From 3f3763486bab8cb181739312823c32bf20b2f574 Mon Sep 17 00:00:00 2001 From: ExponentialML <59846140+ExponentialML@users.noreply.github.com> Date: Fri, 22 Mar 2024 15:24:40 -0700 Subject: [PATCH] Add updated transformer and attention --- modules/attention.py | 676 +++++++++++++++++++ modules/encoders/condition.py | 389 +++++++++++ modules/encoders/resampler.py | 145 ++++ modules/networks/ae_modules.py | 1026 +++++++++++++++++++++++++++++ modules/networks/openaimodel3d.py | 822 +++++++++++++++++++++++ modules/x_transformer.py | 639 ++++++++++++++++++ 6 files changed, 3697 insertions(+) create mode 100644 modules/attention.py create mode 100644 modules/encoders/condition.py create mode 100644 modules/encoders/resampler.py create mode 100644 modules/networks/ae_modules.py create mode 100644 modules/networks/openaimodel3d.py create mode 100644 modules/x_transformer.py diff --git a/modules/attention.py b/modules/attention.py new file mode 100644 index 0000000..7537aa5 --- /dev/null +++ b/modules/attention.py @@ -0,0 +1,676 @@ +import torch +from torch import nn, einsum +import torch.nn.functional as F +from einops import rearrange, repeat +from functools import partial +from ..common import ( + checkpoint, + exists, + default, +) +from ..basics import zero_module +import comfy.ops +ops = comfy.ops.disable_weight_init +from comfy import model_management +from comfy.ldm.modules.attention import optimized_attention, optimized_attention_masked + +if model_management.xformers_enabled(): + import xformers + import xformers.ops + XFORMERS_IS_AVAILBLE = True +else: + XFORMERS_IS_AVAILBLE = False + +class RelativePosition(nn.Module): + """ https://github.com/evelinehong/Transformer_Relative_Position_PyTorch/blob/master/relative_position.py """ + + def __init__(self, num_units, max_relative_position): + super().__init__() + self.num_units = num_units + self.max_relative_position = max_relative_position + self.embeddings_table = nn.Parameter(torch.Tensor(max_relative_position * 2 + 1, num_units)) + nn.init.xavier_uniform_(self.embeddings_table) + + def forward(self, length_q, length_k): + device = self.embeddings_table.device + range_vec_q = torch.arange(length_q, device=device) + range_vec_k = torch.arange(length_k, device=device) + distance_mat = range_vec_k[None, :] - range_vec_q[:, None] + distance_mat_clipped = torch.clamp(distance_mat, -self.max_relative_position, self.max_relative_position) + final_mat = distance_mat_clipped + self.max_relative_position + final_mat = final_mat.long() + embeddings = self.embeddings_table[final_mat] + return embeddings + + +# TODO Add native Comfy optimized attention. +class CrossAttention(nn.Module): + + def __init__( + self, + query_dim, + context_dim=None, + heads=8, + dim_head=64, + dropout=0., + relative_position=False, + temporal_length=None, + video_length=None, + image_cross_attention=False, + image_cross_attention_scale=1.0, + image_cross_attention_scale_learnable=False, + text_context_len=77, + device=None, + dtype=None, + operations=ops + ): + super().__init__() + inner_dim = dim_head * heads + context_dim = default(context_dim, query_dim) + self.scale = dim_head**-0.5 + self.heads = heads + self.dim_head = dim_head + self.to_q = operations.Linear(query_dim, inner_dim, bias=False, device=device, dtype=dtype) + self.to_k = operations.Linear(context_dim, inner_dim, bias=False, device=device, dtype=dtype) + self.to_v = operations.Linear(context_dim, inner_dim, bias=False, device=device, dtype=dtype) + + self.to_out = nn.Sequential( + operations.Linear(inner_dim, query_dim, device=device, dtype=dtype), + nn.Dropout(dropout) + ) + + self.relative_position = relative_position + if self.relative_position: + assert(temporal_length is not None) + self.relative_position_k = RelativePosition(num_units=dim_head, max_relative_position=temporal_length) + self.relative_position_v = RelativePosition(num_units=dim_head, max_relative_position=temporal_length) + else: + ## only used for spatial attention, while NOT for temporal attention + if XFORMERS_IS_AVAILBLE and temporal_length is None: + self.forward = self.efficient_forward + else: + self.forward = self.comfy_efficient_forward + + self.video_length = video_length + self.image_cross_attention = image_cross_attention + self.image_cross_attention_scale = image_cross_attention_scale + self.text_context_len = text_context_len + self.image_cross_attention_scale_learnable = image_cross_attention_scale_learnable + if self.image_cross_attention: + self.to_k_ip = operations.Linear(context_dim, inner_dim, bias=False, device=device, dtype=dtype) + self.to_v_ip = operations.Linear(context_dim, inner_dim, bias=False, device=device, dtype=dtype) + if image_cross_attention_scale_learnable: + self.register_parameter('alpha', nn.Parameter(torch.tensor(0.)) ) + + def comfy_efficient_forward(self, x, context=None, mask=None): + spatial_self_attn = (context is None) + k_ip, v_ip, out_ip = None, None, None + + h = self.heads + q = self.to_q(x) + context = default(context, x) + + if self.image_cross_attention and not spatial_self_attn: + context, context_image = context[:,:self.text_context_len,:], context[:,self.text_context_len:,:] + k = self.to_k(context) + v = self.to_v(context) + k_ip = self.to_k_ip(context_image) + v_ip = self.to_v_ip(context_image) + else: + if not spatial_self_attn: + context = context[:,:self.text_context_len,:] + k = self.to_k(context) + v = self.to_v(context) + + out = optimized_attention(q, k, v, h) + + if exists(mask): + ## feasible for causal attention mask only + out = optimized_attention_masked(q, k, v, h) + + ## for image cross-attention + if k_ip is not None: + q = rearrange(q, 'b n (h d) -> (b h) n d', h=h) + k_ip, v_ip = map(lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h=h), (k_ip, v_ip)) + sim_ip = torch.einsum('b i d, b j d -> b i j', q, k_ip) * self.scale + del k_ip + sim_ip = sim_ip.softmax(dim=-1) + out_ip = torch.einsum('b i j, b j d -> b i d', sim_ip, v_ip) + out_ip = rearrange(out_ip, '(b h) n d -> b n (h d)', h=h) + + if out_ip is not None: + if self.image_cross_attention_scale_learnable: + out = out + self.image_cross_attention_scale * out_ip * (torch.tanh(self.alpha)+1) + else: + out = out + self.image_cross_attention_scale * out_ip + + return self.to_out(out) + + def forward(self, x, context=None, mask=None): + spatial_self_attn = (context is None) + k_ip, v_ip, out_ip = None, None, None + + h = self.heads + q = self.to_q(x) + context = default(context, x) + + if self.image_cross_attention and not spatial_self_attn: + context, context_image = context[:,:self.text_context_len,:], context[:,self.text_context_len:,:] + k = self.to_k(context) + v = self.to_v(context) + k_ip = self.to_k_ip(context_image) + v_ip = self.to_v_ip(context_image) + else: + + # Assumed Spatial Attention (b c h w) + if not spatial_self_attn: + context = context[:,:self.text_context_len,:] + 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)) + + sim = torch.einsum('b i d, b j d -> b i j', q, k) * self.scale + if self.relative_position: + len_q, len_k, len_v = q.shape[1], k.shape[1], v.shape[1] + k2 = self.relative_position_k(len_q, len_k) + sim2 = einsum('b t d, t s d -> b t s', q, k2) * self.scale # TODO check + sim += sim2 + del k + + if exists(mask): + ## feasible for causal attention mask only + max_neg_value = -torch.finfo(sim.dtype).max + mask = repeat(mask, 'b i j -> (b h) i j', h=h) + sim.masked_fill_(~(mask>0.5), 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) + if self.relative_position: + v2 = self.relative_position_v(len_q, len_v) + out2 = einsum('b t s, t s d -> b t d', sim, v2) # TODO check + out += out2 + out = rearrange(out, '(b h) n d -> b n (h d)', h=h) + + + ## for image cross-attention + if k_ip is not None: + k_ip, v_ip = map(lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h=h), (k_ip, v_ip)) + sim_ip = torch.einsum('b i d, b j d -> b i j', q, k_ip) * self.scale + del k_ip + sim_ip = sim_ip.softmax(dim=-1) + out_ip = torch.einsum('b i j, b j d -> b i d', sim_ip, v_ip) + out_ip = rearrange(out_ip, '(b h) n d -> b n (h d)', h=h) + + + if out_ip is not None: + if self.image_cross_attention_scale_learnable: + out = out + self.image_cross_attention_scale * out_ip * (torch.tanh(self.alpha)+1) + else: + out = out + self.image_cross_attention_scale * out_ip + + return self.to_out(out) + + def efficient_forward(self, x, context=None, mask=None): + spatial_self_attn = (context is None) + k_ip, v_ip, out_ip = None, None, None + + q = self.to_q(x) + context = default(context, x) + + if self.image_cross_attention and not spatial_self_attn: + context, context_image = context[:,:self.text_context_len,:], context[:,self.text_context_len:,:] + k = self.to_k(context) + v = self.to_v(context) + k_ip = self.to_k_ip(context_image) + v_ip = self.to_v_ip(context_image) + else: + if not spatial_self_attn: + context = context[:,:self.text_context_len,:] + k = self.to_k(context) + v = self.to_v(context) + + b, _, _ = q.shape + q, k, v = map( + lambda t: t.unsqueeze(3) + .reshape(b, t.shape[1], self.heads, self.dim_head) + .permute(0, 2, 1, 3) + .reshape(b * self.heads, t.shape[1], self.dim_head) + .contiguous(), + (q, k, v), + ) + # actually compute the attention, what we cannot get enough of + out = xformers.ops.memory_efficient_attention(q, k, v, attn_bias=None, op=None) + + ## for image cross-attention + if k_ip is not None: + k_ip, v_ip = map( + lambda t: t.unsqueeze(3) + .reshape(b, t.shape[1], self.heads, self.dim_head) + .permute(0, 2, 1, 3) + .reshape(b * self.heads, t.shape[1], self.dim_head) + .contiguous(), + (k_ip, v_ip), + ) + out_ip = xformers.ops.memory_efficient_attention(q, k_ip, v_ip, attn_bias=None, op=None) + out_ip = ( + out_ip.unsqueeze(0) + .reshape(b, self.heads, out.shape[1], self.dim_head) + .permute(0, 2, 1, 3) + .reshape(b, out.shape[1], self.heads * self.dim_head) + ) + + if exists(mask): + raise NotImplementedError + out = ( + out.unsqueeze(0) + .reshape(b, self.heads, out.shape[1], self.dim_head) + .permute(0, 2, 1, 3) + .reshape(b, out.shape[1], self.heads * self.dim_head) + ) + if out_ip is not None: + if self.image_cross_attention_scale_learnable: + out = out + self.image_cross_attention_scale * out_ip * (torch.tanh(self.alpha)+1) + else: + out = out + self.image_cross_attention_scale * out_ip + + 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, + attention_cls=None, + video_length=None, + image_cross_attention=False, + image_cross_attention_scale=1.0, + image_cross_attention_scale_learnable=False, + text_context_len=77, + device=None, + dtype=None, + operations=ops + ): + super().__init__() + attn_cls = CrossAttention if attention_cls is None else attention_cls + 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=None, device=device, dtype=dtype if self.disable_self_attn else None) + self.ff = FeedForward(dim, dropout=dropout, glu=gated_ff, device=device, dtype=dtype) + self.attn2 = attn_cls( + query_dim=dim, + context_dim=context_dim, + heads=n_heads, + dim_head=d_head, + dropout=dropout, + video_length=video_length, + image_cross_attention=image_cross_attention, + image_cross_attention_scale=image_cross_attention_scale, + image_cross_attention_scale_learnable=image_cross_attention_scale_learnable, + text_context_len=text_context_len, + device=device, + dtype=dtype + ) + self.image_cross_attention = image_cross_attention + + self.norm1 = operations.LayerNorm(dim, device=device, dtype=dtype) + self.norm2 = operations.LayerNorm(dim, device=device, dtype=dtype) + self.norm3 = operations.LayerNorm(dim, device=device, dtype=dtype) + self.checkpoint = checkpoint + + def forward(self, x, context=None, mask=None, **kwargs): + ## implementation tricks: because checkpointing doesn't support non-tensor (e.g. None or scalar) arguments + input_tuple = (x,) ## should not be (x), otherwise *input_tuple will decouple x into multiple arguments + if context is not None: + input_tuple = (x, context) + if mask is not None: + forward_mask = partial(self._forward, mask=mask) + return checkpoint(forward_mask, (x,), self.parameters(), self.checkpoint) + return checkpoint(self._forward, input_tuple, self.parameters(), self.checkpoint) + + + def _forward(self, x, context=None, mask=None): + x = self.attn1(self.norm1(x), context=context if self.disable_self_attn else None, mask=mask) + x + x = self.attn2(self.norm2(x), context=context, mask=mask) + x + x = self.ff(self.norm3(x)) + x + return x + + +class SpatialTransformer(nn.Module): + """ + Transformer block for image-like data in spatial axis. + 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, + use_checkpoint=True, + disable_self_attn=False, + use_linear=False, + video_length=None, + image_cross_attention=False, + image_cross_attention_scale_learnable=False, + device=None, + dtype=None, + operations=ops + ): + super().