# Copyright (c) Alibaba, Inc. and its affiliates. import math import os from functools import partial import torch import torch.nn as nn import torch.nn.functional as F from einops import rearrange from fairscale.nn.checkpoint import checkpoint_wrapper from rotary_embedding_torch import RotaryEmbedding from torch import einsum __all__ = ['UNetSD_temporal'] USE_TEMPORAL_TRANSFORMER = True # load all keys started with prefix and replace them with new_prefix def load_Block(state, prefix, new_prefix=None): if new_prefix is None: new_prefix = prefix state_dict = {} state = {key: value for key, value in state.items() if prefix in key} for key, value in state.items(): new_key = key.replace(prefix, new_prefix) state_dict[new_key] = value return state_dict def load_2d_pretrained_state_dict(state, cfg): new_state_dict = {} dim = cfg.unet_dim num_res_blocks = cfg.unet_res_blocks dim_mult = cfg.unet_dim_mult attn_scales = cfg.unet_attn_scales # params enc_dims = [dim * u for u in [1] + dim_mult] dec_dims = [dim * u for u in [dim_mult[-1]] + dim_mult[::-1]] shortcut_dims = [] scale = 1.0 # embeddings state_dict = load_Block(state, prefix='time_embedding') new_state_dict.update(state_dict) state_dict = load_Block(state, prefix='y_embedding') new_state_dict.update(state_dict) state_dict = load_Block(state, prefix='context_embedding') new_state_dict.update(state_dict) encoder_idx = 0 # init block state_dict = load_Block( state, prefix=f'encoder.{encoder_idx}', new_prefix=f'encoder.{encoder_idx}.0') new_state_dict.update(state_dict) encoder_idx += 1 shortcut_dims.append(dim) for i, (in_dim, out_dim) in enumerate(zip(enc_dims[:-1], enc_dims[1:])): for j in range(num_res_blocks): # residual (+attention) blocks idx = 0 idx_ = 0 # residual (+attention) blocks state_dict = load_Block( state, prefix=f'encoder.{encoder_idx}.{idx}', new_prefix=f'encoder.{encoder_idx}.{idx_}') new_state_dict.update(state_dict) idx += 1 idx_ = 2 if scale in attn_scales: # block.append(AttentionBlock(out_dim, context_dim, num_heads, head_dim)) state_dict = load_Block( state, prefix=f'encoder.{encoder_idx}.{idx}', new_prefix=f'encoder.{encoder_idx}.{idx_}') new_state_dict.update(state_dict) in_dim = out_dim encoder_idx += 1 shortcut_dims.append(out_dim) # downsample if i != len(dim_mult) - 1 and j == num_res_blocks - 1: state_dict = load_Block( state, prefix='encoder.{encoder_idx}', new_prefix='encoder.{encoder_idx}.0') new_state_dict.update(state_dict) shortcut_dims.append(out_dim) scale /= 2.0 encoder_idx += 1 # middle middle_idx = 0 state_dict = load_Block(state, prefix=f'middle.{middle_idx}') new_state_dict.update(state_dict) middle_idx += 2 state_dict = load_Block( state, prefix='middle.1', new_prefix=f'middle.{middle_idx}') new_state_dict.update(state_dict) middle_idx += 1 for _ in range(cfg.temporal_attn_times): middle_idx += 1 state_dict = load_Block( state, prefix='middle.2', new_prefix=f'middle.{middle_idx}') new_state_dict.update(state_dict) middle_idx += 2 decoder_idx = 0 for i, (in_dim, out_dim) in enumerate(zip(dec_dims[:-1], dec_dims[1:])): for j in range(num_res_blocks + 1): idx = 0 idx_ = 0 # residual (+attention) blocks state_dict = load_Block( state, prefix=f'decoder.{decoder_idx}.{idx}', new_prefix=f'decoder.{decoder_idx}.{idx_}') new_state_dict.update(state_dict) idx += 1 idx_ += 2 if scale in attn_scales: state_dict = load_Block( state, prefix=f'decoder.{decoder_idx}.{idx}', new_prefix=f'decoder.{decoder_idx}.{idx_}') new_state_dict.update(state_dict) idx += 1 idx_ += 1 for _ in range(cfg.temporal_attn_times): idx_ += 1 # upsample if i != len(dim_mult) - 1 and j == num_res_blocks: state_dict = load_Block( state, prefix=f'decoder.{decoder_idx}.{idx}', new_prefix=f'decoder.{decoder_idx}.{idx_}') new_state_dict.update(state_dict) idx += 1 idx_ += 2 scale *= 2.0 decoder_idx += 1 state_dict = load_Block(state, prefix='head') new_state_dict.update(state_dict) return new_state_dict def sinusoidal_embedding(timesteps, dim): # check input half = dim // 2 timesteps = timesteps.float() # compute sinusoidal embedding sinusoid = torch.outer( timesteps, torch.pow(10000, -torch.arange(half).to(timesteps).div(half))) x = torch.cat([torch.cos(sinusoid), torch.sin(sinusoid)], dim=1) if dim % 2 != 0: x = torch.cat([x, torch.zeros_like(x[:, :1])], dim=1) return x def exists(x): return x is not None def default(val, d): if exists(val): return val return d() if callable(d) else d def prob_mask_like(shape, prob, device): if prob == 1: return torch.ones(shape, device=device, dtype=torch.bool) elif prob == 0: return torch.zeros(shape, device=device, dtype=torch.bool) else: mask = torch.zeros(shape, device=device).float().uniform_(0, 1) < prob # avoid mask all, which will cause find_unused_parameters error if mask.all(): mask[0] = False return mask class RelativePositionBias(nn.Module): def __init__(self, heads=8, num_buckets=32, max_distance=128): super().__init__() self.num_buckets = num_buckets self.max_distance = max_distance self.relative_attention_bias = nn.Embedding(num_buckets, heads) @staticmethod def _relative_position_bucket(relative_position, num_buckets=32, max_distance=128): ret = 0 n = -relative_position num_buckets //= 2 ret += (n < 0).long() * num_buckets n = torch.abs(n) max_exact = num_buckets // 2 is_small = n < max_exact val_if_large = max_exact + ( torch.log(n.float() / max_exact) / math.log(max_distance / max_exact) * # noqa (num_buckets - max_exact)).long() val_if_large = torch.min( val_if_large, torch.full_like(val_if_large, num_buckets - 1)) ret += torch.where(is_small, n, val_if_large) return ret def forward(self, n, device): q_pos = torch.arange(n, dtype=torch.long, device=device) k_pos = torch.arange(n, dtype=torch.long, device=device) rel_pos = rearrange(k_pos, 'j -> 1 j') - rearrange(q_pos, 'i -> i 1') rp_bucket = self._relative_position_bucket( rel_pos, num_buckets=self.num_buckets, max_distance=self.max_distance) values = self.relative_attention_bias(rp_bucket) return rearrange(values, 'i j h -> h i j') class SpatialTransformer(nn.Module): """ Transformer block for image-like data. First, project the input (aka embedding) and reshape to b, t, d. Then apply standard transformer action. Finally, reshape to image NEW: use_linear for more efficiency instead of the 1x1 convs """ def __init__(self, in_channels, n_heads, d_head, depth=1, dropout=0., context_dim=None, disable_self_attn=False, use_linear=False, use_checkpoint=True): super().