269 lines
11 KiB
Python
269 lines
11 KiB
Python
|
import itertools
|
||
|
|
from time import time
|
||
|
|
|
||
|
|
import torch
|
||
|
|
import torch.nn as nn
|
||
|
|
import torch.nn.functional as F
|
||
|
|
|
||
|
|
from models.arch_util import ResBlock, AttentionBlock
|
||
|
|
from models.audio.music.gpt_music2 import UpperEncoder, GptMusicLower
|
||
|
|
from models.audio.music.music_quantizer2 import MusicQuantizer2
|
||
|
|
from models.audio.tts.lucidrains_dvae import DiscreteVAE
|
||
|
|
from models.diffusion.nn import timestep_embedding, normalization, zero_module, conv_nd, linear
|
||
|
|
from models.diffusion.unet_diffusion import TimestepBlock
|
||
|
|
from models.lucidrains.x_transformers import Encoder, Attention, RMSScaleShiftNorm, RotaryEmbedding, \
|
||
|
|
FeedForward
|
||
|
|
from trainer.networks import register_model
|
||
|
|
from utils.util import checkpoint, print_network
|
||
|
|
|
||
|
|
|
||
|
|
class TimestepRotaryEmbedSequential(nn.Sequential, TimestepBlock):
|
||
|
|
def forward(self, x, emb, rotary_emb):
|
||
|
|
for layer in self:
|
||
|
|
if isinstance(layer, TimestepBlock):
|
||
|
|
x = layer(x, emb, rotary_emb)
|
||
|
|
else:
|
||
|
|
x = layer(x, rotary_emb)
|
||
|
|
return x
|
||
|
|
|
||
|
|
|
||
|
|
class SubBlock(nn.Module):
|
||
|
|
def __init__(self, inp_dim, contraction_dim, heads, dropout):
|
||
|
|
super().__init__()
|
||
|
|
self.attn = Attention(inp_dim, out_dim=contraction_dim, heads=heads, dim_head=contraction_dim//heads, causal=False, dropout=dropout)
|
||
|
|
self.attnorm = nn.LayerNorm(contraction_dim)
|
||
|
|
self.ff = FeedForward(inp_dim+contraction_dim, dim_out=contraction_dim, mult=2, dropout=dropout)
|
||
|
|
self.ffnorm = nn.LayerNorm(contraction_dim)
|
||
|
|
|
||
|
|
def forward(self, x, rotary_emb):
|
||
|
|
ah, _, _, _ = checkpoint(self.attn, x, None, None, None, None, None, rotary_emb)
|
||
|
|
ah = F.gelu(self.attnorm(ah))
|
||
|
|
h = torch.cat([ah, x], dim=-1)
|
||
|
|
hf = checkpoint(self.ff, h)
|
||
|
|
hf = F.gelu(self.ffnorm(hf))
|
||
|
|
h = torch.cat([h, hf], dim=-1)
|
||
|
|
return h
|
||
|
|
|
||
|
|
|
||
|
|
class ConcatAttentionBlock(TimestepBlock):
|
||
|
|
def __init__(self, trunk_dim, contraction_dim, time_embed_dim, cond_dim_in, cond_dim_hidden, heads, dropout):
|
||
|
|
super().__init__()
|
||
|
|
self.prenorm = RMSScaleShiftNorm(trunk_dim, embed_dim=time_embed_dim, bias=False)
|
||
|
|
self.cond_project = nn.Linear(cond_dim_in, cond_dim_hidden)
|
||
|
|
self.block1 = SubBlock(trunk_dim+cond_dim_hidden, contraction_dim, heads, dropout)
|
||
|
|
self.block2 = SubBlock(trunk_dim+cond_dim_hidden+contraction_dim*2, contraction_dim, heads, dropout)
|
||
|
|
self.out = nn.Linear(contraction_dim*4, trunk_dim, bias=False)
|
||
|
|
self.out.weight.data.zero_()
|
||
|
|
|
||
|
|
def forward(self, x, cond, timestep_emb, rotary_emb):
|
||
|
|
h = self.prenorm(x, norm_scale_shift_inp=timestep_emb)
|
||
|
|
cond = self.cond_project(cond)
|
||
|
|
h = torch.cat([h, cond], dim=-1)
|
||
|
|
h = self.block1(h, rotary_emb)
|
||
|
|
h = self.block2(h, rotary_emb)
|
||
|
|
h = self.out(h[:,:,x.shape[-1]+cond.shape[-1]:])
|
||
|
|
return h + x
|
||
|
|
|
||
|
|
|
||
|
|
class TransformerDiffusionWithPointConditioning(nn.Module):
|
||
|
|
"""
|
||
|
|
A diffusion model composed entirely of stacks of transformer layers. Why would you do it any other way?
