olivame
57ba85d147
attention.py — 3处改动： 1. __init__ 添加 _kv_fused = False 标志 2.新增 fuse_kv() 方法：将 to_k + to_v → to_kv，同时处理 _ip/_as/_aa 辅助 KV 对 2. bmm_forward 两个分支加_kv_fused 判断，用to_kv().chunk(2, dim=-1) 替代分别调用
1006 lines
43 KiB
Python
1006 lines
43 KiB
Python
import torch
|
|
import torch.nn.functional as F
|
|
|
|
from torch import nn, einsum
|
|
from einops import rearrange, repeat
|
|
from functools import partial
|
|
|
|
try:
|
|
import xformers
|
|
import xformers.ops
|
|
XFORMERS_IS_AVAILBLE = True
|
|
except:
|
|
XFORMERS_IS_AVAILBLE = False
|
|
|
|
from unifolm_wma.utils.common import (
|
|
checkpoint,
|
|
exists,
|
|
default,
|
|
)
|
|
from unifolm_wma.utils.basics import zero_module
|
|
|
|
|
|
class RelativePosition(nn.Module):
|
|
""" https://github.com/evelinehong/Transformer_Relative_Position_PyTorch/blob/master/relative_position.py """
|
|
|
|
def __init__(self, num_units, max_relative_position):
|
|
super().__init__()
|
|
self.num_units = num_units
|
|
self.max_relative_position = max_relative_position
|
|
self.embeddings_table = nn.Parameter(
|
|
torch.Tensor(max_relative_position * 2 + 1, num_units))
|
|
nn.init.xavier_uniform_(self.embeddings_table)
|
|
|
|
def forward(self, length_q, length_k):
|
|
device = self.embeddings_table.device
|
|
range_vec_q = torch.arange(length_q, device=device)
|
|
range_vec_k = torch.arange(length_k, device=device)
|
|
distance_mat = range_vec_k[None, :] - range_vec_q[:, None]
|
|
distance_mat_clipped = torch.clamp(distance_mat,
|
|
-self.max_relative_position,
|
|
self.max_relative_position)
|
|
final_mat = distance_mat_clipped + self.max_relative_position
|
|
final_mat = final_mat.long()
|
|
embeddings = self.embeddings_table[final_mat]
|
|
return embeddings
|
|
|
|
|
|
class CrossAttention(nn.Module):
|
|
|
|
def __init__(self,
|
|
query_dim,
|
|
context_dim=None,
|
|
heads=8,
|
|
dim_head=64,
|
|
dropout=0.,
|
|
relative_position=False,
|
|
temporal_length=None,
|
|
video_length=None,
|
|
agent_state_context_len=2,
|
|
agent_action_context_len=16,
|
|
image_cross_attention=False,
|
|
image_cross_attention_scale=1.0,
|
|
agent_state_cross_attention_scale=1.0,
|
|
agent_action_cross_attention_scale=1.0,
|
|
cross_attention_scale_learnable=False,
|
|
text_context_len=77):
|
|
super().__init__()
|
|
inner_dim = dim_head * heads
|
|
context_dim = default(context_dim, query_dim)
|
|
|
|
self.scale = dim_head**-0.5
|
|
self.heads = heads
|
|
self.dim_head = dim_head
|
|
self.to_q = nn.Linear(query_dim, inner_dim, bias=False)
|
|
self.to_k = nn.Linear(context_dim, inner_dim, bias=False)
|
|
self.to_v = nn.Linear(context_dim, inner_dim, bias=False)
|
|
|
|
self.to_out = nn.Sequential(nn.Linear(inner_dim, query_dim),
|
|
nn.Dropout(dropout))
|
|
|
|
self.relative_position = relative_position
|
|
if self.relative_position:
|
|
assert (temporal_length is not None)
|
|
self.relative_position_k = RelativePosition(
|
|
num_units=dim_head, max_relative_position=temporal_length)
|
|
self.relative_position_v = RelativePosition(
|
|
num_units=dim_head, max_relative_position=temporal_length)
|
|
else:
|
|
## bmm fused-scale attention for all non-relative-position cases
|
|
self.forward = self.bmm_forward
|
|
|
|
self.video_length = video_length
|
|
self.image_cross_attention = image_cross_attention
|
|
self.image_cross_attention_scale = image_cross_attention_scale
|
|
self.agent_state_cross_attention_scale = agent_state_cross_attention_scale
|
|
self.agent_action_cross_attention_scale = agent_action_cross_attention_scale
|
|
self.text_context_len = text_context_len
|
|
self.agent_state_context_len = agent_state_context_len
|
|
self.agent_action_context_len = agent_action_context_len
|
|
self._kv_cache = {}
|
|
self._kv_cache_enabled = False
|
|
self._kv_fused = False
|
|
|
|
self.cross_attention_scale_learnable = cross_attention_scale_learnable
|
|
if self.image_cross_attention:
|
|
self.to_k_ip = nn.Linear(context_dim, inner_dim, bias=False)
|
|
self.to_v_ip = nn.Linear(context_dim, inner_dim, bias=False)
|
|
self.to_k_as = nn.Linear(context_dim, inner_dim, bias=False)
|
|
self.to_v_as = nn.Linear(context_dim, inner_dim, bias=False)
|
|
self.to_k_aa = nn.Linear(context_dim, inner_dim, bias=False)
|
|
self.to_v_aa = nn.Linear(context_dim, inner_dim, bias=False)
|
|
if cross_attention_scale_learnable:
|
|
self.register_parameter('alpha_ctx',
|
|
nn.Parameter(torch.tensor(0.)))
|
|
self.register_parameter('alpha_cas',
|
|
nn.Parameter(torch.tensor(0.)))
|
|
self.register_parameter('alpha_caa',
|
|
nn.Parameter(torch.tensor(0.)))
