|
|
|
@@ -0,0 +1,163 @@ |
|
|
|
# Copyright 2022 Huawei Technologies Co., Ltd |
|
|
|
# |
|
|
|
# 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. |
|
|
|
# ============================================================================ |
|
|
|
|
|
|
|
""" TasNet """ |
|
|
|
import mindspore |
|
|
|
import mindspore.nn as nn |
|
|
|
import mindspore.ops as ops |
|
|
|
|
|
|
|
EPS = 1e-8 |
|
|
|
|
|
|
|
class TasNet(nn.Cell): |
|
|
|
""" TasNet """ |
|
|
|
def __init__(self, L, N, hidden_size, num_layers, bidirectional=False, nspk=2): |
|
|
|
super(TasNet, self).__init__() |
|
|
|
# hyper-parameter |
|
|
|
self.L = L |
|
|
|
self.N = N |
|
|
|
self.hidden_size = hidden_size |
|
|
|
self.num_layers = num_layers |
|
|
|
self.bidirectional = bidirectional |
|
|
|
self.nspk = nspk |
|
|
|
# Components |
|
|
|
self.encoder = Encoder(L, N) |
|
|
|
self.separator = Separator(N, hidden_size, num_layers, bidirectional=bidirectional, nspk=nspk) |
|
|
|
self.decoder = Decoder(N, L) |
|
|
|
for p in self.get_parameters(): |
|
|
|
if p.ndim > 1: |
|
|
|
mindspore.common.initializer.Uniform(p) |
|
|
|
|
|
|
|
def construct(self, mixture): |
|
|
|
""" |
|
|
|
Args: |
|
|
|
mixture: [B, K, L] |
|
|
|
Returns: |
|
|
|
est_source: [B, nspk, K, L] |
|
|
|
""" |
|
|
|
mixture_w, norm_coef = self.encoder(mixture) |
|
|
|
est_mask = self.separator(mixture_w) |
|
|
|
est_source = self.decoder(mixture_w, est_mask, norm_coef) |
|
|
|
return est_source |
|
|
|
|
|
|
|
|
|
|
|
class Encoder(nn.Cell): |
|
|
|
""" Encoder """ |
|
|
|
def __init__(self, L, N): |
|
|
|
super(Encoder, self).__init__() |
|
|
|
# hyper-parameter |
|
|
|
self.L = L |
|
|
|
self.N = N |
|
|
|
# Components |
|
|
|
self.conv1d_U = nn.Conv1d(L, N, kernel_size=1, stride=1, pad_mode="pad", |
|
|
|
has_bias=True, weight_init="XavierUniform") |
|
|
|
self.conv1d_V = nn.Conv1d(L, N, kernel_size=1, stride=1, pad_mode="pad", |
|
|
|
has_bias=True, weight_init="XavierUniform") |
|
|
|
self.relu = ops.ReLU() |
|
|
|
self.sigmoid = ops.Sigmoid() |
|
|
|
self.expand_dims = ops.ExpandDims() |
|
|
|
self.Norm = nn.Norm(axis=2, keep_dims=True) |
|
|
|
|
|
|
|
def construct(self, mixture): |
|
|
|
""" |
|
|
|
Args: |
|
|
|
mixture: [B, K, L] |
|
|
|
Returns: |
|
|
|
mixture_w: [B, K, N] |
|
|
|
norm_coef: [B, K, 1] |
|
|
|
""" |
|
|
|
B, K, L = mixture.shape |
|
|
|
# L2 Norm along L axis |
|
|
|
norm_coef = self.Norm(mixture) # B x K x 1 |
|
|
|
norm_mixture = mixture / (norm_coef + EPS) # B x K x L |
|
|
|
# 1-D gated conv |
|
|
|
norm_mixture = self.expand_dims(norm_mixture.view(-1, L), 2) # B*K x L x 1 |
|
|
|
conv = self.relu(self.conv1d_U(norm_mixture)) # B*K x N x 1 |
|
|
|
gate = self.sigmoid(self.conv1d_V(norm_mixture)) # B*K x N x 1 |
|
|
|
