""" Alignment code Authors * Elena Rastorgueva 2020 * Loren Lugosch 2020 """ import random import torch from speechbrain.utils.checkpoints import ( mark_as_loader, mark_as_saver, register_checkpoint_hooks, ) from speechbrain.utils.data_utils import undo_padding @register_checkpoint_hooks class HMMAligner(torch.nn.Module): """This class calculates Viterbi alignments in the forward method. It also records alignments and creates batches of them for use in Viterbi training. Arguments --------- states_per_phoneme : int Number of hidden states to use per phoneme. output_folder : str It is the folder that the alignments will be stored in when saved to disk. Not yet implemented. neg_inf : float The float used to represent a negative infinite log probability. Using `-float("Inf")` tends to give numerical instability. A number more negative than -1e5 also sometimes gave errors when the `genbmm` library was used (currently not in use). (default: -1e5) batch_reduction : string One of "none", "sum" or "mean". What kind of batch-level reduction to apply to the loss calculated in the forward method. input_len_norm : bool Whether to normalize the loss in the forward method by the length of the inputs. target_len_norm : bool Whether to normalize the loss in the forward method by the length of the targets. lexicon_path : string The location of the lexicon. Example ------- >>> log_posteriors = torch.tensor([[[ -1., -10., -10.], ... [-10., -1., -10.], ... [-10., -10., -1.]], ... ... [[ -1., -10., -10.], ... [-10., -1., -10.], ... [-10., -10., -10.]]]) >>> lens = torch.tensor([1., 0.66]) >>> phns = torch.tensor([[0, 1, 2], ... [0, 1, 0]]) >>> phn_lens = torch.tensor([1., 0.66]) >>> aligner = HMMAligner() >>> forward_scores = aligner( ... log_posteriors, lens, phns, phn_lens, 'forward' ... ) >>> forward_scores.shape torch.Size([2]) >>> viterbi_scores, alignments = aligner( ... log_posteriors, lens, phns, phn_lens, 'viterbi' ... ) >>> alignments [[0, 1, 2], [0, 1]] >>> viterbi_scores.shape torch.Size([2]) """ def __init__( self, states_per_phoneme=1, output_folder="", neg_inf=-1e5, batch_reduction="none", input_len_norm=False, target_len_norm=False, lexicon_path=None, ): super().__init__() self.states_per_phoneme = states_per_phoneme self.output_folder = output_folder self.neg_inf = neg_inf self.batch_reduction = batch_reduction self.input_len_norm = input_len_norm self.target_len_norm = target_len_norm self.align_dict = {} self.lexicon_path = lexicon_path if self.lexicon_path is not None: with open(self.lexicon_path, "r", encoding="utf-8") as f: lines = f.readlines() for i, line in enumerate(lines): if line[0] != ";": start_index = i break lexicon = {} # {"read": {0: "r eh d", 1: "r iy d"}} lexicon_phones = set() for i in range(start_index, len(lines)): line = lines[i] word = line.split()[0] phones = line.split("/")[1] phones = "".join([p for p in phones if not p.isdigit()]) for p in phones.split(" "): lexicon_phones.add(p) if "~" in word: word = word.split("~")[0] if word in lexicon: number_of_existing_pronunciations = len(lexicon[word]) lexicon[word][number_of_existing_pronunciations] = phones else: lexicon[word] = {0: phones} self.lexicon = lexicon lexicon_phones = list(lexicon_phones) lexicon_phones.sort() self.lex_lab2ind = {p: i + 1 for i, p in enumerate(lexicon_phones)} self.lex_ind2lab = {i + 1: p for i, p in enumerate(lexicon_phones)} # add sil, which is not in the lexicon self.lex_lab2ind["sil"] = 0 self.lex_ind2lab[0] = "sil" def _use_lexicon(self, words, interword_sils, sample_pron): """Do processing using the lexicon to return a sequence of the possible phonemes, the transition/pi probabilities, and the possible final states. Inputs correspond to a single utterance, not a whole batch. Arguments --------- words : list List of the words in the transcript. interword_sils : bool If True, optional silences will be inserted between every word. If False, optional silences will only be placed at the beginning and end of each utterance. sample_pron : bool If True, it will sample a single possible sequence of phonemes. If False, it will return statistics for all possible sequences of phonemes. Returns ------- poss_phns : torch.Tensor (phoneme) The phonemes that are thought to be in each utterance. log_transition_matrix : torch.Tensor (batch, from, to) Tensor containing transition (log) probabilities. start_states : list of ints A list of the possible starting states in each utterance. final_states : list of ints A list of the possible final states for each utterance. """ number_of_states = 0 words_prime = ( [] ) # This will contain one "word" for each optional silence and pronunciation. # structure of each "word_prime": # [word index, [[state sequence 1], [state sequence 2]], ] word_index = 0 phoneme_indices = [] for word in words: if word_index == 0 or interword_sils is True: # optional silence word_prime = [ word_index, [ [ number_of_states + i for i in