From b157b28c7bc9bcb811ec18beecc2e15e38342ef1 Mon Sep 17 00:00:00 2001 From: James Betker Date: Tue, 19 Jul 2022 13:30:05 -0600 Subject: [PATCH] tfd14 hopefully this helps address the positional dependencies of tfd12 --- codes/models/arch_util.py | 17 +- .../audio/music/transformer_diffusion14.py | 303 ++++++++++++++++++ .../injectors/gaussian_diffusion_injector.py | 2 +- 3 files changed, 314 insertions(+), 8 deletions(-) create mode 100644 codes/models/audio/music/transformer_diffusion14.py diff --git a/codes/models/arch_util.py b/codes/models/arch_util.py index e3d9fb9c..d701991c 100644 --- a/codes/models/arch_util.py +++ b/codes/models/arch_util.py @@ -492,6 +492,11 @@ class AttentionBlock(nn.Module): def _forward(self, x, mask=None): b, c, *spatial = x.shape + if len(mask.shape) == 2: + mask = mask.unsqueeze(0).repeat(x.shape[0],1,1) + if mask.shape[1] != x.shape[-1]: + mask = mask[:, :x.shape[-1], :x.shape[-1]] + x = x.reshape(b, c, -1) x = self.norm(x) if self.do_activation: @@ -527,11 +532,10 @@ class QKVAttentionLegacy(nn.Module): weight = torch.einsum( "bct,bcs->bts", q * scale, k * scale ) # More stable with f16 than dividing afterwards - weight = torch.softmax(weight.float(), dim=-1).type(weight.dtype) if mask is not None: - # The proper way to do this is to mask before the softmax using -inf, but that doesn't work properly on CPUs. - mask = mask.repeat(self.n_heads, 1).unsqueeze(1) - weight = weight * mask + mask = mask.repeat(self.n_heads, 1, 1) + weight[mask.logical_not()] = -torch.inf + weight = torch.softmax(weight.float(), dim=-1).type(weight.dtype) a = torch.einsum("bts,bcs->bct", weight, v) return a.reshape(bs, -1, length) @@ -564,9 +568,8 @@ class QKVAttention(nn.Module): (k * scale).view(bs * self.n_heads, ch, length), ) # More stable with f16 than dividing afterwards if mask is not None: - # The proper way to do this is to mask before the softmax using -inf, but that doesn't work properly on CPUs. - mask = mask.repeat(self.n_heads, 1).unsqueeze(1) - weight = weight * mask + mask = mask.repeat(self.n_heads, 1, 1) + weight[mask.logical_not()] = -torch.inf weight = torch.softmax(weight.float(), dim=-1).type(weight.dtype) a = torch.einsum("bts,bcs->bct", weight, v.reshape(bs * self.n_heads, ch, length)) return a.reshape(bs, -1, length) diff --git a/codes/models/audio/music/transformer_diffusion14.py b/codes/models/audio/music/transformer_diffusion14.py new file mode 100644 index 00000000..20c7ba15 --- /dev/null +++ b/codes/models/audio/music/transformer_diffusion14.py @@ -0,0 +1,303 @@ +import itertools + +import torch +import torch.nn as nn +import torch.nn.functional as F + +from models.arch_util import AttentionBlock, TimestepEmbedSequential +from models.audio.music.encoders import ResEncoder16x +from models.diffusion.nn import timestep_embedding, normalization, zero_module, conv_nd, linear +from models.diffusion.unet_diffusion import TimestepBlock +from trainer.networks import register_model +from utils.util import checkpoint, print_network + + +def build_local_attention_mask(n, l, fixed_region): + """ + Builds an attention mask that focuses attention on local region + Includes provisions for a "fixed_region" at the start of the sequence where full attention weights will be applied. + Args: + n: Size of returned matrix (maximum sequence size) + l: Size of local context (uni-directional, e.g. the total context is l*2) + fixed_region: The number of sequence elements at the start of the sequence that get full attention. + Returns: + A mask that can be applied to AttentionBlock to achieve local attention. + """ + assert l*2 < n, f'Local context must be less than global context. {l}, {n}' + o = torch.arange(0,n) + c = o.unsqueeze(-1).repeat(1,n) + r = o.unsqueeze(0).repeat(n,1) + localized = ((-(r-c).abs())+l).clamp(0,l-1) / (l-1) + localized[:fixed_region] = 1 + localized[:, :fixed_region] = 1 + mask = localized > 0 + return mask + + +def test_local_attention_mask(): + print(build_local_attention_mask(9,4,1)) + + +class SubBlock(nn.Module): + def __init__(self, inp_dim, contraction_dim, blk_dim, heads, dropout, enable_attention_masking=False): + super().