2023-04-19 13:36:19 +00:00
|
|
|
import torch
|
2024-01-22 02:51:22 +00:00
|
|
|
from torch import nn
|
2023-05-04 22:07:41 +00:00
|
|
|
from .ldm.modules.attention import CrossAttention
|
2023-04-19 13:36:19 +00:00
|
|
|
from inspect import isfunction
|
2024-02-18 07:20:23 +00:00
|
|
|
import comfy.ops
|
|
|
|
ops = comfy.ops.manual_cast
|
2023-04-19 13:36:19 +00:00
|
|
|
|
|
|
|
def exists(val):
|
|
|
|
return val is not None
|
|
|
|
|
|
|
|
|
|
|
|
def uniq(arr):
|
|
|
|
return{el: True for el in arr}.keys()
|
|
|
|
|
|
|
|
|
|
|
|
def default(val, d):
|
|
|
|
if exists(val):
|
|
|
|
return val
|
|
|
|
return d() if isfunction(d) else d
|
|
|
|
|
|
|
|
|
|
|
|
# feedforward
|
|
|
|
class GEGLU(nn.Module):
|
|
|
|
def __init__(self, dim_in, dim_out):
|
|
|
|
super().__init__()
|
2024-02-18 07:20:23 +00:00
|
|
|
self.proj = ops.Linear(dim_in, dim_out * 2)
|
2023-04-19 13:36:19 +00:00
|
|
|
|
|
|
|
def forward(self, x):
|
|
|
|
x, gate = self.proj(x).chunk(2, dim=-1)
|
|
|
|
return x * torch.nn.functional.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(
|
2024-02-18 07:20:23 +00:00
|
|
|
ops.Linear(dim, inner_dim),
|
2023-04-19 13:36:19 +00:00
|
|
|
nn.GELU()
|
|
|
|
) if not glu else GEGLU(dim, inner_dim)
|
|
|
|
|
|
|
|
self.net = nn.Sequential(
|
|
|
|
project_in,
|
|
|
|
nn.Dropout(dropout),
|
2024-02-18 07:20:23 +00:00
|
|
|
ops.Linear(inner_dim, dim_out)
|
2023-04-19 13:36:19 +00:00
|
|
|
)
|
|
|
|
|
|
|
|
def forward(self, x):
|
|
|
|
return self.net(x)
|
|
|
|
|
|
|
|
|
|
|
|
class GatedCrossAttentionDense(nn.Module):
|
|
|
|
def __init__(self, query_dim, context_dim, n_heads, d_head):
|
|
|
|
super().__init__()
|
|
|
|
|
|
|
|
self.attn = CrossAttention(
|
|
|
|
query_dim=query_dim,
|
|
|
|
context_dim=context_dim,
|
|
|
|
heads=n_heads,
|
2024-02-18 07:20:23 +00:00
|
|
|
dim_head=d_head,
|
|
|
|
operations=ops)
|
2023-04-19 13:36:19 +00:00
|
|
|
self.ff = FeedForward(query_dim, glu=True)
|
|
|
|
|
2024-02-18 07:20:23 +00:00
|
|
|
self.norm1 = ops.LayerNorm(query_dim)
|
|
|
|
self.norm2 = ops.LayerNorm(query_dim)
|
2023-04-19 13:36:19 +00:00
|
|
|
|
|
|
|
self.register_parameter('alpha_attn', nn.Parameter(torch.tensor(0.)))
|
|
|
|
self.register_parameter('alpha_dense', nn.Parameter(torch.tensor(0.)))