__init__() + self.in_channels = in_channels + inner_dim = n_heads * d_head + self.norm = operations.GroupNorm(num_groups=32, num_channels=in_channels, eps=1e-6, affine=True, device=device, dtype=dtype) + if not use_linear: + self.proj_in = opeations.Conv2d(in_channels, inner_dim, kernel_size=1, stride=1, padding=0, device=device, dtype=dtype) + else: + self.proj_in = operations.Linear(in_channels, inner_dim, device=device, dtype=dtype) + + attention_cls = None + self.transformer_blocks = nn.ModuleList([ + BasicTransformerBlock( + inner_dim, + n_heads, + d_head, + dropout=dropout, + context_dim=context_dim, + disable_self_attn=disable_self_attn, + checkpoint=use_checkpoint, + attention_cls=attention_cls, + video_length=video_length, + image_cross_attention=image_cross_attention, + image_cross_attention_scale_learnable=image_cross_attention_scale_learnable, + device=device, + dtype=dtype + ) for d in range(depth) + ]) + if not use_linear: + self.proj_out = zero_module(operations.Conv2d(inner_dim, in_channels, kernel_size=1, stride=1, padding=0, device=device, dtype=dtype)) + else: + self.proj_out = zero_module(operations.Linear(inner_dim, in_channels, device=device, dtype=dtype)) + self.use_linear = use_linear + + def forward(self, x, context=None, **kwargs): + 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, **kwargs) + 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 + + +class TemporalTransformer(nn.Module): + """ + Transformer block for image-like data in temporal axis. + First, 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, + use_checkpoint=True, + use_linear=False, + only_self_att=True, + causal_attention=False, + causal_block_size=1, + relative_position=False, + temporal_length=None, + device=None, + dtype=None, + operations=ops + ): + super().__init__() + self.only_self_att = only_self_att + self.relative_position = relative_position + self.causal_attention = causal_attention + self.causal_block_size = causal_block_size + + if only_self_att: + context_dim = None + + self.in_channels = in_channels + inner_dim = n_heads * d_head + self.norm = operations.GroupNorm(num_groups=32, num_channels=in_channels, eps=1e-6, affine=True, device=device, dtype=dtype) + self.proj_in = nn.Conv1d(in_channels, inner_dim, kernel_size=1, stride=1, padding=0).to(device, dtype) + if not use_linear: + self.proj_in = nn.Conv1d(in_channels, inner_dim, kernel_size=1, stride=1, padding=0).to(device, dtype) + else: + self.proj_in = operations.Linear(in_channels, inner_dim, device=device, dtype=dtype) + + if relative_position: + assert(temporal_length is not None) + attention_cls = partial(CrossAttention, relative_position=True, temporal_length=temporal_length, device=device, dtype=dtype) + else: + attention_cls = partial(CrossAttention, temporal_length=temporal_length, device=device, dtype=dtype) + if self.causal_attention: + assert(temporal_length is not None) + self.mask = torch.tril(torch.ones([1, temporal_length, temporal_length])) + + if self.only_self_att: + context_dim = None + self.transformer_blocks = nn.ModuleList([ + BasicTransformerBlock( + inner_dim, + n_heads, + d_head, + dropout=dropout, + context_dim=context_dim, + attention_cls=attention_cls, + checkpoint=use_checkpoint, + device=device, + dtype=dtype + ) 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).to(device, dtype)) + else: + self.proj_out = zero_module(operations.Linear(inner_dim, in_channels, device=device, dtype=dtype)) + self.use_linear = use_linear + + def forward(self, x, context=None): + b, c, t, h, w = x.shape + x_in = x + x = self.norm(x) + x = rearrange(x, 'b c t h w -> (b h w) c t').contiguous() + if not self.use_linear: + x = self.proj_in(x) + x = rearrange(x, 'bhw c t -> bhw t c').contiguous() + if self.use_linear: + x = self.proj_in(x) + + temp_mask = None + if self.causal_attention: + # slice the from mask map + temp_mask = self.mask[:,:t,:t].to(x.device) + + if temp_mask is not None: + mask = temp_mask.to(x.device) + mask = repeat(mask, 'l i j -> (l bhw) i j', bhw=b*h*w) + else: + mask = None + + if self.only_self_att: + ## note: if no context is given, cross-attention defaults to self-attention + for i, block in enumerate(self.transformer_blocks): + x = block(x, mask=mask) + x = rearrange(x, '(b hw) t c -> b hw t c', b=b).contiguous() + else: + x = rearrange(x, '(b hw) t c -> b hw t c', b=b).contiguous() + context = rearrange(context, '(b t) l con -> b t l con', t=t).contiguous() + for i, block in enumerate(self.transformer_blocks): + # 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_j = repeat( + context[j], + 't l con -> (t r) l con', r=(h * w) // t, t=t).contiguous() + ## note: causal mask will not applied in cross-attention case + x[j] = block(x[j], context=context_j) + + if self.use_linear: + x = self.proj_out(x) + x = rearrange(x, 'b (h w) t c -> b c t h w', h=h, w=w).contiguous() + if not self.use_linear: + x = rearrange(x, 'b hw t c -> (b hw) c t').contiguous() + x = self.proj_out(x) + x = rearrange(x, '(b h w) c t -> b c t h w', b=b, h=h, w=w).contiguous() + + return x + x_in + + +class GEGLU(nn.Module): + def __init__(self, dim_in, dim_out, device=None, dtype=None, operations=ops): + super().__init__() + self.proj = operations.Linear(dim_in, dim_out * 2, device=device, dtype=dtype) + + def forward(self, x): + x, gate = self.proj(x).chunk(2, dim=-1) + return x * F.gelu(gate) + + +class FeedForward(nn.Module): + def __init__(self, dim, dim_out=None, mult=4, glu=False, dropout=0., device=None, dtype=None, operations=ops): + super().__init__() + inner_dim = int(dim * mult) + dim_out = default(dim_out, dim) + project_in = nn.Sequential( + operations.Linear(dim, inner_dim, device=device, dtype=dtype), + nn.GELU() + ) if not glu else GEGLU(dim, inner_dim) + + self.net = nn.Sequential( + project_in, + nn.Dropout(dropout), + operations.Linear(inner_dim, dim_out, device=device, dtype=dtype) + ) + + def forward(self, x): + return self.net(x) + + +class LinearAttention(nn.Module): + def __init__(self, dim, heads=4, dim_head=32, device=None, dtype=None, operations=ops): + super().__init__() + self.heads = heads + hidden_dim = dim_head * heads + self.to_qkv = operations.Conv2d(dim, hidden_dim * 3, 1, bias = False, device=device, dtype=dtype) + self.to_out = operations.Conv2d(hidden_dim, dim, 1, device=device, dtype=dtype) + + def forward(self, x): + b, c, h, w = x.shape + qkv = self.to_qkv(x) + q, k, v = rearrange(qkv, 'b (qkv heads c) h w -> qkv b heads c (h w)', heads = self.heads, qkv=3) + k = k.softmax(dim=-1) + context = torch.einsum('bhdn,bhen->bhde', k, v) + out = torch.einsum('bhde,bhdn->bhen', context, q) + out = rearrange(out, 'b heads c (h w) -> b (heads c) h w', heads=self.heads, h=h, w=w) + return self.to_out(out) + + +class SpatialSelfAttention(nn.Module): + def __init__(self, in_channels, device=None, dtype=None, operations=ops): + super().__init__() + self.in_channels = in_channels + + self.norm = operations.GroupNorm( + num_groups=32, + num_channels=in_channels, + eps=1e-6, + affine=True, + device=device, + dtype=dtype + ) + self.q = operations.Conv2d( + in_channels, + in_channels, + kernel_size=1, + stride=1, + padding=0, + device=device, + dtype=dtype + ) + self.k = operations.Conv2d( + in_channels, + in_channels, + kernel_size=1, + stride=1, + padding=0, + device=device, + dtype=dtype + ) + self.v = operations.Conv2d( + in_channels, + in_channels, + kernel_size=1, + stride=1, + padding=0, + device=device, + dtype=dtype + ) + self.proj_out = operations.Conv2d( + in_channels, + in_channels, + kernel_size=1, + stride=1, + padding=0, + device=device, + dtype=dtype + ) + + def forward(self, x): + h_ = x + h_ = self.norm(h_) + q = self.q(h_) + k = self.k(h_) + v = self.v(h_) + + # compute attention + b,c,h,w = q.shape + q = rearrange(q, 'b c h w -> b (h w) c') + k = rearrange(k, 'b c h w -> b c (h w)') + w_ = torch.einsum('bij,bjk->bik', q, k) + + w_ = w_ * (int(c)**(-0.5)) + w_ = torch.nn.functional.softmax(w_, dim=2) + + # attend to values + v = rearrange(v, 'b c h w -> b c (h w)') + w_ = rearrange(w_, 'b i j -> b j i') + h_ = torch.einsum('bij,bjk->bik', v, w_) + h_ = rearrange(h_, 'b c (h w) -> b c h w', h=h) + h_ = self.proj_out(h_) + + return x+h_ diff --git a/modules/encoders/condition.py b/modules/encoders/condition.py new file mode 100644 index 0000000..610322b --- /dev/null +++ b/modules/encoders/condition.py @@ -0,0 +1,389 @@ +import torch +import torch.nn as nn +import kornia +import open_clip +from torch.utils.checkpoint import checkpoint +from transformers import T5Tokenizer, T5EncoderModel, CLIPTokenizer, CLIPTextModel +from ..common import autocast +from utils.utils import count_params + + +class AbstractEncoder(nn.Module): + def __init__(self): + super().__init__() + + def encode(self, *args, **kwargs): + raise NotImplementedError + + +class IdentityEncoder(AbstractEncoder): + def encode(self, x): + return x + + +class ClassEmbedder(nn.Module): + def __init__(self, embed_dim, n_classes=1000, key='class', ucg_rate=0.1): + super().__init__() + self.key = key + self.embedding = nn.Embedding(n_classes, embed_dim) + self.n_classes = n_classes + self.ucg_rate = ucg_rate + + def forward(self, batch, key=None, disable_dropout=False): + if key is None: + key = self.key + # this is for use in crossattn + c = batch[key][:, None] + if self.ucg_rate > 0. and not disable_dropout: + mask = 1. - torch.bernoulli(torch.ones_like(c) * self.ucg_rate) + c = mask * c + (1 - mask) * torch.ones_like(c) * (self.n_classes - 1) + c = c.long() + c = self.embedding(c) + return c + + def get_unconditional_conditioning(self, bs, device="cuda"): + uc_class = self.n_classes - 1 # 1000 classes --> 0 ... 999, one extra class for ucg (class 1000) + uc = torch.ones((bs,), device=device) * uc_class + uc = {self.key: uc} + return uc + + +def disabled_train(self, mode=True): + """Overwrite model.train with this function to make sure train/eval mode + does not change anymore.""" + return self + + +class FrozenT5Embedder(AbstractEncoder): + """Uses the T5 transformer encoder for text""" + + def __init__(self, version="google/t5-v1_1-large", device="cuda", max_length=77, + freeze=True): # others are google/t5-v1_1-xl and google/t5-v1_1-xxl + super().__init__() + self.tokenizer = T5Tokenizer.from_pretrained(version) + self.transformer = T5EncoderModel.from_pretrained(version) + self.device = device + self.max_length = max_length # TODO: typical value? + if freeze: + self.freeze() + + def freeze(self): + self.transformer = self.transformer.eval() + # self.train = disabled_train + for param in self.parameters(): + param.requires_grad = False + + def forward(self, text): + batch_encoding = self.tokenizer(text, truncation=True, max_length=self.max_length, return_length=True, + return_overflowing_tokens=False, padding="max_length", return_tensors="pt") + tokens = batch_encoding["input_ids"].to(self.device) + outputs = self.transformer(input_ids=tokens) + + z = outputs.last_hidden_state + return z + + def encode(self, text): + return self(text) + + +class FrozenCLIPEmbedder(AbstractEncoder): + """Uses the CLIP transformer encoder for text (from huggingface)""" + LAYERS = [ + "last", + "pooled", + "hidden" + ] + + def __init__(self, version="openai/clip-vit-large-patch14", device="cuda", max_length=77, + freeze=True, layer="last", layer_idx=None): # clip-vit-base-patch32 + super().__init__() + assert layer in self.LAYERS + self.tokenizer = CLIPTokenizer.from_pretrained(version) + self.transformer = CLIPTextModel.from_pretrained(version) + self.device = device + self.max_length = max_length + if freeze: + self.freeze() + self.layer = layer + self.layer_idx = layer_idx + if layer == "hidden": + assert layer_idx is not None + assert 0 <= abs(layer_idx) <= 12 + + def freeze(self): + self.transformer = self.transformer.eval() + # self.train = disabled_train + for param in self.parameters(): + param.requires_grad = False + + def forward(self, text): + batch_encoding = self.tokenizer(text, truncation=True, max_length=self.max_length, return_length=True, + return_overflowing_tokens=False, padding="max_length", return_tensors="pt") + tokens = batch_encoding["input_ids"].to(self.device) + outputs = self.transformer(input_ids=tokens, output_hidden_states=self.layer == "hidden") + if self.layer == "last": + z = outputs.last_hidden_state + elif self.layer == "pooled": + z = outputs.pooler_output[:, None, :] + else: + z = outputs.hidden_states[self.layer_idx] + return z + + def encode(self, text): + return self(text) + + +class ClipImageEmbedder(nn.Module): + def __init__( + self, + model, + jit=False, + device='cuda' if torch.cuda.is_available() else 'cpu', + antialias=True, + ucg_rate=0. + ): + super().