DL-Art-School/codes/models/gpt_voice/unet_diffusion_tts9.py

435 lines
18 KiB
Python
Raw Normal View History

2022-03-06 03:14:36 +00:00
import functools
import random
from collections import OrderedDict
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch import autocast
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.gpt_voice.mini_encoder import AudioMiniEncoder
from scripts.audio.gen.use_diffuse_tts import ceil_multiple
from trainer.networks import register_model
2022-03-08 22:50:45 +00:00
from utils.util import checkpoint, opt_get
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,
):
super().__init__()
self.channels = channels
self.emb_channels = emb_channels
self.dropout = dropout
self.out_channels = out_channels or channels
padding = {1: 0, 3: 1, 5: 2}[kernel_size]
self.in_layers = nn.Sequential(
normalization(channels),
nn.SiLU(),
conv_nd(dims, channels, self.out_channels, 1, padding=0),
)
self.emb_layers = nn.Sequential(
nn.SiLU(),
linear(
emb_channels,
self.out_channels,
),
)
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:
self.skip_connection = conv_nd(dims, channels, self.out_channels, 1)
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]
h = h + emb_out
h = self.out_layers(h)
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,
2022-03-08 22:50:45 +00:00
model_channels=1024,
2022-03-06 03:14:36 +00:00
in_channels=1,
in_latent_channels=1024,
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,
# 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
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),
)
conditioning_dim = model_channels * 8
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))
2022-03-06 03:14:36 +00:00
self.contextual_embedder = AudioMiniEncoder(1, conditioning_dim, base_channels=32, depth=6, resnet_blocks=1,
2022-03-08 22:50:45 +00:00
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(
ResBlock(conditioning_dim, time_embed_dim, dropout, out_channels=conditioning_dim, dims=dims, kernel_size=1),
2022-03-08 22:50:45 +00:00
AttentionBlock(conditioning_dim, num_heads=num_heads, num_head_channels=num_head_channels),
ResBlock(conditioning_dim, time_embed_dim, dropout, out_channels=conditioning_dim, dims=dims, kernel_size=1),
AttentionBlock(conditioning_dim, num_heads=num_heads, num_head_channels=num_head_channels),
2022-03-06 03:14:36 +00:00
ResBlock(conditioning_dim, time_embed_dim, dropout, out_channels=conditioning_dim, dims=dims, kernel_size=1),
)
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,
)
]
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(
ch, conv_resample, dims=dims, out_channels=out_ch, factor=scale_factor, ksize=1, pad=0
)
)
)
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,
),
2022-03-08 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,
),
)
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,
)
]
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)),
)
def get_grad_norm_parameter_groups(self):
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
2022-03-08 22:50:45 +00:00
def forward(self, x, timesteps, aligned_latent, conditioning_input, conditioning_free):
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)
else:
cond_emb = self.contextual_embedder(conditioning_input)
code_emb = self.latent_converter(aligned_latent)
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.
code_emb = self.conditioning_timestep_integrator(code_emb, time_emb)
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:
h = module(h, time_emb)
hs.append(h)
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)
return out
class DiffusionTtsWrapper(nn.Module):
"""
Wraps the above module with some set-up logic such that the above module can be traced by the PyTorch JIT.
"""
def __init__(self, jit_enabled=False, **kwargs):
super().__init__()
self.jit_enabled = jit_enabled
self.jit_forward = None
self.underlying = DiffusionTts(**kwargs)
2022-03-06 03:14:36 +00:00
def forward(self, x, timesteps, aligned_latent, conditioning_input, lr_input=None, conditioning_free=False):
"""
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_latent: an aligned latent providing useful data about the sample to be produced.
: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
2022-03-08 22:50:45 +00:00
if self.underlying.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:
2022-03-08 22:50:45 +00:00
noising_factor = random.uniform(0,self.underlying.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
aligned_latent = aligned_latent.permute(0,2,1)
# 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]
cm = ceil_multiple(x.shape[-1], 2048)
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.
aligned_latent = torch.cat([aligned_latent,
self.underlying.aligned_latent_padding_embedding.repeat(x.shape[0],1,int(pc*aligned_latent.shape[-1]))], dim=-1)
with autocast(x.device.type, enabled=self.underlying.enable_fp16):
if self.jit_enabled:
if self.jit_forward is None:
self.jit_forward = torch.jit.script(self.underlying, (x, timesteps, aligned_latent, conditioning_input, conditioning_free))
out = self.jit_forward(x, timesteps, aligned_latent, conditioning_input, conditioning_free)
2022-03-06 03:14:36 +00:00
else:
2022-03-08 22:50:45 +00:00
out = self.underlying(x, timesteps, aligned_latent, conditioning_input, conditioning_free)
2022-03-06 03:14:36 +00:00
return out[:, :, :orig_x_shape]
@register_model
def register_diffusion_tts9(opt_net, opt):
2022-03-08 22:50:45 +00:00
return DiffusionTtsWrapper(**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)
cond = torch.randn(2, 1, 44000)
ts = torch.LongTensor([600, 600])
2022-03-08 22:50:45 +00:00
model = DiffusionTtsWrapper(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,
super_sampling=False)
o = model(clip, ts, aligned_latent, cond)