|
|
|
|
|
|
|
|
|
""" |
|
temporal_transformers.py |
|
|
|
This module provides classes and functions for implementing Temporal Transformers |
|
in PyTorch, designed for handling video data and temporal sequences within transformer-based models. |
|
|
|
Functions: |
|
zero_module(module) |
|
Zero out the parameters of a module and return it. |
|
|
|
Classes: |
|
TemporalTransformer3DModelOutput(BaseOutput) |
|
Dataclass for storing the output of TemporalTransformer3DModel. |
|
|
|
VanillaTemporalModule(nn.Module) |
|
A Vanilla Temporal Module class for handling temporal data. |
|
|
|
TemporalTransformer3DModel(nn.Module) |
|
A Temporal Transformer 3D Model class for transforming temporal data. |
|
|
|
TemporalTransformerBlock(nn.Module) |
|
A Temporal Transformer Block class for building the transformer architecture. |
|
|
|
PositionalEncoding(nn.Module) |
|
A Positional Encoding module for transformers to encode positional information. |
|
|
|
Dependencies: |
|
math |
|
dataclasses.dataclass |
|
typing (Callable, Optional) |
|
torch |
|
diffusers (FeedForward, Attention, AttnProcessor) |
|
diffusers.utils (BaseOutput) |
|
diffusers.utils.import_utils (is_xformers_available) |
|
einops (rearrange, repeat) |
|
torch.nn |
|
xformers |
|
xformers.ops |
|
|
|
Example Usage: |
|
>>> motion_module = get_motion_module(in_channels=512, motion_module_type="Vanilla", motion_module_kwargs={}) |
|
>>> output = motion_module(input_tensor, temb, encoder_hidden_states) |
|
|
|
This module is designed to facilitate the creation, training, and inference of transformer models |
|
that operate on temporal data, such as videos or time-series. It includes mechanisms for applying temporal attention, |
|
managing positional encoding, and integrating with external libraries for efficient attention operations. |
|
""" |
|
|
|
|
|
|
|
import math |
|
|
|
import torch |
|
import xformers |
|
import xformers.ops |
|
from diffusers.models.attention import FeedForward |
|
from diffusers.models.attention_processor import Attention, AttnProcessor |
|
from diffusers.utils import BaseOutput |
|
from diffusers.utils.import_utils import is_xformers_available |
|
from einops import rearrange, repeat |
|
from torch import nn |
|
|
|
|
|
def zero_module(module): |
|
""" |
|
Zero out the parameters of a module and return it. |
|
|
|
Args: |
|
- module: A PyTorch module to zero out its parameters. |
|
|
|
Returns: |
|
A zeroed out PyTorch module. |
|
""" |
|
for p in module.parameters(): |
|
p.detach().zero_() |
|
return module |
|
|
|
|
|
class TemporalTransformer3DModelOutput(BaseOutput): |
|
""" |
|
Output class for the TemporalTransformer3DModel. |
|
|
|
Attributes: |
|
sample (torch.FloatTensor): The output sample tensor from the model. |
|
""" |
|
sample: torch.FloatTensor |
|
|
|
def get_sample_shape(self): |
|
""" |
|
Returns the shape of the sample tensor. |
|
|
|
Returns: |
|
Tuple: The shape of the sample tensor. |
|
""" |
|
return self.sample.shape |
|
|
|
|
|
def get_motion_module(in_channels, motion_module_type: str, motion_module_kwargs: dict): |
|
""" |
|
This function returns a motion module based on the given type and parameters. |
|
|
|
Args: |
|
- in_channels (int): The number of input channels for the motion module. |
|
- motion_module_type (str): The type of motion module to create. Currently, only "Vanilla" is supported. |
|
- motion_module_kwargs (dict): Additional keyword arguments to pass to the motion module constructor. |
|
|
|
Returns: |
|
VanillaTemporalModule: The created motion module. |
|
|
|
Raises: |
|
ValueError: If an unsupported motion_module_type is provided. |
|
""" |
|
if motion_module_type == "Vanilla": |
|
return VanillaTemporalModule( |
|
in_channels=in_channels, |
|
**motion_module_kwargs, |
|
) |
|
|
|
raise ValueError |
