1074 lines
54 KiB
Python
1074 lines
54 KiB
Python
import os
|
|
import numpy as np
|
|
from torch import nn
|
|
import torch
|
|
from einops import rearrange, repeat
|
|
from typing import Any, Dict, Optional, Tuple
|
|
from torch.nn import functional as F
|
|
from diffusers.loaders import PeftAdapterMixin
|
|
from diffusers.models.transformer_2d import Transformer2DModelOutput
|
|
from diffusers.utils import is_torch_version, deprecate
|
|
from diffusers.configuration_utils import ConfigMixin, register_to_config
|
|
from diffusers.models.modeling_utils import ModelMixin
|
|
from diffusers.models.normalization import AdaLayerNormSingle
|
|
from diffusers.models.embeddings import PixArtAlphaTextProjection
|
|
from opensora.models.diffusion.udit.modules import Upsample2d, Downsample2d, PatchEmbed2D, BasicTransformerBlock, \
|
|
FP32_GELU, FP32_SiLU, FP32_Layernorm
|
|
from opensora.utils.utils import to_2tuple
|
|
import math
|
|
import re
|
|
try:
|
|
import torch_npu
|
|
from opensora.npu_config import npu_config
|
|
except:
|
|
torch_npu = None
|
|
npu_config = None
|
|
|
|
class UDiTT2V(ModelMixin, ConfigMixin, PeftAdapterMixin):
|
|
"""
|
|
A 2D Transformer model for image-like data.
|
|
|
|
Parameters:
|
|
num_attention_heads (`int`, *optional*, defaults to 16): The number of heads to use for multi-head attention.
|
|
attention_head_dim (`int`, *optional*, defaults to 88): The number of channels in each head.
|
|
in_channels (`int`, *optional*):
|
|
The number of channels in the input and output (specify if the input is **continuous**).
|
|
num_layers (`int`, *optional*, defaults to 1): The number of layers of Transformer blocks to use.
|
|
dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
|
|
cross_attention_dim (`int`, *optional*): The number of `encoder_hidden_states` dimensions to use.
|
|
sample_size (`int`, *optional*): The width of the latent images (specify if the input is **discrete**).
|
|
This is fixed during training since it is used to learn a number of position embeddings.
|
|
num_vector_embeds (`int`, *optional*):
|
|
The number of classes of the vector embeddings of the latent pixels (specify if the input is **discrete**).
|
|
Includes the class for the masked latent pixel.
|
|
activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to use in feed-forward.
|
|
num_embeds_ada_norm ( `int`, *optional*):
|
|
The number of diffusion steps used during training. Pass if at least one of the norm_layers is
|
|
`AdaLayerNorm`. This is fixed during training since it is used to learn a number of embeddings that are
|
|
added to the hidden states.
|
|
|
|
During inference, you can denoise for up to but not more steps than `num_embeds_ada_norm`.
|
|
attention_bias (`bool`, *optional*):
|
|
Configure if the `TransformerBlocks` attention should contain a bias parameter.
|
|
"""
|
|
|
|
_supports_gradient_checkpointing = True
|
|
|
|
@register_to_config
|
|
def __init__(
|
|
self,
|
|
num_attention_heads: int = 16,
|
|
attention_head_dim: int = 88,
|
|
in_channels: Optional[int] = None,
|
|
out_channels: Optional[int] = None,
|
|
num_layers: int = 1,
|
|
dropout: float = 0.0,
|
|
norm_num_groups: int = 32,
|
|
cross_attention_dim: Optional[int] = None,
|
|
attention_bias: bool = False,
|
|
sample_size: Optional[int] = None,
|
|
sample_size_t: Optional[int] = None,
|
|
patch_size: Optional[int] = None,
|
|
patch_size_t: Optional[int] = None,
|
|
num_vector_embeds: Optional[int] = None,
|
|
mlp_ratio: int = 4,
|
|
depth: Optional[list] = [2, 5, 8, 5, 2],
|
|
activation_fn: str = "geglu",
|
|
num_embeds_ada_norm: Optional[int] = None,
|
|
use_linear_projection: bool = False,
|
|
only_cross_attention: bool = False,
|
|
double_self_attention: bool = False,
|
|
upcast_attention: bool = False,
|
|
norm_type: str = "layer_norm", # 'layer_norm', 'ada_norm', 'ada_norm_zero', 'ada_norm_single', 'ada_norm_continuous', 'layer_norm_i2vgen'
|
|
norm_elementwise_affine: bool = True,
|
|
norm_eps: float = 1e-5,
|
|
attention_type: str = "default",
|
|
caption_channels: int = None,
|
|
interpolation_scale_h: float = None,
|
|
interpolation_scale_w: float = None,
|
|
interpolation_scale_t: float = None,
|
|
use_additional_conditions: Optional[bool] = None,
|
|
attention_mode: str = 'xformers',
|
|
downsampler: str = 'k333_s222',
|
|
use_rope: bool = False,
|
|
use_stable_fp32: bool = False,
|
|
):
|
|
super().__init__()
|
|
|
|
# Set some common variables used across the board.
|
|
self.use_stable_fp32 = use_stable_fp32
|
|
self.use_rope = use_rope
|
|
self.downsampler = downsampler
|
|
self.use_linear_projection = use_linear_projection
|
|
self.interpolation_scale_t = interpolation_scale_t
|
|
self.interpolation_scale_h = interpolation_scale_h
|
|
self.interpolation_scale_w = interpolation_scale_w
|
|
self.caption_channels = caption_channels
|
|
self.num_attention_heads = num_attention_heads
|
|
self.attention_head_dim = attention_head_dim
|
|
self.inner_dim = self.config.num_attention_heads * self.config.attention_head_dim
|
|
self.in_channels = in_channels
|
|
self.out_channels = in_channels if out_channels is None else out_channels
|
|
self.gradient_checkpointing = False
|
|
use_additional_conditions = False
|
|
# if use_additional_conditions is None:
|
|
# if norm_type == "ada_norm_single" and sample_size == 128:
|
|
# use_additional_conditions = True
|
|
# else:
|
|
# use_additional_conditions = False
|
|
self.use_additional_conditions = use_additional_conditions
|
|
|
|
# 1. Transformer2DModel can process both standard continuous images of shape `(batch_size, num_channels, width, height)` as well as quantized image embeddings of shape `(batch_size, num_image_vectors)`
|
|
# Define whether input is continuous or discrete depending on configuration
|
|
assert in_channels is not None
|
|
|
|
if norm_type == "layer_norm" and num_embeds_ada_norm is not None:
|
|
deprecation_message = (
|
|
f"The configuration file of this model: {self.__class__} is outdated. `norm_type` is either not set or"
|
|
" incorrectly set to `'layer_norm'`. Make sure to set `norm_type` to `'ada_norm'` in the config."
