from abc import abstractmethod import math from einops import rearrange from functools import partial import numpy as np import torch as th import torch.nn as nn import torch.nn.functional as F from omegaconf.listconfig import ListConfig from lvdm.models.modules.util import ( checkpoint, conv_nd, linear, avg_pool_nd, zero_module, normalization, timestep_embedding, nonlinearity, ) # dummy replace def convert_module_to_f16(x): pass def convert_module_to_f32(x): pass ## go # --------------------------------------------------------------------------------------------------- class TimestepBlock(nn.Module): """ Any module where forward() takes timestep embeddings as a second argument. """ @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, **kwargs): for layer in self: if isinstance(layer, TimestepBlock): x = layer(x, emb, **kwargs) elif isinstance(layer, STTransformerClass): x = layer(x, context, **kwargs) else: x = layer(x) return 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. :param dims: determines if the signal is 1D, 2D, or 3D. If 3D, then upsampling occurs in the inner-two dimensions. """ def __init__(self, channels, use_conv, dims=2, out_channels=None, kernel_size_t=3, padding_t=1, ): super().__init__() self.channels = channels self.out_channels = out_channels or channels self.use_conv = use_conv self.dims = dims if use_conv: self.conv = conv_nd(dims, self.channels, self.out_channels, (kernel_size_t, 3,3), padding=(padding_t, 1,1)) def forward(self, x): assert x.shape[1] == self.channels if self.dims == 3: x = F.interpolate( x, (x.shape[2], x.shape[3] * 2, x.shape[4] * 2), mode="nearest" ) else: x = F.interpolate(x, scale_factor=2, mode="nearest") if self.use_conv: x = self.conv(x) return x # --------------------------------------------------------------------------------------------------- class TransposedUpsample(nn.Module): 'Learned 2x upsampling without padding' def __init__(self, channels, out_channels=None, ks=5): super().__init__() self.channels = channels self.out_channels = out_channels or channels self.up = nn.ConvTranspose2d(self.channels,self.out_channels,kernel_size=ks,stride=2) def forward(self,x): return self.up(x) # --------------------------------------------------------------------------------------------------- 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. :param dims: determines if the signal is 1D, 2D, or 3D. If 3D, then downsampling occurs in the inner-two dimensions. """ def __init__(self, channels, use_conv, dims=2, out_channels=None, kernel_size_t=3, padding_t=1, ): super().__init__() self.channels = channels self.out_channels = out_channels or channels self.use_conv = use_conv self.dims = dims stride = 2 if dims != 3 else (1, 2, 2) if use_conv: self.op = conv_nd( dims, self.channels, self.out_channels, (kernel_size_t, 3,3), stride=stride, padding=(padding_t, 1,1) ) else: assert self.channels == self.out_channels self.op = avg_pool_nd(dims, kernel_size=stride, stride=stride) def forward(self, x): assert x.shape[1] == self.channels return self.op(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 dims: determines if the signal is 1D, 2D, or 3D. :param use_checkpoint: if True, use gradient checkpointing on this module. :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, dims=2, use_checkpoint=False, up=False, down=False, # temporal kernel_size_t=3, padding_t=1, nonlinearity_type='silu', **kwargs ): 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_checkpoint = use_checkpoint self.use_scale_shift_norm = use_scale_shift_norm self.nonlinearity_type = nonlinearity_type self.in_layers = nn.Sequential( normalization(channels), nonlinearity(nonlinearity_type), conv_nd(dims, channels, self.out_channels, (kernel_size_t, 3,3), padding=(padding_t, 1,1)), ) self.updown = up or down if up: self.h_upd = Upsample(channels, False, dims, kernel_size_t=kernel_size_t, padding_t=padding_t) self.x_upd = Upsample(channels, False, dims, kernel_size_t=kernel_size_t, padding_t=padding_t) elif down: self.h_upd = Downsample(channels, False, dims, kernel_size_t=kernel_size_t, padding_t=padding_t) self.x_upd = Downsample(channels, False, dims, kernel_size_t=kernel_size_t, padding_t=padding_t) else: self.h_upd = self.x_upd = nn.Identity() self.emb_layers = nn.Sequential( nonlinearity(nonlinearity_type), linear( emb_channels, 2 * self.out_channels if use_scale_shift_norm else self.out_channels, ), ) self.out_layers = nn.Sequential( normalization(self.out_channels), nonlinearity(nonlinearity_type), nn.Dropout(p=dropout), zero_module( conv_nd(dims, self.out_channels, self.out_channels, (kernel_size_t, 3,3), padding=(padding_t, 1,1)) ), ) if self.out_channels == channels: self.skip_connection = nn.Identity() elif use_conv: self.skip_connection = conv_nd( dims, channels, self.out_channels, (kernel_size_t, 3,3), padding=(padding_t, 1,1) ) else: self.skip_connection = conv_nd(dims, channels, self.out_channels, 