DL-Art-School/codes/models/audio/tts/unet_diffusion_tts9.py

537 lines
24 KiB
Python
Raw Normal View History

2022-03-06 03:14:36 +00:00
import random
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch import autocast
from x_transformers import Encoder
2022-03-06 03:14:36 +00:00
from models.diffusion.nn import timestep_embedding, normalization, zero_module, conv_nd, linear
from models.diffusion.unet_diffusion import AttentionBlock, TimestepEmbedSequential, \
Downsample, Upsample, TimestepBlock
from models.audio.tts.mini_encoder import AudioMiniEncoder
from models.audio.tts.unet_diffusion_tts7 import CheckpointedXTransformerEncoder
2022-03-06 03:14:36 +00:00
from scripts.audio.gen.use_diffuse_tts import ceil_multiple
from trainer.networks import register_model
from utils.util import checkpoint
2022-03-06 03:14:36 +00:00
def is_latent(t):
return t.dtype == torch.float
def is_sequence(t):
return t.dtype == torch.long
2022-03-06 03:14:36 +00:00
class ResBlock(TimestepBlock):
def __init__(
self,
channels,
emb_channels,
dropout,
out_channels=None,
dims=2,
kernel_size=3,
2022-03-12 22:10:35 +00:00
efficient_config=True,
2022-03-13 03:42:13 +00:00
use_scale_shift_norm=False,
2022-03-06 03:14:36 +00:00
):
super().__init__()
self.channels = channels
self.emb_channels = emb_channels
self.dropout = dropout
self.out_channels = out_channels or channels
2022-03-13 03:42:13 +00:00
self.use_scale_shift_norm = use_scale_shift_norm
2022-03-06 03:14:36 +00:00
padding = {1: 0, 3: 1, 5: 2}[kernel_size]
2022-03-12 22:10:35 +00:00
eff_kernel = 1 if efficient_config else 3
eff_padding = 0 if efficient_config else 1
2022-03-06 03:14:36 +00:00
self.in_layers = nn.Sequential(
normalization(channels),
nn.SiLU(),
2022-03-12 22:10:35 +00:00
conv_nd(dims, channels, self.out_channels, eff_kernel, padding=eff_padding),
2022-03-06 03:14:36 +00:00
)
self.emb_layers = nn.Sequential(
nn.SiLU(),
linear(
emb_channels,
2022-03-13 03:42:13 +00:00
2 * self.out_channels if use_scale_shift_norm else self.out_channels,
2022-03-06 03:14:36 +00:00
),
)
self.out_layers = nn.Sequential(
normalization(self.out_channels),
nn.SiLU(),
nn.Dropout(p=dropout),
zero_module(
conv_nd(dims, self.out_channels, self.out_channels, kernel_size, padding=padding)
),
)
if self.out_channels == channels:
self.skip_connection = nn.Identity()
else:
2022-03-12 22:10:35 +00:00
self.skip_connection = conv_nd(dims, channels, self.out_channels, eff_kernel, padding=eff_padding)
2022-03-06 03:14:36 +00:00
def forward(self, x, emb):
"""
Apply the block to a Tensor, conditioned on a timestep embedding.
:param x: an [N x C x ...] Tensor of features.
:param emb: an [N x emb_channels] Tensor of timestep embeddings.
:return: an [N x C x ...] Tensor of outputs.
"""
return checkpoint(
self._forward, x, emb
)
def _forward(self, x, emb):
h = self.in_layers(x)
emb_out = self.emb_layers(emb).type(h.dtype)
while len(emb_out.shape) < len(h.shape):
emb_out = emb_out[..., None]
2022-03-13 03:42:13 +00:00
if self.use_scale_shift_norm:
out_norm, out_rest = self.out_layers[0], self.out_layers[1:]
scale, shift = torch.chunk(emb_out, 2, dim=1)
h = out_norm(h) * (1 + scale) + shift
h = out_rest(h)
else:
h = h + emb_out
h = self.out_layers(h)
2022-03-06 03:14:36 +00:00
return self.skip_connection(x) + h
class DiffusionTts(nn.Module):
"""
The full UNet model with attention and timestep embedding.
Customized to be conditioned on an aligned prior derived from a autoregressive
GPT-style model.
:param in_channels: channels in the input Tensor.
:param in_latent_channels: channels from the input latent.