__init__() if exists(context_dim) and not isinstance(context_dim, list): context_dim = [context_dim] self.in_channels = in_channels inner_dim = n_heads * d_head self.norm = torch.nn.GroupNorm( num_groups=32, num_channels=in_channels, eps=1e-6, affine=True) if not use_linear: self.proj_in = nn.Conv2d( in_channels, inner_dim, kernel_size=1, stride=1, padding=0) else: self.proj_in = nn.Linear(in_channels, inner_dim) self.transformer_blocks = nn.ModuleList([ BasicTransformerBlock( inner_dim, n_heads, d_head, dropout=dropout, context_dim=context_dim[d], disable_self_attn=disable_self_attn, checkpoint=use_checkpoint) for d in range(depth) ]) if not use_linear: self.proj_out = zero_module( nn.Conv2d( inner_dim, in_channels, kernel_size=1, stride=1, padding=0)) else: self.proj_out = zero_module(nn.Linear(in_channels, inner_dim)) self.use_linear = use_linear def forward(self, x, context=None): # note: if no context is given, cross-attention defaults to self-attention if not isinstance(context, list): context = [context] b, c, h, w = x.shape x_in = x x = self.norm(x) if not self.use_linear: x = self.proj_in(x) x = rearrange(x, 'b c h w -> b (h w) c').contiguous() if self.use_linear: x = self.proj_in(x) for i, block in enumerate(self.transformer_blocks): x = block(x, context=context[i]) if self.use_linear: x = self.proj_out(x) x = rearrange(x, 'b (h w) c -> b c h w', h=h, w=w).contiguous() if not self.use_linear: x = self.proj_out(x) return x + x_in _ATTN_PRECISION = os.environ.get('ATTN_PRECISION', 'fp32') class CrossAttention(nn.Module): def __init__(self, query_dim, context_dim=None, heads=8, dim_head=64, dropout=0.): super().__init__() inner_dim = dim_head * heads context_dim = default(context_dim, query_dim) self.scale = dim_head**-0.5 self.heads = heads self.to_q = nn.Linear(query_dim, inner_dim, bias=False) self.to_k = nn.Linear(context_dim, inner_dim, bias=False) self.to_v = nn.Linear(context_dim, inner_dim, bias=False) self.to_out = nn.Sequential( nn.Linear(inner_dim, query_dim), nn.Dropout(dropout)) def forward(self, x, context=None, mask=None): h = self.heads q = self.to_q(x) context = default(context, x) k = self.to_k(context) v = self.to_v(context) q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h=h), (q, k, v)) # force cast to fp32 to avoid overflowing if _ATTN_PRECISION == 'fp32': with torch.autocast(enabled=False, device_type='cuda'): q, k = q.float(), k.float() sim = torch.einsum('b i d, b j d -> b i j', q, k) * self.scale else: sim = torch.einsum('b i d, b j d -> b i j', q, k) * self.scale del q, k if exists(mask): mask = rearrange(mask, 'b ... -> b (...)') max_neg_value = -torch.finfo(sim.dtype).max mask = repeat(mask, 'b j -> (b h) () j', h=h) sim.masked_fill_(~mask, max_neg_value) # attention, what we cannot get enough of sim = sim.softmax(dim=-1) out = torch.einsum('b i j, b j d -> b i d', sim, v) out = rearrange(out, '(b h) n d -> b n (h d)', h=h) return self.to_out(out) class BasicTransformerBlock(nn.Module): def __init__(self, dim, n_heads, d_head, dropout=0., context_dim=None, gated_ff=True, checkpoint=True, disable_self_attn=False): super().__init__() attn_cls = CrossAttention self.disable_self_attn = disable_self_attn self.attn1 = attn_cls( query_dim=dim, heads=n_heads, dim_head=d_head, dropout=dropout, context_dim=context_dim if self.disable_self_attn else None) self.ff = FeedForward(dim, dropout=dropout, glu=gated_ff) self.attn2 = attn_cls( query_dim=dim, context_dim=context_dim, heads=n_heads, dim_head=d_head, dropout=dropout) self.norm1 = nn.LayerNorm(dim) self.norm2 = nn.LayerNorm(dim) self.norm3 = nn.LayerNorm(dim) self.checkpoint = checkpoint def forward_(self, x, context=None): return checkpoint(self._forward, (x, context), self.parameters(), self.checkpoint) def forward(self, x, context=None): x = self.attn1( self.norm1(x), context=context if self.disable_self_attn else None) + x x = self.attn2(self.norm2(x), context=context) + x x = self.ff(self.norm3(x)) + x return x # feedforward class GEGLU(nn.Module): def __init__(self, dim_in, dim_out): super().__init__() self.proj = nn.Linear(dim_in, dim_out * 2) def forward(self, x): x, gate = self.proj(x).chunk(2, dim=-1) return x * F.gelu(gate) 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 FeedForward(nn.Module): def __init__(self, dim, dim_out=None, mult=4, glu=False, dropout=0.): super().__init__() inner_dim = int(dim * mult) dim_out = default(dim_out, dim) project_in = nn.Sequential(nn.Linear( dim, inner_dim), nn.GELU()) if not glu else GEGLU(dim, inner_dim) self.net = nn.Sequential(project_in, nn.Dropout(dropout), nn.Linear(inner_dim, dim_out)) def forward(self, x): return self.net(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, padding=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 = nn.Conv2d( self.channels, self.out_channels, 3, padding=padding) 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 ResBlock(nn.Module): """ 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, up=False, down=False, use_temporal_conv=True, use_image_dataset=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.use_temporal_conv = use_temporal_conv self.in_layers = nn.Sequential( nn.GroupNorm(32, channels), nn.SiLU(), nn.Conv2d(channels, self.out_channels, 3, padding=1), ) self.updown = up or down if up: self.h_upd = Upsample(channels, False, dims) self.x_upd = Upsample(channels, False, dims) elif down: self.h_upd = Downsample(channels, False, dims) self.x_upd = Downsample(channels, False, dims) 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( nn.GroupNorm(32, self.out_channels), nn.SiLU(), nn.Dropout(p=dropout), zero_module( nn.Conv2d(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 = conv_nd( dims, channels, self.out_channels, 3, padding=1) else: self.skip_connection = nn.Conv2d(channels, self.out_channels, 1) if self.use_temporal_conv: self.temopral_conv = TemporalConvBlock_v2( self.out_channels, self.out_channels, dropout=0.1, use_image_dataset=use_image_dataset) def forward(self, x, emb, batch_size): """ 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 self._forward(x, emb, batch_size) def _forward(self, x, emb, batch_size): 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) h = self.skip_connection(x) + h if self.use_temporal_conv: h = rearrange(h, '(b f) c h w -> b c f h w', b=batch_size) h = self.temopral_conv(h) h = rearrange(h, 'b c f h w -> (b f) c h w') return h 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, padding=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 = nn.Conv2d( self.channels, self.out_channels, 3, stride=stride, padding=padding) 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 Resample(nn.Module): def __init__(self, in_dim, out_dim, mode): assert mode in ['none', 'upsample', 'downsample'] super(Resample, self).