forked from mrq/DL-Art-School
315 lines
15 KiB
Python
315 lines
15 KiB
Python
import random
|
|
|
|
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, TimestepBlock
|
|
from trainer.networks import register_model
|
|
from utils.util import checkpoint
|
|
|
|
|
|
def is_latent(t):
|
|
return t.dtype == torch.float
|
|
|
|
def is_sequence(t):
|
|
return t.dtype == torch.long
|
|
|
|
|
|
class ResBlock(TimestepBlock):
|
|
def __init__(
|
|
self,
|
|
channels,
|
|
emb_channels,
|
|
dropout,
|
|
out_channels=None,
|
|
dims=2,
|
|
kernel_size=3,
|
|
efficient_config=True,
|
|
use_scale_shift_norm=False,
|
|
):
|
|
super().__init__()
|
|
self.channels = channels
|
|
self.emb_channels = emb_channels
|
|
self.dropout = dropout
|
|
self.out_channels = out_channels or channels
|
|
self.use_scale_shift_norm = use_scale_shift_norm
|
|
padding = {1: 0, 3: 1, 5: 2}[kernel_size]
|
|
eff_kernel = 1 if efficient_config else 3
|
|
eff_padding = 0 if efficient_config else 1
|
|
|
|
self.in_layers = nn.Sequential(
|
|
normalization(channels),
|
|
nn.SiLU(),
|
|
conv_nd(dims, channels, self.out_channels, eff_kernel, padding=eff_padding),
|
|
)
|
|
|
|
self.emb_layers = nn.Sequential(
|
|
nn.SiLU(),
|
|
linear(
|
|
emb_channels,
|
|
2 * self.out_channels if use_scale_shift_norm else 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, eff_kernel, padding=eff_padding)
|
|
|
|
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]
|
|
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)
|
|
return self.skip_connection(x) + h
|
|
|
|
|
|
class DiffusionLayer(TimestepBlock):
|
|
def __init__(self, model_channels, dropout, num_heads):
|
|
super().__init__()
|
|
self.resblk = ResBlock(model_channels, model_channels, dropout, model_channels, dims=1, use_scale_shift_norm=True)
|
|
self.attn = AttentionBlock(model_channels, num_heads, relative_pos_embeddings=True)
|
|
|
|
def forward(self, x, time_emb):
|
|
y = self.resblk(x, time_emb)
|
|
return self.attn(y)
|
|
|
|
|
|
class DiffusionTtsFlat(nn.Module):
|
|
def __init__(
|
|
self,
|
|
model_channels=512,
|
|
num_layers=8,
|
|
in_channels=100,
|
|
in_latent_channels=512,
|
|
in_tokens=8193,
|
|
out_channels=200, # mean and variance
|
|
dropout=0,
|
|
use_fp16=False,
|
|
num_heads=16,
|
|
freeze_everything_except_autoregressive_inputs=False,
|
|
# Parameters for regularization.
|
|
layer_drop=.1,
|
|
unconditioned_percentage=.1, # This implements a mechanism similar to what is used in classifier-free training.
|
|
):
|
|
super().__init__()
|
|
|
|
self.in_channels = in_channels
|
|
self.model_channels = model_channels
|
|
self.out_channels = out_channels
|
|
self.dropout = dropout
|
|
self.num_heads = num_heads
|
|
self.unconditioned_percentage = unconditioned_percentage
|
|
self.enable_fp16 = use_fp16
|
|
self.layer_drop = layer_drop
|
|
|
|
self.inp_block = conv_nd(1, in_channels, model_channels, 3, 1, 1)
|
|
self.time_embed = nn.Sequential(
|
|
linear(model_channels, model_channels),
|
|
nn.SiLU(),
|
|
linear(model_channels, model_channels),
|
|
)
|
|
|
|
# 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_embedding = nn.Embedding(in_tokens, model_channels)
|
|
self.code_converter = nn.Sequential(
|
|
AttentionBlock(model_channels, num_heads, relative_pos_embeddings=True),
|
|
AttentionBlock(model_channels, num_heads, relative_pos_embeddings=True),
|
|
AttentionBlock(model_channels, num_heads, relative_pos_embeddings=True),
|
|
)
|
|
self.code_norm = normalization(model_channels)
|
|
self.latent_converter = nn.Conv1d(in_latent_channels, model_channels, 1)
|
|
self.contextual_embedder = nn.Sequential(nn.Conv1d(in_channels,model_channels,3,padding=1,stride=2),
|
|
nn.Conv1d(model_channels, model_channels*2,3,padding=1,stride=2),
|
|
AttentionBlock(model_channels*2, num_heads, relative_pos_embeddings=True, do_checkpoint=False),
|
|
AttentionBlock(model_channels*2, num_heads, relative_pos_embeddings=True, do_checkpoint=False),
