File size: 9,291 Bytes
338c3c5 | 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 | # Copyright 2026 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
import torch
import torch.nn as nn
import torch.nn.functional as F
from diffusers.configuration_utils import ConfigMixin, register_to_config
from diffusers.models.attention import AttentionMixin, AttentionModuleMixin
from diffusers.models.attention_dispatch import dispatch_attention_fn
from diffusers.models.modeling_utils import ModelMixin
from diffusers.utils import BaseOutput
class MotionBricksMLP(nn.Module):
"""Two linear layers, each followed by LeakyReLU(0.01), then a linear projection."""
def __init__(self, in_features: int, hidden_features: int, out_features: int):
super().__init__()
self.layers = nn.ModuleList(
[nn.Linear(in_features, hidden_features), nn.Linear(hidden_features, hidden_features)]
)
self.proj_out = nn.Linear(hidden_features, out_features)
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
for layer in self.layers:
hidden_states = F.leaky_relu(layer(hidden_states), 0.01)
return self.proj_out(hidden_states)
class MotionBricksAttnProcessor:
_attention_backend = None
_parallel_config = None
def __call__(
self,
attn: "MotionBricksAttention",
hidden_states: torch.Tensor,
attention_mask: torch.Tensor | None = None,
) -> torch.Tensor:
batch_size, seq_len, _ = hidden_states.shape
shape = (batch_size, seq_len, attn.heads, -1)
query = attn.to_q(hidden_states).view(shape)
key = attn.to_k(hidden_states).view(shape)
value = attn.to_v(hidden_states).view(shape)
hidden_states = dispatch_attention_fn(
query,
key,
value,
attn_mask=attention_mask,
backend=self._attention_backend,
parallel_config=self._parallel_config,
)
hidden_states = hidden_states.flatten(2, 3).to(query.dtype)
return attn.to_out[0](hidden_states)
class MotionBricksAttention(nn.Module, AttentionModuleMixin):
"""Multi-head self-attention. `attention_mask` is a boolean `(batch_size, 1, 1, seq_len)` mask of the keys to
attend to."""
_default_processor_cls = MotionBricksAttnProcessor
_available_processors = [MotionBricksAttnProcessor]
def __init__(self, dim: int, heads: int):
super().__init__()
self.heads = heads
self.to_q = nn.Linear(dim, dim)
self.to_k = nn.Linear(dim, dim)
self.to_v = nn.Linear(dim, dim)
self.to_out = nn.ModuleList([nn.Linear(dim, dim), nn.Dropout(0.0)])
self.set_processor(MotionBricksAttnProcessor())
def forward(self, hidden_states: torch.Tensor, attention_mask: torch.Tensor | None = None) -> torch.Tensor:
return self.processor(self, hidden_states, attention_mask)
class MotionBricksTransformerLayer(nn.Module):
"""Post-norm transformer encoder layer with a ReLU feed-forward."""
def __init__(self, dim: int, heads: int, ff_dim: int):
super().__init__()
self.attn = MotionBricksAttention(dim, heads)
self.norm1 = nn.LayerNorm(dim)
self.ff_in = nn.Linear(dim, ff_dim)
self.ff_out = nn.Linear(ff_dim, dim)
self.norm2 = nn.LayerNorm(dim)
def forward(self, hidden_states: torch.Tensor, attention_mask: torch.Tensor | None = None) -> torch.Tensor:
hidden_states = self.norm1(hidden_states + self.attn(hidden_states, attention_mask))
return self.norm2(hidden_states + self.ff_out(F.relu(self.ff_in(hidden_states))))
def key_mask(num_tokens: torch.Tensor, max_tokens: int, prefix: int = 0) -> torch.Tensor:
"""Boolean `(batch_size, 1, 1, prefix + max_tokens)` attention mask: the `prefix` positions, then the first
`num_tokens` of `max_tokens` token positions."""
positions = torch.arange(max_tokens, device=num_tokens.device)
valid = positions[None] < num_tokens[:, None]
valid = F.pad(valid, (prefix, 0), value=True)
return valid[:, None, None]
@dataclass
class MotionBricksPoseTransformerOutput(BaseOutput):
"""
Args:
logits (`torch.Tensor` of shape `(batch_size, max_tokens, num_heads_codebook, codebook_size)`):
The code logits of every token and head.
