File size: 6,833 Bytes
3978e51 |
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 |
from functools import partial
import torch
import torch.nn as nn
from transformers import UperNetForSemanticSegmentation
from utils import prefer_target_instrument
class STFT:
def __init__(self, config):
self.n_fft = config.n_fft
self.hop_length = config.hop_length
self.window = torch.hann_window(window_length=self.n_fft, periodic=True)
self.dim_f = config.dim_f
def __call__(self, x):
window = self.window.to(x.device)
batch_dims = x.shape[:-2]
c, t = x.shape[-2:]
x = x.reshape([-1, t])
x = torch.stft(
x,
n_fft=self.n_fft,
hop_length=self.hop_length,
window=window,
center=True,
return_complex=True
)
x = torch.view_as_real(x)
x = x.permute([0, 3, 1, 2])
x = x.reshape([*batch_dims, c, 2, -1, x.shape[-1]]).reshape([*batch_dims, c * 2, -1, x.shape[-1]])
return x[..., :self.dim_f, :]
def inverse(self, x):
window = self.window.to(x.device)
batch_dims = x.shape[:-3]
c, f, t = x.shape[-3:]
n = self.n_fft // 2 + 1
f_pad = torch.zeros([*batch_dims, c, n - f, t]).to(x.device)
x = torch.cat([x, f_pad], -2)
x = x.reshape([*batch_dims, c // 2, 2, n, t]).reshape([-1, 2, n, t])
x = x.permute([0, 2, 3, 1])
x = x[..., 0] + x[..., 1] * 1.j
x = torch.istft(
x,
n_fft=self.n_fft,
hop_length=self.hop_length,
window=window,
center=True
)
x = x.reshape([*batch_dims, 2, -1])
return x
def get_norm(norm_type):
def norm(c, norm_type):
if norm_type == 'BatchNorm':
return nn.BatchNorm2d(c)
elif norm_type == 'InstanceNorm':
return nn.InstanceNorm2d(c, affine=True)
elif 'GroupNorm' in norm_type:
g = int(norm_type.replace('GroupNorm', ''))
return nn.GroupNorm(num_groups=g, num_channels=c)
else:
return nn.Identity()
return partial(norm, norm_type=norm_type)
def get_act(act_type):
if act_type == 'gelu':
return nn.GELU()
elif act_type == 'relu':
return nn.ReLU()
elif act_type[:3] == 'elu':
alpha = float(act_type.replace('elu', ''))
return nn.ELU(alpha)
else:
raise Exception
class Upscale(nn.Module):
def __init__(self, in_c, out_c, scale, norm, act):
super().__init__()
self.conv = nn.Sequential(
norm(in_c),
act,
nn.ConvTranspose2d(in_channels=in_c, out_channels=out_c, kernel_size=scale, stride=scale, bias=False)
)
def forward(self, x):
return self.conv(x)
class Downscale(nn.Module):
def __init__(self, in_c, out_c, scale, norm, act):
super().__init__()
self.conv = nn.Sequential(
norm(in_c),
act,
nn.Conv2d(in_channels=in_c, out_channels=out_c, kernel_size=scale, stride=scale, bias=False)
)
def forward(self, x):
return self.conv(x)
class TFC_TDF(nn.Module):
def __init__(self, in_c, c, l, f, bn, norm, act):
super().__init__()
self.blocks = nn.ModuleList()
for i in range(l):
block = nn.Module()
block.tfc1 = nn.Sequential(
norm(in_c),
act,
nn.Conv2d(in_c, c, 3, 1, 1, bias=False),
)
block.tdf = nn.Sequential(
norm(c),
act,
nn.Linear(f, f // bn, bias=False),
norm(c),
act,
nn.Linear(f // bn, f, bias=False),
)
block.tfc2 = nn.Sequential(
norm(c),
act,
nn.Conv2d(c, c, 3, 1, 1, bias=False),
)
block.shortcut = nn.Conv2d(in_c, c, 1, 1, 0, bias=False)
self.blocks.append(block)
in_c = c
def forward(self, x):
for block in self.blocks:
s = block.shortcut(x)
x = block.tfc1(x)
x = x + block.tdf(x)
x = block.tfc2(x)
x = x + s
return x
class Swin_UperNet_Model(nn.Module):
def __init__(self, config):
super().__init__()
self.config = config
act = get_act(act_type=config.model.act)
self.num_target_instruments = len(prefer_target_instrument(config))
self.num_subbands = config.model.num_subbands
dim_c = self.num_subbands * config.audio.num_channels * 2
c = config.model.num_channels
f = config.audio.dim_f // self.num_subbands
self.first_conv = nn.Conv2d(dim_c, c, 1, 1, 0, bias=False)
self.swin_upernet_model = UperNetForSemanticSegmentation.from_pretrained("openmmlab/upernet-swin-large")
self.swin_upernet_model.auxiliary_head.classifier = nn.Conv2d(256, c, kernel_size=(1, 1), stride=(1, 1))
self.swin_upernet_model.decode_head.classifier = nn.Conv2d(512, c, kernel_size=(1, 1), stride=(1, 1))
self.swin_upernet_model.backbone.embeddings.patch_embeddings.projection = nn.Conv2d(c, 192, kernel_size=(4, 4), stride=(4, 4))
self.final_conv = nn.Sequential(
nn.Conv2d(c + dim_c, c, 1, 1, 0, bias=False),
act,
nn.Conv2d(c, self.num_target_instruments * dim_c, 1, 1, 0, bias=False)
)
self.stft = STFT(config.audio)
def cac2cws(self, x):
k = self.num_subbands
b, c, f, t = x.shape
x = x.reshape(b, c, k, f // k, t)
x = x.reshape(b, c * k, f // k, t)
return x
def cws2cac(self, x):
k = self.num_subbands
b, c, f, t = x.shape
x = x.reshape(b, c // k, k, f, t)
x = x.reshape(b, c // k, f * k, t)
return x
def forward(self, x):
x = self.stft(x)
mix = x = self.cac2cws(x)
first_conv_out = x = self.first_conv(x)
x = x.transpose(-1, -2)
x = self.swin_upernet_model(x).logits
x = x.transpose(-1, -2)
x = x * first_conv_out # reduce artifacts
x = self.final_conv(torch.cat([mix, x], 1))
x = self.cws2cac(x)
if self.num_target_instruments > 1:
b, c, f, t = x.shape
x = x.reshape(b, self.num_target_instruments, -1, f, t)
x = self.stft.inverse(x)
return x
if __name__ == "__main__":
model = UperNetForSemanticSegmentation.from_pretrained("./results/", ignore_mismatched_sizes=True)
print(model)
print(model.auxiliary_head.classifier)
print(model.decode_head.classifier)
x = torch.zeros((2, 16, 512, 512), dtype=torch.float32)
res = model(x)
print(res.logits.shape)
model.save_pretrained('./results/') |