
Parameters
@@ -300,9 +302,7 @@
self.temperature = temperature
self.eot = eot
- def update(
- self, tokens: Tensor, logits: Tensor, sum_logprobs: Tensor
- ) -> Tuple[Tensor, bool]:
+ def update(self, tokens: Tensor, logits: Tensor, sum_logprobs: Tensor) -> Tuple[Tensor, bool]:
if self.temperature == 0:
next_tokens = logits.argmax(dim=-1)
else:
@@ -339,16 +339,12 @@
self.max_candidates: int = round(beam_size * self.patience)
self.finished_sequences = None
- assert (
- self.max_candidates > 0
- ), f"Invalid beam size ({beam_size}) or patience ({patience})"
+ assert self.max_candidates > 0, f"Invalid beam size ({beam_size}) or patience ({patience})"
def reset(self):
self.finished_sequences = None
- def update(
- self, tokens: Tensor, logits: Tensor, sum_logprobs: Tensor
- ) -> Tuple[Tensor, bool]:
+ def update(self, tokens: Tensor, logits: Tensor, sum_logprobs: Tensor) -> Tuple[Tensor, bool]:
if tokens.shape[0] % self.beam_size != 0:
raise ValueError(f"{tokens.shape}[0] % {self.beam_size} != 0")
@@ -392,9 +388,7 @@
# add newly finished sequences to self.finished_sequences
assert len(self.finished_sequences) == len(finished_sequences)
- for previously_finished, newly_finished in zip(
- self.finished_sequences, finished_sequences
- ):
+ for previously_finished, newly_finished in zip(self.finished_sequences, finished_sequences):
for seq in sorted(newly_finished, key=newly_finished.get, reverse=True):
if len(previously_finished) >= self.max_candidates:
break # the candidate list is full
@@ -402,8 +396,7 @@
# mark as completed if all audio has enough number of samples
completed = all(
- len(sequences) >= self.max_candidates
- for sequences in self.finished_sequences
+ len(sequences) >= self.max_candidates for sequences in self.finished_sequences
)
return tokens, completed
@@ -411,9 +404,7 @@
# collect all finished sequences, including patience, and add unfinished ones if not enough
sum_logprobs = sum_logprobs.cpu()
for i, sequences in enumerate(self.finished_sequences):
- if (
- len(sequences) < self.beam_size
- ): # when not enough sequences are finished
+ if len(sequences) < self.beam_size: # when not enough sequences are finished
for j in list(np.argsort(sum_logprobs[i]))[::-1]:
sequence = preceding_tokens[i, j].tolist() + [self.eot]
sequences[tuple(sequence)] = sum_logprobs[i][j].item()
@@ -421,8 +412,7 @@
break
tokens: List[List[Tensor]] = [
- [torch.tensor(seq) for seq in sequences.keys()]
- for sequences in self.finished_sequences
+ [torch.tensor(seq) for seq in sequences.keys()] for sequences in self.finished_sequences
]
sum_logprobs: List[List[float]] = [
list(sequences.values()) for sequences in self.finished_sequences
@@ -463,8 +453,11 @@
def apply(self, logits: Tensor, tokens: Tensor):
logits[:, self.suppress_tokens] = -np.inf
+
class GainEventToken(LogitFilter):
- def __init__(self, bg_tokens: Sequence[int], ed_tokens:Sequence[int], gain_values: Sequence[float]):
+ def __init__(
+ self, bg_tokens: Sequence[int], ed_tokens: Sequence[int], gain_values: Sequence[float]
+ ):
self.bg_tokens = list(bg_tokens)
self.ed_tokens = list(ed_tokens)
self.gain_value = [np.log(max(ga, 1e-9)) for ga in gain_values]
@@ -477,13 +470,16 @@
sum_bg = sum([1 if x == bg else 0 for x in tokens[i]])
sum_ed = sum([1 if x == ed else 0 for x in tokens[i]])