|
|
|
|
|
class VanillaTemporalModule(nn.Module): |
|
""" |
|
A Vanilla Temporal Module class. |
|
|
|
Args: |
|
- in_channels (int): The number of input channels for the motion module. |
|
- num_attention_heads (int): Number of attention heads. |
|
- num_transformer_block (int): Number of transformer blocks. |
|
- attention_block_types (tuple): Types of attention blocks. |
|
- cross_frame_attention_mode: Mode for cross-frame attention. |
|
- temporal_position_encoding (bool): Flag for temporal position encoding. |
|
- temporal_position_encoding_max_len (int): Maximum length for temporal position encoding. |
|
- temporal_attention_dim_div (int): Divisor for temporal attention dimension. |
|
- zero_initialize (bool): Flag for zero initialization. |
|
""" |
|
|
|
def __init__( |
|
self, |
|
in_channels, |
|
num_attention_heads=8, |
|
num_transformer_block=2, |
|
attention_block_types=("Temporal_Self", "Temporal_Self"), |
|
cross_frame_attention_mode=None, |
|
temporal_position_encoding=False, |
|
temporal_position_encoding_max_len=24, |
|
temporal_attention_dim_div=1, |
|
zero_initialize=True, |
|
): |
|
super().__init__() |
|
|
|
self.temporal_transformer = TemporalTransformer3DModel( |
|
in_channels=in_channels, |
|
num_attention_heads=num_attention_heads, |
|
attention_head_dim=in_channels |
|
// num_attention_heads |
|
// temporal_attention_dim_div, |
|
num_layers=num_transformer_block, |
|
attention_block_types=attention_block_types, |
|
cross_frame_attention_mode=cross_frame_attention_mode, |
|
temporal_position_encoding=temporal_position_encoding, |
|
temporal_position_encoding_max_len=temporal_position_encoding_max_len, |
|
) |
|
|
|
if zero_initialize: |
|
self.temporal_transformer.proj_out = zero_module( |
|
self.temporal_transformer.proj_out |
|
) |
|
|
|
def forward( |
|
self, |
|
input_tensor, |
|
encoder_hidden_states, |
|
attention_mask=None, |
|
): |
|
""" |
|
Forward pass of the TemporalTransformer3DModel. |
|
|
|
Args: |
|
hidden_states (torch.Tensor): The hidden states of the model. |
|
encoder_hidden_states (torch.Tensor, optional): The hidden states of the encoder. |
|
attention_mask (torch.Tensor, optional): The attention mask. |
|
|
|
Returns: |
|
torch.Tensor: The output tensor after the forward pass. |
|
""" |
|
hidden_states = input_tensor |
|
hidden_states = self.temporal_transformer( |
|
hidden_states, encoder_hidden_states |
|
) |
|
|
|
output = hidden_states |
|
return output |
|
|
|
|
|
class TemporalTransformer3DModel(nn.Module): |
|
""" |
|
A Temporal Transformer 3D Model class. |
|
|
|
Args: |
|
- in_channels (int): The number of input channels. |
|
- num_attention_heads (int): Number of attention heads. |
|
- attention_head_dim (int): Dimension of attention heads. |
|
- num_layers (int): Number of transformer layers. |
|
- attention_block_types (tuple): Types of attention blocks. |
|
- dropout (float): Dropout rate. |
|
- norm_num_groups (int): Number of groups for normalization. |
|
- cross_attention_dim (int): Dimension for cross-attention. |
|
- activation_fn (str): Activation function. |
|
- attention_bias (bool): Flag for attention bias. |
|
- upcast_attention (bool): Flag for upcast attention. |
|
- cross_frame_attention_mode: Mode for cross-frame attention. |
|
- temporal_position_encoding (bool): Flag for temporal position encoding. |
|
- temporal_position_encoding_max_len (int): Maximum length for temporal position encoding. |
|
""" |
|
def __init__( |
|
self, |
|
in_channels, |
|
num_attention_heads, |
|
attention_head_dim, |
|
num_layers, |
|
attention_block_types=( |
|
"Temporal_Self", |
|
"Temporal_Self", |
|
), |
|
dropout=0.0, |
|
norm_num_groups=32, |
|
cross_attention_dim=768, |
|
activation_fn="geglu", |
|
attention_bias=False, |
|
upcast_attention=False, |
|