: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,
2022-03-06 03:14:36 +00:00
in_channels=1,
in_latent_channels=1024,
in_tokens=8193,
conditioning_dim_factor=8,
2022-03-13 16:25:55 +00:00
conditioning_expansion=4,
2022-03-06 03:14:36 +00:00
out_channels=2, # mean and variance
dropout=0,
# res 1, 2, 4, 8,16,32,64,128,256,512, 1K, 2K
channel_mult= (1,1.5,2, 3, 4, 6, 8, 12, 16, 24, 32, 48),
num_res_blocks=(1, 1, 1, 1, 1, 2, 2, 2, 2, 2, 2, 2),
# spec_cond: 1, 0, 0, 1, 0, 0, 1, 0, 0, 1, 0, 0)
# attn: 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1
token_conditioning_resolutions=(1,16,),
attention_resolutions=(512,1024,2048),
conv_resample=True,
dims=1,
use_fp16=False,
num_heads=1,
num_head_channels=-1,
num_heads_upsample=-1,
kernel_size=3,
scale_factor=2,
time_embed_dim_multiplier=4,
2022-03-21 23:20:05 +00:00
freeze_main_net=False,
2022-03-12 22:10:35 +00:00
efficient_convs=True, # Uses kernels with width of 1 in several places rather than 3.
2022-03-13 03:42:13 +00:00
use_scale_shift_norm=True,
2022-03-06 03:14:36 +00:00
# Parameters for regularization.
unconditioned_percentage=.1, # This implements a mechanism similar to what is used in classifier-free training.
# Parameters for super-sampling.
super_sampling=False,
super_sampling_max_noising_factor=.1,
):
super().__init__()
if num_heads_upsample == -1:
num_heads_upsample = num_heads
if super_sampling:
in_channels *= 2 # In super-sampling mode, the LR input is concatenated directly onto the input.
self.in_channels = in_channels
self.model_channels = model_channels
self.out_channels = out_channels
self.attention_resolutions = attention_resolutions
self.dropout = dropout
self.channel_mult = channel_mult
self.conv_resample = conv_resample
self.num_heads = num_heads
self.num_head_channels = num_head_channels
self.num_heads_upsample = num_heads_upsample
self.dims = dims
self.super_sampling_enabled = super_sampling
self.super_sampling_max_noising_factor = super_sampling_max_noising_factor
self.unconditioned_percentage = unconditioned_percentage
self.enable_fp16 = use_fp16
2022-03-16 03:43:14 +00:00
self.alignment_size = 2 ** (len(channel_mult)+1)
2022-03-21 23:20:05 +00:00
self.freeze_main_net = freeze_main_net
2022-03-06 03:14:36 +00:00
padding = 1 if kernel_size == 3 else 2
2022-03-12 22:10:35 +00:00
down_kernel = 1 if efficient_convs else 3
2022-03-06 03:14:36 +00:00
time_embed_dim = model_channels * time_embed_dim_multiplier
self.time_embed = nn.Sequential(
linear(model_channels, time_embed_dim),
nn.SiLU(),
linear(time_embed_dim, time_embed_dim),
)
conditioning_dim = model_channels * conditioning_dim_factor
# Either code_converter or latent_converter is used, depending on what type of conditioning data is fed.
# This model is meant to be able to be trained on both for efficiency purposes - it is far less computationally
# complex to generate tokens, while generating latents will normally mean propagating through a deep autoregressive
# transformer network.
self.code_converter = nn.Sequential(
nn.Embedding(in_tokens, conditioning_dim),
CheckpointedXTransformerEncoder(
needs_permute=False,
max_seq_len=-1,
use_pos_emb=False,
attn_layers=Encoder(
dim=conditioning_dim,
depth=3,
heads=num_heads,
ff_dropout=dropout,
attn_dropout=dropout,
use_rmsnorm=True,
ff_glu=True,
2022-03-22 17:41:32 +00:00
rotary_pos_emb=True,
)
))
2022-03-06 03:14:36 +00:00
self.latent_converter = nn.Conv1d(in_latent_channels, conditioning_dim, 1)
2022-03-08 22:50:45 +00:00
self.aligned_latent_padding_embedding = nn.Parameter(torch.randn(1,in_latent_channels,1))
2022-03-15 16:35:11 +00:00
if in_channels > 60: # It's a spectrogram.