logits[i, bg] += ga
- if sum_bg > sum_ed or tokens[i,-1] in [bg, ed]:
+ if sum_bg > sum_ed or tokens[i, -1] in [bg, ed]:
logits[i, bg] = -np.inf
if sum_bg <= sum_ed:
logits[i, ed] = -np.inf
+
class ThresholdEmoToken(LogitFilter):
- def __init__(self, unk_tokens: Sequence[int], emo_tokens:Sequence[int], th_values: Sequence[float]):
+ def __init__(
+ self, unk_tokens: Sequence[int], emo_tokens: Sequence[int], th_values: Sequence[float]
+ ):
self.unk_token = list(unk_tokens)[0]
self.emo_tokens = list(emo_tokens)
self.th_values = list(th_values)
@@ -493,7 +489,7 @@
for i in range(len(tokens)):
for emo, th in zip(self.emo_tokens, self.th_values):
if logits[i].argmax() == emo and logits[i].softmax(dim=-1)[emo] < th:
- logits[i, self.unk_token] = max(logits[i, emo], logits[i, self.unk_token])
+ logits[i, self.unk_token] = max(logits[i, emo], logits[i, self.unk_token])
logits[i, emo] = -np.inf
# for bg, ed, ga in zip(self.bg_tokens, self.ed_tokens, self.gain_value):
@@ -526,12 +522,8 @@
for k in range(tokens.shape[0]):
sampled_tokens = tokens[k, self.sample_begin :]
seq = [t for t in sampled_tokens.tolist()]
- last_was_timestamp = (
- len(seq) >= 1 and seq[-1] >= self.tokenizer.timestamp_begin
- )
- penultimate_was_timestamp = (
- len(seq) < 2 or seq[-2] >= self.tokenizer.timestamp_begin
- )
+ last_was_timestamp = len(seq) >= 1 and seq[-1] >= self.tokenizer.timestamp_begin
+ penultimate_was_timestamp = len(seq) < 2 or seq[-2] >= self.tokenizer.timestamp_begin
if last_was_timestamp:
if penultimate_was_timestamp: # has to be non-timestamp
@@ -539,9 +531,7 @@
else: # cannot be normal text tokens
logits[k, : self.tokenizer.eot] = -np.inf
- timestamps = sampled_tokens[
- sampled_tokens.ge(self.tokenizer.timestamp_begin)
- ]
+ timestamps = sampled_tokens[sampled_tokens.ge(self.tokenizer.timestamp_begin)]
if timestamps.numel() > 0:
# timestamps shouldn't decrease; forbid timestamp tokens smaller than the last
# also force each segment to have a nonzero length, to prevent infinite looping
@@ -557,17 +547,13 @@
# apply the `max_initial_timestamp` option
if self.max_initial_timestamp_index is not None:
- last_allowed = (
- self.tokenizer.timestamp_begin + self.max_initial_timestamp_index
- )
+ last_allowed = self.tokenizer.timestamp_begin + self.max_initial_timestamp_index
logits[:, last_allowed + 1 :] = -np.inf
# if sum of probability over timestamps is above any other token, sample timestamp
logprobs = F.log_softmax(logits.float(), dim=-1)
for k in range(tokens.shape[0]):
- timestamp_logprob = logprobs[k, self.tokenizer.timestamp_begin :].logsumexp(
- dim=-1
- )
+ timestamp_logprob = logprobs[k, self.tokenizer.timestamp_begin :].logsumexp(dim=-1)
max_text_token_logprob = logprobs[k, : self.tokenizer.timestamp_begin].max()
if timestamp_logprob > max_text_token_logprob:
logits[k, : self.tokenizer.timestamp_begin] = -np.inf
@@ -588,7 +574,7 @@
num_languages=model.num_languages,
language=language,
task=options.task,
- vocab_path=options.vocab_path
+ vocab_path=options.vocab_path,
)
self.tokenizer: Tokenizer = tokenizer
self.options: DecodingOptions = self._verify_options(options)
@@ -626,30 +612,28 @@
if self.options.suppress_tokens:
self.logit_filters.append(SuppressTokens(self._get_suppress_tokens()))
if self.options.gain_event:
