opus/dnn/torch/osce/utils/layers/limited_adaptive_comb1d.py
2023-12-20 03:42:44 -05:00

222 lines
8.8 KiB
Python

"""
/* Copyright (c) 2023 Amazon
Written by Jan Buethe */
/*
Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions
are met:
- Redistributions of source code must retain the above copyright
notice, this list of conditions and the following disclaimer.
- Redistributions in binary form must reproduce the above copyright
notice, this list of conditions and the following disclaimer in the
documentation and/or other materials provided with the distribution.
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*/
"""
import torch
from torch import nn
import torch.nn.functional as F
from utils.endoscopy import write_data
class LimitedAdaptiveComb1d(nn.Module):
COUNTER = 1
def __init__(self,
kernel_size,
feature_dim,
frame_size=160,
overlap_size=40,
padding=None,
max_lag=256,
name=None,
gain_limit_db=10,
global_gain_limits_db=[-6, 6],
norm_p=2,
**kwargs):
"""
Parameters:
-----------
feature_dim : int
dimension of features from which kernels, biases and gains are computed
frame_size : int, optional
frame size, defaults to 160
overlap_size : int, optional
overlap size for filter cross-fade. Cross-fade is done on the first overlap_size samples of every frame, defaults to 40
use_bias : bool, optional
if true, biases will be added to output channels. Defaults to True
padding : List[int, int], optional
left and right padding. Defaults to [(kernel_size - 1) // 2, kernel_size - 1 - (kernel_size - 1) // 2]
max_lag : int, optional
maximal pitch lag, defaults to 256
have_a0 : bool, optional
If true, the filter coefficient a0 will be learned as a positive gain (requires in_channels == out_channels). Otherwise, a0 is set to 0. Defaults to False
name: str or None, optional
specifies a name attribute for the module. If None the name is auto generated as comb_1d_COUNT, where COUNT is an instance counter for LimitedAdaptiveComb1d
"""
super(LimitedAdaptiveComb1d, self).__init__()
self.in_channels = 1
self.out_channels = 1
self.feature_dim = feature_dim
self.kernel_size = kernel_size
self.frame_size = frame_size
self.overlap_size = overlap_size
self.max_lag = max_lag
self.limit_db = gain_limit_db
self.norm_p = norm_p
if name is None:
self.name = "limited_adaptive_comb1d_" + str(LimitedAdaptiveComb1d.COUNTER)
LimitedAdaptiveComb1d.COUNTER += 1
else:
self.name = name
# network for generating convolution weights
self.conv_kernel = nn.Linear(feature_dim, kernel_size)
# comb filter gain
self.filter_gain = nn.Linear(feature_dim, 1)
self.log_gain_limit = gain_limit_db * 0.11512925464970229
with torch.no_grad():
self.filter_gain.bias[:] = max(0.1, 4 + self.log_gain_limit)
self.global_filter_gain = nn.Linear(feature_dim, 1)
log_min, log_max = global_gain_limits_db[0] * 0.11512925464970229, global_gain_limits_db[1] * 0.11512925464970229
self.filter_gain_a = (log_max - log_min) / 2
self.filter_gain_b = (log_max + log_min) / 2
if type(padding) == type(None):
self.padding = [kernel_size // 2, kernel_size - 1 - kernel_size // 2]
else:
self.padding = padding
self.overlap_win = nn.Parameter(.5 + .5 * torch.cos((torch.arange(self.overlap_size) + 0.5) * torch.pi / overlap_size), requires_grad=False)
def forward(self, x, features, lags, debug=False):
""" adaptive 1d convolution
Parameters:
-----------
x : torch.tensor
input signal of shape (batch_size, in_channels, num_samples)
feathres : torch.tensor