range(self.states_per_phoneme) ] ], True, ] words_prime.append(word_prime) phoneme_indices += [ self.silence_index * self.states_per_phoneme + i for i in range(self.states_per_phoneme) ] number_of_states += self.states_per_phoneme word_index += 1 # word word_prime = [word_index, [], False] if sample_pron and len(self.lexicon[word]) > 1: random.shuffle(self.lexicon[word]) for pron_idx in range(len(self.lexicon[word])): pronunciation = self.lexicon[word][pron_idx] phonemes = pronunciation.split() word_prime[1].append([]) for p in phonemes: phoneme_indices += [ self.lex_lab2ind[p] * self.states_per_phoneme + i for i in range(self.states_per_phoneme) ] word_prime[1][pron_idx] += [ number_of_states + i for i in range(self.states_per_phoneme) ] number_of_states += self.states_per_phoneme if sample_pron: break words_prime.append(word_prime) word_index += 1 # optional final silence word_prime = [ word_index, [[number_of_states + i for i in range(self.states_per_phoneme)]], True, ] words_prime.append(word_prime) phoneme_indices += [ self.silence_index * self.states_per_phoneme + i for i in range(self.states_per_phoneme) ] number_of_states += self.states_per_phoneme word_index += 1 transition_matrix = 1.0 * torch.eye( number_of_states ) # diagonal = all states have a self-loop final_states = [] for word_prime in words_prime: word_idx = word_prime[0] is_optional_silence = word_prime[-1] next_word_exists = word_idx < len(words_prime) - 2 this_word_last_states = [ word_prime[1][i][-1] for i in range(len(word_prime[1])) ] # create transitions to next state from previous state within each pronunciation for pronunciation in word_prime[1]: for state_idx in range(len(pronunciation) - 1): state = pronunciation[state_idx] next_state = pronunciation[state_idx + 1] transition_matrix[state, next_state] = 1.0 # create transitions to next word's starting states if next_word_exists: if is_optional_silence or not interword_sils: next_word_idx = word_idx + 1 else: next_word_idx = word_idx + 2 next_word_starting_states = [ words_prime[next_word_idx][1][i][0] for i in range(len(words_prime[next_word_idx][1])) ] for this_word_last_state in this_word_last_states: for next_word_starting_state in next_word_starting_states: transition_matrix[ this_word_last_state, next_word_starting_state ] = 1.0 else: final_states += this_word_last_states if not is_optional_silence: next_silence_idx = word_idx + 1 next_silence_starting_state = words_prime[next_silence_idx][1][ 0 ][0] for this_word_last_state in this_word_last_states: transition_matrix[ this_word_last_state, next_silence_starting_state ] = 1.0 log_transition_matrix = transition_matrix.log().log_softmax(1) start_states = [words_prime[0][1][0][0]] start_states += [ words_prime[1][1][i][0] for i in range(len(words_prime[1][1])) ] poss_phns = torch.tensor(phoneme_indices) return poss_phns, log_transition_matrix, start_states, final_states def use_lexicon(self, words, interword_sils=True, sample_pron=False): """Do processing using the lexicon to return a sequence of the possible phonemes, the transition/pi probabilities, and the possible final states. Does processing on an utterance-by-utterance basis. Each utterance in the batch is processed by a helper method `_use_lexicon`. Arguments --------- words : list List of the words in the transcript interword_sils : bool If True, optional silences will be inserted between every word. If False, optional silences will only be placed at the beginning and end of each utterance. sample_pron: bool If True, it will sample a single possible sequence of phonemes. If False, it will return statistics for all possible sequences of phonemes. Returns ------- poss_phns: torch.Tensor (batch, phoneme in possible phn sequence) The phonemes that are thought to be in each utterance. poss_phn_lens: torch.Tensor (batch) The relative length of each possible phoneme sequence in the batch. trans_prob: torch.Tensor (batch, from, to) Tensor containing transition (log) probabilities. pi_prob: torch.Tensor (batch, state) Tensor containing initial (log) probabilities. final_state: list of lists of ints A list of lists of possible final states for each utterance. Example ------- >>> aligner = HMMAligner() >>> aligner.lexicon = { ... "a": {0: "a"}, ... "b": {0: "b", 1: "c"} ... } >>> words = [["a", "b"]] >>> aligner.lex_lab2ind = { ... "sil": 0, ... "a": 1, ... "b": 2, ... "c": 3, ... } >>> poss_phns, poss_phn_lens, trans_prob, pi_prob, final_states = aligner.use_lexicon( ... words, ... interword_sils = True ... ) >>> poss_phns tensor([[0, 1, 0, 2, 3, 0]]) >>> poss_phn_lens tensor([1.]) >>> trans_prob tensor([[[-6.9315e-01, -6.9315e-01, -1.0000e+05, -1.0000e+05, -1.0000e+05, -1.0000e+05], [-1.0000e+05, -1.3863e+00, -1.3863e+00, -1.3863e+00, -1.3863e+00, -1.0000e+05], [-1.0000e+05, -1.0000e+05, -1.0986e+00, -1.0986e+00, -1.0986e+00, -1.0000e+05], [-1.0000e+05, -1.0000e+05, -1.0000e+05, -6.9315e-01, -1.0000e+05, -6.9315e-01], [-1.0000e+05, -1.0000e+05, -1.0000e+05, -1.0000e+05, -6.9315e-01, -6.9315e-01], [-1.0000e+05, -1.0000e+05, -1.0000e+05, -1.0000e+05, -1.0000e+05, 