"""
logits: torch.Tensor
class MotionBricksPoseTransformer(ModelMixin, ConfigMixin, AttentionMixin):
"""Masked token transformer that predicts the pose codes of every token in one pass, from the given poses, the
root of every frame and the length of the motion. Each token covers `frames_per_token` frames and has one code
per head."""
_no_split_modules = ["MotionBricksTransformerLayer"]
_repeated_blocks = ["MotionBricksTransformerLayer"]
_skip_layerwise_casting_patterns = ["norm", "embed"]
@register_to_config
def __init__(
self,
pose_dim: int = 268,
root_dim: int = 4,
num_heads_codebook: int = 5,
codebook_size: int = 16,
code_embed_dim: int = 32,
token_dim: int = 640,
root_embed_dim: int = 256,
length_embed_dim: int = 128,
hidden_dim: int = 1024,
num_heads: int = 16,
ff_dim: int = 2048,
num_layers: int = 16,
frames_per_token: int = 4,
min_tokens: int = 6,
max_tokens: int = 16,
):
super().__init__()
code_dim = num_heads_codebook * code_embed_dim
# Each head has codebook_size codes and a mask code.
self.code_embed = nn.Embedding(num_heads_codebook * (codebook_size + 1), code_embed_dim)
self.code_proj = MotionBricksMLP(code_dim, token_dim, token_dim)
self.pose_proj = nn.Linear(pose_dim, token_dim // frames_per_token)
self.root_proj = nn.Linear(root_dim * frames_per_token, root_embed_dim)
self.length_embed = nn.Embedding(max_tokens - min_tokens + 1, length_embed_dim)
self.proj_in = nn.Linear(token_dim + root_embed_dim + length_embed_dim, hidden_dim)
self.position_embed = nn.Parameter(torch.zeros(max_tokens, hidden_dim))
self.layers = nn.ModuleList(
[MotionBricksTransformerLayer(hidden_dim, num_heads, ff_dim) for _ in range(num_layers)]
)
self.proj_out = nn.Linear(hidden_dim, num_heads_codebook * codebook_size)
def forward(
self,
pose: torch.Tensor,
has_pose: torch.Tensor,
root: torch.Tensor,
num_tokens: torch.Tensor,
return_dict: bool = True,
) -> MotionBricksPoseTransformerOutput | tuple:
"""
Args:
pose (`torch.Tensor` of shape `(batch_size, num_frames, pose_dim)`):
Normalized pose features on the frame timeline.
has_pose (`torch.BoolTensor` of shape `(batch_size, num_frames)`):
The frames whose pose is given.
root (`torch.Tensor` of shape `(batch_size, num_frames, root_dim)`):
Normalized root of every frame: ground position and the cosine and sine of the heading.
num_tokens (`torch.LongTensor` of shape `(batch_size,)`):
The motion length in tokens.
return_dict (`bool`, defaults to `True`):
Whether to return a `MotionBricksPoseTransformerOutput` instead of a tuple.
Returns:
`MotionBricksPoseTransformerOutput` or `tuple`: The code logits; a one-element tuple when `return_dict` is
`False`.
"""
config = self.config
batch_size = pose.shape[0]
heads, size = config.num_heads_codebook, config.codebook_size
mask_codes = size + torch.arange(heads, device=pose.device) * (size + 1)
codes = self.code_embed(mask_codes).flatten()
codes = self.code_proj(codes).expand(batch_size, config.max_tokens, -1)
codes = codes.reshape(batch_size, -1, self.pose_proj.out_features)
has_pose = has_pose[..., None].to(pose.dtype)
frames = self.pose_proj(pose) * has_pose + codes * (1 - has_pose)
frames = frames.reshape(batch_size, config.max_tokens, -1)
root = self.root_proj(root.reshape(batch_size, config.max_tokens, -1))
length = self.length_embed(num_tokens - config.min_tokens)[:, None].expand(-1, config.max_tokens, -1)
hidden_states = F.relu(self.proj_in(torch.cat([frames, root, length], dim=-1))) + self.position_embed
mask = key_mask(num_tokens, config.max_tokens)
for layer in self.layers:
hidden_states = layer(hidden_states, mask)
logits = self.proj_out(hidden_states).unflatten(-1, (heads, size))
if not return_dict:
return (logits,)
return MotionBricksPoseTransformerOutput(logits=logits)
|