File size: 8,836 Bytes
a1739b9
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
# Copyright 2024 The HuggingFace Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from dataclasses import dataclass
from typing import Any, Dict, Optional

import torch
from torch import nn

from diffusers.configuration_utils import ConfigMixin, register_to_config
from diffusers.utils import BaseOutput
from diffusers.models.embeddings import TimestepEmbedding, Timesteps
from diffusers.models.modeling_utils import ModelMixin
from ..resnet import AlphaBlender
from ..attention import (
    BasicTransformerBlock,
    TemporalRopeBasicTransformerBlock,
)
from ..embeddings import rope


@dataclass
class TransformerTemporalModelOutput(BaseOutput):
    """
    The output of [`TransformerTemporalModel`].

    Args:
        sample (`torch.FloatTensor` of shape `(batch_size x num_frames, num_channels, height, width)`):
            The hidden states output conditioned on `encoder_hidden_states` input.
    """

    sample: torch.FloatTensor


class TransformerSpatioTemporalModel(nn.Module):
    """
    A Transformer model for video-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**).
        out_channels (`int`, *optional*):
            The number of channels in the output (specify if the input is **continuous**).
        num_layers (`int`, *optional*, defaults to 1): The number of layers of Transformer blocks to use.
        cross_attention_dim (`int`, *optional*): The number of `encoder_hidden_states` dimensions to use.
    """

    def __init__(
        self,
        num_attention_heads: int = 16,
        attention_head_dim: int = 88,
        in_channels: int = 320,
        out_channels: Optional[int] = None,
        num_layers: int = 1,
        cross_attention_dim: Optional[int] = None,
    ):
        super().__init__()
        self.num_attention_heads = num_attention_heads
        self.attention_head_dim = attention_head_dim

        inner_dim = num_attention_heads * attention_head_dim
        self.inner_dim = inner_dim

        # 2. Define input layers
        self.in_channels = in_channels
        self.norm = torch.nn.GroupNorm(
            num_groups=32, num_channels=in_channels, eps=1e-6
        )
        self.proj_in = nn.Linear(in_channels, inner_dim)

        # 3. Define transformers blocks
        self.transformer_blocks = nn.ModuleList(
            [
                BasicTransformerBlock(
                    inner_dim,
                    num_attention_heads,
                    attention_head_dim,
                    cross_attention_dim=cross_attention_dim,
                )
                for d in range(num_layers)
            ]
        )

        time_mix_inner_dim = inner_dim
        self.temporal_transformer_blocks = nn.ModuleList(
            [
                TemporalRopeBasicTransformerBlock(
                    inner_dim,
                    time_mix_inner_dim,
                    num_attention_heads,
                    attention_head_dim,
                    cross_attention_dim=cross_attention_dim,
                )
                for _ in range(num_layers)
            ]
        )

        self.time_mixer = AlphaBlender(alpha=0.5, merge_strategy="learned_with_images")

        # 4. Define output layers
        self.out_channels = in_channels if out_channels is None else out_channels
        # TODO: should use out_channels for continuous projections
        self.proj_out = nn.Linear(inner_dim, in_channels)

        self.gradient_checkpointing = False

    def forward(
        self,
        hidden_states: torch.Tensor,
        encoder_hidden_states: Optional[torch.Tensor] = None,
        image_only_indicator: Optional[torch.Tensor] = None,
        return_dict: bool = True,
        position_ids: Optional[torch.Tensor] = None,
    ):
        """
        Args:
            hidden_states (`torch.FloatTensor` of shape `(batch size, channel, height, width)`):
                Input hidden_states.
            num_frames (`int`):
                The number of frames to be processed per batch. This is used to reshape the hidden states.
            encoder_hidden_states ( `torch.LongTensor` of shape `(batch size, encoder_hidden_states dim)`, *optional*):
                Conditional embeddings for cross attention layer. If not given, cross-attention defaults to
                self-attention.
            image_only_indicator (`torch.LongTensor` of shape `(batch size, num_frames)`, *optional*):
                A tensor indicating whether the input contains only images. 1 indicates that the input contains only
                images, 0 indicates that the input contains video frames.
            return_dict (`bool`, *optional*, defaults to `True`):
                Whether or not to return a [`~models.transformer_temporal.TransformerTemporalModelOutput`] instead of a plain
                tuple.

        Returns:
            [`~models.transformer_temporal.TransformerTemporalModelOutput`] or `tuple`:
                If `return_dict` is True, an [`~models.transformer_temporal.TransformerTemporalModelOutput`] is
                returned, otherwise a `tuple` where the first element is the sample tensor.
        """
        # 1. Input
        batch_frames, _, height, width = hidden_states.shape
        num_frames = image_only_indicator.shape[-1]
        batch_size = batch_frames // num_frames

        # (B*F, 1, C)
        time_context = encoder_hidden_states
        # (B, 1, C)
        time_context_first_timestep = time_context[None, :].reshape(
            batch_size, num_frames, -1, time_context.shape[-1]
        )[:, 0]

        # (B*N, 1, C)
        time_context = time_context_first_timestep.repeat_interleave(
            height * width, dim=0
        )

        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_frames, height * width, inner_dim
        )
        hidden_states = self.proj_in(hidden_states)

        if position_ids is None:
            # (B, F)
            frame_rotary_emb = torch.arange(num_frames, device=hidden_states.device)
            frame_rotary_emb = frame_rotary_emb[None, :].repeat(batch_size, 1)
        else:
            frame_rotary_emb = position_ids

        # (B, 1, F, d/2, 2, 2)
        frame_rotary_emb = rope(frame_rotary_emb, self.attention_head_dim)
        # (B*N, 1, F, d/2, 2, 2)
        frame_rotary_emb = frame_rotary_emb.repeat_interleave(height * width, dim=0)

        # 2. Blocks
        for block, temporal_block in zip(
            self.transformer_blocks, self.temporal_transformer_blocks
        ):
            if self.training and self.gradient_checkpointing:
                hidden_states = torch.utils.checkpoint.checkpoint(
                    block,
                    hidden_states,
                    None,
                    encoder_hidden_states,
                    None,
                    use_reentrant=False,
                )
            else:
                hidden_states = block(
                    hidden_states,
                    encoder_hidden_states=encoder_hidden_states,
                )

            hidden_states_mix = temporal_block(
                hidden_states,
                num_frames=num_frames,
                encoder_hidden_states=time_context,
                frame_rotary_emb=frame_rotary_emb,
            )
            hidden_states = self.time_mixer(
                x_spatial=hidden_states,
                x_temporal=hidden_states_mix,
                image_only_indicator=image_only_indicator,
            )

        # 3. Output
        hidden_states = self.proj_out(hidden_states)
        hidden_states = (
            hidden_states.reshape(batch_frames, height, width, inner_dim)
            .permute(0, 3, 1, 2)
            .contiguous()
        )

        output = hidden_states + residual

        if not return_dict:
            return (output,)

        return TransformerTemporalModelOutput(sample=output)