|
||
|
|
"""
|
||
|
|
def __init__(
|
||
|
|
self,
|
||
|
|
in_channels=256,
|
||
|
|
out_channels=512, # mean and variance
|
||
|
|
model_channels=1024,
|
||
|
|
contraction_dim=256,
|
||
|
|
time_embed_dim=256,
|
||
|
|
num_layers=8,
|
||
|
|
rotary_emb_dim=32,
|
||
|
|
input_cond_dim=1024,
|
||
|
|
num_heads=8,
|
||
|
|
dropout=0,
|
||
|
|
use_fp16=False,
|
||
|
|
# Parameters for regularization.
|
||
|
|
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.time_embed_dim = time_embed_dim
|
||
|
|
self.out_channels = out_channels
|
||
|
|
self.dropout = dropout
|
||
|
|
self.unconditioned_percentage = unconditioned_percentage
|
||
|
|
self.enable_fp16 = use_fp16
|
||
|
|
|
||
|
|
self.inp_block = conv_nd(1, in_channels, model_channels, 3, 1, 1)
|
||
|
|
|
||
|
|
self.time_embed = nn.Sequential(
|
||
|
|
linear(time_embed_dim, time_embed_dim),
|
||
|
|
nn.SiLU(),
|
||
|
|
linear(time_embed_dim, time_embed_dim),
|
||
|
|
)
|
||
|
|
|
||
|
|
self.unconditioned_embedding = nn.Parameter(torch.randn(1,1,model_channels))
|
||
|
|
self.rotary_embeddings = RotaryEmbedding(rotary_emb_dim)
|
||
|
|
self.layers = TimestepRotaryEmbedSequential(*[ConcatAttentionBlock(model_channels,
|
||
|
|
contraction_dim,
|
||
|
|
time_embed_dim,
|
||
|
|
cond_dim_in=input_cond_dim,
|
||
|
|
cond_dim_hidden=input_cond_dim//2,
|
||
|
|
heads=num_heads,
|
||
|
|
dropout=dropout) for _ in range(num_layers)])
|
||
|
|
|
||
|
|
self.out = nn.Sequential(
|
||
|
|
normalization(model_channels),
|
||
|
|
nn.SiLU(),
|
||
|
|
zero_module(conv_nd(1, model_channels, out_channels, 3, padding=1)),
|
||
|
|
)
|
||
|
|
|
||
|
|
self.debug_codes = {}
|
||
|
|
|
||
|
|
def get_grad_norm_parameter_groups(self):
|
||
|
|
attn1 = list(itertools.chain.from_iterable([lyr.block1.attn.parameters() for lyr in self.layers]))
|
||
|
|
attn2 = list(itertools.chain.from_iterable([lyr.block2.attn.parameters() for lyr in self.layers]))
|
||
|
|
ff1 = list(itertools.chain.from_iterable([lyr.block1.ff.parameters() for lyr in self.layers]))
|
||
|
|
ff2 = list(itertools.chain.from_iterable([lyr.block2.ff.parameters() for lyr in self.layers]))
|
||
|
|
blkout_layers = list(itertools.chain.from_iterable([lyr.out.parameters() for lyr in self.layers]))
|
||
|
|
groups = {
|
||
|
|
'prenorms': list(itertools.chain.from_iterable([lyr.prenorm.parameters() for lyr in self.layers])),
|
||
|
|
'blk1_attention_layers': attn1,
|
||
|
|
'blk2_attention_layers': attn2,
|
||
|
|
'attention_layers': attn1 + attn2,
|
||
|
|
'blk1_ff_layers': ff1,
|
||
|
|
'blk2_ff_layers': ff2,
|
||
|
|
'ff_layers': ff1 + ff2,
|
||
|
|
'block_out_layers': blkout_layers,
|
||
|
|
'rotary_embeddings': list(self.rotary_embeddings.parameters()),
|
||
|
|
'out': list(self.out.parameters()),
|
||
|
|
'x_proj': list(self.inp_block.parameters()),
|
||
|
|
'layers': list(self.layers.parameters()),
|
||
|
|
'time_embed': list(self.time_embed.parameters()),
|
||
|
|
}
|
||
|
|
return groups
|
||
|
|
|
||
|
|
def forward(self, x, timesteps, conditioning_input, conditioning_free=False):
|
||
|
|
unused_params = []
|
||
|
|
if conditioning_free:
|
||
|
|
cond = self.unconditioned_embedding.repeat(x.shape[0], x.shape[-1], 1)
|
||
|
|
else:
|
||
|
|
cond = conditioning_input
|
||
|
|
# 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((cond.shape[0], 1, 1),
|
||
|
|
device=cond.device) < self.unconditioned_percentage
|
||
|
|
cond = torch.where(unconditioned_batches, self.unconditioned_embedding.repeat(cond.shape[0], 1, 1), cond)
|
||
|
|
unused_params.append(self.unconditioned_embedding)
|
||
|
|
cond = cond.repeat(1,x.shape[-1],1)
|
||
|
|
|
||
|
|
with torch.autocast(x.device.type, enabled=self.enable_fp16):
|
||
|
|
blk_emb = self.time_embed(timestep_embedding(timesteps, self.time_embed_dim))
|
||
|
|
x = self.inp_block(x).permute(0,2,1)
