# Copyright (c) 2025, NVIDIA CORPORATION. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import torch import torch.nn.functional as F from torch import nn from nemo.core import NeuralModule from nemo.core.classes import Exportable, NeuralModule, typecheck from nemo.core.neural_types import LabelsType, NeuralType, SpectrogramType class RandomProjectionVectorQuantizer(NeuralModule, Exportable): DIST_FN_LIST = ["l2", "cosine"] def __init__( self, feat_in: int, code_dim: int, num_classes: int, num_books: int, dist_fn: str = "cosine", time_ahead: bool = False, freeze: bool = True, squeeze_single: bool = False, combine_time_steps: int = 1, ): """Vector quantization using random projection proposed in BEST-RQ paper: 'Self-Supervised Learning with Random-Projection Quantizer for Speech Recognition' Args: feat_in: input feature dimension code_dim: dimension of the codebook features num_classes: number of classes num_books: number of codebooks dist_fn: distance function to use, one of "l2" or "cosine" time_ahead: if Ture, the input is of shape (B, T, D), otherwise (B, D, T) freeze: whether to freeze the projection matrix squeeze_single: if True, squeeze codebook dimension if num_books is 1 """ super().__init__() if dist_fn not in self.DIST_FN_LIST: raise ValueError(f"Unknown distance function {dist_fn}, must be one of {self.DIST_FN_LIST}") self.feat_in = feat_in self.code_dim = code_dim self.num_classes = num_classes self.num_books = num_books self.dist_fn = dist_fn self.time_ahead = time_ahead self.squeeze_single = squeeze_single self.combine_time_steps = combine_time_steps # (B, T, D) -> (B, T, num_books, code_dim) self.proj = nn.Linear(self.feat_in * combine_time_steps, self.num_books * self.code_dim, bias=False) torch.nn.init.xavier_normal_(self.proj.weight) # (num_books, num_classes, hid_dim) codebooks = torch.randn(self.num_books, self.num_classes, self.code_dim).double() torch.nn.init.normal_(codebooks, mean=0, std=1) codebooks = F.normalize(codebooks, dim=-1) self.codebooks = nn.Parameter(codebooks) if freeze: self.freeze() @property def input_types(self): """Returns definitions of module input ports.""" if self.time_ahead: return {"input_signal": NeuralType(('B', 'T', 'D'), SpectrogramType())} return {"input_signal": NeuralType(('B', 'D', 'T'), SpectrogramType())} @property def output_types(self): """Returns definitions of module output ports.""" if self.time_ahead: if self.num_books == 1 and self.squeeze_single: return { "xq": NeuralType(('B', 'T', 'D'), SpectrogramType()), "xid": NeuralType(('B', 'T'), LabelsType()), } return { "xq": NeuralType(('B', 'T', 'D', 'H'), SpectrogramType()), "xid": NeuralType(('B', 'T', 'H'), LabelsType()), } if self.num_books == 1 and self.squeeze_single: return { "xq": NeuralType(('B', 'D', 'T'), SpectrogramType()), "xid": NeuralType(('B', 'T'), LabelsType()), } return { "xq": NeuralType(('B', 'D', 'T', 'H'), SpectrogramType()), "xid": NeuralType(('B', 'T', 'H'), LabelsType()), } @typecheck() def forward(self, input_signal): """ Args: input_signal: input features of shape (B, T, D) or (B, D, T) Returns: xq: quantized features of shape (B, T, D, N) or (B, D, T, N) xid: quantized tokens of shape (B, T, N) """ if not self.time_ahead: # (B, D, T) -> (B, T, D) input_signal = input_signal.transpose(1, 2) B, T, _ = input_signal.size() if self.combine_time_steps > 1: input_signal = input_signal.contiguous().reshape(B, T // self.combine_time_steps, -1) T = T // self.combine_time_steps # (B, T, D) -> (B, T, num_books*code_dim) x = self.proj(input_signal) # normalize each feature vector # (B, T, num_books*code_dim) -> (B, T, num_books, code_dim) x = F.normalize(x.view(B, T, self.num_books, self.code_dim), dim=-1) # get tokens (xid) of shape (B, T, num_books) if self.dist_fn == "cosine": # (B, T, num_books, code_dim) -> (B, T, num_books, num_classes) xid = torch.einsum('btdh,dch->btdc', x, self.codebooks) # (B, T, num_books, num_classes) -> (B, T, num_books) xid = xid.max(dim=-1)[1] elif self.dist_fn == "l2": # (B, T, num_books, code_dim) -> (B, T, num_books, code_dim, num_classes) xid = x.unsqueeze(-1) - self.codebooks.transpose(1, 2).unsqueeze(0).unsqueeze(0) xid = xid.norm(dim=-2).argmin(dim=-1) else: raise ValueError(f"Unknown distance function {self.dist_fn}, must be one of {self.DIST_FN_LIST}") # xid2: (B, T, num_books) -> (B, T, num_books) xid2 = xid + self.num_classes * torch.arange(self.num_books, device=xid.device).unsqueeze(0).unsqueeze(0) # xid2: (B, T, num_books) -> (B*num_books, T) xid2 = xid2.transpose(1, 2).contiguous().view(-1, T) # get quantized vector (xq) of shape (B, T, code_dim, num_books) # codebook: (num_books, num_classes, code_dim) -> (num_books*num_classes, code_dim) xq = F.embedding(xid2.view(-1), self.codebooks.view(-1, self.code_dim)).view( B, T, self.code_dim, self.num_books ) if not self.time_ahead: # (B, T, D) -> (B, D, T) xq = xq.transpose(1, 2) if self.num_books == 1 and self.squeeze_single: xq = xq.squeeze(-1) xid = xid.squeeze(-1) return xq, xid