PixArt sigma
Mentioned in #19 Also removes old example sampler since updating it would be too much overhead. Pytorch attention is still wonky.
This commit is contained in:
City
+45
-92
@@ -18,31 +18,31 @@ from timm.models.vision_transformer import PatchEmbed, Mlp
|
||||
|
||||
|
||||
from .utils import auto_grad_checkpoint, to_2tuple
|
||||
from .PixArt_blocks import t2i_modulate, CaptionEmbedder, WindowAttention, MultiHeadCrossAttention, T2IFinalLayer, TimestepEmbedder, LabelEmbedder, FinalLayer
|
||||
from .PixArt_blocks import t2i_modulate, CaptionEmbedder, AttentionKVCompress, MultiHeadCrossAttention, T2IFinalLayer, TimestepEmbedder, LabelEmbedder, FinalLayer
|
||||
|
||||
|
||||
class PixArtBlock(nn.Module):
|
||||
"""
|
||||
A PixArt block with adaptive layer norm (adaLN-single) conditioning.
|
||||
"""
|
||||
|
||||
def __init__(self, hidden_size, num_heads, mlp_ratio=4.0, drop_path=0., window_size=0, input_size=None, use_rel_pos=False, **block_kwargs):
|
||||
def __init__(self, hidden_size, num_heads, mlp_ratio=4.0, drop_path=0, input_size=None, sampling=None, sr_ratio=1, qk_norm=False, **block_kwargs):
|
||||
super().__init__()
|
||||
self.hidden_size = hidden_size
|
||||
self.norm1 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
|
||||
self.attn = WindowAttention(hidden_size, num_heads=num_heads, qkv_bias=True,
|
||||
input_size=input_size if window_size == 0 else (window_size, window_size),
|
||||
use_rel_pos=use_rel_pos, **block_kwargs)
|
||||
self.attn = AttentionKVCompress(
|
||||
hidden_size, num_heads=num_heads, qkv_bias=True, sampling=sampling, sr_ratio=sr_ratio,
|
||||
qk_norm=qk_norm, **block_kwargs
|
||||
)
|
||||
self.cross_attn = MultiHeadCrossAttention(hidden_size, num_heads, **block_kwargs)
|
||||
self.norm2 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
|
||||
# to be compatible with lower version pytorch
|
||||
approx_gelu = lambda: nn.GELU(approximate="tanh")
|
||||
self.mlp = Mlp(in_features=hidden_size, hidden_features=int(hidden_size * mlp_ratio), act_layer=approx_gelu, drop=0)
|
||||
self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
|
||||
self.window_size = window_size
|
||||
self.scale_shift_table = nn.Parameter(torch.randn(6, hidden_size) / hidden_size ** 0.5)
|
||||
self.sampling = sampling
|
||||
self.sr_ratio = sr_ratio
|
||||
|
||||
def forward(self, x, y, t, mask=None):
|
||||
def forward(self, x, y, t, mask=None, **kwargs):
|
||||
B, N, C = x.shape
|
||||
|
||||
shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = (self.scale_shift_table[None] + t.reshape(B, 6, -1)).chunk(6, dim=1)
|
||||
@@ -53,14 +53,11 @@ class PixArtBlock(nn.Module):
|
||||
return x
|
||||
|
||||
|
||||
#############################################################################
|
||||
# Core PixArt Model #
|
||||
#################################################################################
|
||||
### Core PixArt Model ###
|
||||
class PixArt(nn.Module):
|
||||
"""
|
||||
Diffusion model with a Transformer backbone.
|
||||
"""
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
input_size=32,
|
||||
@@ -73,12 +70,12 @@ class PixArt(nn.Module):
|
||||
class_dropout_prob=0.1,
|
||||
pred_sigma=True,
|
||||
drop_path: float = 0.,
|
||||
window_size=0,
|
||||
window_block_indexes=[],
|
||||
use_rel_pos=False,
|
||||
caption_channels=4096,
|
||||
lewei_scale=1.0,
|
||||
pe_interpolation=1.0,
|
||||
config=None,
|
||||
model_max_length=120,
|
||||
qk_norm=False,
|
||||
kv_compress_config=None,
|
||||
**kwargs,
|
||||
):
|
||||
super().__init__()
|
||||
@@ -87,8 +84,8 @@ class PixArt(nn.Module):
|
||||
self.out_channels = in_channels * 2 if pred_sigma else in_channels
|
||||
self.patch_size = patch_size
|