__init__() + from clip import load as load_clip + self.model, _ = load_clip(name=model, device=device, jit=jit) + + self.antialias = antialias + + self.register_buffer('mean', torch.Tensor([0.48145466, 0.4578275, 0.40821073]), persistent=False) + self.register_buffer('std', torch.Tensor([0.26862954, 0.26130258, 0.27577711]), persistent=False) + self.ucg_rate = ucg_rate + + def preprocess(self, x): + # normalize to [0,1] + x = kornia.geometry.resize(x, (224, 224), + interpolation='bicubic', align_corners=True, + antialias=self.antialias) + x = (x + 1.) / 2. + # re-normalize according to clip + x = kornia.enhance.normalize(x, self.mean, self.std) + return x + + def forward(self, x, no_dropout=False): + # x is assumed to be in range [-1,1] + out = self.model.encode_image(self.preprocess(x)) + out = out.to(x.dtype) + if self.ucg_rate > 0. and not no_dropout: + out = torch.bernoulli((1. - self.ucg_rate) * torch.ones(out.shape[0], device=out.device))[:, None] * out + return out + + +class FrozenOpenCLIPEmbedder(AbstractEncoder): + """ + Uses the OpenCLIP transformer encoder for text + """ + LAYERS = [ + # "pooled", + "last", + "penultimate" + ] + + def __init__(self, arch="ViT-H-14", version="laion2b_s32b_b79k", device="cuda", max_length=77, + freeze=True, layer="last"): + super().__init__() + assert layer in self.LAYERS + model, _, _ = open_clip.create_model_and_transforms(arch, device=torch.device('cpu'), pretrained=version) + del model.visual + self.model = model + + self.device = device + self.max_length = max_length + if freeze: + self.freeze() + self.layer = layer + if self.layer == "last": + self.layer_idx = 0 + elif self.layer == "penultimate": + self.layer_idx = 1 + else: + raise NotImplementedError() + + def freeze(self): + self.model = self.model.eval() + for param in self.parameters(): + param.requires_grad = False + + def forward(self, text): + tokens = open_clip.tokenize(text) ## all clip models use 77 as context length + z = self.encode_with_transformer(tokens.to(self.device)) + return z + + def encode_with_transformer(self, text): + x = self.model.token_embedding(text) # [batch_size, n_ctx, d_model] + x = x + self.model.positional_embedding + x = x.permute(1, 0, 2) # NLD -> LND + x = self.text_transformer_forward(x, attn_mask=self.model.attn_mask) + x = x.permute(1, 0, 2) # LND -> NLD + x = self.model.ln_final(x) + return x + + def text_transformer_forward(self, x: torch.Tensor, attn_mask=None): + for i, r in enumerate(self.model.transformer.resblocks): + if i == len(self.model.transformer.resblocks) - self.layer_idx: + break + if self.model.transformer.grad_checkpointing and not torch.jit.is_scripting(): + x = checkpoint(r, x, attn_mask) + else: + x = r(x, attn_mask=attn_mask) + return x + + def encode(self, text): + return self(text) + + +class FrozenOpenCLIPImageEmbedder(AbstractEncoder): + """ + Uses the OpenCLIP vision transformer encoder for images + """ + + def __init__(self, arch="ViT-H-14", version="laion2b_s32b_b79k", device="cuda", max_length=77, + freeze=True, layer="pooled", antialias=True, ucg_rate=0.): + super().__init__() + model, _, _ = open_clip.create_model_and_transforms(arch, device=torch.device('cpu'), + pretrained=version, ) + del model.transformer + self.model = model + # self.mapper = torch.nn.Linear(1280, 1024) + self.device = device + self.max_length = max_length + if freeze: + self.freeze() + self.layer = layer + if self.layer == "penultimate": + raise NotImplementedError() + self.layer_idx = 1 + + self.antialias = antialias + + self.register_buffer('mean', torch.Tensor([0.48145466, 0.4578275, 0.40821073]), persistent=False) + self.register_buffer('std', torch.Tensor([0.26862954, 0.26130258, 0.27577711]), persistent=False) + self.ucg_rate = ucg_rate + + def preprocess(self, x): + # normalize to [0,1] + x = kornia.geometry.resize(x, (224, 224), + interpolation='bicubic', align_corners=True, + antialias=self.antialias) + x = (x + 1.) / 2. + # renormalize according to clip + x = kornia.enhance.normalize(x, self.mean, self.std) + return x + + def freeze(self): + self.model = self.model.eval() + for param in self.model.parameters(): + param.requires_grad = False + + @autocast + def forward(self, image, no_dropout=False): + z = self.encode_with_vision_transformer(image) + if self.ucg_rate > 0. and not no_dropout: + z = torch.bernoulli((1. - self.ucg_rate) * torch.ones(z.shape[0], device=z.device))[:, None] * z + return z + + def encode_with_vision_transformer(self, img): + img = self.preprocess(img) + x = self.model.visual(img) + return x + + def encode(self, text): + return self(text) + +class FrozenOpenCLIPImageEmbedderV2(AbstractEncoder): + """ + Uses the OpenCLIP vision transformer encoder for images + """ + + def __init__(self, arch="ViT-H-14", version="laion2b_s32b_b79k", device="cuda", + freeze=True, layer="pooled", antialias=True): + super().__init__() + model, _, _ = open_clip.create_model_and_transforms(arch, device=torch.device('cpu'), + pretrained=version, ) + del model.transformer + self.model = model + self.device = device + + if freeze: + self.freeze() + self.layer = layer + if self.layer == "penultimate": + raise NotImplementedError() + self.layer_idx = 1 + + self.antialias = antialias + + self.register_buffer('mean', torch.Tensor([0.48145466, 0.4578275, 0.40821073]), persistent=False) + self.register_buffer('std', torch.Tensor([0.26862954, 0.26130258, 0.27577711]), persistent=False) + + + def preprocess(self, x): + # normalize to [0,1] + x = kornia.geometry.resize(x, (224, 224), + interpolation='bicubic', align_corners=True, + antialias=self.antialias) + x = (x + 1.) / 2. + # renormalize according to clip + x = kornia.enhance.normalize(x, self.mean, self.std) + return x + + def freeze(self): + self.model = self.model.eval() + for param in self.model.parameters(): + param.requires_grad = False + + def forward(self, image, no_dropout=False): + ## image: b c h w + z = self.encode_with_vision_transformer(image) + return z + + def encode_with_vision_transformer(self, x): + x = self.preprocess(x) + + # to patches - whether to use dual patchnorm - https://arxiv.org/abs/2302.01327v1 + if self.model.visual.input_patchnorm: + # einops - rearrange(x, 'b c (h p1) (w p2) -> b (h w) (c p1 p2)') + x = x.reshape(x.shape[0], x.shape[1], self.model.visual.grid_size[0], self.model.visual.patch_size[0], self.model.visual.grid_size[1], self.model.visual.patch_size[1]) + x = x.permute(0, 2, 4, 1, 3, 5) + x = x.reshape(x.shape[0], self.model.visual.grid_size[0] * self.model.visual.grid_size[1], -1) + x = self.model.visual.patchnorm_pre_ln(x) + x = self.model.visual.conv1(x) + else: + x = self.model.visual.conv1(x) # shape = [*, width, grid, grid] + x = x.reshape(x.shape[0], x.shape[1], -1) # shape = [*, width, grid ** 2] + x = x.permute(0, 2, 1) # shape = [*, grid ** 2, width] + + # class embeddings and positional embeddings + x = torch.cat( + [self.model.visual.class_embedding.to(x.dtype) + torch.zeros(x.shape[0], 1, x.shape[-1], dtype=x.dtype, device=x.device), + x], dim=1) # shape = [*, grid ** 2 + 1, width] + x = x + self.model.visual.positional_embedding.to(x.dtype) + + # a patch_dropout of 0. would mean it is disabled and this function would do nothing but return what was passed in + x = self.model.visual.patch_dropout(x) + x = self.model.visual.ln_pre(x) + + x = x.permute(1, 0, 2) # NLD -> LND + x = self.model.visual.transformer(x) + x = x.permute(1, 0, 2) # LND -> NLD + + return x + +class FrozenCLIPT5Encoder(AbstractEncoder): + def __init__(self, clip_version="openai/clip-vit-large-patch14", t5_version="google/t5-v1_1-xl", device="cuda", + clip_max_length=77, t5_max_length=77): + super().__init__() + self.clip_encoder = FrozenCLIPEmbedder(clip_version, device, max_length=clip_max_length) + self.t5_encoder = FrozenT5Embedder(t5_version, device, max_length=t5_max_length) + print(f"{self.clip_encoder.__class__.__name__} has {count_params(self.clip_encoder) * 1.e-6:.2f} M parameters, " + f"{self.t5_encoder.__class__.__name__} comes with {count_params(self.t5_encoder) * 1.e-6:.2f} M params.") + + def encode(self, text): + return self(text) + + def forward(self, text): + clip_z = self.clip_encoder.encode(text) + t5_z = self.t5_encoder.encode(text) + return [clip_z, t5_z] diff --git a/modules/encoders/resampler.py b/modules/encoders/resampler.py new file mode 100644 index 0000000..d9abd23 --- /dev/null +++ b/modules/encoders/resampler.py @@ -0,0 +1,145 @@ +# modified from https://github.com/mlfoundations/open_flamingo/blob/main/open_flamingo/src/helpers.py +# and https://github.com/lucidrains/imagen-pytorch/blob/main/imagen_pytorch/imagen_pytorch.py +# and https://github.com/tencent-ailab/IP-Adapter/blob/main/ip_adapter/resampler.py +import math +import torch +import torch.nn as nn + + +class ImageProjModel(nn.Module): + """Projection Model""" + def __init__(self, cross_attention_dim=1024, clip_embeddings_dim=1024, clip_extra_context_tokens=4): + super().__init__() + self.cross_attention_dim = cross_attention_dim + self.clip_extra_context_tokens = clip_extra_context_tokens + self.proj = nn.Linear(clip_embeddings_dim, self.clip_extra_context_tokens * cross_attention_dim) + self.norm = nn.LayerNorm(cross_attention_dim) + + def forward(self, image_embeds): + #embeds = image_embeds + embeds = image_embeds.type(list(self.proj.parameters())[0].dtype) + clip_extra_context_tokens = self.proj(embeds).reshape(-1, self.clip_extra_context_tokens, self.cross_attention_dim) + clip_extra_context_tokens = self.norm(clip_extra_context_tokens) + return clip_extra_context_tokens + + +# FFN +def FeedForward(dim, mult=4): + inner_dim = int(dim * mult) + return nn.Sequential( + nn.LayerNorm(dim), + nn.Linear(dim, inner_dim, bias=False), + nn.GELU(), + nn.Linear(inner_dim, dim, bias=False), + ) + + +def reshape_tensor(x, heads): + bs, length, width = x.shape + #(bs, length, width) --> (bs, length, n_heads, dim_per_head) + x = x.view(bs, length, heads, -1) + # (bs, length, n_heads, dim_per_head) --> (bs, n_heads, length, dim_per_head) + x = x.transpose(1, 2) + # (bs, n_heads, length, dim_per_head) --> (bs*n_heads, length, dim_per_head) + x = x.reshape(bs, heads, length, -1) + return x + + +class PerceiverAttention(nn.Module): + def __init__(self, *, dim, dim_head=64, heads=8): + super().__init__() + self.scale = dim_head**-0.5 + self.dim_head = dim_head + self.heads = heads + inner_dim = dim_head * heads + + self.norm1 = nn.LayerNorm(dim) + self.norm2 = nn.LayerNorm(dim) + + self.to_q = nn.Linear(dim, inner_dim, bias=False) + self.to_kv = nn.Linear(dim, inner_dim * 2, bias=False) + self.to_out = nn.Linear(inner_dim, dim, bias=False) + + + def forward(self, x, latents): + """ + Args: + x (torch.Tensor): image features + shape (b, n1, D) + latent (torch.Tensor): latent features + shape (b, n2, D) + """ + x = self.norm1(x) + latents = self.norm2(latents) + + b, l, _ = latents.shape + + q = self.to_q(latents) + kv_input = torch.cat((x, latents), dim=-2) + k, v = self.to_kv(kv_input).chunk(2, dim=-1) + + q = reshape_tensor(q, self.heads) + k = reshape_tensor(k, self.heads) + v = reshape_tensor(v, self.heads) + + # attention + scale = 1 / math.sqrt(math.sqrt(self.dim_head)) + weight = (q * scale) @ (k * scale).transpose(-2, -1) # More stable with f16 than dividing afterwards + weight = torch.softmax(weight.float(), dim=-1).type(weight.dtype) + out = weight @ v + + out = out.permute(0, 2, 1, 3).reshape(b, l, -1) + + return self.to_out(out) + + +class Resampler(nn.Module): + def __init__( + self, + dim=1024, + depth=8, + dim_head=64, + heads=16, + num_queries=8, + embedding_dim=768, + output_dim=1024, + ff_mult=4, + video_length=None, # using frame-wise version or not + ): + super().__init__() + ## queries for a single frame / image + self.num_queries = num_queries + self.video_length = video_length + + ## queries for each frame + if video_length is not None: + num_queries = num_queries * video_length + + self.latents = nn.Parameter(torch.randn(1, num_queries, dim) / dim**0.5) + self.proj_in = nn.Linear(embedding_dim, dim) + self.proj_out = nn.Linear(dim, output_dim) + self.norm_out = nn.LayerNorm(output_dim) + + self.layers = nn.ModuleList([]) + for _ in range(depth): + self.layers.append( + nn.ModuleList( + [ + PerceiverAttention(dim=dim, dim_head=dim_head, heads=heads), + FeedForward(dim=dim, mult=ff_mult), + ] + ) + ) + + def forward(self, x): + latents = self.latents.repeat(x.size(0), 1, 1) ## B (T L) C + x = self.proj_in(x) + + for attn, ff in self.layers: + latents = attn(x, latents) + latents + latents = ff(latents) + latents + + latents = self.proj_out(latents) + latents = self.norm_out(latents) # B L C or B (T L) C + + return latents \ No newline at end of file diff --git a/modules/networks/ae_modules.py b/modules/networks/ae_modules.py new file mode 100644 index 0000000..35b6817 --- /dev/null +++ b/modules/networks/ae_modules.py @@ -0,0 +1,1026 @@ +# pytorch_diffusion + derived encoder decoder +import math +import torch +import numpy as np +import torch.nn as nn +from einops import rearrange +from utils.utils import instantiate_from_config +from ...modules.attention import LinearAttention + +import comfy.ops +ops = comfy.ops.disable_weight_init + +def nonlinearity(x): + # swish + return x*torch.sigmoid(x) + + +def Normalize(in_channels, num_groups=32, device=None, dtype=None, operations=ops): + return operations.GroupNorm( + num_groups=num_groups, + num_channels=in_channels, + eps=1e-6, + affine=True, + device=device, + dtype=dtype + ) + +class LinAttnBlock(LinearAttention): + """to match AttnBlock usage""" + def __init__(self, in_channels, device=None, dtype=None): + super().