"""Decoding methods for seq2seq autoregressive model. Authors * Adel Moumen 2022, 2023, 2024 * Ju-Chieh Chou 2020 * Peter Plantinga 2020 * Mirco Ravanelli 2020 * Sung-Lin Yeh 2020 """ from functools import cached_property import torch from torch.distributions import Categorical from speechbrain.decoders.utils import ( _update_mem, inflate_tensor, mask_by_condition, ) from speechbrain.utils.data_utils import undo_padding class AlivedHypotheses(torch.nn.Module): """This class handle the data for the hypotheses during the decoding. Arguments --------- alived_seq : torch.Tensor The sequence of tokens for each hypothesis. alived_log_probs : torch.Tensor The log probabilities of each token for each hypothesis. sequence_scores : torch.Tensor The sum of log probabilities for each hypothesis. """ def __init__(self, alived_seq, alived_log_probs, sequence_scores): super().__init__() self.alived_seq = alived_seq self.alived_log_probs = alived_log_probs self.sequence_scores = sequence_scores def __getitem__(self, index): return ( self.alived_seq[index], self.alived_log_probs[index], self.sequence_scores[index], ) def __str__(self): return f"AlivedHypotheses(alived_seq={self.alived_seq}, alived_log_probs={self.alived_log_probs}, sequence_scores={self.sequence_scores})" class S2SBaseSearcher(torch.nn.Module): """S2SBaseSearcher class to be inherited by other decoding approaches for seq2seq model. Arguments --------- bos_index : int The index of the beginning-of-sequence (bos) token. eos_index : int The index of end-of-sequence (eos) token. min_decode_ratio : float The ratio of minimum decoding steps to the length of encoder states. max_decode_ratio : float The ratio of maximum decoding steps to the length of encoder states. """ def __init__( self, bos_index, eos_index, min_decode_ratio, max_decode_ratio ): super().__init__() self.bos_index = bos_index self.eos_index = eos_index self.min_decode_ratio = min_decode_ratio self.max_decode_ratio = max_decode_ratio def forward(self, enc_states, wav_len): """This method should implement the forward algorithm of decoding method. Arguments --------- enc_states : torch.Tensor The precomputed encoder states to be used when decoding. (ex. the encoded speech representation to be attended). wav_len : torch.Tensor The speechbrain-style relative length. Returns ------- hyps The predicted tokens, as a list of lists or, if return_topk is True, a Tensor of shape (batch, topk, max length of token_id sequences). top_lengths The length of each topk sequence in the batch. top_scores This final scores of topk hypotheses. top_log_probs The log probabilities of each hypotheses. """ raise NotImplementedError return def forward_step(self, inp_tokens, memory, enc_states, enc_lens): """This method should implement one step of forwarding operation in the autoregressive model. Arguments --------- inp_tokens : torch.Tensor The input tensor of the current step. memory : No limit The memory variables input for this step. (ex. RNN hidden states). enc_states : torch.Tensor The encoder states to be attended. enc_lens : torch.Tensor The actual length of each enc_states sequence. Returns ------- log_probs : torch.Tensor Log-probabilities of the current step output. memory : No limit The memory variables generated in this step. (ex. RNN hidden states). attn : torch.Tensor The attention weight for doing penalty. """ raise NotImplementedError return def reset_mem(self, batch_size, device): """This method should implement the resetting of memory variables for the seq2seq model. E.g., initializing zero vector as initial hidden states. Arguments --------- batch_size : int The size of the batch. device : torch.device The device to put the initial variables. Return ------ memory : No limit The initial memory variable. """ raise NotImplementedError return def change_max_decoding_length(self, min_decode_steps, max_decode_steps): """set the minimum/maximum length of enc_states to be attended.""" return min_decode_steps, max_decode_steps def set_n_out(self): """set the number of output tokens. Overrides this function if the fc layer is embedded in the model, e.g., Whisper. """ return self.fc.w.out_features def _check_end_condition(self, memory): """This method is supposed to be overridden by the child class. For instance, if the decoder has a maximal number of tokens that it can attend to, this method should return True when the maximal number of tokens is reached. """ return False class S2SGreedySearcher(S2SBaseSearcher): """This class implements the general forward-pass of greedy decoding approach. See also S2SBaseSearcher(). """ @torch.no_grad() def forward(self, enc_states, wav_len): """This method performs a greedy search. Arguments --------- enc_states : torch.Tensor The precomputed encoder states to be used when decoding. (ex. the encoded speech representation to be attended). wav_len : torch.Tensor The speechbrain-style relative length. Returns ------- hyps : List[List[int]] List containing the hypotheses. top_lengths : torch.Tensor (batch) This tensor contains the length of each hypothesis. top_scores : torch.Tensor (batch) The score of each hypotheses. top_log_probs : torch.Tensor (batch, max length of token_id sequences) The log probabilities of each hypotheses. """ enc_lens = torch.round(enc_states.shape[1] * wav_len).int() device = enc_states.device batch_size = enc_states.shape[0] memory = self.reset_mem(batch_size, device=device) # Using bos as the first input inp_tokens = ( enc_states.new_zeros(batch_size).fill_(self.bos_index).long() ) log_probs_lst = [] min_decode_steps = int(enc_states.shape[1] * self.min_decode_ratio) max_decode_steps = int(enc_states.shape[1] * self.max_decode_ratio) min_decode_steps, max_decode_steps = self.change_max_decoding_length( min_decode_steps, max_decode_steps ) has_ended = enc_states.new_zeros(batch_size).bool() for step in range(min_decode_steps, max_decode_steps): logits, memory, _ = self.forward_step( inp_tokens, memory, enc_states, enc_lens ) if self.temperature == 0: inp_tokens = logits.argmax(dim=-1) else: inp_tokens = Categorical( logits=logits / self.temperature ).sample() log_probs = torch.nn.functional.log_softmax(logits.float(), dim=-1) log_probs_lst.append(log_probs) has_ended = has_ended | (inp_tokens == self.eos_index) log_probs[has_ended] = -torch.inf inp_tokens[has_ended] = self.eos_index if has_ended.all() or self._check_end_condition(memory): break log_probs = torch.stack(log_probs_lst, dim=1) scores, predictions = log_probs.max(dim=-1) mask = scores == -torch.inf scores[mask] = 0 predictions[mask] = self.eos_index ( top_hyps, top_lengths, top_scores, top_log_probs, ) = self._get_top_prediction(predictions, scores, log_probs) # Convert best hypothesis to list hyps = undo_padding(top_hyps[:, 0], top_lengths) return hyps, top_lengths, top_scores, top_log_probs def _get_top_prediction(self, hyps, scores, log_probs): """This method sorts the scores and return corresponding hypothesis and log probs. Arguments --------- hyps : torch.Tensor (batch, max length of token_id sequences) This tensor stores the predicted hypothesis. scores : torch.Tensor (batch) The score of each hypotheses. log_probs : torch.Tensor (batch, max