- self.logit_filters.append(GainEventToken(
- self.tokenizer.encode(self.options.gain_tokens_bg, allowed_special="all"),
- self.tokenizer.encode(self.options.gain_tokens_ed, allowed_special="all"),
- self.options.gain_tokens_score
+ self.logit_filters.append(
+ GainEventToken(
+ self.tokenizer.encode(self.options.gain_tokens_bg, allowed_special="all"),
+ self.tokenizer.encode(self.options.gain_tokens_ed, allowed_special="all"),
+ self.options.gain_tokens_score,
)
)
if self.options.use_emo_threshold:
- self.logit_filters.append(ThresholdEmoToken(
- self.tokenizer.encode(self.options.emo_unk_token, allowed_special="all"),
- self.tokenizer.encode(self.options.emo_target_tokens, allowed_special="all"),
- self.options.emo_target_threshold
+ self.logit_filters.append(
+ ThresholdEmoToken(
+ self.tokenizer.encode(self.options.emo_unk_token, allowed_special="all"),
+ self.tokenizer.encode(self.options.emo_target_tokens, allowed_special="all"),
+ self.options.emo_target_threshold,
)
)
if not options.without_timestamps:
precision = CHUNK_LENGTH / model.dims.n_audio_ctx # usually 0.02 seconds
max_initial_timestamp_index = None
if options.max_initial_timestamp:
- max_initial_timestamp_index = round(
- self.options.max_initial_timestamp / precision
- )
+ max_initial_timestamp_index = round(self.options.max_initial_timestamp / precision)
self.logit_filters.append(
- ApplyTimestampRules(
- tokenizer, self.sample_begin, max_initial_timestamp_index
- )
+ ApplyTimestampRules(tokenizer, self.sample_begin, max_initial_timestamp_index)
)
def _verify_options(self, options: DecodingOptions) -> DecodingOptions:
@@ -660,9 +644,7 @@
raise ValueError("best_of with greedy sampling (T=0) is not compatible")
if options.patience is not None and options.beam_size is None:
raise ValueError("patience requires beam_size to be given")
- if options.length_penalty is not None and not (
- 0 <= options.length_penalty <= 1
- ):
+ if options.length_penalty is not None and not (0 <= options.length_penalty <= 1):
raise ValueError("length_penalty (alpha) should be a value between 0 and 1")
return options
@@ -672,9 +654,7 @@
if prefix := self.options.prefix:
prefix_tokens = (
- self.tokenizer.encode(" " + prefix.strip())
- if isinstance(prefix, str)
- else prefix
+ self.tokenizer.encode(" " + prefix.strip()) if isinstance(prefix, str) else prefix
)
if self.sample_len is not None:
max_prefix_len = self.n_ctx // 2 - self.sample_len
@@ -683,16 +663,10 @@
if prompt := self.options.prompt:
prompt_tokens = (
- self.tokenizer.encode(" " + prompt.strip())
- if isinstance(prompt, str)
- else prompt
+ self.tokenizer.encode(" " + prompt.strip()) if isinstance(prompt, str) else prompt
)
- tokens = (
- [self.tokenizer.sot_prev]
- + prompt_tokens[-(self.n_ctx // 2 - 1) :]
- + tokens
- )
- #FIX(funasr): sense vocie
+ tokens = [self.tokenizer.sot_prev] + prompt_tokens[-(self.n_ctx // 2 - 1) :] + tokens
+ # FIX(funasr): sense vocie
if initial_prompt := self.options.initial_prompt:
if self.options.language is not None:
initial_prompt = f"{initial_prompt}<|{self.options.language}|>"
@@ -700,7 +674,6 @@
else:
tokens = self.tokenizer.encode(initial_prompt, allowed_special="all")
tokens += [0]
-
return tuple(tokens)
@@ -746,12 +719,8 @@
else:
audio_features = self.model.encoder(mel)