0.0000e+00]]]) >>> pi_prob tensor([[-6.9315e-01, -6.9315e-01, -1.0000e+05, -1.0000e+05, -1.0000e+05, -1.0000e+05]]) >>> final_states [[3, 4, 5]] >>> # With no optional silences between words >>> poss_phns_, _, trans_prob_, pi_prob_, final_states_ = aligner.use_lexicon( ... words, ... interword_sils = False ... ) >>> poss_phns_ tensor([[0, 1, 2, 3, 0]]) >>> trans_prob_ tensor([[[-6.9315e-01, -6.9315e-01, -1.0000e+05, -1.0000e+05, -1.0000e+05], [-1.0000e+05, -1.0986e+00, -1.0986e+00, -1.0986e+00, -1.0000e+05], [-1.0000e+05, -1.0000e+05, -6.9315e-01, -1.0000e+05, -6.9315e-01], [-1.0000e+05, -1.0000e+05, -1.0000e+05, -6.9315e-01, -6.9315e-01], [-1.0000e+05, -1.0000e+05, -1.0000e+05, -1.0000e+05, 0.0000e+00]]]) >>> pi_prob_ tensor([[-6.9315e-01, -6.9315e-01, -1.0000e+05, -1.0000e+05, -1.0000e+05]]) >>> final_states_ [[2, 3, 4]] >>> # With sampling of a single possible pronunciation >>> import random >>> random.seed(0) >>> poss_phns_, _, trans_prob_, pi_prob_, final_states_ = aligner.use_lexicon( ... words, ... sample_pron = True ... ) >>> poss_phns_ tensor([[0, 1, 0, 2, 0]]) >>> trans_prob_ tensor([[[-6.9315e-01, -6.9315e-01, -1.0000e+05, -1.0000e+05, -1.0000e+05], [-1.0000e+05, -1.0986e+00, -1.0986e+00, -1.0986e+00, -1.0000e+05], [-1.0000e+05, -1.0000e+05, -6.9315e-01, -6.9315e-01, -1.0000e+05], [-1.0000e+05, -1.0000e+05, -1.0000e+05, -6.9315e-01, -6.9315e-01], [-1.0000e+05, -1.0000e+05, -1.0000e+05, -1.0000e+05, 0.0000e+00]]]) """ self.silence_index = self.lex_lab2ind["sil"] poss_phns = [] trans_prob = [] start_states = [] final_states = [] for words_ in words: ( poss_phns_, trans_prob_, start_states_, final_states_, ) = self._use_lexicon(words_, interword_sils, sample_pron) poss_phns.append(poss_phns_) trans_prob.append(trans_prob_) start_states.append(start_states_) final_states.append(final_states_) # pad poss_phns, trans_prob with 0 to have same length poss_phn_lens = [len(poss_phns_) for poss_phns_ in poss_phns] U_max = max(poss_phn_lens) batch_size = len(poss_phns) for index in range(batch_size): phn_pad_length = U_max - len(poss_phns[index]) poss_phns[index] = torch.nn.functional.pad( poss_phns[index], (0, phn_pad_length), value=0 ) trans_prob[index] = torch.nn.functional.pad( trans_prob[index], (0, phn_pad_length, 0, phn_pad_length), value=self.neg_inf, ) # Stack into single tensor poss_phns = torch.stack(poss_phns) trans_prob = torch.stack(trans_prob) trans_prob[trans_prob == -float("Inf")] = self.neg_inf # make pi prob pi_prob = self.neg_inf * torch.ones([batch_size, U_max]) for start_state in start_states: pi_prob[:, start_state] = 1 pi_prob = torch.nn.functional.log_softmax(pi_prob, dim=1) # Convert poss_phn_lens from absolute to relative lengths poss_phn_lens = torch.tensor(poss_phn_lens).float() / U_max return poss_phns, poss_phn_lens, trans_prob, pi_prob, final_states def _make_pi_prob(self, phn_lens_abs): """Creates tensor of initial (log) probabilities (known as 'pi'). Assigns all probability mass to the first phoneme in the sequence. Arguments --------- phn_lens_abs : torch.Tensor (batch) The absolute length of each phoneme sequence in the batch. Returns ------- pi_prob : torch.Tensor (batch, phn) """ batch_size = len(phn_lens_abs) U_max = int(phn_lens_abs.max()) pi_prob = self.neg_inf * torch.ones([batch_size, U_max]) pi_prob[:, 0] = 0 return pi_prob def _make_trans_prob(self, phn_lens_abs): """Creates tensor of transition (log) probabilities. Only allows transitions to the same phoneme (self-loop) or the next phoneme in the phn sequence Arguments --------- phn_lens_abs : torch.Tensor (batch) The absolute length of each phoneme sequence in the batch. Returns ------- trans_prob : torch.Tensor (batch, from, to) """ # Extract useful values for later batch_size = len(phn_lens_abs) U_max = int(phn_lens_abs.max()) device = phn_lens_abs.device ## trans_prob matrix consists of 2 diagonals: ## (1) offset diagonal (next state) & ## (2) main diagonal (self-loop) # make offset diagonal trans_prob_off_diag = torch.eye(U_max - 1) zero_side = torch.zeros([U_max - 1, 1]) zero_bottom = torch.zeros([1, U_max]) trans_prob_off_diag = torch.cat((zero_side, trans_prob_off_diag), 1) trans_prob_off_diag = torch.cat((trans_prob_off_diag, zero_bottom), 0) # make main diagonal trans_prob_main_diag = torch.eye(U_max) # join the diagonals and repeat for whole batch trans_prob = trans_prob_off_diag + trans_prob_main_diag trans_prob = ( trans_prob.reshape(1, U_max, U_max) .repeat(batch_size, 1, 1) .to(device) ) # clear probabilities for too-long sequences mask_a = ( torch.arange(U_max, device=device)[None, :] < phn_lens_abs[:, None] ) mask_a = mask_a.unsqueeze(2) mask_a = mask_a.expand(-1, -1, U_max) mask_b = mask_a.permute(0, 2, 1) trans_prob = trans_prob * (mask_a & mask_b).float() ## put -infs in place of zeros: trans_prob = torch.where( trans_prob == 1, trans_prob, torch.tensor(-float("Inf"), device=device), ) ## normalize