From 741dcd1d6d49fb69ba5ec6f4a55a0c002144e57d Mon Sep 17 00:00:00 2001 From: Daniel Povey Date: Tue, 31 May 2022 12:45:28 +0800 Subject: [PATCH] Move pruned_transducer_stateless4 to pruned_transducer_stateless7 --- .../pruned_transducer_stateless7/__init__.py | 1 + .../asr_datamodule.py | 1 + .../beam_search.py | 1 + .../pruned_transducer_stateless7/conformer.py | 1036 ++++++++++++++++ .../pruned_transducer_stateless7/decode.py | 633 ++++++++++ .../pruned_transducer_stateless7/decoder.py | 102 ++ .../encoder_interface.py | 1 + .../pruned_transducer_stateless7/export.py | 1 + .../pruned_transducer_stateless7/joiner.py | 66 ++ .../ASR/pruned_transducer_stateless7/model.py | 192 +++ .../ASR/pruned_transducer_stateless7/optim.py | 802 +++++++++++++ .../pruned_transducer_stateless7/scaling.py | 415 +++++++ .../test_model.py | 50 + .../ASR/pruned_transducer_stateless7/train.py | 1050 +++++++++++++++++ 14 files changed, 4351 insertions(+) create mode 120000 egs/librispeech/ASR/pruned_transducer_stateless7/__init__.py create mode 120000 egs/librispeech/ASR/pruned_transducer_stateless7/asr_datamodule.py create mode 120000 egs/librispeech/ASR/pruned_transducer_stateless7/beam_search.py create mode 100644 egs/librispeech/ASR/pruned_transducer_stateless7/conformer.py create mode 100755 egs/librispeech/ASR/pruned_transducer_stateless7/decode.py create mode 100644 egs/librispeech/ASR/pruned_transducer_stateless7/decoder.py create mode 120000 egs/librispeech/ASR/pruned_transducer_stateless7/encoder_interface.py create mode 120000 egs/librispeech/ASR/pruned_transducer_stateless7/export.py create mode 100644 egs/librispeech/ASR/pruned_transducer_stateless7/joiner.py create mode 100644 egs/librispeech/ASR/pruned_transducer_stateless7/model.py create mode 100644 egs/librispeech/ASR/pruned_transducer_stateless7/optim.py create mode 100644 egs/librispeech/ASR/pruned_transducer_stateless7/scaling.py create mode 100755 egs/librispeech/ASR/pruned_transducer_stateless7/test_model.py create mode 100755 egs/librispeech/ASR/pruned_transducer_stateless7/train.py diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7/__init__.py b/egs/librispeech/ASR/pruned_transducer_stateless7/__init__.py new file mode 120000 index 000000000..b24e5e357 --- /dev/null +++ b/egs/librispeech/ASR/pruned_transducer_stateless7/__init__.py @@ -0,0 +1 @@ +../pruned_transducer_stateless2/__init__.py \ No newline at end of file diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7/asr_datamodule.py b/egs/librispeech/ASR/pruned_transducer_stateless7/asr_datamodule.py new file mode 120000 index 000000000..a074d6085 --- /dev/null +++ b/egs/librispeech/ASR/pruned_transducer_stateless7/asr_datamodule.py @@ -0,0 +1 @@ +../pruned_transducer_stateless2/asr_datamodule.py \ No newline at end of file diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7/beam_search.py b/egs/librispeech/ASR/pruned_transducer_stateless7/beam_search.py new file mode 120000 index 000000000..8554e44cc --- /dev/null +++ b/egs/librispeech/ASR/pruned_transducer_stateless7/beam_search.py @@ -0,0 +1 @@ +../pruned_transducer_stateless2/beam_search.py \ No newline at end of file diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7/conformer.py b/egs/librispeech/ASR/pruned_transducer_stateless7/conformer.py new file mode 100644 index 000000000..99a6f7ce5 --- /dev/null +++ b/egs/librispeech/ASR/pruned_transducer_stateless7/conformer.py @@ -0,0 +1,1036 @@ +#!/usr/bin/env python3 +# Copyright (c) 2021 University of Chinese Academy of Sciences (author: Han Zhu) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# 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 copy +import math +import warnings +from typing import Optional, Tuple + +import torch +from encoder_interface import EncoderInterface +from scaling import ( + ActivationBalancer, + BasicNorm, + DoubleSwish, + ScaledConv1d, + ScaledConv2d, + ScaledLinear, +) +from torch import Tensor, nn + +from icefall.utils import make_pad_mask + + +class Conformer(EncoderInterface): + """ + Args: + num_features (int): Number of input features + subsampling_factor (int): subsampling factor of encoder (the convolution layers before transformers) + d_model (int): attention dimension, also the output dimension + nhead (int): number of head + dim_feedforward (int): feedforward dimention + num_encoder_layers (int): number of encoder layers + dropout (float): dropout rate + layer_dropout (float): layer-dropout rate. + cnn_module_kernel (int): Kernel size of convolution module + vgg_frontend (bool): whether to use vgg frontend. + """ + + def __init__( + self, + num_features: int, + subsampling_factor: int = 4, + d_model: int = 256, + nhead: int = 4, + dim_feedforward: int = 2048, + num_encoder_layers: int = 12, + dropout: float = 0.1, + layer_dropout: float = 0.075, + cnn_module_kernel: int = 31, + ) -> None: + super(Conformer, self).__init__() + + self.num_features = num_features + self.subsampling_factor = subsampling_factor + if subsampling_factor != 4: + raise NotImplementedError("Support only 'subsampling_factor=4'.") + + # self.encoder_embed converts the input of shape (N, T, num_features) + # to the shape (N, T//subsampling_factor, d_model). + # That is, it does two things simultaneously: + # (1) subsampling: T -> T//subsampling_factor + # (2) embedding: num_features -> d_model + self.encoder_embed = Conv2dSubsampling(num_features, d_model) + + self.encoder_pos = RelPositionalEncoding(d_model, dropout) + + encoder_layer = ConformerEncoderLayer( + d_model, + nhead, + dim_feedforward, + dropout, + layer_dropout, + cnn_module_kernel, + ) + self.encoder = ConformerEncoder(encoder_layer, num_encoder_layers) + + def forward( + self, x: torch.Tensor, x_lens: torch.Tensor, warmup: float = 1.0 + ) -> Tuple[torch.Tensor, torch.Tensor]: + """ + Args: + x: + The input tensor. Its shape is (batch_size, seq_len, feature_dim). + x_lens: + A tensor of shape (batch_size,) containing the number of frames in + `x` before padding. + warmup: + A floating point value that gradually increases from 0 throughout + training; when it is >= 1.0 we are "fully warmed up". It is used + to turn modules on sequentially. + Returns: + Return a tuple containing 2 tensors: + - embeddings: its shape is (batch_size, output_seq_len, d_model) + - lengths, a tensor of shape (batch_size,) containing the number + of frames in `embeddings` before padding. + """ + x = self.encoder_embed(x) + x, pos_emb = self.encoder_pos(x) + x = x.permute(1, 0, 2) # (N, T, C) -> (T, N, C) + + # Caution: We assume the subsampling factor is 4! + + # lengths = ((x_lens - 1) // 2 - 1) // 2 # issue an warning + # + # Note: rounding_mode in torch.div() is available only in torch >= 1.8.0 + lengths = (((x_lens - 1) >> 1) - 1) >> 1 + + assert x.size(0) == lengths.max().item() + mask = make_pad_mask(lengths) + + x = self.encoder( + x, pos_emb, src_key_padding_mask=mask, warmup=warmup + ) # (T, N, C) + + x = x.permute(1, 0, 2) # (T, N, C) ->(N, T, C) + + return x, lengths + + +class ConformerEncoderLayer(nn.Module): + """ + ConformerEncoderLayer is made up of self-attn, feedforward and convolution networks. + See: "Conformer: Convolution-augmented Transformer for Speech Recognition" + + Args: + d_model: the number of expected features in the input (required). + nhead: the number of heads in the multiheadattention models (required). + dim_feedforward: the dimension of the feedforward network model (default=2048). + dropout: the dropout value (default=0.1). + cnn_module_kernel (int): Kernel size of convolution module. + + Examples:: + >>> encoder_layer = ConformerEncoderLayer(d_model=512, nhead=8) + >>> src = torch.rand(10, 32, 512) + >>> pos_emb = torch.rand(32, 19, 512) + >>> out = encoder_layer(src, pos_emb) + """ + + def __init__( + self, + d_model: int, + nhead: int, + dim_feedforward: int = 2048, + dropout: float = 0.1, + layer_dropout: float = 0.075, + cnn_module_kernel: int = 31, + ) -> None: + super(ConformerEncoderLayer, self).__init__() + + self.layer_dropout = layer_dropout + + self.d_model = d_model + + self.self_attn = RelPositionMultiheadAttention( + d_model, nhead, dropout=0.0 + ) + + self.feed_forward = nn.Sequential( + nn.Linear(d_model, dim_feedforward), + ActivationBalancer(channel_dim=-1), + DoubleSwish(), + nn.Dropout(dropout), + ScaledLinear(dim_feedforward, d_model, + initial_scale=0.1), + ) + + self.feed_forward_macaron = nn.Sequential( + nn.Linear(d_model, dim_feedforward), + ActivationBalancer(channel_dim=-1), + DoubleSwish(), + nn.Dropout(dropout), + ScaledLinear(dim_feedforward, d_model, + initial_scale=0.1), + ) + + self.conv_module = ConvolutionModule(d_model, + cnn_module_kernel) + + self.norm_final = BasicNorm(d_model) + + # try to ensure the output is close to zero-mean (or at least, zero-median). + self.balancer = ActivationBalancer( + channel_dim=-1, min_positive=0.45, max_positive=0.55, max_abs=6.0 + ) + + self.dropout = nn.Dropout(dropout) + + def forward( + self, + src: Tensor, + pos_emb: Tensor, + src_mask: Optional[Tensor] = None, + src_key_padding_mask: Optional[Tensor] = None, + warmup: float = 1.0, + ) -> Tensor: + """ + Pass the input through the encoder layer. + + Args: + src: the sequence to the encoder layer (required). + pos_emb: Positional embedding tensor (required). + src_mask: the mask for the src sequence (optional). + src_key_padding_mask: the mask for the src keys per batch (optional). + warmup: controls selective bypass of of layers; if < 1.0, we will + bypass layers more frequently. + + Shape: + src: (S, N, E). + pos_emb: (N, 2*S-1, E) + src_mask: (S, S). + src_key_padding_mask: (N, S). + S is the source sequence length, N is the batch size, E is the feature number + """ + src_orig = src + + warmup_scale = min(0.1 + warmup, 1.0) + # alpha = 1.0 means fully use this encoder layer, 0.0 would mean + # completely bypass it. + if self.training: + alpha = ( + warmup_scale + if torch.rand(()).item() <= (1.0 - self.layer_dropout) + else 0.1 + ) + else: + alpha = 1.0 + + # macaron style feed forward module + src = src + self.dropout(self.feed_forward_macaron(src)) + + # multi-headed self-attention module + src_att = self.self_attn( + src, + src, + src, + pos_emb=pos_emb, + attn_mask=src_mask, + key_padding_mask=src_key_padding_mask, + )[0] + src = src + self.dropout(src_att) + + # convolution module + src = src + self.dropout(self.conv_module(src)) + + # feed forward module + src = src + self.dropout(self.feed_forward(src)) + + src = self.norm_final(self.balancer(src)) + + if alpha != 1.0: + src = alpha * src + (1 - alpha) * src_orig + + return src + + +class ConformerEncoder(nn.Module): + r"""ConformerEncoder is a stack of N encoder layers + + Args: + encoder_layer: an instance of the ConformerEncoderLayer() class (required). + num_layers: the number of sub-encoder-layers in the encoder (required). + + Examples:: + >>> encoder_layer = ConformerEncoderLayer(d_model=512, nhead=8) + >>> conformer_encoder = ConformerEncoder(encoder_layer, num_layers=6) + >>> src = torch.rand(10, 32, 512) + >>> pos_emb = torch.rand(32, 19, 512) + >>> out = conformer_encoder(src, pos_emb) + """ + + def __init__(self, encoder_layer: nn.Module, num_layers: int) -> None: + super().__init__() + self.layers = nn.ModuleList( + [copy.deepcopy(encoder_layer) for i in range(num_layers)] + ) + self.num_layers = num_layers + + def forward( + self, + src: Tensor, + pos_emb: Tensor, + mask: Optional[Tensor] = None, + src_key_padding_mask: Optional[Tensor] = None, + warmup: float = 1.0, + ) -> Tensor: + r"""Pass the input through the encoder layers in turn. + + Args: + src: the sequence to the encoder (required). + pos_emb: Positional embedding tensor (required). + mask: the mask for the src sequence (optional). + src_key_padding_mask: the mask for the src keys per batch (optional). + + Shape: + src: (S, N, E). + pos_emb: (N, 2*S-1, E) + mask: (S, S). + src_key_padding_mask: (N, S). + S is the source sequence length, T is the target sequence length, N is the batch size, E is the feature number + + """ + output = src + + for i, mod in enumerate(self.layers): + output = mod( + output, + pos_emb, + src_mask=mask, + src_key_padding_mask=src_key_padding_mask, + warmup=warmup, + ) + + return output + + +class RelPositionalEncoding(torch.nn.Module): + """Relative positional encoding module. + + See : Appendix B in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context" + Modified from https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/embedding.py + + Args: + d_model: Embedding dimension. + dropout_rate: Dropout rate. + max_len: Maximum input length. + + """ + + def __init__( + self, d_model: int, dropout_rate: float, max_len: int = 5000 + ) -> None: + """Construct an PositionalEncoding object.""" + super(RelPositionalEncoding, self).__init__() + self.d_model = d_model + self.dropout = torch.nn.Dropout(p=dropout_rate) + self.pe = None + self.extend_pe(torch.tensor(0.0).expand(1, max_len)) + + def extend_pe(self, x: Tensor) -> None: + """Reset the positional encodings.""" + if self.pe is not None: + # self.pe contains both positive and negative parts + # the length of self.pe is 2 * input_len - 1 + if self.pe.size(1) >= x.size(1) * 2 - 1: + # Note: TorchScript doesn't implement operator== for torch.Device + if self.pe.dtype != x.dtype or str(self.pe.device) != str( + x.device + ): + self.pe = self.pe.to(dtype=x.dtype, device=x.device) + return + # Suppose `i` means to the position of query vecotr and `j` means the + # position of key vector. We use position relative positions when keys + # are to the left (i>j) and negative relative positions otherwise (i Tuple[Tensor, Tensor]: + """Add positional encoding. + + Args: + x (torch.Tensor): Input tensor (batch, time, `*`). + + Returns: + torch.Tensor: Encoded tensor (batch, time, `*`). + torch.Tensor: Encoded tensor (batch, 2*time-1, `*`). + + """ + self.extend_pe(x) + pos_emb = self.pe[ + :, + self.pe.size(1) // 2 + - x.size(1) + + 1 : self.pe.size(1) // 2 # noqa E203 + + x.size(1), + ] + return self.dropout(x), self.dropout(pos_emb) + + +class RelPositionMultiheadAttention(nn.Module): + r"""Multi-Head Attention layer with relative position encoding + + See reference: "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context" + + Args: + embed_dim: total dimension of the model. + num_heads: parallel attention heads. + dropout: a Dropout layer on attn_output_weights. Default: 0.0. + + Examples:: + + >>> rel_pos_multihead_attn = RelPositionMultiheadAttention(embed_dim, num_heads) + >>> attn_output, attn_output_weights = multihead_attn(query, key, value, pos_emb) + """ + + def __init__( + self, + embed_dim: int, + num_heads: int, + dropout: float = 0.0, + ) -> None: + super(RelPositionMultiheadAttention, self).