|
|
" Please make sure to update the config accordingly as leaving `norm_type` might led to incorrect"
|
|
" results in future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it"
|
|
" would be very nice if you could open a Pull request for the `transformer/config.json` file"
|
|
)
|
|
deprecate("norm_type!=num_embeds_ada_norm", "1.0.0", deprecation_message, standard_warn=False)
|
|
norm_type = "ada_norm"
|
|
|
|
# 2. Initialize the right blocks.
|
|
# Initialize the output blocks and other projection blocks when necessary.
|
|
self._init_patched_inputs(norm_type=norm_type)
|
|
if self.use_stable_fp32:
|
|
self._replace_fp32_layers()
|
|
|
|
def _init_patched_inputs(self, norm_type):
|
|
assert self.config.sample_size_t is not None, "OpenSoraT2V over patched input must provide sample_size_t"
|
|
assert self.config.sample_size is not None, "OpenSoraT2V over patched input must provide sample_size"
|
|
|
|
self.config.sample_size = to_2tuple(self.config.sample_size)
|
|
# self.num_frames = self.config.sample_size_t
|
|
# self.height = self.config.sample_size[0]
|
|
# self.width = self.config.sample_size[1]
|
|
interpolation_scale_t = ((self.config.sample_size_t - 1) // 16 + 1) if self.config.sample_size_t % 2 == 1 else self.config.sample_size_t / 16
|
|
interpolation_scale_t = (
|
|
self.config.interpolation_scale_t if self.config.interpolation_scale_t is not None else interpolation_scale_t
|
|
)
|
|
interpolation_scale = (
|
|
self.config.interpolation_scale_h if self.config.interpolation_scale_h is not None else self.config.sample_size[0] / 30,
|
|
self.config.interpolation_scale_w if self.config.interpolation_scale_w is not None else self.config.sample_size[1] / 40,
|
|
)
|
|
|
|
# down_factor = list(re.search(r's(\d{2,3})', self.downsampler).group(1))
|
|
# down_factor = [int(i) for i in down_factor]
|
|
# down_factor = down_factor if isinstance(self.config.down_factor, list) else [self.config.down_factor] * 5
|
|
# down_factor = [2] * len(self.config.depth)
|
|
is_video_model = False
|
|
# if self.config.downsampler is not None and len(self.config.downsampler) == 9:
|
|
# is_video_model = True # to init weight from image
|
|
# self.pos_embed = OverlapPatchEmbed3D(
|
|
# num_frames=self.config.sample_size_t,
|
|
# height=self.config.sample_size[0],
|
|
# width=self.config.sample_size[1],
|
|
# patch_size_t=self.config.patch_size_t,
|
|
# patch_size=self.config.patch_size,
|
|
# in_channels=self.in_channels,
|
|
# embed_dim=self.inner_dim,
|
|
# interpolation_scale=interpolation_scale,
|
|
# interpolation_scale_t=interpolation_scale_t,
|
|
# use_abs_pos=not self.config.use_rope,
|
|
# )
|
|
# elif self.config.downsampler is not None and len(self.config.downsampler) == 7:
|
|
# is_video_model = False
|
|
self.pos_embed = PatchEmbed2D(
|
|
num_frames=self.config.sample_size_t,
|
|
height=self.config.sample_size[0],
|
|
width=self.config.sample_size[1],
|
|
patch_size_t=self.config.patch_size_t,
|
|
patch_size=self.config.patch_size,
|
|
in_channels=self.in_channels,
|
|
embed_dim=self.inner_dim,
|
|
interpolation_scale=interpolation_scale,
|
|
interpolation_scale_t=interpolation_scale_t,
|
|
use_abs_pos=not self.config.use_rope,
|
|
)
|
|
# layer_thw = [[self.config.sample_size_t//self.config.patch_size_t,
|
|
# (self.config.sample_size[0] + self.config.sample_size[0] % (self.config.patch_size*2))//self.config.patch_size,
|
|
# (self.config.sample_size[1] + self.config.sample_size[1] % (self.config.patch_size*2))//self.config.patch_size]]
|
|
interpolation_scale_thw = (interpolation_scale_t, *interpolation_scale)
|
|
# for i in range((len(self.config.depth)-1)//2):
|
|
# t = layer_thw[i][0] // 2 if layer_thw[i][0] != 1 else 1
|
|
# h = (layer_thw[i][1] + layer_thw[i][1] % 4) // 2 # why mod 4, because downsample and downsampler in attention
|
|
# w = (layer_thw[i][2] + layer_thw[i][2] % 4) // 2
|
|
# layer_thw.append([t, h, w])
|
|
# self.layer_thw = layer_thw
|
|
self.encoder_level_1 = nn.ModuleList(
|
|
[
|
|
BasicTransformerBlock(
|
|
self.inner_dim,
|
|
self.config.num_attention_heads,
|
|
self.config.attention_head_dim,
|
|
downsampler=self.config.downsampler,
|
|
mlp_ratio=self.config.mlp_ratio,
|
|
dropout=self.config.dropout,
|
|
cross_attention_dim=self.inner_dim,
|
|
activation_fn=self.config.activation_fn,
|
|
num_embeds_ada_norm=self.config.num_embeds_ada_norm,
|
|
attention_bias=self.config.attention_bias,
|
|
only_cross_attention=self.config.only_cross_attention,
|
|
double_self_attention=self.config.double_self_attention,
|
|
upcast_attention=self.config.upcast_attention,
|
|
norm_type=norm_type,
|
|
norm_elementwise_affine=self.config.norm_elementwise_affine,
|
|
norm_eps=self.config.norm_eps,
|
|
attention_type=self.config.attention_type,
|
|
attention_mode=self.config.attention_mode,
|
|
use_rope=self.config.use_rope,
|
|
interpolation_scale_thw=interpolation_scale_thw,
|
|
)
|
|
for _ in range(self.config.depth[0])
|
|
]
|
|
)
|
|
# self.down1_2 = Downsample3d(self.inner_dim) if is_video_model else Downsample2d(self.inner_dim)
|
|
self.down1_2 = Downsample2d(self.inner_dim)
|
|
|
|
self.encoder_level_2 = nn.ModuleList(
|
|
[
|
|
BasicTransformerBlock(
|
|
self.inner_dim * 2,
|
|
self.config.num_attention_heads,
|
|
self.config.attention_head_dim * 2,
|
|
downsampler=self.config.downsampler,
|
|
mlp_ratio=self.config.mlp_ratio,
|
|
dropout=self.config.dropout,
|
|
cross_attention_dim=self.inner_dim * 2,
|
|
activation_fn=self.config.activation_fn,
|
|
num_embeds_ada_norm=self.config.num_embeds_ada_norm,
|
|
attention_bias=self.config.attention_bias,
|
|
only_cross_attention=self.config.only_cross_attention,
|
|
double_self_attention=self.config.double_self_attention,
|
|
upcast_attention=self.config.upcast_attention,
|
|
norm_type=norm_type,
|
|
norm_elementwise_affine=self.config.norm_elementwise_affine,
|
|
norm_eps=self.config.norm_eps,
|
|
attention_type=self.config.attention_type,
|
|
attention_mode=self.config.attention_mode,
|
|
use_rope=self.config.use_rope,
|
|
interpolation_scale_thw=interpolation_scale_thw,
|
|
)
|
|
for _ in range(self.config.depth[1])
|
|
]
|
|
)
|
|
# self.down2_3 = Downsample3d(self.inner_dim * 2) if is_video_model else Downsample2d(self.inner_dim * 2)
|
|
self.down2_3 = Downsample2d(self.inner_dim * 2)
|
|
|
|
self.latent = nn.ModuleList(
|
|
[
|
|
BasicTransformerBlock(
|
|
self.inner_dim * 4,
|
|
self.config.num_attention_heads,
|
|
self.config.attention_head_dim * 4,
|
|
downsampler=self.config.downsampler,
|
|
mlp_ratio=self.config.mlp_ratio,
|
|
dropout=self.config.dropout,
|
|
cross_attention_dim=self.inner_dim * 4,
|
|
activation_fn=self.config.activation_fn,
|
|
num_embeds_ada_norm=self.config.num_embeds_ada_norm,
|
|
attention_bias=self.config.attention_bias,
|
|
only_cross_attention=self.config.only_cross_attention,
|
|
double_self_attention=self.config.double_self_attention,