||||
self.num_heads = num_heads
|
||||
self.lewei_scale = lewei_scale,
|
||||
self.dtype = torch.get_default_dtype()
|
||||
self.pe_interpolation = pe_interpolation
|
||||
self.depth = depth
|
||||
|
||||
self.x_embedder = PatchEmbed(input_size, patch_size, in_channels, hidden_size, bias=True)
|
||||
self.t_embedder = TimestepEmbedder(hidden_size)
|
||||
@@ -102,21 +99,32 @@ class PixArt(nn.Module):
|
||||
nn.SiLU(),
|
||||
nn.Linear(hidden_size, 6 * hidden_size, bias=True)
|
||||
)
|
||||
self.y_embedder = CaptionEmbedder(in_channels=caption_channels, hidden_size=hidden_size, uncond_prob=class_dropout_prob, act_layer=approx_gelu)
|
||||
self.y_embedder = CaptionEmbedder(
|
||||
in_channels=caption_channels, hidden_size=hidden_size, uncond_prob=class_dropout_prob,
|
||||
act_layer=approx_gelu, token_num=model_max_length
|
||||
)
|
||||
drop_path = [x.item() for x in torch.linspace(0, drop_path, depth)] # stochastic depth decay rule
|
||||
self.kv_compress_config = kv_compress_config
|
||||
if kv_compress_config is None:
|
||||
self.kv_compress_config = {
|
||||
'sampling': None,
|
||||
'scale_factor': 1,
|
||||
'kv_compress_layer': [],
|
||||
}
|
||||
self.blocks = nn.ModuleList([
|
||||
PixArtBlock(hidden_size, num_heads, mlp_ratio=mlp_ratio, drop_path=drop_path[i],
|
||||
input_size=(input_size // patch_size, input_size // patch_size),
|
||||
window_size=window_size if i in window_block_indexes else 0,
|
||||
use_rel_pos=use_rel_pos if i in window_block_indexes else False)
|
||||
PixArtBlock(
|
||||
hidden_size, num_heads, mlp_ratio=mlp_ratio, drop_path=drop_path[i],
|
||||
input_size=(input_size // patch_size, input_size // patch_size),
|
||||
sampling=self.kv_compress_config['sampling'],
|
||||
sr_ratio=int(
|
||||
self.kv_compress_config['scale_factor']
|
||||
) if i in self.kv_compress_config['kv_compress_layer'] else 1,
|
||||
qk_norm=qk_norm,
|
||||
)
|
||||
for i in range(depth)
|
||||
])
|
||||
self.final_layer = T2IFinalLayer(hidden_size, patch_size, self.out_channels)
|
||||
|
||||
self.initialize_weights()
|
||||
|
||||
print(f'Warning: lewei scale: {self.lewei_scale}, base size: {self.base_size}')
|
||||
|
||||
def forward_raw(self, x, t, y, mask=None, data_info=None):
|
||||
"""
|
||||
Original forward pass of PixArt.
|
||||
@@ -124,9 +132,13 @@ class PixArt(nn.Module):
|
||||
t: (N,) tensor of diffusion timesteps
|
||||
y: (N, 1, 120, C) tensor of class labels
|
||||
"""
|
||||
x = x.to(self.dtype)
|
||||
timestep = t.to(self.dtype)
|
||||
y = y.to(self.dtype)
|
||||
pos_embed = self.pos_embed.to(self.dtype)
|
||||
self.h, self.w = x.shape[-2]//self.patch_size, x.shape[-1]//self.patch_size
|
||||
x = self.x_embedder(x) + self.pos_embed # (N, T, D), where T = H * W / patch_size ** 2
|
||||
t = self.t_embedder(t) # (N, D)
|
||||
x = self.x_embedder(x) + pos_embed # (N, T, D), where T = H * W / patch_size ** 2
|
||||
t = self.t_embedder(timestep.to(x.dtype)) # (N, D)
|
||||
t0 = self.t_block(t)
|
||||
y = self.y_embedder(y, self.training) # (N, 1, L, D)
|
||||
if mask is not None:
|
||||
@@ -168,29 +180,6 @@ class PixArt(nn.Module):
|
||||
eps, rest = out[:, :self.in_channels], out[:, self.in_channels:]
|
||||
return eps
|
||||
|
||||
def forward_with_dpmsolver(self, x, t, y, mask=None, **kwargs):
|
||||
"""
|
||||
dpm solver donnot need variance prediction
|
||||
"""
|
||||
# https://github.com/openai/glide-text2im/blob/main/notebooks/text2im.ipynb
|
||||
model_out = self.forward(x, t, y, mask)
|
||||
return model_out.chunk(2, dim=1)[0]
|
||||
|
||||
def forward_with_cfg(self, x, t, y, cfg_scale, **kwargs):
|
||||
"""
|
||||
Forward pass of PixArt, but also batches the unconditional forward pass for classifier-free guidance.
