DL-Art-School/codes/models/diffusion/unet_diffusion_vocoder.py
2021-09-01 08:33:46 -06:00

312 lines
12 KiB
Python

from models.diffusion.fp16_util import convert_module_to_f32, convert_module_to_f16
from models.diffusion.nn import timestep_embedding, normalization, zero_module, conv_nd, linear
from models.diffusion.unet_diffusion import AttentionPool2d, AttentionBlock, ResBlock, TimestepEmbedSequential, \
Downsample, Upsample
import torch
import torch.nn as nn
from trainer.networks import register_model
class DiffusionVocoder(nn.Module):
"""
The full UNet model with attention and timestep embedding.
Customized to be conditioned on a spectrogram prior.
:param in_channels: channels in the input Tensor.
:param spectrogram_channels: channels in the conditioning spectrogram.
:param model_channels: base channel count for the model.
:param out_channels: channels in the output Tensor.
:param num_res_blocks: number of residual blocks per downsample.
:param attention_resolutions: a collection of downsample rates at which
attention will take place. May be a set, list, or tuple.
For example, if this contains 4, then at 4x downsampling, attention
will be used.
:param dropout: the dropout probability.
:param channel_mult: channel multiplier for each level of the UNet.
:param conv_resample: if True, use learned convolutions for upsampling and
downsampling.
:param dims: determines if the signal is 1D, 2D, or 3D.
:param num_heads: the number of attention heads in each attention layer.
:param num_heads_channels: if specified, ignore num_heads and instead use
a fixed channel width per attention head.
:param num_heads_upsample: works with num_heads to set a different number
of heads for upsampling. Deprecated.
:param use_scale_shift_norm: use a FiLM-like conditioning mechanism.
:param resblock_updown: use residual blocks for up/downsampling.
:param use_new_attention_order: use a different attention pattern for potentially
increased efficiency.
"""
def __init__(
self,
model_channels,
num_res_blocks,
in_channels=1,
out_channels=2, # mean and variance
spectrogram_channels=80,
spectrogram_conditioning_level=3, # Level at which spectrogram conditioning is applied to the waveform.
dropout=0,
# 106496 -> 26624 -> 6656 -> 16664 -> 416 -> 104 -> 26 for ~5secs@22050Hz
channel_mult=(1, 2, 4, 8, 16, 32, 64),
attention_resolutions=(16,32,64),
conv_resample=True,
dims=1,
num_classes=None,
use_fp16=False,
num_heads=1,
num_head_channels=-1,
num_heads_upsample=-1,
use_scale_shift_norm=False,
resblock_updown=False,
use_new_attention_order=False,
kernel_size=5,
):
super().__init__()
if num_heads_upsample == -1:
num_heads_upsample = num_heads
self.in_channels = in_channels
self.spectrogram_channels = spectrogram_channels
self.model_channels = model_channels
self.out_channels = out_channels
self.num_res_blocks = num_res_blocks
self.attention_resolutions = attention_resolutions
self.dropout = dropout
self.channel_mult = channel_mult
self.conv_resample = conv_resample
self.dtype = torch.float16 if use_fp16 else torch.float32
self.num_heads = num_heads
self.num_head_channels = num_head_channels
self.num_heads_upsample = num_heads_upsample
self.dims = dims
padding = 1 if kernel_size == 3 else 2
time_embed_dim = model_channels * 4
self.time_embed = nn.Sequential(
linear(model_channels, time_embed_dim),
nn.SiLU(),
linear(time_embed_dim, time_embed_dim),
)
self.input_blocks = nn.ModuleList(
[
TimestepEmbedSequential(
conv_nd(dims, in_channels, model_channels, kernel_size, padding=padding)
)
]
)
self._feature_size = model_channels
input_block_chans = [model_channels]
ch = model_channels
ds = 1
spec_chs = channel_mult[spectrogram_conditioning_level] * model_channels
self.spectrogram_conditioner = nn.Sequential(
conv_nd(dims, self.spectrogram_channels, spec_chs, 1),
normalization(spec_chs),
nn.SiLU(),
conv_nd(dims, spec_chs, spec_chs, 1)
)
self.convergence_conv = nn.Sequential(
normalization(spec_chs*2),
nn.SiLU(),
conv_nd(dims, spec_chs*2, spec_chs*2, 1)
)
for level, mult in enumerate(channel_mult):
if level == spectrogram_conditioning_level+1:
ch *= 2 # At this level, the spectrogram is concatenated onto the input.