|
|
AttentionBlock(model_channels*2, num_heads, relative_pos_embeddings=True, do_checkpoint=False),
|
|
AttentionBlock(model_channels*2, num_heads, relative_pos_embeddings=True, do_checkpoint=False),
|
|
AttentionBlock(model_channels*2, num_heads, relative_pos_embeddings=True, do_checkpoint=False))
|
|
self.unconditioned_embedding = nn.Parameter(torch.randn(1,model_channels,1))
|
|
self.conditioning_timestep_integrator = TimestepEmbedSequential(
|
|
DiffusionLayer(model_channels, dropout, num_heads),
|
|
DiffusionLayer(model_channels, dropout, num_heads),
|
|
DiffusionLayer(model_channels, dropout, num_heads),
|
|
)
|
|
self.integrating_conv = nn.Conv1d(model_channels*2, model_channels, kernel_size=1)
|
|
self.mel_head = nn.Conv1d(model_channels, in_channels, kernel_size=3, padding=1)
|
|
|
|
self.layers = nn.ModuleList([DiffusionLayer(model_channels, dropout, num_heads) for _ in range(num_layers)] +
|
|
[ResBlock(model_channels, model_channels, dropout, dims=1, use_scale_shift_norm=True) for _ in range(3)])
|
|
|
|
self.out = nn.Sequential(
|
|
normalization(model_channels),
|
|
nn.SiLU(),
|
|
zero_module(conv_nd(1, model_channels, out_channels, 3, padding=1)),
|
|
)
|
|
|
|
if freeze_everything_except_autoregressive_inputs:
|
|
for ap in list(self.latent_converter.parameters()):
|
|
ap.ALLOWED_IN_FLAT = True
|
|
for p in self.parameters():
|
|
if not hasattr(p, 'ALLOWED_IN_FLAT'):
|
|
p.requires_grad = False
|
|
p.DO_NOT_TRAIN = True
|
|
|
|
def get_grad_norm_parameter_groups(self):
|
|
groups = {
|
|
'minicoder': list(self.contextual_embedder.parameters()),
|
|
'layers': list(self.layers.parameters()),
|
|
'code_converters': list(self.code_embedding.parameters()) + list(self.code_converter.parameters()) + list(self.latent_converter.parameters()) + list(self.latent_converter.parameters()),
|
|
'timestep_integrator': list(self.conditioning_timestep_integrator.parameters()) + list(self.integrating_conv.parameters()),
|
|
'time_embed': list(self.time_embed.parameters()),
|
|
}
|
|
return groups
|
|
|
|
def timestep_independent(self, aligned_conditioning, conditioning_input, expected_seq_len, return_code_pred):
|
|
# Shuffle aligned_latent to BxCxS format
|
|
if is_latent(aligned_conditioning):
|
|
aligned_conditioning = aligned_conditioning.permute(0, 2, 1)
|
|
|
|
# Note: this block does not need to repeated on inference, since it is not timestep-dependent or x-dependent.
|
|
speech_conditioning_input = conditioning_input.unsqueeze(1) if len(
|
|
conditioning_input.shape) == 3 else conditioning_input
|
|
conds = []
|
|
for j in range(speech_conditioning_input.shape[1]):
|
|
conds.append(self.contextual_embedder(speech_conditioning_input[:, j]))
|
|
conds = torch.cat(conds, dim=-1)
|
|
cond_emb = conds.mean(dim=-1)
|
|
cond_scale, cond_shift = torch.chunk(cond_emb, 2, dim=1)
|
|
if is_latent(aligned_conditioning):
|
|
code_emb = self.latent_converter(aligned_conditioning)
|
|
else:
|
|
code_emb = self.code_embedding(aligned_conditioning).permute(0, 2, 1)
|
|
code_emb = self.code_converter(code_emb)
|
|
code_emb = self.code_norm(code_emb) * (1 + cond_scale.unsqueeze(-1)) + cond_shift.unsqueeze(-1)
|
|
|
|
unconditioned_batches = torch.zeros((code_emb.shape[0], 1, 1), device=code_emb.device)
|
|
# 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(aligned_conditioning.shape[0], 1, 1),
|
|
code_emb)
|
|
expanded_code_emb = F.interpolate(code_emb, size=expected_seq_len, mode='nearest')
|
|
|
|
if not return_code_pred:
|
|
return expanded_code_emb
|
|
else:
|
|
mel_pred = self.mel_head(expanded_code_emb)
|
|
# Multiply mel_pred by !unconditioned_branches, which drops the gradient on unconditioned branches. This is because we don't want that gradient being used to train parameters through the codes_embedder as it unbalances contributions to that network from the MSE loss.
