import itertools import os import random import torch import torch.nn as nn import torch.nn.functional as F import torchaudio import torchvision 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, load_audio 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, use_conv): 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.use_conv = use_conv if use_conv: self.ff = nn.Conv1d(inp_dim+contraction_dim, contraction_dim, kernel_size=3, padding=1) else: 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.permute(0,2,1) if self.use_conv else h) hf = F.gelu(self.ffnorm(hf.permute(0,2,1) if self.use_conv else 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, cond_projection=True, use_conv=False): super().__init__() self.prenorm = RMSScaleShiftNorm(trunk_dim, embed_dim=time_embed_dim, bias=False) if cond_projection: self.tdim = trunk_dim+cond_dim_hidden self.cond_project = nn.Linear(cond_dim_in, cond_dim_hidden) else: self.tdim = trunk_dim self.block1 = SubBlock(self.tdim, contraction_dim, heads, dropout, use_conv) self.block2 = SubBlock(self.tdim+contraction_dim*2, contraction_dim, heads, dropout, use_conv) 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) if hasattr(self, 'cond_project'): 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[:,:,self.tdim:]) return h + x class ConditioningEncoder(nn.Module): def __init__(self, cond_dim, embedding_dim, time_embed_dim, attn_blocks=6, num_attn_heads=8, dropout=.1, do_checkpointing=False, time_proj=True): super().__init__() self.init = nn.Conv1d(cond_dim, embedding_dim, kernel_size=1) self.time_proj = time_proj if time_proj: self.time_proj = nn.Linear(time_embed_dim, embedding_dim) 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, do_checkpointing=do_checkpointing ) self.dim = embedding_dim def forward(self, x, time_emb): h = self.init(x).permute(0,2,1) if self.time_proj: time_enc = self.time_proj(time_emb) h = torch.cat([time_enc.unsqueeze(1), h], dim=1) h = self.attn(h).permute(0,2,1) return h class TransformerDiffusionWithPointConditioning(nn.Module): 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, time_proj=True, new_cond=False, use_fp16=False, checkpoint_conditioning=True, # This will need to be false for DDP training. :( regularization=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.regularization = regularization self.new_cond = new_cond self.inp_block = conv_nd(1, in_channels, model_channels, 3, 1, 1) self.conditioning_encoder = ConditioningEncoder(256, model_channels, time_embed_dim, do_checkpointing=checkpoint_conditioning, time_proj=time_proj) 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, cond_projection=(k % 3 == 0), use_conv=(k % 3 != 0), ) for k 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()), 'conditioning_encoder': list(self.conditioning_encoder.parameters()), } return groups def process_conditioning(self, conditioning_input, time_emb, N, cond_start, cond_left, cond_right): if self.training and self.regularization: # frequency regularization fstart = random.randint(0, conditioning_input.shape[1] - 1) fclip = random.randint(1, min(conditioning_input.shape[1]-fstart, 16)) conditioning_input[:,fstart:fstart+fclip] = 0 # time regularization for k in range(1, random.randint(2, 4)): tstart = random.randint(0, conditioning_input.shape[-1] - 1) tclip = random.randint(1, min(conditioning_input.shape[-1]-tstart, 10)) conditioning_input[:,:,tstart:tstart+tclip] = 0 if cond_left is None and self.new_cond: assert cond_start > 20 and (cond_start+N+20 <= conditioning_input.shape[-1]), f'{cond_start}, {N}, {conditioning_input.shape}' cond_left = conditioning_input[:,:,:cond_start] left_pt = -1 cond_right = conditioning_input[:,:,cond_start+N:] right_pt = 0 if self.training: # Arbitrarily restrict the context given. We should support short contexts and without this they are never encountered. arb_context_cap = random.randint(50, 100) if cond_left.shape[-1] > arb_context_cap and random.random() > .5: cond_left = cond_left[:,:,-arb_context_cap:] if cond_right.shape[-1] > arb_context_cap and random.random() > .5: cond_right = cond_right[:,:,:arb_context_cap] elif cond_left is None: assert conditioning_input.shape[-1] - cond_start - N >= 0, f'Some sort of conditioning misalignment, {conditioning_input.shape[-1], cond_start, N}' cond_pre = conditioning_input[:,:,:cond_start] cond_aligned = conditioning_input[:,:,cond_start:N+cond_start] cond_post = conditioning_input[:,:,N+cond_start:] # Break up conditioning input into two random segments aligned with the input. MIN_MARGIN = 8 assert N > (MIN_MARGIN*2+4), f"Input size too small. Was {N} but requires at least {MIN_MARGIN*2+4}" break_pt = random.randint(2, N-MIN_MARGIN*2-2) + MIN_MARGIN cond_left = cond_aligned[:,:,:break_pt] cond_right = cond_aligned[:,:,break_pt:] if self.training: # Drop out a random amount of the aligned data. The network will need to figure out how to reconstruct this. to_remove_left = random.randint(1, cond_left.shape[-1]-MIN_MARGIN) cond_left = cond_left[:,:,:-to_remove_left] to_remove_right = random.randint(1, cond_right.shape[-1]-MIN_MARGIN) cond_right = cond_right[:,:,to_remove_right:] # Concatenate the _pre and _post back on. left_pt = cond_start right_pt = cond_right.shape[-1] cond_left = torch.cat([cond_pre, cond_left], dim=-1) cond_right = torch.cat([cond_right, cond_post], dim=-1) else: left_pt = -1 right_pt = 0 # Propagate through the encoder. cond_left_enc = self.conditioning_encoder(cond_left, time_emb) cs = cond_left_enc[:,:,left_pt] cond_right_enc = self.conditioning_encoder(cond_right, time_emb) ce = cond_right_enc[:,:,right_pt] cond_enc = torch.cat([cs.unsqueeze(-1), ce.unsqueeze(-1)], dim=-1) cond = F.interpolate(cond_enc, size=(N,), mode='linear', align_corners=True).permute(0,2,1) return cond def forward(self, x, timesteps, conditioning_input=None, cond_left=None, cond_right=None, conditioning_free=False, cond_start=0): unused_params = [] time_emb = self.time_embed(timestep_embedding(timesteps, self.time_embed_dim)) if conditioning_free: cond = self.unconditioned_embedding cond = cond.repeat(1,x.shape[-1],1) else: cond = self.process_conditioning(conditioning_input, time_emb, x.shape[-1], cond_start, cond_left, cond_right) # 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) with torch.autocast(x.device.type, enabled=self.enable_fp16): 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, time_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 def before_step(self, step): scaled_grad_parameters = list(itertools.chain.from_iterable([lyr.out.parameters() for lyr in self.layers])) + \ list(itertools.chain.from_iterable([lyr.prenorm.parameters() for lyr in self.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_tfdpc5(opt_net, opt): return TransformerDiffusionWithPointConditioning(**opt_net['kwargs']) def test_cheater_model(): clip = torch.randn(2, 256, 350) cl = torch.randn(2, 256, 646) ts = torch.LongTensor([600, 600]) # For music: model = TransformerDiffusionWithPointConditioning(in_channels=256, out_channels=512, model_channels=1024, contraction_dim=512, num_heads=8, num_layers=32, dropout=0, unconditioned_percentage=.4, checkpoint_conditioning=False, regularization=True, new_cond=True) print_network(model) #for cs in range(276,cl.shape[-1]-clip.shape[-1]): # o = model(clip, ts, cl, cond_start=cs) pg = model.get_grad_norm_parameter_groups() def prmsz(lp): sz = 0 for p in lp: q = 1 for s in p.shape: q *= s sz += q return sz for k, v in pg.items(): print(f'{k}: {prmsz(v)/1000000}') def test_conditioning_splitting_logic(): ts = torch.LongTensor([600]) class fake_conditioner(nn.Module): def __init__(self): super().__init__() def forward(self, t, _): print(t[:,0]) return t model = TransformerDiffusionWithPointConditioning(in_channels=256, out_channels=512, model_channels=1024, contraction_dim=512, num_heads=8, num_layers=15, dropout=0, unconditioned_percentage=.4) model.conditioning_encoder = fake_conditioner() BASEDIM=30 for x in range(BASEDIM+1, BASEDIM+20): start = random.randint(0,x-BASEDIM) cl = torch.arange(1, x+1, 1).view(1,1,-1).float().repeat(1,256,1) print("Effective input: " + str(cl[0, 0, start:BASEDIM+start])) res = model.process_conditioning(cl, ts, BASEDIM, start, None) print("Result: " + str(res[0,:,0])) print() def inference_tfdpc5_with_cheater(): with torch.no_grad(): os.makedirs('results/tfdpc_v3', exist_ok=True) # length = 40 * 22050 // 256 // 16 samples = {'electronica1': load_audio('Y:\\split\\yt-music-eval\\00001.wav', 22050), 'electronica2': load_audio('Y:\\split\\yt-music-eval\\00272.wav', 22050), 'e_guitar': load_audio('Y:\\split\\yt-music-eval\\00227.wav', 22050), 'creep': load_audio('Y:\\separated\\bt-music-3\\[2007] MTV Unplugged (Live) (Japan Edition)\\05 - Creep [Cover On Radiohead]\\00001\\no_vocals.wav', 22050), 'rock1': load_audio('Y:\\separated\\bt-music-3\\2016 - Heal My Soul\\01 - Daze Of The Night\\00000\\no_vocals.wav', 22050), 'kiss': load_audio('Y:\\separated\\bt-music-3\\KISS (2001) Box Set CD1\\02 Deuce (Demo Version)\\00000\\no_vocals.wav', 22050), 'purp': load_audio('Y:\\separated\\bt-music-3\\Shades of Deep Purple\\11 Help (Alternate Take)\\00001\\no_vocals.wav', 22050), 'western_stars': load_audio('Y:\\separated\\bt-music-3\\Western Stars\\01 Hitch Hikin\'\\00000\\no_vocals.wav', 22050), 'silk': load_audio('Y:\\separated\\silk\\MonstercatSilkShowcase\\890\\00007\\no_vocals.wav', 22050), 'long_electronica': load_audio('C:\\Users\\James\\Music\\longer_sample.wav', 22050),} for k, sample in samples.items(): sample = sample.cuda() length = sample.shape[0]//256//16 model = TransformerDiffusionWithPointConditioning(in_channels=256, out_channels=512, model_channels=1024, contraction_dim=512, num_heads=8, num_layers=12, dropout=0, use_fp16=False, unconditioned_percentage=0).eval().cuda() model.load_state_dict(torch.load('x:/dlas/experiments/train_music_cheater_gen_v3/models/59000_generator_ema.pth')) from trainer.injectors.audio_injectors import TorchMelSpectrogramInjector spec_fn = TorchMelSpectrogramInjector({'n_mel_channels': 256, 'mel_fmax': 11000, 'filter_length': 16000, 'true_normalization': True, 'normalize': True, 'in': 'in', 'out': 'out'}, {}).cuda() ref_mel = spec_fn({'in': sample.unsqueeze(0)})['out'] from trainer.injectors.audio_injectors import MusicCheaterLatentInjector cheater_encoder = MusicCheaterLatentInjector({'in': 'in', 'out': 'out'}, {}).cuda() ref_cheater = cheater_encoder({'in': ref_mel})['out'] from models.diffusion.respace import SpacedDiffusion from models.diffusion.respace import space_timesteps from models.diffusion.gaussian_diffusion import get_named_beta_schedule diffuser = SpacedDiffusion(use_timesteps=space_timesteps(4000, [128]), model_mean_type='epsilon', model_var_type='learned_range', loss_type='mse', betas=get_named_beta_schedule('linear', 4000), conditioning_free=True, conditioning_free_k=1) # Conventional decoding method: gen_cheater = diffuser.ddim_sample_loop(model, (1,256,length), progress=True, model_kwargs={'true_cheater': ref_cheater}) # Guidance decoding method: #mask = torch.ones_like(ref_cheater) #mask[:,:,15:-15] = 0 #gen_cheater = diffuser.p_sample_loop_with_guidance(model, ref_cheater, mask, model_kwargs={'true_cheater': ref_cheater}) # Just decode the ref. #gen_cheater = ref_cheater from models.audio.music.transformer_diffusion12 import TransformerDiffusionWithCheaterLatent diffuser = SpacedDiffusion(use_timesteps=space_timesteps(4000, [32]), model_mean_type='epsilon', model_var_type='learned_range', loss_type='mse', betas=get_named_beta_schedule('linear', 4000), conditioning_free=True, conditioning_free_k=1) wrap = TransformerDiffusionWithCheaterLatent(in_channels=256, out_channels=512, model_channels=1024, contraction_dim=512, prenet_channels=1024, input_vec_dim=256, prenet_layers=6, num_heads=8, num_layers=16, new_code_expansion=True, dropout=0, unconditioned_percentage=0).eval().cuda() wrap.load_state_dict(torch.load('x:/dlas/experiments/train_music_diffusion_tfd_cheater_from_scratch/models/56500_generator_ema.pth')) cheater_to_mel = wrap.diff gen_mel = diffuser.ddim_sample_loop(cheater_to_mel, (1,256,gen_cheater.shape[-1]*16), progress=True, model_kwargs={'codes': gen_cheater.permute(0,2,1)}) torchvision.utils.save_image((gen_mel + 1)/2, f'results/tfdpc_v3/{k}.png') from utils.music_utils import get_mel2wav_v3_model m2w = get_mel2wav_v3_model().cuda() spectral_diffuser = SpacedDiffusion(use_timesteps=space_timesteps(4000, [32]), model_mean_type='epsilon', model_var_type='learned_range', loss_type='mse', betas=get_named_beta_schedule('linear', 4000), conditioning_free=True, conditioning_free_k=1) from trainer.injectors.audio_injectors import denormalize_mel gen_mel_denorm = denormalize_mel(gen_mel) output_shape = (1,16,gen_mel_denorm.shape[-1]*256//16) gen_wav = spectral_diffuser.ddim_sample_loop(m2w, output_shape, model_kwargs={'codes': gen_mel_denorm}) from trainer.injectors.audio_injectors import pixel_shuffle_1d gen_wav = pixel_shuffle_1d(gen_wav, 16) torchaudio.save(f'results/tfdpc_v3/{k}.wav', gen_wav.squeeze(1).cpu(), 22050) torchaudio.save(f'results/tfdpc_v3/{k}_ref.wav', sample.unsqueeze(0).cpu(), 22050) if __name__ == '__main__': test_cheater_model() #test_conditioning_splitting_logic() #inference_tfdpc5_with_cheater()