length of token_id sequences) The log probabilities of each hypotheses. Returns ------- top_hyps : torch.Tensor (batch, max length of token_id sequences) This tensor stores the best predicted hypothesis. top_lengths : torch.Tensor (batch) This tensor contains the length of each hypothesis. top_scores : torch.Tensor (batch) The score of each hypotheses. top_log_probs : torch.Tensor (batch, max length of token_id sequences) The log probabilities of each hypotheses. """ batch_size = hyps.size(0) max_length = hyps.size(1) top_lengths = [max_length] * batch_size # Collect lengths of top hyps for pred_index in range(batch_size): pred = hyps[pred_index] pred_length = (pred == self.eos_index).nonzero(as_tuple=False) if len(pred_length) > 0: top_lengths[pred_index] = pred_length[0].item() # Convert lists to tensors top_lengths = torch.tensor( top_lengths, dtype=torch.float, device=hyps.device ) # Pick top log probabilities top_log_probs = log_probs # Use SpeechBrain style lengths top_lengths = top_lengths / max_length return ( hyps.unsqueeze(1), top_lengths.unsqueeze(1), scores.unsqueeze(1), top_log_probs.unsqueeze(1), ) class S2STransformerGreedySearcher(S2SGreedySearcher): """This class implements the greedy decoding for Transformer. Arguments --------- modules : list with the following one: model : torch.nn.Module A TransformerASR model. seq_lin : torch.nn.Module A linear output layer for the seq2seq model. temperature : float Temperature to use during decoding. **kwargs Arguments to pass to S2SGreedySearcher """ def __init__(self, modules, temperature=0.0, **kwargs): super().__init__(**kwargs) self.model = modules[0] self.fc = modules[1] self.softmax = torch.nn.LogSoftmax(dim=-1) self.temperature = temperature def reset_mem(self, batch_size, device): """Needed to reset the memory during greedy search.""" return None def forward_step(self, inp_tokens, memory, enc_states, enc_lens): """Performs a step in the implemented greedy searcher.""" memory = _update_mem(inp_tokens, memory) pred, attn = self.model.decode(memory, enc_states, enc_lens) logits = self.fc(pred) return logits[:, -1, :], memory, attn class S2SWhisperGreedySearcher(S2SGreedySearcher): """ This class implements the greedy decoding for Whisper neural nets made by OpenAI in https://cdn.openai.com/papers/whisper.pdf. Arguments --------- model: HuggingFaceWhisper The Whisper model. temperature: float The temperature to use during decoding. use_kv_cache: bool (default: True) Whether to use key-value cache. suppress_blank: bool (default: True) This will suppress blank outputs. suppress_tokens: str or list (default: "-1") list of tokens ids (or comma-separated token ids) to suppress "-1" will suppress a set of symbols as defined in `model.non_speech_tokens()` sample_len: int (default: None) Maximum number of tokens to sample. prefix: str or list (default: None) Prefix to add to the input tokens. See: https://github.com/openai/whisper/discussions/117#discussioncomment-3727051 prompt: str or list (default: None) Prompt to add to the input tokens. See: https://github.com/openai/whisper/discussions/117#discussioncomment-3727051 **kwargs see S2SBaseSearcher, arguments are directly passed. """ def __init__( self, model, temperature=0.0, use_kv_cache=True, suppress_blank=True, suppress_tokens="-1", sample_len=None, prefix=None, prompt=None, **kwargs, ): super().__init__( bos_index=model.bos, eos_index=model.eos, **kwargs, ) self.model = model self.temperature = temperature self.use_kv_cache = use_kv_cache self.kv_cache = None self.suppress_blank = suppress_blank self.suppress_tokens = suppress_tokens self.prefix = prefix self.prompt = prompt self.max_attn_tokens = self.model.model.decoder.config.max_length self.sample_len = sample_len or self.max_attn_tokens // 2 self.initial_tokens = self._get_initial_tokens() self.sample_begin: int = len(self.initial_tokens) self.eos_index: int = self.model.eos self.bos_index: int = self.initial_tokens[-1] self.no_speech_probs = None self.lang_tokens = None def set_lang_tokens(self, lang_tokens): """Set the language to be used during decoding.""" self.lang_tokens = lang_tokens def set_task(self, task): """Set the task to be used during decoding.""" self.model.set_task(task) self.initial_tokens = self._get_initial_tokens() self.sample_begin: int = len(self.initial_tokens) self.bos_index: int = self.initial_tokens[-1] def set_prompt(self, prompt): """Set the prompt to be used during decoding.""" self.prompt = prompt self.initial_tokens = self._get_initial_tokens() self.sample_begin: int = len(self.initial_tokens) self.bos_index: int = self.initial_tokens[-1] @cached_property def get_tokens_to_suppress(self): """Get the tokens to suppress during decoding if self.config.suppress_tokens is None.""" suppress_tokens = self.suppress_tokens if isinstance(suppress_tokens, str): suppress_tokens = [int(t) for t in suppress_tokens.split(",")] if -1 in suppress_tokens: suppress_tokens = [t for t in suppress_tokens if t >= 0] suppress_tokens.extend(self.model.non_speech_tokens) elif suppress_tokens is None or len(suppress_tokens) == 0: suppress_tokens = [] # interpret empty string as an empty list else: assert isinstance( suppress_tokens, list ), "suppress_tokens must be a list" suppress_tokens.extend( [ self.model.transcribe, self.model.translate, self.model.bos, self.model.bos_prev, self.model.bos_lm, ] ) return tuple(sorted(set(suppress_tokens))) def _get_initial_tokens(self): """Get the initial tokens to be used during decoding.""" tokens = self.model.tokenizer.prefix_tokens prefix = self.prefix prompt = self.prompt if prefix: prefix_tokens = ( self.model.tokenizer.encode( " " + prefix.strip(), add_special_tokens=False ) if isinstance(prefix, str) else prefix ) if self.sample_len is not None: max_prefix_len = self.max_attn_tokens // 2 - self.sample_len prefix_tokens = prefix_tokens[-max_prefix_len:] tokens = tokens + prefix_tokens if prompt: prompt_tokens = ( self.model.tokenizer.encode( " " + prompt.strip(), add_special_tokens=False ) if isinstance(prompt, str) else prompt ) tokens = ( [self.model.bos_prev] + prompt_tokens[-(self.max_attn_tokens // 2 - 1) :] + tokens ) return tuple(tokens) def reset_mem(self, batch_size, device): """This method set the first tokens to be decoder_input_tokens during search.""" # reset KV cache if self.use_kv_cache: self.kv_cache = None self.no_speech_probs = [torch.nan] * batch_size # the last token will be used as the first input token # explaining why we are skipping it. memory_tokens = self.initial_tokens[:-1] mem = torch.tensor([memory_tokens] * batch_size).to(device) if self.lang_tokens is not None: mem[:, self.initial_tokens.index(self.model.bos) + 1] = ( self.lang_tokens ) # after using it, reset it. self.lang_token = None return mem def forward_step(self, inp_tokens, memory, enc_states, enc_lens): """Performs a step in the implemented beamsearcher.""" tokens = _update_mem(inp_tokens, memory) logits, attn, kv = self.model.forward_decoder( enc_states, tokens, past_key_values=self.kv_cache ) if tokens.shape[1] == self.sample_begin: probs_at_bos = ( logits[:, self.initial_tokens.index(self.model.bos)] .float() .softmax(dim=-1) ) self.no_speech_probs = probs_at_bos[ :, self.model.no_speech ].tolist() logits = logits[:, -1] if self.use_kv_cache: self.kv_cache = kv if self.suppress_blank: if tokens.shape[1] == self.sample_begin: logits[ :, self.model.tokenizer.encode(" ", add_special_tokens=False) + [self.eos_index], ] = -torch.inf if self.suppress_tokens: if self.model.config.suppress_tokens is None: tokens_to_suppress = self.get_tokens_to_suppress else: tokens_to_suppress = self.model.get_suppress_tokens logits[:, list(tokens_to_suppress)] = -torch.inf return logits, tokens, attn def _check_end_condition(self, memory): """This method checks if the max length is reached.""" return memory.shape[1] >= self.max_attn_tokens - self.sample_begin class S2SRNNGreedySearcher(S2SGreedySearcher): """ This class implements the greedy decoding for AttentionalRNNDecoder (speechbrain/nnet/RNN.py). See also S2SBaseSearcher() and S2SGreedySearcher(). Arguments --------- embedding : torch.nn.Module An embedding layer. decoder : torch.nn.Module Attentional RNN decoder. linear : torch.nn.Module A linear output layer. temperature : float The temperature to use during decoding. **kwargs see S2SBaseSearcher, arguments are directly passed. Example ------- >>> import speechbrain as sb >>> from speechbrain.decoders import S2SRNNGreedySearcher >>> emb = torch.nn.Embedding(5, 3) >>> dec = sb.nnet.RNN.AttentionalRNNDecoder( ... "gru", "content", 3, 3, 1, enc_dim=7, input_size=3 ... ) >>> lin = sb.nnet.linear.Linear(n_neurons=5, input_size=3) >>> searcher = S2SRNNGreedySearcher( ... embedding=emb, ... decoder=dec, ... linear=lin, ... bos_index=0, ... eos_index=1, ... min_decode_ratio=0, ... max_decode_ratio=1, ... ) >>> batch_size = 2 >>> enc = torch.rand([batch_size, 6, 7]) >>> wav_len = torch.ones([batch_size]) >>> top_hyps, top_lengths, _, _ = searcher(enc, wav_len) """ def __init__(self, embedding, decoder, linear, temperature=0.0, **kwargs): super().__init__(**kwargs) self.emb = embedding self.dec = decoder self.fc = linear self.temperature = temperature self.softmax = torch.nn.LogSoftmax(dim=-1) def reset_mem(self, batch_size, device): """When doing greedy search, keep hidden state (hs) and context vector (c) as memory. """ hs = None self.dec.attn.reset() c = torch.zeros(batch_size, self.dec.attn_dim, device=device) return hs, c def forward_step(self, inp_tokens, memory, enc_states, enc_lens): """Performs a step in the implemented beamsearcher.""" hs, c = memory e = self.emb(inp_tokens) dec_out, hs, c, w = self.dec.forward_step( e, hs, c, enc_states, enc_lens ) logits = self.fc(dec_out) return logits, (hs, c), w class S2SBeamSearcher(S2SBaseSearcher): """This class implements the beam-search algorithm for the seq2seq model. See also S2SBaseSearcher(). Arguments --------- bos_index : int The index of beginning-of-sequence token. eos_index : int The index of end-of-sequence token. min_decode_ratio : float The ratio of minimum decoding steps to length of encoder states. max_decode_ratio : float The ratio of maximum decoding steps to length of encoder states. beam_size : int The width of beam. scorer: speechbrain.decoders.scorers.ScorerBuilder Scorer instance. Default: None. return_topk : bool Whether to return topk hypotheses. The topk hypotheses will be padded to the same length. Default: False. topk : int If return_topk is True, then return topk hypotheses. Default: 1. using_eos_threshold : bool Whether to use eos threshold. Default: True. eos_threshold : float The threshold coefficient for eos token. Default: 1.5. See 3.1.2 in reference: https://arxiv.org/abs/1904.02619 length_normalization : bool Whether to divide the scores by the length. Default: True. using_max_attn_shift: bool Whether using the max_attn_shift constraint. Default: False. max_attn_shift: int Beam search will block the beams that attention shift more than max_attn_shift. Default: 60. Reference: https://arxiv.org/abs/1904.02619 minus_inf : float The value of minus infinity to block some path of the search. Default: -1e20. """ def __init__( self, bos_index, eos_index, min_decode_ratio, max_decode_ratio, beam_size, scorer=None, return_topk=False, topk=1, using_eos_threshold=True, eos_threshold=1.5, length_normalization=True, using_max_attn_shift=False, max_attn_shift=60, minus_inf=-1e20, ): super().__init__( bos_index, eos_index, min_decode_ratio, max_decode_ratio ) self.beam_size = beam_size self.scorer = scorer self.return_topk = return_topk self.topk = topk self.length_normalization = length_normalization self.using_eos_threshold = using_eos_threshold self.eos_threshold = eos_threshold self.using_max_attn_shift = using_max_attn_shift self.max_attn_shift = max_attn_shift self.attn_weight = 1.0 self.ctc_weight = 0.0 self.minus_inf = minus_inf if self.scorer is not None: # Check length normalization if length_normalization and self.scorer.weights["length"] > 0.0: raise ValueError( "Length normalization is not compatible with length rewarding." ) if self.scorer.weights["ctc"] > 0.0: # Check indices for ctc all_scorers = { **self.scorer.full_scorers, **self.scorer.partial_scorers, } blank_index = all_scorers["ctc"].blank_index if len({bos_index, eos_index, blank_index}) < 3: raise ValueError( "Set blank, eos and bos to different indexes for joint ATT/CTC or CTC decoding" ) self.ctc_weight = self.scorer.weights["ctc"] self.attn_weight = 1.0 - self.ctc_weight def _check_full_beams(self, hyps): """This method checks whether hyps has been full. Arguments --------- hyps : List This list contains batch_size number. Each inside list contains a list stores all the hypothesis for this sentence. Returns ------- bool Whether the hyps has been full. """ hyps_len = [len(lst) for lst in hyps] beams_size = [self.beam_size for _ in range(len(hyps_len))] return hyps_len == beams_size def _check_attn_shift(self, attn, prev_attn_peak): """This method checks whether attention shift is more than attn_shift. Arguments --------- attn : torch.Tensor The attention to be checked. prev_attn_peak : torch.Tensor The previous attention peak place. Returns ------- cond : torch.BoolTensor Each element represents whether the beam is within the max_shift range. attn_peak : torch.Tensor The peak of the attn tensor. """ # Block the candidates that exceed the max shift _, attn_peak = torch.max(attn, dim=1) lt_cond = attn_peak <= (prev_attn_peak + self.max_attn_shift) mt_cond = attn_peak > (prev_attn_peak - self.max_attn_shift) # True if not exceed limit # Multiplication equals to element-wise and for tensor cond = (lt_cond * mt_cond).unsqueeze(1) return cond, attn_peak def _check_eos_threshold(self, log_probs): """This method checks whether eos log-probabilities exceed threshold. Arguments --------- log_probs : torch.Tensor The log-probabilities. Returns ------- cond : torch.BoolTensor Each element represents whether the eos log-probabilities will be kept. """ max_probs, _ = torch.max(log_probs, dim=-1) eos_probs = log_probs[:, self.eos_index] cond = eos_probs > (self.eos_threshold * max_probs) return cond def init_hypotheses(self): """This method initializes the AlivedHypotheses object. Returns ------- AlivedHypotheses The alived hypotheses filled with the initial values. """ return AlivedHypotheses( alived_seq=torch.empty(self.n_bh, 