__init__() self.in_dim = in_dim self.out_dim = out_dim self.mode = mode def forward(self, x, reference=None): if self.mode == 'upsample': assert reference is not None x = F.interpolate(x, size=reference.shape[-2:], mode='nearest') elif self.mode == 'downsample': x = F.adaptive_avg_pool2d( x, output_size=tuple(u // 2 for u in x.shape[-2:])) return x class ResidualBlock(nn.Module): def __init__(self, in_dim, embed_dim, out_dim, use_scale_shift_norm=True, mode='none', dropout=0.0): super(ResidualBlock, self).__init__() self.in_dim = in_dim self.embed_dim = embed_dim self.out_dim = out_dim self.use_scale_shift_norm = use_scale_shift_norm self.mode = mode # layers self.layer1 = nn.Sequential( nn.GroupNorm(32, in_dim), nn.SiLU(), nn.Conv2d(in_dim, out_dim, 3, padding=1)) self.resample = Resample(in_dim, in_dim, mode) self.embedding = nn.Sequential( nn.SiLU(), nn.Linear(embed_dim, out_dim * 2 if use_scale_shift_norm else out_dim)) self.layer2 = nn.Sequential( nn.GroupNorm(32, out_dim), nn.SiLU(), nn.Dropout(dropout), nn.Conv2d(out_dim, out_dim, 3, padding=1)) self.shortcut = nn.Identity() if in_dim == out_dim else nn.Conv2d( in_dim, out_dim, 1) # zero out the last layer params nn.init.zeros_(self.layer2[-1].weight) def forward(self, x, e, reference=None): identity = self.resample(x, reference) x = self.layer1[-1](self.resample(self.layer1[:-1](x), reference)) e = self.embedding(e).unsqueeze(-1).unsqueeze(-1).type(x.dtype) if self.use_scale_shift_norm: scale, shift = e.chunk(2, dim=1) x = self.layer2[0](x) * (1 + scale) + shift x = self.layer2[1:](x) else: x = x + e x = self.layer2(x) x = x + self.shortcut(identity) return x class AttentionBlock(nn.Module): def __init__(self, dim, context_dim=None, num_heads=None, head_dim=None): # consider head_dim first, then num_heads num_heads = dim // head_dim if head_dim else num_heads head_dim = dim // num_heads assert num_heads * head_dim == dim super(AttentionBlock, self).__init__() self.dim = dim self.context_dim = context_dim self.num_heads = num_heads self.head_dim = head_dim self.scale = math.pow(head_dim, -0.25) # layers self.norm = nn.GroupNorm(32, dim) self.to_qkv = nn.Conv2d(dim, dim * 3, 1) if context_dim is not None: self.context_kv = nn.Linear(context_dim, dim * 2) self.proj = nn.Conv2d(dim, dim, 1) # zero out the last layer params nn.init.zeros_(self.proj.weight) def forward(self, x, context=None): r"""x: [B, C, H, W]. context: [B, L, C] or None. """ identity = x b, c, h, w, n, d = *x.size(), self.num_heads, self.head_dim # compute query, key, value x = self.norm(x) q, k, v = self.to_qkv(x).view(b, n * 3, d, h * w).chunk(3, dim=1) if context is not None: ck, cv = self.context_kv(context).reshape(b, -1, n * 2, d).permute(0, 2, 3, 1).chunk( 2, dim=1) k = torch.cat([ck, k], dim=-1) v = torch.cat([cv, v], dim=-1) # compute attention attn = torch.matmul(q.transpose(-1, -2) * self.scale, k * self.scale) attn = F.softmax(attn, dim=-1) # gather context x = torch.matmul(v, attn.transpose(-1, -2)) x = x.reshape(b, c, h, w) # output x = self.proj(x) return x + identity class TemporalAttentionBlock(nn.Module): def __init__(self, dim, heads=4, dim_head=32, rotary_emb=None, use_image_dataset=False, use_sim_mask=False): super().__init__() # consider num_heads first, as pos_bias needs fixed num_heads dim_head = dim // heads assert heads * dim_head == dim self.use_image_dataset = use_image_dataset self.use_sim_mask = use_sim_mask self.scale = dim_head**-0.5 self.heads = heads hidden_dim = dim_head * heads self.norm = nn.GroupNorm(32, dim) self.rotary_emb = rotary_emb self.to_qkv = nn.Linear(dim, hidden_dim * 3) self.to_out = nn.Linear(hidden_dim, dim) def forward(self, x, pos_bias=None, focus_present_mask=None, video_mask=None): identity = x n, height, device = x.shape[2], x.shape[-2], x.device x = self.norm(x) x = rearrange(x, 'b c f h w -> b (h w) f c') qkv = self.to_qkv(x).chunk(3, dim=-1) if exists(focus_present_mask) and focus_present_mask.all(): # if all batch samples are focusing on present # it would be equivalent to passing that token's values （v=qkv[-1]） through to the output values = qkv[-1] out = self.to_out(values) out = rearrange(out, 'b (h w) f c -> b c f h w', h=height) return out + identity # split out heads q = rearrange(qkv[0], '... n (h d) -> ... h n d', h=self.heads) k = rearrange(qkv[1], '... n (h d) -> ... h n d', h=self.heads) v = rearrange(qkv[2], '... n (h d) -> ... h n d', h=self.heads) # scale q = q * self.scale # rotate positions into queries and keys for time attention if exists(self.rotary_emb): q = self.rotary_emb.rotate_queries_or_keys(q) k = self.rotary_emb.rotate_queries_or_keys(k) # similarity sim = torch.einsum('... h i d, ... h j d -> ... h i j', q, k) # relative positional bias if exists(pos_bias): sim = sim + pos_bias if (focus_present_mask is None and video_mask is not None): mask = video_mask[:, None, :] * video_mask[:, :, None] mask = mask.unsqueeze(1).unsqueeze(1) sim = sim.masked_fill(~mask, -torch.finfo(sim.dtype).max) elif exists(focus_present_mask) and not (~focus_present_mask).all(): attend_all_mask = torch.ones((n, n), device=device, dtype=torch.bool) attend_self_mask = torch.eye(n, device=device, dtype=torch.bool) mask = torch.where( rearrange(focus_present_mask, 'b -> b 1 1 1 1'), rearrange(attend_self_mask, 'i j -> 1 1 1 i j'), rearrange(attend_all_mask, 'i j -> 1 1 1 i j'), ) sim = sim.masked_fill(~mask, -torch.finfo(sim.dtype).max) if