|
|
|
|
|
|
|
|
# this can be useful: we can externally change magnitude of tanh(alpha)
|
|
|
|
# for example, when it is set to 0, then the entire model is same as
|
|
|
|
# original one
|
|
|
|
self.scale = 1
|
|
|
|
|
|
|
|
def forward(self, x, objs):
|
|
|
|
|
|
|
|
x = x + self.scale * \
|
|
|
|
torch.tanh(self.alpha_attn) * self.attn(self.norm1(x), objs, objs)
|
|
|
|
x = x + self.scale * \
|
|
|
|
torch.tanh(self.alpha_dense) * self.ff(self.norm2(x))
|
|
|
|
|
|
|
|
return x
|
|
|
|
|
|
|
|
|
|
|
|
class GatedSelfAttentionDense(nn.Module):
|
|
|
|
def __init__(self, query_dim, context_dim, n_heads, d_head):
|
|
|
|
super().__init__()
|
|
|
|
|
|
|
|
# we need a linear projection since we need cat visual feature and obj
|
|
|
|
# feature
|
2024-02-18 07:20:23 +00:00
|
|
|
self.linear = ops.Linear(context_dim, query_dim)
|
2023-04-19 13:36:19 +00:00
|
|
|
|
|
|
|
self.attn = CrossAttention(
|
|
|
|
query_dim=query_dim,
|
|
|
|
context_dim=query_dim,
|
|
|
|
heads=n_heads,
|
2024-02-18 07:20:23 +00:00
|
|
|
dim_head=d_head,
|
|
|
|
operations=ops)
|
2023-04-19 13:36:19 +00:00
|
|
|
self.ff = FeedForward(query_dim, glu=True)
|
|
|
|
|
2024-02-18 07:20:23 +00:00
|
|
|
self.norm1 = ops.LayerNorm(query_dim)
|
|
|
|
self.norm2 = ops.LayerNorm(query_dim)
|
2023-04-19 13:36:19 +00:00
|
|
|
|
|
|
|
self.register_parameter('alpha_attn', nn.Parameter(torch.tensor(0.)))
|
|
|
|
self.register_parameter('alpha_dense', nn.Parameter(torch.tensor(0.)))
|
|
|
|
|
|
|
|
# this can be useful: we can externally change magnitude of tanh(alpha)
|
|
|
|
# for example, when it is set to 0, then the entire model is same as
|
|
|
|
# original one
|
|
|
|
self.scale = 1
|
|
|
|
|
|
|
|
def forward(self, x, objs):
|
|
|
|
|
|
|
|
N_visual = x.shape[1]
|
|
|
|
objs = self.linear(objs)
|
|
|
|
|
|
|
|
x = x + self.scale * torch.tanh(self.alpha_attn) * self.attn(
|
|
|
|
self.norm1(torch.cat([x, objs], dim=1)))[:, 0:N_visual, :]
|
|
|
|
x = x + self.scale * \
|
|
|
|
torch.tanh(self.alpha_dense) * self.ff(self.norm2(x))
|
|
|
|
|
|
|
|
return x
|
|
|
|
|
|
|
|
|
|
|
|
class GatedSelfAttentionDense2(nn.Module):
|
|
|
|
def __init__(self, query_dim, context_dim, n_heads, d_head):
|
|
|
|
super().__init__()
|
|
|
|
|
|
|
|
# we need a linear projection since we need cat visual feature and obj
|
|
|
|
# feature
|
2024-02-18 07:20:23 +00:00
|
|
|
self.linear = ops.Linear(context_dim, query_dim)
|
2023-04-19 13:36:19 +00:00
|
|
|
|
|
|
|
self.attn = CrossAttention(
|
2024-02-18 07:20:23 +00:00
|
|
|
query_dim=query_dim, context_dim=query_dim, dim_head=d_head, operations=ops)
|
2023-04-19 13:36:19 +00:00
|
|
|
self.ff = FeedForward(query_dim, glu=True)
|
|
|
|
|
2024-02-18 07:20:23 +00:00
|
|
|
self.norm1 = ops.LayerNorm(query_dim)
|
|
|
|
self.norm2 = ops.LayerNorm(query_dim)
|
2023-04-19 13:36:19 +00:00
|
|
|
|
|
|
|
self.register_parameter('alpha_attn', nn.Parameter(torch.tensor(0.)))
|
|
|
|
self.register_parameter('alpha_dense', nn.Parameter(torch.tensor(0.)))