1) def forward(self, x, emb, **kwargs): """ 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), self.parameters(), self.use_checkpoint ) 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) if emb_out.dim() == 3: # btc for video data emb_out = rearrange(emb_out, 'b t c -> b c t') while len(emb_out.shape) < h.dim(): emb_out = emb_out[..., None] # bct -> bct11 or bc -> bc111 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) out = self.skip_connection(x) + h return out # --------------------------------------------------------------------------------------------------- def make_spatialtemporal_transformer(module_name='attention_temporal', class_name='SpatialTemporalTransformer'): module = __import__(f"lvdm.models.modules.{module_name}", fromlist=[class_name]) global STTransformerClass STTransformerClass = getattr(module, class_name) return STTransformerClass # --------------------------------------------------------------------------------------------------- class UNetModel(nn.Module): """ The full UNet model with attention and timestep embedding. :param in_channels: channels in the input Tensor. :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_classes: if specified (as an int), then this model will be class-conditional with `num_classes` classes. :param use_checkpoint: use gradient checkpointing to reduce memory usage. :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, image_size, # not used in UNetModel in_channels, model_channels, out_channels, num_res_blocks, attention_resolutions, dropout=0, channel_mult=(1, 2, 4, 8), conv_resample=True, dims=3, num_classes=None, use_checkpoint=False, use_fp16=False, num_heads=-1, num_head_channels=-1, num_heads_upsample=-1, use_scale_shift_norm=False, resblock_updown=False, transformer_depth=1, # custom transformer support context_dim=None, # custom transformer support legacy=True, # temporal related kernel_size_t=1, padding_t=1, use_temporal_transformer=True, temporal_length=None, use_relative_position=False, cross_attn_on_tempoal=False, temporal_crossattn_type="crossattn", order="stst", nonlinearity_type='silu', temporalcrossfirst=False, split_stcontext=False, temporal_context_dim=None, use_tempoal_causal_attn=False, ST_transformer_module='attention_temporal', ST_transformer_class='SpatialTemporalTransformer', **kwargs, ): super().__init__() assert(use_temporal_transformer) if context_dim is not None: if type(context_dim) == ListConfig: context_dim = list(context_dim) if num_heads_upsample == -1: num_heads_upsample = num_heads if num_heads == -1: assert num_head_channels != -1, 'Either num_heads or num_head_channels has to be set' if num_head_channels == -1: assert num_heads != -1, 'Either num_heads or num_head_channels has to be set' 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.use_checkpoint = use_checkpoint self.dtype = th.float16 if use_fp16 else th.float32 self.num_heads = num_heads self.num_head_channels = num_head_channels self.num_heads_upsample = num_heads_upsample self.use_relative_position = use_relative_position self.temporal_length = temporal_length self.cross_attn_on_tempoal = cross_attn_on_tempoal self.temporal_crossattn_type = temporal_crossattn_type self.order = order self.temporalcrossfirst = temporalcrossfirst self.split_stcontext = split_stcontext self.temporal_context_dim = temporal_context_dim self.nonlinearity_type = nonlinearity_type self.use_tempoal_causal_attn = use_tempoal_causal_attn time_embed_dim = model_channels * 4 self.time_embed_dim = time_embed_dim self.time_embed = nn.Sequential( linear(model_channels, time_embed_dim), nonlinearity(nonlinearity_type), linear(time_embed_dim, time_embed_dim), ) if self.num_classes is not None: self.label_emb = nn.Embedding(num_classes, time_embed_dim) STTransformerClass = make_spatialtemporal_transformer(module_name=ST_transformer_module, class_name=ST_transformer_class) self.input_blocks = nn.ModuleList( [ TimestepEmbedSequential( conv_nd(dims, in_channels, model_channels, (kernel_size_t, 3,3), padding=(padding_t, 1,1)) ) ] ) 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, dims=dims, use_checkpoint=use_checkpoint, use_scale_shift_norm=use_scale_shift_norm, kernel_size_t=kernel_size_t, padding_t=padding_t, nonlinearity_type=nonlinearity_type, **kwargs ) ] ch = mult * model_channels if ds in attention_resolutions: if num_head_channels == -1: dim_head = ch // num_heads else: num_heads = ch // num_head_channels dim_head = num_head_channels if legacy: dim_head = ch // num_heads if use_temporal_transformer else num_head_channels layers.append(STTransformerClass( ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim, # temporal related temporal_length=temporal_length, use_relative_position=use_relative_position, cross_attn_on_tempoal=cross_attn_on_tempoal, temporal_crossattn_type=temporal_crossattn_type, order=order, temporalcrossfirst=temporalcrossfirst, split_stcontext=split_stcontext, temporal_context_dim=temporal_context_dim, use_tempoal_causal_attn=use_tempoal_causal_attn, **kwargs, )) 