import math from abc import abstractmethod import numpy as np import torch import torch as th import torch.nn as nn import torch.nn.functional as F from einops import repeat from models.attention import SpatialTransformer def timestep_embedding(timesteps, dim, max_period=10000, repeat_only=False): """ Create sinusoidal timestep embeddings. :param timesteps: a 1-D Tensor of N indices, one per batch element. These may be fractional. :param dim: the dimension of the output. :param max_period: controls the minimum frequency of the embeddings. :return: an [N x dim] Tensor of positional embeddings. """ if not repeat_only: half = dim // 2 freqs = torch.exp( -math.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32) / half ).to(device=timesteps.device) args = timesteps[:, None].float() * freqs[None] embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1) if dim % 2: embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1) else: embedding = repeat(timesteps, 'b -> b d', d=dim) return embedding def zero_module(module): """ Zero out the parameters of a module and return it. """ for p in module.parameters(): p.detach().zero_() return module class TimestepBlock(nn.Module): @abstractmethod def forward(self, x, emb): """ Apply the module to `x` given `emb` timestep embeddings. """ class TimestepEmbedSequential(nn.Sequential, TimestepBlock): """ A sequential module that passes timestep embeddings to the children that support it as an extra input. """ def forward(self, x, emb, context=None): for layer in self: if isinstance(layer, TimestepBlock): x = layer(x, emb) elif isinstance(layer, SpatialTransformer): x = layer(x, context) else: x = layer(x) return x def Normalize(in_channels): return nn.GroupNorm( num_groups=32, num_channels=in_channels, eps=1e-6, affine=True ) def count_flops_attn(model, _x, y): """ A counter for the `thop` package to count the operations in an attention operation. Meant to be used like: macs, params = thop.profile( model, inputs=(inputs, timestamps), custom_ops={QKVAttention: QKVAttention.count_flops}, ) """ b, c, *spatial = y[0].shape num_spatial = int(np.prod(spatial)) # We perform two matmuls with the same number of ops. # The first computes the weight matrix, the second computes # the combination of the value vectors. matmul_ops = 2 * b * (num_spatial ** 2) * c model.total_ops += th.DoubleTensor([matmul_ops]) class QKVAttentionLegacy(nn.Module): """ A module which performs QKV attention. Matches legacy QKVAttention + input/ouput heads shaping """ def __init__(self, n_heads): super().__init__() self.n_heads = n_heads def forward(self, qkv): """ Apply QKV attention. :param qkv: an [N x (H * 3 * C) x T] tensor of Qs, Ks, and Vs. :return: an [N x (H * C) x T] tensor after attention. """ bs, width, length = qkv.shape assert width % (3 * self.n_heads) == 0 ch = width // (3 * self.n_heads) q, k, v = qkv.reshape(bs * self.n_heads, ch * 3, length).split(ch, dim=1) scale = 1 / math.sqrt(math.sqrt(ch)) weight = th.einsum( "bct,bcs->bts", q * scale, k * scale ) # More stable with f16 than dividing afterwards weight = th.softmax(weight.float(), dim=-1).type(weight.dtype) a = th.einsum("bts,bcs->bct", weight, v) return a.reshape(bs, -1, length) @staticmethod def count_flops(model, _x, y): return count_flops_attn(model, _x, y) class AttentionBlock(nn.Module): """ An attention block that allows spatial positions to attend to each other. Originally ported from here, but adapted to the N-d case. https://github.com/hojonathanho/diffusion/blob/1e0dceb3b3495bbe19116a5e1b3596cd0706c543/diffusion_tf/models/unet.py#L66. """ def __init__( self, channels, num_heads=1, num_head_channels=-1, use_checkpoint=False, ): super().__init__() self.channels = channels if num_head_channels == -1: self.num_heads = num_heads else: assert ( channels % num_head_channels == 0 ), f"q,k,v channels {channels} is not divisible by num_head_channels {num_head_channels}" self.num_heads = channels // num_head_channels self.use_checkpoint = use_checkpoint self.norm = Normalize(channels) self.qkv = nn.Conv1d(channels, channels * 3, 1) self.attention = QKVAttentionLegacy(self.num_heads) self.proj_out = zero_module(nn.Conv1d(channels, channels, 1)) def forward(self, x): return self._forward(x, ) def _forward(self, x): b, c, *spatial = x.shape x = x.reshape(b, c, -1) qkv = self.qkv(self.norm(x)) h = self.attention(qkv) h = self.proj_out(h) return (x + h).reshape(b, c, *spatial) class Downsample(nn.Module): """ A downsampling layer with an optional convolution. :param channels: channels in the inputs and outputs. :param use_conv: a bool determining if a convolution is applied. """ def __init__(self, channels, use_conv, out_channels=None, padding=1): super().