trans_prob = torch.nn.functional.log_softmax(trans_prob, dim=2) ## set nans to v neg numbers trans_prob[trans_prob != trans_prob] = self.neg_inf ## set -infs to v neg numbers trans_prob[trans_prob == -float("Inf")] = self.neg_inf return trans_prob def _make_emiss_pred_useful( self, emission_pred, lens_abs, phn_lens_abs, phns ): """Creates a 'useful' form of the posterior probabilities, rearranged into the order of phoneme appearance in phns. Arguments --------- emission_pred : torch.Tensor (batch, time, phoneme in vocabulary) posterior probabilities from our acoustic model lens_abs : torch.Tensor (batch) The absolute length of each input to the acoustic model, i.e., the number of frames. phn_lens_abs : torch.Tensor (batch) The absolute length of each phoneme sequence in the batch. phns : torch.Tensor (batch, phoneme in phn sequence) The phonemes that are known/thought to be in each utterance. Returns ------- emiss_pred_useful : torch.Tensor Tensor shape (batch, phoneme in phn sequence, time). """ # Extract useful values for later U_max = int(phn_lens_abs.max().item()) fb_max_length = int(lens_abs.max().item()) device = emission_pred.device # apply mask based on lens_abs mask_lens = ( torch.arange(fb_max_length).to(device)[None, :] < lens_abs[:, None] ) emiss_pred_acc_lens = torch.where( mask_lens[:, :, None], emission_pred, torch.tensor([0.0], device=device), ) # manipulate phn tensor, and then 'torch.gather' phns = phns.to(device) phns_copied = phns.unsqueeze(1).expand(-1, fb_max_length, -1) emiss_pred_useful = torch.gather(emiss_pred_acc_lens, 2, phns_copied) # apply mask based on phn_lens_abs mask_phn_lens = ( torch.arange(U_max).to(device)[None, :] < phn_lens_abs[:, None] ) emiss_pred_useful = torch.where( mask_phn_lens[:, None, :], emiss_pred_useful, torch.tensor([self.neg_inf], device=device), ) emiss_pred_useful = emiss_pred_useful.permute(0, 2, 1) return emiss_pred_useful def _dp_forward( self, pi_prob, trans_prob, emiss_pred_useful, lens_abs, phn_lens_abs, phns, ): """Does forward dynamic programming algorithm. Arguments --------- pi_prob : torch.Tensor (batch, phn) Tensor containing initial (log) probabilities. trans_prob : torch.Tensor (batch, from, to) Tensor containing transition (log) probabilities. emiss_pred_useful : torch.Tensor (batch, phoneme in phn sequence, time) A 'useful' form of the posterior probabilities, rearranged into the order of phoneme appearance in phns. lens_abs : torch.Tensor (batch) The absolute length of each input to the acoustic model, i.e., the number of frames. phn_lens_abs : torch.Tensor (batch) The absolute length of each phoneme sequence in the batch. phns : torch.Tensor (batch, phoneme in phn sequence) The phonemes that are known/thought to be in each utterance. Returns ------- sum_alpha_T : torch.Tensor (batch) The (log) likelihood of each utterance in the batch. """ # useful values batch_size = len(phn_lens_abs) U_max = phn_lens_abs.max() fb_max_length = lens_abs.max() device = emiss_pred_useful.device pi_prob = pi_prob.to(device) trans_prob = trans_prob.to(device) # initialise alpha_matrix = self.neg_inf * torch.ones( [batch_size, U_max, fb_max_length], device=device ) alpha_matrix[:, :, 0] = pi_prob + emiss_pred_useful[:, :, 0] for t in range(1, fb_max_length): utt_lens_passed = lens_abs < t if True in utt_lens_passed: n_passed = utt_lens_passed.sum() I_tensor = self.neg_inf * torch.ones(n_passed, U_max, U_max) I_tensor[:, torch.arange(U_max), torch.arange(U_max)] = 0.0 I_tensor = I_tensor.to(device) trans_prob[utt_lens_passed] = I_tensor alpha_times_trans = batch_log_matvecmul( trans_prob.permute(0, 2, 1), alpha_matrix[:, :, t - 1] ) alpha_matrix[:, :, t] = ( alpha_times_trans + emiss_pred_useful[:, :, t] ) sum_alpha_T = torch.logsumexp( alpha_matrix[torch.arange(batch_size), :, -1], dim=1 ) return sum_alpha_T def _dp_viterbi( self, pi_prob, trans_prob, emiss_pred_useful, lens_abs, phn_lens_abs, phns, final_states, ): """Calculates Viterbi alignment using dynamic programming. Arguments --------- pi_prob : torch.Tensor (batch, phn) Tensor containing initial (log) probabilities. trans_prob : torch.Tensor (batch, from, to) Tensor containing transition (log) probabilities. emiss_pred_useful : torch.Tensor (batch, phoneme in phn sequence, time) A 'useful' form of the posterior probabilities, rearranged into the order of phoneme appearance in phns. lens_abs : torch.Tensor (batch) The absolute length of each input to the acoustic model, i.e., the number of frames. phn_lens_abs : torch.Tensor (batch) The absolute length of each phoneme sequence in the batch. phns : torch.Tensor (batch, phoneme in phn sequence) The phonemes that are known/thought to be in each utterance. final_states : list List of final states Returns ------- z_stars : list of lists of int Viterbi alignments for the files in the batch. z_stars_loc : list of lists of int The locations of the Viterbi alignments for the files in the batch. e.g., for a batch