File size: 9,246 Bytes
9c2f1dd | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 | from collections import OrderedDict
from dataclasses import dataclass
from typing import Optional, Tuple
import torch
import torch.nn as nn
import torch.nn.functional as F
from transformers import PreTrainedModel
from transformers.utils import ModelOutput, logging
from .configuration_msp_fusion import MSPFusionConfig
logger = logging.get_logger(__name__)
@dataclass
class MSPFusionOutput(ModelOutput):
last_hidden_state: torch.FloatTensor = None
fusion_padding_mask: Optional[torch.Tensor] = None
audio_hidden_state: Optional[torch.FloatTensor] = None
visual_hidden_state: Optional[torch.FloatTensor] = None
attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
class MSPFusionPreTrainedModel(PreTrainedModel):
config_class = MSPFusionConfig
base_model_prefix = "msp_fusion"
supports_gradient_checkpointing = False
all_tied_weights_keys = OrderedDict()
def _init_weights(self, module):
if isinstance(module, nn.Linear):
module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
if module.bias is not None:
module.bias.data.zero_()
elif isinstance(module, nn.LayerNorm):
module.bias.data.zero_()
module.weight.data.fill_(1.0)
def _align_temporal(
self, x: torch.Tensor, target_len: int, mode: str = "nearest"
) -> torch.Tensor:
if x.size(1) == target_len:
return x
return F.interpolate(x.transpose(1, 2), size=target_len, mode=mode).transpose(
1, 2
)
class MSPFusionModel(MSPFusionPreTrainedModel):
"""Bidirectional cross-attention fusion of audio and visual streams."""
def __init__(self, config: MSPFusionConfig):
super().__init__(config)
if config.fusion_hidden_size % config.num_attention_heads != 0:
raise ValueError(
f"fusion_hidden_size ({config.fusion_hidden_size}) must be divisible "
f"by num_attention_heads ({config.num_attention_heads})."
)
# Modality projection blocks: linear + layer-norm
self.audio_proj = nn.Sequential(
nn.Linear(config.audio_hidden_size, config.fusion_hidden_size),
nn.LayerNorm(config.fusion_hidden_size, eps=config.layer_norm_eps),
)
self.visual_proj = nn.Sequential(
nn.Linear(config.visual_hidden_size, config.fusion_hidden_size),
nn.LayerNorm(config.fusion_hidden_size, eps=config.layer_norm_eps),
)
# Bidirectional cross-attention
self.audio_to_visual_attn = nn.MultiheadAttention(
embed_dim=config.fusion_hidden_size,
num_heads=config.num_attention_heads,
dropout=config.attention_dropout,
batch_first=True,
)
self.visual_to_audio_attn = nn.MultiheadAttention(
embed_dim=config.fusion_hidden_size,
num_heads=config.num_attention_heads,
dropout=config.attention_dropout,
batch_first=True,
)
# Post-attention residual norms
self.audio_norm = nn.LayerNorm(
config.fusion_hidden_size, eps=config.layer_norm_eps
)
self.visual_norm = nn.LayerNorm(
config.fusion_hidden_size, eps=config.layer_norm_eps
)
# Gated fusion: concat → linear → gelu → dropout → layer-norm
self.fusion_gate = nn.Sequential(
nn.Linear(config.fusion_hidden_size * 2, config.fusion_hidden_size),
nn.GELU(),
nn.Dropout(config.dropout),
nn.LayerNorm(config.fusion_hidden_size, eps=config.layer_norm_eps),
)
@property
def dummy_inputs(self) -> dict:
return {
"audio_hidden_states": torch.zeros(2, 50, self.config.audio_hidden_size),
"visual_hidden_states": torch.zeros(2, 10, self.config.visual_hidden_size),
}
def _get_abs_attention_mask(self, attention_mask, dtype):
if attention_mask.dim() == 2:
extended_attention_mask = attention_mask[:, None, None, :]
elif attention_mask.dim() == 3:
extended_attention_mask = attention_mask[:, None, :, :]
else:
extended_attention_mask = attention_mask
extended_attention_mask = extended_attention_mask.to(dtype=dtype)
extended_attention_mask = (1.0 - extended_attention_mask) * torch.finfo(
dtype
).min
return extended_attention_mask
def forward(
self,
audio_hidden_states: Optional[torch.Tensor] = None,
visual_hidden_states: Optional[torch.Tensor] = None,
audio_key_padding_mask: Optional[torch.Tensor] = None,
visual_key_padding_mask: Optional[torch.Tensor] = None,
output_attentions: Optional[bool] = None,
) -> MSPFusionOutput:
"""
Args:
audio_hidden_states: (B, T_audio, audio_hidden_size)
visual_hidden_states: (B, T_visual, visual_hidden_size)
audio_key_padding_mask: (B, T_audio); True = pad position to ignore
visual_key_padding_mask: (B, T_visual); True = pad position to ignore
output_attentions: return attention weight matrices when True
Returns:
MSPFusionOutput with last_hidden_state of shape (B, T_audio, fusion_hidden_size)
"""
has_audio = audio_hidden_states is not None
has_visual = visual_hidden_states is not None
if not has_audio and not has_visual:
raise ValueError(
"At least one of audio_hidden_states or visual_hidden_states must be provided."