|
|
|
|
def fuse_kv(self):
|
|
"""Fuse to_k/to_v into to_kv (2 Linear → 1). Works for all layers."""
|
|
k_w = self.to_k.weight # (inner_dim, context_dim)
|
|
v_w = self.to_v.weight
|
|
self.to_kv = nn.Linear(k_w.shape[1], k_w.shape[0] * 2, bias=False)
|
|
self.to_kv.weight = nn.Parameter(torch.cat([k_w, v_w], dim=0))
|
|
del self.to_k, self.to_v
|
|
if self.image_cross_attention:
|
|
for suffix in ('_ip', '_as', '_aa'):
|
|
k_attr = f'to_k{suffix}'
|
|
v_attr = f'to_v{suffix}'
|
|
kw = getattr(self, k_attr).weight
|
|
vw = getattr(self, v_attr).weight
|
|
fused = nn.Linear(kw.shape[1], kw.shape[0] * 2, bias=False)
|
|
fused.weight = nn.Parameter(torch.cat([kw, vw], dim=0))
|
|
setattr(self, f'to_kv{suffix}', fused)
|
|
delattr(self, k_attr)
|
|
delattr(self, v_attr)
|
|
self._kv_fused = True
|
|
return True
|
|
|
|
def forward(self, x, context=None, mask=None):
|
|
spatial_self_attn = (context is None)
|
|
k_ip, v_ip, out_ip = None, None, None
|
|
k_as, v_as, out_as = None, None, None
|
|
k_aa, v_aa, out_aa = None, None, None
|
|
|
|
h = self.heads
|
|
q = self.to_q(x)
|
|
context = default(context, x)
|
|
|
|
if self.image_cross_attention and not spatial_self_attn:
|
|
# assert 1 > 2, ">>> ERROR: should setup xformers and use efficient_forward ..."
|
|
context_agent_state = context[:, :self.agent_state_context_len, :]
|
|
context_agent_action = context[:,
|
|
self.agent_state_context_len:self.
|
|
agent_state_context_len +
|
|
self.agent_action_context_len, :]
|
|
context_ins = context[:, self.agent_state_context_len +
|
|
self.agent_action_context_len:self.
|
|
agent_state_context_len +
|
|
self.agent_action_context_len +
|
|
self.text_context_len, :]
|
|
context_image = context[:, self.agent_state_context_len +
|
|
self.agent_action_context_len +
|
|
self.text_context_len:, :]
|
|
|
|
k = self.to_k(context_ins)
|
|
v = self.to_v(context_ins)
|
|
k_ip = self.to_k_ip(context_image)
|
|
v_ip = self.to_v_ip(context_image)
|
|
k_as = self.to_k_as(context_agent_state)
|
|
v_as = self.to_v_as(context_agent_state)
|
|
k_aa = self.to_k_aa(context_agent_action)
|
|
v_aa = self.to_v_aa(context_agent_action)
|
|
else:
|
|
if not spatial_self_attn:
|
|
context = context[:, :self.text_context_len, :]
|
|
k = self.to_k(context)
|
|
v = self.to_v(context)
|
|
|
|
q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h=h),
|
|
(q, k, v))
|
|
|
|
sim = torch.einsum('b i d, b j d -> b i j', q, k) * self.scale
|
|
if self.relative_position:
|
|
len_q, len_k, len_v = q.shape[1], k.shape[1], v.shape[1]
|
|
k2 = self.relative_position_k(len_q, len_k)
|
|
sim2 = einsum('b t d, t s d -> b t s', q,
|
|
k2) * self.scale # TODO check
|
|
sim += sim2
|
|
del k
|
|
|
|
if exists(mask):
|
|
## feasible for causal attention mask only
|
|
max_neg_value = -torch.finfo(sim.dtype).max
|
|
mask = repeat(mask, 'b i j -> (b h) i j', h=h)
|
|
sim.masked_fill_(~(mask > 0.5), max_neg_value)
|
|
|
|
# attention, what we cannot get enough of
|
|
with torch.amp.autocast('cuda', enabled=False):
|
|
sim = sim.softmax(dim=-1)
|
|
|
|
out = torch.einsum('b i j, b j d -> b i d', sim, v)
|
|
if self.relative_position:
|
|
v2 = self.relative_position_v(len_q, len_v)
|
|
out2 = einsum('b t s, t s d -> b t d', sim, v2) # TODO check
|
|
out += out2
|
|
out = rearrange(out, '(b h) n d -> b n (h d)', h=h)
|
|
|
|
if k_ip is not None and k_as is not None and k_aa is not None:
|
|
## for image cross-attention
|
|
k_ip, v_ip = map(
|
|
lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h=h),
|
|
(k_ip, v_ip))
|
|
sim_ip = torch.einsum('b i d, b j d -> b i j', q,
|
|
k_ip) * self.scale
|
|
del k_ip
|
|
with torch.amp.autocast('cuda', enabled=False):
|
|
sim_ip = sim_ip.softmax(dim=-1)
|
|
out_ip = torch.einsum('b i j, b j d -> b i d', sim_ip, v_ip)
|
|
out_ip = rearrange(out_ip, '(b h) n d -> b n (h d)', h=h)
|
|
|
|
## for agent state cross-attention
|
|
k_as, v_as = map(
|
|
lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h=h),
|
|
(k_as, v_as))
|
|
sim_as = torch.einsum('b i d, b j d -> b i j', q,
|
|
k_as) * self.scale
|
|
del k_as
|
|
with torch.amp.autocast('cuda', enabled=False):
|
|
sim_as = sim_as.softmax(dim=-1)
|
|
out_as = torch.einsum('b i j, b j d -> b i d', sim_as, v_as)
|
|
out_as = rearrange(out_as, '(b h) n d -> b n (h d)', h=h)
|
|
|
|
## for agent action cross-attention
|
|
k_aa, v_aa = map(
|
|
lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h=h),
|
|
(k_aa, v_aa))
|
|
