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
+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)
|
||||
|
||||
Reference in New Issue
Block a user