|
||||
"""
|
||||
# https://github.com/openai/glide-text2im/blob/main/notebooks/text2im.ipynb
|
||||
half = x[: len(x) // 2]
|
||||
combined = torch.cat([half, half], dim=0)
|
||||
model_out = self.forward(combined, t, y, kwargs)
|
||||
model_out = model_out['x'] if isinstance(model_out, dict) else model_out
|
||||
eps, rest = model_out[:, :3], model_out[:, 3:]
|
||||
cond_eps, uncond_eps = torch.split(eps, len(eps) // 2, dim=0)
|
||||
half_eps = uncond_eps + cfg_scale * (cond_eps - uncond_eps)
|
||||
eps = torch.cat([half_eps, half_eps], dim=0)
|
||||
return torch.cat([eps, rest], dim=1)
|
||||
|
||||
def unpatchify(self, x):
|
||||
"""
|
||||
x: (N, T, patch_size**2 * C)
|
||||
@@ -206,44 +195,8 @@ class PixArt(nn.Module):
|
||||
imgs = x.reshape(shape=(x.shape[0], c, h * p, h * p))
|
||||
return imgs
|
||||
|
||||
def initialize_weights(self):
|
||||
# Initialize transformer layers:
|
||||
def _basic_init(module):
|
||||
if isinstance(module, nn.Linear):
|
||||
torch.nn.init.xavier_uniform_(module.weight)
|
||||
if module.bias is not None:
|
||||
nn.init.constant_(module.bias, 0)
|
||||
|
||||
self.apply(_basic_init)
|
||||
|
||||
# Initialize (and freeze) pos_embed by sin-cos embedding:
|
||||
pos_embed = get_2d_sincos_pos_embed(self.pos_embed.shape[-1], int(self.x_embedder.num_patches ** 0.5), lewei_scale=self.lewei_scale, base_size=self.base_size)
|
||||
self.pos_embed.data.copy_(torch.from_numpy(pos_embed).unsqueeze(0).to(self.dtype))
|
||||
|
||||
# Initialize patch_embed like nn.Linear (instead of nn.Conv2d):
|
||||
w = self.x_embedder.proj.weight.data
|
||||
nn.init.xavier_uniform_(w.view([w.shape[0], -1]))
|
||||
|
||||
# Initialize timestep embedding MLP:
|
||||
nn.init.normal_(self.t_embedder.mlp[0].weight, std=0.02)
|
||||
nn.init.normal_(self.t_embedder.mlp[2].weight, std=0.02)
|
||||
nn.init.normal_(self.t_block[1].weight, std=0.02)
|
||||
|
||||
# Initialize caption embedding MLP:
|
||||
nn.init.normal_(self.y_embedder.y_proj.fc1.weight, std=0.02)
|
||||
nn.init.normal_(self.y_embedder.y_proj.fc2.weight, std=0.02)
|
||||
|
||||
# Zero-out adaLN modulation layers in PixArt blocks:
|
||||
for block in self.blocks:
|
||||
nn.init.constant_(block.cross_attn.proj.weight, 0)
|
||||
nn.init.constant_(block.cross_attn.proj.bias, 0)
|
||||
|
||||
# Zero-out output layers:
|
||||
nn.init.constant_(self.final_layer.linear.weight, 0)
|
||||
nn.init.constant_(self.final_layer.linear.bias, 0)
|
||||
|
||||
|
||||
def get_2d_sincos_pos_embed(embed_dim, grid_size, cls_token=False, extra_tokens=0, lewei_scale=1.0, base_size=16):
|
||||
def get_2d_sincos_pos_embed(embed_dim, grid_size, cls_token=False, extra_tokens=0, pe_interpolation=1.0, base_size=16):
|
||||
"""
|
||||
grid_size: int of the grid height and width
|
||||
return:
|
||||
@@ -251,8 +204,8 @@ def get_2d_sincos_pos_embed(embed_dim, grid_size, cls_token=False, extra_tokens=
|
||||
"""
|
||||
if isinstance(grid_size, int):
|
||||
grid_size = to_2tuple(grid_size)
|
||||
grid_h = np.arange(grid_size[0], dtype=np.float32) / (grid_size[0]/base_size) / lewei_scale
|
||||
grid_w = np.arange(grid_size[1], dtype=np.float32) / (grid_size[1]/base_size) / lewei_scale
|
||||
grid_h = np.arange(grid_size[0], dtype=np.float32) / (grid_size[0]/base_size) / pe_interpolation
|
||||
grid_w = np.arange(grid_size[1], dtype=np.float32) / (grid_size[1]/base_size) / pe_interpolation
|
||||
grid = np.meshgrid(grid_w, grid_h) # here w goes first
|
||||
grid = np.stack(grid, axis=0)
|
||||
grid = grid.reshape([2, 1, grid_size[1], grid_size[0]])
|
||||
|
||||
+61
-96
@@ -15,12 +15,13 @@ from timm.models.layers import DropPath
|
||||
from timm.models.vision_transformer import Mlp
|
||||
|
||||
from .utils import auto_grad_checkpoint, to_2tuple
|
||||
from .PixArt_blocks import t2i_modulate, CaptionEmbedder, WindowAttention, MultiHeadCrossAttention, T2IFinalLayer, TimestepEmbedder, SizeEmbedder
|
||||
from .PixArt_blocks import t2i_modulate, CaptionEmbedder, AttentionKVCompress, MultiHeadCrossAttention, T2IFinalLayer, TimestepEmbedder, SizeEmbedder
|
||||
from .PixArt import PixArt, get_2d_sincos_pos_embed
|
||||
|
||||
|
||||
class PatchEmbed(nn.Module):
|
||||
""" 2D Image to Patch Embedding
|
||||
"""
|
||||
2D Image to Patch Embedding
|
||||
"""
|
||||
def __init__(
|
||||
self,
|
||||
@@ -50,42 +51,39 @@ class PixArtMSBlock(nn.Module):
|
||||
"""
|
||||
A PixArt block with adaptive layer norm zero (adaLN-Zero) conditioning.