sim_aa = torch.einsum('b i d, b j d -> b i j', q,
|
|
k_aa) * self.scale
|
|
del k_aa
|
|
with torch.amp.autocast('cuda', enabled=False):
|
|
sim_aa = sim_aa.softmax(dim=-1)
|
|
out_aa = torch.einsum('b i j, b j d -> b i d', sim_aa, v_aa)
|
|
out_aa = rearrange(out_aa, '(b h) n d -> b n (h d)', h=h)
|
|
|
|
if out_ip is not None and out_as is not None and out_aa is not None:
|
|
if self.cross_attention_scale_learnable:
|
|
out = out + \
|
|
self.image_cross_attention_scale * out_ip * (torch.tanh(self.alpha_ctx) + 1) + \
|
|
self.agent_state_cross_attention_scale * out_as * (torch.tanh(self.alpha_cas) + 1) + \
|
|
self.agent_action_cross_attention_scale * out_aa * (torch.tanh(self.alpha_caa) + 1)
|
|
else:
|
|
out = out + \
|
|
self.image_cross_attention_scale * out_ip + \
|
|
self.agent_state_cross_attention_scale * out_as + \
|
|
self.agent_action_cross_attention_scale * out_aa
|
|
|
|
return self.to_out(out)
|
|
|
|
def bmm_forward(self, x, context=None, mask=None):
|
|
spatial_self_attn = (context is None)
|
|
k_ip, v_ip, out_ip = None, None, None
|
|
k_as, v_as, out_as = None, None, None
|
|
k_aa, v_aa, out_aa = None, None, None
|
|
|
|
h = self.heads
|
|
q = self.to_q(x)
|
|
context = default(context, x)
|
|
|
|
use_cache = self._kv_cache_enabled and not spatial_self_attn
|
|
cache_hit = use_cache and len(self._kv_cache) > 0
|
|
|
|
if cache_hit:
|
|
# Reuse cached K/V (already in (b*h, n, d) shape)
|
|
k = self._kv_cache['k']
|
|
v = self._kv_cache['v']
|
|
if 'k_ip' in self._kv_cache:
|
|
k_ip = self._kv_cache['k_ip']
|
|
v_ip = self._kv_cache['v_ip']
|
|
k_as = self._kv_cache['k_as']
|
|
v_as = self._kv_cache['v_as']
|
|
k_aa = self._kv_cache['k_aa']
|
|
v_aa = self._kv_cache['v_aa']
|
|
q = rearrange(q, 'b n (h d) -> (b h) n d', h=h)
|
|
elif self.image_cross_attention and not spatial_self_attn:
|
|
context_agent_state = context[:, :self.agent_state_context_len, :]
|
|
context_agent_action = context[:,
|
|
self.agent_state_context_len:self.
|
|
agent_state_context_len +
|
|
self.agent_action_context_len, :]
|
|
context_ins = context[:, self.agent_state_context_len +
|
|
self.agent_action_context_len:self.
|
|
agent_state_context_len +
|
|
self.agent_action_context_len +
|
|
self.text_context_len, :]
|
|
context_image = context[:, self.agent_state_context_len +
|
|
self.agent_action_context_len +
|
|
self.text_context_len:, :]
|
|
|
|
if self._kv_fused:
|
|
k, v = self.to_kv(context_ins).chunk(2, dim=-1)
|
|
k_ip, v_ip = self.to_kv_ip(context_image).chunk(2, dim=-1)
|
|
k_as, v_as = self.to_kv_as(context_agent_state).chunk(2, dim=-1)
|
|
k_aa, v_aa = self.to_kv_aa(context_agent_action).chunk(2, dim=-1)
|
|
else:
|
|
k = self.to_k(context_ins)
|
|
v = self.to_v(context_ins)
|
|
k_ip = self.to_k_ip(context_image)
|
|
v_ip = self.to_v_ip(context_image)
|
|
k_as = self.to_k_as(context_agent_state)
|
|
v_as = self.to_v_as(context_agent_state)
|
|
k_aa = self.to_k_aa(context_agent_action)
|
|
v_aa = self.to_v_aa(context_agent_action)
|
|
|
|
q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h=h),
|
|
(q, k, v))
|
|
k_ip, v_ip = map(lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h=h),
|
|
(k_ip, v_ip))
|
|
k_as, v_as = map(lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h=h),
|
|
(k_as, v_as))
|
|
k_aa, v_aa = map(lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h=h),
|
|
(k_aa, v_aa))
|
|
|
|
if use_cache:
|
|
self._kv_cache = {
|
|
'k': k, 'v': v,
|
|
'k_ip': k_ip, 'v_ip': v_ip,
|
|
'k_as': k_as, 'v_as': v_as,
|
|
'k_aa': k_aa, 'v_aa': v_aa,
|
|
}
|
|
else:
|
|
if not spatial_self_attn:
|
|
context = context[:, :self.text_context_len, :]
|
|
if self._kv_fused:
|
|
k, v = self.to_kv(context).chunk(2, dim=-1)
|
|
else:
|
|
k = self.to_k(context)
|
|
v = self.to_v(context)
|
|
|
|
q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h=h),
|
|
(q, k, v))
|
|
|
|
if use_cache:
|
|
self._kv_cache = {'k': k, 'v': v}
|
|
|
|
# baddbmm: fuse scale into GEMM → one kernel instead of matmul + mul
|
|
sim = torch.baddbmm(
|
|
torch.empty(q.shape[0], q.shape[1], k.shape[1], dtype=q.dtype, device=q.device),
|
|
q, k.transpose(-1, -2), beta=0, alpha=self.scale)
|
|
|
|
if exists(mask):
|
|
max_neg_value = -torch.finfo(sim.dtype).max
|
|
mask = repeat(mask, 'b i j -> (b h) i j', h=h)
|
|
sim.masked_fill_(~(mask > 0.5), max_neg_value)
|
|
|
|
with torch.amp.autocast('cuda', enabled=False):
|
|
sim = sim.softmax(dim=-1)
|
|
|
|
out = torch.bmm(sim, v)