|
|
upcast_attention=self.config.upcast_attention,
|
|
norm_type=norm_type,
|
|
norm_elementwise_affine=self.config.norm_elementwise_affine,
|
|
norm_eps=self.config.norm_eps,
|
|
attention_type=self.config.attention_type,
|
|
attention_mode=self.config.attention_mode,
|
|
use_rope=self.config.use_rope,
|
|
interpolation_scale_thw=interpolation_scale_thw,
|
|
)
|
|
for _ in range(self.config.depth[2])
|
|
]
|
|
)
|
|
|
|
# self.up3_2 = Upsample3d(int(self.inner_dim * 4)) if is_video_model else Upsample2d(self.inner_dim * 4) ## From Level 4 to Level 3
|
|
self.up3_2 = Upsample2d(self.inner_dim * 4) ## From Level 4 to Level 3
|
|
|
|
self.reduce_chan_level2_norm = nn.LayerNorm(int(self.inner_dim * 4), elementwise_affine=True, eps=1e-6)
|
|
self.reduce_chan_level2 = nn.Linear(int(self.inner_dim * 4), int(self.inner_dim * 2), bias=True)
|
|
self.decoder_level_2 = nn.ModuleList(
|
|
[
|
|
BasicTransformerBlock(
|
|
self.inner_dim * 2,
|
|
self.config.num_attention_heads,
|
|
self.config.attention_head_dim * 2,
|
|
downsampler=self.config.downsampler,
|
|
mlp_ratio=self.config.mlp_ratio,
|
|
dropout=self.config.dropout,
|
|
cross_attention_dim=self.inner_dim * 2,
|
|
activation_fn=self.config.activation_fn,
|
|
num_embeds_ada_norm=self.config.num_embeds_ada_norm,
|
|
attention_bias=self.config.attention_bias,
|
|
only_cross_attention=self.config.only_cross_attention,
|
|
double_self_attention=self.config.double_self_attention,
|
|
upcast_attention=self.config.upcast_attention,
|
|
norm_type=norm_type,
|
|
norm_elementwise_affine=self.config.norm_elementwise_affine,
|
|
norm_eps=self.config.norm_eps,
|
|
attention_type=self.config.attention_type,
|
|
attention_mode=self.config.attention_mode,
|
|
use_rope=self.config.use_rope,
|
|
interpolation_scale_thw=interpolation_scale_thw,
|
|
)
|
|
for _ in range(self.config.depth[3])
|
|
]
|
|
)
|
|
|
|
# self.up2_1 = Upsample3d(int(self.inner_dim * 2)) if is_video_model else Upsample2d(self.inner_dim * 2) ## From Level 4 to Level 3
|
|
self.up2_1 = Upsample2d(self.inner_dim * 2) ## From Level 4 to Level 3
|
|
|
|
self.reduce_chan_level1_norm = nn.LayerNorm(int(self.inner_dim * 2), elementwise_affine=True, eps=1e-6)
|
|
self.reduce_chan_level1 = nn.Linear(int(self.inner_dim * 2), int(self.inner_dim * 1), bias=True)
|
|
self.decoder_level_1 = nn.ModuleList(
|
|
[
|
|
BasicTransformerBlock(
|
|
self.inner_dim,
|
|
self.config.num_attention_heads,
|
|
self.config.attention_head_dim,
|
|
downsampler=self.config.downsampler,
|
|
mlp_ratio=self.config.mlp_ratio,
|
|
dropout=self.config.dropout,
|
|
cross_attention_dim=self.inner_dim,
|
|
activation_fn=self.config.activation_fn,
|
|
num_embeds_ada_norm=self.config.num_embeds_ada_norm,
|
|
attention_bias=self.config.attention_bias,
|
|
only_cross_attention=self.config.only_cross_attention,
|
|
double_self_attention=self.config.double_self_attention,
|
|
upcast_attention=self.config.upcast_attention,
|
|
norm_type=norm_type,
|
|
norm_elementwise_affine=self.config.norm_elementwise_affine,
|
|
norm_eps=self.config.norm_eps,
|
|
attention_type=self.config.attention_type,
|
|
attention_mode=self.config.attention_mode,
|
|
use_rope=self.config.use_rope,
|
|
interpolation_scale_thw=interpolation_scale_thw,
|
|
)
|
|
for _ in range(self.config.depth[4])
|
|
]
|
|
)
|
|
|
|
if self.config.norm_type != "ada_norm_single":
|
|
self.norm_out = nn.LayerNorm(2 * self.inner_dim, elementwise_affine=False, eps=1e-6)
|
|
self.proj_out_1 = nn.Linear(2 * self.inner_dim, 2 * self.inner_dim)
|
|
self.proj_out_2 = nn.Linear(
|
|
2 * self.inner_dim, self.config.patch_size_t * self.config.patch_size * self.config.patch_size * self.out_channels
|
|
)
|
|
elif self.config.norm_type == "ada_norm_single":
|
|
self.norm_out = nn.LayerNorm(self.inner_dim, elementwise_affine=False, eps=1e-6)
|
|
self.scale_shift_table = nn.Parameter(torch.randn(2, self.inner_dim) / (self.inner_dim)**0.5)
|
|
self.proj_out = nn.Linear(
|
|
self.inner_dim, self.config.patch_size_t * self.config.patch_size * self.config.patch_size * self.out_channels
|
|
)
|
|
|
|
# PixArt-Alpha blocks.
|
|
# self.adaln_single = None
|
|
# if self.config.norm_type == "ada_norm_single":
|
|
# TODO(Sayak, PVP) clean this, for now we use sample size to determine whether to use
|
|
# additional conditions until we find better name
|
|
self.adaln_single_1 = AdaLayerNormSingle(
|
|
self.inner_dim, use_additional_conditions=self.use_additional_conditions
|
|
)
|
|
self.adaln_single_2 = AdaLayerNormSingle(
|
|
self.inner_dim * 2, use_additional_conditions=self.use_additional_conditions
|
|
)
|
|
self.adaln_single_3 = AdaLayerNormSingle(
|
|
self.inner_dim * 4, use_additional_conditions=self.use_additional_conditions
|
|
)
|
|
|
|
# self.caption_projection = None
|
|
# if self.caption_channels is not None:
|
|
self.caption_projection_1 = PixArtAlphaTextProjection(
|
|
in_features=self.caption_channels, hidden_size=self.inner_dim
|
|
)
|
|
self.caption_projection_2 = PixArtAlphaTextProjection(
|
|
in_features=self.caption_channels, hidden_size=self.inner_dim * 2
|
|
)
|
|
self.caption_projection_3 = PixArtAlphaTextProjection(
|
|
in_features=self.caption_channels, hidden_size=self.inner_dim * 4
|
|
)
|
|
|
|
def _replace_fp32_layers(self, module=None):
|
|
if module is None:
|
|
module = self
|
|
for name, submodule in module.named_children():
|
|
if isinstance(submodule, nn.LayerNorm):
|
|
# print(f"Replacing LayerNorm in {name}")
|
|
new_layer = FP32_Layernorm(submodule.normalized_shape, submodule.eps, submodule.elementwise_affine)
|
|
if submodule.elementwise_affine:
|
|
new_layer.weight.data.copy_(submodule.weight.data.float())
|
|
if submodule.bias is not None:
|
|
new_layer.bias.data.copy_(submodule.bias.data.float())
|
|
setattr(module, name, new_layer)
|
|
elif isinstance(submodule, nn.SiLU):
|
|
# print(f"Replacing SiLU in {name}")
|
|
setattr(module, name, FP32_SiLU(submodule.inplace))
|
|
elif isinstance(submodule, nn.GELU):
|
|
# print(f"Replacing GELU in {name}")
|
|
setattr(module, name, FP32_GELU(submodule.approximate))
|
|
else:
|
|
self._replace_fp32_layers(submodule)
|
|
|
|
def _set_gradient_checkpointing(self, module, value=False):
|
|
if hasattr(module, "gradient_checkpointing"):
|
|
module.gradient_checkpointing = value
|
|
|
|
def forward(
|
|
self,
|
|
hidden_states: torch.Tensor,
|
|
timestep: Optional[torch.LongTensor] = None,
|
|
encoder_hidden_states: Optional[torch.Tensor] = None,
|
|
added_cond_kwargs: Dict[str, torch.Tensor] = None,
|
|
class_labels: Optional[torch.LongTensor] = None,
|
|
cross_attention_kwargs: Dict[str, Any] = None,
|
|
attention_mask: Optional[torch.Tensor] = None,
|
|
encoder_attention_mask: Optional[torch.Tensor] = None,
|
|
use_image_num: Optional[int] = 0,
|
|
return_dict: bool = True,
|
|
):
|
|
"""
|
|
The [`Transformer2DModel`] forward method.
