File size: 3,546 Bytes
22d5f88 |
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 |
# Copyright (c) Kyutai, all rights reserved.
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.
import typing as tp
from einops import rearrange
import torch
from torch import nn
from .conv import StreamingConv1d, StreamingConvTranspose1d
class ConvDownsample1d(nn.Module):
"""
Downsampling by some integer amount `stride` using convolutions
with a kernel size of twice the stride.
If `causal` is True, the output uses a causal convolution.
"""
def __init__(
self,
stride: int,
dimension: tp.Optional[int] = None,
causal: bool = False,
learnt: bool = False,
channel_wise: bool = False,
):
super().__init__()
self.learnt = learnt
self.channel_wise = channel_wise
groups = 1
if learnt:
assert dimension is not None, "Dimension required for learnt convolutions."
in_channels = dimension
out_channels = dimension
if channel_wise:
groups = dimension
else:
in_channels = 1
out_channels = 1
self.conv = StreamingConv1d(
in_channels,
out_channels,
kernel_size=2 * stride,
stride=stride,
causal=causal,
groups=groups,
bias=False,
pad_mode="replicate",
)
if not learnt:
actual_conv = self.conv.conv.conv
actual_conv.weight.requires_grad_(False)
actual_conv.weight.data.fill_(1.0 / (2 * stride))
def forward(self, x: torch.Tensor):
batch_size = len(x)
if not self.learnt:
x = rearrange(x, "b c t -> (b c) () t")
y = self.conv(x)
if not self.learnt:
y = rearrange(y, "(b c) () t -> b c t", b=batch_size)
return y
class ConvTrUpsample1d(nn.Module):
"""
Upsample by some integer amount `stride` using transposed convolutions.
"""
def __init__(
self,
stride: int,
dimension: tp.Optional[int] = None,
causal: bool = False,
learnt: bool = False,
channel_wise: bool = False,
):
super().__init__()
self.learnt = learnt
self.channel_wise = channel_wise
groups = 1
if learnt:
assert dimension is not None, "Dimension required for learnt convolutions."
in_channels = dimension
out_channels = dimension
if channel_wise:
groups = dimension
else:
in_channels = 1
out_channels = 1
self.convtr = StreamingConvTranspose1d(
in_channels,
out_channels,
kernel_size=2 * stride,
stride=stride,
causal=causal,
groups=groups,
bias=False,
)
if not learnt:
actual_convtr = self.convtr.convtr.convtr
actual_convtr.weight.requires_grad_(False)
actual_convtr.weight.data.fill_(1.0)
def forward(self, x: torch.Tensor):
batch_size = len(x)
if not self.learnt:
x = rearrange(x, "b c t -> (b c) () t")
y = self.convtr(x)
if not self.learnt:
x_for_normalization = torch.ones_like(x[:1])
normalization = self.convtr(x_for_normalization)
y = y / normalization
y = rearrange(y, "(b c) () t -> b c t", b=batch_size)
return y
|