|
|
out = rearrange(out, '(b h) n d -> b n (h d)', h=h)
|
|
|
|
if k_ip is not None and k_as is not None and k_aa is not None:
|
|
## image cross-attention (k_ip/v_ip already in (b*h, n, d) shape)
|
|
sim_ip = torch.baddbmm(
|
|
torch.empty(q.shape[0], q.shape[1], k_ip.shape[1], dtype=q.dtype, device=q.device),
|
|
q, k_ip.transpose(-1, -2), beta=0, alpha=self.scale)
|
|
with torch.amp.autocast('cuda', enabled=False):
|
|
sim_ip = sim_ip.softmax(dim=-1)
|
|
out_ip = torch.bmm(sim_ip, v_ip)
|
|
out_ip = rearrange(out_ip, '(b h) n d -> b n (h d)', h=h)
|
|
|
|
## agent state cross-attention (k_as/v_as already in (b*h, n, d) shape)
|
|
sim_as = torch.baddbmm(
|
|
torch.empty(q.shape[0], q.shape[1], k_as.shape[1], dtype=q.dtype, device=q.device),
|
|
q, k_as.transpose(-1, -2), beta=0, alpha=self.scale)
|
|
with torch.amp.autocast('cuda', enabled=False):
|
|
sim_as = sim_as.softmax(dim=-1)
|
|
out_as = torch.bmm(sim_as, v_as)
|
|
out_as = rearrange(out_as, '(b h) n d -> b n (h d)', h=h)
|
|
|
|
## agent action cross-attention (k_aa/v_aa already in (b*h, n, d) shape)
|
|
sim_aa = torch.baddbmm(
|
|
torch.empty(q.shape[0], q.shape[1], k_aa.shape[1], dtype=q.dtype, device=q.device),
|
|
q, k_aa.transpose(-1, -2), beta=0, alpha=self.scale)
|
|
with torch.amp.autocast('cuda', enabled=False):
|
|
sim_aa = sim_aa.softmax(dim=-1)
|
|
out_aa = torch.bmm(sim_aa, v_aa)
|
|
out_aa = rearrange(out_aa, '(b h) n d -> b n (h d)', h=h)
|
|
|
|
if out_ip is not None and out_as is not None and out_aa is not None:
|
|
if self.cross_attention_scale_learnable:
|
|
out = out + \
|
|
self.image_cross_attention_scale * out_ip * (torch.tanh(self.alpha_ctx) + 1) + \
|
|
self.agent_state_cross_attention_scale * out_as * (torch.tanh(self.alpha_cas) + 1) + \
|
|
self.agent_action_cross_attention_scale * out_aa * (torch.tanh(self.alpha_caa) + 1)
|
|
else:
|
|
out = out + \
|
|
self.image_cross_attention_scale * out_ip + \
|
|
self.agent_state_cross_attention_scale * out_as + \
|
|
self.agent_action_cross_attention_scale * out_aa
|
|
|
|
return self.to_out(out)
|
|
|
|
def efficient_forward(self, x, context=None, mask=None):
|
|
spatial_self_attn = (context is None)
|
|
k, v, out = None, None, None
|
|
k_ip, v_ip, out_ip = None, None, None
|
|
k_as, v_as, out_as = None, None, None
|
|
k_aa, v_aa, out_aa = None, None, None
|
|
|
|
q = self.to_q(x)
|
|
context = default(context, x)
|
|
|
|
if self.image_cross_attention and not spatial_self_attn:
|
|
if context.shape[1] == self.text_context_len + self.video_length:
|
|
context_ins, context_image = context[:, :self.text_context_len, :], context[:,self.text_context_len:, :]
|
|
k = self.to_k(context)
|
|
v = self.to_v(context)
|
|
k_ip = self.to_k_ip(context_image)
|
|
v_ip = self.to_v_ip(context_image)
|
|
elif context.shape[1] == self.agent_state_context_len + self.text_context_len + self.video_length:
|
|
context_agent_state = context[:, :self.agent_state_context_len, :]
|
|
context_ins = context[:, self.agent_state_context_len:self.agent_state_context_len+self.text_context_len, :]
|
|
context_image = context[:, self.agent_state_context_len+self.text_context_len:, :]
|
|
k = self.to_k(context_ins)
|
|
v = self.to_v(context_ins)
|
|
k_ip = self.to_k_ip(context_image)
|
|
v_ip = self.to_v_ip(context_image)
|
|
k_as = self.to_k_as(context_agent_state)
|
|
v_as = self.to_v_as(context_agent_state)
|
|
else:
|
|
context_agent_state = context[:, :self.agent_state_context_len, :]
|
|
context_agent_action = context[:, self.agent_state_context_len:self.agent_state_context_len+self.agent_action_context_len, :]
|
|
context_ins = context[:, self.agent_state_context_len+self.agent_action_context_len:self.agent_state_context_len+self.agent_action_context_len+self.text_context_len, :]
|
|
context_image = context[:, self.agent_state_context_len+self.agent_action_context_len+self.text_context_len:, :]
|
|
|
|
k = self.to_k(context_ins)
|
|
v = self.to_v(context_ins)
|
|
k_ip = self.to_k_ip(context_image)
|
|
v_ip = self.to_v_ip(context_image)
|
|
k_as = self.to_k_as(context_agent_state)
|
|
v_as = self.to_v_as(context_agent_state)
|
|
k_aa = self.to_k_aa(context_agent_action)
|
|
v_aa = self.to_v_aa(context_agent_action)
|
|
|
|
attn_mask_aa = self._get_attn_mask_aa(x.shape[0],
|
|
q.shape[1],
|
|
k_aa.shape[1],
|
|
block_size=16,
|
|
device=k_aa.device)
|
|
else:
|
|
if not spatial_self_attn:
|
|
assert 1 > 2, ">>> ERROR: you should never go into here ..."