__init__() self.channels = channels self.out_channels = out_channels or channels self.use_conv = use_conv if use_conv: self.op = nn.Conv3d( self.channels, self.out_channels, 3, stride=2, padding=padding ) else: assert self.channels == self.out_channels self.op = nn.AvgPool3d(kernel_size=2, stride=2) def forward(self, x): assert x.shape[1] == self.channels return self.op(x) class Upsample(nn.Module): """ An upsampling layer with an optional convolution. :param channels: channels in the inputs and outputs. :param use_conv: a bool determining if a convolution is applied. """ def __init__(self, channels, use_conv, out_channels=None, padding=1): super().__init__() self.channels = channels self.out_channels = out_channels or channels self.use_conv = use_conv if use_conv: self.conv = nn.Conv3d(self.channels, self.out_channels, 3, padding=padding) def forward(self, x): assert x.shape[1] == self.channels x = F.interpolate(x, scale_factor=2, mode="nearest") if self.use_conv: x = self.conv(x) return x class ResBlock(TimestepBlock): """ A residual block that can optionally change the number of channels. :param channels: the number of input channels. :param emb_channels: the number of timestep embedding channels. :param dropout: the rate of dropout. :param out_channels: if specified, the number of out channels. :param use_conv: if True and out_channels is specified, use a spatial convolution instead of a smaller 1x1 convolution to change the channels in the skip connection. :param up: if True, use this block for upsampling. :param down: if True, use this block for downsampling. """ def __init__( self, channels, emb_channels, dropout, out_channels=None, use_conv=False, use_scale_shift_norm=False, up=False, down=False, ): super().__init__() self.channels = channels self.emb_channels = emb_channels self.dropout = dropout self.out_channels = out_channels or channels self.use_conv = use_conv self.use_scale_shift_norm = use_scale_shift_norm self.in_layers = nn.Sequential( Normalize(channels), nn.SiLU(), nn.Conv3d(channels, self.out_channels, 3, padding=1), ) self.updown = up or down if up: self.h_upd = Upsample(channels, False) self.x_upd = Upsample(channels, False) elif down: self.h_upd = Downsample(channels, False) self.x_upd = Downsample(channels, False) else: self.h_upd = self.x_upd = nn.Identity() self.emb_layers = nn.Sequential( nn.SiLU(), nn.Linear( emb_channels, 2 * self.out_channels if use_scale_shift_norm else self.out_channels, ), ) self.out_layers = nn.Sequential( Normalize(self.out_channels), nn.SiLU(), nn.Dropout(p=dropout), zero_module( nn.Conv3d(self.out_channels, self.out_channels, 3, padding=1) ), ) if self.out_channels == channels: self.skip_connection = nn.Identity() elif use_conv: self.skip_connection = nn.Conv3d( channels, self.out_channels, 3, padding=1 ) else: self.skip_connection = nn.Conv3d(channels, self.out_channels, 1) def forward(self, x, emb): return self._forward(x, emb) def _forward(self, x, emb): if self.updown: in_rest, in_conv = self.in_layers[:-1], self.in_layers[-1] h = in_rest(x) h = self.h_upd(h) x = self.x_upd(x) h = in_conv(h) else: 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] if self.use_scale_shift_norm: out_norm, out_rest = self.out_layers[0], self.out_layers[1:] scale, shift = th.chunk(emb_out, 2, dim=1) h = out_norm(h) * (1 + scale) + shift h = out_rest(h) else: h = h + emb_out h = self.out_layers(h) return self.skip_connection(x) + h class UNetModel(nn.Module): def __init__( self, image_size, in_channels, model_channels, out_channels, num_res_blocks, attention_resolutions, dropout=0, channel_mult=(1, 2, 4, 8), conv_resample=True, num_classes=None, num_heads=1, num_head_channels=-1, num_heads_upsample=-1, use_scale_shift_norm=False, resblock_updown=False, use_spatial_transformer=False, # custom transformer support transformer_depth=1, # custom transformer support context_dim=None, # custom transformer support n_embed=None # custom support for prediction of discrete ids into codebook of first stage vq model ): super().