with a single utterance with 5 phonemes, `z_stars_loc` will look like: [[0, 0, 0, 1, 1, 2, 3, 3, 3, 4, 4]]. viterbi_scores : torch.Tensor (batch) The (log) likelihood of the Viterbi path for each utterance. """ # useful values batch_size = len(phn_lens_abs) U_max = phn_lens_abs.max() fb_max_length = lens_abs.max() device = emiss_pred_useful.device pi_prob = pi_prob.to(device) trans_prob = trans_prob.to(device) v_matrix = self.neg_inf * torch.ones( [batch_size, U_max, fb_max_length], device=device ) backpointers = -99 * torch.ones( [batch_size, U_max, fb_max_length], device=device ) # initialise v_matrix[:, :, 0] = pi_prob + emiss_pred_useful[:, :, 0] for t in range(1, fb_max_length): x, argmax = batch_log_maxvecmul( trans_prob.permute(0, 2, 1), v_matrix[:, :, t - 1] ) v_matrix[:, :, t] = x + emiss_pred_useful[:, :, t] backpointers[:, :, t] = argmax.type(torch.FloatTensor) z_stars = [] z_stars_loc = [] for utterance_in_batch in range(batch_size): len_abs = lens_abs[utterance_in_batch] if final_states is not None: final_states_utter = final_states[utterance_in_batch] # Pick most probable of the final states viterbi_finals = v_matrix[ utterance_in_batch, final_states_utter, len_abs - 1 ] final_state_chosen = torch.argmax(viterbi_finals).item() U = final_states_utter[final_state_chosen] else: U = phn_lens_abs[utterance_in_batch].long().item() - 1 z_star_i_loc = [U] z_star_i = [phns[utterance_in_batch, z_star_i_loc[0]].item()] for time_step in range(len_abs, 1, -1): current_best_loc = z_star_i_loc[0] earlier_best_loc = ( backpointers[ utterance_in_batch, current_best_loc, time_step - 1 ] .long() .item() ) earlier_z_star = phns[ utterance_in_batch, earlier_best_loc ].item() z_star_i_loc.insert(0, earlier_best_loc) z_star_i.insert(0, earlier_z_star) z_stars.append(z_star_i) z_stars_loc.append(z_star_i_loc) # picking out viterbi_scores viterbi_scores = v_matrix[ torch.arange(batch_size), phn_lens_abs - 1, lens_abs - 1 ] return z_stars, z_stars_loc, viterbi_scores def _loss_reduction(self, loss, input_lens, target_lens): """Applies reduction to loss as specified during object initialization. Arguments --------- loss : torch.Tensor (batch) The loss tensor to be reduced. input_lens : torch.Tensor (batch) The absolute durations of the inputs. target_lens : torch.Tensor (batch) The absolute durations of the targets. Returns ------- loss : torch.Tensor (batch, or scalar) The loss with reduction applied if it is specified. """ if self.input_len_norm is True: loss = torch.div(loss, input_lens) if self.target_len_norm is True: loss = torch.div(loss, target_lens) if self.batch_reduction == "none": pass elif self.batch_reduction == "sum": loss = loss.sum() elif self.batch_reduction == "mean": loss = loss.mean() else: raise ValueError( "`batch_reduction` parameter must be one of 'none', 'sum' or 'mean'" ) return loss def forward( self, emission_pred, lens, phns, phn_lens, dp_algorithm, prob_matrices=None, ): """Prepares relevant (log) probability tensors and does dynamic programming: either the forward or the Viterbi algorithm. Applies reduction as specified during object initialization. Arguments --------- emission_pred : torch.Tensor (batch, time, phoneme in vocabulary) Posterior probabilities from our acoustic model. lens : torch.Tensor (batch) The relative duration of each utterance sound file. phns : torch.Tensor (batch, phoneme in phn sequence) The phonemes that are known/thought to be in each utterance phn_lens : torch.Tensor (batch) The relative length of each phoneme sequence in the batch. dp_algorithm : string Either "forward" or "viterbi". prob_matrices : dict (Optional) Must contain keys 'trans_prob', 'pi_prob' and 'final_states'. Used to override the default forward and viterbi operations which force traversal over all of the states in the `phns` sequence. Returns ------- tensor (1) if dp_algorithm == "forward". ``forward_scores`` : torch.Tensor (batch, or scalar) The (log) likelihood of each utterance in the batch, with reduction applied if specified. (OR) (2) if dp_algorithm == "viterbi". ``viterbi_scores`` : torch.Tensor (batch, or scalar) The (log) likelihood of the Viterbi path for each utterance, with reduction applied if specified. ``alignments`` : list of lists of int Viterbi alignments for the files in the batch. """ lens_abs = torch.round(emission_pred.shape[1] * lens).long() phn_lens_abs = torch.round(phns.shape[1] * phn_lens).long() phns = phns.long() if prob_matrices is None: pi_prob = self._make_pi_prob(phn_lens_abs) trans_prob = self._make_trans_prob(phn_lens_abs) final_states = None else: if ( ("pi_prob" in prob_matrices) and ("trans_prob" in prob_matrices) and ("final_states" in prob_matrices) ): pi_prob = prob_matrices["pi_prob"] trans_prob = prob_matrices["trans_prob"] final_states = prob_matrices["final_states"] else: ValueError( """`prob_matrices` must contain the keys `pi_prob`, `trans_prob` and `final_states`""" ) emiss_pred_useful = self._make_emiss_pred_useful( emission_pred, lens_abs, phn_lens_abs, phns ) if dp_algorithm == "forward": # do forward training forward_scores = self._dp_forward( pi_prob, trans_prob, emiss_pred_useful, lens_abs, phn_lens_abs, phns, ) forward_scores = self._loss_reduction( forward_scores, lens_abs, phn_lens_abs ) return forward_scores elif dp_algorithm == "viterbi": alignments, _, viterbi_scores = self._dp_viterbi( pi_prob, trans_prob, emiss_pred_useful, lens_abs, phn_lens_abs, phns, final_states, ) viterbi_scores = self._loss_reduction( viterbi_scores, lens_abs, phn_lens_abs ) return viterbi_scores, alignments else: raise ValueError( "dp_algorithm input must be either 'forward' or 'viterbi'" ) def expand_phns_by_states_per_phoneme(self, phns, phn_lens): """Expands each phoneme in the phn sequence by the number of hidden states per phoneme defined in the HMM. Arguments --------- phns : torch.Tensor (batch, phoneme in phn sequence) The phonemes that are known/thought to be in each utterance. phn_lens : torch.Tensor (batch) The relative length of each phoneme sequence in the batch. Returns ------- expanded_phns : torch.Tensor (batch, phoneme in expanded phn sequence) Example ------- >>> phns = torch.tensor([[0., 3., 5., 0.], ... [0., 2., 0., 0.]]) >>> phn_lens = torch.tensor([1., 0.75]) >>> aligner = HMMAligner(states_per_phoneme = 3) >>> expanded_phns = aligner.expand_phns_by_states_per_phoneme( ... phns, phn_lens ... ) >>> expanded_phns tensor([[ 0., 1., 2., 9., 10., 11., 15., 16., 17., 0., 1., 2.], [ 0., 1., 2., 6., 7., 8., 0., 1., 2., 0., 0., 0.]]) """ # Initialise expanded_phns expanded_phns = torch.zeros( phns.shape[0], phns.shape[1] * self.states_per_phoneme ) expanded_phns = expanded_phns.to(phns.device) phns = undo_padding(phns, phn_lens) for i, phns_utt in enumerate(phns): expanded_phns_utt = [] for phoneme_index in phns_utt: expanded_phns_utt += [ self.states_per_phoneme * phoneme_index + i_ for i_ in range(self.states_per_phoneme) ] expanded_phns[i, : len(expanded_phns_utt)] = torch.tensor( expanded_phns_utt ) return expanded_phns def store_alignments(self, ids, alignments): """Records Viterbi alignments in `self.align_dict`. Arguments --------- ids : list of str IDs of the files in the batch. alignments : list of lists of int Viterbi alignments for the files in the batch. Without padding. Example ------- >>> aligner = HMMAligner() >>> ids = ['id1', 'id2'] >>> alignments = [[0, 2, 4], [1, 2, 3, 4]] >>> aligner.store_alignments(ids, alignments) >>> aligner.align_dict.keys() dict_keys(['id1', 'id2']) >>> aligner.align_dict['id1'] tensor([0, 2, 4], dtype=torch.int16) """ for i, id in enumerate(ids): alignment_i = alignments[i] alignment_i = torch.tensor(alignment_i, dtype=torch.int16).cpu() self.align_dict[id] = alignment_i def _get_flat_start_batch(self, lens_abs, phn_lens_abs, phns): """Prepares flat start alignments (with zero padding) for every utterance in the batch. Every phoneme will have an equal duration, except for the final phoneme potentially. E.g. if 104 frames and 10 phonemes, 9 phonemes will have duration of 10 frames, and one phoneme will have a duration of 14 frames. Arguments --------- lens_abs : torch.Tensor (batch) The absolute length of each input to the acoustic model, i.e., the number of frames. phn_lens_abs : torch.Tensor (batch) The absolute length of each phoneme sequence in the batch. phns : torch.Tensor (batch, phoneme in phn sequence) The phonemes that are known/thought to be in each utterance. Returns ------- flat_start_batch : torch.Tensor (batch, time) Flat start alignments for utterances in the batch, with zero padding. """ phns = phns.long() batch_size = len(lens_abs) fb_max_length = torch.max(lens_abs) flat_start_batch = torch.zeros( batch_size, fb_max_length, device=phns.device ).long() for i in range(batch_size): utter_phns = phns[i] utter_phns = utter_phns[: phn_lens_abs[i]] # crop out zero padding repeat_amt = int(lens_abs[i].item() / len(utter_phns)) # make sure repeat_amt is at least 1. (the code above # may make repeat_amt==0 if self.states_per_phoneme is too large). if repeat_amt == 0: repeat_amt = 1 # repeat each phoneme in utter_phns by repeat_amt utter_phns = utter_phns.repeat_interleave(repeat_amt) # len(utter_phns) may be <, == or > lens_abs[i], so # make sure len(utter_phns) == lens_abs[i] utter_phns = utter_phns[: lens_abs[i]] utter_phns = torch.nn.functional.pad( utter_phns, (0, int(lens_abs[i]) - len(utter_phns)), value=utter_phns[-1], # pad out with final phoneme ) flat_start_batch[i, : len(utter_phns)] = utter_phns return flat_start_batch def _get_viterbi_batch(self, ids, lens_abs): """Retrieves Viterbi alignments stored in `self.align_dict` and creates a batch of them, with zero padding. Arguments --------- ids : list of str IDs of the files in the batch. lens_abs : torch.Tensor (batch) The absolute length of each input to the acoustic model, i.e., the number of frames. Returns ------- viterbi_batch : torch.Tensor (batch, time) The previously-recorded