|
|
context = context[:, :self.text_context_len, :]
|
|
k = self.to_k(context)
|
|
v = self.to_v(context)
|
|
|
|
b, _, _ = q.shape
|
|
q = q.unsqueeze(3).reshape(b, q.shape[1], self.heads, self.dim_head).permute(0, 2, 1, 3).reshape(b * self.heads, q.shape[1], self.dim_head).contiguous()
|
|
if k is not None:
|
|
k, v = map(
|
|
lambda t: t.unsqueeze(3).reshape(b, t.shape[
|
|
1], self.heads, self.dim_head).permute(0, 2, 1, 3).reshape(
|
|
b * self.heads, t.shape[1], self.dim_head).contiguous(),
|
|
(k, v),
|
|
)
|
|
out = xformers.ops.memory_efficient_attention(q,
|
|
k,
|
|
v,
|
|
attn_bias=None,
|
|
op=None)
|
|
out = (out.unsqueeze(0).reshape(
|
|
b, self.heads, out.shape[1],
|
|
self.dim_head).permute(0, 2, 1,
|
|
3).reshape(b, out.shape[1],
|
|
self.heads * self.dim_head))
|
|
|
|
if k_ip is not None:
|
|
# For image cross-attention
|
|
k_ip, v_ip = map(
|
|
lambda t: t.unsqueeze(3).reshape(b, t.shape[
|
|
1], self.heads, self.dim_head).permute(0, 2, 1, 3).reshape(
|
|
b * self.heads, t.shape[1], self.dim_head).contiguous(
|
|
),
|
|
(k_ip, v_ip),
|
|
)
|
|
out_ip = xformers.ops.memory_efficient_attention(q,
|
|
k_ip,
|
|
v_ip,
|
|
attn_bias=None,
|
|
op=None)
|
|
out_ip = (out_ip.unsqueeze(0).reshape(
|
|
b, self.heads, out_ip.shape[1],
|
|
self.dim_head).permute(0, 2, 1,
|
|
3).reshape(b, out_ip.shape[1],
|
|
self.heads * self.dim_head))
|
|
|
|
if k_as is not None:
|
|
# For agent state cross-attention
|
|
k_as, v_as = map(
|
|
lambda t: t.unsqueeze(3).reshape(b, t.shape[
|
|
1], self.heads, self.dim_head).permute(0, 2, 1, 3).reshape(
|
|
b * self.heads, t.shape[1], self.dim_head).contiguous(
|
|
),
|
|
(k_as, v_as),
|
|
)
|
|
out_as = xformers.ops.memory_efficient_attention(q,
|
|
k_as,
|
|
v_as,
|
|
attn_bias=None,
|
|
op=None)
|
|
out_as = (out_as.unsqueeze(0).reshape(
|
|
b, self.heads, out_as.shape[1],
|
|
self.dim_head).permute(0, 2, 1,
|
|
3).reshape(b, out_as.shape[1],
|
|
self.heads * self.dim_head))
|
|
if k_aa is not None:
|
|
# For agent action cross-attention
|
|
k_aa, v_aa = map(
|
|
lambda t: t.unsqueeze(3).reshape(b, t.shape[
|
|
1], self.heads, self.dim_head).permute(0, 2, 1, 3).reshape(
|
|
b * self.heads, t.shape[1], self.dim_head).contiguous(
|
|
),
|
|
(k_aa, v_aa),
|
|
)
|
|
|
|
attn_mask_aa = attn_mask_aa.unsqueeze(1).repeat(1,self.heads,1,1).reshape(
|
|
b * self.heads, attn_mask_aa.shape[1], attn_mask_aa.shape[2])
|
|
attn_mask_aa = attn_mask_aa.to(q.dtype)
|
|
|
|
out_aa = xformers.ops.memory_efficient_attention(
|
|
q, k_aa, v_aa, attn_bias=attn_mask_aa, op=None)
|
|
|
|
out_aa = (out_aa.unsqueeze(0).reshape(
|
|
b, self.heads, out_aa.shape[1],
|
|
self.dim_head).permute(0, 2, 1,
|
|
3).reshape(b, out_aa.shape[1],
|
|
self.heads * self.dim_head))
|
|
if exists(mask):
|
|
raise NotImplementedError
|
|
|
|
out = 0.0 if out is None else out
|
|
out_ip = 0.0 if out_ip is None else out_ip
|
|
out_as = 0.0 if out_as is None else out_as
|
|
out_aa = 0.0 if out_aa is None else out_aa
|
|
|
|
if self.cross_attention_scale_learnable:
|
|
out = out + \
|
|
self.image_cross_attention_scale * out_ip * (torch.tanh(self.alpha_ctx) + 1) + \
|
|
self.agent_state_cross_attention_scale * out_as * (torch.tanh(self.alpha_cas) + 1) + \
|
|
self.agent_action_cross_attention_scale * out_aa * (torch.tanh(self.alpha_caa) + 1)
|
|
|
|
else:
|
|
out = out + \
|
|
self.image_cross_attention_scale * out_ip + \
|
|
self.agent_state_cross_attention_scale * out_as + \
|
|
self.agent_action_cross_attention_scale * out_aa
|
|
|
|
return self.to_out(out)
|
|
|
|
def _get_attn_mask_aa(self, b, l1, l2, block_size=16, device=None):
|
|
cache_key = (b, l1, l2, block_size)
|
|
if hasattr(self, '_attn_mask_aa_cache_key') and self._attn_mask_aa_cache_key == cache_key:
|
|
cached = self._attn_mask_aa_cache
