Add a diffusion network that takes aligned text instead of MELs
This commit is contained in:
James Betker
parent
87c83e4957
commit
9100e7fa9b
312
codes/models/gpt_voice/unet_diffusion_tts.py
Normal file
312
codes/models/gpt_voice/unet_diffusion_tts.py
Normal file
|
|
@ -0,0 +1,312 @@
|
|||
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
|
||||
import torch.nn.functional as F
|
||||
|
||||
from models.gpt_voice.mini_encoder import AudioMiniEncoder, EmbeddingCombiner
|
||||
from trainer.networks import register_model
|
||||
from utils.util import get_mask_from_lengths
|
||||
|
||||
|
||||
class DiffusionTts(nn.Module):
|
||||
"""
|
||||
The full UNet model with attention and timestep embedding.
|
||||
|
||||
Customized to be conditioned on an aligned token prior.
|
||||
|
||||
:param in_channels: channels in the input Tensor.
|
||||
:param num_tokens: number of tokens (e.g. characters) which can be provided.
|
||||
: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,
|
||||
in_channels=1,
|
||||
num_tokens=30,
|
||||
out_channels=2, # mean and variance
|
||||
discrete_codes=512,
|
||||
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,
|
||||
use_scale_shift_norm=False,
|
||||
resblock_updown=False,
|
||||
use_new_attention_order=False,
|
||||
kernel_size=3,
|
||||
scale_factor=2,
|
||||
conditioning_inputs_provided=True,
|
||||
conditioning_input_dim=80,
|
||||
time_embed_dim_multiplier=4,
|
||||
only_train_dvae_connection_layers=False,
|
||||
):
|
||||
super().__init__()
|
||||
|
||||
if num_heads_upsample == -1:
|
||||
num_heads_upsample = num_heads
|
||||
|
||||
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.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 * time_embed_dim_multiplier
|
||||
self.time_embed = nn.Sequential(
|
||||
linear(model_channels, time_embed_dim),
|
||||
nn.SiLU(),
|
||||
linear(time_embed_dim, time_embed_dim),
|
||||
)
|
||||
|
||||
self.conditioning_enabled = conditioning_inputs_provided
|
||||
if conditioning_inputs_provided:
|
||||
self.contextual_embedder = AudioMiniEncoder(in_channels, time_embed_dim, base_channels=32, depth=6, resnet_blocks=1,
|
||||
attn_blocks=2, num_attn_heads=2, dropout=dropout, downsample_factor=4, kernel_size=5)
|
||||
|
||||
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.Embedding(num_tokens, ch)
|
||||
token_conditioning_block.weight.data.normal_(mean=0.0, std=.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,
|
||||
use_scale_shift_norm=use_scale_shift_norm,
|
||||
kernel_size=kernel_size,
|
||||
)
|
||||
]
|
||||
ch = int(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, factor=scale_factor
|
||||
)
|
||||
)
|
||||
)
|
||||
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,
|
||||
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, 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,
|
||||
use_scale_shift_norm=use_scale_shift_norm,
|
||||
kernel_size=kernel_size,
|
||||
)
|
||||
]
|
||||
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,
|
||||
use_new_attention_order=use_new_attention_order,
|
||||
)
|
||||
)
|
||||
if level and i == num_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, 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)),
|
||||
)
|
||||
|
||||
if only_train_dvae_connection_layers:
|
||||
for p in self.parameters():
|
||||
p.DO_NOT_TRAIN = True
|
||||
p.requires_grad = False
|
||||
for sb in token_conditioning_blocks:
|
||||
for p in sb.parameters():
|
||||
del p.DO_NOT_TRAIN
|
||||
p.requires_grad = True
|
||||
|
||||
def forward(self, x, timesteps, tokens, conditioning_input=None):
|
||||
"""
|
||||
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 tokens: an aligned text input.
|
||||
: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.
|
||||
if self.conditioning_enabled:
|
||||
assert conditioning_input is not None
|
||||
|
||||
hs = []
|
||||
emb1 = self.time_embed(timestep_embedding(timesteps, self.model_channels))
|
||||
if self.conditioning_enabled:
|
||||
emb2 = self.contextual_embedder(conditioning_input)
|
||||
emb = emb1 + emb2
|
||||
else:
|
||||
emb = emb1
|
||||
|
||||
h = x.type(self.dtype)
|
||||
for k, module in enumerate(self.input_blocks):
|
||||
if isinstance(module, nn.Embedding):
|
||||
h_tok = F.interpolate(module(tokens).permute(0,2,1), size=(h.shape[-1]), mode='nearest')
|
||||
h = h + h_tok
|
||||
else:
|
||||
h = module(h, emb)
|
||||
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_diffusion_tts(opt_net, opt):
|
||||
return DiffusionTts(**opt_net['kwargs'])
|
||||
|
||||
|
||||
# Test for ~4 second audio clip at 22050Hz
|
||||
if __name__ == '__main__':
|
||||
clip = torch.randn(2, 1, 40960)
|
||||
tok = torch.randint(0,30, (2,200))
|
||||
cond = torch.randn(2, 1, 40960)
|
||||
ts = torch.LongTensor([555, 556])
|
||||
model = DiffusionTts(32, conditioning_inputs_provided=True, time_embed_dim_multiplier=8)
|
||||
print(model(clip, ts, tok, cond).shape)
|
||||
Loading…
Reference in New Issue
Block a user