|
|
mel_pred = mel_pred * unconditioned_batches.logical_not()
|
|
return expanded_code_emb, mel_pred
|
|
|
|
|
|
def forward(self, x, timesteps, aligned_conditioning=None, conditioning_input=None, precomputed_aligned_embeddings=None, conditioning_free=False, return_code_pred=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_conditioning: an aligned latent or sequence of tokens 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 precomputed_aligned_embeddings: Embeddings returned from self.timestep_independent()
|
|
: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 precomputed_aligned_embeddings is not None or (aligned_conditioning is not None and conditioning_input is not None)
|
|
assert not (return_code_pred and precomputed_aligned_embeddings is not None) # These two are mutually exclusive.
|
|
|
|
unused_params = []
|
|
if conditioning_free:
|
|
code_emb = self.unconditioned_embedding.repeat(x.shape[0], 1, x.shape[-1])
|
|
unused_params.extend(list(self.code_converter.parameters()) + list(self.code_embedding.parameters()))
|
|
unused_params.extend(list(self.latent_converter.parameters()))
|
|
else:
|
|
if precomputed_aligned_embeddings is not None:
|
|
code_emb = precomputed_aligned_embeddings
|
|
else:
|
|
code_emb, mel_pred = self.timestep_independent(aligned_conditioning, conditioning_input, x.shape[-1], True)
|
|
if is_latent(aligned_conditioning):
|
|
unused_params.extend(list(self.code_converter.parameters()) + list(self.code_embedding.parameters()))
|
|
else:
|
|
unused_params.extend(list(self.latent_converter.parameters()))
|
|
|
|
unused_params.append(self.unconditioned_embedding)
|
|
|
|
time_emb = self.time_embed(timestep_embedding(timesteps, self.model_channels))
|
|
code_emb = self.conditioning_timestep_integrator(code_emb, time_emb)
|
|
x = self.inp_block(x)
|
|
x = torch.cat([x, code_emb], dim=1)
|
|
x = self.integrating_conv(x)
|
|
for i, lyr in enumerate(self.layers):
|
|
# Do layer drop where applicable. Do not drop first and last layers.
|
|
if self.training and self.layer_drop > 0 and i != 0 and i != (len(self.layers)-1) and random.random() < self.layer_drop:
|
|
unused_params.extend(list(lyr.parameters()))
|
|
else:
|
|
# First and last blocks will have autocast disabled for improved precision.
|
|
with autocast(x.device.type, enabled=self.enable_fp16 and i != 0):
|
|
x = lyr(x, time_emb)
|
|
|
|
x = x.float()
|
|
out = self.out(x)
|
|
|
|
# Involve probabilistic or possibly unused parameters in loss so we don't get DDP errors.
|
|
extraneous_addition = 0
|
|
for p in unused_params:
|
|
extraneous_addition = extraneous_addition + p.mean()
|
|
out = out + extraneous_addition * 0
|
|
|
|
if return_code_pred:
|
|
return out, mel_pred
|
|
return out
|
|
|
|
|
|
@register_model
|
|
def register_diffusion_tts_flat(opt_net, opt):
|
|
return DiffusionTtsFlat(**opt_net['kwargs'])
|
|
|
|
|
|
if __name__ == '__main__':
|
|
clip = torch.randn(2, 100, 400)
|
|
aligned_latent = torch.randn(2,388,512)
|
|
aligned_sequence = torch.randint(0,8192,(2,100))
|
|
cond = torch.randn(2, 100, 400)
|
|
ts = torch.LongTensor([600, 600])
|
|
model = DiffusionTtsFlat(512, layer_drop=.3, unconditioned_percentage=.5, freeze_everything_except_autoregressive_inputs=True)
|
|
# Test with latent aligned conditioning
|
|
#o = model(clip, ts, aligned_latent, cond)
|
|
# Test with sequence aligned conditioning
|
|
o = model(clip, ts, aligned_sequence, cond)
|
|
|