self.use_sim_mask: sim_mask = torch.tril( torch.ones((n, n), device=device, dtype=torch.bool), diagonal=0) sim = sim.masked_fill(~sim_mask, -torch.finfo(sim.dtype).max) # numerical stability sim = sim - sim.amax(dim=-1, keepdim=True).detach() attn = sim.softmax(dim=-1) # aggregate values out = torch.einsum('... h i j, ... h j d -> ... h i d', attn, v) out = rearrange(out, '... h n d -> ... n (h d)') out = self.to_out(out) out = rearrange(out, 'b (h w) f c -> b c f h w', h=height) if self.use_image_dataset: out = identity + 0 * out else: out = identity + out return out class TemporalTransformer(nn.Module): """ Transformer block for image-like data. First, project the input (aka embedding) and reshape to b, t, d. Then apply standard transformer action. Finally, reshape to image """ def __init__(self, in_channels, n_heads, d_head, depth=1, dropout=0., context_dim=None, disable_self_attn=False, use_linear=False, use_checkpoint=True, only_self_att=True, multiply_zero=False): super().__init__() self.multiply_zero = multiply_zero self.only_self_att = only_self_att self.use_adaptor = False if self.only_self_att: context_dim = None if not isinstance(context_dim, list): context_dim = [context_dim] self.in_channels = in_channels inner_dim = n_heads * d_head self.norm = torch.nn.GroupNorm( num_groups=32, num_channels=in_channels, eps=1e-6, affine=True) if not use_linear: self.proj_in = nn.Conv1d( in_channels, inner_dim, kernel_size=1, stride=1, padding=0) else: self.proj_in = nn.Linear(in_channels, inner_dim) if self.use_adaptor: self.adaptor_in = nn.Linear(frames, frames) self.transformer_blocks = nn.ModuleList([ BasicTransformerBlock( inner_dim, n_heads, d_head, dropout=dropout, context_dim=context_dim[d], checkpoint=use_checkpoint) for d in range(depth) ]) if not use_linear: self.proj_out = zero_module( nn.Conv1d( inner_dim, in_channels, kernel_size=1, stride=1, padding=0)) else: self.proj_out = zero_module(nn.Linear(in_channels, inner_dim)) if self.use_adaptor: self.adaptor_out = nn.Linear(frames, frames) self.use_linear = use_linear def forward(self, x, context=None): # note: if no context is given, cross-attention defaults to self-attention if self.only_self_att: context = None if not isinstance(context, list): context = [context] b, c, f, h, w = x.shape x_in = x x = self.norm(x) if not self.use_linear: x = rearrange(x, 'b c f h w -> (b h w) c f').contiguous() x = self.proj_in(x) if self.use_linear: x = rearrange( x, '(b f) c h w -> b (h w) f c', f=self.frames).contiguous() x = self.proj_in(x) if self.only_self_att: x = rearrange(x, 'bhw c f -> bhw f c').contiguous() for i, block in enumerate(self.transformer_blocks): x = block(x) x = rearrange(x, '(b hw) f c -> b hw f c', b=b).contiguous() else: x = rearrange(x, '(b hw) c f -> b hw f c', b=b).contiguous() for i, block in enumerate(self.transformer_blocks): context[i] = rearrange( context[i], '(b f) l con -> b f l con', f=self.frames).contiguous() # calculate each batch one by one # (since number in shape could not greater then 65,535 for some package) for j in range(b): context_i_j = repeat( context[i][j], 'f l con -> (f r) l con', r=(h * w) // self.frames, f=self.frames).contiguous() x[j] = block(x[j], context=context_i_j) if self.use_linear: x = self.proj_out(x) x = rearrange(x, 'b (h w) f c -> b f c h w', h=h, w=w).contiguous() if not self.use_linear: x = rearrange(x, 'b hw f c -> (b hw) c f').contiguous() x = self.proj_out(x) x = rearrange( x, '(b h w) c f -> b c f h w', b=b, h=h, w=w).contiguous() if self.multiply_zero: x = 0.0 * x + x_in else: x = x + x_in return x class TemporalAttentionMultiBlock(nn.Module): def __init__( self, dim, heads=4, dim_head=32, rotary_emb=None, use_image_dataset=False, use_sim_mask=False, temporal_attn_times=1, ): super().__init__() self.att_layers = nn.ModuleList([ TemporalAttentionBlock(dim, heads, dim_head, rotary_emb, use_image_dataset, use_sim_mask) for _ in range(temporal_attn_times) ]) def forward(self, x, pos_bias=None, focus_present_mask=None, video_mask=None): for layer in self.att_layers: x = layer(x, pos_bias, focus_present_mask, video_mask) return x class InitTemporalConvBlock(nn.Module): def __init__(self, in_dim, out_dim=None, dropout=0.0, use_image_dataset=False): super(InitTemporalConvBlock, self).__init__() if out_dim is None: out_dim = in_dim self.in_dim = in_dim self.out_dim = out_dim self.use_image_dataset = use_image_dataset # conv layers self.conv = nn.Sequential( nn.GroupNorm(32, out_dim), nn.SiLU(), nn.Dropout(dropout), nn.Conv3d(out_dim, in_dim, (3, 1, 1), padding=(1, 0, 0))) # zero out the last layer params,so the conv block is identity nn.init.zeros_(self.conv[-1].weight) nn.init.zeros_(self.conv[-1].bias) def forward(self, x): identity = x x = self.conv(x) if self.use_image_dataset: x = identity + 0 * x else: x = identity + x return x class TemporalConvBlock(nn.Module): def __init__(self, in_dim, out_dim=None, dropout=0.0, use_image_dataset=False): super(TemporalConvBlock, self).__init__() if out_dim is None: out_dim = in_dim self.in_dim = in_dim self.out_dim = out_dim self.use_image_dataset = use_image_dataset # conv layers self.conv1 = nn.Sequential( nn.GroupNorm(32, in_dim), nn.SiLU(), nn.Conv3d(in_dim, out_dim, (3, 1, 1), padding=(1, 0, 0))) self.conv2 = nn.Sequential( nn.GroupNorm(32, out_dim), nn.SiLU(), nn.Dropout(dropout), nn.Conv3d(out_dim, in_dim, (3, 1, 1), padding=(1, 0, 0))) # zero out the last layer params,so the conv block is identity nn.init.zeros_(self.conv2[-1].weight) nn.init.zeros_(self.conv2[-1].bias) def forward(self, x): identity = x x = self.conv1(x) x = self.conv2(x) if self.use_image_dataset: x = identity + 0 * x else: x = identity + x return x class TemporalConvBlock_v2(nn.Module): def __init__(self, in_dim, out_dim=None, dropout=0.0, use_image_dataset=False): super(TemporalConvBlock_v2, self).