__init__(dim=in_channels, heads=1, dim_head=in_channels, device=device, dtype=dtype) + + +class AttnBlock(nn.Module): + def __init__(self, in_channels, device=None, dtype=None, operations=ops): + super().__init__() + self.in_channels = in_channels + + self.norm = Normalize(in_channels, device=device, dtype=dtype) + self.q = operations.Conv2d( + in_channels, + in_channels, + kernel_size=1, + stride=1, + padding=0, + device=device, + dtype=dtype + ) + self.k = operations.Conv2d( + in_channels, + in_channels, + kernel_size=1, + stride=1, + padding=0, + device=device, + dtype=dtype + ) + self.v = operations.Conv2d( + in_channels, + in_channels, + kernel_size=1, + stride=1, + padding=0, + device=device, + dtype=dtype + ) + self.proj_out = operations.Conv2d( + in_channels, + in_channels, + kernel_size=1, + stride=1, + padding=0, + device=device, + dtype=dtype + ) + + def forward(self, x): + h_ = x + h_ = self.norm(h_) + q = self.q(h_) + k = self.k(h_) + v = self.v(h_) + + # compute attention + b,c,h,w = q.shape + q = q.reshape(b,c,h*w) # bcl + q = q.permute(0,2,1) # bcl -> blc l=hw + k = k.reshape(b,c,h*w) # bcl + + w_ = torch.bmm(q,k) # b,hw,hw w[b,i,j]=sum_c q[b,i,c]k[b,c,j] + w_ = w_ * (int(c)**(-0.5)) + w_ = torch.nn.functional.softmax(w_, dim=2) + + # attend to values + v = v.reshape(b,c,h*w) + w_ = w_.permute(0,2,1) # b,hw,hw (first hw of k, second of q) + h_ = torch.bmm(v,w_) # b, c,hw (hw of q) h_[b,c,j] = sum_i v[b,c,i] w_[b,i,j] + h_ = h_.reshape(b,c,h,w) + + h_ = self.proj_out(h_) + + return x+h_ + +def make_attn(in_channels, attn_type="vanilla", device=None, dtype=None): + assert attn_type in ["vanilla", "linear", "none"], f'attn_type {attn_type} unknown' + #print(f"making attention of type '{attn_type}' with {in_channels} in_channels") + if attn_type == "vanilla": + return AttnBlock(in_channels, device=device, dtype=dtype) + elif attn_type == "none": + return nn.Identity(in_channels) + else: + return LinAttnBlock(in_channels, device=device, dtype=dtype) + +class Downsample(nn.Module): + def __init__(self, in_channels, with_conv, device=None, dtype=None, operations=ops): + super().__init__() + self.with_conv = with_conv + self.in_channels = in_channels + if self.with_conv: + # no asymmetric padding in torch conv, must do it ourselves + self.conv = operations.Conv2d( + in_channels, + in_channels, + kernel_size=3, + stride=2, + padding=0, + device=device, + dtype=dtype + ) + def forward(self, x): + if self.with_conv: + pad = (0,1,0,1) + x = torch.nn.functional.pad(x, pad, mode="constant", value=0) + x = self.conv(x) + else: + x = torch.nn.functional.avg_pool2d(x, kernel_size=2, stride=2) + return x + +class Upsample(nn.Module): + def __init__(self, in_channels, with_conv, device=None, dtype=None, operations=ops): + super().__init__() + self.with_conv = with_conv + self.in_channels = in_channels + if self.with_conv: + self.conv = operations.Conv2d( + in_channels, + in_channels, + kernel_size=3, + stride=1, + padding=1, + device=device, + dtype=dtype + ) + + def forward(self, x): + x = torch.nn.functional.interpolate(x, scale_factor=2.0, mode="nearest") + if self.with_conv: + x = self.conv(x) + return x + +def get_timestep_embedding(timesteps, embedding_dim): + """ + This matches the implementation in Denoising Diffusion Probabilistic Models: + From Fairseq. + Build sinusoidal embeddings. + This matches the implementation in tensor2tensor, but differs slightly + from the description in Section 3.5 of "Attention Is All You Need". + """ + assert len(timesteps.shape) == 1 + + half_dim = embedding_dim // 2 + emb = math.log(10000) / (half_dim - 1) + emb = torch.exp(torch.arange(half_dim, dtype=torch.float32) * -emb) + emb = emb.to(device=timesteps.device) + emb = timesteps.float()[:, None] * emb[None, :] + emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1) + if embedding_dim % 2 == 1: # zero pad + emb = torch.nn.functional.pad(emb, (0,1,0,0)) + return emb + + + +class ResnetBlock(nn.Module): + def __init__( + self, + *, + in_channels, + out_channels=None, + conv_shortcut=False, + dropout, + temb_channels=512, + device=None, + dtype=None, + operations=ops + ): + super().__init__() + self.in_channels = in_channels + out_channels = in_channels if out_channels is None else out_channels + self.out_channels = out_channels + self.use_conv_shortcut = conv_shortcut + + self.norm1 = Normalize(in_channels, device=device, dtype=dtype) + self.conv1 = operations.Conv2d( + in_channels, + out_channels, + kernel_size=3, + stride=1, + padding=1, + device=device, + dtype=dtype + ) + if temb_channels > 0: + self.temb_proj = operations.Linear( + temb_channels, + out_channels, + device=device, + dtype=dtype + ) + self.norm2 = Normalize(out_channels, device=device, dtype=dtype) + self.dropout = torch.nn.Dropout(dropout) + self.conv2 = operations.Conv2d( + out_channels, + out_channels, + kernel_size=3, + stride=1, + padding=1, + device=device, + dtype=dtype + ) + if self.in_channels != self.out_channels: + if self.use_conv_shortcut: + self.conv_shortcut = operations.Conv2d( + in_channels, + out_channels, + kernel_size=3, + stride=1, + padding=1, + device=device, + dtype=device + ) + else: + self.nin_shortcut = operations.Conv2d( + in_channels, + out_channels, + kernel_size=1, + stride=1, + padding=0, + device=device, + dtype=dtype + ) + + def forward(self, x, temb): + h = x + h = self.norm1(h) + h = nonlinearity(h) + h = self.conv1(h) + + if temb is not None: + h = h + self.temb_proj(nonlinearity(temb))[:,:,None,None] + + h = self.norm2(h) + h = nonlinearity(h) + h = self.dropout(h) + h = self.conv2(h) + + if self.in_channels != self.out_channels: + if self.use_conv_shortcut: + x = self.conv_shortcut(x) + else: + x = self.nin_shortcut(x) + + return x+h + +class Model(nn.Module): + def __init__( + self, + *, + ch, + out_ch, + ch_mult=(1,2,4,8), + num_res_blocks, + attn_resolutions, + dropout=0.0, + resamp_with_conv=True, + in_channels, + resolution, + use_timestep=True, + use_linear_attn=False, + attn_type="vanilla", + device=None, + dtype=None, + operations=ops + ): + super().__init__() + if use_linear_attn: attn_type = "linear" + self.ch = ch + self.temb_ch = self.ch*4 + self.num_resolutions = len(ch_mult) + self.num_res_blocks = num_res_blocks + self.resolution = resolution + self.in_channels = in_channels + + self.use_timestep = use_timestep + if self.use_timestep: + # timestep embedding + self.temb = nn.Module() + self.temb.dense = nn.ModuleList([ + operations.Linear( + self.ch, + self.temb_ch, + device=device, + dtype=dtype + ), + operations.Linear( + self.temb_ch, + self.temb_ch, + device=device, + dtype=dtype + ), + ]) + + # downsampling + self.conv_in = operations.Conv2d( + in_channels, + self.ch, + kernel_size=3, + stride=1, + padding=1, + device=device, + dtype=dtype + ) + + curr_res = resolution + in_ch_mult = (1,)+tuple(ch_mult) + self.down = nn.ModuleList() + for i_level in range(self.num_resolutions): + block = nn.ModuleList() + attn = nn.ModuleList() + block_in = ch*in_ch_mult[i_level] + block_out = ch*ch_mult[i_level] + for i_block in range(self.num_res_blocks): + block.append(ResnetBlock(in_channels=block_in, + out_channels=block_out, + temb_channels=self.temb_ch, + dropout=dropout, + device=device, + dtype=dtype)) + block_in = block_out + if curr_res in attn_resolutions: + attn.append(make_attn(block_in, attn_type=attn_type, device=device, dtype=dtype)) + down = nn.Module() + down.block = block + down.attn = attn + if i_level != self.num_resolutions-1: + down.downsample = Downsample(block_in, resamp_with_conv, device=device, dtype=dtype) + curr_res = curr_res // 2 + self.down.append(down) + + # middle + self.mid = nn.Module() + self.mid.block_1 = ResnetBlock(in_channels=block_in, + out_channels=block_in, + temb_channels=self.temb_ch, + dropout=dropout, + device=device, + dtype=dtype) + self.mid.attn_1 = make_attn(block_in, attn_type=attn_type, device=device, dtype=dtype) + self.mid.block_2 = ResnetBlock(in_channels=block_in, + out_channels=block_in, + temb_channels=self.temb_ch, + dropout=dropout, + device=device, + dtype=dtype) + + # upsampling + self.up = nn.ModuleList() + for i_level in reversed(range(self.num_resolutions)): + block = nn.ModuleList() + attn = nn.ModuleList() + block_out = ch*ch_mult[i_level] + skip_in = ch*ch_mult[i_level] + for i_block in range(self.num_res_blocks+1): + if i_block == self.num_res_blocks: + skip_in = ch*in_ch_mult[i_level] + block.append(ResnetBlock(in_channels=block_in+skip_in, + out_channels=block_out, + temb_channels=self.temb_ch, + dropout=dropout, + device=device, + dtype=dtype)) + block_in = block_out + if curr_res in attn_resolutions: + attn.append(make_attn(block_in, attn_type=attn_type, device=device, dtype=dtype)) + up = nn.Module() + up.block = block + up.attn = attn + if i_level != 0: + up.upsample = Upsample(block_in, resamp_with_conv, device=device, dtype=dtype) + curr_res = curr_res * 2 + self.up.insert(0, up) # prepend to get consistent order + + # end + self.norm_out = Normalize(block_in, device=device, dtype=device) + self.conv_out = torch.nn.Conv2d(block_in, + out_ch, + kernel_size=3, + stride=1, + padding=1, + device=device, + dtype=dtype) + + def forward(self, x, t=None, context=None): + #assert x.shape[2] == x.shape[3] == self.resolution + if context is not None: + # assume aligned context, cat along channel axis + x = torch.cat((x, context), dim=1) + if self.use_timestep: + # timestep embedding + assert t is not None + temb = get_timestep_embedding(t, self.ch) + temb = self.temb.dense[0](temb) + temb = nonlinearity(temb) + temb = self.temb.dense[1](temb) + else: + temb = None + + # downsampling + hs = [self.conv_in(x)] + for i_level in range(self.num_resolutions): + for i_block in range(self.num_res_blocks): + h = self.down[i_level].block[i_block](hs[-1], temb) + if len(self.down[i_level].attn) > 0: + h = self.down[i_level].attn[i_block](h) + hs.append(h) + if i_level != self.num_resolutions-1: + hs.append(self.down[i_level].downsample(hs[-1])) + + # middle + h = hs[-1] + h = self.mid.block_1(h, temb) + h = self.mid.attn_1(h) + h = self.mid.block_2(h, temb) + + # upsampling + for i_level in reversed(range(self.num_resolutions)): + for i_block in range(self.num_res_blocks+1): + h = self.up[i_level].block[i_block]( + torch.cat([h, hs.pop()], dim=1), temb) + if len(self.up[i_level].attn) > 0: + h = self.up[i_level].attn[i_block](h) + if i_level != 0: + h = self.up[i_level].upsample(h) + + # end + h = self.norm_out(h) + h = nonlinearity(h) + h = self.conv_out(h) + return h + + def get_last_layer(self): + return self.conv_out.weight + + +class Encoder(nn.Module): + def __init__( + self, + *, + ch, + out_ch, + ch_mult=(1,2,4,8), + num_res_blocks, + attn_resolutions, + dropout=0.0, + resamp_with_conv=True, + in_channels, + resolution, + z_channels, + double_z=True, + use_linear_attn=False, + attn_type="vanilla", + device=None, + dtype=None, + operations=ops, + **ignore_kwargs + ): + super().