|
|
|
|
Args:
|
|
hidden_states (`torch.LongTensor` of shape `(batch size, num latent pixels)` if discrete, `torch.FloatTensor` of shape `(batch size, channel, height, width)` if continuous):
|
|
Input `hidden_states`.
|
|
encoder_hidden_states ( `torch.FloatTensor` of shape `(batch size, sequence len, embed dims)`, *optional*):
|
|
Conditional embeddings for cross attention layer. If not given, cross-attention defaults to
|
|
self-attention.
|
|
timestep ( `torch.LongTensor`, *optional*):
|
|
Used to indicate denoising step. Optional timestep to be applied as an embedding in `AdaLayerNorm`.
|
|
class_labels ( `torch.LongTensor` of shape `(batch size, num classes)`, *optional*):
|
|
Used to indicate class labels conditioning. Optional class labels to be applied as an embedding in
|
|
`AdaLayerZeroNorm`.
|
|
cross_attention_kwargs ( `Dict[str, Any]`, *optional*):
|
|
A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
|
|
`self.processor` in
|
|
[diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
|
|
attention_mask ( `torch.Tensor`, *optional*):
|
|
An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask
|
|
is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large
|
|
negative values to the attention scores corresponding to "discard" tokens.
|
|
encoder_attention_mask ( `torch.Tensor`, *optional*):
|
|
Cross-attention mask applied to `encoder_hidden_states`. Two formats supported:
|
|
|
|
* Mask `(batch, sequence_length)` True = keep, False = discard.
|
|
* Bias `(batch, 1, sequence_length)` 0 = keep, -10000 = discard.
|
|
|
|
If `ndim == 2`: will be interpreted as a mask, then converted into a bias consistent with the format
|
|
above. This bias will be added to the cross-attention scores.
|
|
return_dict (`bool`, *optional*, defaults to `True`):
|
|
Whether or not to return a [`~models.unets.unet_2d_condition.UNet2DConditionOutput`] instead of a plain
|
|
tuple.
|
|
|
|
Returns:
|
|
If `return_dict` is True, an [`~models.transformer_2d.Transformer2DModelOutput`] is returned, otherwise a
|
|
`tuple` where the first element is the sample tensor.
|
|
"""
|
|
batch_size, c, frame, height, width = hidden_states.shape
|
|
frame = frame - use_image_num # 21-4=17
|
|
if cross_attention_kwargs is not None:
|
|
if cross_attention_kwargs.get("scale", None) is not None:
|
|
print.warning("Passing `scale` to `cross_attention_kwargs` is deprecated. `scale` will be ignored.")
|
|
# ensure attention_mask is a bias, and give it a singleton query_tokens dimension.
|
|
# we may have done this conversion already, e.g. if we came here via UNet2DConditionModel#forward.
|
|
# we can tell by counting dims; if ndim == 2: it's a mask rather than a bias.
|
|
# expects mask of shape:
|
|
# [batch, key_tokens]
|
|
# adds singleton query_tokens dimension:
|
|
# [batch, 1, key_tokens]
|
|
# this helps to broadcast it as a bias over attention scores, which will be in one of the following shapes:
|
|
# [batch, heads, query_tokens, key_tokens] (e.g. torch sdp attn)
|
|
# [batch * heads, query_tokens, key_tokens] (e.g. xformers or classic attn)
|
|
|
|
if attention_mask is not None and attention_mask.ndim == 4:
|
|
# assume that mask is expressed as:
|
|
# (1 = keep, 0 = discard)
|
|
# convert mask into a bias that can be added to attention scores:
|
|
# (keep = +0, discard = -10000.0)
|
|
# b, frame+use_image_num, h, w -> a video with images
|
|
# b, 1, h, w -> only images
|
|
pad_h_0, pad_w_0 = height % (self.config.patch_size * 2), width % (self.config.patch_size * 2)
|
|
|
|
hidden_states = F.pad(hidden_states, (0, pad_w_0, 0, pad_h_0, 0, 0), mode='reflect')
|
|
attention_mask = attention_mask.to(self.dtype)
|
|
attention_mask = attention_mask.unsqueeze(1) # b 1 t h w
|
|
attention_mask = F.pad(attention_mask, (0, pad_w_0, 0, pad_h_0, 0, 0))
|
|
attention_mask = F.max_pool3d(attention_mask, kernel_size=(self.config.patch_size_t, self.config.patch_size, self.config.patch_size),
|
|
stride=(self.config.patch_size_t, self.config.patch_size, self.config.patch_size))
|
|
attention_mask = rearrange(attention_mask, 'b 1 t h w -> b 1 (t h w)')
|
|
|
|
attention_bias = (1 - attention_mask.bool().to(self.dtype)) * -10000.0
|
|
|
|
# convert encoder_attention_mask to a bias the same way we do for attention_mask
|
|
if encoder_attention_mask is not None and encoder_attention_mask.ndim == 3:
|
|
# b, 1, l -> only video
|
|
encoder_attention_mask = (1 - encoder_attention_mask.to(self.dtype)) * -10000.0
|
|
|
|
|
|
# 1. Input
|
|
added_cond_kwargs = {"resolution": None, "aspect_ratio": None}
|
|
hidden_states, encoder_hidden_states_1, encoder_hidden_states_2, encoder_hidden_states_3, \
|
|
timestep_1, timestep_2, timestep_3, \
|
|
embedded_timestep_1, embedded_timestep_2, embedded_timestep_3 = self._operate_on_patched_inputs(
|
|
hidden_states, encoder_hidden_states, timestep, added_cond_kwargs, batch_size, frame, use_image_num
|
|
)
|
|
frame, height, width = frame // self.config.patch_size_t, \
|
|
(height + pad_h_0) // (self.config.patch_size), (width + pad_w_0) // (self.config.patch_size)
|
|
|
|
|
|
assert not torch.any(torch.isnan(hidden_states)), 'after _operate_on_patched_inputs'
|
|
def create_custom_forward(module, return_dict=None):
|
|
def custom_forward(*inputs):
|
|
if return_dict is not None:
|
|
return module(*inputs, return_dict=return_dict)
|
|
else:
|
|
return module(*inputs)
|
|
|
|
return custom_forward
|
|
|
|
# encoder_1
|
|
out_enc_level1 = hidden_states
|
|
# import ipdb;ipdb.set_trace()
|
|
if self.training and self.gradient_checkpointing:
|
|
|
|
ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
|
|
for block in self.encoder_level_1:
|
|
out_enc_level1 = torch.utils.checkpoint.checkpoint(