0, device=self.device).long(), alived_log_probs=torch.empty(self.n_bh, 0, device=self.device), sequence_scores=torch.empty(self.n_bh, device=self.device) .fill_(float("-inf")) .index_fill_(0, self.beam_offset, 0.0), ) def _attn_weight_step( self, inp_tokens, memory, enc_states, enc_lens, attn, log_probs ): """This method computes a forward_step if attn_weight is superior to 0. Arguments --------- inp_tokens : torch.Tensor The input tensor of the current step. memory : No limit The memory variables input for this step. (ex. RNN hidden states). enc_states : torch.Tensor The encoder states to be attended. enc_lens : torch.Tensor The actual length of each enc_states sequence. attn : torch.Tensor The attention weight. log_probs : torch.Tensor The log-probabilities of the current step output. Returns ------- log_probs : torch.Tensor Log-probabilities of the current step output. memory : No limit The memory variables generated in this step. (ex. RNN hidden states). attn : torch.Tensor The attention weight. """ if self.attn_weight > 0: log_probs, memory, attn = self.forward_step( inp_tokens, memory, enc_states, enc_lens ) log_probs = self.attn_weight * log_probs return log_probs, memory, attn def _max_attn_shift_step(self, attn, prev_attn_peak, log_probs): """This method will block the beams that attention shift more than max_attn_shift. Arguments --------- attn : torch.Tensor The attention weight. prev_attn_peak : torch.Tensor The previous attention peak place. log_probs : torch.Tensor The log-probabilities of the current step output. Returns ------- log_probs : torch.Tensor Log-probabilities of the current step output. prev_attn_peak : torch.Tensor The previous attention peak place. """ if self.using_max_attn_shift: cond, prev_attn_peak = self._check_attn_shift(attn, prev_attn_peak) log_probs = mask_by_condition( log_probs, cond, fill_value=self.minus_inf ) return log_probs, prev_attn_peak def _scorer_step(self, inp_tokens, scorer_memory, attn, log_probs): """This method call the scorers if scorer is not None. Arguments --------- inp_tokens : torch.Tensor The input tensor of the current step. scorer_memory : No limit The memory variables input for this step. (ex. RNN hidden states). attn : torch.Tensor The attention weight. log_probs : torch.Tensor The log-probabilities of the current step output. Returns ------- log_probs : torch.Tensor Log-probabilities of the current step output. scorer_memory : No limit The memory variables generated in this step. """ if self.scorer is not None: log_probs, scorer_memory = self.scorer.score( inp_tokens, scorer_memory, attn, log_probs, self.beam_size ) return log_probs, scorer_memory def _set_eos_minus_inf_step(self, log_probs, step, min_decode_steps): """This method set the log_probs of eos to minus infinity if the step is less than min_decode_steps. Arguments --------- log_probs : torch.Tensor The log-probabilities of the current step output. step : int The current decoding step. min_decode_steps : int The minimum decoding steps. Returns ------- log_probs : torch.Tensor Log-probabilities of the current step output. """ if step < min_decode_steps: log_probs[:, self.eos_index] = self.minus_inf return log_probs def _eos_threshold_step(self, log_probs): """This method set the log_probs of eos to minus infinity if the eos log-probabilities is less than eos_threshold. Arguments --------- log_probs : torch.Tensor The log-probabilities of the current step output. Returns ------- log_probs : torch.Tensor Log-probabilities of the current step output. """ if self.using_eos_threshold: cond = self._check_eos_threshold(log_probs) log_probs[:, self.eos_index] = mask_by_condition( log_probs[:, self.eos_index], cond, fill_value=self.minus_inf ) return log_probs def _attn_weight_permute_memory_step(self, memory, predecessors): """This method permute the memory if attn_weight is superior to 0. Arguments --------- memory : No limit The memory variables input for this step. (ex. RNN hidden states). predecessors : torch.Tensor The index of which beam the current top-K output came from in (t-1) steps. Returns ------- memory : No limit The memory variables generated in this step. (ex. RNN hidden states). """ if self.attn_weight > 0: memory = self.permute_mem(memory, index=predecessors) return memory def _scorer_permute_memory_step( self, scorer_memory, predecessors, candidates ): """This method permute the scorer_memory if scorer is not None. Arguments --------- scorer_memory : No limit The memory variables input for this step. (ex. RNN hidden states). predecessors : torch.Tensor The index of which beam the current top-K output came from in (t-1) steps. candidates : torch.Tensor The index of the current top-K output. Returns ------- scorer_memory : No limit The memory variables generated in this step. """ if self.scorer is not None: scorer_memory = self.scorer.permute_scorer_mem( scorer_memory, index=predecessors, candidates=candidates ) return scorer_memory def _max_attn_shift_permute_memory_step(self, prev_attn_peak, predecessors): """This method permute the prev_attn_peak if using_max_attn_shift is True. Arguments --------- prev_attn_peak : torch.Tensor The previous attention peak place. predecessors : torch.Tensor The index of which beam the current top-K output came from in (t-1) steps. Returns ------- prev_attn_peak : torch.Tensor The previous attention peak place. """ if self.using_max_attn_shift: prev_attn_peak = torch.index_select( prev_attn_peak, dim=0, index=predecessors ) return prev_attn_peak def _update_reset_memory(self, enc_states, enc_lens): """Call reset memory for each module. Arguments --------- enc_states : torch.Tensor The encoder states to be attended. enc_lens : torch.Tensor The actual length of each enc_states sequence. Returns ------- memory : No limit The memory variables generated in this step. scorer_memory : No limit The memory variables generated in this step. """ memory = self.reset_mem(self.n_bh, device=self.device) scorer_memory = None if self.scorer is not None: scorer_memory = self.scorer.reset_scorer_mem(enc_states, enc_lens) return memory, scorer_memory def _update_permute_memory( self, memory, scorer_memory, predecessors, candidates, prev_attn_peak ): """Call permute memory for each module. It allows us to synchronize the memory with the output. Arguments --------- memory : No limit The memory variables input for this step. (ex. RNN hidden states). scorer_memory : No limit The memory variables input for this step. (ex. RNN hidden states). predecessors : torch.Tensor The index of which beam the current top-K output came from in (t-1) steps. candidates : torch.Tensor The index of the current top-K output. prev_attn_peak : torch.Tensor The previous attention peak place. Returns ------- memory : No limit The memory variables generated in this step. scorer_memory : No limit The memory variables generated in this step. prev_attn_peak : torch.Tensor The previous attention peak place. """ memory = self._attn_weight_permute_memory_step(memory, predecessors) scorer_memory = self._scorer_permute_memory_step( scorer_memory, predecessors, candidates ) # If using_max_attn_shift, then the previous attn peak has to be