__init__() + self.embed_dim = embed_dim + self.num_heads = num_heads + self.dropout = dropout + self.head_dim = embed_dim // num_heads + assert ( + self.head_dim * num_heads == self.embed_dim + ), "embed_dim must be divisible by num_heads" + + self.in_proj = nn.Linear(embed_dim, 3 * embed_dim, bias=True) + self.out_proj = ScaledLinear( + embed_dim, embed_dim, bias=True, initial_scale=0.5 + ) + + # linear transformation for positional encoding. + self.linear_pos = nn.Linear(embed_dim, embed_dim, bias=False) + # these two learnable bias are used in matrix c and matrix d + # as described in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context" Section 3.3 + self.pos_bias_u = nn.Parameter(torch.Tensor(num_heads, self.head_dim)) + self.pos_bias_v = nn.Parameter(torch.Tensor(num_heads, self.head_dim)) + self._reset_parameters() + + def _reset_parameters(self) -> None: + nn.init.uniform_(self.pos_bias_u, -0.05, 0.05) + nn.init.uniform_(self.pos_bias_v, -0.05, 0.05) + + def forward( + self, + query: Tensor, + key: Tensor, + value: Tensor, + pos_emb: Tensor, + key_padding_mask: Optional[Tensor] = None, + need_weights: bool = True, + attn_mask: Optional[Tensor] = None, + ) -> Tuple[Tensor, Optional[Tensor]]: + r""" + Args: + query, key, value: map a query and a set of key-value pairs to an output. + pos_emb: Positional embedding tensor + key_padding_mask: if provided, specified padding elements in the key will + be ignored by the attention. When given a binary mask and a value is True, + the corresponding value on the attention layer will be ignored. When given + a byte mask and a value is non-zero, the corresponding value on the attention + layer will be ignored + need_weights: output attn_output_weights. + attn_mask: 2D or 3D mask that prevents attention to certain positions. A 2D mask will be broadcasted for all + the batches while a 3D mask allows to specify a different mask for the entries of each batch. + + Shape: + - Inputs: + - query: :math:`(L, N, E)` where L is the target sequence length, N is the batch size, E is + the embedding dimension. + - key: :math:`(S, N, E)`, where S is the source sequence length, N is the batch size, E is + the embedding dimension. + - value: :math:`(S, N, E)` where S is the source sequence length, N is the batch size, E is + the embedding dimension. + - pos_emb: :math:`(N, 2*L-1, E)` where L is the target sequence length, N is the batch size, E is + the embedding dimension. + - key_padding_mask: :math:`(N, S)` where N is the batch size, S is the source sequence length. + If a ByteTensor is provided, the non-zero positions will be ignored while the position + with the zero positions will be unchanged. If a BoolTensor is provided, the positions with the + value of ``True`` will be ignored while the position with the value of ``False`` will be unchanged. + - attn_mask: 2D mask :math:`(L, S)` where L is the target sequence length, S is the source sequence length. + 3D mask :math:`(N*num_heads, L, S)` where N is the batch size, L is the target sequence length, + S is the source sequence length. attn_mask ensure that position i is allowed to attend the unmasked + positions. If a ByteTensor is provided, the non-zero positions are not allowed to attend + while the zero positions will be unchanged. If a BoolTensor is provided, positions with ``True`` + is not allowed to attend while ``False`` values will be unchanged. If a FloatTensor + is provided, it will be added to the attention weight. + + - Outputs: + - attn_output: :math:`(L, N, E)` where L is the target sequence length, N is the batch size, + E is the embedding dimension. + - attn_output_weights: :math:`(N, L, S)` where N is the batch size, + L is the target sequence length, S is the source sequence length. + """ + return self.multi_head_attention_forward( + query, + key, + value, + pos_emb, + self.embed_dim, + self.num_heads, + self.in_proj.weight, + self.in_proj.bias, + self.dropout, + self.out_proj.weight, + self.out_proj.bias, + training=self.training, + key_padding_mask=key_padding_mask, + need_weights=need_weights, + attn_mask=attn_mask, + ) + + def rel_shift(self, x: Tensor) -> Tensor: + """Compute relative positional encoding. + + Args: + x: Input tensor (batch, head, time1, 2*time1-1). + time1 means the length of query vector. + + Returns: + Tensor: tensor of shape (batch, head, time1, time2) + (note: time2 has the same value as time1, but it is for + the key, while time1 is for the query). + """ + (batch_size, num_heads, time1, n) = x.shape + assert n == 2 * time1 - 1 + # Note: TorchScript requires explicit arg for stride() + batch_stride = x.stride(0) + head_stride = x.stride(1) + time1_stride = x.stride(2) + n_stride = x.stride(3) + return x.as_strided( + (batch_size, num_heads, time1, time1), + (batch_stride, head_stride, time1_stride - n_stride, n_stride), + storage_offset=n_stride * (time1 - 1), + ) + + def multi_head_attention_forward( + self, + query: Tensor, + key: Tensor, + value: Tensor, + pos_emb: Tensor, + embed_dim_to_check: int, + num_heads: int, + in_proj_weight: Tensor, + in_proj_bias: Tensor, + dropout_p: float, + out_proj_weight: Tensor, + out_proj_bias: Tensor, + training: bool = True, + key_padding_mask: Optional[Tensor] = None, + need_weights: bool = True, + attn_mask: Optional[Tensor] = None, + ) -> Tuple[Tensor, Optional[Tensor]]: + r""" + Args: + query, key, value: map a query and a set of key-value pairs to an output. + pos_emb: Positional embedding tensor + embed_dim_to_check: total dimension of the model. + num_heads: parallel attention heads. + in_proj_weight, in_proj_bias: input projection weight and bias. + dropout_p: probability of an element to be zeroed. + out_proj_weight, out_proj_bias: the output projection weight and bias. + training: apply dropout if is ``True``. + key_padding_mask: if provided, specified padding elements in the key will + be ignored by the attention. This is an binary mask. When the value is True, + the corresponding value on the attention layer will be filled with -inf. + need_weights: output attn_output_weights. + attn_mask: 2D or 3D mask that prevents attention to certain positions. A 2D mask will be broadcasted for all + the batches while a 3D mask allows to specify a different mask for the entries of each batch. + + Shape: + Inputs: + - query: :math:`(L, N, E)` where L is the target sequence length, N is the batch size, E is + the embedding dimension. + - key: :math:`(S, N, E)`, where S is the source sequence length, N is the batch size, E is + the embedding dimension. + - value: :math:`(S, N, E)` where S is the source sequence length, N is the batch size, E is + the embedding dimension. + - pos_emb: :math:`(N, 2*L-1, E)` or :math:`(1, 2*L-1, E)` where L is the target sequence + length, N is the batch size, E is the embedding dimension. + - key_padding_mask: :math:`(N, S)` where N is the batch size, S is the source sequence length. + If a ByteTensor is provided, the non-zero positions will be ignored while the zero positions + will be unchanged. If a BoolTensor is provided, the positions with the + value of ``True`` will be ignored while the position with the value of ``False`` will be unchanged. + - attn_mask: 2D mask :math:`(L, S)` where L is the target sequence length, S is the source sequence length. + 3D mask :math:`(N*num_heads, L, S)` where N is the batch size, L is the target sequence length, + S is the source sequence length. attn_mask ensures that position i is allowed to attend the unmasked + positions. If a ByteTensor is provided, the non-zero positions are not allowed to attend + while the zero positions will be unchanged. If a BoolTensor is provided, positions with ``True`` + are not allowed to attend while ``False`` values will be unchanged. If a FloatTensor + is provided, it will be added to the attention weight. + + Outputs: + - attn_output: :math:`(L, N, E)` where L is the target sequence length, N is the batch size, + E is the embedding dimension. + - attn_output_weights: :math:`(N, L, S)` where N is the batch size, + L is the target sequence length, S is the source sequence length. + """ + + tgt_len, bsz, embed_dim = query.size() + assert embed_dim == embed_dim_to_check + assert key.size(0) == value.size(0) and key.size(1) == value.size(1) + + head_dim = embed_dim // num_heads + assert ( + head_dim * num_heads == embed_dim + ), "embed_dim must be divisible by num_heads" + + scaling = float(head_dim) ** -0.5 + + if torch.equal(query, key) and torch.equal(key, value): + # self-attention + q, k, v = nn.functional.linear( + query, in_proj_weight, in_proj_bias + ).chunk(3, dim=-1) + + elif torch.equal(key, value): + # encoder-decoder attention + # This is inline in_proj function with in_proj_weight and in_proj_bias + _b = in_proj_bias + _start = 0 + _end = embed_dim + _w = in_proj_weight[_start:_end, :] + if _b is not None: + _b = _b[_start:_end] + q = nn.functional.linear(query, _w, _b) + + # This is inline in_proj function with in_proj_weight and in_proj_bias + _b = in_proj_bias + _start = embed_dim + _end = None + _w = in_proj_weight[_start:, :] + if _b is not None: + _b = _b[_start:] + k, v = nn.functional.linear(key, _w, _b).chunk(2, dim=-1) + + else: + # This is inline in_proj function with in_proj_weight and in_proj_bias + _b = in_proj_bias + _start = 0 + _end = embed_dim + _w = in_proj_weight[_start:_end, :] + if _b is not None: + _b = _b[_start:_end] + q = nn.functional.linear(query, _w, _b) + + # This is inline in_proj function with in_proj_weight and in_proj_bias + _b = in_proj_bias + _start = embed_dim + _end = embed_dim * 2 + _w = in_proj_weight[_start:_end, :] + if _b is not None: + _b = _b[_start:_end] + k = nn.functional.linear(key, _w, _b) + + # This is inline in_proj function with in_proj_weight and in_proj_bias + _b = in_proj_bias + _start = embed_dim * 2 + _end = None + _w = in_proj_weight[_start:, :] + if _b is not None: + _b = _b[_start:] + v = nn.functional.linear(value, _w, _b) + + if attn_mask is not None: + assert ( + attn_mask.dtype == torch.float32 + or attn_mask.dtype == torch.float64 + or attn_mask.dtype == torch.float16 + or attn_mask.dtype == torch.uint8 + or attn_mask.dtype == torch.bool + ), "Only float, byte, and bool types are supported for attn_mask, not {}".format( + attn_mask.dtype + ) + if attn_mask.dtype == torch.uint8: + warnings.warn( + "Byte tensor for attn_mask is deprecated. Use bool tensor instead." + ) + attn_mask = attn_mask.to(torch.bool) + + if attn_mask.dim() == 2: + attn_mask = attn_mask.unsqueeze(0) + if list(attn_mask.size()) != [1, query.size(0), key.size(0)]: + raise RuntimeError( + "The size of the 2D attn_mask is not correct." + ) + elif attn_mask.dim() == 3: + if list(attn_mask.size()) != [ + bsz * num_heads, + query.size(0), + key.size(0), + ]: + raise RuntimeError( + "The size of the 3D attn_mask is not correct." + ) + else: + raise RuntimeError( + "attn_mask's dimension {} is not supported".format( + attn_mask.dim() + ) + ) + # attn_mask's dim is 3 now. + + # convert ByteTensor key_padding_mask to bool + if ( + key_padding_mask is not None + and key_padding_mask.dtype == torch.uint8 + ): + warnings.warn( + "Byte tensor for key_padding_mask is deprecated. Use bool tensor instead." + ) + key_padding_mask = key_padding_mask.to(torch.bool) + + q = (q * scaling).contiguous().view(tgt_len, bsz, num_heads, head_dim) + k = k.contiguous().view(-1, bsz, num_heads, head_dim) + v = v.contiguous().view(-1, bsz * num_heads, head_dim).transpose(0, 1) + + src_len = k.size(0) + + if key_padding_mask is not None: + assert key_padding_mask.size(0) == bsz, "{} == {}".format( + key_padding_mask.size(0), bsz + ) + assert key_padding_mask.size(1) == src_len, "{} == {}".format( + key_padding_mask.size(1), src_len + ) + + q = q.transpose(0, 1) # (batch, time1, head, d_k) + + pos_emb_bsz = pos_emb.size(0) + assert pos_emb_bsz in (1, bsz) # actually it is 1 + p = self.linear_pos(pos_emb).view(pos_emb_bsz, -1, num_heads, head_dim) + p = p.transpose(1, 2) # (batch, head, 2*time1-1, d_k) + + q_with_bias_u = (q + self.pos_bias_u).transpose( + 1, 2 + ) # (batch, head, time1, d_k) + + q_with_bias_v = (q + self.pos_bias_v).transpose( + 1, 2 + ) # (batch, head, time1, d_k) + + # compute attention score + # first compute matrix a and matrix c + # as described in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context" Section 3.3 + k = k.permute(1, 2, 3, 0) # (batch, head, d_k, time2) + matrix_ac = torch.matmul( + q_with_bias_u, k + ) # (batch, head, time1, time2) + + # compute matrix b and matrix d + matrix_bd = torch.matmul( + q_with_bias_v, p.transpose(-2, -1) + ) # (batch, head, time1, 2*time1-1) + matrix_bd = self.rel_shift(matrix_bd) + + attn_output_weights = ( + matrix_ac + matrix_bd + ) # (batch, head, time1, time2) + + attn_output_weights = attn_output_weights.view( + bsz * num_heads, tgt_len, -1 + ) + + assert list(attn_output_weights.size()) == [ + bsz * num_heads, + tgt_len, + src_len, + ] + + if attn_mask is not None: + if attn_mask.dtype == torch.bool: + attn_output_weights.masked_fill_(attn_mask, float("-inf")) + else: + attn_output_weights += attn_mask + + if key_padding_mask is not None: + attn_output_weights = attn_output_weights.view( + bsz, num_heads, tgt_len, src_len + ) + attn_output_weights = attn_output_weights.masked_fill( + key_padding_mask.unsqueeze(1).unsqueeze(2), + float("-inf"), + ) + attn_output_weights = attn_output_weights.view( + bsz * num_heads, tgt_len, src_len + ) + + attn_output_weights = nn.functional.softmax(attn_output_weights, dim=-1) + attn_output_weights = nn.functional.dropout( + attn_output_weights, p=dropout_p, training=training + ) + + attn_output = torch.bmm(attn_output_weights, v) + assert list(attn_output.size()) == [bsz * num_heads, tgt_len, head_dim] + attn_output = ( + attn_output.transpose(0, 1) + .contiguous() + .view(tgt_len, bsz, embed_dim) + ) + attn_output = nn.functional.linear( + attn_output, out_proj_weight, out_proj_bias + ) + + if need_weights: + # average attention weights over heads + attn_output_weights = attn_output_weights.view( + bsz, num_heads, tgt_len, src_len + ) + return attn_output, attn_output_weights.sum(dim=1) / num_heads + else: + return attn_output, None + + +class ConvolutionModule(nn.Module): + """ConvolutionModule in Conformer model. + Modified from https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/conformer/convolution.py + + Args: + channels (int): The number of channels of conv layers. + kernel_size (int): Kernerl size of conv layers. + bias (bool): Whether to use bias in conv layers (default=True). + + """ + + def __init__( + self, channels: int, kernel_size: int, bias: bool = True + ) -> None: + """Construct an ConvolutionModule object.""" + super(ConvolutionModule, self).__init__() + # kernerl_size should be a odd number for 'SAME' padding + assert (kernel_size - 1) % 2 == 0 + + self.pointwise_conv1 = nn.Conv1d( + channels, + 2 * channels, + kernel_size=1, + stride=1, + padding=0, + bias=bias, + ) + + # after pointwise_conv1 we put x through a gated linear unit (nn.functional.glu). + # For most layers the normal rms value of channels of x seems to be in the range 1 to 4, + # but sometimes, for some reason, for layer 0 the rms ends up being very large, + # between 50 and 100 for different channels. This will cause very peaky and + # sparse derivatives for the sigmoid gating function, which will tend to make + # the loss function not learn effectively. (for most layers the average absolute values + # are in the range 0.5..9.0, and the average p(x>0), i.e. positive proportion, + # at the output of pointwise_conv1.output is around 0.35 to 0.45 for different + # layers, which likely breaks down as 0.5 for the "linear" half and + # 0.2 to 0.3 for the part that goes into the sigmoid. The idea is that if we + # constrain the rms values to a reasonable range via a constraint of max_abs=10.0, + # it will be in a better position to start learning something, i.e. to latch onto + # the correct range. + self.deriv_balancer1 = ActivationBalancer( + channel_dim=1, max_abs=10.0, min_positive=0.05, max_positive=1.0 + ) + + self.depthwise_conv = nn.Conv1d( + channels, + channels, + kernel_size, + stride=1, + padding=(kernel_size - 1) // 2, + groups=channels, + bias=bias, + ) + + self.deriv_balancer2 = ActivationBalancer( + channel_dim=1, min_positive=0.05, max_positive=1.0 + ) + + self.activation = DoubleSwish() + + self.pointwise_conv2 = ScaledConv1d( + channels, + channels, + kernel_size=1, + stride=1, + padding=0, + bias=bias, + initial_scale=0.5, + ) + + def forward(self, x: Tensor) -> Tensor: + """Compute convolution module. + + Args: + x: Input tensor (#time, batch, channels). + + Returns: + Tensor: Output tensor (#time, batch, channels). + + """ + # exchange the temporal dimension and the feature dimension + x = x.permute(1, 2, 0) # (#batch, channels, time). + + # GLU mechanism + x = self.pointwise_conv1(x) # (batch, 2*channels, time) + + x = self.deriv_balancer1(x) + x = nn.functional.glu(x, dim=1) # (batch, channels, time) + + # 1D Depthwise Conv + x = self.depthwise_conv(x) + + x = self.deriv_balancer2(x) + x = self.activation(x) + + x = self.pointwise_conv2(x) # (batch, channel, time) + + return x.permute(2, 0, 1) + + +class Conv2dSubsampling(nn.Module): + """Convolutional 2D subsampling (to 1/4 length). + + Convert an input of shape (N, T, idim) to an output + with shape (N, T', odim), where + T' = ((T-1)//2 - 1)//2, which approximates T' == T//4 + + It is based on + https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/subsampling.py # noqa + """ + + def __init__( + self, + in_channels: int, + out_channels: int, + layer1_channels: int = 8, + layer2_channels: int = 32, + layer3_channels: int = 128, + ) -> None: + """ + Args: + in_channels: + Number of channels in. The input shape is (N, T, in_channels). + Caution: It requires: T >=7, in_channels >=7 + out_channels + Output dim. The output shape is (N, ((T-1)//2 - 1)//2, out_channels) + layer1_channels: + Number of channels in layer1 + layer1_channels: + Number of channels in layer2 + """ + assert in_channels >= 7 + super().