self.contextual_embedder = nn.Sequential(nn.Conv1d(in_channels,conditioning_dim,3,padding=1,stride=2),
CheckpointedXTransformerEncoder(
needs_permute=True,
max_seq_len=-1,
use_pos_emb=False,
attn_layers=Encoder(
dim=conditioning_dim,
depth=4,
heads=num_heads,
ff_dropout=dropout,
attn_dropout=dropout,
use_rmsnorm=True,
ff_glu=True,
2022-03-22 17:41:32 +00:00
rotary_pos_emb=True,
)
))
else:
self.contextual_embedder = AudioMiniEncoder(1, conditioning_dim, base_channels=32, depth=6, resnet_blocks=1,
attn_blocks=3, num_attn_heads=8, dropout=dropout, downsample_factor=4, kernel_size=5)
2022-03-06 03:14:36 +00:00
self.conditioning_conv = nn.Conv1d(conditioning_dim*2, conditioning_dim, 1)
self.unconditioned_embedding = nn.Parameter(torch.randn(1,conditioning_dim,1))
self.conditioning_timestep_integrator = TimestepEmbedSequential(
2022-03-13 03:42:13 +00:00
ResBlock(conditioning_dim, time_embed_dim, dropout, out_channels=conditioning_dim, dims=dims, kernel_size=1, use_scale_shift_norm=use_scale_shift_norm),
2022-03-08 22:50:45 +00:00
AttentionBlock(conditioning_dim, num_heads=num_heads, num_head_channels=num_head_channels),
2022-03-13 03:42:13 +00:00
ResBlock(conditioning_dim, time_embed_dim, dropout, out_channels=conditioning_dim, dims=dims, kernel_size=1, use_scale_shift_norm=use_scale_shift_norm),
2022-03-08 22:50:45 +00:00
AttentionBlock(conditioning_dim, num_heads=num_heads, num_head_channels=num_head_channels),
2022-03-13 03:42:13 +00:00
ResBlock(conditioning_dim, time_embed_dim, dropout, out_channels=conditioning_dim, dims=dims, kernel_size=1, use_scale_shift_norm=use_scale_shift_norm),
2022-03-06 03:14:36 +00:00
)
2022-03-13 16:25:55 +00:00
self.conditioning_expansion = conditioning_expansion
2022-03-06 03:14:36 +00:00
self.input_blocks = nn.ModuleList(
[
TimestepEmbedSequential(
conv_nd(dims, in_channels, model_channels, kernel_size, padding=padding)
)
]
)
token_conditioning_blocks = []
self._feature_size = model_channels
input_block_chans = [model_channels]
ch = model_channels
ds = 1
for level, (mult, num_blocks) in enumerate(zip(channel_mult, num_res_blocks)):
if ds in token_conditioning_resolutions:
token_conditioning_block = nn.Conv1d(conditioning_dim, ch, 1)
token_conditioning_block.weight.data *= .02
self.input_blocks.append(token_conditioning_block)
token_conditioning_blocks.append(token_conditioning_block)
for _ in range(num_blocks):
layers = [
ResBlock(
ch,
time_embed_dim,
dropout,
out_channels=int(mult * model_channels),
dims=dims,
kernel_size=kernel_size,
2022-03-13 03:42:13 +00:00
efficient_config=efficient_convs,
use_scale_shift_norm=use_scale_shift_norm,
2022-03-06 03:14:36 +00:00
)
]
ch = int(mult * model_channels)
if ds in attention_resolutions:
layers.append(
AttentionBlock(
ch,
num_heads=num_heads,
num_head_channels=num_head_channels,
)
)
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(
Downsample(
2022-03-12 22:10:35 +00:00
ch, conv_resample, dims=dims, out_channels=out_ch, factor=scale_factor, ksize=down_kernel, pad=0 if down_kernel == 1 else 1
2022-03-06 03:14:36 +00:00
)
)
)
ch = out_ch
input_block_chans.append(ch)
ds *= 2
self._feature_size += ch
self.middle_block = TimestepEmbedSequential(
ResBlock(
ch,
time_embed_dim,
dropout,
dims=dims,
kernel_size=kernel_size,
2022-03-12 22:10:35 +00:00
efficient_config=efficient_convs,
2022-03-13 03:42:13 +00:00
use_scale_shift_norm=use_scale_shift_norm,
2022-03-06 03:14:36 +00:00
),
2022-03-08 22:50:45 +00:00
AttentionBlock(
ch,
num_heads=num_heads,
num_head_channels=num_head_channels,
),
2022-03-06 03:14:36 +00:00
ResBlock(
ch,
time_embed_dim,
dropout,
dims=dims,
kernel_size=kernel_size,
2022-03-12 22:10:35 +00:00
efficient_config=efficient_convs,
2022-03-13 03:42:13 +00:00
use_scale_shift_norm=use_scale_shift_norm,
2022-03-06 03:14:36 +00:00
),
)
self._feature_size += ch