permuted too. prev_attn_peak = self._max_attn_shift_permute_memory_step( prev_attn_peak, predecessors ) return memory, scorer_memory, prev_attn_peak def _update_sequences_and_log_probs( self, log_probs, inp_tokens, predecessors, candidates, alived_hyps ): """This method update sequences and log probabilities by adding the new inp_tokens. Arguments --------- log_probs : torch.Tensor The log-probabilities of the current step output. inp_tokens : torch.Tensor The input tensor of the current step. predecessors : torch.Tensor The index of which beam the current top-K output came from in (t-1) steps. candidates : torch.Tensor The index of the current top-K output. alived_hyps : AlivedHypotheses The alived hypotheses. Returns ------- alived_hyps : AlivedHypotheses The alived hypotheses. """ # Update alived_seq alived_hyps.alived_seq = torch.cat( [ torch.index_select( alived_hyps.alived_seq, dim=0, index=predecessors ), inp_tokens.unsqueeze(1), ], dim=-1, ) # Takes the log-probabilities beam_log_probs = log_probs[ torch.arange(self.batch_size).unsqueeze(1), candidates ].reshape(self.n_bh) # Update alived_log_probs alived_hyps.alived_log_probs = torch.cat( [ torch.index_select( alived_hyps.alived_log_probs, dim=0, index=predecessors ), beam_log_probs.unsqueeze(1), ], dim=-1, ) return alived_hyps def _compute_scores_and_next_inp_tokens(self, alived_hyps, log_probs, step): """Compute scores and next input tokens. Arguments --------- alived_hyps : AlivedHypotheses The alived hypotheses. log_probs : torch.Tensor The log-probabilities of the current step output. step : int The current decoding step. Returns ------- scores : torch.Tensor The scores of the current step output. candidates : torch.Tensor The index of the current top-K output. predecessors : torch.Tensor The index of which beam the current top-K output came from in (t-1) steps. inp_tokens : torch.Tensor The input tensor of the current step. alived_hyps : AlivedHypotheses The alived hypotheses. """ scores = alived_hyps.sequence_scores.unsqueeze(1).expand(-1, self.n_out) scores = scores + log_probs # length normalization if self.length_normalization: scores = scores / (step + 1) # keep topk beams scores, candidates = scores.view(self.batch_size, -1).topk( self.beam_size, dim=-1 ) # The input for the next step, also the output of current step. inp_tokens = (candidates % self.n_out).view(self.n_bh) scores = scores.view(self.n_bh) alived_hyps.sequence_scores = scores # recover the length normalization if self.length_normalization: alived_hyps.sequence_scores = alived_hyps.sequence_scores * ( step + 1 ) # The index of which beam the current top-K output came from in (t-1) steps. predecessors = ( torch.div(candidates, self.n_out, rounding_mode="floor") + self.beam_offset.unsqueeze(1).expand_as(candidates) ).view(self.n_bh) return ( scores, candidates, predecessors, inp_tokens, alived_hyps, ) def init_beam_search_data(self, enc_states, wav_len): """Initialize the beam search data. Arguments --------- enc_states : torch.Tensor The encoder states to be attended. wav_len : torch.Tensor The actual length of each enc_states sequence. Returns ------- alived_hyps : AlivedHypotheses The alived hypotheses. inp_tokens : torch.Tensor The input tensor of the current step. log_probs : torch.Tensor The log-probabilities of the current step output. eos_hyps_and_log_probs_scores : list Generated hypotheses (the ones that have reached eos) and log probs scores. memory : No limit The memory variables generated in this step. scorer_memory : No limit The memory variables generated in this step. attn : torch.Tensor The attention weight. prev_attn_peak : torch.Tensor The previous attention peak place. enc_states : torch.Tensor The encoder states to be attended. enc_lens : torch.Tensor The actual length of each enc_states sequence. """ enc_lens = torch.round(enc_states.shape[1] * wav_len).int() self.device = enc_states.device self.batch_size = enc_states.shape[0] self.n_bh = self.batch_size * self.beam_size self.n_out = self.set_n_out() memory, scorer_memory = self._update_reset_memory(enc_states, enc_lens) # Inflate the enc_states and enc_len by beam_size times enc_states = inflate_tensor(enc_states, times=self.beam_size, dim=0) enc_lens = inflate_tensor(enc_lens, times=self.beam_size, dim=0) # Using bos as the first input inp_tokens = ( torch.zeros(self.n_bh, device=self.device) .fill_(self.bos_index) .long() ) # The first index of each sentence. self.beam_offset = ( torch.arange(self.batch_size, device=self.device) * self.beam_size ) # initialize sequence scores variables. sequence_scores = torch.empty(self.n_bh, device=self.device).fill_( self.minus_inf ) # keep only the first to make sure no redundancy. sequence_scores.index_fill_(0, self.beam_offset, 0.0) # keep the hypothesis that reaches eos and their corresponding score and log_probs. eos_hyps_and_log_probs_scores = [[] for _ in range(self.batch_size)] self.min_decode_steps = int(enc_states.shape[1] * self.min_decode_ratio) self.max_decode_steps = int(enc_states.shape[1] * self.max_decode_ratio) # the decoding steps can be based on the max number of tokens that a decoder can process # (e.g., 448 for Whisper). ( self.min_decode_steps, self.max_decode_steps, ) = self.change_max_decoding_length( self.min_decode_steps, self.max_decode_steps ) # Initialize the previous attention peak to zero # This variable will be used when using_max_attn_shift=True prev_attn_peak = torch.zeros(self.n_bh, device=self.device) attn = None log_probs = torch.full((self.n_bh, self.n_out), 0.0, device=self.device) alived_hyps = self.init_hypotheses() return ( alived_hyps, inp_tokens, log_probs, eos_hyps_and_log_probs_scores, memory, scorer_memory, attn, prev_attn_peak, enc_states, enc_lens, ) def _update_hyps_and_scores_if_eos_token( self, inp_tokens, alived_hyps, eos_hyps_and_log_probs_scores, scores ): """This method will update hyps and scores if inp_tokens are eos. Arguments --------- inp_tokens : torch.Tensor The current output. alived_hyps : AlivedHypotheses alived_seq : torch.Tensor alived_log_probs : torch.Tensor eos_hyps_and_log_probs_scores : list Generated hypotheses (the ones that have reached eos) and log probs scores. scores : torch.Tensor Scores at the current step. Returns ------- is_eos : torch.BoolTensor Each element represents whether the token is eos. """ is_eos = inp_tokens.eq(self.eos_index) (eos_indices,) = torch.nonzero(is_eos, as_tuple=True) # Store the hypothesis and their scores when reaching eos. if eos_indices.shape[0] > 0: for index in eos_indices: # convert to int index = index.item() batch_id = torch.div( index, self.beam_size, rounding_mode="floor" ) if ( len(eos_hyps_and_log_probs_scores[batch_id]) == self.beam_size ): continue hyp = alived_hyps.alived_seq[index, :] log_probs = alived_hyps.alived_log_probs[index, :] final_scores = scores[index].clone() eos_hyps_and_log_probs_scores[batch_id].append( (hyp, log_probs, final_scores) ) return is_eos def _get_topk_prediction(self, eos_hyps_and_log_probs_scores): """This method sorts the scores and return corresponding hypothesis and log probs. Arguments --------- eos_hyps_and_log_probs_scores : list Generated hypotheses (the ones that