V2Dial/models/backbones/moes.py
2025-06-24 08:38:09 +02:00

287 lines
No EOL
11 KiB
Python

import torch
import torch.nn as nn
from transformers.models.llama.modeling_llama import LlamaRMSNorm
from timm.models.layers import DropPath
class Mlp(nn.Module):
def __init__(
self,
in_features,
hidden_features=None,
out_features=None,
act_layer=nn.GELU,
drop=0.0,
):
super().__init__()
out_features = out_features or in_features
hidden_features = hidden_features or in_features
self.fc1 = nn.Linear(in_features, hidden_features)
self.act = act_layer()
self.fc2 = nn.Linear(hidden_features, out_features)
self.drop = nn.Dropout(drop)
def forward(self, x):
x = self.fc1(x)
x = self.act(x)
x = self.drop(x)
x = self.fc2(x)
x = self.drop(x)
return x
class Attention(nn.Module):
def __init__(
self,
dim,
num_heads=8,
qkv_bias=False,
qk_scale=None,
attn_drop=0.0,
proj_drop=0.0,
):
super().__init__()
self.num_heads = num_heads
head_dim = dim // num_heads
# NOTE scale factor was wrong in my original version, can set manually to be compat with prev weights
self.scale = qk_scale or head_dim ** -0.5
self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
self.attn_drop = nn.Dropout(attn_drop)
self.proj = nn.Linear(dim, dim)
self.proj_drop = nn.Dropout(proj_drop)
def forward(self, x, mask=None):
B, N, C = x.shape
qkv = (
self.qkv(x)
.reshape(B, N, 3, self.num_heads, C // self.num_heads)
.permute(2, 0, 3, 1, 4)
)
q, k, v = (
qkv[0],
qkv[1],
qkv[2],
) # make torchscript happy (cannot use tensor as tuple)
attn = (q @ k.transpose(-2, -1)) * self.scale
if mask is not None:
# if mask.dim() != x.dim():
# expanded_mask = mask[:, None, None, :].expand(B, 1, N, N)
# else:
# expanded_mask = mask
mask = mask.bool()
attn = attn.masked_fill(~mask, float("-inf"))
attn = attn.softmax(dim=-1)
attn = self.attn_drop(attn)
x = (attn @ v).transpose(1, 2).reshape(B, N, C)
x = self.proj(x)
x = self.proj_drop(x)
return x, attn
class MoELayer(nn.Module):
def __init__(
self,
dim,
num_heads,
expert_type,
use_sep_spatial_temp_experts=True,
has_hist=False,
mlp_ratio=4.0,
qkv_bias=False,
qk_scale=None,
drop=0.0,
attn_drop=0.0,
drop_path=0.0,
act_layer=nn.GELU,
norm_layer=LlamaRMSNorm,
):
super().__init__()
self.has_hist = has_hist
self.use_sep_spatial_temp_experts = use_sep_spatial_temp_experts
self.norm_att = norm_layer(dim)
self.attn = Attention(
dim,
num_heads=num_heads,
qkv_bias=qkv_bias,
qk_scale=qk_scale,
attn_drop=attn_drop,
proj_drop=drop,
)
# NOTE: drop path for stochastic depth, we shall see if this is better than dropout here
self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()
mlp_hidden_dim = int(dim * mlp_ratio)
if expert_type == 'modalities':
# EXPERT CONSTRUCTION
if use_sep_spatial_temp_experts:
# Spatial expert
self.norm_spatial = norm_layer(dim)
self.mlp_spatial = Mlp(
in_features=dim,
hidden_features=mlp_hidden_dim,
act_layer=act_layer,
drop=drop,
)
# Temporal expert
self.norm_temp = norm_layer(dim)
self.mlp_temp = Mlp(
in_features=dim,
hidden_features=mlp_hidden_dim,
act_layer=act_layer,
drop=drop,
)
# Vis expert
self.norm_vis = norm_layer(dim)
self.mlp_vis = Mlp(
in_features=dim,
hidden_features=mlp_hidden_dim,
act_layer=act_layer,
drop=drop,
)
# caption expert
self.norm_cap = norm_layer(dim)
self.mlp_cap = Mlp(
in_features=dim,
hidden_features=mlp_hidden_dim,
act_layer=act_layer,
drop=drop,
)
# history expert
if has_hist:
self.norm_hist = norm_layer(dim)
self.mlp_hist = Mlp(
in_features=dim,
hidden_features=mlp_hidden_dim,
act_layer=act_layer,
drop=drop,
)
elif expert_type == 'fusion':
# Fusion expert
self.norm_fusion = norm_layer(dim)
self.mlp_fusion = Mlp(
in_features=dim,
hidden_features=mlp_hidden_dim,
act_layer=act_layer,
drop=drop,
)
else:
raise ValueError
def forward(self, x, vis_feat_len, cap_feat_len, expert_flag, hist_feat_len=None, is_vid=False, mask=None, only_text=False, expert_permutation=None):
if self.has_hist:
assert hist_feat_len is not None
x_shortcut, attn = self.attn(self.norm_att(x), mask=mask)
x = x + self.drop_path(x_shortcut)
len_init = x.size(1)