self.output_blocks = nn.ModuleList([])
for level, (mult, num_blocks) in list(enumerate(zip(channel_mult, num_res_blocks)))[::-1]:
for i in range(num_blocks + 1):
ich = input_block_chans.pop()
layers = [
ResBlock(
ch + ich,
time_embed_dim,
dropout,
out_channels=int(model_channels * mult),
dims=dims,
kernel_size=kernel_size,
2022-03-12 22:10:35 +00:00
efficient_config=efficient_convs,
2022-03-13 03:42:13 +00:00
use_scale_shift_norm=use_scale_shift_norm,
2022-03-06 03:14:36 +00:00
)
]
ch = int(model_channels * mult)
if ds in attention_resolutions:
layers.append(
AttentionBlock(
ch,
num_heads=num_heads_upsample,
num_head_channels=num_head_channels,
)
)
if level and i == num_blocks:
out_ch = ch
layers.append(
Upsample(ch, conv_resample, dims=dims, out_channels=out_ch, factor=scale_factor)
)
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)),
)
2022-03-21 23:20:05 +00:00
if self.freeze_main_net:
mains = [self.time_embed, self.contextual_embedder, self.conditioning_conv, self.unconditioned_embedding, self.conditioning_timestep_integrator,
self.input_blocks, self.middle_block, self.output_blocks, self.out]
for m in mains:
for p in m.parameters():
p.requires_grad = False
p.DO_NOT_TRAIN = True
2022-03-06 03:14:36 +00:00
def get_grad_norm_parameter_groups(self):
2022-03-21 23:20:05 +00:00
if self.freeze_main_net:
return {}
2022-03-06 03:14:36 +00:00
groups = {
'minicoder': list(self.contextual_embedder.parameters()),
'input_blocks': list(self.input_blocks.parameters()),
'output_blocks': list(self.output_blocks.parameters()),
'middle_transformer': list(self.middle_block.parameters()),
}
return groups
2022-03-16 16:16:34 +00:00
def fix_alignment(self, x, aligned_conditioning):
"""
The UNet requires that the input <x> is a certain multiple of 2, defined by the UNet depth. Enforce this by
padding both <x> and <aligned_conditioning> before forward propagation and removing the padding before returning.
"""
cm = ceil_multiple(x.shape[-1], self.alignment_size)
if cm != 0:
pc = (cm-x.shape[-1])/x.shape[-1]
x = F.pad(x, (0,cm-x.shape[-1]))
# Also fix aligned_latent, which is aligned to x.
if is_latent(aligned_conditioning):
aligned_conditioning = torch.cat([aligned_conditioning,
self.aligned_latent_padding_embedding.repeat(x.shape[0], 1, int(pc * aligned_conditioning.shape[-1]))], dim=-1)
else:
aligned_conditioning = F.pad(aligned_conditioning, (0,int(pc*aligned_conditioning.shape[-1])))
return x, aligned_conditioning
def forward(self, x, timesteps, aligned_conditioning, conditioning_input, lr_input=None, conditioning_free=False):
2022-03-06 03:14:36 +00:00
"""
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 aligned_conditioning: an aligned latent or sequence of tokens providing useful data about the sample to be produced.
2022-03-06 03:14:36 +00:00
:param conditioning_input: a full-resolution audio clip that is used as a reference to the style you want decoded.
:param lr_input: for super-sampling models, a guidance audio clip at a lower sampling rate.
:param conditioning_free: When set, all conditioning inputs (including tokens and conditioning_input) will not be considered.
:return: an [N x C x ...] Tensor of outputs.
"""
assert conditioning_input is not None
if self.super_sampling_enabled:
2022-03-06 03:14:36 +00:00
assert lr_input is not None
if self.training and self.super_sampling_max_noising_factor > 0:
noising_factor = random.uniform(0,self.super_sampling_max_noising_factor)
2022-03-06 03:14:36 +00:00
lr_input = torch.randn_like(lr_input) * noising_factor + lr_input
lr_input = F.interpolate(lr_input, size=(x.shape[-1],), mode='nearest')
x = torch.cat([x, lr_input], dim=1)
2022-03-08 22:50:45 +00:00
# Shuffle aligned_latent to BxCxS format
if is_latent(aligned_conditioning):
aligned_conditioning = aligned_conditioning.permute(0, 2, 1)