have reached eos) and log probs scores. Returns ------- topk_hyps : torch.Tensor (batch, topk, max length of token_id sequences) This tensor stores the topk predicted hypothesis. topk_lengths : torch.Tensor (batch, topk) This tensor contains the final scores of topk hypotheses. topk_scores : torch.Tensor (batch, topk) The length of each topk sequence in the batch. topk_log_probs : torch.Tensor (batch, topk, max length of token_id sequences) The log probabilities of each hypotheses. """ top_hyps, top_log_probs, top_scores, top_lengths = [], [], [], [] batch_size = len(eos_hyps_and_log_probs_scores) # Collect hypotheses for i in range(len(eos_hyps_and_log_probs_scores)): hyps, log_probs, scores = zip(*eos_hyps_and_log_probs_scores[i]) top_hyps += hyps top_scores += scores top_log_probs += log_probs top_lengths += [len(hyp) for hyp in hyps] # Convert lists to tensors top_hyps = torch.nn.utils.rnn.pad_sequence( top_hyps, batch_first=True, padding_value=0 ) top_log_probs = torch.nn.utils.rnn.pad_sequence( top_log_probs, batch_first=True, padding_value=0 ) top_lengths = torch.tensor( top_lengths, dtype=torch.float, device=top_hyps.device ) top_scores = torch.stack((top_scores), dim=0).view(batch_size, -1) # Use SpeechBrain style lengths top_lengths = (top_lengths - 1) / top_hyps.size(1) # Get topk indices topk_scores, indices = top_scores.topk(self.topk, dim=-1) indices = (indices + self.beam_offset.unsqueeze(1)).view( batch_size * self.topk ) # Select topk hypotheses topk_hyps = torch.index_select(top_hyps, dim=0, index=indices) topk_hyps = topk_hyps.view(batch_size, self.topk, -1) topk_lengths = torch.index_select(top_lengths, dim=0, index=indices) topk_lengths = topk_lengths.view(batch_size, self.topk) topk_log_probs = torch.index_select(top_log_probs, dim=0, index=indices) topk_log_probs = topk_log_probs.view(batch_size, self.topk, -1) return topk_hyps, topk_lengths, topk_scores, topk_log_probs def search_step( self, alived_hyps, inp_tokens, log_probs, eos_hyps_and_log_probs_scores, memory, scorer_memory, attn, prev_attn_peak, enc_states, enc_lens, step, ): """A search step for the next most likely tokens. Arguments --------- alived_hyps : AlivedHypotheses The alived hypotheses. inp_tokens : torch.Tensor The input tensor of the current step. log_probs : torch.Tensor The log-probabilities of the current step output. eos_hyps_and_log_probs_scores : list Generated hypotheses (the ones that have reached eos) and log probs scores. memory : No limit The memory variables input for this step. (ex. RNN hidden states). scorer_memory : No limit The memory variables input for this step. (ex. RNN hidden states). attn : torch.Tensor The attention weight. prev_attn_peak : torch.Tensor The previous attention peak place. enc_states : torch.Tensor The encoder states to be attended. enc_lens : torch.Tensor The actual length of each enc_states sequence. step : int The current decoding step. Returns ------- alived_hyps : AlivedHypotheses The alived hypotheses. inp_tokens : torch.Tensor The input tensor of the current step. log_probs : torch.Tensor The log-probabilities of the current step output. eos_hyps_and_log_probs_scores : list Generated hypotheses (the ones that have reached eos) and log probs scores. memory : No limit The memory variables generated in this step. scorer_memory : No limit The memory variables generated in this step. attn : torch.Tensor The attention weight. prev_attn_peak : torch.Tensor The previous attention peak place. scores : torch.Tensor The scores of the current step output. """ (log_probs, memory, attn) = self._attn_weight_step( inp_tokens, memory, enc_states, enc_lens, attn, log_probs ) # Keep the original value log_probs_clone = log_probs.clone().reshape(self.batch_size, -1) (log_probs, prev_attn_peak) = self._max_attn_shift_step( attn, prev_attn_peak, log_probs ) log_probs = self._set_eos_minus_inf_step( log_probs, step, self.min_decode_steps ) log_probs = self._eos_threshold_step(log_probs) (log_probs, scorer_memory) = self._scorer_step( inp_tokens, scorer_memory, attn, log_probs ) ( scores, candidates, predecessors, inp_tokens, alived_hyps, ) = self._compute_scores_and_next_inp_tokens( alived_hyps, log_probs, step ) memory, scorer_memory, prev_attn_peak = self._update_permute_memory( memory, scorer_memory, predecessors, candidates, prev_attn_peak ) alived_hyps = self._update_sequences_and_log_probs( log_probs_clone, inp_tokens, predecessors, candidates, alived_hyps ) is_eos = self._update_hyps_and_scores_if_eos_token( inp_tokens, alived_hyps, eos_hyps_and_log_probs_scores, scores ) # Block the paths that have reached eos. alived_hyps.sequence_scores.masked_fill_(is_eos, float("-inf")) return ( alived_hyps, inp_tokens, log_probs, eos_hyps_and_log_probs_scores, memory, scorer_memory, attn, prev_attn_peak, scores, ) def _fill_alived_hyps_with_eos_token( self, alived_hyps, eos_hyps_and_log_probs_scores, scores ): """Fill the alived_hyps that have not reached eos with eos. Arguments --------- alived_hyps : AlivedHypotheses The alived hypotheses. eos_hyps_and_log_probs_scores : list Generated hypotheses (the ones that have reached eos) and log probs scores. scores : torch.Tensor The scores of the current step output. Returns ------- eos_hyps_and_log_probs_scores : list Generated hypotheses (the ones that have reached eos) and log probs scores. """ if not self._check_full_beams(eos_hyps_and_log_probs_scores): # Using all eos to fill-up the hyps. inp_tokens = ( torch.zeros(self.n_bh, device=self.device) .fill_(self.eos_index) .long() ) self._update_hyps_and_scores_if_eos_token( inp_tokens, alived_hyps, eos_hyps_and_log_probs_scores, scores ) return eos_hyps_and_log_probs_scores def forward(self, enc_states, wav_len): # noqa: C901 """Applies beamsearch and returns the predicted tokens. Arguments --------- enc_states : torch.Tensor The encoder states to be attended. wav_len : torch.Tensor The actual length of each enc_states sequence. Returns ------- hyps : list The predicted tokens. best_lens : torch.Tensor The length of each predicted tokens. best_scores : torch.Tensor The scores of each predicted tokens. best_log_probs : torch.Tensor The log probabilities of each predicted tokens. """ ( alived_hyps, inp_tokens, log_probs, eos_hyps_and_log_probs_scores, memory, scorer_memory, attn, prev_attn_peak, enc_states, enc_lens, ) = self.init_beam_search_data(enc_states, wav_len) for step in range(self.max_decode_steps): # terminate condition if self._check_full_beams(eos_hyps_and_log_probs_scores): break ( alived_hyps, inp_tokens, log_probs, eos_hyps_and_log_probs_scores, memory, scorer_memory, attn, prev_attn_peak, scores, ) = self.search_step( alived_hyps, inp_tokens, log_probs, eos_hyps_and_log_probs_scores, memory, scorer_memory, attn, prev_attn_peak, enc_states, enc_lens, step, ) if self._check_end_condition(alived_hyps): break finals_hyps_and_log_probs_scores = ( self._fill_alived_hyps_with_eos_token( alived_hyps, eos_hyps_and_log_probs_scores, scores ) ) ( topk_hyps, topk_lengths, topk_scores, topk_log_probs, ) = self._get_topk_prediction(finals_hyps_and_log_probs_scores) if self.return_topk: return topk_hyps, topk_lengths, topk_scores, topk_log_probs else: # select