|
||
|
|
|
||
|
|
rotary_pos_emb = self.rotary_embeddings(x.shape[1]+1, x.device)
|
||
|
|
for layer in self.layers:
|
||
|
|
x = checkpoint(layer, x, cond, blk_emb, rotary_pos_emb)
|
||
|
|
|
||
|
|
x = x.float().permute(0,2,1)
|
||
|
|
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
|
||
|
|
|
||
|
|
return out
|
||
|
|
|
||
|
|
|
||
|
|
class ConditioningEncoder(nn.Module):
|
||
|
|
def __init__(self,
|
||
|
|
cond_dim,
|
||
|
|
embedding_dim,
|
||
|
|
attn_blocks=6,
|
||
|
|
num_attn_heads=8,
|
||
|
|
dropout=.1,
|
||
|
|
do_checkpointing=False):
|
||
|
|
super().__init__()
|
||
|
|
attn = []
|
||
|
|
self.init = nn.Conv1d(cond_dim, embedding_dim, kernel_size=1)
|
||
|
|
self.attn = Encoder(
|
||
|
|
dim=embedding_dim,
|
||
|
|
depth=attn_blocks,
|
||
|
|
heads=num_attn_heads,
|
||
|
|
ff_dropout=dropout,
|
||
|
|
attn_dropout=dropout,
|
||
|
|
use_rmsnorm=True,
|
||
|
|
ff_glu=True,
|
||
|
|
rotary_pos_emb=True,
|
||
|
|
zero_init_branch_output=True,
|
||
|
|
ff_mult=2,
|
||
|
|
)
|
||
|
|
self.dim = embedding_dim
|
||
|
|
self.do_checkpointing = do_checkpointing
|
||
|
|
|
||
|
|
def forward(self, x):
|
||
|
|
h = self.init(x).permute(0,2,1)
|
||
|
|
h = self.attn(h).permute(0,2,1)
|
||
|
|
return h.mean(dim=2).unsqueeze(1)
|
||
|
|
|
||
|
|
|
||
|
|
class TransformerDiffusionWithConditioningEncoder(nn.Module):
|
||
|
|
def __init__(self, **kwargs):
|
||
|
|
super().__init__()
|
||
|
|
self.internal_step = 0
|
||
|
|
self.diff = TransformerDiffusionWithPointConditioning(**kwargs)
|
||
|
|
self.conditioning_encoder = ConditioningEncoder(256, kwargs['model_channels'])
|
||
|
|
|
||
|
|
def forward(self, x, timesteps, true_cheater, conditioning_input=None, disable_diversity=False, conditioning_free=False):
|
||
|
|
cond = self.conditioning_encoder(true_cheater)
|
||
|
|
diff = self.diff(x, timesteps, conditioning_input=cond, conditioning_free=conditioning_free)
|
||
|
|
return diff
|
||
|
|
|
||
|
|
def get_debug_values(self, step, __):
|
||
|
|
self.internal_step = step
|
||
|
|
return {}
|
||
|
|
|
||
|
|
def get_grad_norm_parameter_groups(self):
|
||
|
|
groups = self.diff.get_grad_norm_parameter_groups()
|
||
|
|
groups['conditioning_encoder'] = list(self.conditioning_encoder.parameters())
|
||
|
|
return groups
|
||
|
|
|
||
|
|
def before_step(self, step):
|
||
|
|
scaled_grad_parameters = list(itertools.chain.from_iterable([lyr.out.parameters() for lyr in self.diff.layers])) + \
|
||
|
|
list(itertools.chain.from_iterable([lyr.prenorm.parameters() for lyr in self.diff.layers]))
|
||
|
|
# Scale back the gradients of the blkout and prenorm layers by a constant factor. These get two orders of magnitudes
|
||
|
|
# higher gradients. Ideally we would use parameter groups, but ZeroRedundancyOptimizer makes this trickier than
|
||
|
|
# directly fiddling with the gradients.
|
||
|
|
for p in scaled_grad_parameters:
|
||
|
|
if hasattr(p, 'grad') and p.grad is not None:
|
||
|
|
p.grad *= .2
|
||
|
|
|
||
|
|
|
||
|
|
@register_model
|
||
|
|
def register_tfdpc2(opt_net, opt):
|
||
|
|
return TransformerDiffusionWithPointConditioning(**opt_net['kwargs'])
|
||
|
|
|
||
|
|
|
||
|
|
@register_model
|
||
|
|
def register_tfdpc3_with_conditioning_encoder(opt_net, opt):
|
||
|
|
return TransformerDiffusionWithConditioningEncoder(**opt_net['kwargs'])
|
||
|
|
|
||
|
|
|
||
|
|
def test_cheater_model():
|
||
|
|
clip = torch.randn(2, 256, 400)
|
||
|
|
cl = torch.randn(2, 256, 400)
|
||
|
|
ts = torch.LongTensor([600, 600])
|
||
|
|
|
||
|
|
# For music:
|
||
|
|
model = TransformerDiffusionWithConditioningEncoder(model_channels=1024)
|
||
|
|
print_network(model)
|
||
|
|
o = model(clip, ts, cl)
|
||
|
|
pg = model.get_grad_norm_parameter_groups()
|
||
|
|
|
||
|
|
|
||
|
|
if __name__ == '__main__':
|
||
|
|
test_cheater_model()
|