|
|
|
|
|
|
|
|
# this can be useful: we can externally change magnitude of tanh(alpha)
|
|
|
|
# for example, when it is set to 0, then the entire model is same as
|
|
|
|
# original one
|
|
|
|
self.scale = 1
|
|
|
|
|
|
|
|
def forward(self, x, objs):
|
|
|
|
|
|
|
|
B, N_visual, _ = x.shape
|
|
|
|
B, N_ground, _ = objs.shape
|
|
|
|
|
|
|
|
objs = self.linear(objs)
|
|
|
|
|
|
|
|
# sanity check
|
|
|
|
size_v = math.sqrt(N_visual)
|
|
|
|
size_g = math.sqrt(N_ground)
|
|
|
|
assert int(size_v) == size_v, "Visual tokens must be square rootable"
|
|
|
|
assert int(size_g) == size_g, "Grounding tokens must be square rootable"
|
|
|
|
size_v = int(size_v)
|
|
|
|
size_g = int(size_g)
|
|
|
|
|
|
|
|
# select grounding token and resize it to visual token size as residual
|
|
|
|
out = self.attn(self.norm1(torch.cat([x, objs], dim=1)))[
|
|
|
|
:, N_visual:, :]
|
|
|
|
out = out.permute(0, 2, 1).reshape(B, -1, size_g, size_g)
|
|
|
|
out = torch.nn.functional.interpolate(
|
|
|
|
out, (size_v, size_v), mode='bicubic')
|
|
|
|
residual = out.reshape(B, -1, N_visual).permute(0, 2, 1)
|
|
|
|
|
|
|
|
# add residual to visual feature
|
|
|
|
x = x + self.scale * torch.tanh(self.alpha_attn) * residual
|
|
|
|
x = x + self.scale * \
|
|
|
|
torch.tanh(self.alpha_dense) * self.ff(self.norm2(x))
|
|
|
|
|
|
|
|
return x
|
|
|
|
|
|
|
|
|
|
|
|
class FourierEmbedder():
|
|
|
|
def __init__(self, num_freqs=64, temperature=100):
|
|
|
|
|
|
|
|
self.num_freqs = num_freqs
|
|
|
|
self.temperature = temperature
|
|
|
|
self.freq_bands = temperature ** (torch.arange(num_freqs) / num_freqs)
|
|
|
|
|
|
|
|
@torch.no_grad()
|
|
|
|
def __call__(self, x, cat_dim=-1):
|
|
|
|
"x: arbitrary shape of tensor. dim: cat dim"
|
|
|
|
out = []
|
|
|
|
for freq in self.freq_bands:
|
|
|
|
out.append(torch.sin(freq * x))
|
|
|
|
out.append(torch.cos(freq * x))
|
|
|
|
return torch.cat(out, cat_dim)
|
|
|
|
|
|
|
|
|
|
|
|
class PositionNet(nn.Module):
|
|
|
|
def __init__(self, in_dim, out_dim, fourier_freqs=8):
|
|
|
|
super().__init__()
|
|
|
|
self.in_dim = in_dim
|
|
|
|
self.out_dim = out_dim
|
|
|
|
|
|
|
|
self.fourier_embedder = FourierEmbedder(num_freqs=fourier_freqs)
|
|
|
|
self.position_dim = fourier_freqs * 2 * 4 # 2 is sin&cos, 4 is xyxy
|
|
|
|
|
|
|
|
self.linears = nn.Sequential(
|
2024-02-18 07:20:23 +00:00
|
|
|
ops.Linear(self.in_dim + self.position_dim, 512),
|
2023-04-19 13:36:19 +00:00
|
|
|
nn.SiLU(),
|
2024-02-18 07:20:23 +00:00
|
|
|
ops.Linear(512, 512),
|
2023-04-19 13:36:19 +00:00
|
|
|
nn.SiLU(),
|
2024-02-18 07:20:23 +00:00
|
|
|
ops.Linear(512, out_dim),
|
2023-04-19 13:36:19 +00:00
|
|
|
)
|
|
|
|