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, dims=dims, use_checkpoint=use_checkpoint, use_scale_shift_norm=use_scale_shift_norm, down=True, kernel_size_t=kernel_size_t, padding_t=padding_t, nonlinearity_type=nonlinearity_type, **kwargs ) if resblock_updown else Downsample( ch, conv_resample, dims=dims, out_channels=out_ch, kernel_size_t=kernel_size_t, padding_t=padding_t ) ) ) ch = out_ch input_block_chans.append(ch) ds *= 2 self._feature_size += ch if num_head_channels == -1: dim_head = ch // num_heads else: num_heads = ch // num_head_channels dim_head = num_head_channels if legacy: dim_head = ch // num_heads if use_temporal_transformer else num_head_channels self.middle_block = TimestepEmbedSequential( ResBlock( ch, time_embed_dim, dropout, dims=dims, use_checkpoint=use_checkpoint, use_scale_shift_norm=use_scale_shift_norm, kernel_size_t=kernel_size_t, padding_t=padding_t, nonlinearity_type=nonlinearity_type, **kwargs ), STTransformerClass( ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim, # temporal related temporal_length=temporal_length, use_relative_position=use_relative_position, cross_attn_on_tempoal=cross_attn_on_tempoal, temporal_crossattn_type=temporal_crossattn_type, order=order, temporalcrossfirst=temporalcrossfirst, split_stcontext=split_stcontext, temporal_context_dim=temporal_context_dim, use_tempoal_causal_attn=use_tempoal_causal_attn, **kwargs, ), ResBlock( ch, time_embed_dim, dropout, dims=dims, use_checkpoint=use_checkpoint, use_scale_shift_norm=use_scale_shift_norm, kernel_size_t=kernel_size_t, padding_t=padding_t, nonlinearity_type=nonlinearity_type, **kwargs ), ) 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, dims=dims, use_checkpoint=use_checkpoint, use_scale_shift_norm=use_scale_shift_norm, kernel_size_t=kernel_size_t, padding_t=padding_t, nonlinearity_type=nonlinearity_type, **kwargs ) ] ch = model_channels * mult if ds in attention_resolutions: if num_head_channels == -1: dim_head = ch // num_heads else: num_heads = ch // num_head_channels dim_head = num_head_channels if legacy: dim_head = ch // num_heads if use_temporal_transformer else num_head_channels layers.append( STTransformerClass( ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim, # temporal related temporal_length=temporal_length, use_relative_position=use_relative_position, cross_attn_on_tempoal=cross_attn_on_tempoal, temporal_crossattn_type=temporal_crossattn_type, order=order, temporalcrossfirst=temporalcrossfirst, split_stcontext=split_stcontext, temporal_context_dim=temporal_context_dim, use_tempoal_causal_attn=use_tempoal_causal_attn, **kwargs, ) ) if level and i == num_res_blocks: out_ch = ch layers.append( ResBlock( ch, time_embed_dim, dropout, out_channels=out_ch, dims=dims, use_checkpoint=use_checkpoint, use_scale_shift_norm=use_scale_shift_norm, up=True, kernel_size_t=kernel_size_t, padding_t=padding_t, nonlinearity_type=nonlinearity_type, **kwargs ) if resblock_updown else Upsample(ch, conv_resample, dims=dims, out_channels=out_ch, kernel_size_t=kernel_size_t, padding_t=padding_t) ) ds //= 2 self.output_blocks.append(TimestepEmbedSequential(*layers)) self._feature_size += ch self.out = nn.Sequential( normalization(ch), nonlinearity(nonlinearity_type), zero_module(conv_nd(dims, model_channels, out_channels, (kernel_size_t, 3,3), padding=(padding_t, 1,1))), ) def convert_to_fp16(self): """ Convert the torso of the model to float16. """ self.input_blocks.apply(convert_module_to_f16) self.middle_block.apply(convert_module_to_f16) self.output_blocks.apply(convert_module_to_f16) def convert_to_fp32(self): """ Convert the torso of the model to float32. """ self.input_blocks.apply(convert_module_to_f32) self.middle_block.apply(convert_module_to_f32) self.output_blocks.apply(convert_module_to_f32) def forward(self, x, timesteps=None, time_emb_replace=None, context=None, features_adapter=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. """ hs = [] if time_emb_replace is None: t_emb = timestep_embedding(timesteps, self.model_channels, repeat_only=False) emb = self.time_embed(t_emb) else: emb = time_emb_replace if y is not None: # if class-conditional model, inject class labels assert y.shape == (x.shape[0],) emb = emb + self.label_emb(y) h = x.type(self.dtype) adapter_idx = 0 for id, module in enumerate(self.input_blocks): h = module(h, emb, context, **kwargs) ## plug-in adapter features if ((id+1)%3 == 0) and features_adapter is not None: h = h + features_adapter[adapter_idx] adapter_idx += 1 hs.append(h) if features_adapter is not None: assert len(features_adapter)==adapter_idx, 'Mismatch features adapter' h = self.middle_block(h, emb, context, **kwargs) for module in self.output_blocks: h = th.cat([h, hs.pop()], dim=1) h = module(h, emb, context, **kwargs) h = h.type(x.dtype) return self.out(h)