|
|
if device is not None and cached.device != torch.device(device):
|
|
cached = cached.to(device)
|
|
self._attn_mask_aa_cache = cached
|
|
return cached
|
|
|
|
target_device = device if device is not None else 'cpu'
|
|
num_token = l2 // block_size
|
|
start_positions = ((torch.arange(b, device=target_device) % block_size) + 1) * num_token
|
|
col_indices = torch.arange(l2, device=target_device)
|
|
mask_2d = col_indices.unsqueeze(0) >= start_positions.unsqueeze(1)
|
|
mask = mask_2d.unsqueeze(1).expand(b, l1, l2)
|
|
attn_mask = torch.zeros(b, l1, l2, dtype=torch.bfloat16, device=target_device)
|
|
attn_mask[mask] = float('-inf')
|
|
|
|
self._attn_mask_aa_cache_key = cache_key
|
|
self._attn_mask_aa_cache = attn_mask
|
|
return attn_mask
|
|
|
|
|
|
def enable_cross_attn_kv_cache(module):
|
|
for m in module.modules():
|
|
if isinstance(m, CrossAttention):
|
|
m._kv_cache_enabled = True
|
|
m._kv_cache = {}
|
|
|
|
|
|
def disable_cross_attn_kv_cache(module):
|
|
for m in module.modules():
|
|
if isinstance(m, CrossAttention):
|
|
m._kv_cache_enabled = False
|
|
m._kv_cache = {}
|
|
|
|
|
|
class BasicTransformerBlock(nn.Module):
|
|
|
|
def __init__(self,
|
|
dim,
|
|
n_heads,
|
|
d_head,
|
|
dropout=0.,
|
|
context_dim=None,
|
|
gated_ff=True,
|
|
checkpoint=True,
|
|
disable_self_attn=False,
|
|
attention_cls=None,
|
|
video_length=None,
|
|
agent_state_context_len=2,
|
|
agent_action_context_len=16,
|
|
image_cross_attention=False,
|
|
image_cross_attention_scale=1.0,
|
|
cross_attention_scale_learnable=False,
|
|
text_context_len=77):
|
|
super().__init__()
|
|
attn_cls = CrossAttention if attention_cls is None else attention_cls
|
|
self.disable_self_attn = disable_self_attn
|
|
self.attn1 = attn_cls(
|
|
query_dim=dim,
|
|
heads=n_heads,
|
|
dim_head=d_head,
|
|
dropout=dropout,
|
|
context_dim=context_dim if self.disable_self_attn else None)
|
|
self.ff = FeedForward(dim, dropout=dropout, glu=gated_ff)
|
|
self.attn2 = attn_cls(
|
|
query_dim=dim,
|
|
context_dim=context_dim,
|
|
heads=n_heads,
|
|
dim_head=d_head,
|
|
dropout=dropout,
|
|
video_length=video_length,
|
|
agent_state_context_len=agent_state_context_len,
|
|
agent_action_context_len=agent_action_context_len,
|
|
image_cross_attention=image_cross_attention,
|
|
image_cross_attention_scale=image_cross_attention_scale,
|
|
cross_attention_scale_learnable=cross_attention_scale_learnable,
|
|
text_context_len=text_context_len)
|
|
self.image_cross_attention = image_cross_attention
|
|
|
|
self.norm1 = nn.LayerNorm(dim)
|
|
self.norm2 = nn.LayerNorm(dim)
|
|
self.norm3 = nn.LayerNorm(dim)
|
|
self.checkpoint = checkpoint
|
|
|
|
def forward(self, x, context=None, mask=None, **kwargs):
|
|
# implementation tricks: because checkpointing doesn't support non-tensor (e.g. None or scalar) arguments
|
|
input_tuple = (
|
|
x,
|
|
) # should not be (x), otherwise *input_tuple will decouple x into multiple arguments
|
|
if context is not None:
|
|
input_tuple = (x, context)
|
|
if mask is not None:
|
|
forward_mask = partial(self._forward, mask=mask)
|
|
return checkpoint(forward_mask, (x, ), self.parameters(),
|
|
self.checkpoint)
|
|
return checkpoint(self._forward, input_tuple, self.parameters(),
|
|
self.checkpoint)
|
|
|
|
def _forward(self, x, context=None, mask=None):
|
|
x = self.attn1(self.norm1(x),
|
|
context=context if self.disable_self_attn else None,
|
|
mask=mask) + x
|
|
x = self.attn2(self.norm2(x), context=context, mask=mask) + x
|
|
x = self.ff(self.norm3(x)) + x
|
|
return x
|
|
|
|
|
|
class SpatialTransformer(nn.Module):
|
|
"""
|
|
Transformer block for image-like data in spatial axis.
|
|
First, project the input (aka embedding)
|
|
and reshape to b, t, d.
|
|
Then apply standard transformer action.