__init__() + if use_linear_attn: attn_type = "linear" + self.ch = ch + self.temb_ch = 0 + self.num_resolutions = len(ch_mult) + self.num_res_blocks = num_res_blocks + self.resolution = resolution + self.in_channels = in_channels + + # downsampling + self.conv_in = operations.Conv2d(in_channels, + self.ch, + kernel_size=3, + stride=1, + padding=1, + device=device, + dtype=dtype) + + curr_res = resolution + in_ch_mult = (1,)+tuple(ch_mult) + self.in_ch_mult = in_ch_mult + self.down = nn.ModuleList() + for i_level in range(self.num_resolutions): + block = nn.ModuleList() + attn = nn.ModuleList() + block_in = ch*in_ch_mult[i_level] + block_out = ch*ch_mult[i_level] + for i_block in range(self.num_res_blocks): + block.append(ResnetBlock(in_channels=block_in, + out_channels=block_out, + temb_channels=self.temb_ch, + dropout=dropout, + device=device, + dtype=dtype)) + block_in = block_out + if curr_res in attn_resolutions: + attn.append(make_attn(block_in, attn_type=attn_type, device=device, dtype=dtype)) + down = nn.Module() + down.block = block + down.attn = attn + if i_level != self.num_resolutions-1: + down.downsample = Downsample(block_in, resamp_with_conv, device=device, dtype=dtype) + curr_res = curr_res // 2 + self.down.append(down) + + # middle + self.mid = nn.Module() + self.mid.block_1 = ResnetBlock(in_channels=block_in, + out_channels=block_in, + temb_channels=self.temb_ch, + dropout=dropout, + device=device, + dtype=dtype) + self.mid.attn_1 = make_attn(block_in, attn_type=attn_type, device=device, dtype=dtype) + self.mid.block_2 = ResnetBlock(in_channels=block_in, + out_channels=block_in, + temb_channels=self.temb_ch, + dropout=dropout, + device=device, + dtype=dtype) + + # end + self.norm_out = Normalize(block_in, device=device, dtype=dtype) + self.conv_out = operations.Conv2d(block_in, + 2*z_channels if double_z else z_channels, + kernel_size=3, + stride=1, + padding=1, + device=device, + dtype=dtype) + + def forward(self, x): + # timestep embedding + temb = None + + # print(f'encoder-input={x.shape}') + # downsampling + hs = [self.conv_in(x)] + # print(f'encoder-conv in feat={hs[0].shape}') + for i_level in range(self.num_resolutions): + for i_block in range(self.num_res_blocks): + h = self.down[i_level].block[i_block](hs[-1], temb) + # print(f'encoder-down feat={h.shape}') + if len(self.down[i_level].attn) > 0: + h = self.down[i_level].attn[i_block](h) + hs.append(h) + if i_level != self.num_resolutions-1: + # print(f'encoder-downsample (input)={hs[-1].shape}') + hs.append(self.down[i_level].downsample(hs[-1])) + # print(f'encoder-downsample (output)={hs[-1].shape}') + + # middle + h = hs[-1] + h = self.mid.block_1(h, temb) + # print(f'encoder-mid1 feat={h.shape}') + h = self.mid.attn_1(h) + h = self.mid.block_2(h, temb) + # print(f'encoder-mid2 feat={h.shape}') + + # end + h = self.norm_out(h) + h = nonlinearity(h) + h = self.conv_out(h) + # print(f'end feat={h.shape}') + return h + + +class Decoder(nn.Module): + def __init__( + self, + *, + ch, + out_ch, + ch_mult=(1,2,4,8), + num_res_blocks, + attn_resolutions, + dropout=0.0, + resamp_with_conv=True, + in_channels, + resolution, + z_channels, + give_pre_end=False, + tanh_out=False, + use_linear_attn=False, + attn_type="vanilla", + device=None, + dtype=None, + operations=ops, + **ignorekwargs + ): + super().__init__() + if use_linear_attn: attn_type = "linear" + self.ch = ch + self.temb_ch = 0 + self.num_resolutions = len(ch_mult) + self.num_res_blocks = num_res_blocks + self.resolution = resolution + self.in_channels = in_channels + self.give_pre_end = give_pre_end + self.tanh_out = tanh_out + + # compute in_ch_mult, block_in and curr_res at lowest res + in_ch_mult = (1,)+tuple(ch_mult) + block_in = ch*ch_mult[self.num_resolutions-1] + curr_res = resolution // 2**(self.num_resolutions-1) + self.z_shape = (1,z_channels,curr_res,curr_res) + print("AE working on z of shape {} = {} dimensions.".format( + self.z_shape, np.prod(self.z_shape))) + + # z to block_in + self.conv_in = torch.nn.Conv2d(z_channels, + block_in, + kernel_size=3, + stride=1, + padding=1, + device=device, + dtype=dtype) + + # middle + self.mid = nn.Module() + self.mid.block_1 = ResnetBlock(in_channels=block_in, + out_channels=block_in, + temb_channels=self.temb_ch, + dropout=dropout, + device=device, + dtype=dtype) + self.mid.attn_1 = make_attn(block_in, attn_type=attn_type, device=device, dtype=dtype) + self.mid.block_2 = ResnetBlock(in_channels=block_in, + out_channels=block_in, + temb_channels=self.temb_ch, + dropout=dropout, + device=device, + dtype=dtype) + + # upsampling + self.up = nn.ModuleList() + for i_level in reversed(range(self.num_resolutions)): + block = nn.ModuleList() + attn = nn.ModuleList() + block_out = ch*ch_mult[i_level] + for i_block in range(self.num_res_blocks+1): + block.append(ResnetBlock(in_channels=block_in, + out_channels=block_out, + temb_channels=self.temb_ch, + dropout=dropout, + device=device, + dtype=dtype)) + block_in = block_out + if curr_res in attn_resolutions: + attn.append(make_attn(block_in, attn_type=attn_type, device=device, dtype=dtype)) + up = nn.Module() + up.block = block + up.attn = attn + if i_level != 0: + up.upsample = Upsample(block_in, resamp_with_conv, device=device, dtype=dtype) + curr_res = curr_res * 2 + self.up.insert(0, up) # prepend to get consistent order + + # end + self.norm_out = Normalize(block_in, device=device, dtype=dtype) + self.conv_out = operations.Conv2d(block_in, + out_ch, + kernel_size=3, + stride=1, + padding=1, + device=device, + dtype=dtype) + + def forward(self, z): + #assert z.shape[1:] == self.z_shape[1:] + self.last_z_shape = z.shape + + # print(f'decoder-input={z.shape}') + # timestep embedding + temb = None + + # z to block_in + h = self.conv_in(z) + # print(f'decoder-conv in feat={h.shape}') + + # middle + h = self.mid.block_1(h, temb) + h = self.mid.attn_1(h) + h = self.mid.block_2(h, temb) + # print(f'decoder-mid feat={h.shape}') + + # upsampling + for i_level in reversed(range(self.num_resolutions)): + for i_block in range(self.num_res_blocks+1): + h = self.up[i_level].block[i_block](h, temb) + if len(self.up[i_level].attn) > 0: + h = self.up[i_level].attn[i_block](h) + # print(f'decoder-up feat={h.shape}') + if i_level != 0: + h = self.up[i_level].upsample(h) + # print(f'decoder-upsample feat={h.shape}') + + # end + if self.give_pre_end: + return h + + h = self.norm_out(h) + h = nonlinearity(h) + h = self.conv_out(h) + # print(f'decoder-conv_out feat={h.shape}') + if self.tanh_out: + h = torch.tanh(h) + return h + + +class SimpleDecoder(nn.Module): + def __init__(self, in_channels, out_channels, device=None, dtype=None, operations=ops, *args, **kwargs): + super().__init__() + self.model = nn.ModuleList([nn.Conv2d(in_channels, in_channels, 1), + ResnetBlock(in_channels=in_channels, + out_channels=2 * in_channels, + temb_channels=0, dropout=0.0, + device=device, + dtype=dtype), + ResnetBlock(in_channels=2 * in_channels, + out_channels=4 * in_channels, + temb_channels=0, dropout=0.0, + device=device, + dtype=dtype), + ResnetBlock(in_channels=4 * in_channels, + out_channels=2 * in_channels, + temb_channels=0, dropout=0.0, + device=device, + dtype=dtype), + operations.Conv2d(2*in_channels, in_channels, 1), + Upsample(in_channels, with_conv=True, device=device, dtype=dtype)]) + # end + self.norm_out = Normalize(in_channels, device=device, dtype=dtype) + self.conv_out = torch.nn.Conv2d(in_channels, + out_channels, + kernel_size=3, + stride=1, + padding=1, + device=device, + dtype=dtype) + + def forward(self, x): + for i, layer in enumerate(self.model): + if i in [1,2,3]: + x = layer(x, None) + else: + x = layer(x) + + h = self.norm_out(x) + h = nonlinearity(h) + x = self.conv_out(h) + return x + + +class UpsampleDecoder(nn.Module): + def __init__(self, in_channels, out_channels, ch, num_res_blocks, resolution, + ch_mult=(2,2), dropout=0.0, device=None, dtype=None, operations=ops): + super().__init__() + # upsampling + self.temb_ch = 0 + self.num_resolutions = len(ch_mult) + self.num_res_blocks = num_res_blocks + block_in = in_channels + curr_res = resolution // 2 ** (self.num_resolutions - 1) + self.res_blocks = nn.ModuleList() + self.upsample_blocks = nn.ModuleList() + for i_level in range(self.num_resolutions): + res_block = [] + block_out = ch * ch_mult[i_level] + for i_block in range(self.num_res_blocks + 1): + res_block.append(ResnetBlock(in_channels=block_in, + out_channels=block_out, + temb_channels=self.temb_ch, + dropout=dropout, + device=device, + dtype=dtype)) + block_in = block_out + self.res_blocks.append(nn.ModuleList(res_block)) + if i_level != self.num_resolutions - 1: + self.upsample_blocks.append(Upsample(block_in, True, device=device, dtype=dtype)) + curr_res = curr_res * 2 + + # end + self.norm_out = Normalize(block_in, device=device, dtype=dtype) + self.conv_out = torch.nn.Conv2d(block_in, + out_channels, + kernel_size=3, + stride=1, + padding=1, + device=device, + dtype=dtype) + + def forward(self, x): + # upsampling + h = x + for k, i_level in enumerate(range(self.num_resolutions)): + for i_block in range(self.num_res_blocks + 1): + h = self.res_blocks[i_level][i_block](h, None) + if i_level != self.num_resolutions - 1: + h = self.upsample_blocks[k](h) + h = self.norm_out(h) + h = nonlinearity(h) + h = self.conv_out(h) + return h + + +class LatentRescaler(nn.Module): + def __init__(self, factor, in_channels, mid_channels, out_channels, depth=2, device=None, dtype=None, operations=ops): + super().__init__() + # residual block, interpolate, residual block + self.factor = factor + self.conv_in = operations.Conv2d(in_channels, + mid_channels, + kernel_size=3, + stride=1, + padding=1, + device=device, + dtype=dtype) + self.res_block1 = nn.ModuleList([ResnetBlock(in_channels=mid_channels, + out_channels=mid_channels, + temb_channels=0, + dropout=0.0, + device=device, + dtype=dtype) for _ in range(depth)]) + self.attn = AttnBlock(mid_channels, device=device, dtype=dtype) + self.res_block2 = nn.ModuleList([ResnetBlock(in_channels=mid_channels, + out_channels=mid_channels, + temb_channels=0, + dropout=0.0, + device=device, + dtype=dtype) for _ in range(depth)]) + + self.conv_out = operations.Conv2d(mid_channels, + out_channels, + kernel_size=1, + device=device, + dtype=dtype + ) + + def forward(self, x): + x = self.conv_in(x) + for block in self.res_block1: + x = block(x, None) + x = torch.nn.functional.interpolate(x, size=(int(round(x.shape[2]*self.factor)), int(round(x.shape[3]*self.factor)))) + x = self.attn(x) + for block in self.res_block2: + x = block(x, None) + x = self.conv_out(x) + return x + + +class MergedRescaleEncoder(nn.Module): + def __init__(self, in_channels, ch, resolution, out_ch, num_res_blocks, + attn_resolutions, dropout=0.0, resamp_with_conv=True, + ch_mult=(1,2,4,8), rescale_factor=1.0, rescale_module_depth=1, device=None, dtype=None, operations=ops): + super().__init__() + intermediate_chn = ch * ch_mult[-1] + self.encoder = Encoder(in_channels=in_channels, num_res_blocks=num_res_blocks, ch=ch, ch_mult=ch_mult, + z_channels=intermediate_chn, double_z=False, resolution=resolution, + attn_resolutions=attn_resolutions, dropout=dropout, resamp_with_conv=resamp_with_conv, + out_ch=None, device=device, dtype=dtype) + self.rescaler = LatentRescaler(factor=rescale_factor, in_channels=intermediate_chn, + mid_channels=intermediate_chn, out_channels=out_ch, depth=rescale_module_depth, + device=device, dtype=dtype) + + def forward(self, x): + x = self.encoder(x) + x = self.rescaler(x) + return x + + +class MergedRescaleDecoder(nn.Module): + def __init__(self, z_channels, out_ch, resolution, num_res_blocks, attn_resolutions, ch, ch_mult=(1,2,4,8), + dropout=0.0, resamp_with_conv=True, rescale_factor=1.0, rescale_module_depth=1, + device=None, dtype=None, operations=ops): + super().__init__() + tmp_chn = z_channels*ch_mult[-1] + self.decoder = Decoder(out_ch=out_ch, z_channels=tmp_chn, attn_resolutions=attn_resolutions, dropout=dropout, + resamp_with_conv=resamp_with_conv, in_channels=None, num_res_blocks=num_res_blocks, + ch_mult=ch_mult, resolution=resolution, ch=ch, device=device, operations=ops) + self.rescaler = LatentRescaler(factor=rescale_factor, in_channels=z_channels, mid_channels=tmp_chn, + out_channels=tmp_chn, depth=rescale_module_depth, device=device, operations=ops) + + def forward(self, x): + x = self.rescaler(x) + x = self.decoder(x) + return x + + +class Upsampler(nn.Module): + def __init__(self, in_size, out_size, in_channels, out_channels, ch_mult=2, device=None, dtype=None, operations=ops): + super().__init__() + assert out_size >= in_size + num_blocks = int(np.log2(out_size//in_size))+1 + factor_up = 1.+ (out_size % in_size) + print(f"Building {self.__class__.__name__} with in_size: {in_size} --> out_size {out_size} and factor {factor_up}") + self.rescaler = LatentRescaler(factor=factor_up, in_channels=in_channels, mid_channels=2*in_channels, + out_channels=in_channels, device=device, dtype=dtype) + self.decoder = Decoder(out_ch=out_channels, resolution=out_size, z_channels=in_channels, num_res_blocks=2, + attn_resolutions=[], in_channels=None, ch=in_channels, device=device, dtype=dtype, + ch_mult=[ch_mult for _ in range(num_blocks)]) + + def forward(self, x): + x = self.rescaler(x) + x = self.decoder(x) + return x + + +class Resize(nn.Module): + def __init__(self, in_channels=None, learned=False, mode="bilinear", device=None, dtype=None, operations=ops): + super().__init__() + self.with_conv = learned + self.mode = mode + if self.with_conv: + print(f"Note: {self.__class__.__name} uses learned downsampling and will ignore the fixed {mode} mode") + raise NotImplementedError() + assert in_channels is not None + # no asymmetric padding in torch conv, must do it ourselves + self.conv = operations.Conv2d(in_channels, + in_channels, + kernel_size=4, + stride=2, + padding=1, + device=device, + dtype=dtype) + + def forward(self, x, scale_factor=1.0): + if scale_factor==1.0: + return x + else: + x = torch.nn.functional.interpolate(x, mode=self.mode, align_corners=False, scale_factor=scale_factor) + return x + +class FirstStagePostProcessor(nn.Module): + + def __init__(self, ch_mult:list, in_channels, + pretrained_model:nn.Module=None, + reshape=False, + n_channels=None, + dropout=0., + pretrained_config=None, + device=None, + dtype=None, + operations=ops): + super().