2022-03-08 22:50:45 +00:00
# Fix input size to the proper multiple of 2 so we don't get alignment errors going down and back up the U-net.
orig_x_shape = x.shape[-1]
2022-03-16 16:16:34 +00:00
x, aligned_conditioning = self.fix_alignment(x, aligned_conditioning)
with autocast(x.device.type, enabled=self.enable_fp16):
hs = []
time_emb = self.time_embed(timestep_embedding(timesteps, self.model_channels))
# Note: this block does not need to repeated on inference, since it is not timestep-dependent.
if conditioning_free:
code_emb = self.unconditioned_embedding.repeat(x.shape[0], 1, 1)
2022-03-06 03:14:36 +00:00
else:
cond_emb = self.contextual_embedder(conditioning_input)
if len(cond_emb.shape) == 3: # Just take the first element.
cond_emb = cond_emb[:, :, 0]
if is_latent(aligned_conditioning):
code_emb = self.latent_converter(aligned_conditioning)
else:
code_emb = self.code_converter(aligned_conditioning)
cond_emb = cond_emb.unsqueeze(-1).repeat(1, 1, code_emb.shape[-1])
code_emb = self.conditioning_conv(torch.cat([cond_emb, code_emb], dim=1))
# Mask out the conditioning branch for whole batch elements, implementing something similar to classifier-free guidance.
if self.training and self.unconditioned_percentage > 0:
unconditioned_batches = torch.rand((code_emb.shape[0], 1, 1),
device=code_emb.device) < self.unconditioned_percentage
code_emb = torch.where(unconditioned_batches, self.unconditioned_embedding.repeat(x.shape[0], 1, 1),
code_emb)
# Everything after this comment is timestep dependent.
2022-03-13 16:25:55 +00:00
code_emb = torch.repeat_interleave(code_emb, self.conditioning_expansion, dim=-1)
code_emb = self.conditioning_timestep_integrator(code_emb, time_emb)
first = True
time_emb = time_emb.float()
h = x
for k, module in enumerate(self.input_blocks):
if isinstance(module, nn.Conv1d):
h_tok = F.interpolate(module(code_emb), size=(h.shape[-1]), mode='nearest')
h = h + h_tok
else:
with autocast(x.device.type, enabled=self.enable_fp16 and not first):
# First block has autocast disabled to allow a high precision signal to be properly vectorized.
h = module(h, time_emb)
hs.append(h)
first = False
h = self.middle_block(h, time_emb)
for module in self.output_blocks:
h = torch.cat([h, hs.pop()], dim=1)
h = module(h, time_emb)
# Last block also has autocast disabled for high-precision outputs.
h = h.float()
out = self.out(h)
2022-03-13 16:25:55 +00:00
# Involve probabilistic or possibly unused parameters in loss so we don't get DDP errors.
extraneous_addition = 0
params = [self.aligned_latent_padding_embedding, self.unconditioned_embedding] + list(self.latent_converter.parameters())
2022-03-13 16:25:55 +00:00
for p in params:
extraneous_addition = extraneous_addition + p.mean()
out = out + extraneous_addition * 0
2022-03-06 03:14:36 +00:00
return out[:, :, :orig_x_shape]
@register_model
def register_diffusion_tts9(opt_net, opt):
return DiffusionTts(**opt_net['kwargs'])
2022-03-06 03:14:36 +00:00
if __name__ == '__main__':
clip = torch.randn(2, 1, 32868)
aligned_latent = torch.randn(2,388,1024)
aligned_sequence = torch.randint(0,8192,(2,388))
2022-03-06 03:14:36 +00:00
cond = torch.randn(2, 1, 44000)
ts = torch.LongTensor([600, 600])
model = DiffusionTts(128,
2022-03-06 03:14:36 +00:00
channel_mult=[1,1.5,2, 3, 4, 6, 8],
num_res_blocks=[2, 2, 2, 2, 2, 2, 1],
token_conditioning_resolutions=[1,4,16,64],
attention_resolutions=[],
num_heads=8,
kernel_size=3,
scale_factor=2,
time_embed_dim_multiplier=4,
2022-03-12 22:10:35 +00:00
super_sampling=False,
efficient_convs=False)
# Test with latent aligned conditioning
2022-03-06 03:14:36 +00:00
o = model(clip, ts, aligned_latent, cond)
# Test with sequence aligned conditioning
o = model(clip, ts, aligned_sequence, cond)
2022-03-06 03:14:36 +00:00