__init__() + + self.conv = nn.Sequential( + nn.Conv2d( + in_channels=1, + out_channels=layer1_channels, + kernel_size=3, + padding=1, + ), + ActivationBalancer(channel_dim=1), + DoubleSwish(), + nn.Conv2d( + in_channels=layer1_channels, + out_channels=layer2_channels, + kernel_size=3, + stride=2, + ), + ActivationBalancer(channel_dim=1), + DoubleSwish(), + nn.Conv2d( + in_channels=layer2_channels, + out_channels=layer3_channels, + kernel_size=3, + stride=2, + ), + ActivationBalancer(channel_dim=1), + DoubleSwish(), + ) + self.out = nn.Linear( + layer3_channels * (((in_channels - 1) // 2 - 1) // 2), out_channels + ) + # set learn_eps=False because out_norm is preceded by `out`, and `out` + # itself has learned scale, so the extra degree of freedom is not + # needed. + self.out_norm = BasicNorm(out_channels, learn_eps=False) + # constrain median of output to be close to zero. + self.out_balancer = ActivationBalancer( + channel_dim=-1, min_positive=0.45, max_positive=0.55 + ) + + def forward(self, x: torch.Tensor) -> torch.Tensor: + """Subsample x. + + Args: + x: + Its shape is (N, T, idim). + + Returns: + Return a tensor of shape (N, ((T-1)//2 - 1)//2, odim) + """ + # On entry, x is (N, T, idim) + x = x.unsqueeze(1) # (N, T, idim) -> (N, 1, T, idim) i.e., (N, C, H, W) + x = self.conv(x) + # Now x is of shape (N, odim, ((T-1)//2 - 1)//2, ((idim-1)//2 - 1)//2) + b, c, t, f = x.size() + x = self.out(x.transpose(1, 2).contiguous().view(b, t, c * f)) + # Now x is of shape (N, ((T-1)//2 - 1))//2, odim) + x = self.out_norm(x) + x = self.out_balancer(x) + return x + + +if __name__ == "__main__": + feature_dim = 50 + c = Conformer(num_features=feature_dim, d_model=128, nhead=4) + batch_size = 5 + seq_len = 20 + # Just make sure the forward pass runs. + f = c( + torch.randn(batch_size, seq_len, feature_dim), + torch.full((batch_size,), seq_len, dtype=torch.int64), + warmup=0.5, + ) diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7/decode.py b/egs/librispeech/ASR/pruned_transducer_stateless7/decode.py new file mode 100755 index 000000000..d1af63aaa --- /dev/null +++ b/egs/librispeech/ASR/pruned_transducer_stateless7/decode.py @@ -0,0 +1,633 @@ +#!/usr/bin/env python3 +# +# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang, +# Zengwei Yao) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# 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. +""" +Usage: +(1) greedy search +./pruned_transducer_stateless4/decode.py \ + --epoch 30 \ + --avg 15 \ + --exp-dir ./pruned_transducer_stateless4/exp \ + --max-duration 600 \ + --decoding-method greedy_search + +(2) beam search (not recommended) +./pruned_transducer_stateless4/decode.py \ + --epoch 30 \ + --avg 15 \ + --exp-dir ./pruned_transducer_stateless4/exp \ + --max-duration 600 \ + --decoding-method beam_search \ + --beam-size 4 + +(3) modified beam search +./pruned_transducer_stateless4/decode.py \ + --epoch 30 \ + --avg 15 \ + --exp-dir ./pruned_transducer_stateless4/exp \ + --max-duration 600 \ + --decoding-method modified_beam_search \ + --beam-size 4 + +(4) fast beam search +./pruned_transducer_stateless4/decode.py \ + --epoch 30 \ + --avg 15 \ + --exp-dir ./pruned_transducer_stateless4/exp \ + --max-duration 600 \ + --decoding-method fast_beam_search \ + --beam 4 \ + --max-contexts 4 \ + --max-states 8 +""" + + +import argparse +import logging +from collections import defaultdict +from pathlib import Path +from typing import Dict, List, Optional, Tuple + +import k2 +import sentencepiece as spm +import torch +import torch.nn as nn +from asr_datamodule import LibriSpeechAsrDataModule +from beam_search import ( + beam_search, + fast_beam_search_one_best, + greedy_search, + greedy_search_batch, + modified_beam_search, +) +from train import get_params, get_transducer_model + +from icefall.checkpoint import ( + average_checkpoints, + average_checkpoints_with_averaged_model, + find_checkpoints, + load_checkpoint, +) +from icefall.utils import ( + AttributeDict, + setup_logger, + store_transcripts, + str2bool, + write_error_stats, +) + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--epoch", + type=int, + default=30, + help="""It specifies the checkpoint to use for decoding. + Note: Epoch counts from 1. + You can specify --avg to use more checkpoints for model averaging.""", + ) + + parser.add_argument( + "--iter", + type=int, + default=0, + help="""If positive, --epoch is ignored and it + will use the checkpoint exp_dir/checkpoint-iter.pt. + You can specify --avg to use more checkpoints for model averaging. + """, + ) + + parser.add_argument( + "--avg", + type=int, + default=15, + help="Number of checkpoints to average. Automatically select " + "consecutive checkpoints before the checkpoint specified by " + "'--epoch' and '--iter'", + ) + + parser.add_argument( + "--use-averaged-model", + type=str2bool, + default=False, + help="Whether to load averaged model. Currently it only supports " + "using --epoch. If True, it would decode with the averaged model " + "over the epoch range from `epoch-avg` (excluded) to `epoch`." + "Actually only the models with epoch number of `epoch-avg` and " + "`epoch` are loaded for averaging. ", + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="pruned_transducer_stateless4/exp", + help="The experiment dir", + ) + + parser.add_argument( + "--bpe-model", + type=str, + default="data/lang_bpe_500/bpe.model", + help="Path to the BPE model", + ) + + parser.add_argument( + "--decoding-method", + type=str, + default="greedy_search", + help="""Possible values are: + - greedy_search + - beam_search + - modified_beam_search + - fast_beam_search + """, + ) + + parser.add_argument( + "--beam-size", + type=int, + default=4, + help="""An integer indicating how many candidates we will keep for each + frame. Used only when --decoding-method is beam_search or + modified_beam_search.""", + ) + + parser.add_argument( + "--beam", + type=float, + default=4, + help="""A floating point value to calculate the cutoff score during beam + search (i.e., `cutoff = max-score - beam`), which is the same as the + `beam` in Kaldi. + Used only when --decoding-method is fast_beam_search""", + ) + + parser.add_argument( + "--max-contexts", + type=int, + default=4, + help="""Used only when --decoding-method is + fast_beam_search""", + ) + + parser.add_argument( + "--max-states", + type=int, + default=8, + help="""Used only when --decoding-method is + fast_beam_search""", + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; " + "2 means tri-gram", + ) + parser.add_argument( + "--max-sym-per-frame", + type=int, + default=1, + help="""Maximum number of symbols per frame. + Used only when --decoding_method is greedy_search""", + ) + + return parser + + +def decode_one_batch( + params: AttributeDict, + model: nn.Module, + sp: spm.SentencePieceProcessor, + batch: dict, + decoding_graph: Optional[k2.Fsa] = None, +) -> Dict[str, List[List[str]]]: + """Decode one batch and return the result in a dict. The dict has the + following format: + + - key: It indicates the setting used for decoding. For example, + if greedy_search is used, it would be "greedy_search" + If beam search with a beam size of 7 is used, it would be + "beam_7" + - value: It contains the decoding result. `len(value)` equals to + batch size. `value[i]` is the decoding result for the i-th + utterance in the given batch. + Args: + params: + It's the return value of :func:`get_params`. + model: + The neural model. + sp: + The BPE model. + batch: + It is the return value from iterating + `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation + for the format of the `batch`. + decoding_graph: + The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used + only when --decoding_method is fast_beam_search. + Returns: + Return the decoding result. See above description for the format of + the returned dict. + """ + device = next(model.parameters()).device + feature = batch["inputs"] + assert feature.ndim == 3 + + feature = feature.to(device) + # at entry, feature is (N, T, C) + + supervisions = batch["supervisions"] + feature_lens = supervisions["num_frames"].to(device) + + encoder_out, encoder_out_lens = model.encoder( + x=feature, x_lens=feature_lens + ) + hyps = [] + + if params.decoding_method == "fast_beam_search": + hyp_tokens = fast_beam_search_one_best( + model=model, + decoding_graph=decoding_graph, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam, + max_contexts=params.max_contexts, + max_states=params.max_states, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif ( + params.decoding_method == "greedy_search" + and params.max_sym_per_frame == 1 + ): + hyp_tokens = greedy_search_batch( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif params.decoding_method == "modified_beam_search": + hyp_tokens = modified_beam_search( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam_size, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + else: + batch_size = encoder_out.size(0) + + for i in range(batch_size): + # fmt: off + encoder_out_i = encoder_out[i:i + 1, :encoder_out_lens[i]] + # fmt: on + if params.decoding_method == "greedy_search": + hyp = greedy_search( + model=model, + encoder_out=encoder_out_i, + max_sym_per_frame=params.max_sym_per_frame, + ) + elif params.decoding_method == "beam_search": + hyp = beam_search( + model=model, + encoder_out=encoder_out_i, + beam=params.beam_size, + ) + else: + raise ValueError( + f"Unsupported decoding method: {params.decoding_method}" + ) + hyps.append(sp.decode(hyp).split()) + + if params.decoding_method == "greedy_search": + return {"greedy_search": hyps} + elif params.decoding_method == "fast_beam_search": + return { + ( + f"beam_{params.beam}_" + f"max_contexts_{params.max_contexts}_" + f"max_states_{params.max_states}" + ): hyps + } + else: + return {f"beam_size_{params.beam_size}": hyps} + + +def decode_dataset( + dl: torch.utils.data.DataLoader, + params: AttributeDict, + model: nn.Module, + sp: spm.SentencePieceProcessor, + decoding_graph: Optional[k2.Fsa] = None, +) -> Dict[str, List[Tuple[List[str], List[str]]]]: + """Decode dataset. + + Args: + dl: + PyTorch's dataloader containing the dataset to decode. + params: + It is returned by :func:`get_params`. + model: + The neural model. + sp: + The BPE model. + decoding_graph: + The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used + only when --decoding_method is fast_beam_search. + Returns: + Return a dict, whose key may be "greedy_search" if greedy search + is used, or it may be "beam_7" if beam size of 7 is used. + Its value is a list of tuples. Each tuple contains two elements: + The first is the reference transcript, and the second is the + predicted result. + """ + num_cuts = 0 + + try: + num_batches = len(dl) + except TypeError: + num_batches = "?" + + if params.decoding_method == "greedy_search": + log_interval = 50 + else: + log_interval = 10 + + results = defaultdict(list) + for batch_idx, batch in enumerate(dl): + texts = batch["supervisions"]["text"] + + hyps_dict = decode_one_batch( + params=params, + model=model, + sp=sp, + decoding_graph=decoding_graph, + batch=batch, + ) + + for name, hyps in hyps_dict.items(): + this_batch = [] + assert len(hyps) == len(texts) + for hyp_words, ref_text in zip(hyps, texts): + ref_words = ref_text.split() + this_batch.append((ref_words, hyp_words)) + + results[name].extend(this_batch) + + num_cuts += len(texts) + + if batch_idx % log_interval == 0: + batch_str = f"{batch_idx}/{num_batches}" + + logging.info( + f"batch {batch_str}, cuts processed until now is {num_cuts}" + ) + return results + + +def save_results( + params: AttributeDict, + test_set_name: str, + results_dict: Dict[str, List[Tuple[List[int], List[int]]]], +): + test_set_wers = dict() + for key, results in results_dict.items(): + recog_path = ( + params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt" + ) + store_transcripts(filename=recog_path, texts=results) + logging.info(f"The transcripts are stored in {recog_path}") + + # The following prints out WERs, per-word error statistics and aligned + # ref/hyp pairs. + errs_filename = ( + params.res_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt" + ) + with open(errs_filename, "w") as f: + wer = write_error_stats( + f, f"{test_set_name}-{key}", results, enable_log=True + ) + test_set_wers[key] = wer + + logging.info("Wrote detailed error stats to {}".format(errs_filename)) + + test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1]) + errs_info = ( + params.res_dir + / f"wer-summary-{test_set_name}-{key}-{params.suffix}.txt" + ) + with open(errs_info, "w") as f: + print("settings\tWER", file=f) + for key, val in test_set_wers: + print("{}\t{}".format(key, val), file=f) + + s = "\nFor {}, WER of different settings are:\n".format(test_set_name) + note = "\tbest for {}".format(test_set_name) + for key, val in test_set_wers: + s += "{}\t{}{}\n".format(key, val, note) + note = "" + logging.info(s) + + +@torch.no_grad() +def main(): + parser = get_parser() + LibriSpeechAsrDataModule.add_arguments(parser) + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + + params = get_params() + params.update(vars(args)) + + assert params.decoding_method in ( + "greedy_search", + "beam_search", + "fast_beam_search", + "modified_beam_search", + ) + params.res_dir = params.exp_dir / params.decoding_method + + if params.iter > 0: + params.suffix = f"iter-{params.iter}-avg-{params.avg}" + else: + params.suffix = f"epoch-{params.epoch}-avg-{params.avg}" + + if "fast_beam_search" in params.decoding_method: + params.suffix += f"-beam-{params.beam}" + params.suffix += f"-max-contexts-{params.max_contexts}" + params.suffix += f"-max-states-{params.max_states}" + elif "beam_search" in params.decoding_method: + params.suffix += ( + f"-{params.decoding_method}-beam-size-{params.beam_size}" + ) + else: + params.suffix += f"-context-{params.context_size}" + params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}" + + if params.use_averaged_model: + params.suffix += "-use-averaged-model" + + setup_logger(f"{params.res_dir}/log-decode-{params.suffix}") + logging.info("Decoding started") + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", 0) + + logging.info(f"Device: {device}") + + sp = spm.SentencePieceProcessor() + sp.load(params.bpe_model) + + # and are defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.unk_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + + logging.info(params) + + logging.info("About to