import torch import torch.nn as nn import torch.nn.functional as F from models.diffusion.nn import timestep_embedding, normalization, zero_module, conv_nd, linear from models.diffusion.unet_diffusion import TimestepEmbedSequential, TimestepBlock from models.lucidrains.x_transformers import Encoder, Attention, FeedForward, RMSScaleShiftNorm, RotaryEmbedding 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 MultiGroupEmbedding(nn.Module): def __init__(self, tokens, groups, dim): super().__init__() self.m = nn.ModuleList([nn.Embedding(tokens, dim // groups) for _ in range(groups)]) def forward(self, x): h = [embedding(x[:, :, i]) for i, embedding in enumerate(self.m)] return torch.cat(h, dim=-1) 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 AttentionBlock(TimestepBlock): def __init__(self, dim, heads, dropout): super().__init__() self.attn = Attention(dim, heads=heads, causal=False, dropout=dropout, zero_init_output=False) self.ff = FeedForward(dim, mult=2, dropout=dropout, zero_init_output=True) self.rms_scale_norm = RMSScaleShiftNorm(dim) def forward(self, x, timestep_emb, rotary_emb): h = self.rms_scale_norm(x, norm_scale_shift_inp=timestep_emb) h, _, _, _ = checkpoint(self.attn, h, None, None, None, None, None, rotary_emb) h = checkpoint(self.ff, h) return h + x class TransformerDiffusion(nn.Module): """ A diffusion model composed entirely of stacks of transformer layers. Why would you do it any other way? """ def __init__( self, model_channels=512, num_layers=8, in_channels=256, in_latent_channels=512, rotary_emb_dim=32, token_count=8, in_groups=None, out_channels=512, # mean and variance 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.out_channels = out_channels self.dropout = dropout self.unconditioned_percentage = unconditioned_percentage self.enable_fp16 = use_fp16 heads = model_channels//64 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), ) self.conditioning_embedder = nn.Sequential(nn.Conv1d(in_channels, model_channels // 2, 3, padding=1, stride=2), nn.Conv1d(model_channels//2, model_channels,3,padding=1,stride=2)) self.conditioning_encoder = Encoder( dim=model_channels, depth=4, heads=heads, ff_dropout=dropout, attn_dropout=dropout, use_rmsnorm=True, ff_glu=True, rotary_pos_emb=True, ) # 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. if in_groups is None: self.embeddings = nn.Embedding(token_count, model_channels) else: self.embeddings = MultiGroupEmbedding(token_count, in_groups, model_channels) self.latent_conditioner = nn.Sequential( nn.Conv1d(in_latent_channels, model_channels, 3, padding=1), Encoder( dim=model_channels, depth=2, heads=heads, ff_dropout=dropout, attn_dropout=dropout, use_rmsnorm=True, ff_glu=True, rotary_pos_emb=True, ) ) self.latent_fade = nn.Parameter(torch.zeros(1,1,model_channels)) self.code_converter = Encoder( dim=model_channels, depth=3, heads=heads, ff_dropout=dropout, attn_dropout=dropout, use_rmsnorm=True, ff_glu=True, rotary_pos_emb=True, ) self.unconditioned_embedding = nn.Parameter(torch.randn(1,1,model_channels)) self.mel_head = nn.Conv1d(model_channels, in_channels, kernel_size=3, padding=1) self.rotary_embeddings = RotaryEmbedding(rotary_emb_dim) self.intg = nn.Linear(model_channels*2, model_channels) self.layers = TimestepRotaryEmbedSequential(*[AttentionBlock(model_channels, model_channels//64, 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): groups = { 'contextual_embedder': list(self.conditioning_embedder.parameters()), 'layers': list(self.layers.parameters()) + list(self.inp_block.parameters()), 'code_converters': list(self.embeddings.parameters()) + list(self.code_converter.parameters()) + list(self.latent_conditioner.parameters()), 'time_embed': list(self.time_embed.parameters()), } return groups def timestep_independent(self, codes, conditioning_input, expected_seq_len, prenet_latent=None, return_code_pred=False): cond_emb = self.conditioning_embedder(conditioning_input).permute(0,2,1) cond_emb = self.conditioning_encoder(cond_emb)[:, 0] code_emb = self.embeddings(codes) if prenet_latent is not None: latent_conditioning = self.latent_conditioner(prenet_latent) code_emb = code_emb + latent_conditioning * self.latent_fade 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(codes.shape[0], 1, 1), code_emb) code_emb = self.code_converter(code_emb) expanded_code_emb = F.interpolate(code_emb.permute(0,2,1), size=expected_seq_len, mode='nearest').permute(0,2,1) if not return_code_pred: return expanded_code_emb, cond_emb else: # Perform the mel_head computation on the pre-exanded code embeddings, then interpolate it separately. mel_pred = self.mel_head(code_emb.permute(0,2,1)) mel_pred = F.interpolate(mel_pred, size=expected_seq_len, mode='nearest') # 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, cond_emb, mel_pred def forward(self, x, timesteps, codes=None, conditioning_input=None, prenet_latent=None, precomputed_code_embeddings=None, precomputed_cond_embeddings=None, conditioning_free=False, return_code_pred=False): if precomputed_code_embeddings is not None: assert precomputed_cond_embeddings is not None, "Must specify both precomputed embeddings if one is specified" assert codes is None and conditioning_input is None and prenet_latent is None, "Do not provide precomputed embeddings and the other parameters. It is unclear what you want me to do here." assert not (return_code_pred and precomputed_code_embeddings is not None), "I cannot compute a code_pred output for you." unused_params = [] if not return_code_pred: unused_params.extend(list(self.mel_head.parameters())) 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_conditioner.parameters())) else: if precomputed_code_embeddings is not None: code_emb = precomputed_code_embeddings cond_emb = precomputed_cond_embeddings else: code_emb, cond_emb, mel_pred = self.timestep_independent(codes, conditioning_input, x.shape[-1], prenet_latent, True) if prenet_latent is None: unused_params.extend(list(self.latent_conditioner.parameters()) + [self.latent_fade]) unused_params.append(self.unconditioned_embedding) blk_emb = self.time_embed(timestep_embedding(timesteps, self.model_channels)) + cond_emb x = self.inp_block(x).permute(0,2,1) rotary_pos_emb = self.rotary_embeddings(x.shape[1], x.device) x = self.intg(torch.cat([x, code_emb], dim=-1)) x = self.layers(x, 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 if return_code_pred: return out, mel_pred return out @register_model def register_transformer_diffusion3(opt_net, opt): return TransformerDiffusion(**opt_net['kwargs']) if __name__ == '__main__': clip = torch.randn(2, 256, 400) aligned_latent = torch.randn(2,100,512) aligned_sequence = torch.randint(0,8,(2,100,8)) cond = torch.randn(2, 256, 400) ts = torch.LongTensor([600, 600]) model = TransformerDiffusion(512, unconditioned_percentage=.5, in_groups=8) o = model(clip, ts, aligned_sequence, cond, return_code_pred=True) #o = model(clip, ts, aligned_sequence, cond, aligned_latent)