File size: 6,521 Bytes
9e14838 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 | import math
from collections import OrderedDict
import fairseq
import numpy as np
import scipy.io as sio
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch import Tensor
from torch.autograd import Variable
from torch.nn.parameter import Parameter
from torch.utils import data
___author__ = "Hemlata Tak"
__email__ = "tak@eurecom.fr"
# from losses_anti_spoofing import AMSoftmax
############################
## FOR fine-tuning SSL MODEL
############################
class SSLModel(nn.Module):
def __init__(self, device):
super(SSLModel, self).__init__()
cp_path = "/change_to_path_to_pre_trained_model_XLR_300M/xlsr2_300m.pt"
model, cfg, task = fairseq.checkpoint_utils.load_model_ensemble_and_task(
[cp_path]
)
self.model = model[0]
self.device = device
self.out_dim = 1024
return
def extract_feat(self, input_data):
# put the model to GPU if it not there
if (
next(self.model.parameters()).device != input_data.device
or next(self.model.parameters()).dtype != input_data.dtype
):
self.model.to(input_data.device, dtype=input_data.dtype)
self.model.train()
if True:
# input should be in shape (batch, length)
if input_data.ndim == 3:
input_tmp = input_data[:, :, 0]
else:
input_tmp = input_data
# [batch, length, dim]
emb = self.model(input_tmp, mask=False, features_only=True)["x"]
return emb
# ---------Graph attention simple back-end------------------------#
"""
Hemlata Tak, Jee-weon Jung, Jose Patino, Madhu Kamble, Massimiliano Todisco, Nicholas Evans.
End-to-end spectro-temporal graph attention networks for speaker verification anti-spoofing and speech deepfake detection.
In Proc. Automatic Speaker Verification and Spoofing Countermeasures Challenge 2021 Interspeech 2021 satellite workshop.
"""
class GraphAttentionLayer(nn.Module):
def __init__(self, in_dim, out_dim, **kwargs):
super(GraphAttentionLayer, self).__init__()
# attention map
self.att_proj = nn.Linear(in_dim, out_dim)
self.att_weight = self._init_new_params(out_dim, 1)
# project
self.proj_with_att = nn.Linear(in_dim, out_dim)
self.proj_without_att = nn.Linear(in_dim, out_dim)
# batch norm
self.bn = nn.BatchNorm1d(out_dim)
# dropout for inputs
self.input_drop = nn.Dropout(p=0.2)
self.act = nn.SELU(inplace=True)
def forward(self, x):
"""
x :(#bs, #node, #dim)
"""
# apply input dropout
x = self.input_drop(x)
# derive attention map
att_map = self._derive_att_map(x)
# projection
x = self._project(x, att_map)
# apply batch norm
x = self._apply_BN(x)
x = self.act(x)
return x
def _pairwise_mul_nodes(self, x):
"""
Calculates pairwise multiplication of nodes.
- for attention map
x :(#bs, #node, #dim)
out_shape :(#bs, #node, #node, #dim)
"""
nb_nodes = x.size(1)
x = x.unsqueeze(2).expand(-1, -1, nb_nodes, -1)
x_mirror = x.transpose(1, 2)
return x * x_mirror
def _derive_att_map(self, x):
"""
x :(#bs, #node, #dim)
out_shape :(#bs, #node, #node, 1)
"""
att_map = self._pairwise_mul_nodes(x)
att_map = torch.tanh(
self.att_proj(att_map)
) # size: (#bs, #node, #node, #dim_out)
att_map = torch.matmul(att_map, self.att_weight) # size: (#bs, #node, #node, 1)
att_map = F.softmax(att_map, dim=-2)
return att_map
def _project(self, x, att_map):
x1 = self.proj_with_att(torch.matmul(att_map.squeeze(-1), x))
x2 = self.proj_without_att(x)
return x1 + x2
def _apply_BN(self, x):
org_size = x.size()
x = x.view(-1, org_size[-1])
x = self.bn(x)
x = x.view(org_size)
return x
def _init_new_params(self, *size):
out = nn.Parameter(torch.FloatTensor(*size))
nn.init.xavier_normal_(out)
return out
class GraphPool(nn.Module):
def __init__(self, k: float, in_dim: int, p):
super().__init__()
self.k = k
self.sigmoid = nn.Sigmoid()
self.proj = nn.Linear(in_dim, 1)
self.drop = nn.Dropout(p=p) if p > 0 else nn.Identity()
self.in_dim = in_dim
def forward(self, h):
Z = self.drop(h)
weights = self.proj(Z)
scores = self.sigmoid(weights)
new_h = self.top_k_graph(scores, h, self.k)
return new_h
def top_k_graph(self, scores, h, k):
"""
args
=====
scores: attention-based weights (#bs, #node, 1)
h: graph data (#bs, #node, #dim)
k: ratio of remaining nodes, (float)
returns
=====
h: graph pool applied data (#bs, #node', #dim)
"""
_, n_nodes, n_feat = h.size()
n_nodes = max(int(n_nodes * k), 1)
_, idx = torch.topk(scores, n_nodes, dim=1)
idx = idx.expand(-1, -1, n_feat)
h = h * scores
h = torch.gather(h, 1, idx)
return h
class Model(nn.Module):
def __init__(self, d_args, device):
super(Model, self).__init__()
# SSL model
self.device = device
self.ssl_model = SSLModel(self.device)
self.LL = nn.Linear(self.ssl_model.out_dim, 128)
self.first_bn = nn.BatchNorm1d(num_features=128)
self.selu = nn.SELU(inplace=True)
# graph module layer
self.GAT_layer = GraphAttentionLayer(128, 64)
self.proj = nn.Linear(64, 1)
self.pool = GraphPool(0.8, 64, 0.3)
# classifier head
self.proj_node = nn.Linear(53, 2)
def forward(self, x_inp, Freq_aug=False):
# SSL wav2vec 2.0 model
x_ssl_feat = self.ssl_model.extract_feat(x_inp.squeeze(-1))
x_SSL = self.LL(x_ssl_feat) # (bs,frame_number,feat_out_dim)
x_SSL = x_SSL.transpose(1, 2) # (bs,feat_out_dim,frame_number)
x = F.max_pool1d(x_SSL, (3))
x = self.first_bn(x)
x = self.selu(x)
x = self.GAT_layer(x.transpose(1, 2))
x = self.pool(x)
x = self.proj(x).flatten(1)
output = self.proj_node(x)
return output
|