create model") + model = get_transducer_model(params) + + if not params.use_averaged_model: + if params.iter > 0: + filenames = find_checkpoints( + params.exp_dir, iteration=-params.iter + )[: params.avg] + if len(filenames) == 0: + raise ValueError( + f"No checkpoints found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + elif len(filenames) < params.avg: + raise ValueError( + f"Not enough checkpoints ({len(filenames)}) found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + logging.info(f"averaging {filenames}") + model.to(device) + model.load_state_dict(average_checkpoints(filenames, device=device)) + elif params.avg == 1: + load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model) + else: + start = params.epoch - params.avg + 1 + filenames = [] + for i in range(start, params.epoch + 1): + if i >= 1: + filenames.append(f"{params.exp_dir}/epoch-{i}.pt") + logging.info(f"averaging {filenames}") + model.to(device) + model.load_state_dict(average_checkpoints(filenames, device=device)) + else: + if params.iter > 0: + filenames = find_checkpoints( + params.exp_dir, iteration=-params.iter + )[: params.avg + 1] + if len(filenames) == 0: + raise ValueError( + f"No checkpoints found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + elif len(filenames) < params.avg + 1: + raise ValueError( + f"Not enough checkpoints ({len(filenames)}) found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + filename_start = filenames[-1] + filename_end = filenames[0] + logging.info( + "Calculating the averaged model over iteration checkpoints" + f" from {filename_start} (excluded) to {filename_end}" + ) + model.to(device) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + device=device, + ) + ) + else: + assert params.avg > 0, params.avg + start = params.epoch - params.avg + assert start >= 1, start + filename_start = f"{params.exp_dir}/epoch-{start}.pt" + filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt" + logging.info( + f"Calculating the averaged model over epoch range from " + f"{start} (excluded) to {params.epoch}" + ) + model.to(device) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + device=device, + ) + ) + + model.to(device) + model.eval() + + if params.decoding_method == "fast_beam_search": + decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device) + else: + decoding_graph = None + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + librispeech = LibriSpeechAsrDataModule(args) + + test_clean_cuts = librispeech.test_clean_cuts() + test_other_cuts = librispeech.test_other_cuts() + + test_clean_dl = librispeech.test_dataloaders(test_clean_cuts) + test_other_dl = librispeech.test_dataloaders(test_other_cuts) + + test_sets = ["test-clean", "test-other"] + test_dl = [test_clean_dl, test_other_dl] + + for test_set, test_dl in zip(test_sets, test_dl): + results_dict = decode_dataset( + dl=test_dl, + params=params, + model=model, + sp=sp, + decoding_graph=decoding_graph, + ) + + save_results( + params=params, + test_set_name=test_set, + results_dict=results_dict, + ) + + logging.info("Done!") + + +if __name__ == "__main__": + main() diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7/decoder.py b/egs/librispeech/ASR/pruned_transducer_stateless7/decoder.py new file mode 100644 index 000000000..a1c755d73 --- /dev/null +++ b/egs/librispeech/ASR/pruned_transducer_stateless7/decoder.py @@ -0,0 +1,102 @@ +# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# 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 as nn +import torch.nn.functional as F + + +class Decoder(nn.Module): + """This class modifies the stateless decoder from the following paper: + + RNN-transducer with stateless prediction network + https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=9054419 + + It removes the recurrent connection from the decoder, i.e., the prediction + network. Different from the above paper, it adds an extra Conv1d + right after the embedding layer. + + TODO: Implement https://arxiv.org/pdf/2109.07513.pdf + """ + + def __init__( + self, + vocab_size: int, + decoder_dim: int, + blank_id: int, + context_size: int, + ): + """ + Args: + vocab_size: + Number of tokens of the modeling unit including blank. + decoder_dim: + Dimension of the input embedding, and of the decoder output. + blank_id: + The ID of the blank symbol. + context_size: + Number of previous words to use to predict the next word. + 1 means bigram; 2 means trigram. n means (n+1)-gram. + """ + super().__init__() + + self.embedding = nn.Embedding( + num_embeddings=vocab_size, + embedding_dim=decoder_dim, + padding_idx=blank_id, + ) + self.blank_id = blank_id + + assert context_size >= 1, context_size + self.context_size = context_size + self.vocab_size = vocab_size + if context_size > 1: + self.conv = nn.Conv1d( + in_channels=decoder_dim, + out_channels=decoder_dim, + kernel_size=context_size, + padding=0, + groups=decoder_dim, + bias=False, + ) + + def forward(self, y: torch.Tensor, need_pad: bool = True) -> torch.Tensor: + """ + Args: + y: + A 2-D tensor of shape (N, U). + need_pad: + True to left pad the input. Should be True during training. + False to not pad the input. Should be False during inference. + Returns: + Return a tensor of shape (N, U, decoder_dim). + """ + y = y.to(torch.int64) + embedding_out = self.embedding(y) + if self.context_size > 1: + embedding_out = embedding_out.permute(0, 2, 1) + if need_pad is True: + embedding_out = F.pad( + embedding_out, pad=(self.context_size - 1, 0) + ) + else: + # During inference time, there is no need to do extra padding + # as we only need one output + assert embedding_out.size(-1) == self.context_size + embedding_out = self.conv(embedding_out) + embedding_out = embedding_out.permute(0, 2, 1) + embedding_out = F.relu(embedding_out) + return embedding_out diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7/encoder_interface.py b/egs/librispeech/ASR/pruned_transducer_stateless7/encoder_interface.py new file mode 120000 index 000000000..b9aa0ae08 --- /dev/null +++ b/egs/librispeech/ASR/pruned_transducer_stateless7/encoder_interface.py @@ -0,0 +1 @@ +../pruned_transducer_stateless2/encoder_interface.py \ No newline at end of file diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7/export.py b/egs/librispeech/ASR/pruned_transducer_stateless7/export.py new file mode 120000 index 000000000..19c56a722 --- /dev/null +++ b/egs/librispeech/ASR/pruned_transducer_stateless7/export.py @@ -0,0 +1 @@ +../pruned_transducer_stateless2/export.py \ No newline at end of file diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7/joiner.py b/egs/librispeech/ASR/pruned_transducer_stateless7/joiner.py new file mode 100644 index 000000000..afcd690e9 --- /dev/null +++ b/egs/librispeech/ASR/pruned_transducer_stateless7/joiner.py @@ -0,0 +1,66 @@ +# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# 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 as nn + + +class Joiner(nn.Module): + def __init__( + self, + encoder_dim: int, + decoder_dim: int, + joiner_dim: int, + vocab_size: int, + ): + super().__init__() + + self.encoder_proj = nn.Linear(encoder_dim, joiner_dim) + self.decoder_proj = nn.Linear(decoder_dim, joiner_dim) + self.output_linear = nn.Linear(joiner_dim, vocab_size) + + def forward( + self, + encoder_out: torch.Tensor, + decoder_out: torch.Tensor, + project_input: bool = True, + ) -> torch.Tensor: + """ + Args: + encoder_out: + Output from the encoder. Its shape is (N, T, s_range, C). + decoder_out: + Output from the decoder. Its shape is (N, T, s_range, C). + project_input: + If true, apply input projections encoder_proj and decoder_proj. + If this is false, it is the user's responsibility to do this + manually. + Returns: + Return a tensor of shape (N, T, s_range, C). + """ + assert encoder_out.ndim == decoder_out.ndim == 4 + assert encoder_out.shape[:-1] == decoder_out.shape[:-1] + + if project_input: + logit = self.encoder_proj(encoder_out) + self.decoder_proj( + decoder_out + ) + else: + logit = encoder_out + decoder_out + + logit = self.output_linear(torch.tanh(logit)) + + return logit diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7/model.py b/egs/librispeech/ASR/pruned_transducer_stateless7/model.py new file mode 100644 index 000000000..24898ed09 --- /dev/null +++ b/egs/librispeech/ASR/pruned_transducer_stateless7/model.py @@ -0,0 +1,192 @@ +# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang, Wei Kang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# 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 k2 +import torch +import torch.nn as nn +from encoder_interface import EncoderInterface + +from icefall.utils import add_sos + + +class Transducer(nn.Module): + """It implements https://arxiv.org/pdf/1211.3711.pdf + "Sequence Transduction with Recurrent Neural Networks" + """ + + def __init__( + self, + encoder: EncoderInterface, + decoder: nn.Module, + joiner: nn.Module, + encoder_dim: int, + decoder_dim: int, + joiner_dim: int, + vocab_size: int, + ): + """ + Args: + encoder: + It is the transcription network in the paper. Its accepts + two inputs: `x` of (N, T, encoder_dim) and `x_lens` of shape (N,). + It returns two tensors: `logits` of shape (N, T, encoder_dm) and + `logit_lens` of shape (N,). + decoder: + It is the prediction network in the paper. Its input shape + is (N, U) and its output shape is (N, U, decoder_dim). + It should contain one attribute: `blank_id`. + joiner: + It has two inputs with shapes: (N, T, encoder_dim) and (N, U, decoder_dim). + Its output shape is (N, T, U, vocab_size). Note that its output contains + unnormalized probs, i.e., not processed by log-softmax. + """ + super().__init__() + assert isinstance(encoder, EncoderInterface), type(encoder) + assert hasattr(decoder, "blank_id") + + self.encoder = encoder + self.decoder = decoder + self.joiner = joiner + + self.simple_am_proj = nn.Linear( + encoder_dim, vocab_size, + ) + self.simple_lm_proj = nn.Linear(decoder_dim, vocab_size) + + def forward( + self, + x: torch.Tensor, + x_lens: torch.Tensor, + y: k2.RaggedTensor, + prune_range: int = 5, + am_scale: float = 0.0, + lm_scale: float = 0.0, + warmup: float = 1.0, + ) -> torch.Tensor: + """ + Args: + x: + A 3-D tensor of shape (N, T, C). + x_lens: + A 1-D tensor of shape (N,). It contains the number of frames in `x` + before padding. + y: + A ragged tensor with 2 axes [utt][label]. It contains labels of each + utterance. + prune_range: + The prune range for rnnt loss, it means how many symbols(context) + we are considering for each frame to compute the loss. + am_scale: + The scale to smooth the loss with am (output of encoder network) + part + lm_scale: + The scale to smooth the loss with lm (output of predictor network) + part + warmup: + A value warmup >= 0 that determines which modules are active, values + warmup > 1 "are fully warmed up" and all modules will be active. + Returns: + Return the transducer loss. + + Note: + Regarding am_scale & lm_scale, it will make the loss-function one of + the form: + lm_scale * lm_probs + am_scale * am_probs + + (1-lm_scale-am_scale) * combined_probs + """ + assert x.ndim == 3, x.shape + assert x_lens.ndim == 1, x_lens.shape + assert y.num_axes == 2, y.num_axes + + assert x.size(0) == x_lens.size(0) == y.dim0 + + encoder_out, x_lens = self.encoder(x, x_lens, warmup=warmup) + assert torch.all(x_lens > 0) + + # Now for the decoder, i.e., the prediction network + row_splits = y.shape.row_splits(1) + y_lens = row_splits[1:] - row_splits[:-1] + + blank_id = self.decoder.blank_id + sos_y = add_sos(y, sos_id=blank_id) + + # sos_y_padded: [B, S + 1], start with SOS. + sos_y_padded = sos_y.pad(mode="constant", padding_value=blank_id) + + # decoder_out: [B, S + 1, decoder_dim] + decoder_out = self.decoder(sos_y_padded) + + # Note: y does not start with SOS + # y_padded : [B, S] + y_padded = y.pad(mode="constant", padding_value=0) + + y_padded = y_padded.to(torch.int64) + boundary = torch.zeros( + (x.size(0), 4), dtype=torch.int64, device=x.device + ) + boundary[:, 2] = y_lens + boundary[:, 3] = x_lens + + lm = self.simple_lm_proj(decoder_out) + am = self.simple_am_proj(encoder_out) + + with torch.cuda.amp.autocast(enabled=False): + simple_loss, (px_grad, py_grad) = k2.rnnt_loss_smoothed( + lm=lm.float(), + am=am.float(), + symbols=y_padded, + termination_symbol=blank_id, + lm_only_scale=lm_scale, + am_only_scale=am_scale, + boundary=boundary, + reduction="sum", + return_grad=True, + ) + + # ranges : [B, T, prune_range] + ranges = k2.get_rnnt_prune_ranges( + px_grad=px_grad, + py_grad=py_grad, + boundary=boundary, + s_range=prune_range, + ) + + # am_pruned : [B, T, prune_range, encoder_dim] + # lm_pruned : [B, T, prune_range, decoder_dim] + am_pruned, lm_pruned = k2.do_rnnt_pruning( + am=self.joiner.encoder_proj(encoder_out), + lm=self.joiner.decoder_proj(decoder_out), + ranges=ranges, + ) + + # logits : [B, T, prune_range, vocab_size] + + # project_input=False since we applied the decoder's input projections + # prior to do_rnnt_pruning (this is an optimization for speed). + logits = self.joiner(am_pruned, lm_pruned, project_input=False) + + with torch.cuda.amp.autocast(enabled=False): + pruned_loss = k2.rnnt_loss_pruned( + logits=logits.float(), + symbols=y_padded, + ranges=ranges, + termination_symbol=blank_id, + boundary=boundary, + reduction="sum", + ) + + return (simple_loss, pruned_loss) diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7/optim.py b/egs/librispeech/ASR/pruned_transducer_stateless7/optim.py new file mode 100644 index 000000000..d7e0193dc --- /dev/null +++ b/egs/librispeech/ASR/pruned_transducer_stateless7/optim.py @@ -0,0 +1,802 @@ +# Copyright 2022 Xiaomi Corp. (authors: Daniel Povey) +# +# See ../LICENSE for clarification regarding multiple authors +# +# 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. + + +from typing import List, Optional, Union, Tuple, List +from lhotse.utils import fix_random_seed +import torch +from torch import Tensor +from torch.optim import Optimizer + + + + + + +class Cain(Optimizer): + """Implements Cain algorithm, which is a substantially modified form of the + Adam optimizer. + The modifications can be summarized as follows: + + (i) change Adam to what we called "Eve", which adds an extra configuration + value, "target_rms" (default value: 0.1); and for non-scalar parameters, + if their root-mean-square value exceeds "target_rms", we apply weight + decay (aggressive enough weight decay that the rms stays at target_rms). + This weight decay is applied in the same way as in AdamW, i.e. by + parameter shrinkage rather than by modifying the gradients. + (ii) By limiting the rms to 0.1, we reduce our modeling power; we restore + it by replacing all weights and biases-- in fact, all model parameters-- + with expressions like + weight * weight_scale.exp() or bias * bias_scale.exp(). + This has the benefit of making the learnable offsets and scales of + BatchNorm/LayerNorm unnecessary. + Below we describe the changes from Eve to Cain: + (iii) We removed all the weight_scale and bias_scale parameters from + the model and "moved them into the optimizer". Now we allow + parameters to have any rms value (that's <= 10), but we: + (a) make the update speed for a parameter proportional to + its rms value, using target_rms (0.1) as the reference point + where we just use the standard Adam-type formula. + (b) simulate the effect of learning the weight_scale or bias_scale + in log space using standard Adam. (see scale_exp_avg_sq in the + code below). + (iv) Periodically (every 2000 batches) perform a "change of basis" on + all non-convolutional axes of each parameter tensor, and reset the + Adam statistics. This means an orthogonal change of co-ordinates + in the space corresponding to each dimension of a