|
|
create_custom_forward(block),
|
|
out_enc_level1,
|
|
attention_bias,
|
|
encoder_hidden_states_1,
|
|
encoder_attention_mask,
|
|
timestep_1,
|
|
cross_attention_kwargs,
|
|
class_labels,
|
|
frame,
|
|
height,
|
|
width,
|
|
**ckpt_kwargs,
|
|
)
|
|
else:
|
|
for block in self.encoder_level_1:
|
|
out_enc_level1 = block(
|
|
out_enc_level1,
|
|
attention_mask=attention_bias,
|
|
encoder_hidden_states=encoder_hidden_states_1,
|
|
encoder_attention_mask=encoder_attention_mask,
|
|
timestep=timestep_1,
|
|
cross_attention_kwargs=cross_attention_kwargs,
|
|
class_labels=class_labels,
|
|
frame=frame,
|
|
height=height,
|
|
width=width,
|
|
)
|
|
pad_h_1, pad_w_1 = height % 4, width % 4
|
|
|
|
inp_enc_level2, attention_bias, attention_mask = self.down1_2(out_enc_level1, attention_mask, frame, height, width, pad_h=pad_h_1, pad_w=pad_w_1)
|
|
# frame, height, width = frame // 2 if frame != 1 else frame, (height + pad_h_1) // 2, (width + pad_w_1) // 2
|
|
height, width = (height + pad_h_1) // 2, (width + pad_w_1) // 2
|
|
|
|
# encoder_2
|
|
out_enc_level2 = inp_enc_level2
|
|
|
|
if self.training and self.gradient_checkpointing:
|
|
|
|
ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
|
|
|
|
for block in self.encoder_level_2:
|
|
out_enc_level2 = torch.utils.checkpoint.checkpoint(
|
|
create_custom_forward(block),
|
|
out_enc_level2,
|
|
attention_bias,
|
|
encoder_hidden_states_2,
|
|
encoder_attention_mask,
|
|
timestep_2,
|
|
cross_attention_kwargs,
|
|
class_labels,
|
|
frame,
|
|
height,
|
|
width,
|
|
**ckpt_kwargs,
|
|
)
|
|
else:
|
|
for block in self.encoder_level_2:
|
|
out_enc_level2 = block(
|
|
out_enc_level2,
|
|
attention_mask=attention_bias,
|
|
encoder_hidden_states=encoder_hidden_states_2,
|
|
encoder_attention_mask=encoder_attention_mask,
|
|
timestep=timestep_2,
|
|
cross_attention_kwargs=cross_attention_kwargs,
|
|
class_labels=class_labels,
|
|
frame=frame,
|
|
height=height,
|
|
width=width,
|
|
)
|
|
pad_h_2, pad_w_2 = height % 4, width % 4
|
|
|
|
# import ipdb;ipdb.set_trace()
|
|
inp_enc_level3, attention_bias, attention_mask = self.down2_3(out_enc_level2, attention_mask, frame, height, width, pad_h=pad_h_2, pad_w=pad_w_2)
|
|
# frame, height, width = frame // 2 if frame != 1 else frame, (height + pad_h_2) // 2, (width + pad_w_2) // 2
|
|
height, width = (height + pad_h_2) // 2, (width + pad_w_2) // 2
|
|
|
|
# latent
|
|
latent = inp_enc_level3
|
|
if self.training and self.gradient_checkpointing:
|
|
|
|
ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
|
|
|
|
for block in self.latent:
|
|
latent = torch.utils.checkpoint.checkpoint(
|
|
create_custom_forward(block),
|
|
latent,
|
|
attention_bias,
|
|
encoder_hidden_states_3,
|
|
encoder_attention_mask,
|
|
timestep_3,
|
|
cross_attention_kwargs,
|
|
class_labels,
|
|
frame,
|
|
height,
|
|
width,
|
|
**ckpt_kwargs,
|
|
)
|
|
else:
|
|
for block in self.latent:
|
|
latent = block(
|
|
latent,
|
|
attention_mask=attention_bias,
|
|
encoder_hidden_states=encoder_hidden_states_3,
|
|
encoder_attention_mask=encoder_attention_mask,
|
|
timestep=timestep_3,
|
|
cross_attention_kwargs=cross_attention_kwargs,
|
|
class_labels=class_labels,
|
|
frame=frame,
|
|
height=height,
|
|
width=width,
|
|
)
|
|
|
|
# decoder_2
|
|
|
|
# import ipdb;ipdb.set_trace()
|
|
inp_dec_level2, attention_bias, attention_mask = self.up3_2(latent, attention_mask, frame, height, width, pad_h=pad_h_2, pad_w=pad_w_2)
|
|
# frame, height, width = frame * 2 if frame != 1 else frame, height * 2 - pad_h_2, width * 2 - pad_w_2
|
|
height, width = height * 2 - pad_h_2, width * 2 - pad_w_2
|
|
inp_dec_level2 = torch.cat([inp_dec_level2, out_enc_level2], 2)
|
|
inp_dec_level2 = self.reduce_chan_level2_norm(inp_dec_level2)
|
|
inp_dec_level2 = self.reduce_chan_level2(inp_dec_level2)
|
|
out_dec_level2 = inp_dec_level2
|
|
|
|
if self.training and self.gradient_checkpointing:
|
|
|
|
ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
|
|
|
|
for block in self.decoder_level_2:
|
|
out_dec_level2 = torch.utils.checkpoint.checkpoint(
|
|
create_custom_forward(block),
|
|
out_dec_level2,
|
|
attention_bias,
|
|
encoder_hidden_states_2,
|
|
encoder_attention_mask,
|
|
timestep_2,
|
|
cross_attention_kwargs,
|
|
class_labels,
|
|
frame,
|
|
height,
|
|
width,
|
|
**ckpt_kwargs,
|
|
)
|
|
else:
|
|
for block in self.decoder_level_2:
|
|
out_dec_level2 = block(
|
|
out_dec_level2,
|
|
attention_mask=attention_bias,
|
|
encoder_hidden_states=encoder_hidden_states_2,
|
|
encoder_attention_mask=encoder_attention_mask,
|
|
timestep=timestep_2,
|
|
cross_attention_kwargs=cross_attention_kwargs,
|
|
class_labels=class_labels,
|
|
frame=frame,
|
|
height=height,
|
|
width=width,
|
|
)
|
|
|
|
# decoder_1
|
|
|
|
# import ipdb;ipdb.set_trace()
|
|
inp_dec_level1, attention_bias, attention_mask = self.up2_1(out_dec_level2, attention_mask, frame, height, width, pad_h=pad_h_1, pad_w=pad_w_1)
|
|
# frame, height, width = frame * 2 if frame != 1 else frame, height * 2 - pad_h_1, width * 2 - pad_w_1
|
|
height, width = height * 2 - pad_h_1, width * 2 - pad_w_1
|
|
inp_dec_level1 = torch.cat([inp_dec_level1, out_enc_level1], 2)
|
|
inp_dec_level1 = self.reduce_chan_level1_norm(inp_dec_level1)
|
|
inp_dec_level1 = self.reduce_chan_level1(inp_dec_level1)
|
|
out_dec_level1 = inp_dec_level1
|
|
|
|
if self.training and self.gradient_checkpointing:
|
|
|
|
ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
|
|
|
|
for block in self.decoder_level_1:
|
|
out_dec_level1 = torch.utils.checkpoint.checkpoint(
|
|
create_custom_forward(block),
|
|
out_dec_level1,
|
|
attention_bias,
|
|
encoder_hidden_states_1,
|
|
encoder_attention_mask,
|
|
timestep_1,