|
||||
"""
|
||||
|
||||
def __init__(self, hidden_size, num_heads, mlp_ratio=4.0, drop_path=0., window_size=0, input_size=None, use_rel_pos=False, **block_kwargs):
|
||||
def __init__(self, hidden_size, num_heads, mlp_ratio=4.0, drop_path=0., input_size=None,
|
||||
sampling=None, sr_ratio=1, qk_norm=False, **block_kwargs):
|
||||
super().__init__()
|
||||
self.hidden_size = hidden_size
|
||||
self.norm1 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
|
||||
self.attn = WindowAttention(hidden_size, num_heads=num_heads, qkv_bias=True,
|
||||
input_size=input_size if window_size == 0 else (window_size, window_size),
|
||||
use_rel_pos=use_rel_pos, **block_kwargs)
|
||||
self.attn = AttentionKVCompress(
|
||||
hidden_size, num_heads=num_heads, qkv_bias=True, sampling=sampling, sr_ratio=sr_ratio,
|
||||
qk_norm=qk_norm, **block_kwargs
|
||||
)
|
||||
self.cross_attn = MultiHeadCrossAttention(hidden_size, num_heads, **block_kwargs)
|
||||
self.norm2 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
|
||||
# to be compatible with lower version pytorch
|
||||
approx_gelu = lambda: nn.GELU(approximate="tanh")
|
||||
self.mlp = Mlp(in_features=hidden_size, hidden_features=int(hidden_size * mlp_ratio), act_layer=approx_gelu, drop=0)
|
||||
self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
|
||||
self.window_size = window_size
|
||||
self.scale_shift_table = nn.Parameter(torch.randn(6, hidden_size) / hidden_size ** 0.5)
|
||||
|
||||
def forward(self, x, y, t, mask=None, **kwargs):
|
||||
def forward(self, x, y, t, mask=None, HW=None, **kwargs):
|
||||
B, N, C = x.shape
|
||||
|
||||
shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = (self.scale_shift_table[None] + t.reshape(B, 6, -1)).chunk(6, dim=1)
|
||||
x = x + self.drop_path(gate_msa * self.attn(t2i_modulate(self.norm1(x), shift_msa, scale_msa)))
|
||||
x = x + self.drop_path(gate_msa * self.attn(t2i_modulate(self.norm1(x), shift_msa, scale_msa), HW=HW))
|
||||
x = x + self.cross_attn(x, y, mask)
|
||||
x = x + self.drop_path(gate_mlp * self.mlp(t2i_modulate(self.norm2(x), shift_mlp, scale_mlp)))
|
||||
|
||||
return x
|
||||
|
||||
|
||||
#############################################################################
|
||||
# Core PixArt Model #
|
||||
#################################################################################
|
||||
### Core PixArt Model ###
|
||||
class PixArtMS(PixArt):
|
||||
"""
|
||||
Diffusion model with a Transformer backbone.
|
||||
"""
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
input_size=32,
|
||||
@@ -99,12 +97,13 @@ class PixArtMS(PixArt):
|
||||
learn_sigma=True,
|
||||
pred_sigma=True,
|
||||
drop_path: float = 0.,
|
||||
window_size=0,
|
||||
window_block_indexes=[],
|
||||
use_rel_pos=False,
|
||||
caption_channels=4096,
|
||||
lewei_scale=1.,
|
||||
pe_interpolation=1.,
|
||||
config=None,
|
||||
model_max_length=120,
|
||||
micro_condition=True,
|
||||
qk_norm=False,
|
||||
kv_compress_config=None,
|
||||
**kwargs,
|
||||
):
|
||||
super().__init__(
|
||||
@@ -119,11 +118,11 @@ class PixArtMS(PixArt):
|
||||
learn_sigma=learn_sigma,
|
||||
pred_sigma=pred_sigma,
|
||||
drop_path=drop_path,
|
||||
window_size=window_size,
|
||||
window_block_indexes=window_block_indexes,
|
||||
use_rel_pos=use_rel_pos,
|
||||
lewei_scale=lewei_scale,
|
||||
pe_interpolation=pe_interpolation,
|
||||
config=config,
|
||||
model_max_length=model_max_length,
|
||||
qk_norm=qk_norm,
|
||||
kv_compress_config=kv_compress_config,
|
||||
**kwargs,
|
||||
)
|
||||
self.dtype = torch.get_default_dtype()
|
||||
@@ -134,20 +133,29 @@ class PixArtMS(PixArt):
|
||||
nn.Linear(hidden_size, 6 * hidden_size, bias=True)
|
||||
)
|
||||
self.x_embedder = PatchEmbed(patch_size, in_channels, hidden_size, bias=True)
|
||||
self.y_embedder = CaptionEmbedder(in_channels=caption_channels, hidden_size=hidden_size, uncond_prob=class_dropout_prob, act_layer=approx_gelu)
|
||||
self.csize_embedder = SizeEmbedder(hidden_size//3) # c_size embed
|
||||
self.ar_embedder = SizeEmbedder(hidden_size//3) # aspect ratio embed
|
||||
self.y_embedder = CaptionEmbedder(in_channels=caption_channels, hidden_size=hidden_size, uncond_prob=class_dropout_prob, act_layer=approx_gelu, token_num=model_max_length)
|
||||
self.micro_conditioning = micro_condition
|
||||
if self.micro_conditioning:
|
||||
