From e58dab14c347370807417fc1bd9c4fcf94bc5f98 Mon Sep 17 00:00:00 2001
From: James Betker <jbetker@gmail.com>
Date: Sat, 29 Jan 2022 11:01:01 -0700
Subject: [PATCH] new diffusion updates from testing
---
codes/models/gpt_voice/unet_diffusion_tts3.py | 135 +++--
codes/models/gpt_voice/unet_diffusion_tts4.py | 406 ++++++++++++++++
codes/models/gpt_voice/unet_diffusion_tts5.py | 460 ++++++++++++++++++
codes/models/lucidrains/dalle/transformer.py | 5 +-
codes/train.py | 2 +-
5 files changed, 965 insertions(+), 43 deletions(-)
create mode 100644 codes/models/gpt_voice/unet_diffusion_tts4.py
create mode 100644 codes/models/gpt_voice/unet_diffusion_tts5.py
diff --git a/codes/models/gpt_voice/unet_diffusion_tts3.py b/codes/models/gpt_voice/unet_diffusion_tts3.py
index 071cbaa4..f1958d83 100644
--- a/codes/models/gpt_voice/unet_diffusion_tts3.py
+++ b/codes/models/gpt_voice/unet_diffusion_tts3.py
@@ -1,17 +1,41 @@
-import operator
+import functools
from collections import OrderedDict
-from models.diffusion.nn import timestep_embedding, normalization, zero_module, conv_nd, linear
-from models.diffusion.unet_diffusion import AttentionPool2d, AttentionBlock, TimestepEmbedSequential, \
- Downsample, Upsample, TimestepBlock
import torch
import torch.nn as nn
import torch.nn.functional as F
+from torch import autocast
+from x_transformers.x_transformers import AbsolutePositionalEmbedding, AttentionLayers
-from models.gpt_voice.mini_encoder import AudioMiniEncoder, EmbeddingCombiner
+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.gpt_voice.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
+
+
+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, **kwargs):
+ kw_requires_grad = {}
+ kw_no_grad = {}
+ for k, v in kwargs.items():
+ if v is not None and isinstance(v, torch.Tensor) and v.requires_grad:
+ kw_requires_grad[k] = v
+ else:
+ kw_no_grad[k] = v
+ partial = functools.partial(self.wrap, **kw_no_grad)
+ return torch.utils.checkpoint.checkpoint(partial, x, **kw_requires_grad)
class ResBlock(TimestepBlock):
@@ -186,6 +210,12 @@ class DiffusionTts(nn.Module):
ch = model_channels
ds = 1
+ class Permute(nn.Module):
+ def __init__(self):
+ super().__init__()
+ def forward(self, x):
+ return x.permute(0,2,1)
+
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)
@@ -230,6 +260,26 @@ class DiffusionTts(nn.Module):
ds *= 2
self._feature_size += ch
+ mid_transformer = ContinuousTransformerWrapper(
+ max_seq_len=-1, # Should be unused
+ use_pos_emb=False,
+ attn_layers=Encoder(
+ dim=ch,
+ depth=8,
+ heads=num_heads,
+ ff_dropout=dropout,
+ attn_dropout=dropout,
+ use_rmsnorm=True,
+ ff_glu=True,
+ rotary_pos_emb=True,
+ )
+ )
+
+ for i in range(len(mid_transformer.attn_layers.layers)):
+ n, b, r = mid_transformer.attn_layers.layers[i]
+ mid_transformer.attn_layers.layers[i] = nn.ModuleList([n, CheckpointedLayer(b), r])
+
+
self.middle_block = TimestepEmbedSequential(
ResBlock(
ch,
@@ -238,11 +288,9 @@ class DiffusionTts(nn.Module):
dims=dims,
kernel_size=kernel_size,
),
- AttentionBlock(
- ch,
- num_heads=num_heads,
- num_head_channels=num_head_channels,
- ),
+ Permute(),
+ mid_transformer,
+ Permute(),
ResBlock(
ch,
time_embed_dim,
@@ -318,41 +366,48 @@ class DiffusionTts(nn.Module):
:param tokens: an aligned text input.
:return: an [N x C x ...] Tensor of outputs.
"""
- 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]))
- tokens = F.pad(tokens, (0,int(pc*tokens.shape[-1])))
- if self.conditioning_enabled:
- assert conditioning_input is not None
+ with autocast(x.device.type):
+ 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]))
+ tokens = F.pad(tokens, (0,int(pc*tokens.shape[-1])))
+ if self.conditioning_enabled:
+ assert conditioning_input is not None
- hs = []
- time_emb = self.time_embed(timestep_embedding(timesteps, self.model_channels))
+ hs = []
+ time_emb = self.time_embed(timestep_embedding(timesteps, self.model_channels))
- # Mask out guidance tokens for un-guided diffusion.
