import functools 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, \ Downsample, Upsample, TimestepBlock from models.audio.tts.mini_encoder import AudioMiniEncoder from scripts.audio.gen.use_diffuse_tts import ceil_multiple from trainer.networks import register_model from utils.util import checkpoint from x_transformers import Encoder, ContinuousTransformerWrapper def clustered_mask(probability, shape, dev, lateral_expansion_radius_max=3, inverted=False): """ Produces a masking vector of the specified shape where each element has probability to be zero. lateral_expansion_radius_max neighbors of any element that is zero also have a 50% chance to be zero. Effectively, this produces clusters of masks tending to be lateral_expansion_radius_max wide. """ # Each masked token spreads out to 1+lateral_expansion_radius_max on average, therefore reduce the probability in # kind probability = probability / (1+lateral_expansion_radius_max) mask = torch.rand(shape, device=dev) mask = (mask < probability).float() kernel = torch.tensor([.5 for _ in range(lateral_expansion_radius_max)] + [1] + [.5 for _ in range(lateral_expansion_radius_max)], device=dev) mask = F.conv1d(mask.unsqueeze(1), kernel.view(1,1,2*lateral_expansion_radius_max+1), padding=lateral_expansion_radius_max).squeeze(1) if inverted: return torch.bernoulli(torch.clamp(mask, 0, 1)) != 0 else: return torch.bernoulli(torch.clamp(mask, 0, 1)) == 0 class CheckpointedLayer(nn.Module): """ Wraps a module. When forward() is called, passes kwargs that require_grad through torch.checkpoint() and bypasses checkpoint for all other args. """ def __init__(self, wrap): super().__init__() self.wrap = wrap def forward(self, x, *args, **kwargs): for k, v in kwargs.items(): assert not (isinstance(v, torch.Tensor) and v.requires_grad) # This would screw up checkpointing. partial = functools.partial(self.wrap, **kwargs) return torch.utils.checkpoint.checkpoint(partial, x, *args) class CheckpointedXTransformerEncoder(nn.Module): """ Wraps a ContinuousTransformerWrapper and applies CheckpointedLayer to each layer and permutes from channels-mid to channels-last that XTransformer expects. """ def __init__(self, needs_permute=True, **xtransformer_kwargs): super().__init__() self.transformer = ContinuousTransformerWrapper(**xtransformer_kwargs) self.needs_permute = needs_permute for i in range(len(self.transformer.attn_layers.layers)): n, b, r = self.transformer.attn_layers.layers[i] self.transformer.attn_layers.layers[i] = nn.ModuleList([n, CheckpointedLayer(b), r]) def forward(self, x, **kwargs): if self.needs_permute: x = x.permute(0,2,1) h = self.transformer(x, **kwargs) return h.permute(0,2,1) 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 if kernel_size == 3 else 2 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 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=32, 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, cond_transformer_depth=8, mid_transformer_depth=8, # Parameters for regularization. nil_guidance_fwd_proportion=.3, 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, # Parameters for unaligned inputs. enabled_unaligned_inputs=False, num_unaligned_tokens=164, unaligned_encoder_depth=8, # Experimental parameters component_gradient_boosting=False, ): 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.nil_guidance_fwd_proportion = nil_guidance_fwd_proportion self.mask_token_id = num_tokens 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 self.component_gradient_boosting = component_gradient_boosting 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), ) embedding_dim = model_channels * 8 self.code_embedding = nn.Embedding(num_tokens+1, embedding_dim) self.contextual_embedder = AudioMiniEncoder(1, embedding_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.conditioning_conv = nn.Conv1d(embedding_dim*3, embedding_dim, 1) self.enable_unaligned_inputs = enabled_unaligned_inputs if enabled_unaligned_inputs: self.unaligned_embedder = nn.Embedding(num_unaligned_tokens, embedding_dim) self.unaligned_encoder = CheckpointedXTransformerEncoder( max_seq_len=-1, use_pos_emb=False, attn_layers=Encoder( dim=embedding_dim, depth=unaligned_encoder_depth, heads=num_heads, ff_dropout=dropout, attn_dropout=dropout, use_rmsnorm=True, ff_glu=True, rotary_emb_dim=True, ) ) self.conditioning_encoder = CheckpointedXTransformerEncoder( max_seq_len=-1, # Should be unused use_pos_emb=False, attn_layers=Encoder( dim=embedding_dim, depth=cond_transformer_depth, heads=num_heads, ff_dropout=dropout, attn_dropout=dropout, use_rmsnorm=True, ff_glu=True, rotary_pos_emb=True, cross_attend=self.enable_unaligned_inputs, ) ) self.unconditioned_embedding = nn.Parameter(torch.randn(1,embedding_dim,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(embedding_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 mid_transformer = CheckpointedXTransformerEncoder( max_seq_len=-1, # Should be unused