__init__() + if pretrained_config is None: + assert pretrained_model is not None, 'Either "pretrained_model" or "pretrained_config" must not be None' + self.pretrained_model = pretrained_model + else: + assert pretrained_config is not None, 'Either "pretrained_model" or "pretrained_config" must not be None' + self.instantiate_pretrained(pretrained_config) + + self.do_reshape = reshape + + if n_channels is None: + n_channels = self.pretrained_model.encoder.ch + + self.proj_norm = Normalize(in_channels,num_groups=in_channels//2, device=device, dtype=dtype) + self.proj = nn.Conv2d(in_channels,n_channels,kernel_size=3, + stride=1,padding=1, device=device, dtype=dtype) + + blocks = [] + downs = [] + ch_in = n_channels + for m in ch_mult: + blocks.append(ResnetBlock(in_channels=ch_in,out_channels=m*n_channels,dropout=dropout, device=device, dtype=dtype)) + ch_in = m * n_channels + downs.append(Downsample(ch_in, with_conv=False, device=device, dtype=dtype)) + + self.model = nn.ModuleList(blocks) + self.downsampler = nn.ModuleList(downs) + + + def instantiate_pretrained(self, config): + model = instantiate_from_config(config) + self.pretrained_model = model.eval() + # self.pretrained_model.train = False + for param in self.pretrained_model.parameters(): + param.requires_grad = False + + + @torch.no_grad() + def encode_with_pretrained(self,x): + c = self.pretrained_model.encode(x) + if isinstance(c, DiagonalGaussianDistribution): + c = c.mode() + return c + + def forward(self,x): + z_fs = self.encode_with_pretrained(x) + z = self.proj_norm(z_fs) + z = self.proj(z) + z = nonlinearity(z) + + for submodel, downmodel in zip(self.model,self.downsampler): + z = submodel(z,temb=None) + z = downmodel(z) + + if self.do_reshape: + z = rearrange(z,'b c h w -> b (h w) c') + return z \ No newline at end of file diff --git a/modules/networks/openaimodel3d.py b/modules/networks/openaimodel3d.py new file mode 100644 index 0000000..838b6fe --- /dev/null +++ b/modules/networks/openaimodel3d.py @@ -0,0 +1,822 @@ +from functools import partial +from abc import abstractmethod +import torch +import torch.nn as nn +from einops import rearrange +import torch.nn.functional as F +from ...models.utils_diffusion import timestep_embedding +from ...common import checkpoint +from ...basics import ( + zero_module, + conv_nd, + linear, + avg_pool_nd, + normalization +) +from ...modules.attention import SpatialTransformer, TemporalTransformer +import comfy.ops +import logging + +ops = comfy.ops.disable_weight_init + +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. + """ + +#This is needed because accelerate makes a copy of transformer_options which breaks "transformer_index" +def forward_timestep_embed(ts, x, emb, context=None, batch_size=None, transformer_options={}): + for layer in ts: + if isinstance(layer, TimestepBlock): + x = layer(x, emb, batch_size=batch_size) + elif isinstance(layer, SpatialTransformer): + x = layer(x, context) + if "transformer_index" in transformer_options: + transformer_options["transformer_index"] += 1 + elif isinstance(layer, TemporalTransformer): + x = rearrange(x, '(b f) c h w -> b c f h w', b=batch_size) + x = layer(x, context) + if "transformer_index" in transformer_options: + transformer_options["transformer_index"] += 1 + x = rearrange(x, 'b c f h w -> (b f) c h w') + else: + x = layer(x) + return x + +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, *args, **kwargs): + return forward_timestep_embed(self, *args, **kwargs) + +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, dtype=None, device=None, operations=ops): + 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 = operations.conv_nd( + dims, 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 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, dtype=None, device=None, operations=ops): + 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 = operations.conv_nd(dims, self.channels, self.out_channels, 3, padding=padding, dtype=dtype, device=device) + + 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(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 up: if True, use this block for upsampling. + :param down: if True, use this block for downsampling. + :param use_temporal_conv: if True, use the temporal convolution. + :param use_image_dataset: if True, the temporal parameters will not be optimized. + """ + + def __init__( + self, + channels, + emb_channels, + dropout, + out_channels=None, + use_scale_shift_norm=False, + dims=2, + use_checkpoint=False, + use_conv=False, + up=False, + down=False, + kernel_size=3, + use_temporal_conv=False, + tempspatial_aware=False, + dtype=None, + device=None, + operations=ops + ): + 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.use_temporal_conv = use_temporal_conv + + if isinstance(kernel_size, list): + padding =[k // 2 for k in kernel_size] + else: + padding = kernel_size // 2 + + # operations used in normalization function + self.in_layers = nn.Sequential( + normalization(channels, dtype=dtype, device=device), + nn.SiLU(), + operations.conv_nd(dims, channels, self.out_channels, 3, padding=1, dtype=dtype, device=device), + ) + + self.updown = up or down + + if up: + self.h_upd = Upsample(channels, False, dims, dtype=dtype, device=device) + self.x_upd = Upsample(channels, False, dims, dtype=dtype, device=device) + elif down: + self.h_upd = Downsample(channels, False, dims, dtype=dtype, device=device) + self.x_upd = Downsample(channels, False, dims, dtype=dtype, device=device) + else: + self.h_upd = self.x_upd = nn.Identity() + + self.emb_layers = nn.Sequential( + nn.SiLU(), + operations.Linear( + emb_channels, + 2 * self.out_channels if use_scale_shift_norm else self.out_channels, + dtype=dtype, + device=device + ), + ) + self.out_layers = nn.Sequential( + normalization(self.out_channels, dtype=dtype, device=device), + nn.SiLU(), + nn.Dropout(p=dropout), + zero_module(operations.Conv2d(self.out_channels, self.out_channels, 3, padding=1, dtype=dtype, device=device)), + ) + + if self.out_channels == channels: + self.skip_connection = nn.Identity() + elif use_conv: + self.skip_connection = operations.conv_nd(dims, channels, self.out_channels, 3, padding=1, dtype=dtype, device=device) + else: + self.skip_connection = operations.conv_nd(dims, channels, self.out_channels, 1, dtype=dtype, device=device) + + if self.use_temporal_conv: + self.temopral_conv = TemporalConvBlock( + self.out_channels, + self.out_channels, + dropout=0.1, + spatial_aware=tempspatial_aware, + dtype=dtype, + device=device + ) + + def forward(self, x, emb, batch_size=None): + """ + 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. + """ + input_tuple = (x, emb) + if batch_size: + forward_batchsize = partial(self._forward, batch_size=batch_size) + return checkpoint(forward_batchsize, input_tuple, self.parameters(), self.use_checkpoint) + return checkpoint(self._forward, input_tuple, self.parameters(), self.use_checkpoint) + + def _forward(self, x, emb, batch_size=None): + 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 = torch.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 and batch_size: + h = rearrange(h, '(b t) c h w -> b c t h w', b=batch_size) + h = self.temopral_conv(h) + h = rearrange(h, 'b c t h w -> (b t) c h w') + return h + +class TemporalConvBlock(nn.Module): + """ + Adapted from modelscope: https://github.com/modelscope/modelscope/blob/master/modelscope/models/multi_modal/video_synthesis/unet_sd.py + """ + def __init__( + self, + in_channels, + out_channels=None, + dropout=0.0, + spatial_aware=False, + dtype=None, + device=None, + operations=ops + ): + super(TemporalConvBlock, self).__init__() + if out_channels is None: + out_channels = in_channels + self.in_channels = in_channels + self.out_channels = out_channels + th_kernel_shape = (3, 1, 1) if not spatial_aware else (3, 3, 1) + th_padding_shape = (1, 0, 0) if not spatial_aware else (1, 1, 0) + tw_kernel_shape = (3, 1, 1) if not spatial_aware else (3, 1, 3) + tw_padding_shape = (1, 0, 0) if not spatial_aware else (1, 0, 1) + + # conv layers + self.conv1 = nn.Sequential( + operations.GroupNorm(32, in_channels, device=device, dtype=dtype), nn.SiLU(), + operations.Conv3d(in_channels, out_channels, th_kernel_shape, padding=th_padding_shape, device=device, dtype=dtype)) + self.conv2 = nn.Sequential( + operations.GroupNorm(32, out_channels, device=device, dtype=dtype), nn.SiLU(), nn.Dropout(dropout), + operations.Conv3d(out_channels, in_channels, tw_kernel_shape, padding=tw_padding_shape, device=device, dtype=dtype)) + self.conv3 = nn.Sequential( + operations.GroupNorm(32, out_channels, device=device, dtype=dtype), nn.SiLU(), nn.Dropout(dropout), + operations.Conv3d(out_channels, in_channels, th_kernel_shape, padding=th_padding_shape, device=device, dtype=dtype)) + self.conv4 = nn.Sequential( + operations.GroupNorm(32, out_channels, device=device, dtype=dtype), nn.SiLU(), nn.Dropout(dropout), + operations.Conv3d(out_channels, in_channels, tw_kernel_shape, padding=tw_padding_shape, device=device, dtype=dtype)) + + # 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) + + return identity + x + +def context_processor(context, t, img_emb=None, temporal_size=16, concat_only=False, disable_concat=False): + if disable_concat: + return context + + ## repeat t times for context [(b t) 77 768] & time embedding + ## check if we use per-frame image conditioning + + if img_emb is not None: + context = torch.cat([context, img_emb.to(context.device, context.dtype)], dim=1) + + if concat_only: + return context + + b, l_context, _ = context.shape + if l_context == 77 + t * temporal_size: + context_text, context_img = context[:,:77,:], context[:,77:,:] + context_text = context_text.repeat_interleave(repeats=t, dim=0) + context_img = rearrange(context_img, 'b (t l) c -> (b t) l c', t=t) + context = torch.cat([context_text, context_img], dim=1) + else: + context = context.repeat_interleave(repeats=t, dim=0) + + return context + +def apply_control(h, control, name, cond_idx=None): + if control is not None and name in control and len(control[name]) > 0: + frames = h.shape[0] + ctrl = control[name].pop() + if ctrl is not None: + try: + if cond_idx is not None and ctrl.shape[0] > frames: + ctrl_frames_list = list(range(ctrl.shape[0])) + ctrl_frames = len(ctrl_frames_list) + + idxs = ( + ctrl_frames_list[:ctrl_frames // 2] if cond_idx == 0 else \ + ctrl_frames_list[ctrl_frames // 2:] + ) + + ctrl = ctrl[idxs] + + h += ctrl + except Exception as e: + if h.shape != ctrl.shape: + logging.warning( + "warning control could not be applied {} {}".format(h.shape, ctrl.shape) + ) + logging.warning(e) + return h + +class UNetModel(nn.Module): + """ + The full UNet model with attention and timestep embedding. + :param in_channels: in_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, + in_channels, + model_channels, + out_channels, + num_res_blocks, + attention_resolutions, + dropout=0.0, + channel_mult=(1, 2, 4, 8), + conv_resample=True, + dims=2, + context_dim=None, + use_scale_shift_norm=False, + resblock_updown=False, + num_heads=-1, + num_head_channels=-1, + transformer_depth=1, + use_linear=False, + use_checkpoint=False, + temporal_conv=False, + tempspatial_aware=False, + temporal_attention=True, + use_relative_position=True, + use_causal_attention=False, + temporal_length=None, + use_fp16=False, + addition_attention=False, + temporal_selfatt_only=True, + image_cross_attention=False, + image_cross_attention_scale_learnable=False, + default_fs=4, + fs_condition=False, + device=None, + dtype=torch.float16, + operations=ops + ): + super(UNetModel, self).__init__() + 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.