tensor, e.g. in + a 256 x 512 parameter tensor we'll have a 256x256 orthogonal matrix + and a 512x512 one. This does a better job of separating "important" + from "unimportant" directions in activation space, and empirically + improved convergence, final loss function and WERs. + + The original Adam algorithm was proposed in `Adam: A Method for Stochastic Optimization`_. + The AdamW variant was proposed in `Decoupled Weight Decay Regularization`_. + Eve is unpublished so far. + + Arguments: + params (iterable): iterable of parameters to optimize or dicts defining + parameter groups + lr (float, optional): learning rate (default: 1e-3) + betas (Tuple[float, float], optional): coefficients used for computing + running averages of gradient and its square (default: (0.9, 0.999)) + eps (float, optional): term added to the denominator to improve + numerical stability (default: 1e-8) + target_rms (float, optional): target root-mean-square value of + parameters, if they fall below this we will stop applying weight decay. + rms_eps (float, optional): epsilon value such that when parameter + rms value goes below this, we stop learning slower for that parameter, + i.e. we treat the rms as if it were rms_eps. In practice, rms values + will not go much below this. + rms_max (float, optional): maximum root-mean-square value for + non-scalar parameters, such that we don't let the parameter + values exceed this; we'll shrink any parameter matrices that becomes + larger than this. + + + + + .. _Adam\: A Method for Stochastic Optimization: + https://arxiv.org/abs/1412.6980 + .. _Decoupled Weight Decay Regularization: + https://arxiv.org/abs/1711.05101 + .. _On the Convergence of Adam and Beyond: + https://openreview.net/forum?id=ryQu7f-RZ + + """ + + def __init__( + self, + params, + lr=1e-3, + betas=(0.9, 0.98), + eps=1e-8, + target_rms=0.1, + rms_eps=1.0e-05, + rms_max=10.0, + ): + + if not 0.0 <= lr: + raise ValueError("Invalid learning rate: {}".format(lr)) + if not 0.0 <= eps: + raise ValueError("Invalid epsilon value: {}".format(eps)) + if not 0.0 <= betas[0] < 1.0: + raise ValueError( + "Invalid beta parameter at index 0: {}".format(betas[0]) + ) + if not 0.0 <= betas[1] < 1.0: + raise ValueError( + "Invalid beta parameter at index 1: {}".format(betas[1]) + ) + if not 0 < target_rms <= 10.0: + raise ValueError("Invalid target_rms value: {}".format(target_rms)) + if not 0.1 < rms_max <= 1000.0: + raise ValueError("Invalid rms_max value: {}".format(rms_max)) + if not 0.0 < rms_eps < 0.1: + raise ValueError("Invalid rms_eps value: {}".format(rms_eps)) + + defaults = dict( + lr=lr, + betas=betas, + eps=eps, + target_rms=target_rms, + rms_eps=rms_eps, + rms_max=rms_max, + ) + super(Cain, self).__init__(params, defaults) + + def __setstate__(self, state): + super(Cain, self).__setstate__(state) + + @torch.no_grad() + def step(self, closure=None): + """Performs a single optimization step. + + Arguments: + closure (callable, optional): A closure that reevaluates the model + and returns the loss. + """ + loss = None + if closure is not None: + with torch.enable_grad(): + loss = closure() + + for group in self.param_groups: + beta1, beta2 = group["betas"] + lr = group["lr"] + target_rms = group["target_rms"] + rms_eps = group["rms_eps"] + eps = group["eps"] + rms_max = group["rms_max"] + + for p in group["params"]: + if p.grad is None: + continue + + # Perform optimization step + grad = p.grad + if grad.is_sparse: + raise RuntimeError( + "AdamW does not support sparse gradients" + ) + + state = self.state[p] + + # State initialization + if len(state) == 0: + state["step"] = 0 + # Exponential moving average of gradient values + state["delta"] = torch.zeros_like( + p, memory_format=torch.preserve_format + ) + # Exponential moving average of squared gradient values + state["exp_avg_sq"] = torch.zeros_like( + p, memory_format=torch.preserve_format + ) + if p.numel() > 1: + # we keep track of the RMS value of the parameters to + # help set the learning rate... this emulates Eve. + state["param_rms"] = (p**2).mean().sqrt() + # we learn the scale on the parameters in a way + # that also emulates Eve. + state["scale_exp_avg_sq"] = torch.zeros((), device=p.device, + dtype=p.dtype) + # alpha is a scale related to the natural-gradient update, + # see self._get_denom() + state["alpha"] = torch.ones((), device=p.device, + dtype=p.dtype) + + + + delta, exp_avg_sq = state["delta"], state["exp_avg_sq"] + step = state["step"] + + delta.mul_(beta1) + + if p.numel() > 1: + # This part learns the scale on the parameters. + # scale_deriv is the derivative w.r.t. a log-space scale on the parameters, i.e. + # if we imagine: p = underlying_param * scale.exp(), this is the derivative + # w.r.t. scale (it does not actually matter what value `scale` currently has). + scale_deriv = (p * grad).sum() + scale_exp_avg_sq = state["scale_exp_avg_sq"] + scale_exp_avg_sq.mul_(beta2).addcmul_(scale_deriv, scale_deriv, + value=1 - beta2) + + # should actually be step + 1, so on 1st minibatch we are not learning + # anything here. May fix this at some point. + scale_bias_correction2 = 1 - beta2 ** (step + 1) + + scale_denom = (scale_exp_avg_sq.sqrt()).add_(eps) + + scale_alpha = -lr * (scale_bias_correction2 ** 0.5) * (1-beta1) + + # scale_delta is going to be the change in log-scale (before momentum), i.e. if + # p = underlying_param * scale.exp(), + # delta is the change in `scale`. + scale_delta = scale_alpha * (scale_deriv / scale_denom) + + # the following is equivalent to: + # if torch.all(state["param_rms"] > rms_max): + # delta = -1.0e-03 + # which means: if the parameter root-mean-square value goes above the + # specified limit, start to decay the parameters (and ignore the computed + # delta). + scale_delta = torch.where(state["param_rms"] > rms_max, + torch.tensor(-1.0e-03 * (1-beta1), + device=scale_delta.device, + dtype=scale_delta.dtype), + scale_delta) + delta.add_(p, alpha=scale_delta) + + + # forget bias_correction1. We use normal momentum. delta really + # just stores the moving-average gradient step. + bias_correction2 = 1 - beta2 ** self._step_for_bias_correction(step) + + grad = self._change_coordinates(grad, state, forward=True) + + # Decay the second moment running average coefficient + exp_avg_sq.mul_(beta2).addcmul_(grad, grad, value=1 - beta2) + + denom = (exp_avg_sq.sqrt()).add_(eps) + + this_delta = grad / denom + + this_delta = self._change_coordinates(this_delta, state, forward=False) + + alpha = -lr*(1-beta1)*(bias_correction2 ** 0.5) + if p.numel() > 1: + if step % 10 == 0: + state["param_rms"].fill_((p**2).mean().sqrt()) + # imagine param was normalized to rms=target_rms and we stored the + # scale as a separate scalar. This is how fast we'd be learning. + alpha = (alpha / target_rms) * state["param_rms"].clamp(min=rms_eps) + + delta.add_(this_delta, alpha=alpha) + + p.add_(delta) + state["step"] += 1 + + return loss + + + def _change_coordinates(self, + grad: Tensor, + state: dict, + forward: bool): + """ + Possibly performs some magnitude-preserving changes of co-ordinates on `grad` + """ + ndim = grad.ndim + numel = grad.numel() + step = state["step"] + for dim in range(grad.ndim): + # see for each dim in turn whether we want to perform any changes in co-ordinates, + # or store any stats. + size = grad.shape[dim] + if size <= 3 or (size % 2 == 1 and size < 128) or size >= 2048 or size == numel: + # we don't do any such co-ordinate changes in dims with sizes + # that are too small (no point) or large (too slow), or that are + # assumed convolutional (because they are odd and not too huge). + # We can revisit this later. + continue + grad = self._change_coordinates_for_dim(grad, state, dim, forward) + return grad + + + def _change_coordinates_for_dim(self, + grad: Tensor, + state: dict, + dim: int, + forward: bool, + orig_dim: int = -1): + assert grad.ndim > 1 + if not (grad.ndim == 2 and dim == 1): + # normalize the format and recurse. + dims_order = [] + rev_dims_order = [] + ndim = grad.ndim + for i in range(dim): + dims_order.append(i) + rev_dims_order.append(i) + rev_dims_order.append(ndim-1) + for i in range(dim+1, ndim): + dims_order.append(i) + rev_dims_order.append(i-1) + dims_order.append(dim) + # e.g. ndim=4, dim=1, dims_order=(0,2,3,1), rev_dims_order=(0,3,1,2) + new_grad = grad.permute(*dims_order) + assert new_grad.shape[-1] == grad.shape[dim] + new_shape = new_grad.shape + new_grad = new_grad.reshape(-1, new_grad.shape[-1]) + new_grad = self._change_coordinates_for_dim(new_grad, state, 1, forward, + orig_dim=dim) + return new_grad.reshape(new_shape).permute(*rev_dims_order) + + # OK: grad.ndim == 2 and dim == 1 + size = grad.shape[1] + if orig_dim == -1: + orig_dim = dim + step = state["step"] + must_store_stats, must_zero_stats = self._must_store_stats(step) + if must_store_stats and forward: + # store stats for 200 iters preceding when we estimate the + # transform. + stats_name = f"cov_{orig_dim}" + if not stats_name in state: + state[stats_name] = torch.zeros(size, size, dtype=grad.dtype, + device=grad.device) + cov = state[stats_name] + if must_zero_stats: + cov.zero_() + cov += torch.matmul(grad.t(), grad) * (1/grad.shape[0]) + + transform_name = f"t_{orig_dim}" + if transform_name not in state: + # make sure they're all allocated at the start, may be better for + # fragmention and debugging reasons (although may not). + state[transform_name] = torch.eye(size, device=grad.device, + dtype=grad.dtype) + + must_estimate_transform = self._must_estimate_transform(step) + if must_estimate_transform and forward: + stats_name = f"cov_{orig_dim}" + cov = state[stats_name] + l, U = cov.symeig(eigenvectors=True) + #print("estimate, l = ", l[::20]) + t = U.t() # t is indexed (new_dim, old_dim) a.k.a. (transformed_dim, real_dim) + state[transform_name][:] = t + state["exp_avg_sq"].zero_() # zero exp-avg-sq stats, we'll restart these from scratch. + + t = state[transform_name] + if forward: + return torch.matmul(grad, t.t()) + else: + return torch.matmul(grad, t) + + def _must_store_stats(self, step: int) -> Tuple[bool, bool]: + """ + Returns (must_store_stats, must_zero_stats), where + must_store_stats: bool, is True for the 100 or 200 steps preceding + a step on which we will estimate the transform. + must_zero_stats: bool, is True for only the 1st of the 100 or 200 + steps mentioned above. + """ + if step < 4000: + if step < 2000: + return (step % 500 >= 400, step % 500 == 400) + else: + return (step % 1000 >= 800, step % 1000 == 800) + else: + return (step % 2000 >= 1800, step % 2000 == 1800) + + def _must_estimate_transform(self, step: int) -> bool: + """ + Returns True if we will re-estimate the transform on this step + """ + if step < 4000: + if step < 2000: + return step % 500 == 0 and step > 0 + else: + return step % 1000 == 0 and step > 0 + else: + return step % 2000 == 0 + + def _step_for_bias_correction(self, step: int) -> bool: + """ + Return a modified step for bias-correction (actually just to give us the + right denominator for the exp-avg-sq stats).. this needs to take into + account how long it has been since we re-estimated the transforms, because + we zero the exp-avg-sq stats at those times. + + The + 1 is for the "current step". We post-increment the step so it + starts from 0. + """ + if step < 4000: + if step < 2000: + return step % 500 + 1 + else: + return step % 1000 + 1 + else: + return step % 2000 + 1 + + + +class LRScheduler(object): + """ + Base-class for learning rate schedulers where the learning-rate depends on both the + batch and the epoch. + """ + + def __init__(self, optimizer: Optimizer, verbose: bool = False): + # Attach optimizer + if not isinstance(optimizer, Optimizer): + raise TypeError( + "{} is not an Optimizer".format(type(optimizer).__name__) + ) + self.optimizer = optimizer + self.verbose = verbose + + for group in optimizer.param_groups: + group.setdefault("initial_lr", group["lr"]) + + self.base_lrs = [ + group["initial_lr"] for group in optimizer.param_groups + ] + + self.epoch = 0 + self.batch = 0 + + def state_dict(self): + """Returns the state of the scheduler as a :class:`dict`. + + It contains an entry for every variable in self.__dict__ which + is not the optimizer. + """ + return { + "base_lrs": self.base_lrs, + "epoch": self.epoch, + "batch": self.batch, + } + + def load_state_dict(self, state_dict): + """Loads the schedulers state. + + Args: + state_dict (dict): scheduler state. Should be an object returned + from a call to :meth:`state_dict`. + """ + self.__dict__.update(state_dict) + + def get_last_lr(self) -> List[float]: + """Return last computed learning rate by current scheduler. Will be a list of float.""" + return self._last_lr + + def get_lr(self): + # Compute list of learning rates from self.epoch and self.batch and + # self.base_lrs; this must be overloaded by the user. + # e.g. return [some_formula(self.batch, self.epoch, base_lr) for base_lr in self.base_lrs ] + raise NotImplementedError + + def step_batch(self, batch: Optional[int] = None) -> None: + # Step the batch index, or just set it. If `batch` is specified, it + # must be the batch index from the start of training, i.e. summed over + # all epochs. + # You can call this in any order; if you don't provide 'batch', it should + # of course be called once per batch. + if batch is not None: + self.batch = batch + else: + self.batch = self.batch + 1 + self._set_lrs() + + def step_epoch(self, epoch: Optional[int] = None): + # Step the epoch index, or just set it. If you provide the 'epoch' arg, + # you should call this at the start of the epoch; if you don't provide the 'epoch' + # arg, you should call it at the end of the epoch. + if epoch is not None: + self.epoch = epoch + else: + self.epoch = self.epoch + 1 + self._set_lrs() + + def _set_lrs(self): + values = self.get_lr() + assert len(values) == len(self.optimizer.param_groups) + + for i, data in enumerate(zip(self.optimizer.param_groups, values)): + param_group, lr = data + param_group["lr"] = lr + self.print_lr(self.verbose, i, lr) + self._last_lr = [group["lr"] for group in self.optimizer.param_groups] + + def print_lr(self, is_verbose, group, lr): + """Display the current learning rate.""" + if is_verbose: + print( + f"Epoch={self.epoch}, batch={self.batch}: adjusting learning rate" + f" of group {group} to {lr:.4e}." + ) + +class Eve(Optimizer): + """ + Implements Eve algorithm. This is a modified version of AdamW with a special + way of setting the weight-decay / shrinkage-factor, which is designed to make the + rms of the parameters approach a particular target_rms (default: 0.1). This is + for use with networks with 'scaled' versions of modules (see scaling.py), which + will be close to invariant to the absolute scale on the parameter matrix. + + The original Adam algorithm was proposed in `Adam: A Method for Stochastic Optimization`_. + The AdamW variant was proposed in `Decoupled Weight Decay Regularization`_. + Eve is unpublished so far. + + Arguments: + params (iterable): iterable of parameters to optimize or dicts defining + parameter groups + lr (float, optional): learning rate (default: 1e-3) + betas (Tuple[float, float], optional): coefficients used for computing + running averages of gradient and its square (default: (0.9, 0.999)) + eps (float, optional): term added to the denominator to improve + numerical stability (default: 1e-8) + weight_decay (float, optional): weight decay coefficient (default: 3e-4; + this value means that the weight would decay significantly after + about 3k minibatches. Is not multiplied by learning rate, but + is conditional on RMS-value of parameter being > target_rms. + target_rms (float, optional): target root-mean-square value of + parameters, if they fall below this we will stop applying weight decay. + + + .. _Adam\: A Method for Stochastic Optimization: + https://arxiv.org/abs/1412.6980 + .. _Decoupled Weight Decay Regularization: + https://arxiv.org/abs/1711.05101 + .. _On the Convergence of Adam and Beyond: + https://openreview.net/forum?id=ryQu7f-RZ + """ + + def __init__( + self, + params, + lr=1e-3, + betas=(0.9, 0.98), + eps=1e-8, + weight_decay=1e-3, + target_rms=0.1, + ): + + if not 0.0 <= lr: + raise ValueError("Invalid learning rate: {}".format(lr)) + if not 0.0 <= eps: + raise ValueError("Invalid epsilon value: {}".format(eps)) + if not 0.0 <= betas[0] < 1.0: + raise ValueError( + "Invalid beta parameter at index 0: {}".format(betas[0]) + ) + if not 0.0 <= betas[1] < 1.0: + raise ValueError( + "Invalid beta parameter at index 1: {}".format(betas[1]) + ) + if not 0 <= weight_decay <= 0.1: + raise ValueError( + "Invalid weight_decay value: {}".format(weight_decay) + ) + if not 0 < target_rms <= 10.0: + raise ValueError("Invalid target_rms value: {}".format(target_rms)) + defaults = dict( + lr=lr, + betas=betas, + eps=eps, + weight_decay=weight_decay, + target_rms=target_rms, + ) + super(Eve, self).