__init__() if out_dim is None: out_dim = in_dim self.in_dim = in_dim self.out_dim = out_dim self.use_image_dataset = use_image_dataset # conv layers self.conv1 = nn.Sequential( nn.GroupNorm(32, in_dim), nn.SiLU(), nn.Conv3d(in_dim, out_dim, (3, 1, 1), padding=(1, 0, 0))) self.conv2 = nn.Sequential( nn.GroupNorm(32, out_dim), nn.SiLU(), nn.Dropout(dropout), nn.Conv3d(out_dim, in_dim, (3, 1, 1), padding=(1, 0, 0))) self.conv3 = nn.Sequential( nn.GroupNorm(32, out_dim), nn.SiLU(), nn.Dropout(dropout), nn.Conv3d(out_dim, in_dim, (3, 1, 1), padding=(1, 0, 0))) self.conv4 = nn.Sequential( nn.GroupNorm(32, out_dim), nn.SiLU(), nn.Dropout(dropout), nn.Conv3d(out_dim, in_dim, (3, 1, 1), padding=(1, 0, 0))) # zero out the last layer params,so the conv block is identity nn.init.zeros_(self.conv4[-1].weight) nn.init.zeros_(self.conv4[-1].bias) def forward(self, x): identity = x x = self.conv1(x) x = self.conv2(x) x = self.conv3(x) x = self.conv4(x) if self.use_image_dataset: x = identity + 0.0 * x else: x = identity + x return x class UNetSD_temporal(nn.Module): def __init__( self, cfg, in_dim=7, dim=512, y_dim=512, context_dim=512, hist_dim=156, concat_dim=8, out_dim=6, dim_mult=[1, 2, 3, 4], num_heads=None, head_dim=64, num_res_blocks=3, attn_scales=[1 / 2, 1 / 4, 1 / 8], use_scale_shift_norm=True, dropout=0.1, temporal_attn_times=1, temporal_attention=True, use_checkpoint=False, use_image_dataset=False, use_fps_condition=False, use_sim_mask=False, misc_dropout=0.5, training=True, inpainting=True, video_compositions=['text', 'mask'], p_all_zero=0.1, p_all_keep=0.1, zero_y=None, black_image_feature=None, ): embed_dim = dim * 4 num_heads = num_heads if num_heads else dim // 32 super(UNetSD_temporal, self).__init__() self.zero_y = zero_y self.black_image_feature = black_image_feature self.cfg = cfg self.in_dim = in_dim self.dim = dim self.y_dim = y_dim self.context_dim = context_dim self.hist_dim = hist_dim self.concat_dim = concat_dim self.embed_dim = embed_dim self.out_dim = out_dim self.dim_mult = dim_mult # for temporal attention self.num_heads = num_heads # for spatial attention self.head_dim = head_dim self.num_res_blocks = num_res_blocks self.attn_scales = attn_scales self.use_scale_shift_norm = use_scale_shift_norm self.temporal_attn_times = temporal_attn_times self.temporal_attention = temporal_attention self.use_checkpoint = use_checkpoint self.use_image_dataset = use_image_dataset self.use_fps_condition = use_fps_condition self.use_sim_mask = use_sim_mask self.training = training self.inpainting = inpainting self.video_compositions = video_compositions self.misc_dropout = misc_dropout self.p_all_zero = p_all_zero self.p_all_keep = p_all_keep use_linear_in_temporal = False transformer_depth = 1 disabled_sa = False # params enc_dims = [dim * u for u in [1] + dim_mult] dec_dims = [dim * u for u in [dim_mult[-1]] + dim_mult[::-1]] shortcut_dims = [] scale = 1.0 if hasattr(cfg, 'adapter_transformer_layers' ) and cfg.adapter_transformer_layers: adapter_transformer_layers = cfg.adapter_transformer_layers else: adapter_transformer_layers = 1 # embeddings self.time_embed = nn.Sequential( nn.Linear(dim, embed_dim), nn.SiLU(), nn.Linear(embed_dim, embed_dim)) self.pre_image_condition = nn.Sequential( nn.Linear(1024, 1024), nn.SiLU(), nn.Linear(1024, 1024)) # depth embedding if 'depthmap' in self.video_compositions: self.depth_embedding = nn.Sequential( nn.Conv2d(1, concat_dim * 4, 3, padding=1), nn.SiLU(), nn.AdaptiveAvgPool2d((128, 128)), nn.Conv2d( concat_dim * 4, concat_dim * 4, 3, stride=2, padding=1), nn.SiLU(), nn.Conv2d(concat_dim * 4, concat_dim, 3, stride=2, padding=1)) self.depth_embedding_after = Transformer_v2( heads=2, dim=concat_dim, dim_head_k=concat_dim, dim_head_v=concat_dim, dropout_atte=0.05, mlp_dim=concat_dim, dropout_ffn=0.05, depth=adapter_transformer_layers) if 'motion' in self.video_compositions: self.motion_embedding = nn.Sequential( nn.Conv2d(2, concat_dim * 4, 3, padding=1), nn.SiLU(), nn.AdaptiveAvgPool2d((128, 128)), nn.Conv2d( concat_dim * 4, concat_dim * 4, 3, stride=2, padding=1), nn.SiLU(), nn.Conv2d(concat_dim * 4, concat_dim, 3, stride=2, padding=1)) self.motion_embedding_after = Transformer_v2( heads=2, dim=concat_dim, dim_head_k=concat_dim, dim_head_v=concat_dim, dropout_atte=0.05, mlp_dim=concat_dim, dropout_ffn=0.05, depth=adapter_transformer_layers) # canny embedding if 'canny' in self.video_compositions: self.canny_embedding = nn.Sequential( nn.Conv2d(1, concat_dim * 4, 3, padding=1), nn.SiLU(), nn.AdaptiveAvgPool2d((128, 128)), nn.Conv2d( concat_dim * 4, concat_dim * 4, 3, stride=2, padding=1), nn.SiLU(), nn.Conv2d(concat_dim * 4, concat_dim, 3, stride=2, padding=1)) self.canny_embedding_after = Transformer_v2( heads=2, dim=concat_dim, dim_head_k=concat_dim, dim_head_v=concat_dim, dropout_atte=0.05, mlp_dim=concat_dim, dropout_ffn=0.05, depth=adapter_transformer_layers) # masked-image embedding if 'mask' in self.video_compositions: self.masked_embedding = nn.Sequential( nn.Conv2d(4, concat_dim * 4, 3, padding=1), nn.SiLU(), nn.AdaptiveAvgPool2d((128, 128)), nn.Conv2d( concat_dim * 4, concat_dim * 4, 3, stride=2, padding=1), nn.SiLU(), nn.Conv2d(concat_dim * 4, concat_dim, 3, stride=2, padding=1)) if inpainting else None self.mask_embedding_after = Transformer_v2( heads=2, dim=concat_dim, dim_head_k=concat_dim, dim_head_v=concat_dim, dropout_atte=0.05, mlp_dim=concat_dim, dropout_ffn=0.05, depth=adapter_transformer_layers) # sketch embedding if 'sketch' in self.video_compositions: self.sketch_embedding = nn.Sequential( nn.Conv2d(1, concat_dim * 4, 3, padding=1), nn.SiLU(), nn.AdaptiveAvgPool2d((128, 128)), nn.Conv2d( concat_dim * 4, concat_dim * 4, 3, stride=2, padding=1), nn.SiLU(), nn.Conv2d(concat_dim * 4, concat_dim, 3, stride=2, padding=1)) self.sketch_embedding_after = Transformer_v2( heads=2, dim=concat_dim, dim_head_k=concat_dim, dim_head_v=concat_dim, dropout_atte=0.05, mlp_dim=concat_dim, dropout_ffn=0.05, depth=adapter_transformer_layers) if 'single_sketch' in self.video_compositions: self.single_sketch_embedding = nn.Sequential( nn.Conv2d(1, concat_dim * 4, 3, padding=1), nn.SiLU(), nn.AdaptiveAvgPool2d((128, 128)), nn.Conv2d( concat_dim * 4, concat_dim * 4, 3, stride=2, padding=1), nn.SiLU(), nn.Conv2d(concat_dim * 4, concat_dim, 3, stride=2, padding=1)) self.single_sketch_embedding_after = Transformer_v2( heads=2, dim=concat_dim, dim_head_k=concat_dim, dim_head_v=concat_dim, dropout_atte=0.05, mlp_dim=concat_dim, dropout_ffn=0.05, depth=adapter_transformer_layers) if 'local_image' in self.video_compositions: self.local_image_embedding = nn.Sequential( nn.Conv2d(3, concat_dim * 4, 3, padding=1), nn.SiLU(), nn.AdaptiveAvgPool2d((128, 128)), nn.Conv2d( concat_dim * 4, concat_dim * 4, 3, stride=2, padding=1), nn.SiLU(), nn.Conv2d(concat_dim * 4, concat_dim, 3, stride=2, padding=1)) self.local_image_embedding_after = Transformer_v2( heads=2, dim=concat_dim, dim_head_k=concat_dim, dim_head_v=concat_dim, dropout_atte=0.05, mlp_dim=concat_dim, dropout_ffn=0.05, depth=adapter_transformer_layers) # Condition Dropout self.misc_dropout = DropPath(misc_dropout) if temporal_attention and not USE_TEMPORAL_TRANSFORMER: self.rotary_emb = RotaryEmbedding(min(32, head_dim)) self.time_rel_pos_bias = RelativePositionBias( heads=num_heads, max_distance=32) if self.use_fps_condition: self.fps_embedding = nn.Sequential( nn.Linear(dim, embed_dim), nn.SiLU(), nn.Linear(embed_dim, embed_dim)) nn.init.zeros_(self.fps_embedding[-1].weight) nn.init.zeros_(self.fps_embedding[-1].bias) # encoder self.input_blocks = nn.ModuleList() if cfg.resume: self.pre_image = nn.Sequential() init_block = nn.ModuleList( [nn.Conv2d(self.in_dim + concat_dim, dim, 3, padding=1)]) else: self.pre_image = nn.Sequential( nn.Conv2d(self.in_dim + concat_dim, self.in_dim, 1, padding=0)) init_block = nn.ModuleList( [nn.Conv2d(self.in_dim, dim, 3, padding=1)]) # need an initial temporal attention? if temporal_attention: if USE_TEMPORAL_TRANSFORMER: init_block.append( TemporalTransformer( dim, num_heads, head_dim, depth=transformer_depth, context_dim=context_dim, disable_self_attn=disabled_sa, use_linear=use_linear_in_temporal, multiply_zero=use_image_dataset)) else: init_block.append( TemporalAttentionMultiBlock( dim, num_heads, head_dim, rotary_emb=self.rotary_emb, temporal_attn_times=temporal_attn_times, use_image_dataset=use_image_dataset)) self.input_blocks.append(init_block) shortcut_dims.append(dim) for i, (in_dim, out_dim) in enumerate(zip(enc_dims[:-1], enc_dims[1:])): for j in range(num_res_blocks): # residual (+attention) blocks block = nn.ModuleList([ ResBlock( in_dim, embed_dim, dropout, out_channels=out_dim, use_scale_shift_norm=False, use_image_dataset=use_image_dataset, ) ]) if scale in attn_scales: # block.append( SpatialTransformer( out_dim, out_dim // head_dim, head_dim, depth=1, context_dim=self.context_dim, disable_self_attn=False, use_linear=True)) if self.temporal_attention: if USE_TEMPORAL_TRANSFORMER: block.append( TemporalTransformer( out_dim, out_dim // head_dim, head_dim, depth=transformer_depth, context_dim=context_dim, disable_self_attn=disabled_sa, use_linear=use_linear_in_temporal, multiply_zero=use_image_dataset)) else: block.append( TemporalAttentionMultiBlock( out_dim, num_heads, head_dim, rotary_emb=self.rotary_emb, use_image_dataset=use_image_dataset, use_sim_mask=use_sim_mask, temporal_attn_times=temporal_attn_times)) in_dim = out_dim self.input_blocks.append(block) shortcut_dims.append(out_dim) # downsample if i != len(dim_mult) - 1 and j == num_res_blocks - 1: downsample = Downsample( out_dim, True, dims=2, out_channels=out_dim) shortcut_dims.append(out_dim) scale /= 2.0 self.input_blocks.append(downsample) # middle self.middle_block = nn.ModuleList([ ResBlock( out_dim, embed_dim, dropout, use_scale_shift_norm=False, use_image_dataset=use_image_dataset, ), SpatialTransformer( out_dim, out_dim // head_dim, head_dim, depth=1, context_dim=self.context_dim, disable_self_attn=False, use_linear=True) ]) if self.temporal_attention: if USE_TEMPORAL_TRANSFORMER: self.middle_block.append( TemporalTransformer( out_dim, out_dim // head_dim, head_dim, depth=transformer_depth, context_dim=context_dim, disable_self_attn=disabled_sa, use_linear=use_linear_in_temporal, multiply_zero=use_image_dataset, )) else: self.middle_block.append( TemporalAttentionMultiBlock( out_dim, num_heads, head_dim, rotary_emb=self.rotary_emb, use_image_dataset=use_image_dataset, use_sim_mask=use_sim_mask, temporal_attn_times=temporal_attn_times)) self.middle_block.append( ResBlock(out_dim, embed_dim, dropout, use_scale_shift_norm=False)) # decoder self.output_blocks = nn.ModuleList() for i, (in_dim, out_dim) in enumerate(zip(dec_dims[:-1], dec_dims[1:])): for j in range(num_res_blocks + 1): block = nn.ModuleList([ ResBlock( in_dim + shortcut_dims.pop(), embed_dim, dropout, out_dim, use_scale_shift_norm=False, use_image_dataset=use_image_dataset, ) ]) if scale in attn_scales: block.append( SpatialTransformer( out_dim, out_dim // head_dim, head_dim, depth=1, context_dim=1024, disable_self_attn=False, use_linear=True)) if self.temporal_attention: if USE_TEMPORAL_TRANSFORMER: block.append( TemporalTransformer( out_dim, out_dim // head_dim, head_dim, depth=transformer_depth, context_dim=context_dim, disable_self_attn=disabled_sa, use_linear=use_linear_in_temporal, multiply_zero=use_image_dataset)) else: block.append( TemporalAttentionMultiBlock( out_dim, num_heads, head_dim, rotary_emb=self.rotary_emb, use_image_dataset=use_image_dataset, use_sim_mask=use_sim_mask, temporal_attn_times=temporal_attn_times)) in_dim = out_dim # upsample if i != len(dim_mult) - 1 and j == num_res_blocks: upsample = Upsample( out_dim, True, dims=2.0, out_channels=out_dim) scale *= 2.0 block.append(upsample) self.output_blocks.append(block) # head self.out = nn.Sequential( nn.GroupNorm(32, out_dim), nn.SiLU(), nn.Conv2d(out_dim, self.out_dim, 3, padding=1)) # zero out the last layer params nn.init.zeros_(self.out[-1].weight) def forward( self, x, t, y=None, depth=None, image=None, motion=None, local_image=None, single_sketch=None, masked=None, canny=None, sketch=None, histogram=None, fps=None, video_mask=None, focus_present_mask=None, prob_focus_present=0., mask_last_frame_num=0 # mask last frame num ): assert self.inpainting or masked is None, 'inpainting is not supported' batch, c, f, h, w = x.shape device = x.device self.batch = batch # image and video joint training, if mask_last_frame_num is set, prob_focus_present will be ignored if mask_last_frame_num > 0: focus_present_mask = None video_mask[-mask_last_frame_num:] = False else: focus_present_mask = default( focus_present_mask, lambda: prob_mask_like( (batch, ), prob_focus_present, device=device)) if self.temporal_attention and not USE_TEMPORAL_TRANSFORMER: time_rel_pos_bias = self.time_rel_pos_bias( x.shape[2], device=x.device) else: time_rel_pos_bias = None # all-zero and all-keep masks zero = torch.zeros(batch, dtype=torch.bool).to(x.device) keep = torch.zeros(batch, dtype=torch.bool).to(x.device) if self.training: nzero = (torch.rand(batch) < self.p_all_zero).sum() nkeep = (torch.rand(batch) < self.p_all_keep).sum() index = torch.randperm(batch) zero[index[0:nzero]] = True keep[index[nzero:nzero + nkeep]] = True assert not (zero & keep).any() misc_dropout = partial(self.misc_dropout, zero=zero, keep=keep) concat = x.new_zeros(batch, self.concat_dim, f, h, w) if depth is not None: # DropPath mask depth = rearrange(depth, 'b c f h w -> (b f) c h w') depth = self.depth_embedding(depth) h = depth.shape[2] depth = self.depth_embedding_after( rearrange(depth, '(b f) c h w -> (b h w) f c', b=batch)) # depth = rearrange(depth, '(b h w) f c -> b c f h w', b=batch, h=h) concat = concat + misc_dropout(depth) # local_image_embedding if local_image is not None: local_image = rearrange(local_image, 'b c f h w -> (b f) c h w') local_image = self.local_image_embedding(local_image) h = local_image.shape[2] local_image = self.local_image_embedding_after( rearrange(local_image, '(b f) c h w -> (b h w) f c', b=batch)) local_image = rearrange( local_image, '(b h w) f c -> b c f h w', b=batch, h=h) concat = concat + misc_dropout(local_image) if motion is not None: motion = rearrange(motion, 'b c f h w -> (b f) c h w') motion = self.motion_embedding(motion) h = motion.shape[2] motion = self.motion_embedding_after( rearrange(motion, '(b f) c h w -> (b h w) f c', b=batch)) motion = rearrange( motion, '(b h w) f c -> b c f h w', b=batch, h=h) if hasattr(self.cfg, 'p_zero_motion_alone' ) and self.cfg.p_zero_motion_alone and self.training: motion_d = torch.rand(batch) < self.cfg.p_zero_motion motion_d = motion_d[:, None, None, None, None] motion = motion.masked_fill(motion_d.cuda(), 0) concat = concat + motion else: concat = concat + misc_dropout(motion) if canny is not None: # DropPath mask canny = rearrange(canny, 'b c f h w -> (b f) c h w') canny = self.canny_embedding(canny) h = canny.shape[2] canny = self.canny_embedding_after( rearrange(canny, '(b f) c h w -> (b h w) f c', b=batch)) canny = rearrange(canny, '(b h w) f c -> b c f h w', b=batch, h=h) concat = concat + misc_dropout(canny) if sketch is not None: # DropPath mask sketch = rearrange(sketch, 'b c f h w -> (b f) c h w') sketch = self.sketch_embedding(sketch) h = sketch.shape[2] sketch = self.sketch_embedding_after( rearrange(sketch, '(b f) c h w -> (b h w) f c', b=batch)) sketch = rearrange( sketch, '(b h w) f c -> b c f h w', b=batch, h=h) concat = concat + misc_dropout(sketch) if single_sketch is not None: # DropPath mask single_sketch = rearrange(single_sketch, 'b c f h w -> (b f) c h w') single_sketch = self.single_sketch_embedding(single_sketch) h = single_sketch.shape[2] single_sketch = self.single_sketch_embedding_after( rearrange( single_sketch, '(b f) c h w -> (b h w) f c', b=batch)) single_sketch = rearrange( single_sketch, '(b h w) f c -> b c f h w', b=batch, h=h) concat = concat + misc_dropout(single_sketch) if masked is not None: # DropPath mask masked = rearrange(masked, 'b c f h w -> (b f) c h w') masked = self.masked_embedding(masked) h = masked.shape[2] masked = self.mask_embedding_after( rearrange(masked, '(b f) c h w -> (b h w) f c', b=batch)) masked = rearrange( masked, '(b h w) f c -> b c f h w', b=batch, h=h) concat = concat + misc_dropout(masked) x = torch.cat([x, concat], dim=1) x = rearrange(x, 'b c f h w -> (b f) c h w') x = self.pre_image(x) x = rearrange(x, '(b f) c h w -> b c f h w', b=batch) # embeddings if self.use_fps_condition and fps is not None: e = self.time_embed(sinusoidal_embedding( t, self.dim)) + self.fps_embedding( sinusoidal_embedding(fps, self.dim)) else: e = self.time_embed(sinusoidal_embedding(t, self.dim)) context = x.new_zeros(batch, 0, self.context_dim) if y is not None: y_context = misc_dropout(y) context = torch.cat([context, y_context], dim=1) else: y_context = self.zero_y.repeat(batch, 1, 1) context = torch.cat([context, y_context], dim=1) if image is not None: image_context = misc_dropout(self.pre_image_condition(image)) context = torch.cat([context, image_context], dim=1) # repeat f times for spatial e and context e = e.repeat_interleave(repeats=f, dim=0) context = context.repeat_interleave(repeats=f, dim=0) # always in shape (b f) c h w, except for temporal layer x = rearrange(x, 'b c f h w -> (b f) c h w') # encoder xs = [] for block in self.input_blocks: x = self._forward_single(block, x, e, context, time_rel_pos_bias, focus_present_mask, video_mask) xs.append(x) # middle for block in self.middle_block: x = self._forward_single(block, x, e, context, time_rel_pos_bias, focus_present_mask, video_mask) # decoder for block in self.output_blocks: x = torch.cat([x, xs.pop()], dim=1) x = self._forward_single( block, x, e, context, time_rel_pos_bias, focus_present_mask, video_mask, reference=xs[-1] if len(xs) > 0 else None) # head x = self.out(x) # reshape back to (b c f h w) x = rearrange(x, '(b f) c h w -> b c f h w', b=batch) return x def _forward_single(self, module, x, e, context, time_rel_pos_bias, focus_present_mask, video_mask, reference=None): if isinstance(module, ResidualBlock): module = checkpoint_wrapper( module) if self.use_checkpoint else module x = x.contiguous() x = module(x, e, reference) elif isinstance(module, ResBlock): module = checkpoint_wrapper( module) if self.use_checkpoint else module x = x.contiguous() x = module(x, e, self.batch) elif isinstance(module, SpatialTransformer): module = checkpoint_wrapper( module) if self.use_checkpoint else module x = module(x, context) elif isinstance(module, TemporalTransformer): module = checkpoint_wrapper( module) if self.use_checkpoint else module x = rearrange(x, '(b f) c h w -> b c f h w', b=self.batch) x = module(x, context) x = rearrange(x, 'b c f h w -> (b f) c h w') elif isinstance(module, CrossAttention): module = checkpoint_wrapper( module) if self.use_checkpoint else module x = module(x, context) elif isinstance(module, BasicTransformerBlock): module = checkpoint_wrapper( module) if self.use_checkpoint