- if audio_features.dtype != (
- torch.float16 if self.options.fp16 else torch.float32
- ):
- return TypeError(
- f"audio_features has an incorrect dtype: {audio_features.dtype}"
- )
+ if audio_features.dtype != (torch.float16 if self.options.fp16 else torch.float32):
+ return TypeError(f"audio_features has an incorrect dtype: {audio_features.dtype}")
return audio_features
@@ -766,15 +735,18 @@
languages = [max(probs, key=probs.get) for probs in lang_probs]
# FIX(funasr): sense vocie
# if self.options.language is None:
- # tokens[:, self.sot_index + 1] = lang_tokens # write language tokens
+ # tokens[:, self.sot_index + 1] = lang_tokens # write language tokens
if self.options.language is None:
# tokens[:, self.sot_index + 1] = lang_tokens # write language tokens
languages = "".join([f"<|{language}|>" for language in languages])
n_audio = audio_features.shape[0]
- lang_tokens = torch.tensor([self.tokenizer.encode(languages, allowed_special="all")] * n_audio).to(
- audio_features.device) # [n_audio, 1]
-
+ lang_tokens = torch.tensor(
+ [self.tokenizer.encode(languages, allowed_special="all")] * n_audio
+ ).to(
+ audio_features.device
+ ) # [n_audio, 1]
+
tokens[:, -1:] = lang_tokens[:, :]
languages = [languages]
@@ -789,9 +761,7 @@
for i in range(self.sample_len):
logits = self.inference.logits(tokens, audio_features)
- if (
- i == 0 and self.tokenizer.no_speech is not None
- ): # save no_speech_probs
+ if i == 0 and self.tokenizer.no_speech is not None: # save no_speech_probs
probs_at_sot = logits[:, self.sot_index].float().softmax(dim=-1)
no_speech_probs = probs_at_sot[:, self.tokenizer.no_speech].tolist()
@@ -825,12 +795,8 @@
languages, language_probs = self._detect_language(audio_features, tokens)
if self.options.task == "lang_id":
return [
- DecodingResult(
- audio_features=features, language=language, language_probs=probs
- )
- for features, language, probs in zip(
- audio_features, languages, language_probs
- )
+ DecodingResult(audio_features=features, language=language, language_probs=probs)
+ for features, language, probs in zip(audio_features, languages, language_probs)
]
# repeat text tensors by the group size, for beam search or best-of-n sampling
@@ -850,8 +816,7 @@
# get the final candidates for each group, and slice between the first sampled token and EOT
tokens, sum_logprobs = self.decoder.finalize(tokens, sum_logprobs)
tokens: List[List[Tensor]] = [
- [t[self.sample_begin : (t == tokenizer.eot).nonzero()[0, 0]] for t in s]
- for s in tokens
+ [t[self.sample_begin : (t == tokenizer.eot).nonzero()[0, 0]] for t in s] for s in tokens
]
# select the top-ranked sample in each group
@@ -860,9 +825,7 @@
texts: List[str] = [tokenizer.decode(t).strip() for t in tokens]
sum_logprobs: List[float] = [lp[i] for i, lp in zip(selected, sum_logprobs)]
- avg_logprobs: List[float] = [
- lp / (len(t) + 1) for t, lp in zip(tokens, sum_logprobs)
- ]
+ avg_logprobs: List[float] = [lp / (len(t) + 1) for t, lp in zip(tokens, sum_logprobs)]
fields = (
texts,
@@ -886,9 +849,7 @@
temperature=self.options.temperature,
compression_ratio=compression_ratio(text),
)
- for text, language, tokens, features, avg_logprob, no_speech_prob in zip(
- *fields
- )
+ for text, language, tokens, features, avg_logprob, no_speech_prob in zip(*fields)
]
--
Gitblit v1.9.1