__init__(params, defaults) + + def __setstate__(self, state): + super(Eve, self).__setstate__(state) + + @torch.no_grad() + def step(self, closure=None): + """Performs a single optimization step. + + Arguments: + closure (callable, optional): A closure that reevaluates the model + and returns the loss. + """ + loss = None + if closure is not None: + with torch.enable_grad(): + loss = closure() + + for group in self.param_groups: + for p in group["params"]: + if p.grad is None: + continue + + # Perform optimization step + grad = p.grad + if grad.is_sparse: + raise RuntimeError( + "AdamW does not support sparse gradients" + ) + + state = self.state[p] + + # State initialization + if len(state) == 0: + state["step"] = 0 + # Exponential moving average of gradient values + state["exp_avg"] = torch.zeros_like( + p, memory_format=torch.preserve_format + ) + # Exponential moving average of squared gradient values + state["exp_avg_sq"] = torch.zeros_like( + p, memory_format=torch.preserve_format + ) + + exp_avg, exp_avg_sq = state["exp_avg"], state["exp_avg_sq"] + + beta1, beta2 = group["betas"] + + state["step"] += 1 + bias_correction1 = 1 - beta1 ** state["step"] + bias_correction2 = 1 - beta2 ** state["step"] + + # Decay the first and second moment running average coefficient + exp_avg.mul_(beta1).add_(grad, alpha=1 - beta1) + exp_avg_sq.mul_(beta2).addcmul_(grad, grad, value=1 - beta2) + denom = (exp_avg_sq.sqrt() * (bias_correction2 ** -0.5)).add_( + group["eps"] + ) + + step_size = group["lr"] / bias_correction1 + target_rms = group["target_rms"] + weight_decay = group["weight_decay"] + + if p.numel() > 1: + # avoid applying this weight-decay on "scaling factors" + # (which are scalar). + is_above_target_rms = p.norm() > ( + target_rms * (p.numel() ** 0.5) + ) + p.mul_(1 - (weight_decay * is_above_target_rms)) + + p.addcdiv_(exp_avg, denom, value=-step_size) + + return loss + + +class Eden(LRScheduler): + """ + Eden scheduler. + lr = initial_lr * (((batch**2 + lr_batches**2) / lr_batches**2) ** -0.25 * + (((epoch**2 + lr_epochs**2) / lr_epochs**2) ** -0.25)) + + E.g. suggest initial-lr = 0.003 (passed to optimizer). + + Args: + optimizer: the optimizer to change the learning rates on + lr_batches: the number of batches after which we start significantly + decreasing the learning rate, suggest 5000. + lr_epochs: the number of epochs after which we start significantly + decreasing the learning rate, suggest 6 if you plan to do e.g. + 20 to 40 epochs, but may need smaller number if dataset is huge + and you will do few epochs. + """ + + def __init__( + self, + optimizer: Optimizer, + lr_batches: Union[int, float], + lr_epochs: Union[int, float], + verbose: bool = False, + ): + super(Eden, self).__init__(optimizer, verbose) + self.lr_batches = lr_batches + self.lr_epochs = lr_epochs + + def get_lr(self): + factor = ( + (self.batch ** 2 + self.lr_batches ** 2) / self.lr_batches ** 2 + ) ** -0.25 * ( + ((self.epoch ** 2 + self.lr_epochs ** 2) / self.lr_epochs ** 2) + ** -0.25 + ) + return [x * factor for x in self.base_lrs] + + +def _test_eden(): + m = torch.nn.Linear(100, 100) + optim = Cain(m.parameters(), lr=0.003) + + scheduler = Eden(optim, lr_batches=30, lr_epochs=2, verbose=True) + + for epoch in range(10): + scheduler.step_epoch(epoch) # sets epoch to `epoch` + + for step in range(20): + x = torch.randn(200, 100).detach() + x.requires_grad = True + y = m(x) + dy = torch.randn(200, 100).detach() + f = (y * dy).sum() + f.backward() + + optim.step() + scheduler.step_batch() + optim.zero_grad() + + print("last lr = ", scheduler.get_last_lr()) + print("state dict = ", scheduler.state_dict()) + + +def _test_eve_cain(): + import timeit + from scaling import ScaledLinear + E = 100 + B = 4 + T = 2 + print("in test_eve_cain") + device = torch.device('cuda') + dtype = torch.float32 + + # these input_magnitudes and output_magnitudes are to test that + # Abel is working as we expect and is able to adjust scales of + # different dims differently. + input_magnitudes = (1.0 * torch.randn(E, dtype=dtype, device=device)).exp() + output_magnitudes = (1.0 * torch.randn(E, dtype=dtype, device=device)).exp() + + for iter in [1,0]: + fix_random_seed(42) + Linear = torch.nn.Linear if iter == 0 else ScaledLinear + m = torch.nn.Sequential(Linear(E, 200), + torch.nn.ReLU(), + Linear(200, E)).to(device) + + train_pairs = [ (100.0 * torch.randn(B, T, E, device=device, dtype=dtype) * input_magnitudes, + torch.randn(B, T, E, device=device, dtype=dtype) * output_magnitudes) for _ in range(20) ] + + if iter == 0: optim = Eve(m.parameters(), lr=0.003) + else: optim = Cain(m.parameters(), lr=0.003) + scheduler = Eden(optim, lr_batches=200, lr_epochs=10, verbose=False) + + start = timeit.default_timer() + for epoch in range(150): + scheduler.step_epoch() + for n, (x,y) in enumerate(train_pairs): + y_out = m(x) + loss = ((y_out - y)**2).mean() * 100.0 + if n == 0 and epoch % 10 == 0: + norm1 = '%.2e' % (m[0].weight**2).mean().sqrt().item() + norm1b = '%.2e' % (m[0].bias**2).mean().sqrt().item() + norm2 = '%.2e' % (m[2].weight**2).mean().sqrt().item() + norm2b = '%.2e' % (m[2].bias**2).mean().sqrt().item() + #scale1 = '%.2e' % (m[0].weight_scale.exp().item()) + #scale1b = '%.2e' % (m[0].bias_scale.exp().item()) + #scale2 = '%.2e' % (m[2].weight_scale.exp().item()) + #scale2b = '%.2e' % (m[2].bias_scale.exp().item()) + print(f"Iter {iter}, epoch {epoch}, batch {n}, loss {loss.item()}, norms={norm1,norm1b,norm2,norm2b}") # scales={scale1,scale1b,scale2,scale2b} + loss.log().backward() + optim.step() + optim.zero_grad() + scheduler.step_batch() + + stop = timeit.default_timer() + print(f"Iter={iter}, Time taken: {stop - start}") + + print("last lr = ", scheduler.get_last_lr()) + #print("state dict = ", scheduler.state_dict()) + #print("optim state_dict = ", optim.state_dict()) + print("input_magnitudes = ", input_magnitudes) + print("output_magnitudes = ", output_magnitudes) + + def stddev(x): + return ((x-x.mean())**2).mean().sqrt() + print("Un-normalized input col magnitudes log-stddev: ", stddev((m[0].weight**2).sum(dim=0).sqrt().log())) + print("Normalized input col magnitudes log-stddev: ", stddev(((m[0].weight**2).sum(dim=0).sqrt() * input_magnitudes).log())) + + print("Un-normalized 0-output row magnitudes log-stddev: ", stddev((m[0].weight**2).sum(dim=1).sqrt().log())) + print("Un-normalized 2-input col magnitudes log-stddev: ", stddev((m[2].weight**2).sum(dim=0).sqrt().log())) + print("Un-normalized 2-output row magnitudes log-stddev: ", stddev((m[2].weight**2).sum(dim=1).sqrt().log())) + + print("Normalized output row magnitudes log-stddev: ", stddev(((m[2].weight**2).sum(dim=1).sqrt() / output_magnitudes).log())) + + + +if __name__ == "__main__": + torch.set_num_threads(1) + torch.set_num_interop_threads(1) + _test_eve_cain() + #_test_eden() diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7/scaling.py b/egs/librispeech/ASR/pruned_transducer_stateless7/scaling.py new file mode 100644 index 000000000..7ba71a94a --- /dev/null +++ b/egs/librispeech/ASR/pruned_transducer_stateless7/scaling.py @@ -0,0 +1,415 @@ +# Copyright 2022 Xiaomi Corp. (authors: Daniel Povey) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# 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 collections +from itertools import repeat +from typing import Optional, Tuple + +import torch +import torch.nn as nn +from torch import Tensor +from torch.nn import Embedding as ScaledEmbedding + + +def _ntuple(n): + def parse(x): + if isinstance(x, collections.Iterable): + return x + return tuple(repeat(x, n)) + + return parse + + +_single = _ntuple(1) +_pair = _ntuple(2) + + +class ActivationBalancerFunction(torch.autograd.Function): + @staticmethod + def forward( + ctx, + x: Tensor, + channel_dim: int, + min_positive: float, # e.g. 0.05 + max_positive: float, # e.g. 0.95 + max_factor: float, # e.g. 0.01 + min_abs: float, # e.g. 0.2 + max_abs: float, # e.g. 100.0 + ) -> Tensor: + if x.requires_grad: + if channel_dim < 0: + channel_dim += x.ndim + sum_dims = [d for d in range(x.ndim) if d != channel_dim] + xgt0 = x > 0 + proportion_positive = torch.mean( + xgt0.to(x.dtype), dim=sum_dims, keepdim=True + ) + factor1 = ( + (min_positive - proportion_positive).relu() + * (max_factor / min_positive) + if min_positive != 0.0 + else 0.0 + ) + factor2 = ( + (proportion_positive - max_positive).relu() + * (max_factor / (max_positive - 1.0)) + if max_positive != 1.0 + else 0.0 + ) + factor = factor1 + factor2 + if isinstance(factor, float): + factor = torch.zeros_like(proportion_positive) + + mean_abs = torch.mean(x.abs(), dim=sum_dims, keepdim=True) + below_threshold = mean_abs < min_abs + above_threshold = mean_abs > max_abs + + ctx.save_for_backward( + factor, xgt0, below_threshold, above_threshold + ) + ctx.max_factor = max_factor + ctx.sum_dims = sum_dims + return x + + @staticmethod + def backward( + ctx, x_grad: Tensor + ) -> Tuple[Tensor, None, None, None, None, None, None]: + factor, xgt0, below_threshold, above_threshold = ctx.saved_tensors + dtype = x_grad.dtype + scale_factor = ( + (below_threshold.to(dtype) - above_threshold.to(dtype)) + * (xgt0.to(dtype) - 0.5) + * (ctx.max_factor * 2.0) + ) + + neg_delta_grad = x_grad.abs() * (factor + scale_factor) + return x_grad - neg_delta_grad, None, None, None, None, None, None + + +class BasicNorm(torch.nn.Module): + """ + This is intended to be a simpler, and hopefully cheaper, replacement for + LayerNorm. The observation this is based on, is that Transformer-type + networks, especially with pre-norm, sometimes seem to set one of the + feature dimensions to a large constant value (e.g. 50), which "defeats" + the LayerNorm because the output magnitude is then not strongly dependent + on the other (useful) features. Presumably the weight and bias of the + LayerNorm are required to allow it to do this. + + So the idea is to introduce this large constant value as an explicit + parameter, that takes the role of the "eps" in LayerNorm, so the network + doesn't have to do this trick. We make the "eps" learnable. + + Args: + num_channels: the number of channels, e.g. 512. + channel_dim: the axis/dimension corresponding to the channel, + interprted as an offset from the input's ndim if negative. + shis is NOT the num_channels; it should typically be one of + {-2, -1, 0, 1, 2, 3}. + eps: the initial "epsilon" that we add as ballast in: + scale = ((input_vec**2).mean() + epsilon)**-0.5 + Note: our epsilon is actually large, but we keep the name + to indicate the connection with conventional LayerNorm. + learn_eps: if true, we learn epsilon; if false, we keep it + at the initial value. + """ + + def __init__( + self, + num_channels: int, + channel_dim: int = -1, # CAUTION: see documentation. + eps: float = 0.25, + learn_eps: bool = True, + ) -> None: + super(BasicNorm, self).__init__() + self.num_channels = num_channels + self.channel_dim = channel_dim + if learn_eps: + self.eps = nn.Parameter(torch.tensor(eps).log().detach()) + else: + self.register_buffer("eps", torch.tensor(eps).log().detach()) + + def forward(self, x: Tensor) -> Tensor: + assert x.shape[self.channel_dim] == self.num_channels + scales = ( + torch.mean(x ** 2, dim=self.channel_dim, keepdim=True) + + self.eps.exp() + ) ** -0.5 + return x * scales + + +class ScaledLinear(nn.Linear): + """ + A modified version of nn.Linear that gives an easy way to set the + default initial parameter scale. + + Args: + Accepts the standard args and kwargs that nn.Linear accepts + e.g. in_features, out_features, bias=False. + + initial_scale: you can override this if you want to increase + or decrease the initial magnitude of the module's output + (affects the initialization of weight_scale and bias_scale). + Another option, if you want to do something like this, is + to re-initialize the parameters. + """ + + def __init__( + self, + *args, + initial_scale: float = 1.0, + **kwargs + ): + super(ScaledLinear, self).__init__(*args, **kwargs) + with torch.no_grad(): + self.weight[:] *= initial_scale + if self.bias is not None: + torch.nn.init.uniform_(self.bias, + -0.1 * initial_scale, + 0.1 * initial_scale) + + + +class ScaledConv1d(nn.Conv1d): + # See docs for ScaledLinear + def __init__( + self, + *args, + initial_scale: float = 1.0, + **kwargs + ): + super(ScaledConv1d, self).__init__(*args, **kwargs) + with torch.no_grad(): + self.weight[:] *= initial_scale + if self.bias is not None: + torch.nn.init.uniform_(self.bias, + -0.1 * initial_scale, + 0.1 * initial_scale) + + def get_weight(self): # TODO: delete + return self.weight + + def get_bias(self): # TODO: delete + return self.bias + + +class ScaledConv2d(nn.Conv2d): + # See docs for ScaledLinear + def __init__( + self, + *args, + initial_scale: float = 1.0, + **kwargs + ): + super(ScaledConv2d, self).__init__(*args, **kwargs) + with torch.no_grad(): + self.weight[:] *= initial_scale + if self.bias is not None: + torch.nn.init.uniform_(self.bias, + -0.1 * initial_scale, + 0.1 * initial_scale) + + def get_weight(self): + return self.weight + + def get_bias(self): + return self.bias + + + +class ActivationBalancer(torch.nn.Module): + """ + Modifies the backpropped derivatives of a function to try to encourage, for + each channel, that it is positive at least a proportion `threshold` of the + time. It does this by multiplying negative derivative values by up to + (1+max_factor), and positive derivative values by up to (1-max_factor), + interpolated from 1 at the threshold to those extremal values when none + of the inputs are positive. + + + Args: + channel_dim: the dimension/axis corresponding to the channel, e.g. + -1, 0, 1, 2; will be interpreted as an offset from x.ndim if negative. + min_positive: the minimum, per channel, of the proportion of the time + that (x > 0), below which we start to modify the derivatives. + max_positive: the maximum, per channel, of the proportion of the time + that (x > 0), above which we start to modify the derivatives. + max_factor: the maximum factor by which we modify the derivatives for + either the sign constraint or the magnitude constraint; + e.g. with max_factor=0.02, the the derivatives would be multiplied by + values in the range [0.98..1.02]. + min_abs: the minimum average-absolute-value per channel, which + we allow, before we start to modify the derivatives to prevent + this. + max_abs: the maximum average-absolute-value per channel, which + we allow, before we start to modify the derivatives to prevent + this. + """ + + def __init__( + self, + channel_dim: int, + min_positive: float = 0.05, + max_positive: float = 0.95, + max_factor: float = 0.01, + min_abs: float = 0.2, + max_abs: float = 100.0, + ): + super(ActivationBalancer, self).