- if self.training and self.nil_guidance_fwd_proportion > 0:
- token_mask = torch.rand(tokens.shape, device=tokens.device) < self.nil_guidance_fwd_proportion
- tokens = torch.where(token_mask, self.mask_token_id, tokens)
- code_emb = self.code_embedding(tokens).permute(0,2,1)
- if self.conditioning_enabled:
- cond_emb = self.contextual_embedder(conditioning_input)
- code_emb = cond_emb.unsqueeze(-1) * code_emb
+ # Mask out guidance tokens for un-guided diffusion.
+ if self.training and self.nil_guidance_fwd_proportion > 0:
+ token_mask = torch.rand(tokens.shape, device=tokens.device) < self.nil_guidance_fwd_proportion
+ tokens = torch.where(token_mask, self.mask_token_id, tokens)
+ code_emb = self.code_embedding(tokens).permute(0,2,1)
+ if self.conditioning_enabled:
+ cond_emb = self.contextual_embedder(conditioning_input)
+ code_emb = cond_emb.unsqueeze(-1) * code_emb
- h = x.type(self.dtype)
+ first = False # First block has autocast disabled.
+ 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=not first):
+ if isinstance(module, nn.Conv1d):
+ h_tok = F.interpolate(module(code_emb), size=(h.shape[-1]), mode='nearest')
+ h = h + h_tok
+ else:
+ h = module(h, time_emb)
+ hs.append(h)
+ first = True
+ with autocast(x.device.type):
+ 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)
- hs.append(h)
- 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)
- h = h.type(x.dtype)
- out = self.out(h)
+ h = h.type(x.dtype)
+ h = h.float()
+ out = self.out(h) # Last block also has autocast disabled.
return out[:, :, :orig_x_shape]
diff --git a/codes/models/gpt_voice/unet_diffusion_tts4.py b/codes/models/gpt_voice/unet_diffusion_tts4.py
new file mode 100644
index 00000000..3046e83c
--- /dev/null
+++ b/codes/models/gpt_voice/unet_diffusion_tts4.py
@@ -0,0 +1,406 @@
+import functools
+from collections import OrderedDict
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torch import autocast
+from x_transformers.x_transformers import AbsolutePositionalEmbedding, AttentionLayers
+
+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.gpt_voice.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
+
+
+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, **kwargs):
+ kw_requires_grad = {}
+ kw_no_grad = {}
+ for k, v in kwargs.items():
+ if v is not None and isinstance(v, torch.Tensor) and v.requires_grad:
+ kw_requires_grad[k] = v
+ else:
+ kw_no_grad[k] = v
+ partial = functools.partial(self.wrap, **kw_no_grad)
+ return torch.utils.checkpoint.checkpoint(partial, x, **kw_requires_grad)
+
+
+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,
+ conditioning_inputs_provided=True,
+ time_embed_dim_multiplier=4,
+ nil_guidance_fwd_proportion=.3,
+ ):
+ super().__init__()
+
+ if num_heads_upsample == -1:
+ num_heads_upsample = num_heads
+
+ 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.dtype = torch.float16 if use_fp16 else torch.float32
+ 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
+
+ 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 * 4
+ self.code_embedding = nn.Embedding(num_tokens+1, embedding_dim)
+ self.conditioning_enabled = conditioning_inputs_provided
+ if conditioning_inputs_provided:
+ self.contextual_embedder = AudioMiniEncoder(in_channels, 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.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
+
+ class Permute(nn.Module):
+ def __init__(self):
+ super().__init__()
+ def forward(self, x):
+ return x.permute(0,2,1)
+
+ 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
+
+ self.middle_block = TimestepEmbedSequential(
+ ResBlock(
+ ch,
+ time_embed_dim,
+ dropout,
+ dims=dims,
+ kernel_size=kernel_size,
+ ),
+ 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 load_state_dict(self, state_dict: 'OrderedDict[str, Tensor]',
+ strict: bool = True):
+ # Temporary hack to allow the addition of nil-guidance token embeddings to the existing guidance embeddings.
+ lsd = self.state_dict()
+ revised = 0
+ for i, blk in enumerate(self.input_blocks):
+ if isinstance(blk, nn.Embedding):
+ key = f'input_blocks.{i}.weight'
+ if state_dict[key].shape[0] != lsd[key].shape[0]:
+ t = torch.randn_like(lsd[key]) * .02
+ t[:state_dict[key].shape[0]] = state_dict[key]
+ state_dict[key] = t
+ revised += 1
+ print(f"Loaded experimental unet_diffusion_net with {revised} modifications.")
+ return super().load_state_dict(state_dict, strict)
+
+
+
+ def forward(self, x, timesteps, tokens, conditioning_input=None):
+ """
+ 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.
+ :return: an [N x C x ...] Tensor of outputs.