|
|
Finally, reshape to image
|
|
NEW: use_linear for more efficiency instead of the 1x1 convs
|
|
"""
|
|
|
|
def __init__(self,
|
|
in_channels,
|
|
n_heads,
|
|
d_head,
|
|
depth=1,
|
|
dropout=0.,
|
|
context_dim=None,
|
|
use_checkpoint=True,
|
|
disable_self_attn=False,
|
|
use_linear=False,
|
|
video_length=None,
|
|
agent_state_context_len=2,
|
|
agent_action_context_len=16,
|
|
image_cross_attention=False,
|
|
cross_attention_scale_learnable=False):
|
|
super().__init__()
|
|
self.in_channels = in_channels
|
|
inner_dim = n_heads * d_head
|
|
self.norm = torch.nn.GroupNorm(num_groups=32,
|
|
num_channels=in_channels,
|
|
eps=1e-6,
|
|
affine=True)
|
|
if not use_linear:
|
|
self.proj_in = nn.Conv2d(in_channels,
|
|
inner_dim,
|
|
kernel_size=1,
|
|
stride=1,
|
|
padding=0)
|
|
else:
|
|
self.proj_in = nn.Linear(in_channels, inner_dim)
|
|
|
|
attention_cls = None
|
|
self.transformer_blocks = nn.ModuleList([
|
|
BasicTransformerBlock(
|
|
inner_dim,
|
|
n_heads,
|
|
d_head,
|
|
dropout=dropout,
|
|
context_dim=context_dim,
|
|
disable_self_attn=disable_self_attn,
|
|
checkpoint=use_checkpoint,
|
|
attention_cls=attention_cls,
|
|
video_length=video_length,
|
|
agent_state_context_len=agent_state_context_len,
|
|
agent_action_context_len=agent_action_context_len,
|
|
image_cross_attention=image_cross_attention,
|
|
cross_attention_scale_learnable=cross_attention_scale_learnable,
|
|
) for d in range(depth)
|
|
])
|
|
if not use_linear:
|
|
self.proj_out = zero_module(
|
|
nn.Conv2d(inner_dim,
|
|
in_channels,
|
|
kernel_size=1,
|
|
stride=1,
|
|
padding=0))
|
|
else:
|
|
self.proj_out = zero_module(nn.Linear(inner_dim, in_channels))
|
|
self.use_linear = use_linear
|
|
|
|
def forward(self, x, context=None, **kwargs):
|
|
b, c, h, w = x.shape
|
|
x_in = x
|
|
x = self.norm(x)
|
|
if not self.use_linear:
|
|
x = self.proj_in(x)
|
|
x = rearrange(x, 'b c h w -> b (h w) c').contiguous()
|
|
if self.use_linear:
|
|
x = self.proj_in(x)
|
|
for i, block in enumerate(self.transformer_blocks):
|
|
x = block(x, context=context, **kwargs)
|
|
if self.use_linear:
|
|
x = self.proj_out(x)
|
|
x = rearrange(x, 'b (h w) c -> b c h w', h=h, w=w).contiguous()
|
|
if not self.use_linear:
|
|
x = self.proj_out(x)
|
|
return x + x_in
|
|
|
|
|
|
class TemporalTransformer(nn.Module):
|
|
"""
|
|
Transformer block for image-like data in temporal axis.
|
|
First, reshape to b, t, d.
|
|
Then apply standard transformer action.
|
|
Finally, reshape to image
|
|
"""
|
|
|
|
def __init__(self,
|
|
in_channels,
|
|
n_heads,
|
|
d_head,
|
|
depth=1,
|
|
dropout=0.,
|
|
context_dim=None,
|
|
use_checkpoint=True,
|
|
use_linear=False,
|
|
only_self_att=True,
|
|
causal_attention=False,
|
|
causal_block_size=1,
|
|
relative_position=False,
|
|
temporal_length=None):
|
|
super().__init__()
|
|
self.only_self_att = only_self_att
|
|
self.relative_position = relative_position
|
|
self.causal_attention = causal_attention
|
|
self.causal_block_size = causal_block_size
|
|
|
|
self.in_channels = in_channels
|
|
inner_dim = n_heads * d_head
|
|
self.norm = torch.nn.GroupNorm(num_groups=32,
|
|
num_channels=in_channels,
|
|
eps=1e-6,
|
|
affine=True)
|
|
self.proj_in = nn.Conv1d(in_channels,
|
|
inner_dim,
|
|
kernel_size=1,
|
|
stride=1,
|
|
padding=0)
|
|
if not use_linear:
|
|
self.proj_in = nn.Conv1d(in_channels,
|
|
inner_dim,
|
|
kernel_size=1,
|
|
stride=1,
|
|
padding=0)
|
|
else:
|
|
self.proj_in = nn.Linear(in_channels, inner_dim)
|
|
|
|
if relative_position:
|
|
assert (temporal_length is not None)
|
|
attention_cls = partial(CrossAttention,
|
|
relative_position=True,
|
|
temporal_length=temporal_length)
|
|
else:
|
|
attention_cls = partial(CrossAttention,
|
|
temporal_length=temporal_length)
|
|
if self.causal_attention:
|
|
assert (temporal_length is not None)
|
|
self.mask = torch.tril(
|
|
torch.ones([1, temporal_length, temporal_length]))
|
|
|
|
if self.only_self_att:
|
|
context_dim = None
|
|
self.transformer_blocks = nn.ModuleList([
|
|
BasicTransformerBlock(inner_dim,
|
|
n_heads,
|
|
d_head,
|
|
dropout=dropout,
|
|
context_dim=context_dim,
|
|
attention_cls=attention_cls,
|
|
checkpoint=use_checkpoint)
|
|
for d in range(depth)
|
|
])
|
|
if not use_linear:
|
|
self.proj_out = zero_module(
|
|
nn.Conv1d(inner_dim,
|
|
in_channels,
|
|
kernel_size=1,
|
|
stride=1,
|
|
padding=0))
|
|
else:
|
|
self.proj_out = zero_module(nn.Linear(inner_dim, in_channels))
|
|
self.use_linear = use_linear