Viterbi alignments for the utterances in the batch. """ batch_size = len(lens_abs) fb_max_length = torch.max(lens_abs) viterbi_batch = torch.zeros( batch_size, fb_max_length, device=lens_abs.device ).long() for i in range(batch_size): viterbi_preds = self.align_dict[ids[i]] viterbi_preds = torch.nn.functional.pad( viterbi_preds, (0, fb_max_length - len(viterbi_preds)) ) viterbi_batch[i] = viterbi_preds.long() return viterbi_batch def get_prev_alignments(self, ids, emission_pred, lens, phns, phn_lens): """Fetches previously recorded Viterbi alignments if they are available. If not, fetches flat start alignments. Currently, assumes that if a Viterbi alignment is not available for the first utterance in the batch, it will not be available for the rest of the utterances. Arguments --------- ids : list of str IDs of the files in the batch. emission_pred : torch.Tensor (batch, time, phoneme in vocabulary) Posterior probabilities from our acoustic model. Used to infer the duration of the longest utterance in the batch. lens : torch.Tensor (batch) The relative duration of each utterance sound file. phns : torch.Tensor (batch, phoneme in phn sequence) The phonemes that are known/thought to be in each utterance. phn_lens : torch.Tensor (batch) The relative length of each phoneme sequence in the batch. Returns ------- torch.Tensor (batch, time) Zero-padded alignments. Example ------- >>> ids = ['id1', 'id2'] >>> emission_pred = torch.tensor([[[ -1., -10., -10.], ... [-10., -1., -10.], ... [-10., -10., -1.]], ... ... [[ -1., -10., -10.], ... [-10., -1., -10.], ... [-10., -10., -10.]]]) >>> lens = torch.tensor([1., 0.66]) >>> phns = torch.tensor([[0, 1, 2], ... [0, 1, 0]]) >>> phn_lens = torch.tensor([1., 0.66]) >>> aligner = HMMAligner() >>> alignment_batch = aligner.get_prev_alignments( ... ids, emission_pred, lens, phns, phn_lens ... ) >>> alignment_batch tensor([[0, 1, 2], [0, 1, 0]]) """ lens_abs = torch.round(emission_pred.shape[1] * lens).long() phn_lens_abs = torch.round(phns.shape[1] * phn_lens).long() if ids[0] in self.align_dict: return self._get_viterbi_batch(ids, lens_abs) else: return self._get_flat_start_batch(lens_abs, phn_lens_abs, phns) def _calc_accuracy_sent(self, alignments_, ends_, phns_): """Calculates the accuracy between predicted alignments and ground truth alignments for a single sentence/utterance. Arguments --------- alignments_ : list of ints The predicted alignments for the utterance. ends_ : list of ints A list of the sample indices where each ground truth phoneme ends, according to the transcription. phns_ : list of ints The unpadded list of ground truth phonemes in the utterance. Returns ------- mean_acc : float The mean percentage of times that the upsampled predicted alignment matches the ground truth alignment. """ # Create array containing the true alignment at each sample ends_ = [0] + [int(end) for end in ends_] true_durations = [ends_[i] - ends_[i - 1] for i in range(1, len(ends_))] true_alignments = [] for i in range(len(phns_)): true_alignments += [phns_[i]] * (true_durations[i]) true_alignments = torch.tensor(true_alignments) # Upsample the predicted alignment array # and make sure length matches that of `true_alignment` upsample_factor = int( torch.round(torch.tensor(len(true_alignments) / len(alignments_))) ) alignments_ = torch.tensor(alignments_) alignments_upsampled = alignments_.repeat_interleave(upsample_factor) alignments_upsampled = alignments_upsampled[: len(true_alignments)] if len(true_alignments) > len(alignments_upsampled): alignments_upsampled = torch.nn.functional.pad( alignments_upsampled, (0, len(true_alignments) - len(alignments_upsampled)), ) # Measure sample-wise accuracy accuracy = ( alignments_upsampled == true_alignments ).float().mean().item() * 100 return accuracy def calc_accuracy(self, alignments, ends, phns, ind2labs=None): """Calculates mean accuracy between predicted alignments and ground truth alignments. Ground truth alignments are derived from ground truth phns and their ends in the audio sample. Arguments --------- alignments : list of lists of ints/floats The predicted alignments for each utterance in the batch. ends : list of lists of ints A list of lists of sample indices where each ground truth phoneme ends, according to the transcription. Note: current implementation assumes that 'ends' mark the index where the next phoneme begins. phns : list of lists of ints/floats The unpadded list of lists of ground truth phonemes in the batch. ind2labs : tuple (Optional) Contains the original index-to-label dicts for the first and second sequence of phonemes. Returns ------- mean_acc : float The mean percentage of times that the upsampled predicted alignment matches the ground truth alignment. Example ------- >>> aligner = HMMAligner() >>> alignments = [[0., 0., 0., 1.]] >>> phns = [[0., 1.]] >>> ends = [[2, 4]] >>> mean_acc = aligner.calc_accuracy(alignments, ends, phns) >>> mean_acc.item() 75.0 """ acc_hist = [] # Do conversion if states_per_phoneme > 1 if self.states_per_phoneme > 1: alignments = [ [i // self.states_per_phoneme for i in utt] for utt in alignments ] # convert to common alphabet if need be if ind2labs is not None: alignments, phns = map_inds_to_intersect(alignments, phns, ind2labs) for alignments_, ends_, phns_ in zip(alignments, ends, phns): acc = self._calc_accuracy_sent(alignments_, ends_, phns_) acc_hist.append(acc) acc_hist = torch.tensor(acc_hist) mean_acc = acc_hist.mean() return mean_acc.unsqueeze(0) def collapse_alignments(self, alignments): """ Converts alignments to 1 state per phoneme style. Arguments --------- alignments : list of ints Predicted alignments for a single utterance. Returns ------- sequence : list of ints The predicted alignments converted to a 1 state per phoneme style. Example ------- >>> aligner = HMMAligner(states_per_phoneme = 3) >>> alignments = [0, 1, 2, 3, 4, 5, 3, 4, 5, 0, 1, 2] >>> sequence = aligner.collapse_alignments(alignments) >>> sequence [0, 1, 1, 0] """ # Filter the repetitions sequence = [ v for i, v in enumerate(alignments) if i == 0 or v != alignments[i - 1] ] # Pick out only multiples of self.states_per_phoneme sequence = [v for v in sequence if v % self.states_per_phoneme == 0] # Divide by self.states_per_phoneme sequence = [v // self.states_per_phoneme for v in sequence] return sequence @mark_as_saver def _save(self, path): torch.save(self.align_dict, path) @mark_as_loader def _load(self, path, end_of_epoch=False): del end_of_epoch # Not used here. self.align_dict = torch.load(path) def map_inds_to_intersect(lists1, lists2, ind2labs): """Converts 2 lists containing indices for phonemes from different phoneme sets to a single phoneme so that comparing the equality of the indices of the resulting lists will yield the correct accuracy. Arguments --------- lists1 : list of lists of ints Contains the indices of the first sequence of phonemes. lists2 : list of lists of ints Contains the indices of the second sequence of phonemes. ind2labs : tuple (dict, dict) Contains the original index-to-label dicts for the first and second sequence of phonemes. Returns ------- lists1_new : list of lists of ints Contains the indices of the first sequence of phonemes, mapped to the new phoneme set. lists2_new : list of lists of ints Contains the indices of the second sequence of phonemes, mapped to the new phoneme set. Example ------- >>> lists1 = [[0, 1]] >>> lists2 = [[0, 1]] >>> ind2lab1 = { ... 0: "a", ... 1: "b", ... } >>> ind2lab2 = { ... 0: "a", ... 1: "c", ... } >>> ind2labs = (ind2lab1, ind2lab2) >>> out1, out2 = map_inds_to_intersect(lists1, lists2, ind2labs) >>> out1 [[0, 1]] >>> out2 [[0, 2]] """ ind2lab1, ind2lab2 = ind2labs # Form 3 sets: # (1) labs in both mappings # (2) labs in only 1st mapping # (3) labs in only 2nd mapping set1, set2 = set(ind2lab1.values()), set(ind2lab2.values()) intersect = set1.intersection(set2) set1_only = set1.difference(set2) set2_only = set2.difference(set1) new_lab2ind = {lab: i for i, lab in enumerate(intersect)} new_lab2ind.update( {lab: len(new_lab2ind) + i for i, lab in enumerate(set1_only)} ) new_lab2ind.update( {lab: len(new_lab2ind) + i for i, lab in enumerate(set2_only)} ) # Map lists to labels and apply new_lab2ind lists1_lab = [[ind2lab1[ind] for ind in utt] for utt in lists1] lists2_lab = [[ind2lab2[ind] for ind in utt] for utt in lists2] lists1_new = [[new_lab2ind[lab] for lab in utt] for utt in lists1_lab] lists2_new = [[new_lab2ind[lab] for lab in utt] for utt in lists2_lab] return lists1_new, lists2_new def batch_log_matvecmul(A, b): """For each 'matrix' and 'vector' pair in the batch, do matrix-vector multiplication in the log domain, i.e., logsumexp instead of add, add instead of multiply. Arguments --------- A : torch.Tensor (batch, dim1, dim2) Tensor b : torch.Tensor (batch, dim1) Tensor. Returns ------- x : torch.Tensor (batch, dim1) Example ------- >>> A = torch.tensor([[[ 0., 0.], ... [ -1e5, 0.]]]) >>> b = torch.tensor([[0., 0.,]]) >>> x = batch_log_matvecmul(A, b) >>> x tensor([[0.6931, 0.0000]]) >>> >>> # non-log domain equivalent without batching functionality >>> A_ = torch.tensor([[1., 1.], ... [0., 1.]]) >>> b_ = torch.tensor([1., 1.,]) >>> x_ = torch.matmul(A_, b_) >>> x_ tensor([2., 1.]) """ b = b.unsqueeze(1) x = torch.logsumexp(A + b, dim=2) return x def batch_log_maxvecmul(A, b): """Similar to batch_log_matvecmul, but takes a maximum instead of logsumexp. Returns both the max and the argmax. Arguments --------- A : torch.Tensor (batch, dim1, dim2) Tensor. b : torch.Tensor (batch, dim1) Tensor Returns ------- x : torch.Tensor (batch, dim1) Tensor. argmax : torch.Tensor (batch, dim1) Tensor. Example ------- >>> A = torch.tensor([[[ 0., -1.], ... [ -1e5, 0.]]]) >>> b = torch.tensor([[0., 0.,]]) >>> x, argmax = batch_log_maxvecmul(A, b) >>> x tensor([[0., 0.]]) >>> argmax tensor([[0, 1]]) """ b = b.unsqueeze(1) x, argmax = torch.max(A + b, dim=2) return x, argmax