__init__() if use_spatial_transformer: assert context_dim is not None, 'Fool!! You forgot to include the dimension of your cross-attention conditioning...' if context_dim is not None: assert use_spatial_transformer, 'Fool!! You forgot to use the spatial transformer for your cross-attention conditioning...' if num_heads_upsample == -1: num_heads_upsample = num_heads self.image_size = image_size self.in_channels = in_channels self.model_channels = model_channels self.out_channels = out_channels self.num_res_blocks = num_res_blocks self.attention_resolutions = attention_resolutions self.dropout = dropout self.channel_mult = channel_mult self.conv_resample = conv_resample self.num_classes = num_classes self.num_heads = num_heads self.num_head_channels = num_head_channels self.num_heads_upsample = num_heads_upsample self.predict_codebook_ids = n_embed is not None time_embed_dim = model_channels * 4 self.time_embed = nn.Sequential( nn.Linear(model_channels, time_embed_dim), nn.SiLU(), nn.Linear(time_embed_dim, time_embed_dim), ) if self.num_classes is not None: self.label_emb = nn.Embedding(num_classes, time_embed_dim) self.input_blocks = nn.ModuleList( [ TimestepEmbedSequential( nn.Conv3d(in_channels, model_channels, 3, padding=1) ) ] ) self._feature_size = model_channels input_block_chans = [model_channels] ch = model_channels ds = 1 for level, mult in enumerate(channel_mult): for _ in range(num_res_blocks): layers = [ ResBlock( ch, time_embed_dim, dropout, out_channels=mult * model_channels, use_scale_shift_norm=use_scale_shift_norm, ) ] ch = mult * model_channels if ds in attention_resolutions: dim_head = ch // num_heads layers.append( AttentionBlock( ch, num_heads=num_heads, num_head_channels=num_head_channels, ) if not use_spatial_transformer else SpatialTransformer( ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim ) ) 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( ResBlock( ch, time_embed_dim, dropout, out_channels=out_ch, use_scale_shift_norm=use_scale_shift_norm, down=True, ) if resblock_updown else Downsample( ch, conv_resample, out_channels=out_ch ) ) ) ch = out_ch input_block_chans.append(ch) ds *= 2 self._feature_size += ch dim_head = ch // num_heads self.middle_block = TimestepEmbedSequential( ResBlock( ch, time_embed_dim, dropout, use_scale_shift_norm=use_scale_shift_norm, ), AttentionBlock( ch, num_heads=num_heads, num_head_channels=num_head_channels, ) if not use_spatial_transformer else SpatialTransformer( ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim ), ResBlock( ch, time_embed_dim, dropout, use_scale_shift_norm=use_scale_shift_norm, ), ) self._feature_size += ch self.output_blocks = nn.ModuleList([]) for level, mult in list(enumerate(channel_mult))[::-1]: for i in range(num_res_blocks + 1): ich = input_block_chans.pop() layers = [ ResBlock( ch + ich, time_embed_dim, dropout, out_channels=model_channels * mult, use_scale_shift_norm=use_scale_shift_norm, ) ] ch = model_channels * mult if ds in attention_resolutions: dim_head = ch // num_heads layers.append( AttentionBlock( ch, num_heads=num_heads_upsample, num_head_channels=num_head_channels, ) if not use_spatial_transformer else SpatialTransformer( ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim ) ) if level and i == num_res_blocks: out_ch = ch layers.append( ResBlock( ch, time_embed_dim, dropout, out_channels=out_ch, use_scale_shift_norm=use_scale_shift_norm, up=True, ) if resblock_updown else Upsample(ch, conv_resample, out_channels=out_ch) ) ds //= 2 self.output_blocks.append(TimestepEmbedSequential(*layers)) self._feature_size += ch self.out = nn.Sequential( Normalize(ch), nn.SiLU(), zero_module(nn.Conv3d(model_channels, out_channels, 3, padding=1)), ) if self.predict_codebook_ids: self.id_predictor = nn.Sequential( Normalize(ch), nn.Conv3d(model_channels, n_embed, 1), ) def forward(self, x, timesteps=None, context=None, y=None, **kwargs): """ 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 context: conditioning plugged in via crossattn :param y: an [N] Tensor of labels, if class-conditional. :return: an [N x C x ...] Tensor of outputs. """ assert (y is not None) == ( self.num_classes is not None ), "must specify y if and only if the model is class-conditional" assert timesteps is not None, 'need to implement no-timestep usage' hs = [] t_emb = timestep_embedding(timesteps, self.model_channels, repeat_only=False) emb = self.time_embed(t_emb) if self.num_classes is not None: assert y.shape == (x.shape[0],) emb = emb + self.label_emb(y) h = x for module in self.input_blocks: h = module(h, emb, context) hs.append(h) h = self.middle_block(h, emb, context) for module in self.output_blocks: h = th.cat([h, hs.pop()], dim=1) h = module(h, emb, context) if self.predict_codebook_ids: # return self.out(h), self.id_predictor(h) return self.id_predictor(h) else: return self.out(h)