else module x = module(x, context) elif isinstance(module, FeedForward): x = module(x, context) elif isinstance(module, Upsample): x = module(x) elif isinstance(module, Downsample): x = module(x) elif isinstance(module, Resample): x = module(x, reference) elif isinstance(module, TemporalAttentionBlock): module = checkpoint_wrapper( module) if self.use_checkpoint else module x = rearrange(x, '(b f) c h w -> b c f h w', b=self.batch) x = module(x, time_rel_pos_bias, focus_present_mask, video_mask) x = rearrange(x, 'b c f h w -> (b f) c h w') elif isinstance(module, TemporalAttentionMultiBlock): module = checkpoint_wrapper( module) if self.use_checkpoint else module x = rearrange(x, '(b f) c h w -> b c f h w', b=self.batch) x = module(x, time_rel_pos_bias, focus_present_mask, video_mask) x = rearrange(x, 'b c f h w -> (b f) c h w') elif isinstance(module, InitTemporalConvBlock): module = checkpoint_wrapper( module) if self.use_checkpoint else module x = rearrange(x, '(b f) c h w -> b c f h w', b=self.batch) x = module(x) x = rearrange(x, 'b c f h w -> (b f) c h w') elif isinstance(module, TemporalConvBlock): module = checkpoint_wrapper( module) if self.use_checkpoint else module x = rearrange(x, '(b f) c h w -> b c f h w', b=self.batch) x = module(x) x = rearrange(x, 'b c f h w -> (b f) c h w') elif isinstance(module, nn.ModuleList): for block in module: x = self._forward_single(block, x, e, context, time_rel_pos_bias, focus_present_mask, video_mask, reference) else: x = module(x) return x class PreNormattention(nn.Module): def __init__(self, dim, fn): super().__init__() self.norm = nn.LayerNorm(dim) self.fn = fn def forward(self, x, **kwargs): return self.fn(self.norm(x), **kwargs) + x class PreNormattention_qkv(nn.Module): def __init__(self, dim, fn): super().__init__() self.norm = nn.LayerNorm(dim) self.fn = fn def forward(self, q, k, v, **kwargs): return self.fn(self.norm(q), self.norm(k), self.norm(v), **kwargs) + q class Attention(nn.Module): def __init__(self, dim, heads=8, dim_head=64, dropout=0.): super().__init__() inner_dim = dim_head * heads project_out = not (heads == 1 and dim_head == dim) self.heads = heads self.scale = dim_head**-0.5 self.attend = nn.Softmax(dim=-1) self.to_qkv = nn.Linear(dim, inner_dim * 3, bias=False) self.to_out = nn.Sequential( nn.Linear(inner_dim, dim), nn.Dropout(dropout)) if project_out else nn.Identity() def forward(self, x): _, _, _, h = *x.shape, self.heads qkv = self.to_qkv(x).chunk(3, dim=-1) q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> b h n d', h=h), qkv) dots = einsum('b h i d, b h j d -> b h i j', q, k) * self.scale attn = self.attend(dots) out = einsum('b h i j, b h j d -> b h i d', attn, v) out = rearrange(out, 'b h n d -> b n (h d)') return self.to_out(out) class Attention_qkv(nn.Module): def __init__(self, dim, heads=8, dim_head=64, dropout=0.): super().__init__() inner_dim = dim_head * heads project_out = not (heads == 1 and dim_head == dim) self.heads = heads self.scale = dim_head**-0.5 self.attend = nn.Softmax(dim=-1) self.to_q = nn.Linear(dim, inner_dim, bias=False) self.to_k = nn.Linear(dim, inner_dim, bias=False) self.to_v = nn.Linear(dim, inner_dim, bias=False) self.to_out = nn.Sequential( nn.Linear(inner_dim, dim), nn.Dropout(dropout)) if project_out else nn.Identity() def forward(self, q, k, v): _, _, _, h = *q.shape, self.heads bk = k.shape[0] q = self.to_q(q) k = self.to_k(k) v = self.to_v(v) q = rearrange(q, 'b n (h d) -> b h n d', h=h) k = rearrange(k, 'b n (h d) -> b h n d', b=bk, h=h) v = rearrange(v, 'b n (h d) -> b h n d', b=bk, h=h) dots = einsum('b h i d, b h j d -> b h i j', q, k) * self.scale attn = self.attend(dots) out = einsum('b h i j, b h j d -> b h i d', attn, v) out = rearrange(out, 'b h n d -> b n (h d)') return self.to_out(out) class PostNormattention(nn.Module): def __init__(self, dim, fn): super().__init__() self.norm = nn.LayerNorm(dim) self.fn = fn def forward(self, x, **kwargs): return self.norm(self.fn(x, **kwargs) + x) class Transformer_v2(nn.Module): def __init__(self, heads=8, dim=2048, dim_head_k=256, dim_head_v=256, dropout_atte=0.05, mlp_dim=2048, dropout_ffn=0.05, depth=1): super().__init__() self.layers = nn.ModuleList([]) self.depth = depth for _ in range(depth): self.layers.append( nn.ModuleList([ PreNormattention( dim, Attention( dim, heads=heads, dim_head=dim_head_k, dropout=dropout_atte)), FeedForward(dim, mlp_dim, dropout=dropout_ffn), ])) def forward(self, x): for attn, ff in self.layers[:1]: x = attn(x) x = ff(x) + x if self.depth > 1: for attn, ff in self.layers[1:]: x = attn(x) x = ff(x) + x return x class DropPath(nn.Module): r"""DropPath but without rescaling and supports optional all-zero and/or all-keep. """ def __init__(self, p): super(DropPath, self).__init__() self.p = p def forward(self, *args, zero=None, keep=None): if not self.training: return args[0] if len(args) == 1 else args # params x = args[0] b = x.size(0) n = (torch.rand(b) < self.p).sum() # non-zero and non-keep mask mask = x.new_ones(b, dtype=torch.bool) if keep is not None: mask[keep] = False if zero is not None: mask[zero] = False # drop-path index index = torch.where(mask)[0] index = index[torch.randperm(len(index))[:n]] if zero is not None: index = torch.cat([index, torch.where(zero)[0]], dim=0) # drop-path multiplier multiplier = x.new_ones(b) multiplier[index] = 0.0 output = tuple(u * self.broadcast(multiplier, u) for u in args) return output[0] if len(args) == 1 else output def broadcast(self, src, dst): assert src.size(0) == dst.size(0) shape = (dst.size(0), ) + (1, ) * (dst.ndim - 1) return src.view(shape) if __name__ == '__main__': from config import cfg # [model] unet model = UNetSD_temporal( in_dim=cfg.unet_in_dim, dim=cfg.unet_dim, y_dim=cfg.unet_y_dim, context_dim=cfg.unet_context_dim, out_dim=cfg.unet_out_dim, dim_mult=cfg.unet_dim_mult, num_heads=cfg.unet_num_heads, head_dim=cfg.unet_head_dim, num_res_blocks=cfg.unet_res_blocks, attn_scales=cfg.unet_attn_scales, dropout=cfg.unet_dropout, temporal_attn_times=0, use_checkpoint=cfg.use_checkpoint, use_image_dataset=True, use_fps_condition=cfg.use_fps_condition) print( int(sum(p.numel() for k, p in model.named_parameters()) / (1024**2)), 'M parameters')