|
|
|
|
self.null_positive_feature = torch.nn.Parameter(
|
|
|
|
torch.zeros([self.in_dim]))
|
|
|
|
self.null_position_feature = torch.nn.Parameter(
|
|
|
|
torch.zeros([self.position_dim]))
|
|
|
|
|
|
|
|
def forward(self, boxes, masks, positive_embeddings):
|
|
|
|
B, N, _ = boxes.shape
|
2024-02-18 07:20:23 +00:00
|
|
|
masks = masks.unsqueeze(-1)
|
|
|
|
positive_embeddings = positive_embeddings
|
2023-04-19 13:36:19 +00:00
|
|
|
|
|
|
|
# embedding position (it may includes padding as placeholder)
|
2024-02-18 07:20:23 +00:00
|
|
|
xyxy_embedding = self.fourier_embedder(boxes) # B*N*4 --> B*N*C
|
2023-04-19 13:36:19 +00:00
|
|
|
|
|
|
|
# learnable null embedding
|
2024-02-18 07:20:23 +00:00
|
|
|
positive_null = self.null_positive_feature.to(device=boxes.device, dtype=boxes.dtype).view(1, 1, -1)
|
|
|
|
xyxy_null = self.null_position_feature.to(device=boxes.device, dtype=boxes.dtype).view(1, 1, -1)
|
2023-04-19 13:36:19 +00:00
|
|
|
|
|
|
|
# replace padding with learnable null embedding
|
|
|
|
positive_embeddings = positive_embeddings * \
|
|
|
|
masks + (1 - masks) * positive_null
|
|
|
|
xyxy_embedding = xyxy_embedding * masks + (1 - masks) * xyxy_null
|
|
|
|
|
|
|
|
objs = self.linears(
|
|
|
|
torch.cat([positive_embeddings, xyxy_embedding], dim=-1))
|
|
|
|
assert objs.shape == torch.Size([B, N, self.out_dim])
|
|
|
|
return objs
|
|
|
|
|
|
|
|
|
|
|
|
class Gligen(nn.Module):
|
|
|
|
def __init__(self, modules, position_net, key_dim):
|
|
|
|
super().__init__()
|
|
|
|
self.module_list = nn.ModuleList(modules)
|
|
|
|
self.position_net = position_net
|
|
|
|
self.key_dim = key_dim
|
|
|
|
self.max_objs = 30
|
2023-08-17 05:06:34 +00:00
|
|
|
self.current_device = torch.device("cpu")
|
2023-04-19 13:36:19 +00:00
|
|
|
|
|
|
|
def _set_position(self, boxes, masks, positive_embeddings):
|
|
|
|
objs = self.position_net(boxes, masks, positive_embeddings)
|
2023-08-17 05:06:34 +00:00
|
|
|
def func(x, extra_options):
|
|
|
|
key = extra_options["transformer_index"]
|
|
|
|
module = self.module_list[key]
|
2024-02-18 07:20:23 +00:00
|
|
|
return module(x, objs.to(device=x.device, dtype=x.dtype))
|
2023-08-17 05:06:34 +00:00
|
|
|
return func
|
2023-04-19 13:36:19 +00:00
|
|
|
|
|
|
|
def set_position(self, latent_image_shape, position_params, device):
|
|
|
|
batch, c, h, w = latent_image_shape
|
|
|
|
masks = torch.zeros([self.max_objs], device="cpu")
|
|
|
|
boxes = []
|
|
|
|
positive_embeddings = []
|
|
|
|
for p in position_params:
|
|
|
|
x1 = (p[4]) / w
|
|
|
|
y1 = (p[3]) / h
|
|
|
|
x2 = (p[4] + p[2]) / w
|
|
|
|
y2 = (p[3] + p[1]) / h
|
|
|
|
masks[len(boxes)] = 1.0
|
|
|
|