for _ in range(num_res_blocks):
layers = [
ResBlock(
ch,
time_embed_dim,
dropout,
out_channels=mult * model_channels,
dims=dims,
use_scale_shift_norm=use_scale_shift_norm,
kernel_size=kernel_size,
)
]
ch = mult * model_channels
if ds in attention_resolutions:
layers.append(
AttentionBlock(
ch,
num_heads=num_heads,
num_head_channels=num_head_channels,
use_new_attention_order=use_new_attention_order,
)
)
self.input_blocks.append(TimestepEmbedSequential(*layers))
self._feature_size += ch
input_block_chans.append(ch)
if level != len(channel_mult) - 1:
out_ch = ch
self.input_blocks.append(
TimestepEmbedSequential(
ResBlock(
ch,
time_embed_dim,
dropout,
out_channels=out_ch,
dims=dims,
use_scale_shift_norm=use_scale_shift_norm,
down=True,
kernel_size=kernel_size,
)
if resblock_updown
else Downsample(
ch, conv_resample, dims=dims, out_channels=out_ch
)
)
)
ch = out_ch
input_block_chans.append(ch)
ds *= 2
self._feature_size += ch
if level == spectrogram_conditioning_level:
self.input_block_injection_point = len(self.input_blocks)-1
input_block_chans[-1] *= 2
self.middle_block = TimestepEmbedSequential(
ResBlock(
ch,
time_embed_dim,
dropout,
dims=dims,
use_scale_shift_norm=use_scale_shift_norm,
kernel_size=kernel_size,
),
AttentionBlock(
ch,
num_heads=num_heads,
num_head_channels=num_head_channels,
use_new_attention_order=use_new_attention_order,
),
ResBlock(
ch,
time_embed_dim,
dropout,
dims=dims,
use_scale_shift_norm=use_scale_shift_norm,
kernel_size=kernel_size,
),
)
self._feature_size += ch
self.output_blocks = nn.ModuleList([])
for level, mult in list(enumerate(channel_mult))[::-1]:
for i in range(num_res_blocks + 1):
ich = input_block_chans.pop()
layers = [
ResBlock(
ch + ich,
time_embed_dim,
dropout,
out_channels=model_channels * mult,
dims=dims,
use_scale_shift_norm=use_scale_shift_norm,
kernel_size=kernel_size,
)
]
ch = model_channels * mult
if ds in attention_resolutions:
layers.append(
AttentionBlock(
ch,
num_heads=num_heads_upsample,
num_head_channels=num_head_channels,
use_new_attention_order=use_new_attention_order,
)
)
if level and i == num_res_blocks:
out_ch = ch
layers.append(
ResBlock(
ch,
time_embed_dim,
dropout,
out_channels=out_ch,
dims=dims,
use_scale_shift_norm=use_scale_shift_norm,
up=True,
kernel_size=kernel_size,
)
if resblock_updown
else Upsample(ch, conv_resample, dims=dims, out_channels=out_ch)
)
ds //= 2
self.output_blocks.append(TimestepEmbedSequential(*layers))
self._feature_size += ch
self.out = nn.Sequential(
normalization(ch),
nn.SiLU(),
zero_module(conv_nd(dims, model_channels, out_channels, kernel_size, padding=padding)),
)
def convert_to_fp16(self):
"""
Convert the torso of the model to float16.
"""
self.input_blocks.apply(convert_module_to_f16)
self.middle_block.apply(convert_module_to_f16)
self.output_blocks.apply(convert_module_to_f16)
def convert_to_fp32(self):
"""
Convert the torso of the model to float32.
"""
self.input_blocks.apply(convert_module_to_f32)
self.middle_block.apply(convert_module_to_f32)
self.output_blocks.apply(convert_module_to_f32)
def forward(self, x, timesteps, spectrogram):
"""
Apply the model to an input batch.
:param x: an [N x C x ...] Tensor of inputs.
:param timesteps: a 1-D batch of timesteps.
:param y: an [N] Tensor of labels, if class-conditional.
:return: an [N x C x ...] Tensor of outputs.
"""
assert x.shape[-1] % 4096 == 0 # This model operates at base//4096 at it's bottom levels, thus this requirement.
hs = []
emb = self.time_embed(timestep_embedding(timesteps, self.model_channels))
conditioning = self.spectrogram_conditioner(spectrogram)
h = x.type(self.dtype)
for k, module in enumerate(self.input_blocks):
h = module(h, emb)
if k == self.input_block_injection_point:
cond = nn.functional.interpolate(conditioning, size=h.shape[-self.dims:], mode='nearest')
h = torch.cat([h, cond], dim=1)
h = self.convergence_conv(h)
hs.append(h)
h = self.middle_block(h, emb)
for module in self.output_blocks:
h = torch.cat([h, hs.pop()], dim=1)
h = module(h, emb)
h = h.type(x.dtype)
return self.out(h)
@register_model
def register_unet_diffusion_vocoder(opt_net, opt):
return DiffusionVocoder(**opt_net['kwargs'])
# Test for ~4 second audio clip at 22050Hz
if __name__ == '__main__':
clip = torch.randn(1, 1, 81920)
spec = torch.randn(1, 80, 416)
ts = torch.LongTensor([555])
model = DiffusionVocoder(16, 2)
print(model(clip, ts, spec).shape)