mixture_w = conv * gate # B*K x N x 1 |
|
|
|
mixture_w = mixture_w.view(B, K, self.N) # B x K x N |
|
|
|
return mixture_w, norm_coef |
|
|
|
|
|
|
|
class Separator(nn.Cell): |
|
|
|
""" Estimation of source masks """ |
|
|
|
def __init__(self, N, hidden_size, num_layers, bidirectional=False, nspk=2): |
|
|
|
super(Separator, self).__init__() |
|
|
|
# hyper-parameter |
|
|
|
self.N = N |
|
|
|
self.hidden_size = hidden_size |
|
|
|
self.num_layers = num_layers |
|
|
|
self.bidirectional = bidirectional |
|
|
|
self.nspk = nspk |
|
|
|
# Components |
|
|
|
self.layer_norm = nn.LayerNorm([N]) |
|
|
|
self.lstm = nn.LSTM(N, hidden_size, num_layers, |
|
|
|
batch_first=True, |
|
|
|
bidirectional=bidirectional) |
|
|
|
# self.fc = nn.Linear(hidden_size, nspk * N) |
|
|
|
self.fc = nn.Dense(hidden_size, nspk * N, weight_init="XavierUniform") |
|
|
|
self.new_lstm = nn.Dense(500, 512) |
|
|
|
self.softmax = ops.Softmax(axis=2) |
|
|
|
|
|
|
|
def construct(self, mixture_w): |
|
|
|
""" |
|
|
|
Args: |
|
|
|
mixture_w: [B, K, N], padded |
|
|
|
Returns: |
|
|
|
est_mask: [B, K, nspk, N] |
|
|
|
""" |
|
|
|
B, K, N = mixture_w.shape |
|
|
|
# layer norm |
|
|
|
norm_mixture_w = self.layer_norm(mixture_w) |
|
|
|
# norm_mixture_w = nn.LayerNorm(mixture_w[-1:]) |
|
|
|
# LSTM |
|
|
|
# output, _ = self.lstm(norm_mixture_w) |
|
|
|
output = norm_mixture_w |
|
|
|
# fc |
|
|
|
score = self.fc(output) # B x K x nspk*N |
|
|
|
score = score.view(B, K, self.nspk, N) |
|
|
|
# softmax |
|
|
|
est_mask = self.softmax(score) |
|
|
|
return est_mask |
|
|
|
|
|
|
|
class Decoder(nn.Cell): |
|
|
|
""" Decoder """ |
|
|
|
def __init__(self, N, L): |
|
|
|
super(Decoder, self).__init__() |
|
|
|
# hyper-parameter |
|
|
|
self.N, self.L = N, L |
|
|
|
# Components |
|
|
|
# self.basis_signals = nn.Linear(N, L, bias=False) |
|
|
|
self.basis_signals = nn.Dense(N, L, weight_init="XavierUniform") |
|
|
|
# self.basis_signals = nn.Dense(N, L) |
|
|
|
self.expand_dims = ops.ExpandDims() |
|
|
|
self.transpose = ops.Transpose() |
|
|
|
|
|
|
|
def construct(self, mixture_w, est_mask, norm_coef): |
|
|
|
""" |
|
|
|
Args: |
|
|
|
mixture_w: [B, K, N] |
|
|
|
est_mask: [B, K, nspk, N] |
|
|
|
norm_coef: [B, K, 1] |
|
|
|
Returns: |
|
|
|
est_source: [B, nspk, K, L] |
|
|
|
""" |
|
|
|
# D = W * M |
|
|
|
source_w = self.expand_dims(mixture_w, 2) * est_mask # B x K x nspk x N |
|
|
|
# S = DB |
|
|
|
est_source = self.basis_signals(source_w) # B x K x nspk x L |
|
|
|
# reverse L2 norm |
|
|
|
norm_coef = self.expand_dims(norm_coef, 2) # B x K x 1 x1 |
|
|
|
est_source = est_source * norm_coef # B x K x nspk x L |
|
|
|
est_source = self.transpose(est_source, (0, 2, 1, 3)) # B x nspk x K x L |
|
|
|
return est_source |