self.csize_embedder = SizeEmbedder(hidden_size//3) # c_size embed
|
||||
self.ar_embedder = SizeEmbedder(hidden_size//3) # aspect ratio embed
|
||||
drop_path = [x.item() for x in torch.linspace(0, drop_path, depth)] # stochastic depth decay rule
|
||||
if kv_compress_config is None:
|
||||
kv_compress_config = {
|
||||
'sampling': None,
|
||||
'scale_factor': 1,
|
||||
'kv_compress_layer': [],
|
||||
}
|
||||
self.blocks = nn.ModuleList([
|
||||
PixArtMSBlock(hidden_size, num_heads, mlp_ratio=mlp_ratio, drop_path=drop_path[i],
|
||||
input_size=(input_size // patch_size, input_size // patch_size),
|
||||
window_size=window_size if i in window_block_indexes else 0,
|
||||
use_rel_pos=use_rel_pos if i in window_block_indexes else False)
|
||||
PixArtMSBlock(
|
||||
hidden_size, num_heads, mlp_ratio=mlp_ratio, drop_path=drop_path[i],
|
||||
input_size=(input_size // patch_size, input_size // patch_size),
|
||||
sampling=kv_compress_config['sampling'],
|
||||
sr_ratio=int(kv_compress_config['scale_factor']) if i in kv_compress_config['kv_compress_layer'] else 1,
|
||||
qk_norm=qk_norm,
|
||||
)
|
||||
for i in range(depth)
|
||||
])
|
||||
self.final_layer = T2IFinalLayer(hidden_size, patch_size, self.out_channels)
|
||||
self.training = False
|
||||
self.initialize()
|
||||
|
||||
def forward_raw(self, x, t, y, mask=None, data_info=None, **kwargs):
|
||||
"""
|
||||
@@ -157,16 +165,29 @@ class PixArtMS(PixArt):
|
||||
y: (N, 1, 120, C) tensor of class labels
|
||||
"""
|
||||
bs = x.shape[0]
|
||||
c_size, ar = data_info['img_hw'], data_info['aspect_ratio']
|
||||
x = x.to(self.dtype)
|
||||
timestep = t.to(self.dtype)
|
||||
y = y.to(self.dtype)
|
||||
self.h, self.w = x.shape[-2]//self.patch_size, x.shape[-1]//self.patch_size
|
||||
pos_embed = torch.from_numpy(get_2d_sincos_pos_embed(self.pos_embed.shape[-1], (self.h, self.w), lewei_scale=self.lewei_scale, base_size=self.base_size)).unsqueeze(0).to(x.device).to(self.dtype)
|
||||
pos_embed = torch.from_numpy(
|
||||
get_2d_sincos_pos_embed(
|
||||
self.pos_embed.shape[-1], (self.h, self.w), pe_interpolation=self.pe_interpolation,
|
||||
base_size=self.base_size
|
||||
)
|
||||
).unsqueeze(0).to(x.device).to(self.dtype)
|
||||
|
||||
x = self.x_embedder(x) + pos_embed # (N, T, D), where T = H * W / patch_size ** 2
|
||||
t = self.t_embedder(t) # (N, D)
|
||||
csize = self.csize_embedder(c_size, bs) # (N, D)
|
||||
ar = self.ar_embedder(ar, bs) # (N, D)
|
||||
t = t + torch.cat([csize, ar], dim=1)
|
||||
t = self.t_embedder(timestep) # (N, D)
|
||||
|
||||
if self.micro_conditioning:
|
||||
c_size, ar = data_info['img_hw'].to(self.dtype), data_info['aspect_ratio'].to(self.dtype)
|
||||
csize = self.csize_embedder(c_size, bs) # (N, D)
|
||||
ar = self.ar_embedder(ar, bs) # (N, D)
|
||||
t = t + torch.cat([csize, ar], dim=1)
|
||||
|
||||
t0 = self.t_block(t)
|
||||
y = self.y_embedder(y, self.training) # (N, D)
|
||||
|
||||
if mask is not None:
|
||||
if mask.shape[0] != y.shape[0]:
|
||||
mask = mask.repeat(y.shape[0] // mask.shape[0], 1)
|
||||
@@ -177,9 +198,11 @@ class PixArtMS(PixArt):
|
||||
y_lens = [y.shape[2]] * y.shape[0]
|
||||
y = y.squeeze(1).view(1, -1, x.shape[-1])
|
||||
for block in self.blocks:
|
||||
x = auto_grad_checkpoint(block, x, y, t0, y_lens, **kwargs) # (N, T, D) #support grad checkpoint
|
||||
x = auto_grad_checkpoint(block, x, y, t0, y_lens, (self.h, self.w), **kwargs) # (N, T, D) #support grad checkpoint
|
||||
|
||||
x = self.final_layer(x, t) # (N, T, patch_size ** 2 * out_channels)
|
||||
x = self.unpatchify(x) # (N, out_channels, H, W)
|
||||
|
||||
return x
|
||||
|
||||
def forward(self, x, timesteps, context, img_hw=None, aspect_ratio=None, **kwargs):
|
||||
@@ -228,28 +251,6 @@ class PixArtMS(PixArt):
|
||||
eps, rest = out[:, :self.in_channels], out[:, self.in_channels:]
|
||||
return eps
|
||||
|
||||
def forward_with_dpmsolver(self, x, t, y, data_info, **kwargs):
|
||||
"""
|
||||
dpm solver donnot need variance prediction
|
||||
"""
|
||||
# https://github.com/openai/glide-text2im/blob/main/notebooks/text2im.ipynb
|
||||
model_out = self.forward_raw(x, t, y, data_info=data_info, **kwargs)
|
||||
return model_out.chunk(2, dim=1)[0]
|
||||
|
||||
def forward_with_cfg(self, x, t, y, cfg_scale, data_info, **kwargs):
|
||||
"""
|
||||
Forward pass of PixArt, but also batches the unconditional forward pass for classifier-free guidance.