cross_frame_attention_mode=None, |
|
temporal_position_encoding=False, |
|
temporal_position_encoding_max_len=24, |
|
): |
|
super().__init__() |
|
|
|
inner_dim = num_attention_heads * attention_head_dim |
|
|
|
self.norm = torch.nn.GroupNorm( |
|
num_groups=norm_num_groups, num_channels=in_channels, eps=1e-6, affine=True |
|
) |
|
self.proj_in = nn.Linear(in_channels, inner_dim) |
|
|
|
self.transformer_blocks = nn.ModuleList( |
|
[ |
|
TemporalTransformerBlock( |
|
dim=inner_dim, |
|
num_attention_heads=num_attention_heads, |
|
attention_head_dim=attention_head_dim, |
|
attention_block_types=attention_block_types, |
|
dropout=dropout, |
|
cross_attention_dim=cross_attention_dim, |
|
activation_fn=activation_fn, |
|
attention_bias=attention_bias, |
|
upcast_attention=upcast_attention, |
|
cross_frame_attention_mode=cross_frame_attention_mode, |
|
temporal_position_encoding=temporal_position_encoding, |
|
temporal_position_encoding_max_len=temporal_position_encoding_max_len, |
|
) |
|
for d in range(num_layers) |
|
] |
|
) |
|
self.proj_out = nn.Linear(inner_dim, in_channels) |
|
|
|
def forward(self, hidden_states, encoder_hidden_states=None): |
|
""" |
|
Forward pass for the TemporalTransformer3DModel. |
|
|
|
Args: |
|
hidden_states (torch.Tensor): The input hidden states with shape (batch_size, sequence_length, in_channels). |
|
encoder_hidden_states (torch.Tensor, optional): The encoder hidden states with shape (batch_size, encoder_sequence_length, in_channels). |
|
|
|
Returns: |
|
torch.Tensor: The output hidden states with shape (batch_size, sequence_length, in_channels). |
|
""" |
|
assert ( |
|
hidden_states.dim() == 5 |
|
), f"Expected hidden_states to have ndim=5, but got ndim={hidden_states.dim()}." |
|
video_length = hidden_states.shape[2] |
|
hidden_states = rearrange(hidden_states, "b c f h w -> (b f) c h w") |
|
|
|
batch, _, height, weight = hidden_states.shape |
|
residual = hidden_states |
|
|
|
hidden_states = self.norm(hidden_states) |
|
inner_dim = hidden_states.shape[1] |
|
hidden_states = hidden_states.permute(0, 2, 3, 1).reshape( |
|
batch, height * weight, inner_dim |
|
) |
|
hidden_states = self.proj_in(hidden_states) |
|
|
|
|
|
for block in self.transformer_blocks: |
|
hidden_states = block( |
|
hidden_states, |
|
encoder_hidden_states=encoder_hidden_states, |
|
video_length=video_length, |
|
) |
|
|
|
|
|
hidden_states = self.proj_out(hidden_states) |
|
hidden_states = ( |
|
hidden_states.reshape(batch, height, weight, inner_dim) |
|
.permute(0, 3, 1, 2) |
|
.contiguous() |
|
) |
|
|
|
output = hidden_states + residual |
|
output = rearrange(output, "(b f) c h w -> b c f h w", f=video_length) |
|
|
|
return output |
|
|
|
|
|
class TemporalTransformerBlock(nn.Module): |
|
""" |
|
A Temporal Transformer Block class. |
|
|
|
Args: |
|
- dim (int): Dimension of the block. |
|
- num_attention_heads (int): Number of attention heads. |
|
- attention_head_dim (int): Dimension of attention heads. |
|
- attention_block_types (tuple): Types of attention blocks. |
|
- dropout (float): Dropout rate. |
|
- cross_attention_dim (int): Dimension for cross-attention. |
|
- activation_fn (str): Activation function. |
|
- attention_bias (bool): Flag for attention bias. |
|
- upcast_attention (bool): Flag for upcast attention. |
|
- cross_frame_attention_mode: Mode for cross-frame attention. |
|
- temporal_position_encoding (bool): Flag for temporal position encoding. |
|
- temporal_position_encoding_max_len (int): Maximum length for temporal position encoding. |
|
""" |
|
def __init__( |
|
self, |
|
dim, |
|
num_attention_heads, |
|
attention_head_dim, |
|
attention_block_types=( |
|
"Temporal_Self", |
|
"Temporal_Self", |
|
), |
|
dropout=0.0, |
|
cross_attention_dim=768, |
|
activation_fn="geglu", |
|
attention_bias=False, |