__init__() + self.enable_attention_masking = enable_attention_masking + self.dropout = nn.Dropout(p=dropout) + self.blk_emb_proj = nn.Conv1d(blk_dim, inp_dim, 1) + self.attn = AttentionBlock(inp_dim, out_channels=contraction_dim, num_heads=heads) + self.attnorm = nn.GroupNorm(8, contraction_dim) + self.ff = nn.Conv1d(inp_dim+contraction_dim, contraction_dim, kernel_size=3, padding=1) + self.ffnorm = nn.GroupNorm(8, contraction_dim) + if self.enable_attention_masking: + # All regions can attend to the first token, which will be the timestep embedding. Hence, fixed_region. + self.mask = build_local_attention_mask(n=2000, l=48, fixed_region=1) + else: + self.mask = None + + def forward(self, x, blk_emb): + if self.mask is not None: + self.mask = self.mask.to(x.device) + blk_enc = self.blk_emb_proj(blk_emb) + ah = self.dropout(self.attn(torch.cat([blk_enc, x], dim=-1), mask=self.mask)) + ah = ah[:,:,blk_emb.shape[-1]:] # Strip off the blk_emb and re-align with x. + ah = F.gelu(self.attnorm(ah)) + h = torch.cat([ah, x], dim=1) + hf = self.dropout(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, heads, dropout, enable_attention_masking=False): + super().__init__() + self.prenorm = nn.GroupNorm(8, trunk_dim) + self.block1 = SubBlock(trunk_dim, contraction_dim, trunk_dim, heads, dropout, + enable_attention_masking=enable_attention_masking) + self.block2 = SubBlock(trunk_dim+contraction_dim*2, contraction_dim, trunk_dim, heads, dropout, + enable_attention_masking=enable_attention_masking) + self.out = nn.Conv1d(contraction_dim*4, trunk_dim, kernel_size=1, bias=False) + self.out.weight.data.zero_() + + def forward(self, x, blk_emb): + h = self.prenorm(x) + h = self.block1(h, blk_emb) + h = self.block2(h, blk_emb) + h = self.out(h[:,x.shape[1]:]) + 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, + time_embed_dim=256, + model_channels=1024, + contraction_dim=256, + num_layers=8, + in_channels=256, + input_vec_dim=1024, + out_channels=512, # mean and variance + num_heads=4, + dropout=0, + use_corner_alignment=False, # This is an interpolation parameter only provided for backwards compatibility. ALL NEW TRAINS SHOULD SET THIS TO TRUE. + use_fp16=False, + new_code_expansion=False, + # Parameters for regularization. + unconditioned_percentage=.1, # This implements a mechanism similar to what is used in classifier-free training. + # Parameters for re-training head + freeze_except_code_converters=False, + ): + 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.new_code_expansion = new_code_expansion + self.use_corner_alignment = use_corner_alignment + 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, model_channels), + ) + + self.input_converter = nn.Conv1d(input_vec_dim, model_channels, 1) + self.unconditioned_embedding = nn.Parameter(torch.randn(1,model_channels,1)) + self.intg = nn.Conv1d(model_channels*2, model_channels, 1) + self.layers = TimestepEmbedSequential(*[ConcatAttentionBlock(model_channels, contraction_dim, num_heads, dropout, enable_attention_masking=True) 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)), + ) + + if freeze_except_code_converters: + for p in self.parameters(): + p.DO_NOT_TRAIN = True + p.requires_grad = False + for m in [self.code_converter and self.input_converter]: + for p in m.parameters(): + del p.DO_NOT_TRAIN + p.requires_grad = True + + 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, + '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, prior=None, conditioning_free=False): + if conditioning_free: + code_emb = self.unconditioned_embedding.repeat(x.shape[0], x.shape[-1], 1) + else: + code_emb = self.input_converter(prior) + + # 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) + + code_emb = F.interpolate(code_emb, size=x.shape[-1], mode='nearest') + + with torch.autocast(x.device.type, enabled=self.enable_fp16): + blk_emb = self.time_embed(timestep_embedding(timesteps, self.time_embed_dim)).unsqueeze(-1) + x = self.inp_block(x) + + x = self.intg(torch.cat([x, code_emb], dim=1)) + for layer in self.layers: + x = checkpoint(layer, x, blk_emb) + + x = x.float() + out = self.out(x) + + # Defensively involve probabilistic or possibly unused parameters in loss so we don't get DDP errors. + unused_params = [self.unconditioned_embedding] + extraneous_addition = 0 + for p in unused_params: + extraneous_addition = extraneous_addition + p.mean() + out = out + extraneous_addition * 0 + + return out + + +class TransformerDiffusionWithCheaterLatent(nn.Module): + def __init__(self, freeze_encoder_until=None, checkpoint_encoder=True, **kwargs): + super().