+ """
+ with autocast(x.device.type):
+ 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]))
+ tokens = F.pad(tokens, (0,int(pc*tokens.shape[-1])))
+ if self.conditioning_enabled:
+ assert conditioning_input is not None
+
+ hs = []
+ time_emb = self.time_embed(timestep_embedding(timesteps, self.model_channels))
+
+ # Mask out guidance tokens for un-guided diffusion.
+ if self.training and self.nil_guidance_fwd_proportion > 0:
+ token_mask = torch.rand(tokens.shape, device=tokens.device) < self.nil_guidance_fwd_proportion
+ tokens = torch.where(token_mask, self.mask_token_id, tokens)
+ code_emb = self.code_embedding(tokens).permute(0,2,1)
+ if self.conditioning_enabled:
+ cond_emb = self.contextual_embedder(conditioning_input)
+ code_emb = cond_emb.unsqueeze(-1) * code_emb
+
+ first = False # First block has autocast disabled.
+ time_emb = time_emb.float()
+ h = x
+ for k, module in enumerate(self.input_blocks):
+ with autocast(x.device.type, enabled=not first):
+ if isinstance(module, nn.Conv1d):
+ h_tok = F.interpolate(module(code_emb), size=(h.shape[-1]), mode='nearest')
+ h = h + h_tok
+ else:
+ h = module(h, time_emb)
+ hs.append(h)
+ first = True
+ with autocast(x.device.type):
+ 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)
+ h = h.type(x.dtype)
+ h = h.float()
+ out = self.out(h) # Last block also has autocast disabled.
+ return out[:, :, :orig_x_shape]
+
+
+@register_model
+def register_diffusion_tts4(opt_net, opt):
+ return DiffusionTts(**opt_net['kwargs'])
+
+
+# Test for ~4 second audio clip at 22050Hz
+if __name__ == '__main__':
+ clip = torch.randn(4, 1, 86016)
+ tok = torch.randint(0,30, (4,388))
+ cond = torch.randn(4, 1, 44000)
+ ts = torch.LongTensor([555, 556, 600, 600])
+ model = DiffusionTts(64, channel_mult=[1,1.5,2, 3, 4, 6, 8, 8, 8, 8], num_res_blocks=[2, 2, 2, 2, 2, 2, 2, 4, 4, 4],
+ token_conditioning_resolutions=[1,4,16,64], attention_resolutions=[256,512], num_heads=4, kernel_size=3,
+ scale_factor=2, conditioning_inputs_provided=True, time_embed_dim_multiplier=4)
+ model(clip, ts, tok, cond)
+
diff --git a/codes/models/gpt_voice/unet_diffusion_tts5.py b/codes/models/gpt_voice/unet_diffusion_tts5.py
new file mode 100644
index 00000000..66654ff1
--- /dev/null
+++ b/codes/models/gpt_voice/unet_diffusion_tts5.py
@@ -0,0 +1,460 @@
+import functools
+from collections import OrderedDict
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torch import autocast
+from x_transformers.x_transformers import AbsolutePositionalEmbedding, AttentionLayers
+
+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.gpt_voice.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
+
+
+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, **kwargs):
+ kw_requires_grad = {}
+ kw_no_grad = {}
+ for k, v in kwargs.items():
+ if v is not None and isinstance(v, torch.Tensor) and v.requires_grad:
+ kw_requires_grad[k] = v
+ else:
+ kw_no_grad[k] = v
+ partial = functools.partial(self.wrap, **kw_no_grad)
+ return torch.utils.checkpoint.checkpoint(partial, x, **kw_requires_grad)
+
+
+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, **xtransformer_kwargs):
+ super().__init__()
+ self.transformer = ContinuousTransformerWrapper(**xtransformer_kwargs)
+
+ 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):
+ x = x.permute(0,2,1)
+ h = self.transformer(x)
+ 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,
+ conditioning_inputs_provided=True,
+ time_embed_dim_multiplier=4,
+ nil_guidance_fwd_proportion=.3,
+ ):
+ super().__init__()
+
+ if num_heads_upsample == -1:
+ num_heads_upsample = num_heads
+
+ 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.dtype = torch.float16 if use_fp16 else torch.float32
+ 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
+
+ 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.conditioning_enabled = conditioning_inputs_provided
+ if conditioning_inputs_provided:
+ self.contextual_embedder = AudioMiniEncoder(in_channels, 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_encoder = CheckpointedXTransformerEncoder(
+ max_seq_len=-1, # Should be unused
+ use_pos_emb=False,
+ attn_layers=Encoder(
+ dim=embedding_dim,
+ depth=8,
+ heads=num_heads,
+ ff_dropout=dropout,
+ attn_dropout=dropout,
+ use_rmsnorm=True,
+ ff_glu=True,
+ rotary_pos_emb=True,
+ )
+ )
+
+ 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=8,
+ 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 load_state_dict(self, state_dict: 'OrderedDict[str, Tensor]',
+ strict: bool = True):
+ # Temporary hack to allow the addition of nil-guidance token embeddings to the existing guidance embeddings.