# bs, h_dim = x.size(0), x.size(-1)
# device = x.device
# if only_text:
# # end_idx_caption = special_toks_indices.get('<history>', special_toks_indices['</s>'] + 1)
# # x = x[:, special_toks_indices['<caption>']: end_idx_caption, :]
# x = x + self.drop_path(self.mlp_cap(self.norm_cap(x)))
if expert_flag == 'modalities':
if self.use_sep_spatial_temp_experts:
x_spatial = x[:, :vis_feat_len]
if expert_permutation is not None:
if expert_permutation['spatial'] == 'temporal':
x_spatial = x_spatial + self.drop_path(self.mlp_temp(self.norm_temp(x_spatial)))
elif expert_permutation['spatial'] == 'caption':
x_spatial = x_spatial + self.drop_path(self.mlp_cap(self.norm_cap(x_spatial)))
elif expert_permutation['spatial'] == 'history':
x_spatial = x_spatial + self.drop_path(self.mlp_hist(self.norm_hist(x_spatial)))
elif expert_permutation['spatial'] == 'spatial':
x_spatial = x_spatial + self.drop_path(self.mlp_spatial(self.norm_spatial(x_spatial)))
x_vis = x_spatial
else:
x_spatial = x_spatial + self.drop_path(self.mlp_spatial(self.norm_spatial(x_spatial)))
x_vis = x_spatial
if is_vid:
x_temporal = x[:, vis_feat_len:2*vis_feat_len]
if expert_permutation is not None:
if expert_permutation['temporal'] == 'spatial':
x_temporal = x_temporal + self.drop_path(self.mlp_spatial(self.norm_spatial(x_temporal)))
elif expert_permutation['temporal'] == 'caption':
x_temporal = x_temporal + self.drop_path(self.mlp_cap(self.norm_cap(x_temporal)))
elif expert_permutation['temporal'] == 'history':
x_temporal = x_temporal + self.drop_path(self.mlp_hist(self.norm_hist(x_temporal)))
elif expert_permutation['temporal'] == 'temporal':
x_temporal = x_temporal + self.drop_path(self.mlp_temp(self.norm_temp(x_temporal)))
else:
x_temporal = x_temporal + self.drop_path(self.mlp_temp(self.norm_temp(x_temporal)))
x_vis = torch.concat([x_spatial, x_temporal], dim=1)
x_vis = x_vis + self.drop_path(self.mlp_vis(self.norm_vis(x_vis)))
else:
x_vis = x[:, :vis_feat_len]
x_vis = x_vis + self.drop_path(self.mlp_vis(self.norm_vis(x_vis)))
if self.has_hist:
x_caption = x[:, -(cap_feat_len + hist_feat_len): -hist_feat_len]
if expert_permutation is not None:
if expert_permutation['caption'] == 'spatial':
x_caption = x_caption + self.drop_path(self.mlp_spatial(self.norm_spatial(x_caption)))
elif expert_permutation['caption'] == 'temporal':
x_caption = x_caption + self.drop_path(self.mlp_temp(self.norm_temp(x_caption)))
elif expert_permutation['caption'] == 'history':
x_caption = x_caption + self.drop_path(self.mlp_hist(self.norm_hist(x_caption)))
elif expert_permutation['caption'] == 'caption':
x_caption = x_caption + self.drop_path(self.mlp_cap(self.norm_cap(x_caption)))
else:
x_caption = x_caption + self.drop_path(self.mlp_cap(self.norm_cap(x_caption)))
x_history = x[:, -hist_feat_len:]
if expert_permutation is not None:
if expert_permutation['history'] == 'spatial':
x_history = x_history + self.drop_path(self.mlp_spatial(self.norm_spatial(x_history)))
elif expert_permutation['history'] == 'temporal':
x_history = x_history + self.drop_path(self.mlp_temp(self.norm_temp(x_history)))
elif expert_permutation['history'] == 'caption':
x_history = x_history + self.drop_path(self.mlp_cap(self.norm_cap(x_history)))
elif expert_permutation['history'] == 'history':
x_history = x_history + self.drop_path(self.mlp_hist(self.norm_hist(x_history)))
else:
x_history = x_history + self.drop_path(self.mlp_hist(self.norm_hist(x_history)))
# concat the features back
x = torch.cat([x_vis, x_caption, x_history], dim=1)
else:
x_caption = x[:, -cap_feat_len:]
x_caption = x_caption + self.drop_path(self.mlp_cap(self.norm_cap(x_caption)))
x = torch.cat([x_vis, x_caption], dim=1)
assert x.size(1) == len_init, 'Reconstructed features length is {} != original features len = {}'.format(
x.size(1), len_init
)
elif expert_flag == 'fusion':
x = x + self.drop_path(self.mlp_fusion(self.norm_fusion(x)))
return x
class Pooler(nn.Module):
def __init__(self, hidden_size):
super(Pooler, self).__init__()
self.dense = nn.Linear(hidden_size, hidden_size)
self.activation = nn.Tanh()
def forward(self, hidden_states):
pooled_states = hidden_states[:, 0]
pooled_output = self.dense(pooled_states)
pooled_output = self.activation(pooled_output)
return pooled_output