__init__() + self.internal_step = 0 + self.freeze_encoder_until = freeze_encoder_until + self.diff = TransformerDiffusion(**kwargs) + self.encoder = ResEncoder16x(256, 1024, 256, checkpointing_enabled=checkpoint_encoder) + + def forward(self, x, timesteps, truth_mel, conditioning_free=False): + unused_parameters = [] + encoder_grad_enabled = self.freeze_encoder_until is not None and self.internal_step > self.freeze_encoder_until + if not encoder_grad_enabled: + unused_parameters.extend(list(self.encoder.parameters())) + with torch.set_grad_enabled(encoder_grad_enabled): + proj = self.encoder(truth_mel) + + for p in unused_parameters: + proj = proj + p.mean() * 0 + + diff = self.diff(x, timesteps, prior=proj, 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['encoder'] = list(self.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_transformer_diffusion14(opt_net, opt): + return TransformerDiffusion(**opt_net['kwargs']) + + +@register_model +def register_transformer_diffusion_14_with_cheater_latent(opt_net, opt): + return TransformerDiffusionWithCheaterLatent(**opt_net['kwargs']) + + +def test_tfd(): + clip = torch.randn(2,256,400) + ts = torch.LongTensor([600, 600]) + model = TransformerDiffusion(in_channels=256, model_channels=1024, contraction_dim=512, + num_heads=3, input_vec_dim=256, num_layers=12, dropout=.1) + model(clip, ts, clip) + + +def test_cheater_model(): + clip = torch.randn(2, 256, 400) + ts = torch.LongTensor([600, 600]) + + # For music: + model = TransformerDiffusionWithCheaterLatent(in_channels=256, out_channels=512, + model_channels=1024, contraction_dim=512, num_heads=8, + input_vec_dim=256, num_layers=16, + dropout=.1, new_code_expansion=True, + ) + #diff_weights = torch.load('extracted_diff.pth') + #model.diff.load_state_dict(diff_weights, strict=False) + #model.encoder.load_state_dict(torch.load('../experiments/music_cheater_encoder_256.pth', map_location=torch.device('cpu')), strict=True) + #torch.save(model.state_dict(), 'sample.pth') + + print_network(model) + o = model(clip, ts, clip) + pg = model.get_grad_norm_parameter_groups() + + +def extract_cheater_encoder(in_f, out_f): + p = torch.load(in_f) + out = {} + for k, v in p.items(): + if k.startswith('encoder.'): + out[k] = v + torch.save(out, out_f) + + +if __name__ == '__main__': + #test_local_attention_mask() + extract_cheater_encoder('X:\\dlas\\experiments\\train_music_diffusion_tfd_and_cheater\\models\\104500_generator_ema.pth', 'X:\\dlas\\experiments\\tfd12_self_learned_cheater_enc.pth', True) + test_cheater_model() + #extract_diff('X:\\dlas\experiments\\train_music_diffusion_tfd_cheater_from_scratch\\models\\56500_generator_ema.pth', 'extracted.pth', remove_head=True) diff --git a/codes/trainer/injectors/gaussian_diffusion_injector.py b/codes/trainer/injectors/gaussian_diffusion_injector.py index 6d461ba2..b2e9222a 100644 --- a/codes/trainer/injectors/gaussian_diffusion_injector.py +++ b/codes/trainer/injectors/gaussian_diffusion_injector.py @@ -80,7 +80,7 @@ class GaussianDiffusionInjector(Injector): def forward(self, state): gen = self.env['generators'][self.opt['generator']] hq = state[self.input] - assert hq.max() < 1 or hq.min() > -1, f"Attempting to train gaussian diffusion on un-normalized inputs. This won't work, silly! {hq.min()} {hq.max()}" + assert hq.max() < 1.000001 or hq.min() > -1.00001, f"Attempting to train gaussian diffusion on un-normalized inputs. This won't work, silly! {hq.min()} {hq.max()}" with autocast(enabled=self.env['opt']['fp16']): if not gen.training or (self.deterministic_timesteps_every != 0 and self.env['step'] % self.deterministic_timesteps_every == 0):