the best hyps best_hyps = topk_hyps[:, 0, :] best_lens = topk_lengths[:, 0] best_scores = topk_scores[:, 0] best_log_probs = topk_log_probs[:, 0, :] # Convert best hypothesis to list hyps = undo_padding(best_hyps, best_lens) return hyps, best_lens, best_scores, best_log_probs def _check_end_condition(self, alived_hyps): """This method is supposed to be overridden by the child class. For instance, if the decoder has a maximal number of tokens that it can attend to, this method should return True when the maximal number of tokens is reached. """ return False def permute_mem(self, memory, index): """This method permutes the seq2seq model memory to synchronize the memory index with the current output. Arguments --------- memory : No limit The memory variable to be permuted. index : torch.Tensor The index of the previous path. Returns ------- The variable of the memory being permuted. """ raise NotImplementedError return class S2SRNNBeamSearcher(S2SBeamSearcher): """ This class implements the beam search decoding for AttentionalRNNDecoder (speechbrain/nnet/RNN.py). See also S2SBaseSearcher(), S2SBeamSearcher(). Arguments --------- embedding : torch.nn.Module An embedding layer. decoder : torch.nn.Module Attentional RNN decoder. linear : torch.nn.Module A linear output layer. temperature : float Temperature factor applied to softmax. It changes the probability distribution, being softer when T>1 and sharper with T<1. **kwargs see S2SBeamSearcher, arguments are directly passed. Example ------- >>> import speechbrain as sb >>> vocab_size = 5 >>> emb = torch.nn.Embedding(vocab_size, 3) >>> dec = sb.nnet.RNN.AttentionalRNNDecoder( ... "gru", "content", 3, 3, 1, enc_dim=7, input_size=3 ... ) >>> lin = sb.nnet.linear.Linear(n_neurons=vocab_size, input_size=3) >>> coverage_scorer = sb.decoders.scorer.CoverageScorer(vocab_size) >>> scorer = sb.decoders.scorer.ScorerBuilder( ... full_scorers = [coverage_scorer], ... partial_scorers = [], ... weights= dict(coverage=1.5) ... ) >>> searcher = S2SRNNBeamSearcher( ... embedding=emb, ... decoder=dec, ... linear=lin, ... bos_index=4, ... eos_index=4, ... min_decode_ratio=0, ... max_decode_ratio=1, ... beam_size=2, ... scorer=scorer, ... ) >>> batch_size = 2 >>> enc = torch.rand([batch_size, 6, 7]) >>> wav_len = torch.ones([batch_size]) >>> hyps, _, _, _ = searcher(enc, wav_len) """ def __init__(self, embedding, decoder, linear, temperature=1.0, **kwargs): super().__init__(**kwargs) self.emb = embedding self.dec = decoder self.fc = linear self.softmax = torch.nn.LogSoftmax(dim=-1) self.temperature = temperature def reset_mem(self, batch_size, device): """Needed to reset the memory during beamsearch.""" hs = None self.dec.attn.reset() c = torch.zeros(batch_size, self.dec.attn_dim, device=device) return hs, c def forward_step(self, inp_tokens, memory, enc_states, enc_lens): """Performs a step in the implemented beamsearcher.""" with torch.no_grad(): hs, c = memory e = self.emb(inp_tokens) dec_out, hs, c, w = self.dec.forward_step( e, hs, c, enc_states, enc_lens ) log_probs = self.softmax(self.fc(dec_out) / self.temperature) # average attn weight of heads when attn_type is multiheadlocation if self.dec.attn_type == "multiheadlocation": w = torch.mean(w, dim=1) return log_probs, (hs, c), w def permute_mem(self, memory, index): """Memory permutation during beamsearch.""" hs, c = memory # shape of hs: [num_layers, batch_size, n_neurons] if isinstance(hs, tuple): hs_0 = torch.index_select(hs[0], dim=1, index=index) hs_1 = torch.index_select(hs[1], dim=1, index=index) hs = (hs_0, hs_1) else: hs = torch.index_select(hs, dim=1, index=index) c = torch.index_select(c, dim=0, index=index) if self.dec.attn_type == "location": self.dec.attn.prev_attn = torch.index_select( self.dec.attn.prev_attn, dim=0, index=index ) return (hs, c) class S2STransformerBeamSearcher(S2SBeamSearcher): """This class implements the beam search decoding for Transformer. See also S2SBaseSearcher(), S2SBeamSearcher(). Arguments --------- modules : list with the following one: model : torch.nn.Module A Transformer model. seq_lin : torch.nn.Module A linear output layer. temperature : float Temperature factor applied to softmax. It changes the probability distribution, being softer when T>1 and sharper with T<1. **kwargs Arguments to pass to S2SBeamSearcher Example ------- >>> from speechbrain.nnet.linear import Linear >>> from speechbrain.lobes.models.transformer.TransformerASR import TransformerASR >>> from speechbrain.decoders import S2STransformerBeamSearcher >>> batch_size=8 >>> n_channels=6 >>> input_size=40 >>> d_model=128 >>> tgt_vocab=140 >>> src = torch.rand([batch_size, n_channels, input_size]) >>> tgt = torch.randint(0, tgt_vocab, [batch_size, n_channels]) >>> net = TransformerASR( ... tgt_vocab, input_size, d_model, 8, 1, 1, 1024, activation=torch.nn.GELU ... ) >>> ctc_lin = Linear(input_shape=(1, 40, d_model), n_neurons=tgt_vocab) >>> lin = Linear(input_shape=(1, 40, d_model), n_neurons=tgt_vocab) >>> searcher = S2STransformerBeamSearcher( ... modules=[net, lin], ... bos_index=1, ... eos_index=2, ... min_decode_ratio=0.0, ... max_decode_ratio=1.0, ... using_eos_threshold=False, ... beam_size=7, ... temperature=1.15, ... ) >>> enc, dec = net.forward(src, tgt) >>> hyps, _, _, _ = searcher(enc, torch.ones(batch_size)) """ def __init__(self, modules, temperature=1.0, **kwargs): super().__init__(**kwargs) self.model = modules[0] self.fc = modules[1] self.softmax = torch.nn.LogSoftmax(dim=-1) self.temperature = temperature def reset_mem(self, batch_size, device): """Needed to reset the memory during beamsearch.""" return None def permute_mem(self, memory, index): """Memory permutation during beamsearch.""" memory = torch.index_select(memory, dim=0, index=index) return memory def forward_step(self, inp_tokens, memory, enc_states, enc_lens): """Performs a step in the implemented beamsearcher.""" memory = _update_mem(inp_tokens, memory) pred, attn = self.model.decode(memory, enc_states, enc_lens) prob_dist = self.softmax(self.fc(pred) / self.temperature) return prob_dist[:, -1, :], memory, attn class S2SWhisperBeamSearcher(S2SBeamSearcher): """This class implements the beam search decoding for Whisper neural nets made by OpenAI in https://cdn.openai.com/papers/whisper.pdf. The beam search is stateful, meaning that some variables are stored in the searcher. If you want to reuse the searcher in different contexts, you should make sure that the variables are updated accordingly. Arguments --------- module : list with the following one: model : torch.nn.Module A whisper model. It should have a decode() method. temperature: float The temperature to use during decoding. use_kv_cache: bool (default: True) Whether to use key-value cache. suppress_blank: bool (default: True) This will suppress blank outputs. suppress_tokens: str or list (default: "-1") list of tokens ids (or comma-separated token ids) to suppress "-1" will suppress a set of symbols as defined in `model.non_speech_tokens()` sample_len: int (default: None) Maximum number of tokens to sample. prefix: str or list (default: None) Prefix to add to the input tokens. See: https://github.com/openai/whisper/discussions/117#discussioncomment-3727051 prompt: str or list (default: None) Prompt to add to the input tokens. See: https://github.com/openai/whisper/discussions/117#discussioncomment-3727051 **kwargs see S2SBeamSearcher, arguments are directly passed. """ def __init__( self, module, temperature=1.0, use_kv_cache=True, suppress_blank=True, suppress_tokens="-1", sample_len=None, prefix=None, prompt=None, **kwargs, ): super().