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.temporal_attention = temporal_attention + time_embed_dim = model_channels * 4 + self.use_checkpoint = use_checkpoint + temporal_self_att_only = True + self.addition_attention = addition_attention + self.temporal_length = temporal_length + self.image_cross_attention = image_cross_attention + self.image_cross_attention_scale_learnable = image_cross_attention_scale_learnable + self.default_fs = default_fs + self.fs_condition = fs_condition + self.device = device + #self.dtype = dtype + self.dtype = torch.float32 + + ## Time embedding blocks + self.time_embed = nn.Sequential( + linear(model_channels, time_embed_dim, device=device, dtype=self.dtype), + nn.SiLU(), + linear(time_embed_dim, time_embed_dim, device=device, dtype=self.dtype), + ) + if fs_condition: + self.fps_embedding = nn.Sequential( + linear(model_channels, time_embed_dim, device=device, dtype=self.dtype), + nn.SiLU(), + linear(time_embed_dim, time_embed_dim, device=device, dtype=self.dtype), + ) + nn.init.zeros_(self.fps_embedding[-1].weight) + nn.init.zeros_(self.fps_embedding[-1].bias) + ## Input Block + self.input_blocks = nn.ModuleList( + [ + TimestepEmbedSequential( + operations.conv_nd( + dims, + in_channels, + model_channels, + 3, + padding=1, + device=device, + dtype=self.dtype + )) + ] + ) + if self.addition_attention: + self.init_attn=TimestepEmbedSequential( + TemporalTransformer( + model_channels, + n_heads=8, + d_head=num_head_channels, + depth=transformer_depth, + context_dim=context_dim, + use_checkpoint=use_checkpoint, only_self_att=temporal_selfatt_only, + causal_attention=False, relative_position=use_relative_position, + temporal_length=temporal_length, + device=device, + dtype=self.dtype + )) + + 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, tempspatial_aware=tempspatial_aware, + use_temporal_conv=temporal_conv, + device=device, + dtype=self.dtype + ) + ] + 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 + layers.append( + SpatialTransformer(ch, num_heads, dim_head, + depth=transformer_depth, context_dim=context_dim, use_linear=use_linear, + use_checkpoint=use_checkpoint, disable_self_attn=False, + video_length=temporal_length, image_cross_attention=self.image_cross_attention, + image_cross_attention_scale_learnable=self.image_cross_attention_scale_learnable, + device=device, + dtype=self.dtype + ) + ) + if self.temporal_attention: + layers.append( + TemporalTransformer(ch, num_heads, dim_head, + depth=transformer_depth, context_dim=context_dim, use_linear=use_linear, + use_checkpoint=use_checkpoint, only_self_att=temporal_self_att_only, + causal_attention=use_causal_attention, relative_position=use_relative_position, + temporal_length=temporal_length, + device=device, + dtype=self.dtype + ) + ) + self.input_blocks.append(TimestepEmbedSequential(*layers)) + 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, + device=device, + dtype=self.dtype + ) + if resblock_updown + else Downsample( + ch, + conv_resample, + dims=dims, + out_channels=out_ch, + device=device, + dtype=self.dtype + ) + ) + ) + ch = out_ch + input_block_chans.append(ch) + ds *= 2 + + if num_head_channels == -1: + dim_head = ch // num_heads + else: + num_heads = ch // num_head_channels + dim_head = num_head_channels + layers = [ + ResBlock(ch, time_embed_dim, dropout, + dims=dims, use_checkpoint=use_checkpoint, + use_scale_shift_norm=use_scale_shift_norm, tempspatial_aware=tempspatial_aware, + use_temporal_conv=temporal_conv, + device=device, + dtype=self.dtype + ), + SpatialTransformer(ch, num_heads, dim_head, + depth=transformer_depth, context_dim=context_dim, use_linear=use_linear, + use_checkpoint=use_checkpoint, disable_self_attn=False, video_length=temporal_length, + image_cross_attention=self.image_cross_attention,image_cross_attention_scale_learnable=self.image_cross_attention_scale_learnable, + device=device, + dtype=self.dtype + ) + ] + if self.temporal_attention: + layers.append( + TemporalTransformer(ch, num_heads, dim_head, + depth=transformer_depth, context_dim=context_dim, use_linear=use_linear, + use_checkpoint=use_checkpoint, only_self_att=temporal_self_att_only, + causal_attention=use_causal_attention, relative_position=use_relative_position, + temporal_length=temporal_length, + device=device, + dtype=self.dtype + ) + ) + layers.append( + ResBlock(ch, time_embed_dim, dropout, + dims=dims, use_checkpoint=use_checkpoint, + use_scale_shift_norm=use_scale_shift_norm, tempspatial_aware=tempspatial_aware, + use_temporal_conv=temporal_conv, + device=device, + dtype=self.dtype + ) + ) + + ## Middle Block + self.middle_block = TimestepEmbedSequential(*layers) + + ## Output Block + 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=mult * model_channels, dims=dims, use_checkpoint=use_checkpoint, + use_scale_shift_norm=use_scale_shift_norm, tempspatial_aware=tempspatial_aware, + use_temporal_conv=temporal_conv, + device=device, + dtype=self.dtype + ) + ] + 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 + layers.append( + SpatialTransformer(ch, num_heads, dim_head, + depth=transformer_depth, context_dim=context_dim, use_linear=use_linear, + use_checkpoint=use_checkpoint, disable_self_attn=False, video_length=temporal_length, + image_cross_attention=self.image_cross_attention,image_cross_attention_scale_learnable=self.image_cross_attention_scale_learnable, + device=device, + dtype=self.dtype + ) + ) + if self.temporal_attention: + layers.append( + TemporalTransformer(ch, num_heads, dim_head, + depth=transformer_depth, context_dim=context_dim, use_linear=use_linear, + use_checkpoint=use_checkpoint, only_self_att=temporal_self_att_only, + causal_attention=use_causal_attention, relative_position=use_relative_position, + temporal_length=temporal_length, + device=device, + dtype=self.dtype + ) + ) + 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, + device=device, + dtype=self.dtype + ) + if resblock_updown + else Upsample(ch, conv_resample, dims=dims, out_channels=out_ch) + ) + ds //= 2 + self.output_blocks.append(TimestepEmbedSequential(*layers)) + + self.out = nn.Sequential( + normalization(ch, device=device, dtype=self.dtype), + nn.SiLU(), + zero_module( + operations.conv_nd( + dims, + model_channels, + out_channels, + 3, + padding=1, + device=device, + dtype=self.dtype + ) + ), + ) + + # TODO Add Transformer options to leverage the usage of patches. + def forward( + self, + x, + timesteps, + context=None, + context_in=None, + cc_concat=None, + num_video_frames=16, + features_adapter=None, + fs=None, + img_emb=None, + control=None, + transformer_options={}, + cond_idx=None, + **kwargs + ): + + if any([fs is None, img_emb is None, cc_concat is None]): + raise ValueError("One or more of the required inputs for UNet Forward is None.") + + cond_idx = transformer_options.get("cond_idx", None) + transformer_options['original_shape'] = list(x.shape) + transformer_options['transformer_index'] = 0 + transformer_patches = transformer_options.get("patches", {}) + + # In ComfyUI, the frames are always with the batch, so we deconstruct it here. + # This is mandatory as this is a video based model. + # We usually denote "f" as frames, but will use "t" (time) to be consistent with DynamiCrafter. + b,_,t,_,_ = x.shape + + context = context_in + cc_concat = cc_concat.to(x.device, x.dtype) + x = torch.cat([x, cc_concat], dim=1) + + fs = fs.to(x.device, x.dtype) + + timestep = timesteps + context = context_processor(context, num_video_frames, img_emb=img_emb) + + t_emb = timestep_embedding(timestep, self.model_channels, repeat_only=False, dtype=self.dtype) + emb = self.time_embed(t_emb) + emb = emb.repeat_interleave(repeats=t, dim=0) + + ## always in shape (b t) c h w, except for temporal layer + x = rearrange(x, 'b c t h w -> (b t) c h w') + + ## combine emb + if self.fs_condition: + if fs is None: + fs = torch.tensor( + [self.default_fs] * b, dtype=torch.long, device=x.device) + fs_emb = timestep_embedding(fs, self.model_channels, repeat_only=False, dtype=self.dtype).type(x.dtype) + + fs_embed = self.fps_embedding(fs_emb) + fs_embed = fs_embed.repeat_interleave(repeats=t, dim=0) + + emb = emb + fs_embed + + h = x.type(self.dtype) + adapter_idx = 0 + hs = [] + + for id, module in enumerate(self.input_blocks): + transformer_options["block"] = ("input", id) + #h = module(h, emb, context=context, batch_size=b) + h = forward_timestep_embed( + module, + h, + emb, + context=context, + batch_size=b, + transformer_options=transformer_options + ) + h = apply_control(h, control, 'input', cond_idx) + + if "input_block_patch" in transformer_patches: + patch = transformer_patches["input_block_patch"] + for p in patch: + h = p(h, transformer_options) + + if id ==0 and self.addition_attention: + h = forward_timestep_embed( + self.init_attn, + h, + emb, + context=context, + batch_size=b, + transformer_options=transformer_options + ) + ## 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 "input_block_patch_after_skip" in transformer_patches: + patch = transformer_patches["input_block_patch_after_skip"] + for p in patch: + h = p(h, transformer_options) + if features_adapter is not None: + assert len(features_adapter)==adapter_idx, 'Wrong features_adapter' + transformer_options["block"] = ("middle", 0) + h = forward_timestep_embed( + self.middle_block, + h, + emb, + context=context, + batch_size=b, + transformer_options=transformer_options + ) + h = apply_control(h, control, 'middle', cond_idx) + for id, module in enumerate(self.output_blocks): + transformer_options["block"] = ("output", id) + hsp = hs.pop() + hsp = apply_control(hsp, control, 'output', cond_idx) + + if "output_block_patch" in transformer_patches: + patch = transformer_patches["output_block_patch"] + for p in patch: + h, hsp = p(h, hsp, transformer_options) + + h = torch.cat([h, hsp], dim=1) + del hsp + h = forward_timestep_embed( + module, + h, + emb, + context=context, + batch_size=b, + transformer_options=transformer_options + ) + h = h.type(x.dtype) + h = self.out(h) + + # We output with the tensor unfolded framewise, then reshape them to batched using ComfyUI nodes. + h = rearrange(h, '(b t) c h w -> b c t h w', t=num_video_frames) + + return h \ No newline at end of file diff --git a/modules/x_transformer.py b/modules/x_transformer.py new file mode 100644 index 0000000..5321012 --- /dev/null +++ b/modules/x_transformer.py @@ -0,0 +1,639 @@ +"""shout-out to https://github.com/lucidrains/x-transformers/tree/main/x_transformers""" +from functools import partial +from inspect import isfunction +from collections import namedtuple +from einops import rearrange, repeat +import torch +from torch import nn, einsum +import torch.nn.functional as F + +# constants +DEFAULT_DIM_HEAD = 64 + +Intermediates = namedtuple('Intermediates', [ + 'pre_softmax_attn', + 'post_softmax_attn' +]) + +LayerIntermediates = namedtuple('Intermediates', [ + 'hiddens', + 'attn_intermediates' +]) + + +class AbsolutePositionalEmbedding(nn.Module): + def __init__(self, dim, max_seq_len): + super().__init__() + self.emb = nn.Embedding(max_seq_len, dim) + self.init_() + + def init_(self): + nn.init.normal_(self.emb.weight, std=0.02) + + def forward(self, x): + n = torch.arange(x.shape[1], device=x.device) + return self.emb(n)[None, :, :] + + +class FixedPositionalEmbedding(nn.Module): + def __init__(self, dim): + super().__init__() + inv_freq = 1. / (10000 ** (torch.arange(0, dim, 2).float() / dim)) + self.register_buffer('inv_freq', inv_freq) + + def forward(self, x, seq_dim=1, offset=0): + t = torch.arange(x.shape[seq_dim], device=x.device).type_as(self.inv_freq) + offset + sinusoid_inp = torch.einsum('i , j -> i j', t, self.inv_freq) + emb = torch.cat((sinusoid_inp.sin(), sinusoid_inp.cos()), dim=-1) + return emb[None, :, :] + + +# helpers + +def exists(val): + return val is not None + + +def default(val, d): + if exists(val): + return val + return d() if isfunction(d) else d + + +def always(val): + def inner(*args, **kwargs): + return val + return inner + + +def not_equals(val): + def inner(x): + return x != val + return inner + + +def equals(val): + def inner(x): + return x == val + return inner + + +def max_neg_value(tensor): + return -torch.finfo(tensor.dtype).max + + +# keyword argument helpers + +def pick_and_pop(keys, d): + values = list(map(lambda key: d.pop(key), keys)) + return dict(zip(keys, values)) + + +def group_dict_by_key(cond, d): + return_val = [dict(), dict()] + for key in d.keys(): + match = bool(cond(key)) + ind = int(not match) + return_val[ind][key] = d[key] + return (*return_val,) + + +def string_begins_with(prefix, str): + return str.startswith(prefix) + + +def group_by_key_prefix(prefix, d): + return group_dict_by_key(partial(string_begins_with, prefix), d) + + +def groupby_prefix_and_trim(prefix, d): + kwargs_with_prefix, kwargs = group_dict_by_key(partial(string_begins_with, prefix), d) + kwargs_without_prefix = dict(map(lambda x: (x[0][len(prefix):], x[1]), tuple(kwargs_with_prefix.items()))) + return kwargs_without_prefix, kwargs + + +# classes +class Scale(nn.Module): + def __init__(self, value, fn): + super().__init__() + self.value = value + self.fn = fn + + def forward(self, x, **kwargs): + x, *rest = self.fn(x, **kwargs) + return (x * self.value, *rest) + + +class Rezero(nn.Module): + def __init__(self, fn): + super().__init__() + self.fn = fn + self.g = nn.Parameter(torch.zeros(1)) + + def forward(self, x, **kwargs): + x, *rest = self.fn(x, **kwargs) + return (x * self.g, *rest) + + +class ScaleNorm(nn.Module): + def __init__(self, dim, eps=1e-5): + super().__init__() + self.scale = dim ** -0.5 + self.eps = eps + self.g = nn.Parameter(torch.ones(1)) + + def forward(self, x): + norm = torch.norm(x, dim=-1, keepdim=True) * self.scale + return x / norm.clamp(min=self.eps) * self.g + + +class RMSNorm(nn.Module): + def __init__(self, dim, eps=1e-8): + super().