|
|
cross_attention_kwargs,
|
|
class_labels,
|
|
frame,
|
|
height,
|
|
width,
|
|
**ckpt_kwargs,
|
|
)
|
|
else:
|
|
for block in self.decoder_level_1:
|
|
out_dec_level1 = block(
|
|
out_dec_level1,
|
|
attention_mask=attention_bias,
|
|
encoder_hidden_states=encoder_hidden_states_1,
|
|
encoder_attention_mask=encoder_attention_mask,
|
|
timestep=timestep_1,
|
|
cross_attention_kwargs=cross_attention_kwargs,
|
|
class_labels=class_labels,
|
|
frame=frame,
|
|
height=height,
|
|
width=width,
|
|
)
|
|
|
|
assert not torch.any(torch.isnan(out_dec_level1)), 'after out_dec_level1'
|
|
# 3. Output
|
|
output = self._get_output_for_patched_inputs(
|
|
hidden_states=out_dec_level1,
|
|
timestep=timestep_1,
|
|
class_labels=class_labels,
|
|
embedded_timestep=embedded_timestep_1,
|
|
num_frames=frame,
|
|
height=height,
|
|
width=width,
|
|
) # b c t h w
|
|
|
|
assert not torch.any(torch.isnan(output)), 'after output'
|
|
frame, height, width = frame * self.config.patch_size_t, height * self.config.patch_size - pad_h_0, width * self.config.patch_size - pad_w_0
|
|
output = output[:, :, :frame, :height, :width]
|
|
if not return_dict:
|
|
return (output,)
|
|
|
|
return Transformer2DModelOutput(sample=output)
|
|
|
|
|
|
def _operate_on_patched_inputs(self, hidden_states, encoder_hidden_states, timestep, added_cond_kwargs, batch_size, frame, use_image_num):
|
|
# batch_size = hidden_states.shape[0]
|
|
hidden_states = self.pos_embed(hidden_states.to(self.dtype), frame)
|
|
|
|
if self.use_additional_conditions and added_cond_kwargs is None:
|
|
raise ValueError(
|
|
"`added_cond_kwargs` cannot be None when using additional conditions for `adaln_single`."
|
|
)
|
|
timestep_1, embedded_timestep_1 = self.adaln_single_1(
|
|
timestep, added_cond_kwargs, batch_size=batch_size, hidden_dtype=self.dtype
|
|
) # b 6d, b d
|
|
timestep_2, embedded_timestep_2 = self.adaln_single_2(
|
|
timestep, added_cond_kwargs, batch_size=batch_size, hidden_dtype=self.dtype
|
|
) # b 6d, b d
|
|
timestep_3, embedded_timestep_3 = self.adaln_single_3(
|
|
timestep, added_cond_kwargs, batch_size=batch_size, hidden_dtype=self.dtype
|
|
) # b 6d, b d
|
|
|
|
encoder_hidden_states_1 = self.caption_projection_1(encoder_hidden_states) # b, 1+use_image_num, l, d or b, 1, l, d
|
|
encoder_hidden_states_1 = rearrange(encoder_hidden_states_1[:, :1], 'b 1 l d -> (b 1) l d')
|
|
encoder_hidden_states_2 = self.caption_projection_2(encoder_hidden_states) # b, 1+use_image_num, l, d or b, 1, l, d
|
|
encoder_hidden_states_2 = rearrange(encoder_hidden_states_2[:, :1], 'b 1 l d -> (b 1) l d')
|
|
encoder_hidden_states_3 = self.caption_projection_3(encoder_hidden_states) # b, 1+use_image_num, l, d or b, 1, l, d
|
|
encoder_hidden_states_3 = rearrange(encoder_hidden_states_3[:, :1], 'b 1 l d -> (b 1) l d')
|
|
|
|
|
|
return hidden_states, encoder_hidden_states_1, encoder_hidden_states_2, encoder_hidden_states_3, \
|
|
timestep_1, timestep_2, timestep_3, embedded_timestep_1, embedded_timestep_2, embedded_timestep_3
|
|
|
|
|
|
|
|
def _get_output_for_patched_inputs(
|
|
self, hidden_states, timestep, class_labels, embedded_timestep, num_frames, height=None, width=None
|
|
):
|
|
if self.config.norm_type != "ada_norm_single":
|
|
conditioning = self.transformer_blocks[0].norm1.emb(
|
|
timestep, class_labels, hidden_dtype=self.dtype
|
|
)
|
|
shift, scale = self.proj_out_1(F.silu(conditioning)).chunk(2, dim=1)
|
|
hidden_states = self.norm_out(hidden_states) * (1 + scale[:, None]) + shift[:, None]
|
|
hidden_states = self.proj_out_2(hidden_states)
|
|
elif self.config.norm_type == "ada_norm_single":
|
|
shift, scale = (self.scale_shift_table[None] + embedded_timestep[:, None]).chunk(2, dim=1)
|
|
hidden_states = self.norm_out(hidden_states)
|
|
# Modulation
|
|
hidden_states = hidden_states * (1 + scale) + shift
|
|
hidden_states = self.proj_out(hidden_states)
|
|
hidden_states = hidden_states.squeeze(1)
|
|
|
|
# # unpatchify
|
|
# hidden_states = hidden_states.reshape(
|
|
# shape=(-1, num_frames, height, width, self.out_channels)
|
|
# )
|
|
# output = torch.einsum("nthwc->ncthw", hidden_states)
|
|
# return output
|
|
# unpatchify
|
|
hidden_states = hidden_states.reshape(
|
|
shape=(-1, num_frames, height, width, self.config.patch_size_t, self.config.patch_size, self.config.patch_size, self.config.out_channels)
|
|
)
|
|
hidden_states = torch.einsum("nthwopqc->nctohpwq", hidden_states)
|
|
output = hidden_states.reshape(
|
|
shape=(-1, self.config.out_channels, num_frames * self.config.patch_size_t, height * self.config.patch_size, width * self.config.patch_size)
|
|
)
|
|
# import ipdb;ipdb.set_trace()
|
|
# if output.shape[2] % 2 == 0:
|
|
# output = output[:, :, 1:]
|
|
return output
|
|
|
|
|
|
def UDiTT2V_S_111(**kwargs):
|
|
return UDiTT2V(depth=[2, 5, 8, 5, 2], attention_head_dim=24, num_attention_heads=16, patch_size_t=1, patch_size=1,
|
|
mlp_ratio=2, norm_type="ada_norm_single", caption_channels=4096, **kwargs)
|
|
|
|
def UDiTT2V_S_122(**kwargs):
|
|
return UDiTT2V(depth=[2, 5, 8, 5, 2], attention_head_dim=24, num_attention_heads=16, patch_size_t=1, patch_size=2,
|
|
mlp_ratio=2, norm_type="ada_norm_single", caption_channels=4096, **kwargs)
|
|
|
|
def UDiTT2V_B_111(**kwargs):
|
|
return UDiTT2V(depth=[2, 5, 10, 5, 2], attention_head_dim=24, num_attention_heads=24, patch_size_t=1, patch_size=1,
|
|
mlp_ratio=2, norm_type="ada_norm_single", caption_channels=4096, **kwargs)
|
|
|
|
def UDiTT2V_B_122(**kwargs):
|
|
return UDiTT2V(depth=[2, 5, 10, 5, 2], attention_head_dim=24, num_attention_heads=24, patch_size_t=1, patch_size=2,
|
|
mlp_ratio=2, norm_type="ada_norm_single", caption_channels=4096, **kwargs)
|
|
|
