# noqa E501: Ported from https://github.com/BUTSpeechFIT/speakerbeam/blob/main/src/models/adapt_layers.py
# Copyright (c) 2021 Brno University of Technology
# Copyright (c) 2021 Nippon Telegraph and Telephone corporation (NTT).
# All rights reserved
# By Katerina Zmolikova, August 2021.
from functools import partial
import torch
import torch.nn as nn
[docs]def make_adapt_layer(type, indim, enrolldim, ninputs=1):
adapt_class = adaptation_layer_types.get(type)
return adapt_class(indim, enrolldim, ninputs)
[docs]def into_tuple(x):
"""Transforms tensor/list/tuple into tuple."""
if isinstance(x, list):
return tuple(x)
elif isinstance(x, torch.Tensor):
return (x,)
elif isinstance(x, tuple):
return x
else:
raise ValueError("x should be tensor, list of tuple")
[docs]def into_orig_type(x, orig_type):
"""Inverts into_tuple function."""
if orig_type is tuple:
return x
if orig_type is list:
return list(x)
if orig_type is torch.Tensor:
return x[0]
else:
assert False
[docs]class ConcatAdaptLayer(nn.Module):
def __init__(self, indim, enrolldim, ninputs=1):
super().__init__()
self.ninputs = ninputs
self.transform = nn.ModuleList(
[nn.Linear(indim + enrolldim, indim) for _ in range(ninputs)]
)
[docs] def forward(self, main, enroll):
"""ConcatAdaptLayer forward.
Args:
main: tensor or tuple or list
activations in the main neural network, which are adapted
tuple/list may be useful when we want to apply the adaptation
to both normal and skip connection at once
enroll: tensor or tuple or list
embedding extracted from enrollment
tuple/list may be useful when we want to apply the adaptation
to both normal and skip connection at once
"""
assert type(main) is type(enroll)
orig_type = type(main)
main, enroll = into_tuple(main), into_tuple(enroll)
assert len(main) == len(enroll) == self.ninputs
out = []
for transform, main0, enroll0 in zip(self.transform, main, enroll):
out.append(
transform(
torch.cat(
(main0, enroll0[:, :, None].expand(main0.shape)), dim=1
).permute(0, 2, 1)
).permute(0, 2, 1)
)
return into_orig_type(tuple(out), orig_type)
[docs]class MulAddAdaptLayer(nn.Module):
def __init__(self, indim, enrolldim, ninputs=1, do_addition=True):
super().__init__()
self.ninputs = ninputs
self.do_addition = do_addition
if do_addition:
assert enrolldim == 2 * indim, (enrolldim, indim)
else:
assert enrolldim == indim, (enrolldim, indim)
[docs] def forward(self, main, enroll):
"""MulAddAdaptLayer Forward.
Args:
main: tensor or tuple or list
activations in the main neural network, which are adapted
tuple/list may be useful when we want to apply the adaptation
to both normal and skip connection at once
enroll: tensor or tuple or list
embedding extracted from enrollment
tuple/list may be useful when we want to apply the adaptation
to both normal and skip connection at once
"""
assert type(main) is type(enroll)
orig_type = type(main)
main, enroll = into_tuple(main), into_tuple(enroll)
assert len(main) == len(enroll) == self.ninputs, (
len(main),
len(enroll),
self.ninputs,
)
out = []
for main0, enroll0 in zip(main, enroll):
if self.do_addition:
enroll0_mul, enroll0_add = torch.chunk(enroll0, 2, dim=1)
out.append(enroll0_mul[:, :, None] * main0 + enroll0_add[:, :, None])
else:
out.append(enroll0[:, :, None] * main0)
return into_orig_type(tuple(out), orig_type)
# aliases for possible adaptation layer types
adaptation_layer_types = {
"concat": ConcatAdaptLayer,
"muladd": MulAddAdaptLayer,
"mul": partial(MulAddAdaptLayer, do_addition=False),
}