__init__() + self.scale = dim ** -0.5 + self.eps = eps + self.g = nn.Parameter(torch.ones(dim)) + + def forward(self, x): + norm = torch.norm(x, dim=-1, keepdim=True) * self.scale + return x / norm.clamp(min=self.eps) * self.g + + +class Residual(nn.Module): + def forward(self, x, residual): + return x + residual + + +class GRUGating(nn.Module): + def __init__(self, dim): + super().__init__() + self.gru = nn.GRUCell(dim, dim) + + def forward(self, x, residual): + gated_output = self.gru( + rearrange(x, 'b n d -> (b n) d'), + rearrange(residual, 'b n d -> (b n) d') + ) + + return gated_output.reshape_as(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) + + +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) + + +# attention. +class Attention(nn.Module): + def __init__( + self, + dim, + dim_head=DEFAULT_DIM_HEAD, + heads=8, + causal=False, + mask=None, + talking_heads=False, + sparse_topk=None, + use_entmax15=False, + num_mem_kv=0, + dropout=0., + on_attn=False + ): + super().__init__() + if use_entmax15: + raise NotImplementedError("Check out entmax activation instead of softmax activation!") + self.scale = dim_head ** -0.5 + self.heads = heads + self.causal = causal + self.mask = mask + + inner_dim = dim_head * heads + + 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.dropout = nn.Dropout(dropout) + + # talking heads + self.talking_heads = talking_heads + if talking_heads: + self.pre_softmax_proj = nn.Parameter(torch.randn(heads, heads)) + self.post_softmax_proj = nn.Parameter(torch.randn(heads, heads)) + + # explicit topk sparse attention + self.sparse_topk = sparse_topk + + # entmax + #self.attn_fn = entmax15 if use_entmax15 else F.softmax + self.attn_fn = F.softmax + + # add memory key / values + self.num_mem_kv = num_mem_kv + if num_mem_kv > 0: + self.mem_k = nn.Parameter(torch.randn(heads, num_mem_kv, dim_head)) + self.mem_v = nn.Parameter(torch.randn(heads, num_mem_kv, dim_head)) + + # attention on attention + self.attn_on_attn = on_attn + self.to_out = nn.Sequential(nn.Linear(inner_dim, dim * 2), nn.GLU()) if on_attn else nn.Linear(inner_dim, dim) + + def forward( + self, + x, + context=None, + mask=None, + context_mask=None, + rel_pos=None, + sinusoidal_emb=None, + prev_attn=None, + mem=None + ): + b, n, _, h, talking_heads, device = *x.shape, self.heads, self.talking_heads, x.device + kv_input = default(context, x) + + q_input = x + k_input = kv_input + v_input = kv_input + + if exists(mem): + k_input = torch.cat((mem, k_input), dim=-2) + v_input = torch.cat((mem, v_input), dim=-2) + + if exists(sinusoidal_emb): + # in shortformer, the query would start at a position offset depending on the past cached memory + offset = k_input.shape[-2] - q_input.shape[-2] + q_input = q_input + sinusoidal_emb(q_input, offset=offset) + k_input = k_input + sinusoidal_emb(k_input) + + q = self.to_q(q_input) + k = self.to_k(k_input) + v = self.to_v(v_input) + + q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> b h n d', h=h), (q, k, v)) + + input_mask = None + if any(map(exists, (mask, context_mask))): + q_mask = default(mask, lambda: torch.ones((b, n), device=device).bool()) + k_mask = q_mask if not exists(context) else context_mask + k_mask = default(k_mask, lambda: torch.ones((b, k.shape[-2]), device=device).bool()) + q_mask = rearrange(q_mask, 'b i -> b () i ()') + k_mask = rearrange(k_mask, 'b j -> b () () j') + input_mask = q_mask * k_mask + + if self.num_mem_kv > 0: + mem_k, mem_v = map(lambda t: repeat(t, 'h n d -> b h n d', b=b), (self.mem_k, self.mem_v)) + k = torch.cat((mem_k, k), dim=-2) + v = torch.cat((mem_v, v), dim=-2) + if exists(input_mask): + input_mask = F.pad(input_mask, (self.num_mem_kv, 0), value=True) + + dots = einsum('b h i d, b h j d -> b h i j', q, k) * self.scale + mask_value = max_neg_value(dots) + + if exists(prev_attn): + dots = dots + prev_attn + + pre_softmax_attn = dots + + if talking_heads: + dots = einsum('b h i j, h k -> b k i j', dots, self.pre_softmax_proj).contiguous() + + if exists(rel_pos): + dots = rel_pos(dots) + + if exists(input_mask): + dots.masked_fill_(~input_mask, mask_value) + del input_mask + + if self.causal: + i, j = dots.shape[-2:] + r = torch.arange(i, device=device) + mask = rearrange(r, 'i -> () () i ()') < rearrange(r, 'j -> () () () j') + mask = F.pad(mask, (j - i, 0), value=False) + dots.masked_fill_(mask, mask_value) + del mask + + if exists(self.sparse_topk) and self.sparse_topk < dots.shape[-1]: + top, _ = dots.topk(self.sparse_topk, dim=-1) + vk = top[..., -1].unsqueeze(-1).expand_as(dots) + mask = dots < vk + dots.masked_fill_(mask, mask_value) + del mask + + attn = self.attn_fn(dots, dim=-1) + post_softmax_attn = attn + + attn = self.dropout(attn) + + if talking_heads: + attn = einsum('b h i j, h k -> b k i j', attn, self.post_softmax_proj).contiguous() + + 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)') + + intermediates = Intermediates( + pre_softmax_attn=pre_softmax_attn, + post_softmax_attn=post_softmax_attn + ) + + return self.to_out(out), intermediates + + +class AttentionLayers(nn.Module): + def __init__( + self, + dim, + depth, + heads=8, + causal=False, + cross_attend=False, + only_cross=False, + use_scalenorm=False, + use_rmsnorm=False, + use_rezero=False, + rel_pos_num_buckets=32, + rel_pos_max_distance=128, + position_infused_attn=False, + custom_layers=None, + sandwich_coef=None, + par_ratio=None, + residual_attn=False, + cross_residual_attn=False, + macaron=False, + pre_norm=True, + gate_residual=False, + **kwargs + ): + super().__init__() + ff_kwargs, kwargs = groupby_prefix_and_trim('ff_', kwargs) + attn_kwargs, _ = groupby_prefix_and_trim('attn_', kwargs) + + dim_head = attn_kwargs.get('dim_head', DEFAULT_DIM_HEAD) + + self.dim = dim + self.depth = depth + self.layers = nn.ModuleList([]) + + self.has_pos_emb = position_infused_attn + self.pia_pos_emb = FixedPositionalEmbedding(dim) if position_infused_attn else None + self.rotary_pos_emb = always(None) + + assert rel_pos_num_buckets <= rel_pos_max_distance, 'number of relative position buckets must be less than the relative position max distance' + self.rel_pos = None + + self.pre_norm = pre_norm + + self.residual_attn = residual_attn + self.cross_residual_attn = cross_residual_attn + + norm_class = ScaleNorm if use_scalenorm else nn.LayerNorm + norm_class = RMSNorm if use_rmsnorm else norm_class + norm_fn = partial(norm_class, dim) + + norm_fn = nn.Identity if use_rezero else norm_fn + branch_fn = Rezero if use_rezero else None + + if cross_attend and not only_cross: + default_block = ('a', 'c', 'f') + elif cross_attend and only_cross: + default_block = ('c', 'f') + else: + default_block = ('a', 'f') + + if macaron: + default_block = ('f',) + default_block + + if exists(custom_layers): + layer_types = custom_layers + elif exists(par_ratio): + par_depth = depth * len(default_block) + assert 1 < par_ratio <= par_depth, 'par ratio out of range' + default_block = tuple(filter(not_equals('f'), default_block)) + par_attn = par_depth // par_ratio + depth_cut = par_depth * 2 // 3 # 2 / 3 attention layer cutoff suggested by PAR paper + par_width = (depth_cut + depth_cut // par_attn) // par_attn + assert len(default_block) <= par_width, 'default block is too large for par_ratio' + par_block = default_block + ('f',) * (par_width - len(default_block)) + par_head = par_block * par_attn + layer_types = par_head + ('f',) * (par_depth - len(par_head)) + elif exists(sandwich_coef): + assert sandwich_coef > 0 and sandwich_coef <= depth, 'sandwich coefficient should be less than the depth' + layer_types = ('a',) * sandwich_coef + default_block * (depth - sandwich_coef) + ('f',) * sandwich_coef + else: + layer_types = default_block * depth + + self.layer_types = layer_types + self.num_attn_layers = len(list(filter(equals('a'), layer_types))) + + for layer_type in self.layer_types: + if layer_type == 'a': + layer = Attention(dim, heads=heads, causal=causal, **attn_kwargs) + elif layer_type == 'c': + layer = Attention(dim, heads=heads, **attn_kwargs) + elif layer_type == 'f': + layer = FeedForward(dim, **ff_kwargs) + layer = layer if not macaron else Scale(0.5, layer) + else: + raise Exception(f'invalid layer type {layer_type}') + + if isinstance(layer, Attention) and exists(branch_fn): + layer = branch_fn(layer) + + if gate_residual: + residual_fn = GRUGating(dim) + else: + residual_fn = Residual() + + self.layers.append(nn.ModuleList([ + norm_fn(), + layer, + residual_fn + ])) + + def forward( + self, + x, + context=None, + mask=None, + context_mask=None, + mems=None, + return_hiddens=False + ): + hiddens = [] + intermediates = [] + prev_attn = None + prev_cross_attn = None + + mems = mems.copy() if exists(mems) else [None] * self.num_attn_layers + + for ind, (layer_type, (norm, block, residual_fn)) in enumerate(zip(self.layer_types, self.layers)): + is_last = ind == (len(self.layers) - 1) + + if layer_type == 'a': + hiddens.append(x) + layer_mem = mems.pop(0) + + residual = x + + if self.pre_norm: + x = norm(x) + + if layer_type == 'a': + out, inter = block(x, mask=mask, sinusoidal_emb=self.pia_pos_emb, rel_pos=self.rel_pos, + prev_attn=prev_attn, mem=layer_mem) + elif layer_type == 'c': + out, inter = block(x, context=context, mask=mask, context_mask=context_mask, prev_attn=prev_cross_attn) + elif layer_type == 'f': + out = block(x) + + x = residual_fn(out, residual) + + if layer_type in ('a', 'c'): + intermediates.append(inter) + + if layer_type == 'a' and self.residual_attn: + prev_attn = inter.pre_softmax_attn + elif layer_type == 'c' and self.cross_residual_attn: + prev_cross_attn = inter.pre_softmax_attn + + if not self.pre_norm and not is_last: + x = norm(x) + + if return_hiddens: + intermediates = LayerIntermediates( + hiddens=hiddens, + attn_intermediates=intermediates + ) + + return x, intermediates + + return x + + +class Encoder(AttentionLayers): + def __init__(self, **kwargs): + assert 'causal' not in kwargs, 'cannot set causality on encoder' + super().__init__(causal=False, **kwargs) + + + +class TransformerWrapper(nn.Module): + def __init__( + self, + *, + num_tokens, + max_seq_len, + attn_layers, + emb_dim=None, + max_mem_len=0., + emb_dropout=0., + num_memory_tokens=None, + tie_embedding=False, + use_pos_emb=True + ): + super().__init__() + assert isinstance(attn_layers, AttentionLayers), 'attention layers must be one of Encoder or Decoder' + + dim = attn_layers.dim + emb_dim = default(emb_dim, dim) + + self.max_seq_len = max_seq_len + self.max_mem_len = max_mem_len + self.num_tokens = num_tokens + + self.token_emb = nn.Embedding(num_tokens, emb_dim) + self.pos_emb = AbsolutePositionalEmbedding(emb_dim, max_seq_len) if ( + use_pos_emb and not attn_layers.has_pos_emb) else always(0) + self.emb_dropout = nn.Dropout(emb_dropout) + + self.project_emb = nn.Linear(emb_dim, dim) if emb_dim != dim else nn.Identity() + self.attn_layers = attn_layers + self.norm = nn.LayerNorm(dim) + + self.init_() + + self.to_logits = nn.Linear(dim, num_tokens) if not tie_embedding else lambda t: t @ self.token_emb.weight.t() + + # memory tokens (like [cls]) from Memory Transformers paper + num_memory_tokens = default(num_memory_tokens, 0) + self.num_memory_tokens = num_memory_tokens + if num_memory_tokens > 0: + self.memory_tokens = nn.Parameter(torch.randn(num_memory_tokens, dim)) + + # let funnel encoder know number of memory tokens, if specified + if hasattr(attn_layers, 'num_memory_tokens'): + attn_layers.num_memory_tokens = num_memory_tokens + + def init_(self): + nn.init.normal_(self.token_emb.weight, std=0.02) + + def forward( + self, + x, + return_embeddings=False, + mask=None, + return_mems=False, + return_attn=False, + mems=None, + **kwargs + ): + b, n, device, num_mem = *x.shape, x.device, self.num_memory_tokens + x = self.token_emb(x) + x += self.pos_emb(x) + x = self.emb_dropout(x) + + x = self.project_emb(x) + + if num_mem > 0: + mem = repeat(self.memory_tokens, 'n d -> b n d', b=b) + x = torch.cat((mem, x), dim=1) + + # auto-handle masking after appending memory tokens + if exists(mask): + mask = F.pad(mask, (num_mem, 0), value=True) + + x, intermediates = self.attn_layers(x, mask=mask, mems=mems, return_hiddens=True, **kwargs) + x = self.norm(x) + + mem, x = x[:, :num_mem], x[:, num_mem:] + + out = self.to_logits(x) if not return_embeddings else x + + if return_mems: + hiddens = intermediates.hiddens + new_mems = list(map(lambda pair: torch.cat(pair, dim=-2), zip(mems, hiddens))) if exists(mems) else hiddens + new_mems = list(map(lambda t: t[..., -self.max_mem_len:, :].detach(), new_mems)) + return out, new_mems + + if return_attn: + attn_maps = list(map(lambda t: t.post_softmax_attn, intermediates.attn_intermediates)) + return out, attn_maps + + return out \ No newline at end of file