__init__( bos_index=module[0].bos, eos_index=module[0].eos, **kwargs, ) self.model = module[0] self.temperature = temperature self.use_kv_cache = use_kv_cache self.kv_cache = None self.suppress_blank = suppress_blank self.suppress_tokens = suppress_tokens self.prefix = prefix self.prompt = prompt self.max_attn_tokens = self.model.model.decoder.config.max_length self.sample_len = sample_len or self.max_attn_tokens // 2 self.initial_tokens = self._get_initial_tokens() self.sample_begin: int = len(self.initial_tokens) self.eos_index: int = self.model.eos self.bos_index: int = self.initial_tokens[-1] self.no_speech_probs = None self.lang_tokens = None def set_lang_tokens(self, lang_tokens): """Set the language to be used during decoding.""" self.lang_tokens = lang_tokens def set_task(self, task): """Set the task to be used during decoding.""" self.model.set_task(task) self.initial_tokens = self._get_initial_tokens() self.sample_begin: int = len(self.initial_tokens) self.bos_index: int = self.initial_tokens[-1] def set_prompt(self, prompt): """Set the prompt to be used during decoding.""" self.prompt = prompt self.initial_tokens = self._get_initial_tokens() self.sample_begin: int = len(self.initial_tokens) self.bos_index: int = self.initial_tokens[-1] @cached_property def get_tokens_to_suppress(self): """Get the tokens to suppress during decoding if self.config.suppress_tokens is None.""" suppress_tokens = self.suppress_tokens if isinstance(suppress_tokens, str): suppress_tokens = [int(t) for t in suppress_tokens.split(",")] if -1 in suppress_tokens: suppress_tokens = [t for t in suppress_tokens if t >= 0] suppress_tokens.extend(self.model.non_speech_tokens) elif suppress_tokens is None or len(suppress_tokens) == 0: suppress_tokens = [] # interpret empty string as an empty list else: assert isinstance( suppress_tokens, list ), "suppress_tokens must be a list" suppress_tokens.extend( [ self.model.transcribe, self.model.translate, self.model.bos, self.model.bos_prev, self.model.bos_lm, ] ) return tuple(sorted(set(suppress_tokens))) def _get_initial_tokens(self): """Get the initial tokens to be used during decoding.""" tokens = self.model.tokenizer.prefix_tokens prefix = self.prefix prompt = self.prompt if prefix: prefix_tokens = ( self.model.tokenizer.encode( " " + prefix.strip(), add_special_tokens=False ) if isinstance(prefix, str) else prefix ) if self.sample_len is not None: max_prefix_len = self.max_attn_tokens // 2 - self.sample_len prefix_tokens = prefix_tokens[-max_prefix_len:] tokens = tokens + prefix_tokens if prompt: prompt_tokens = ( self.model.tokenizer.encode( " " + prompt.strip(), add_special_tokens=False ) if isinstance(prompt, str) else prompt ) tokens = ( [self.model.bos_prev] + prompt_tokens[-(self.max_attn_tokens // 2 - 1) :] + tokens ) return tuple(tokens) def reset_mem(self, batch_size, device): """This method set the first tokens to be decoder_input_tokens during search.""" # reset KV cache if self.use_kv_cache: self.kv_cache = None self.no_speech_probs = [torch.nan] * batch_size # the last token will be used as the first input token # explaining why we are skipping it. memory_tokens = self.initial_tokens[:-1] mem = torch.tensor([memory_tokens] * batch_size).to(device) if self.lang_tokens is not None: mem[:, self.initial_tokens.index(self.model.bos) + 1] = ( self.lang_tokens ) # after using it, reset it. self.lang_token = None return mem def permute_mem(self, memory, index): """Permutes the memory.""" memory = torch.index_select(memory, dim=0, index=index) # if using kv_cache, we need to permute the kv_cache as well if self.use_kv_cache: self.kv_cache = self._reorder_cache(self.kv_cache, index) return memory def _reorder_cache(self, past_key_values, beam_idx): """Reorder the key-value cache. Arguments --------- past_key_values : tuple The key-value cache. beam_idx : torch.Tensor The index of the previous path. Returns ------- The reordered key-value cache. """ reordered_past = () for layer_past in past_key_values: reordered_past += ( tuple( past_state.index_select(0, beam_idx) for past_state in layer_past ), ) return reordered_past def set_n_out(self): """set the number of output tokens.""" return self.model.model.decoder.embed_tokens.weight.shape[0] def forward_step(self, inp_tokens, memory, enc_states, enc_lens): """Performs a step in the implemented beamsearcher.""" tokens = _update_mem(inp_tokens, memory) logits, attn, kv = self.model.forward_decoder( enc_states, tokens, past_key_values=self.kv_cache ) if tokens.shape[1] == self.sample_begin: probs_at_bos = ( logits[:, self.initial_tokens.index(self.model.bos)] .float() .softmax(dim=-1) ) self.no_speech_probs = probs_at_bos[ :, self.model.no_speech ].tolist() logits = logits[:, -1] if self.use_kv_cache: self.kv_cache = kv if self.suppress_blank: if tokens.shape[1] == self.sample_begin: logits[ :, self.model.tokenizer.encode(" ", add_special_tokens=False) + [self.eos_index], ] = -torch.inf if self.suppress_tokens: if self.model.config.suppress_tokens is None: tokens_to_suppress = self.get_tokens_to_suppress else: tokens_to_suppress = self.model.get_suppress_tokens logits[:, list(tokens_to_suppress)] = -torch.inf log_probs = ( torch.nn.functional.log_softmax(logits.float(), dim=-1) / self.temperature ) return log_probs, tokens, attn def _check_end_condition(self, alived_hyps): """This method checks if the max length is reached.""" return ( alived_hyps.alived_seq.shape[1] >= self.max_attn_tokens - self.sample_begin ) class S2SHFTextBasedBeamSearcher(S2STransformerBeamSearcher): """This class implements the beam search decoding for the text-based HF seq2seq models, such as mBART or NLLB. It is NOT significantly different from S2STransformerBeamSearcher. This is why it inherits S2STransformerBeamSearcher. The main difference might arise when one wishes to use directly the lm_head of the text-based HF model rather than making a new projection layer (self.fc = None). Arguments --------- modules : list with the following one: model : torch.nn.Module A Transformer model. seq_lin : torch.nn.Module A linear output layer. Normally set to None for this usecase. vocab_size : int The dimension of the lm_head. **kwargs Arguments to pass to S2SBeamSearcher """ def __init__(self, modules, vocab_size, **kwargs): super().__init__(modules, **kwargs) self.vocab_size = vocab_size def forward_step(self, inp_tokens, memory, enc_states, enc_lens): """Performs a step in the implemented beamsearcher.""" memory = _update_mem(inp_tokens, memory) pred, attn = self.model.decode(memory, enc_states, enc_lens) if self.fc is not None: pred = self.fc(pred) prob_dist = self.softmax(pred / self.temperature) return prob_dist[:, -1, :], memory, attn def set_n_out(self): """set the number of output tokens.""" return self.vocab_size