|
upcast_attention=False, |
|
cross_frame_attention_mode=None, |
|
temporal_position_encoding=False, |
|
temporal_position_encoding_max_len=24, |
|
): |
|
super().__init__() |
|
|
|
attention_blocks = [] |
|
norms = [] |
|
|
|
for block_name in attention_block_types: |
|
attention_blocks.append( |
|
VersatileAttention( |
|
attention_mode=block_name.split("_", maxsplit=1)[0], |
|
cross_attention_dim=cross_attention_dim |
|
if block_name.endswith("_Cross") |
|
else None, |
|
query_dim=dim, |
|
heads=num_attention_heads, |
|
dim_head=attention_head_dim, |
|
dropout=dropout, |
|
bias=attention_bias, |
|
upcast_attention=upcast_attention, |
|
cross_frame_attention_mode=cross_frame_attention_mode, |
|
temporal_position_encoding=temporal_position_encoding, |
|
temporal_position_encoding_max_len=temporal_position_encoding_max_len, |
|
) |
|
) |
|
norms.append(nn.LayerNorm(dim)) |
|
|
|
self.attention_blocks = nn.ModuleList(attention_blocks) |
|
self.norms = nn.ModuleList(norms) |
|
|
|
self.ff = FeedForward(dim, dropout=dropout, |
|
activation_fn=activation_fn) |
|
self.ff_norm = nn.LayerNorm(dim) |
|
|
|
def forward( |
|
self, |
|
hidden_states, |
|
encoder_hidden_states=None, |
|
video_length=None, |
|
): |
|
""" |
|
Forward pass for the TemporalTransformerBlock. |
|
|
|
Args: |
|
hidden_states (torch.Tensor): The input hidden states with shape |
|
(batch_size, video_length, in_channels). |
|
encoder_hidden_states (torch.Tensor, optional): The encoder hidden states |
|
with shape (batch_size, encoder_length, in_channels). |
|
video_length (int, optional): The length of the video. |
|
|
|
Returns: |
|
torch.Tensor: The output hidden states with shape |
|
(batch_size, video_length, in_channels). |
|
""" |
|
for attention_block, norm in zip(self.attention_blocks, self.norms): |
|
norm_hidden_states = norm(hidden_states) |
|
hidden_states = ( |
|
attention_block( |
|
norm_hidden_states, |
|
encoder_hidden_states=encoder_hidden_states |
|
if attention_block.is_cross_attention |
|
else None, |
|
video_length=video_length, |
|
) |
|
+ hidden_states |
|
) |
|
|
|
hidden_states = self.ff(self.ff_norm(hidden_states)) + hidden_states |
|
|
|
output = hidden_states |
|
return output |
|
|
|
|
|
class PositionalEncoding(nn.Module): |
|
""" |
|
Positional Encoding module for transformers. |
|
|
|
Args: |
|
- d_model (int): Model dimension. |
|
- dropout (float): Dropout rate. |
|
- max_len (int): Maximum length for positional encoding. |
|
""" |
|
def __init__(self, d_model, dropout=0.0, max_len=24): |
|
super().__init__() |
|
self.dropout = nn.Dropout(p=dropout) |
|
position = torch.arange(max_len).unsqueeze(1) |
|
div_term = torch.exp( |
|
torch.arange(0, d_model, 2) * (-math.log(10000.0) / d_model) |
|
) |
|
pe = torch.zeros(1, max_len, d_model) |
|
pe[0, :, 0::2] = torch.sin(position * div_term) |
|
pe[0, :, 1::2] = torch.cos(position * div_term) |
|
self.register_buffer("pe", pe) |
|
|
|
def forward(self, x): |
|
""" |
|
Forward pass of the PositionalEncoding module. |
|
|
|
This method takes an input tensor `x` and adds the positional encoding to it. The positional encoding is |
|
generated based on the input tensor's shape and is added to the input tensor element-wise. |
|
|
|
Args: |
|
x (torch.Tensor): The input tensor to be positionally encoded. |
|
|
|
Returns: |
|
torch.Tensor: The positionally encoded tensor. |
|
""" |
|
x = x + self.pe[:, : x.size(1)] |
|
return self.dropout(x) |
|
|
|
|
|
class VersatileAttention(Attention): |
|
""" |
|
Versatile Attention class. |
|
|
|
Args: |
|
- attention_mode: Attention mode. |
|
- temporal_position_encoding (bool): Flag for temporal position encoding. |
|
- temporal_position_encoding_max_len (int): Maximum length for temporal position encoding. |
|
""" |
|
def __init__( |
|
self, |
|