|
||||
"""
|
||||
# https://github.com/openai/glide-text2im/blob/main/notebooks/text2im.ipynb
|
||||
half = x[: len(x) // 2]
|
||||
combined = torch.cat([half, half], dim=0)
|
||||
model_out = self.forward_raw(combined, t, y, data_info=data_info)
|
||||
eps, rest = model_out[:, :3], model_out[:, 3:]
|
||||
cond_eps, uncond_eps = torch.split(eps, len(eps) // 2, dim=0)
|
||||
half_eps = uncond_eps + cfg_scale * (cond_eps - uncond_eps)
|
||||
eps = torch.cat([half_eps, half_eps], dim=0)
|
||||
return torch.cat([eps, rest], dim=1)
|
||||
|
||||
def unpatchify(self, x):
|
||||
"""
|
||||
x: (N, T, patch_size**2 * C)
|
||||
@@ -263,39 +264,3 @@ class PixArtMS(PixArt):
|
||||
x = torch.einsum('nhwpqc->nchpwq', x)
|
||||
imgs = x.reshape(shape=(x.shape[0], c, self.h * p, self.w * p))
|
||||
return imgs
|
||||
|
||||
def initialize(self):
|
||||
# Initialize transformer layers:
|
||||
def _basic_init(module):
|
||||
if isinstance(module, nn.Linear):
|
||||
torch.nn.init.xavier_uniform_(module.weight)
|
||||
if module.bias is not None:
|
||||
nn.init.constant_(module.bias, 0)
|
||||
|
||||
self.apply(_basic_init)
|
||||
|
||||
# Initialize patch_embed like nn.Linear (instead of nn.Conv2d):
|
||||
w = self.x_embedder.proj.weight.data
|
||||
nn.init.xavier_uniform_(w.view([w.shape[0], -1]))
|
||||
|
||||
# Initialize timestep embedding MLP:
|
||||
nn.init.normal_(self.t_embedder.mlp[0].weight, std=0.02)
|
||||
nn.init.normal_(self.t_embedder.mlp[2].weight, std=0.02)
|
||||
nn.init.normal_(self.t_block[1].weight, std=0.02)
|
||||
nn.init.normal_(self.csize_embedder.mlp[0].weight, std=0.02)
|
||||
nn.init.normal_(self.csize_embedder.mlp[2].weight, std=0.02)
|
||||
nn.init.normal_(self.ar_embedder.mlp[0].weight, std=0.02)
|
||||
nn.init.normal_(self.ar_embedder.mlp[2].weight, std=0.02)
|
||||
|
||||
# Initialize caption embedding MLP:
|
||||
nn.init.normal_(self.y_embedder.y_proj.fc1.weight, std=0.02)
|
||||
nn.init.normal_(self.y_embedder.y_proj.fc2.weight, std=0.02)
|
||||
|
||||
# Zero-out adaLN modulation layers in PixArt blocks:
|
||||
for block in self.blocks:
|
||||
nn.init.constant_(block.cross_attn.proj.weight, 0)
|
||||
nn.init.constant_(block.cross_attn.proj.bias, 0)
|
||||
|
||||
# Zero-out output layers:
|
||||
nn.init.constant_(self.final_layer.linear.weight, 0)
|
||||
nn.init.constant_(self.final_layer.linear.bias, 0)
|
||||
|
||||
+116
-80
@@ -11,25 +11,29 @@
|
||||
import math
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
import torch.nn.functional as F
|
||||
from timm.models.vision_transformer import Mlp, Attention as Attention_
|
||||
from einops import rearrange, repeat
|
||||
from einops import rearrange
|
||||
|
||||
from .utils import add_decomposed_rel_pos
|
||||
from comfy import model_management
|
||||
|
||||
if model_management.xformers_enabled():
|
||||
import xformers
|
||||
import xformers.ops
|
||||
else:
|
||||
print("""
|
||||
########################################
|
||||
PixArt: Not using xformers!
|
||||
Expect images to be non-deterministic!
|
||||
Batch sizes > 1 are most likely broken
|
||||
########################################
|
||||
""")
|
||||
|
||||
def modulate(x, shift, scale):
|
||||
return x * (1 + scale.unsqueeze(1)) + shift.unsqueeze(1)
|
||||
|
||||
|
||||
def t2i_modulate(x, shift, scale):
|
||||
return x * (1 + scale) + shift
|
||||
|
||||
competent_attention_implementation = False
|
||||
|
||||
class MultiHeadCrossAttention(nn.Module):
|
||||
def __init__(self, d_model, num_heads, attn_drop=0., proj_drop=0., **block_kwargs):
|
||||
super(MultiHeadCrossAttention, self).__init__()
|
||||
@@ -49,32 +53,24 @@ class MultiHeadCrossAttention(nn.Module):
|
||||
# query/value: img tokens; key: condition; mask: if padding tokens
|
||||
B, N, C = x.shape
|
||||
|
||||
q = self.q_linear(x).view(1, -1, self.num_heads, self.head_dim)
|
||||
kv = self.kv_linear(cond).view(1, -1, 2, self.num_heads, self.head_dim)
|
||||
k, v = kv.unbind(2)
|
||||
|
||||
if model_management.xformers_enabled():
|
||||
q = self.q_linear(x).view(1, -1, self.num_heads, self.head_dim)
|
||||
kv = self.kv_linear(cond).view(1, -1, 2, self.num_heads, self.head_dim)
|
||||
k, v = kv.unbind(2)
|
||||
attn_bias = None
|
||||
if mask is not None:
|
||||
attn_bias = xformers.ops.fmha.BlockDiagonalMask.from_seqlens([N] * B, mask)
|
||||
x = xformers.ops.memory_efficient_attention(q, k, v, p=self.attn_drop.p, attn_bias=attn_bias)
|
||||
x = x.view(B, -1, C)
|
||||
x = self.proj(x)
|
||||
x = self.proj_drop(x)
|