use_pos_emb=False, attn_layers=Encoder( dim=ch, depth=mid_transformer_depth, heads=num_heads, ff_dropout=dropout, attn_dropout=dropout, use_rmsnorm=True, ff_glu=True, rotary_pos_emb=True, ) ) self.middle_block = TimestepEmbedSequential( ResBlock( ch, time_embed_dim, dropout, dims=dims, kernel_size=kernel_size, ), mid_transformer, 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()), 'conditioning_encoder': list(self.conditioning_encoder.parameters()) } if self.enable_unaligned_inputs: groups['unaligned_encoder'] = list(self.unaligned_encoder.parameters()) return groups def before_step(self, it): if not self.component_gradient_boosting: return MIN_PROPORTIONAL_BOOST_LEVEL = .5 MAX_MULTIPLIER = 100 components = [list(self.contextual_embedder.parameters()), list(self.middle_block.parameters()), list(self.conditioning_encoder.parameters()), list(self.unaligned_encoder.parameters())] input_norm = torch.norm(torch.stack([torch.norm(p.grad.detach(), 2) for p in self.input_blocks.parameters()]), 2) output_norm = torch.norm(torch.stack([torch.norm(p.grad.detach(), 2) for p in self.output_blocks.parameters()]), 2) diffusion_norm = (input_norm + output_norm) / 2 min_norm = diffusion_norm * MIN_PROPORTIONAL_BOOST_LEVEL for component in components: norm = torch.norm(torch.stack([torch.norm(p.grad.detach(), 2) for p in component]), 2) if norm < min_norm: mult = min_norm / (norm + 1e-8) mult = min(mult, MAX_MULTIPLIER) for p in component: p.grad.data.mul_(mult) def forward(self, x, timesteps, tokens=None, conditioning_input=None, lr_input=None, unaligned_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 tokens: an aligned text input. :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 unaligned_input: A structural input that is not properly aligned with the output of the diffusion model. Can be combined with a conditioning input to produce more robust conditioning. :param conditioning_free: When set, all conditioning inputs (including tokens, conditioning_input and unaligned_input) will not be considered. :return: an [N x C x ...] Tensor of outputs. """ assert conditioning_input is not None if self.super_sampling_enabled: assert lr_input is not None if self.training and self.super_sampling_max_noising_factor > 0: noising_factor = random.uniform(0,self.super_sampling_max_noising_factor) 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) with autocast(x.device.type, enabled=self.enable_fp16): 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])) if tokens is not None: tokens = F.pad(tokens, (0,int(pc*tokens.shape[-1]))) hs = [] time_emb = self.time_embed(timestep_embedding(timesteps, self.model_channels)) if conditioning_free: code_emb = self.unconditioned_embedding.repeat(x.shape[0], 1, 1) else: if self.enable_unaligned_inputs: assert unaligned_input is not None unaligned_h = self.unaligned_embedder(unaligned_input).permute(0,2,1) unaligned_h = self.unaligned_encoder(unaligned_h).permute(0,2,1) cond_emb = self.contextual_embedder(conditioning_input) if tokens is not None: # Mask out guidance tokens for un-guided diffusion. if self.training and self.nil_guidance_fwd_proportion > 0: token_mask = clustered_mask(self.nil_guidance_fwd_proportion, tokens.shape, tokens.device, inverted=True) tokens = torch.where(token_mask, self.mask_token_id, tokens) code_emb = self.code_embedding(tokens).permute(0,2,1) cond_emb = cond_emb.unsqueeze(-1).repeat(1,1,code_emb.shape[-1]) cond_time_emb = timestep_embedding(torch.zeros_like(timesteps), code_emb.shape[1]) # This was something I was doing (adding timesteps into this computation), but removed on second thought. TODO: completely remove. cond_time_emb = cond_time_emb.unsqueeze(-1).repeat(1,1,code_emb.shape[-1]) code_emb = self.conditioning_conv(torch.cat([cond_emb, code_emb, cond_time_emb], dim=1)) else: code_emb = cond_emb.unsqueeze(-1) if self.enable_unaligned_inputs: code_emb = self.conditioning_encoder(code_emb, context=unaligned_h) else: code_emb = self.conditioning_encoder(code_emb) # 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) first = True 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: with autocast(x.device.type, enabled=self.enable_fp16 and not first): # First block has autocast disabled to allow a high precision signal to be properly vectorized. h = module(h, time_emb) hs.append(h) first = False 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[:, :, :orig_x_shape] @register_model def register_diffusion_tts7(opt_net, opt): return DiffusionTts(**opt_net['kwargs']) # Test for ~4 second audio clip at 22050Hz if __name__ == '__main__': clip = torch.randn(2, 1, 32768) tok = torch.randint(0,30, (2,388)) cond = torch.randn(2, 1, 44000) ts = torch.LongTensor([600, 600]) lr = torch.randn(2,1,10000) un = torch.randint(0,120, (2,100)) model = DiffusionTts(128, 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, enabled_unaligned_inputs=True, component_gradient_boosting=True) o = model(clip, ts, tok, cond, lr, un) o.sum().backward() model.before_step(0) torch.save(model.state_dict(), 'test_out.pth')