frame-wise features of shape (batch_size, num_frames, feature_dim)
lags: torch.LongTensor
frame-wise lags for comb-filtering
"""
batch_size = x.size(0)
num_frames = features.size(1)
num_samples = x.size(2)
frame_size = self.frame_size
overlap_size = self.overlap_size
kernel_size = self.kernel_size
win1 = torch.flip(self.overlap_win, [0])
win2 = self.overlap_win
if num_samples // self.frame_size != num_frames:
raise ValueError('non matching sizes in AdaptiveConv1d.forward')
conv_kernels = self.conv_kernel(features).reshape((batch_size, num_frames, self.out_channels, self.in_channels, self.kernel_size))
conv_kernels = conv_kernels / (1e-6 + torch.norm(conv_kernels, p=self.norm_p, dim=-1, keepdim=True))
conv_gains = torch.exp(- torch.relu(self.filter_gain(features).permute(0, 2, 1)) + self.log_gain_limit)
# calculate gains
global_conv_gains = torch.exp(self.filter_gain_a * torch.tanh(self.global_filter_gain(features).permute(0, 2, 1)) + self.filter_gain_b)
if debug and batch_size == 1:
key = self.name + "_gains"
write_data(key, conv_gains.detach().squeeze().cpu().numpy(), 16000 // self.frame_size)
key = self.name + "_kernels"
write_data(key, conv_kernels.detach().squeeze().cpu().numpy(), 16000 // self.frame_size)
key = self.name + "_lags"
write_data(key, lags.detach().squeeze().cpu().numpy(), 16000 // self.frame_size)
key = self.name + "_global_conv_gains"
write_data(key, global_conv_gains.detach().squeeze().cpu().numpy(), 16000 // self.frame_size)
# frame-wise convolution with overlap-add
output_frames = []
overlap_mem = torch.zeros((batch_size, self.out_channels, self.overlap_size), device=x.device)
x = F.pad(x, self.padding)
x = F.pad(x, [self.max_lag, self.overlap_size])
idx = torch.arange(frame_size + kernel_size - 1 + overlap_size).to(x.device).view(1, 1, -1)
idx = torch.repeat_interleave(idx, batch_size, 0)
idx = torch.repeat_interleave(idx, self.in_channels, 1)
for i in range(num_frames):
cidx = idx + i * frame_size + self.max_lag - lags[..., i].view(batch_size, 1, 1)
xx = torch.gather(x, -1, cidx).reshape((1, batch_size * self.in_channels, -1))
new_chunk = torch.conv1d(xx, conv_kernels[:, i, ...].reshape((batch_size * self.out_channels, self.in_channels, self.kernel_size)), groups=batch_size).reshape(batch_size, self.out_channels, -1)
offset = self.max_lag + self.padding[0]
new_chunk = global_conv_gains[:, :, i : i + 1] * (new_chunk * conv_gains[:, :, i : i + 1] + x[..., offset + i * frame_size : offset + (i + 1) * frame_size + overlap_size])
# overlapping part
output_frames.append(new_chunk[:, :, : overlap_size] * win1 + overlap_mem * win2)
# non-overlapping part
output_frames.append(new_chunk[:, :, overlap_size : frame_size])
# mem for next frame
overlap_mem = new_chunk[:, :, frame_size :]
# concatenate chunks
output = torch.cat(output_frames, dim=-1)
return output
def flop_count(self, rate):
frame_rate = rate / self.frame_size
overlap = self.overlap_size
overhead = overlap / self.frame_size
count = 0
# kernel computation and filtering
count += 2 * (frame_rate * self.feature_dim * self.kernel_size)
count += 2 * (self.in_channels * self.out_channels * self.kernel_size * (1 + overhead) * rate)
count += 2 * (frame_rate * self.feature_dim * self.out_channels) + rate * (1 + overhead) * self.out_channels
# a0 computation
count += 2 * (frame_rate * self.feature_dim * self.out_channels) + rate * (1 + overhead) * self.out_channels
# windowing
count += overlap * frame_rate * 3 * self.out_channels
return count