||||
return x
|
||||
x = xformers.ops.memory_efficient_attention(
|
||||
q, k, v,
|
||||
p=self.attn_drop.p,
|
||||
attn_bias=attn_bias
|
||||
)
|
||||
else:
|
||||
global competent_attention_implementation
|
||||
if not competent_attention_implementation:
|
||||
print("""\nYou should REALLY consider installing/enabling xformers.\nAlternatively, open up ExtraModels/PixArt/models/PixArt_blocks.py and\n- Fix the attention map on line 77 if you know how to\n- Add scaled_dot_product_attention on line 150\n- Send a PR and remove this message on line 32/66-69\n""")
|
||||
competent_attention_implementation = True
|
||||
|
||||
q = self.q_linear(x).view(1, -1, self.num_heads, self.head_dim)
|
||||
kv = self.kv_linear(cond).view(1, -1, 2, self.num_heads, self.head_dim)
|
||||
k, v = kv.unbind(2)
|
||||
q, k, v = map(lambda t: t.permute(0, 2, 1, 3),(q, k, v),)
|
||||
|
||||
attn_mask = None
|
||||
if mask is not None and len(mask) > 1:
|
||||
# This is probably wrong
|
||||
# This is most definitely wrong, especially for B>1
|
||||
attn_mask = torch.zeros(
|
||||
[1, q.shape[1], q.shape[2], v.shape[2]],
|
||||
dtype=q.dtype,
|
||||
@@ -82,26 +78,28 @@ class MultiHeadCrossAttention(nn.Module):
|
||||
)
|
||||
attn_mask[:, :, (q.shape[2]//2):, mask[0]:] = True
|
||||
attn_mask[:, :, :(q.shape[2]//2), :mask[1]] = True
|
||||
|
||||
x = torch.nn.functional.scaled_dot_product_attention(q, k, v, attn_mask=attn_mask, dropout_p=self.attn_drop.p)
|
||||
x = x.permute(0, 2, 1, 3).contiguous()
|
||||
x = x.view(B, -1, C)
|
||||
x = self.proj(x)
|
||||
x = self.proj_drop(x)
|
||||
return x
|
||||
x = torch.nn.functional.scaled_dot_product_attention(
|
||||
q, k, v,
|
||||
attn_mask=attn_mask,
|
||||
dropout_p=self.attn_drop.p
|
||||
).permute(0, 2, 1, 3).contiguous()
|
||||
x = x.view(B, -1, C)
|
||||
x = self.proj(x)
|
||||
x = self.proj_drop(x)
|
||||
return x
|
||||
|
||||
|
||||
class WindowAttention(Attention_):
|
||||
"""Multi-head Attention block with relative position embeddings."""
|
||||
class AttentionKVCompress(Attention_):
|
||||
"""Multi-head Attention block with KV token compression and qk norm."""
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
dim,
|
||||
num_heads=8,
|
||||
qkv_bias=True,
|
||||
use_rel_pos=False,
|
||||
rel_pos_zero_init=True,
|
||||
input_size=None,
|
||||
sampling='conv',
|
||||
sr_ratio=1,
|
||||
qk_norm=False,
|
||||
**block_kwargs,
|
||||
):
|
||||
"""
|
||||
@@ -109,56 +107,94 @@ class WindowAttention(Attention_):
|
||||
dim (int): Number of input channels.
|
||||
num_heads (int): Number of attention heads.
|
||||
qkv_bias (bool: If True, add a learnable bias to query, key, value.
|
||||
rel_pos (bool): If True, add relative positional embeddings to the attention map.
|
||||
rel_pos_zero_init (bool): If True, zero initialize relative positional parameters.
|
||||
input_size (int or None): Input resolution for calculating the relative positional
|
||||
parameter size.
|
||||
"""
|
||||
super().__init__(dim, num_heads=num_heads, qkv_bias=qkv_bias, **block_kwargs)
|
||||
|
||||
self.use_rel_pos = use_rel_pos
|
||||
if self.use_rel_pos:
|
||||
# initialize relative positional embeddings
|
||||
self.rel_pos_h = nn.Parameter(torch.zeros(2 * input_size[0] - 1, self.head_dim))
|
||||
self.rel_pos_w = nn.Parameter(torch.zeros(2 * input_size[1] - 1, self.head_dim))
|
||||
|
||||
if not rel_pos_zero_init:
|
||||
nn.init.trunc_normal_(self.rel_pos_h, std=0.02)
|
||||
nn.init.trunc_normal_(self.rel_pos_w, std=0.02)
|
||||
|
||||
def forward(self, x, mask=None):
|
||||
B, N, C = x.shape # 2 4096 1152
|
||||
qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads)
|
||||
|
||||
if model_management.xformers_enabled():
|
||||
q, k, v = qkv.unbind(2)
|
||||
|
||||
if getattr(self, 'fp32_attention', False):
|
||||
q, k, v = q.float(), k.float(), v.float()
|
||||
|
||||
attn_bias = None
|
||||
if mask is not None:
|
||||
attn_bias = torch.zeros([B * self.num_heads, q.shape[1], k.shape[1]], dtype=q.dtype, device=q.device)
|
||||
attn_bias.masked_fill_(mask.squeeze(1).repeat(self.num_heads, 1, 1) == 0, float('-inf'))
|
||||
# Switch between torch / xformers attention
|
||||
x = xformers.ops.memory_efficient_attention(q, k, v, p=self.attn_drop.p, attn_bias=attn_bias)
|
||||
x = x.view(B, N, C)
|
||||
x = self.proj(x)
|
||||
x = self.proj_drop(x)
|
||||
return x
|
||||
self.sampling=sampling # ['conv', 'ave', 'uniform', 'uniform_every']
|
||||
self.sr_ratio = sr_ratio
|
||||
if sr_ratio > 1 and sampling == 'conv':