|
def UDiTT2V_L_111(**kwargs):
|
|
return UDiTT2V(depth=[4, 8, 12, 8, 4], attention_head_dim=24, num_attention_heads=32, patch_size_t=1, patch_size=1,
|
|
mlp_ratio=2, norm_type="ada_norm_single", caption_channels=4096, **kwargs)
|
|
|
|
def UDiTT2V_L_211(**kwargs):
|
|
return UDiTT2V(depth=[4, 8, 12, 8, 4], attention_head_dim=24, num_attention_heads=32, patch_size_t=2, patch_size=1,
|
|
mlp_ratio=2, norm_type="ada_norm_single", caption_channels=4096, **kwargs)
|
|
|
|
def UDiTT2V_L_122(**kwargs):
|
|
return UDiTT2V(depth=[4, 8, 12, 8, 4], attention_head_dim=24, num_attention_heads=32, patch_size_t=1, patch_size=2,
|
|
mlp_ratio=2, norm_type="ada_norm_single", caption_channels=4096, **kwargs)
|
|
|
|
def UDiTT2V_L_222(**kwargs):
|
|
return UDiTT2V(depth=[4, 8, 12, 8, 4], attention_head_dim=24, num_attention_heads=32, patch_size_t=2, patch_size=2,
|
|
mlp_ratio=2, norm_type="ada_norm_single", caption_channels=4096, **kwargs)
|
|
|
|
def UDiTT2V_XL_111(**kwargs):
|
|
return UDiTT2V(depth=[4, 8, 12, 8, 4], attention_head_dim=24, num_attention_heads=48, patch_size_t=1, patch_size=1,
|
|
mlp_ratio=2, norm_type="ada_norm_single", caption_channels=4096, **kwargs)
|
|
|
|
|
|
UDiT_models = {
|
|
"UDiTT2V-S/111": UDiTT2V_S_111, # 0.4B 0.7B if video
|
|
"UDiTT2V-S/122": UDiTT2V_S_122, # 0.4B 0.7B if video
|
|
"UDiTT2V-B/111": UDiTT2V_B_111, # 1.0B 1.6B if video
|
|
"UDiTT2V-B/122": UDiTT2V_B_122, # 1.0B 1.6B if video
|
|
"UDiTT2V-L/111": UDiTT2V_L_111, # 2.2B 3.3B if video
|
|
"UDiTT2V-L/211": UDiTT2V_L_211, # 2.2B 3.3B if video
|
|
"UDiTT2V-L/122": UDiTT2V_L_122, # 2.2B 3.3B if video
|
|
"UDiTT2V-L/222": UDiTT2V_L_222, # 2.2B 3.3B if video
|
|
"UDiTT2V-XL/111": UDiTT2V_XL_111, # 5.0B 7.4B if video
|
|
}
|
|
|
|
UDiT_models_class = {
|
|
"UDiTT2V-S/111": UDiTT2V,
|
|
"UDiTT2V-S/122": UDiTT2V,
|
|
"UDiTT2V-B/111": UDiTT2V,
|
|
"UDiTT2V-B/122": UDiTT2V,
|
|
"UDiTT2V-L/111": UDiTT2V,
|
|
"UDiTT2V-L/211": UDiTT2V,
|
|
"UDiTT2V-L/122": UDiTT2V,
|
|
"UDiTT2V-L/222": UDiTT2V,
|
|
"UDiTT2V-XL/111": UDiTT2V,
|
|
}
|
|
|
|
|
|
if __name__ == '__main__':
|
|
import sys
|
|
from copy import deepcopy
|
|
from opensora.models.causalvideovae import ae_channel_config, ae_stride_config
|
|
from opensora.models.causalvideovae import getae, getae_wrapper
|
|
from opensora.models.causalvideovae import CausalVAEModelWrapper
|
|
|
|
args = type('args', (),
|
|
{
|
|
'ae': 'CausalVAEModel_4x8x8',
|
|
'attention_mode': 'xformers',
|
|
'use_rope': True,
|
|
'model_max_length': 300,
|
|
'max_height': 240,
|
|
'max_width': 320,
|
|
'num_frames': 1,
|
|
'use_image_num': 0,
|
|
'compress_kv_factor': 1,
|
|
'interpolation_scale_t': 1,
|
|
'interpolation_scale_h': 1,
|
|
'interpolation_scale_w': 1,
|
|
}
|
|
)
|
|
b = 16
|
|
c = 4
|
|
cond_c = 4096
|
|
num_timesteps = 1000
|
|
ae_stride_t, ae_stride_h, ae_stride_w = ae_stride_config[args.ae]
|
|
latent_size = (args.max_height // ae_stride_h, args.max_width // ae_stride_w)
|
|
num_frames = (args.num_frames - 1) // ae_stride_t + 1
|
|
|
|
device = torch.device('cuda:0')
|
|
|
|
|
|
|
|
model = UDiTT2V_L_122(in_channels=c,
|
|
out_channels=c,
|
|
sample_size=latent_size,
|
|
sample_size_t=num_frames,
|
|
activation_fn="gelu-approximate",
|
|
attention_bias=True,
|
|
attention_type="default",
|
|
double_self_attention=False,
|
|
norm_elementwise_affine=False,
|
|
norm_eps=1e-06,
|
|
norm_num_groups=32,
|
|
num_vector_embeds=None,
|
|
only_cross_attention=False,
|
|
upcast_attention=False,
|
|
use_linear_projection=False,
|
|
use_additional_conditions=False,
|
|
downsampler=None,
|
|
interpolation_scale_t=args.interpolation_scale_t,
|
|
interpolation_scale_h=args.interpolation_scale_h,
|
|
interpolation_scale_w=args.interpolation_scale_w,
|
|
use_rope=args.use_rope).to(device)
|
|
|
|
print(model)
|
|
print(f'{sum(p.numel() for p in model.parameters() if p.requires_grad)/1e9} B')
|
|
|
|
|
|
model_state_dict = model.state_dict()
|
|
pretrained = "/storage/ongoing/new/Open-Sora-Plan/bs2_20node_73000k_480p_61x480p_lr5e-5_snr5_noioff0.02_ema_rope_uditultra122_qknorm_ds222_mt5xxl_sucai288w/checkpoint-11500/model_ema/diffusion_pytorch_model.safetensors"
|
|
try:
|
|
if 'safetensors' in pretrained: # pixart series
|
|
from safetensors.torch import load_file as safe_load
|
|
# import ipdb;ipdb.set_trace()
|
|
pretrained_checkpoint = safe_load(pretrained, device="cpu")
|
|
pretrained_keys = set(list(pretrained_checkpoint.keys()))
|
|
model_keys = set(list(model_state_dict.keys()))
|
|
common_keys = list(pretrained_keys & model_keys)
|
|
checkpoint = {k: pretrained_checkpoint[k] for k in common_keys if model_state_dict[k].numel() == pretrained_checkpoint[k].numel()}
|
|
# if checkpoint['pos_embed.proj.weight'].shape != model.pos_embed.proj.weight.shape and checkpoint['pos_embed.proj.weight'].ndim == 4:
|
|
# logger.info(f"Resize pos_embed, {checkpoint['pos_embed.proj.weight'].shape} -> {model.pos_embed.proj.weight.shape}")
|
|
# repeat = model.pos_embed.proj.weight.shape[2]
|
|
# checkpoint['pos_embed.proj.weight'] = checkpoint['pos_embed.proj.weight'].unsqueeze(2).repeat(1, 1, repeat, 1, 1) / float(repeat)
|
|
# del checkpoint['proj_out.weight'], checkpoint['proj_out.bias']
|
|
else: # latest stage training weight
|
|
checkpoint = torch.load(pretrained, map_location='cpu')
|
|
if 'model' in checkpoint:
|
|
checkpoint = checkpoint['model']
|
|
# import ipdb;ipdb.set_trace()
|
|
missing_keys, unexpected_keys = model.load_state_dict(checkpoint, strict=False)
|
|
print(f'missing_keys {len(missing_keys)} {missing_keys}, unexpected_keys {len(unexpected_keys)}')
|
|
print(f'Successfully load {len(model_state_dict) - len(missing_keys)}/{len(model_state_dict)} keys from {args.pretrained}!')