)
output_attentions = (
output_attentions if output_attentions is not None else False
)
# Handle cases where only one modality is present
# only audio
if has_audio and not has_visual:
audio_states = self.audio_proj(audio_hidden_states)
visual_states = torch.zeros_like(audio_states)
fusion_key_padding_mask = audio_key_padding_mask
fused = self.fusion_gate(torch.cat([audio_states, visual_states], dim=-1))
return MSPFusionOutput(
last_hidden_state=fused,
fusion_padding_mask=fusion_key_padding_mask,
audio_hidden_state=audio_states,
visual_hidden_state=None,
attentions=None,
)
# only visual
if has_visual and not has_audio:
visual_states = self.visual_proj(visual_hidden_states)
audio_states = torch.zeros_like(visual_states)
fusion_key_padding_mask = visual_key_padding_mask
fused = self.fusion_gate(torch.cat([audio_states, visual_states], dim=-1))
return MSPFusionOutput(
last_hidden_state=fused,
fusion_padding_mask=fusion_key_padding_mask,
audio_hidden_state=None,
visual_hidden_state=visual_states,
attentions=None,
)
# Both modalities are present
T_audio = audio_hidden_states.size(1)
visual_hidden_states = self._align_temporal(
visual_hidden_states, T_audio, mode="nearest"
)
audio_states = self.audio_proj(audio_hidden_states)
visual_states = self.visual_proj(visual_hidden_states)
# fusion key padding mask for calc ctc loss
if audio_key_padding_mask is not None:
fusion_key_padding_mask = audio_key_padding_mask
elif visual_key_padding_mask is not None:
fusion_key_padding_mask = visual_key_padding_mask[:, : audio_states.size(1)]
else:
fusion_key_padding_mask = None
a2v_attn = None
v2a_attn = None
audio_residual = audio_states
visual_residual = visual_states
# Bidirectional cross-attention
# Audio attends to visual (audio queries, visual keys/values)
audio_attended, a2v_attn = self.audio_to_visual_attn.forward(
query=audio_states,
key=visual_states,
value=visual_states,
need_weights=output_attentions,
average_attn_weights=False,
)
audio_states = self.audio_norm(audio_residual + audio_attended)
# Visual attends to audio (visual queries, audio keys/values)
visual_attended, v2a_attn = self.visual_to_audio_attn.forward(
query=visual_states,
key=audio_residual,
value=audio_residual,
need_weights=output_attentions,
average_attn_weights=False,
)
visual_states = self.visual_norm(visual_residual + visual_attended)
# Gated fusion: concatenate both streams and project to fusion space
fused = self.fusion_gate(torch.cat([audio_states, visual_states], dim=-1))
attentions = (a2v_attn, v2a_attn) if output_attentions else None
return MSPFusionOutput(
last_hidden_state=fused,
fusion_padding_mask=fusion_key_padding_mask,
audio_hidden_state=audio_states,
visual_hidden_state=visual_states,
attentions=attentions,
)
|