Spaces:
Running
on
Zero
Running
on
Zero
File size: 5,067 Bytes
bc752b1 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 |
import random
import torch
import copy
import re
from torch import nn
from torch.nn.utils.rnn import pad_sequence
import torch.nn.functional as F
class CNNAdapter(torch.nn.Module):
def __init__(
self,
enc_out_dim: int = 512,
llm_embed_dim: int = 4096,
kernel_size: int = 5,
):
super().__init__()
self.left_padding1 = nn.ConstantPad1d((kernel_size - 1, 0), 0.0)
self.left_padding2 = nn.ConstantPad1d((kernel_size - 1, 0), 0.0)
self.conv1d1 = nn.Conv1d(enc_out_dim, 2 * enc_out_dim, kernel_size, 1, 0)
self.conv1d2 = nn.Conv1d(2 * enc_out_dim, 4 * enc_out_dim, kernel_size, 1, 0)
self.bn1 = nn.BatchNorm1d(2 * enc_out_dim, eps=1e-3, momentum=0.99)
self.bn2 = nn.BatchNorm1d(4 * enc_out_dim, eps=1e-3, momentum=0.99)
self.relu1 = nn.ReLU()
self.relu2 = nn.ReLU()
self.project = nn.Linear(4 * enc_out_dim, llm_embed_dim)
def forward(self, x, mask_pad):
"""
x: B, T, enc_out_dim
mask: (B, T) or (B, 1, T)
"""
x = x.transpose(1, 2) # B, channels, T
# mask batch padding
if mask_pad.size(2) > 0: # time > 0
x.masked_fill_(~mask_pad, 0.0)
x = self.left_padding1(x)
x = self.conv1d1(x)
x = self.bn1(x)
x = self.relu1(x)
x = self.left_padding2(x)
x = self.conv1d2(x)
x = self.bn2(x)
x = self.relu2(x)
x = x.transpose(1, 2)
x = self.project(x)
return x, mask_pad
class LinearAdapter(torch.nn.Module):
def __init__(
self,
enc_out_dim: int = 512,
llm_embed_dim: int = 4096,
):
super().__init__()
self.adpter = torch.nn.Linear(enc_out_dim, llm_embed_dim)
def forward(self, x, mask_pad):
return self.adpter(x), mask_pad
class CNNSubsampling(torch.nn.Module):
def __init__(
self,
enc_out_dim: int = 512,
llm_embed_dim: int = 4096,
kernel_size: int = 5,
activation_func: str = 'relu',
norm: str = 'batch',
):
super().__init__()
self.kernel_size = kernel_size
if enc_out_dim * 4 < llm_embed_dim:
self.left_padding1 = nn.ConstantPad1d((kernel_size - 1, 0), 0.0)
self.conv1d1 = nn.Conv1d(enc_out_dim, 2 * enc_out_dim, kernel_size, 1, 0)
self.bn1 = nn.BatchNorm1d(2 * enc_out_dim, eps=1e-3, momentum=0.99)
self.relu1 = nn.ReLU()
self.left_padding2 = nn.ConstantPad1d((kernel_size - 1, 0), 0.0)
self.conv1d2 = nn.Conv1d(2 * enc_out_dim, 4 * enc_out_dim, kernel_size, 2, 0)
self.bn2 = nn.BatchNorm1d(4 * enc_out_dim, eps=1e-3, momentum=0.99)
self.relu2 = nn.ReLU()
self.project = nn.Linear(4 * enc_out_dim, llm_embed_dim)
self.cnn_num = 2
else:
self.left_padding2 = nn.ConstantPad1d((kernel_size - 1, 0), 0.0)
self.conv1d2 = nn.Conv1d(enc_out_dim, 2 * enc_out_dim, kernel_size, 2, 0)
if norm == 'batch':
self.bn2 = nn.BatchNorm1d(2 * enc_out_dim, eps=1e-3, momentum=0.99)
elif norm == 'layer':
self.bn2 = nn.LayerNorm(2 * enc_out_dim, eps=1e-3)
if activation_func == 'gelu':
self.relu2 = nn.GELU()
else:
self.relu2 = nn.ReLU()
self.project = nn.Linear(2 * enc_out_dim, llm_embed_dim)
self.cnn_num = 1
def forward(self, x, mask_pad, cache=None, return_cache=False):
"""
x: B, T, enc_out_dim
mask: (B, T) or (B, 1, T)
"""
x = x.transpose(1, 2) # B, channels, T
# mask batch padding
if mask_pad.size(2) > 0: # time > 0
x.masked_fill_(~mask_pad, 0.0)
if self.cnn_num == 2:
if cache is None:
x = self.left_padding1(x)
else:
x = torch.cat((cache[1], x), dim=2)
if cache is not None:
cache[1] = x[:, :, 1-self.kernel_size:]
else:
cache = [None, x[:, :, 1-self.kernel_size:]]
x = self.conv1d1(x)
x = self.bn1(x)
x = self.relu1(x)
if cache is None or cache[0] is None:
x = self.left_padding2(x)
else:
x = torch.cat((cache[0], x), dim=2)
if cache is not None:
cache[0] = x[:, :, 1-self.kernel_size:]
else:
cache = [x[:, :, 1-self.kernel_size:]]
x = self.conv1d2(x)
if isinstance(self.bn2, nn.LayerNorm):
x = x.transpose(1, 2)
x = self.bn2(x)
if isinstance(self.bn2, nn.LayerNorm):
x = x.transpose(1, 2)
x = self.relu2(x)
x = x.transpose(1, 2)
x = self.project(x)
if return_cache:
return x, mask_pad[:, :, 0::2], cache
return x, mask_pad[:, :, 0::2]
|