|
|
except Exception as e:
|
|
print(e)
|
|
|
|
# import sys;sys.exit()
|
|
# try:
|
|
# path = "bs32_1node_480p_lr1e-4_snr5_noioff0.02_ema_uditultra22_ds22_mt5xxl/checkpoint-500/model/diffusion_pytorch_model.safetensors"
|
|
# from safetensors.torch import load_file as safe_load
|
|
# ckpt = safe_load(path, device="cpu")
|
|
# new_ckpt = {}
|
|
# k_size = 3
|
|
# t_stride = 1
|
|
# for k, v in ckpt.items():
|
|
# if 'pos_embed.proj.weight' in k:
|
|
# new_v = v.unsqueeze(-3).repeat(1, 1, k_size, 1, 1) # 768, 4, 3, 3 -> 768, 4, 3, 3, 3
|
|
# elif 'attn1.downsampler.layer.weight' in k:
|
|
# new_v = v.unsqueeze(-3).repeat(1, 1, k_size, 1, 1) # 768, 4, 3, 3 -> 768, 4, 3, 3, 3
|
|
# elif 'body.0.weight' in k and 'down' in k:
|
|
# in_c = v.shape[0]
|
|
# new_v = v[:in_c//2].unsqueeze(-3).repeat(1, 1, k_size, 1, 1) # 384, 768, 3, 3 -> 192, 768, 3, 3, 3
|
|
# elif 'body.0.weight' in k and 'up' in k:
|
|
# new_v = v.unsqueeze(-3).repeat(2, 1, k_size, 1, 1) # 6144, 3072, 3, 3 -> 12288, 3072, 3, 3, 3
|
|
# elif 'proj_out' in k:
|
|
# if 'weight' in k:
|
|
# new_v = v.repeat(t_stride, 1) # 16, 768 -> 32, 768
|
|
# elif 'bias' in k:
|
|
# new_v = v.repeat(t_stride) # 16 -> 32
|
|
# else:
|
|
# new_v = v
|
|
# new_ckpt[k] = new_v
|
|
# msg = model.load_state_dict(new_ckpt, strict=False)
|
|
# # print(msg)
|
|
# except Exception as e:
|
|
# print(e)
|
|
x = torch.randn(b, c, 1+(args.num_frames-1)//ae_stride_t+args.use_image_num, args.max_height//ae_stride_h, args.max_width//ae_stride_w).to(device)
|
|
cond = torch.randn(b, 1+args.use_image_num, args.model_max_length, cond_c).to(device)
|
|
attn_mask = torch.randint(0, 2, (b, 1+(args.num_frames-1)//ae_stride_t+args.use_image_num, args.max_height//ae_stride_h, args.max_width//ae_stride_w)).to(device) # B L or B 1+num_images L
|
|
cond_mask = torch.randint(0, 2, (b, 1+args.use_image_num, args.model_max_length)).to(device) # B L or B 1+num_images L
|
|
timestep = torch.randint(0, 1000, (b,), device=device)
|
|
model_kwargs = dict(hidden_states=x, encoder_hidden_states=cond, attention_mask=attn_mask,
|
|
encoder_attention_mask=cond_mask, use_image_num=args.use_image_num, timestep=timestep)
|
|
with torch.no_grad():
|
|
output = model(**model_kwargs)[0]
|
|
import ipdb;ipdb.set_trace()
|
|
print(output.shape)
|
|
|
|
|
|
|
|
|
|
# from peft import LoraConfig, PeftModel, get_peft_model
|
|
# from opensora.utils.lora_utils import EMAModel_LoRA, maybe_zero_3, get_peft_state_maybe_zero_3
|
|
# lora_save_path = '/storage/ongoing/new/Open-Sora-Plan/debug_lora/model_lora'
|
|
# ema_lora_save_path = '/storage/ongoing/new/Open-Sora-Plan/debug_lora/ema_model_lora'
|
|
# origin_model_path = '/storage/ongoing/new/Open-Sora-Plan/bs16_4node_240p_lr1e-4_snr5_noioff0.02_ema_rope_uditultra22_ds22_mt5xxl/checkpoint-500/model_ema'
|
|
# model = UDiTUltraT2V.from_pretrained(origin_model_path)
|
|
# lora_config = LoraConfig(
|
|
# r=64,
|
|
# lora_alpha=64,
|
|
# init_lora_weights="gaussian",
|
|
# target_modules=["to_k", "to_q", "to_v", "to_out.0"],
|
|
# )
|
|
# model_lora = get_peft_model(model, lora_config)
|
|
# # --------------------ema lora_model----------------------------------
|
|
# # create ema lora_model
|
|
# ema_model = deepcopy(model_lora)
|
|
# ema_model_lora = EMAModel_LoRA(lora_config, parameters=ema_model.parameters(), update_after_step=0,
|
|
# model_cls=UDiTUltraT2V, model_config=ema_model.config)
|
|
# ema_model_lora.save_pretrained(ema_lora_save_path)
|
|
# ema_model_load_lora = EMAModel_LoRA.from_pretrained(ema_lora_save_path, UDiTUltraT2V, lora_config, origin_model_path)
|
|
# ema_model_lora.load_state_dict(ema_model_load_lora.state_dict())
|
|
# ema_model_lora.to(device)
|
|
|
|
# # -----------------lora model---------------------------------
|
|
# # get lora weight
|
|
# model_lora.save_pretrained(lora_save_path)
|
|
# # ----------------load lora model------------------------------
|
|
# # load lora weight
|
|
# model = UDiTUltraT2V.from_pretrained(origin_model_path)
|
|
# import ipdb;ipdb.set_trace()
|
|
# model_load_lora = PeftModel.from_pretrained(model, lora_save_path)
|
|
# for k, v in model_load_lora.state_dict().items():
|
|
# assert torch.allclose(v, model_lora.state_dict()[k])
|
|
# # for k, v in zip(ema_model_lora.shadow_params, model_lora.parameters()):
|
|
# # assert torch.allclose(v, k)
|
|
# print('Merging LoRA weights...')
|
|
# import ipdb;ipdb.set_trace()
|
|
# model_load_lora_merge = model_load_lora.merge_and_unload()
|
|
# with torch.no_grad():
|
|
# output = model_load_lora_merge(**model_kwargs)
|
|
# print(output[0].shape)
|
|
|
|
|
|
|
|
|