*args, |
|
attention_mode=None, |
|
cross_frame_attention_mode=None, |
|
temporal_position_encoding=False, |
|
temporal_position_encoding_max_len=24, |
|
**kwargs, |
|
): |
|
super().__init__(*args, **kwargs) |
|
assert attention_mode == "Temporal" |
|
|
|
self.attention_mode = attention_mode |
|
self.is_cross_attention = kwargs.get("cross_attention_dim") is not None |
|
|
|
self.pos_encoder = ( |
|
PositionalEncoding( |
|
kwargs["query_dim"], |
|
dropout=0.0, |
|
max_len=temporal_position_encoding_max_len, |
|
) |
|
if (temporal_position_encoding and attention_mode == "Temporal") |
|
else None |
|
) |
|
|
|
def extra_repr(self): |
|
""" |
|
Returns a string representation of the module with information about the attention mode and whether it is cross-attention. |
|
|
|
Returns: |
|
str: A string representation of the module. |
|
""" |
|
return f"(Module Info) Attention_Mode: {self.attention_mode}, Is_Cross_Attention: {self.is_cross_attention}" |
|
|
|
def set_use_memory_efficient_attention_xformers( |
|
self, |
|
use_memory_efficient_attention_xformers: bool, |
|
): |
|
""" |
|
Sets the use of memory-efficient attention xformers for the VersatileAttention class. |
|
|
|
Args: |
|
use_memory_efficient_attention_xformers (bool): A boolean flag indicating whether to use memory-efficient attention xformers or not. |
|
|
|
Returns: |
|
None |
|
|
|
""" |
|
if use_memory_efficient_attention_xformers: |
|
if not is_xformers_available(): |
|
raise ModuleNotFoundError( |
|
( |
|
"Refer to https://github.com/facebookresearch/xformers for more information on how to install" |
|
" xformers" |
|
), |
|
name="xformers", |
|
) |
|
|
|
if not torch.cuda.is_available(): |
|
raise ValueError( |
|
"torch.cuda.is_available() should be True but is False. xformers' memory efficient attention is" |
|
" only available for GPU " |
|
) |
|
|
|
try: |
|
|
|
_ = xformers.ops.memory_efficient_attention( |
|
torch.randn((1, 2, 40), device="cuda"), |
|
torch.randn((1, 2, 40), device="cuda"), |
|
torch.randn((1, 2, 40), device="cuda"), |
|
) |
|
except Exception as e: |
|
raise e |
|
processor = AttnProcessor() |
|
else: |
|
processor = AttnProcessor() |
|
|
|
self.set_processor(processor) |
|
|
|
def forward( |
|
self, |
|
hidden_states, |
|
encoder_hidden_states=None, |
|
attention_mask=None, |
|
video_length=None, |
|
**cross_attention_kwargs, |
|
): |
|
""" |
|
Args: |
|
hidden_states (`torch.Tensor`): |
|
The hidden states to be passed through the model. |
|
encoder_hidden_states (`torch.Tensor`, optional): |
|
The encoder hidden states to be passed through the model. |
|
attention_mask (`torch.Tensor`, optional): |
|
The attention mask to be used in the model. |
|
video_length (`int`, optional): |
|
The length of the video. |
|
cross_attention_kwargs (`dict`, optional): |
|
Additional keyword arguments to be used for cross-attention. |
|
|
|
Returns: |
|
`torch.Tensor`: |
|
The output tensor after passing through the model. |
|
|
|
""" |
|
if self.attention_mode == "Temporal": |
|
d = hidden_states.shape[1] |
|
hidden_states = rearrange( |
|
hidden_states, "(b f) d c -> (b d) f c", f=video_length |
|
) |
|
|
|
if self.pos_encoder is not None: |
|
hidden_states = self.pos_encoder(hidden_states) |
|
|
|
encoder_hidden_states = ( |
|
repeat(encoder_hidden_states, "b n c -> (b d) n c", d=d) |
|
if encoder_hidden_states is not None |
|
else encoder_hidden_states |
|
) |
|
|
|
else: |
|
raise NotImplementedError |
|
|
|
hidden_states = self.processor( |
|
self, |
|
hidden_states, |
|
encoder_hidden_states=encoder_hidden_states, |
|
attention_mask=attention_mask, |
|
**cross_attention_kwargs, |
|
) |
|
|
|
if self.attention_mode == "Temporal": |
|
hidden_states = rearrange( |
|
hidden_states, "(b d) f c -> (b f) d c", d=d) |
|
|
|
return hidden_states |
|
|