boxes += [torch.tensor((x1, y1, x2, y2)).unsqueeze(0)]
|
|
|
|
positive_embeddings += [p[0]]
|
|
|
|
append_boxes = []
|
|
|
|
append_conds = []
|
|
|
|
if len(boxes) < self.max_objs:
|
|
|
|
append_boxes = [torch.zeros(
|
|
|
|
[self.max_objs - len(boxes), 4], device="cpu")]
|
|
|
|
append_conds = [torch.zeros(
|
|
|
|
[self.max_objs - len(boxes), self.key_dim], device="cpu")]
|
|
|
|
|
|
|
|
box_out = torch.cat(
|
|
|
|
boxes + append_boxes).unsqueeze(0).repeat(batch, 1, 1)
|
|
|
|
masks = masks.unsqueeze(0).repeat(batch, 1)
|
|
|
|
conds = torch.cat(positive_embeddings +
|
|
|
|
append_conds).unsqueeze(0).repeat(batch, 1, 1)
|
|
|
|
return self._set_position(
|
|
|
|
box_out.to(device),
|
|
|
|
masks.to(device),
|
|
|
|
conds.to(device))
|
|
|
|
|
|
|
|
def set_empty(self, latent_image_shape, device):
|
|
|
|
batch, c, h, w = latent_image_shape
|
|
|
|
masks = torch.zeros([self.max_objs], device="cpu").repeat(batch, 1)
|
|
|
|
box_out = torch.zeros([self.max_objs, 4],
|
|
|
|
device="cpu").repeat(batch, 1, 1)
|
|
|
|
conds = torch.zeros([self.max_objs, self.key_dim],
|
|
|
|
device="cpu").repeat(batch, 1, 1)
|
|
|
|
return self._set_position(
|
|
|
|
box_out.to(device),
|
|
|
|
masks.to(device),
|
|
|
|
conds.to(device))
|
|
|
|
|
|
|
|
|
|
|
|
def load_gligen(sd):
|
|
|
|
sd_k = sd.keys()
|
|
|
|
output_list = []
|
|
|
|
key_dim = 768
|
|
|
|
for a in ["input_blocks", "middle_block", "output_blocks"]:
|
|
|
|
for b in range(20):
|
|
|
|
k_temp = filter(lambda k: "{}.{}.".format(a, b)
|
|
|
|
in k and ".fuser." in k, sd_k)
|
|
|
|
k_temp = map(lambda k: (k, k.split(".fuser.")[-1]), k_temp)
|
|
|
|
|
|
|
|
n_sd = {}
|
|
|
|
for k in k_temp:
|
|
|
|
n_sd[k[1]] = sd[k[0]]
|
|
|
|
if len(n_sd) > 0:
|
|
|
|
query_dim = n_sd["linear.weight"].shape[0]
|
|
|
|
key_dim = n_sd["linear.weight"].shape[1]
|
|
|
|
|
|
|
|
if key_dim == 768: # SD1.x
|
|
|
|
n_heads = 8
|
|
|
|
d_head = query_dim // n_heads
|
|
|
|
else:
|
|
|
|
d_head = 64
|
|
|
|
n_heads = query_dim // d_head
|
|
|
|
|
|
|
|
gated = GatedSelfAttentionDense(
|
|
|
|
query_dim, key_dim, n_heads, d_head)
|
|
|
|
gated.load_state_dict(n_sd, strict=False)
|
|
|
|
output_list.append(gated)
|
|
|
|
|
|
|
|
if "position_net.null_positive_feature" in sd_k:
|
|
|
|
in_dim = sd["position_net.null_positive_feature"].shape[0]
|
|
|
|
out_dim = sd["position_net.linears.4.weight"].shape[0]
|
|
|
|
|
|
|
|
class WeightsLoader(torch.nn.Module):
|
|
|
|
pass
|
|
|
|
w = WeightsLoader()
|
|
|
|
w.position_net = PositionNet(in_dim, out_dim)
|
|
|
|
w.load_state_dict(sd, strict=False)
|
|
|
|
|
|
|
|
gligen = Gligen(output_list, w.position_net, key_dim)
|
|
|
|
return gligen
|