__init__() + self.channel_dim = channel_dim + self.min_positive = min_positive + self.max_positive = max_positive + self.max_factor = max_factor + self.min_abs = min_abs + self.max_abs = max_abs + + def forward(self, x: Tensor) -> Tensor: + if torch.jit.is_scripting(): + return x + + return ActivationBalancerFunction.apply( + x, + self.channel_dim, + self.min_positive, + self.max_positive, + self.max_factor, + self.min_abs, + self.max_abs, + ) + + +class DoubleSwishFunction(torch.autograd.Function): + """ + double_swish(x) = x * torch.sigmoid(x-1) + This is a definition, originally motivated by its close numerical + similarity to swish(swish(x)), where swish(x) = x * sigmoid(x). + + Memory-efficient derivative computation: + double_swish(x) = x * s, where s(x) = torch.sigmoid(x-1) + double_swish'(x) = d/dx double_swish(x) = x * s'(x) + x' * s(x) = x * s'(x) + s(x). + Now, s'(x) = s(x) * (1-s(x)). + double_swish'(x) = x * s'(x) + s(x). + = x * s(x) * (1-s(x)) + s(x). + = double_swish(x) * (1-s(x)) + s(x) + ... so we just need to remember s(x) but not x itself. + """ + + @staticmethod + def forward(ctx, x: Tensor) -> Tensor: + x = x.detach() + s = torch.sigmoid(x - 1.0) + y = x * s + ctx.save_for_backward(s, y) + return y + + @staticmethod + def backward(ctx, y_grad: Tensor) -> Tensor: + s, y = ctx.saved_tensors + return (y * (1 - s) + s) * y_grad + + +class DoubleSwish(torch.nn.Module): + def forward(self, x: Tensor) -> Tensor: + """Return double-swish activation function which is an approximation to Swish(Swish(x)), + that we approximate closely with x * sigmoid(x-1). + """ + if torch.jit.is_scripting(): + return x * torch.sigmoid(x - 1.0) + return DoubleSwishFunction.apply(x) + + + + +def _test_activation_balancer_sign(): + probs = torch.arange(0, 1, 0.01) + N = 1000 + x = 1.0 * (torch.rand(probs.numel(), N) < probs.unsqueeze(-1)) + x = x.detach() + x.requires_grad = True + m = ActivationBalancer( + channel_dim=0, + min_positive=0.05, + max_positive=0.95, + max_factor=0.2, + min_abs=0.0, + ) + + y_grad = torch.sign(torch.randn(probs.numel(), N)) + + y = m(x) + y.backward(gradient=y_grad) + print("_test_activation_balancer_sign: x = ", x) + print("_test_activation_balancer_sign: y grad = ", y_grad) + print("_test_activation_balancer_sign: x grad = ", x.grad) + + +def _test_activation_balancer_magnitude(): + magnitudes = torch.arange(0, 1, 0.01) + N = 1000 + x = torch.sign(torch.randn(magnitudes.numel(), N)) * magnitudes.unsqueeze( + -1 + ) + x = x.detach() + x.requires_grad = True + m = ActivationBalancer( + channel_dim=0, + min_positive=0.0, + max_positive=1.0, + max_factor=0.2, + min_abs=0.2, + max_abs=0.8, + ) + + y_grad = torch.sign(torch.randn(magnitudes.numel(), N)) + + y = m(x) + y.backward(gradient=y_grad) + print("_test_activation_balancer_magnitude: x = ", x) + print("_test_activation_balancer_magnitude: y grad = ", y_grad) + print("_test_activation_balancer_magnitude: x grad = ", x.grad) + + +def _test_basic_norm(): + num_channels = 128 + m = BasicNorm(num_channels=num_channels, channel_dim=1) + + x = torch.randn(500, num_channels) + + y = m(x) + + assert y.shape == x.shape + x_rms = (x ** 2).mean().sqrt() + y_rms = (y ** 2).mean().sqrt() + print("x rms = ", x_rms) + print("y rms = ", y_rms) + assert y_rms < x_rms + assert y_rms > 0.5 * x_rms + + +def _test_double_swish_deriv(): + x = torch.randn(10, 12, dtype=torch.double) * 0.5 + x.requires_grad = True + m = DoubleSwish() + torch.autograd.gradcheck(m, x) + + +if __name__ == "__main__": + _test_activation_balancer_sign() + _test_activation_balancer_magnitude() + _test_basic_norm() + _test_double_swish_deriv() diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7/test_model.py b/egs/librispeech/ASR/pruned_transducer_stateless7/test_model.py new file mode 100755 index 000000000..b1832d0ec --- /dev/null +++ b/egs/librispeech/ASR/pruned_transducer_stateless7/test_model.py @@ -0,0 +1,50 @@ +#!/usr/bin/env python3 +# Copyright 2022 Xiaomi Corp. (authors: Fangjun Kuang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# 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. + + +""" +To run this file, do: + + cd icefall/egs/librispeech/ASR + python ./pruned_transducer_stateless4/test_model.py +""" + +import torch +from train import get_params, get_transducer_model + + +def test_model(): + params = get_params() + params.vocab_size = 500 + params.blank_id = 0 + params.context_size = 2 + params.unk_id = 2 + + model = get_transducer_model(params) + + num_param = sum([p.numel() for p in model.parameters()]) + print(f"Number of model parameters: {num_param}") + model.__class__.forward = torch.jit.ignore(model.__class__.forward) + torch.jit.script(model) + + +def main(): + test_model() + + +if __name__ == "__main__": + main() diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7/train.py b/egs/librispeech/ASR/pruned_transducer_stateless7/train.py new file mode 100755 index 000000000..2e8d64971 --- /dev/null +++ b/egs/librispeech/ASR/pruned_transducer_stateless7/train.py @@ -0,0 +1,1050 @@ +#!/usr/bin/env python3 +# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang, +# Wei Kang, +# Mingshuang Luo,) +# Zengwei Yao) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# 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. +""" +Usage: + +export CUDA_VISIBLE_DEVICES="0,1,2,3" + +./pruned_transducer_stateless4/train.py \ + --world-size 4 \ + --num-epochs 30 \ + --start-epoch 1 \ + --exp-dir pruned_transducer_stateless2/exp \ + --full-libri 1 \ + --max-duration 300 + +# For mix precision training: + +./pruned_transducer_stateless4/train.py \ + --world-size 4 \ + --num-epochs 30 \ + --start-epoch 1 \ + --use-fp16 1 \ + --exp-dir pruned_transducer_stateless2/exp \ + --full-libri 1 \ + --max-duration 550 + +""" + + +import argparse +import copy +import logging +import warnings +from pathlib import Path +from shutil import copyfile +from typing import Any, Dict, Optional, Tuple, Union + +import k2 +import optim +import sentencepiece as spm +import torch +import torch.multiprocessing as mp +import torch.nn as nn +from asr_datamodule import LibriSpeechAsrDataModule +from conformer import Conformer +from decoder import Decoder +from joiner import Joiner +from lhotse.cut import Cut +from lhotse.dataset.sampling.base import CutSampler +from lhotse.utils import fix_random_seed +from model import Transducer +from optim import Eden, Cain +from torch import Tensor +from torch.cuda.amp import GradScaler +from torch.nn.parallel import DistributedDataParallel as DDP +from torch.utils.tensorboard import SummaryWriter + +from icefall import diagnostics +from icefall.checkpoint import load_checkpoint, remove_checkpoints +from icefall.checkpoint import save_checkpoint as save_checkpoint_impl +from icefall.checkpoint import ( + save_checkpoint_with_global_batch_idx, + update_averaged_model, +) +from icefall.dist import cleanup_dist, setup_dist +from icefall.env import get_env_info +from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool + +LRSchedulerType = Union[ + torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler +] + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--world-size", + type=int, + default=1, + help="Number of GPUs for DDP training.", + ) + + parser.add_argument( + "--master-port", + type=int, + default=12354, + help="Master port to use for DDP training.", + ) + + parser.add_argument( + "--tensorboard", + type=str2bool, + default=True, + help="Should various information be logged in tensorboard.", + ) + + parser.add_argument( + "--num-epochs", + type=int, + default=30, + help="Number of epochs to train.", + ) + + parser.add_argument( + "--start-epoch", + type=int, + default=1, + help="""Resume training from this epoch. It should be positive. + If larger than 1, it will load checkpoint from + exp-dir/epoch-{start_epoch-1}.pt + """, + ) + + parser.add_argument( + "--start-batch", + type=int, + default=0, + help="""If positive, --start-epoch is ignored and + it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt + """, + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="pruned_transducer_stateless2/exp", + help="""The experiment dir. + It specifies the directory where all training related + files, e.g., checkpoints, log, etc, are saved + """, + ) + + parser.add_argument( + "--bpe-model", + type=str, + default="data/lang_bpe_500/bpe.model", + help="Path to the BPE model", + ) + + parser.add_argument( + "--initial-lr", + type=float, + default=0.003, + help="""The initial learning rate. This value should not need to be + changed.""", + ) + + parser.add_argument( + "--lr-batches", + type=float, + default=5000, + help="""Number of steps that affects how rapidly the learning rate decreases. + We suggest not to change this.""", + ) + + parser.add_argument( + "--lr-epochs", + type=float, + default=6, + help="""Number of epochs that affects how rapidly the learning rate decreases. + """, + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; " + "2 means tri-gram", + ) + + parser.add_argument( + "--prune-range", + type=int, + default=5, + help="The prune range for rnnt loss, it means how many symbols(context)" + "we are using to compute the loss", + ) + + parser.add_argument( + "--lm-scale", + type=float, + default=0.25, + help="The scale to smooth the loss with lm " + "(output of prediction network) part.", + ) + + parser.add_argument( + "--am-scale", + type=float, + default=0.0, + help="The scale to smooth the loss with am (output of encoder network)" + "part.", + ) + + parser.add_argument( + "--simple-loss-scale", + type=float, + default=0.5, + help="To get pruning ranges, we will calculate a simple version" + "loss(joiner is just addition), this simple loss also uses for" + "training (as a regularization item). We will scale the simple loss" + "with this parameter before adding to the final loss.", + ) + + parser.add_argument( + "--seed", + type=int, + default=42, + help="The seed for random generators intended for reproducibility", + ) + + parser.add_argument( + "--print-diagnostics", + type=str2bool, + default=False, + help="Accumulate stats on activations, print them and exit.", + ) + + parser.add_argument( + "--save-every-n", + type=int, + default=8000, + help="""Save checkpoint after processing this number of batches" + periodically. We save checkpoint to exp-dir/ whenever + params.batch_idx_train % save_every_n == 0. The checkpoint filename + has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt' + Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the + end of each epoch where `xxx` is the epoch number counting from 0. + """, + ) + + parser.add_argument( + "--keep-last-k", + type=int, + default=20, + help="""Only keep this number of checkpoints on disk. + For instance, if it is 3, there are only 3 checkpoints + in the exp-dir with filenames `checkpoint-xxx.pt`. + It does not affect checkpoints with name `epoch-xxx.pt`. + """, + ) + + parser.add_argument( + "--average-period", + type=int, + default=100, + help="""Update the averaged model, namely `model_avg`, after processing + this number of batches. `model_avg` is a separate version of model, + in which each floating-point parameter is the average of all the + parameters from the start of training. Each time we take the average, + we do: `model_avg = model * (average_period / batch_idx_train) + + model_avg * ((batch_idx_train - average_period) / batch_idx_train)`. + """, + ) + + parser.add_argument( + "--use-fp16", + type=str2bool, + default=False, + help="Whether to use half precision training.", + ) + + return parser + + +def get_params() -> AttributeDict: + """Return a dict containing training parameters. + + All training related parameters that are not passed from the commandline + are saved in the variable `params`. + + Commandline options are merged into `params` after they are parsed, so + you can also access them via `params`. + + Explanation of options saved in `params`: + + - best_train_loss: Best training loss so far. It is used to select + the model that has the lowest training loss. It is + updated during the training. + + - best_valid_loss: Best validation loss so far. It is used to select + the model that has the lowest validation loss. It is + updated during the training. + + - best_train_epoch: It is the epoch that has the best training loss. + + - best_valid_epoch: It is the epoch that has the best validation loss. + + - batch_idx_train: Used to writing statistics to tensorboard. It + contains number of batches trained so far across + epochs. + + - log_interval: Print training loss if batch_idx % log_interval` is 0 + + - reset_interval: Reset statistics if batch_idx % reset_interval is 0 + + - valid_interval: Run validation if batch_idx % valid_interval is 0 + + - feature_dim: The model input dim. It has to match the one used + in computing features. + + - subsampling_factor: The subsampling factor for the model. + + - encoder_dim: Hidden dim for multi-head attention model. + + - num_decoder_layers: Number of decoder layer of transformer decoder. + + - warm_step: The warm_step for Noam optimizer. + """ + params = AttributeDict( + { + "best_train_loss": float("inf"), + "best_valid_loss": float("inf"), + "best_train_epoch": -1, + "best_valid_epoch": -1, + "batch_idx_train": 0, + "log_interval": 50, + "reset_interval": 200, + "valid_interval": 3000, # For the 100h subset, use 800 + # parameters for conformer + "feature_dim": 80, + "subsampling_factor": 4, + "encoder_dim": 512, + "nhead": 8, + "dim_feedforward": 2048, + "num_encoder_layers": 12, + # parameters for decoder + "decoder_dim": 512, + # parameters for joiner + "joiner_dim": 512, + # parameters for Noam + "model_warm_step": 3000, # arg given to model, not for lrate + "env_info": get_env_info(), + } + ) + + return params + + +def get_encoder_model(params: AttributeDict) -> nn.Module: + # TODO: We can add an option to switch between Conformer and Transformer + encoder = Conformer( + num_features=params.feature_dim, + subsampling_factor=params.subsampling_factor, + d_model=params.encoder_dim, + nhead=params.nhead, + dim_feedforward=params.dim_feedforward, + num_encoder_layers=params.num_encoder_layers, + ) + return encoder + + +def get_decoder_model(params: AttributeDict) -> nn.Module: + decoder = Decoder( + vocab_size=params.vocab_size, + decoder_dim=params.decoder_dim, + blank_id=params.blank_id, + context_size=params.context_size, + ) + return decoder + + +def get_joiner_model(params: AttributeDict) -> nn.Module: + joiner = Joiner( + encoder_dim=params.encoder_dim, + decoder_dim=params.decoder_dim, + joiner_dim=params.joiner_dim, + vocab_size=params.vocab_size, + ) + return joiner + + +def get_transducer_model(params: AttributeDict) -> nn.Module: + encoder = get_encoder_model(params) + decoder = get_decoder_model(params) + joiner = get_joiner_model(params) + + model = Transducer( + encoder=encoder, + decoder=decoder, + joiner=joiner, + encoder_dim=params.encoder_dim, + decoder_dim=params.decoder_dim, + joiner_dim=params.joiner_dim, + vocab_size=params.vocab_size, + ) + return model + + +def load_checkpoint_if_available( + params: AttributeDict, + model: nn.Module, + model_avg: nn.Module = None, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[LRSchedulerType] = None, +) -> Optional[Dict[str, Any]]: + """Load checkpoint from file. + + If params.start_batch is positive, it will load the checkpoint from + `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if + params.start_epoch is larger than 1, it will load the checkpoint from + `params.start_epoch - 1`. + + Apart from loading state dict for `model` and `optimizer` it also updates + `best_train_epoch`, `best_train_loss`, `best_valid_epoch`, + and `best_valid_loss` in `params`. + + Args: + params: + The return value of :func:`get_params`. + model: + The training model. + model_avg: + The stored model averaged from the start of training. + optimizer: + The optimizer that we are using. + scheduler: + The scheduler that we are using. + Returns: + Return a dict containing previously saved training info. + """ + if params.start_batch > 0: + filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt" + elif params.start_epoch > 1: + filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt" + else: + return None + + assert filename.is_file(), f"{filename} does not exist!" + + saved_params = load_checkpoint( + filename, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + ) + + keys = [ + "best_train_epoch", + "best_valid_epoch", + "batch_idx_train", + "best_train_loss", + "best_valid_loss", + ] + for k in keys: + params[k] = saved_params[k] + + if params.start_batch > 0: + if "cur_epoch" in saved_params: + params["start_epoch"] = saved_params["cur_epoch"] + + if "cur_batch_idx" in saved_params: + params["cur_batch_idx"] = saved_params["cur_batch_idx"] + + return saved_params + + +def save_checkpoint( + params: AttributeDict, + model: Union[nn.Module, DDP], + model_avg: Optional[nn.Module] = None, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[LRSchedulerType] = None, + sampler: Optional[CutSampler] = None, + scaler: Optional[GradScaler] = None, + rank: int = 0, +) -> None: + """Save model, optimizer, scheduler and training stats to file. + + Args: + params: + It is returned by :func:`get_params`. + model: + The training model. + model_avg: + The stored model averaged from the start of training. + optimizer: + The optimizer used in the training. + sampler: + The sampler for the training dataset. + scaler: + The scaler used for mix precision training. + """ + if rank != 0: + return + filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt" + save_checkpoint_impl( + filename=filename, + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=sampler, + scaler=scaler, + rank=rank, + ) + + if params.best_train_epoch == params.cur_epoch: + best_train_filename = params.exp_dir / "best-train-loss.pt" + copyfile(src=filename, dst=best_train_filename) + + if params.best_valid_epoch == params.cur_epoch: + best_valid_filename = params.exp_dir / "best-valid-loss.pt" + copyfile(src=filename, dst=best_valid_filename) + + +def compute_loss( + params: AttributeDict, + model: Union[nn.Module, DDP], + sp: spm.SentencePieceProcessor, + batch: dict, + is_training: bool, + warmup: float = 1.0, +) -> Tuple[Tensor, MetricsTracker]: + """ + Compute CTC loss given the model and its inputs. + + Args: + params: + Parameters for training. See :func:`get_params`. + model: + The model for training. It is an instance of Conformer in our case. + batch: + A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()` + for the content in it. + is_training: + True for training. False for validation. When it is True, this + function enables autograd during computation; when it is False, it + disables autograd. + warmup: a floating point value which increases throughout training; + values >= 1.0 are fully warmed up and have all modules present. + """ + device = ( + model.device + if isinstance(model, DDP) + else next(model.parameters()).device + ) + feature = batch["inputs"] + # at entry, feature is (N, T, C) + assert feature.ndim == 3 + feature = feature.to(device) + + supervisions = batch["supervisions"] + feature_lens = supervisions["num_frames"].to(device) + + texts = batch["supervisions"]["text"] + y = sp.encode(texts, out_type=int) + y = k2.RaggedTensor(y).to(device) + + with torch.set_grad_enabled(is_training): + simple_loss, pruned_loss = model( + x=feature, + x_lens=feature_lens, + y=y, + prune_range=params.prune_range, + am_scale=params.am_scale, + lm_scale=params.lm_scale, + warmup=warmup, + ) + # after the main warmup step, we keep pruned_loss_scale small + # for the same amount of time (model_warm_step), to avoid + # overwhelming the simple_loss and causing it to diverge, + # in case it had not fully learned the alignment yet. + pruned_loss_scale = ( + 0.0 + if warmup < 1.0 + else (0.1 if warmup > 1.0 and warmup < 2.0 else 1.0) + ) + loss = ( + params.simple_loss_scale * simple_loss + + pruned_loss_scale * pruned_loss + ) + + assert loss.requires_grad == is_training + + info = MetricsTracker() + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + info["frames"] = ( + (feature_lens // params.subsampling_factor).sum().item() + ) + + # Note: We use reduction=sum while computing the loss. + info["loss"] = loss.detach().cpu().item() + info["simple_loss"] = simple_loss.detach().cpu().item() + info["pruned_loss"] = pruned_loss.detach().cpu().item() + + return loss, info + + +def compute_validation_loss( + params: AttributeDict, + model: Union[nn.Module, DDP], + sp: spm.SentencePieceProcessor, + valid_dl: torch.utils.data.DataLoader, + world_size: int = 1, +) -> MetricsTracker: + """Run the validation process.""" + model.eval() + + tot_loss = MetricsTracker() + + for batch_idx, batch in enumerate(valid_dl): + loss, loss_info = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=False, + ) + assert loss.requires_grad is False + tot_loss = tot_loss + loss_info + + if world_size > 1: + tot_loss.reduce(loss.device) + + loss_value = tot_loss["loss"] / tot_loss["frames"] + if loss_value < params.best_valid_loss: + params.best_valid_epoch = params.cur_epoch + params.best_valid_loss = loss_value + + return tot_loss + + +def train_one_epoch( + params: AttributeDict, + model: Union[nn.Module, DDP], + optimizer: torch.optim.Optimizer, + scheduler: LRSchedulerType, + sp: spm.SentencePieceProcessor, + train_dl: torch.utils.data.DataLoader, + valid_dl: torch.utils.data.DataLoader, + scaler: GradScaler, + model_avg: Optional[nn.Module] = None, + tb_writer: Optional[SummaryWriter] = None, + world_size: int = 1, + rank: int = 0, +) -> None: + """Train the model for one epoch. + + The training loss from the mean of all frames is saved in + `params.train_loss`. It runs the validation process every + `params.valid_interval` batches. + + Args: + params: + It is returned by :func:`get_params`. + model: + The model for training. + optimizer: + The optimizer we are using. + scheduler: + The learning rate scheduler, we call step() every step. + train_dl: + Dataloader for the training dataset. + valid_dl: + Dataloader for the validation dataset. + scaler: + The scaler used for mix precision training. + model_avg: + The stored model averaged from the start of training. + tb_writer: + Writer to write log messages to tensorboard. + world_size: + Number of nodes in DDP training. If it is 1, DDP is disabled. + rank: + The rank of the node in DDP training. If no DDP is used, it should + be set to 0. + """ + model.train() + + tot_loss = MetricsTracker() + + cur_batch_idx = params.get("cur_batch_idx", 0) + + for batch_idx, batch in enumerate(train_dl): + if batch_idx < cur_batch_idx: + continue + cur_batch_idx = batch_idx + + params.batch_idx_train += 1 + batch_size = len(batch["supervisions"]["text"]) + + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, loss_info = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=True, + warmup=(params.batch_idx_train / params.model_warm_step), + ) + # summary stats + tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info + + # NOTE: We use reduction==sum and loss is computed over utterances + # in the batch and there is no normalization to it so far. + scaler.scale(loss).backward() + scheduler.step_batch(params.batch_idx_train) + scaler.step(optimizer) + scaler.update() + optimizer.zero_grad() + + if params.print_diagnostics and batch_idx == 30: + return + + if ( + rank == 0 + and params.batch_idx_train > 0 + and params.batch_idx_train % params.average_period == 0 + ): + update_averaged_model( + params=params, + model_cur=model, + model_avg=model_avg, + ) + + if ( + params.batch_idx_train > 0 + and params.batch_idx_train % params.save_every_n == 0 + ): + params.cur_batch_idx = batch_idx + save_checkpoint_with_global_batch_idx( + out_dir=params.exp_dir, + global_batch_idx=params.batch_idx_train, + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + del params.cur_batch_idx + remove_checkpoints( + out_dir=params.exp_dir, + topk=params.keep_last_k, + rank=rank, + ) + + if batch_idx % params.log_interval == 0: + cur_lr = scheduler.get_last_lr()[0] + logging.info( + f"Epoch {params.cur_epoch}, " + f"batch {batch_idx}, loss[{loss_info}], " + f"tot_loss[{tot_loss}], batch size: {batch_size}, " + f"lr: {cur_lr:.2e}" + ) + + if tb_writer is not None: + tb_writer.add_scalar( + "train/learning_rate", cur_lr, params.batch_idx_train + ) + + loss_info.write_summary( + tb_writer, "train/current_", params.batch_idx_train + ) + tot_loss.write_summary( + tb_writer, "train/tot_", params.batch_idx_train + ) + + if batch_idx > 0 and batch_idx % params.valid_interval == 0: + logging.info("Computing validation loss") + valid_info = compute_validation_loss( + params=params, + model=model, + sp=sp, + valid_dl=valid_dl, + world_size=world_size, + ) + model.train() + logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}") + if tb_writer is not None: + valid_info.write_summary( + tb_writer, "train/valid_", params.batch_idx_train + ) + + loss_value = tot_loss["loss"] / tot_loss["frames"] + params.train_loss = loss_value + if params.train_loss < params.best_train_loss: + params.best_train_epoch = params.cur_epoch + params.best_train_loss = params.train_loss + + +def run(rank, world_size, args): + """ + Args: + rank: + It is a value between 0 and `world_size-1`, which is + passed automatically by `mp.spawn()` in :func:`main`. + The node with rank 0 is responsible for saving checkpoint. + world_size: + Number of GPUs for DDP training. + args: + The return value of get_parser().parse_args() + """ + params = get_params() + params.update(vars(args)) + if params.full_libri is False: + params.valid_interval = 1600 + + fix_random_seed(params.seed) + if world_size > 1: + setup_dist(rank, world_size, params.master_port) + + setup_logger(f"{params.exp_dir}/log/log-train") + logging.info("Training started") + + if args.tensorboard and rank == 0: + tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard") + else: + tb_writer = None + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", rank) + logging.info(f"Device: {device}") + + sp = spm.SentencePieceProcessor() + sp.load(params.bpe_model) + + # is defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + + logging.info(params) + + logging.info("About to create model") + model = get_transducer_model(params) + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + assert params.save_every_n >= params.average_period + model_avg: Optional[nn.Module] = None + if rank == 0: + # model_avg is only used with rank 0 + model_avg = copy.deepcopy(model).to(torch.float64) + + assert params.start_epoch > 0, params.start_epoch + checkpoints = load_checkpoint_if_available( + params=params, model=model, model_avg=model_avg + ) + + model.to(device) + if world_size > 1: + logging.info("Using DDP") + model = DDP(model, device_ids=[rank]) + + optimizer = Cain(model.parameters(), lr=params.initial_lr) + + scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs) + + if checkpoints and "optimizer" in checkpoints: + logging.info("Loading optimizer state dict") + optimizer.load_state_dict(checkpoints["optimizer"]) + + if ( + checkpoints + and "scheduler" in checkpoints + and checkpoints["scheduler"] is not None + ): + logging.info("Loading scheduler state dict") + scheduler.load_state_dict(checkpoints["scheduler"]) + + if params.print_diagnostics: + diagnostic = diagnostics.attach_diagnostics(model) + + librispeech = LibriSpeechAsrDataModule(args) + + train_cuts = librispeech.train_clean_100_cuts() + if params.full_libri: + train_cuts += librispeech.train_clean_360_cuts() + train_cuts += librispeech.train_other_500_cuts() + + def remove_short_and_long_utt(c: Cut): + # Keep only utterances with duration between 1 second and 20 seconds + # + # Caution: There is a reason to select 20.0 here. Please see + # ../local/display_manifest_statistics.py + # + # You should use ../local/display_manifest_statistics.py to get + # an utterance duration distribution for your dataset to select + # the threshold + return 1.0 <= c.duration <= 20.0 + + train_cuts = train_cuts.filter(remove_short_and_long_utt) + + if params.start_batch > 0 and checkpoints and "sampler" in checkpoints: + # We only load the sampler's state dict when it loads a checkpoint + # saved in the middle of an epoch + sampler_state_dict = checkpoints["sampler"] + else: + sampler_state_dict = None + + train_dl = librispeech.train_dataloaders( + train_cuts, sampler_state_dict=sampler_state_dict + ) + + valid_cuts = librispeech.dev_clean_cuts() + valid_cuts += librispeech.dev_other_cuts() + valid_dl = librispeech.valid_dataloaders(valid_cuts) + + if not params.print_diagnostics: + scan_pessimistic_batches_for_oom( + model=model, + train_dl=train_dl, + optimizer=optimizer, + sp=sp, + params=params, + ) + + scaler = GradScaler(enabled=params.use_fp16) + if checkpoints and "grad_scaler" in checkpoints: + logging.info("Loading grad scaler state dict") + scaler.load_state_dict(checkpoints["grad_scaler"]) + + for epoch in range(params.start_epoch, params.num_epochs + 1): + scheduler.step_epoch(epoch - 1) + fix_random_seed(params.seed + epoch - 1) + train_dl.sampler.set_epoch(epoch - 1) + + if tb_writer is not None: + tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train) + + params.cur_epoch = epoch + + train_one_epoch( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sp=sp, + train_dl=train_dl, + valid_dl=valid_dl, + scaler=scaler, + tb_writer=tb_writer, + world_size=world_size, + rank=rank, + ) + + if params.print_diagnostics: + diagnostic.print_diagnostics() + break + + save_checkpoint( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + + logging.info("Done!") + + if world_size > 1: + torch.distributed.barrier() + cleanup_dist() + + +def scan_pessimistic_batches_for_oom( + model: Union[nn.Module, DDP], + train_dl: torch.utils.data.DataLoader, + optimizer: torch.optim.Optimizer, + sp: spm.SentencePieceProcessor, + params: AttributeDict, +): + from lhotse.dataset import find_pessimistic_batches + + logging.info( + "Sanity check -- see if any of the batches in epoch 1 would cause OOM." + ) + batches, crit_values = find_pessimistic_batches(train_dl.sampler) + for criterion, cuts in batches.items(): + batch = train_dl.dataset[cuts] + try: + # warmup = 0.0 is so that the derivs for the pruned loss stay zero + # (i.e. are not remembered by the decaying-average in adam), because + # we want to avoid these params being subject to shrinkage in adam. + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, _ = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=True, + warmup=0.0, + ) + loss.backward() + optimizer.step() + optimizer.zero_grad() + except RuntimeError as e: + if "CUDA out of memory" in str(e): + logging.error( + "Your GPU ran out of memory with the current " + "max_duration setting. We recommend decreasing " + "max_duration and trying again.\n" + f"Failing criterion: {criterion} " + f"(={crit_values[criterion]}) ..." + ) + raise + + +def main(): + parser = get_parser() + LibriSpeechAsrDataModule.add_arguments(parser) + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + + world_size = args.world_size + assert world_size >= 1 + if world_size > 1: + mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True) + else: + run(rank=0, world_size=1, args=args) + + +torch.set_num_threads(1) +torch.set_num_interop_threads(1) + +if __name__ == "__main__": + main()