|
||||
# Avg Conv Init.
|
||||
self.sr = nn.Conv2d(dim, dim, groups=dim, kernel_size=sr_ratio, stride=sr_ratio)
|
||||
self.sr.weight.data.fill_(1/sr_ratio**2)
|
||||
self.sr.bias.data.zero_()
|
||||
self.norm = nn.LayerNorm(dim)
|
||||
if qk_norm:
|
||||
self.q_norm = nn.LayerNorm(dim)
|
||||
self.k_norm = nn.LayerNorm(dim)
|
||||
else:
|
||||
q, k, v = qkv.permute(2, 0, 3, 1, 4).unbind(0)
|
||||
self.q_norm = nn.Identity()
|
||||
self.k_norm = nn.Identity()
|
||||
|
||||
q = q * self.scale
|
||||
attn = q @ k.transpose(-2, -1)
|
||||
attn = attn.softmax(dim=-1)
|
||||
attn = self.attn_drop(attn)
|
||||
x = attn @ v
|
||||
def downsample_2d(self, tensor, H, W, scale_factor, sampling=None):
|
||||
if sampling is None or scale_factor == 1:
|
||||
return tensor
|
||||
B, N, C = tensor.shape
|
||||
|
||||
x = x.transpose(1, 2).reshape(B, N, C)
|
||||
x = self.proj(x)
|
||||
x = self.proj_drop(x)
|
||||
return x
|
||||
if sampling == 'uniform_every':
|
||||
return tensor[:, ::scale_factor], int(N // scale_factor)
|
||||
|
||||
tensor = tensor.reshape(B, H, W, C).permute(0, 3, 1, 2)
|
||||
new_H, new_W = int(H / scale_factor), int(W / scale_factor)
|
||||
new_N = new_H * new_W
|
||||
|
||||
if sampling == 'ave':
|
||||
tensor = F.interpolate(
|
||||
tensor, scale_factor=1 / scale_factor, mode='nearest'
|
||||
).permute(0, 2, 3, 1)
|
||||
elif sampling == 'uniform':
|
||||
tensor = tensor[:, :, ::scale_factor, ::scale_factor].permute(0, 2, 3, 1)
|
||||
elif sampling == 'conv':
|
||||
tensor = self.sr(tensor).reshape(B, C, -1).permute(0, 2, 1)
|
||||
tensor = self.norm(tensor)
|
||||
else:
|
||||
raise ValueError
|
||||
|
||||
return tensor.reshape(B, new_N, C).contiguous(), new_N
|
||||
|
||||
def forward(self, x, mask=None, HW=None, block_id=None):
|
||||
B, N, C = x.shape # 2 4096 1152
|
||||
new_N = N
|
||||
if HW is None:
|
||||
H = W = int(N ** 0.5)
|
||||
else:
|
||||
H, W = HW
|
||||
qkv = self.qkv(x).reshape(B, N, 3, C)
|
||||
|
||||
q, k, v = qkv.unbind(2)
|
||||
dtype = q.dtype
|
||||
q = self.q_norm(q)
|
||||
k = self.k_norm(k)
|
||||
|
||||
# KV compression
|
||||
if self.sr_ratio > 1:
|
||||
k, new_N = self.downsample_2d(k, H, W, self.sr_ratio, sampling=self.sampling)
|
||||
v, new_N = self.downsample_2d(v, H, W, self.sr_ratio, sampling=self.sampling)
|
||||
|
||||
q = q.reshape(B, N, self.num_heads, C // self.num_heads).to(dtype)
|
||||
k = k.reshape(B, new_N, self.num_heads, C // self.num_heads).to(dtype)
|
||||
v = v.reshape(B, new_N, self.num_heads, C // self.num_heads).to(dtype)
|
||||
|
||||
attn_bias = None
|
||||
if mask is not None:
|
||||
attn_bias = torch.zeros([B * self.num_heads, q.shape[1], k.shape[1]], dtype=q.dtype, device=q.device)
|
||||
attn_bias.masked_fill_(mask.squeeze(1).repeat(self.num_heads, 1, 1) == 0, float('-inf'))
|
||||
# Switch between torch / xformers attention
|
||||
if model_management.xformers_enabled():
|
||||
x = xformers.ops.memory_efficient_attention(
|
||||
q, k, v,
|
||||
p=self.attn_drop.p,
|
||||
attn_bias=attn_bias
|
||||
)
|
||||
else:
|
||||
q, k, v = map(lambda t: t.transpose(1, 2),(q, k, v),)
|
||||
x = torch.nn.functional.scaled_dot_product_attention(
|
||||
q, k, v,
|
||||
attn_mask=attn_bias
|
||||
).transpose(1, 2).contiguous()
|
||||
x = x.view(B, N, C)
|
||||
x = self.proj(x)
|
||||
x = self.proj_drop(x)
|
||||
return x
|
||||
|
||||
|
||||
#################################################################################
|
||||
|
||||
Reference in New Issue
Block a user