+ lsd = self.state_dict()
+ revised = 0
+ for i, blk in enumerate(self.input_blocks):
+ if isinstance(blk, nn.Embedding):
+ key = f'input_blocks.{i}.weight'
+ if state_dict[key].shape[0] != lsd[key].shape[0]:
+ t = torch.randn_like(lsd[key]) * .02
+ t[:state_dict[key].shape[0]] = state_dict[key]
+ state_dict[key] = t
+ revised += 1
+ print(f"Loaded experimental unet_diffusion_net with {revised} modifications.")
+ return super().load_state_dict(state_dict, strict)
+
+
+
+ def forward(self, x, timesteps, tokens, conditioning_input=None):
+ """
+ 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.
+ :return: an [N x C x ...] Tensor of outputs.
+ """
+ with autocast(x.device.type):
+ 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]))
+ tokens = F.pad(tokens, (0,int(pc*tokens.shape[-1])))
+ if self.conditioning_enabled:
+ assert conditioning_input is not None
+
+ hs = []
+ time_emb = self.time_embed(timestep_embedding(timesteps, self.model_channels))
+
+ # Mask out guidance tokens for un-guided diffusion.
+ if self.training and self.nil_guidance_fwd_proportion > 0:
+ token_mask = torch.rand(tokens.shape, device=tokens.device) < self.nil_guidance_fwd_proportion
+ tokens = torch.where(token_mask, self.mask_token_id, tokens)
+ code_emb = self.code_embedding(tokens).permute(0,2,1)
+ if self.conditioning_enabled:
+ cond_emb = self.contextual_embedder(conditioning_input)
+ code_emb = cond_emb.unsqueeze(-1) * code_emb
+ code_emb = self.conditioning_encoder(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=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_tts5(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])
+ 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,
+ conditioning_inputs_provided=True,
+ time_embed_dim_multiplier=4)
+ model(clip, ts, tok, cond)
+ torch.save(model.state_dict(), 'test_out.pth')
+
diff --git a/codes/models/lucidrains/dalle/transformer.py b/codes/models/lucidrains/dalle/transformer.py
index 27f5f2bd..389e6849 100644
--- a/codes/models/lucidrains/dalle/transformer.py
+++ b/codes/models/lucidrains/dalle/transformer.py
@@ -146,6 +146,7 @@ class Transformer(nn.Module):
ff_dropout = 0.,
attn_types = None,
image_fmap_size = None,
+ oned_fmap_size = None,
sparse_attn = False,
stable = False,
sandwich_norm = False,
@@ -204,7 +205,7 @@ class Transformer(nn.Module):
assert 'mlp' not in attn_types, 'you cannot use gMLPs if rotary embedding is turned on'
rot_dim = dim_head // 3
- img_seq_len = (image_fmap_size ** 2)
+ img_seq_len = (image_fmap_size ** 2) if image_fmap_size is not None else oned_fmap_size
text_len = seq_len - img_seq_len + 1
text_pos_emb = RotaryEmbedding(dim = rot_dim)
@@ -214,7 +215,7 @@ class Transformer(nn.Module):
img_to_text_freqs = text_pos_emb(torch.full((img_seq_len,), 8192)) # image is given a position far away from text
text_freqs = torch.cat((text_freqs, img_to_text_freqs), dim = 0)
- img_freqs_axial = img_axial_pos_emb(torch.linspace(-1, 1, steps = image_fmap_size))
+ img_freqs_axial = img_axial_pos_emb(torch.linspace(-1, 1, steps = image_fmap_size if image_fmap_size is not None else oned_fmap_size))
img_freqs = broadcat((rearrange(img_freqs_axial, 'i d -> i () d'), rearrange(img_freqs_axial, 'j d -> () j d')), dim = -1)
img_freqs = rearrange(img_freqs, 'h w d -> (h w) d')
diff --git a/codes/train.py b/codes/train.py
index a3c4edd2..0a159119 100644
--- a/codes/train.py
+++ b/codes/train.py
@@ -299,7 +299,7 @@ class Trainer:
if __name__ == '__main__':
parser = argparse.ArgumentParser()
- parser.add_argument('-opt', type=str, help='Path to option YAML file.', default='../experiments/train_diffusion_tts_experimental_fp16/train_diffusion_tts.yml')
+ parser.add_argument('-opt', type=str, help='Path to option YAML file.', default='../options/train_diffusion_tts5_medium.yml')
parser.add_argument('--launcher', choices=['none', 'pytorch'], default='none', help='job launcher')
parser.add_argument('--local_rank', type=int, default=0)
args = parser.parse_args()