|
|
|
|
def forward(self, x, context=None):
|
|
b, c, t, h, w = x.shape
|
|
x_in = x
|
|
x = self.norm(x)
|
|
x = rearrange(x, 'b c t h w -> (b h w) c t').contiguous()
|
|
if not self.use_linear:
|
|
x = self.proj_in(x)
|
|
x = rearrange(x, 'bhw c t -> bhw t c').contiguous()
|
|
if self.use_linear:
|
|
x = self.proj_in(x)
|
|
|
|
temp_mask = None
|
|
if self.causal_attention:
|
|
# Slice the from mask map
|
|
temp_mask = self.mask[:, :t, :t].to(x.device)
|
|
|
|
if temp_mask is not None:
|
|
mask = temp_mask.to(x.device)
|
|
mask = repeat(mask, 'l i j -> (l bhw) i j', bhw=b * h * w)
|
|
else:
|
|
mask = None
|
|
|
|
if self.only_self_att:
|
|
# NOTE: if no context is given, cross-attention defaults to self-attention
|
|
for i, block in enumerate(self.transformer_blocks):
|
|
x = block(x, mask=mask)
|
|
x = rearrange(x, '(b hw) t c -> b hw t c', b=b).contiguous()
|
|
else:
|
|
x = rearrange(x, '(b hw) t c -> b hw t c', b=b).contiguous()
|
|
context = rearrange(context, '(b t) l con -> b t l con',
|
|
t=t).contiguous()
|
|
for i, block in enumerate(self.transformer_blocks):
|
|
# Calculate each batch one by one (since number in shape could not greater then 65,535 for some package)
|
|
for j in range(b):
|
|
context_j = repeat(context[j],
|
|
't l con -> (t r) l con',
|
|
r=(h * w) // t,
|
|
t=t).contiguous()
|
|
# Note: causal mask will not applied in cross-attention case
|
|
x[j] = block(x[j], context=context_j)
|
|
|
|
if self.use_linear:
|
|
x = self.proj_out(x)
|
|
x = rearrange(x, 'b (h w) t c -> b c t h w', h=h, w=w).contiguous()
|
|
if not self.use_linear:
|
|
x = rearrange(x, 'b hw t c -> (b hw) c t').contiguous()
|
|
x = self.proj_out(x)
|
|
x = rearrange(x, '(b h w) c t -> b c t h w', b=b, h=h,
|
|
w=w).contiguous()
|
|
|
|
return x + x_in
|
|
|
|
|
|
class GEGLU(nn.Module):
|
|
|
|
def __init__(self, dim_in, dim_out):
|
|
super().__init__()
|
|
self.proj = nn.Linear(dim_in, dim_out * 2)
|
|
|
|
def forward(self, x):
|
|
x, gate = self.proj(x).chunk(2, dim=-1)
|
|
return x * F.gelu(gate)
|
|
|
|
|
|
class FeedForward(nn.Module):
|
|
|
|
def __init__(self, dim, dim_out=None, mult=4, glu=False, dropout=0.):
|
|
super().__init__()
|
|
inner_dim = int(dim * mult)
|
|
dim_out = default(dim_out, dim)
|
|
project_in = nn.Sequential(nn.Linear(
|
|
dim, inner_dim), nn.GELU()) if not glu else GEGLU(dim, inner_dim)
|
|
|
|
self.net = nn.Sequential(project_in, nn.Dropout(dropout),
|
|
nn.Linear(inner_dim, dim_out))
|
|
|
|
def forward(self, x):
|
|
return self.net(x)
|
|
|
|
|
|
class LinearAttention(nn.Module):
|
|
|
|
def __init__(self, dim, heads=4, dim_head=32):
|
|
super().__init__()
|
|
self.heads = heads
|
|
hidden_dim = dim_head * heads
|
|
self.to_qkv = nn.Conv2d(dim, hidden_dim * 3, 1, bias=False)
|
|
self.to_out = nn.Conv2d(hidden_dim, dim, 1)
|
|
|
|
def forward(self, x):
|
|
b, c, h, w = x.shape
|
|
qkv = self.to_qkv(x)
|
|
q, k, v = rearrange(qkv,
|
|
'b (qkv heads c) h w -> qkv b heads c (h w)',
|
|
heads=self.heads,
|
|
qkv=3)
|
|
k = k.softmax(dim=-1)
|
|
context = torch.einsum('bhdn,bhen->bhde', k, v)
|
|
out = torch.einsum('bhde,bhdn->bhen', context, q)
|
|
out = rearrange(out,
|
|
'b heads c (h w) -> b (heads c) h w',
|
|
heads=self.heads,
|
|
h=h,
|
|
w=w)
|
|
return self.to_out(out)
|
|
|
|
|
|
class SpatialSelfAttention(nn.Module):
|
|
|
|
def __init__(self, in_channels):
|
|
super().__init__()
|
|
self.in_channels = in_channels
|
|
|
|
self.norm = torch.nn.GroupNorm(num_groups=32,
|
|
num_channels=in_channels,
|
|
eps=1e-6,
|
|
affine=True)
|
|
self.q = torch.nn.Conv2d(in_channels,
|
|
in_channels,
|
|
kernel_size=1,
|
|
stride=1,
|
|
padding=0)
|
|
self.k = torch.nn.Conv2d(in_channels,
|
|
in_channels,
|
|
kernel_size=1,
|
|
stride=1,
|
|
padding=0)
|
|
self.v = torch.nn.Conv2d(in_channels,
|
|
in_channels,
|
|
kernel_size=1,
|
|
stride=1,
|
|
padding=0)
|
|
self.proj_out = torch.nn.Conv2d(in_channels,
|
|
in_channels,
|
|
kernel_size=1,
|
|
stride=1,
|
|
padding=0)
|
|
|
|
def forward(self, x):
|
|
h_ = x
|
|
h_ = self.norm(h_)
|
|
q = self.q(h_)
|
|
k = self.k(h_)
|
|
v = self.v(h_)
|
|
|
|
# Compute attention
|
|
b, c, h, w = q.shape
|
|
q = rearrange(q, 'b c h w -> b (h w) c')
|
|
k = rearrange(k, 'b c h w -> b c (h w)')
|
|
w_ = torch.einsum('bij,bjk->bik', q, k)
|
|
|
|
w_ = w_ * (int(c)**(-0.5))
|
|
w_ = torch.nn.functional.softmax(w_, dim=2)
|
|
|
|
# Attend to values
|
|
v = rearrange(v, 'b c h w -> b c (h w)')
|
|
w_ = rearrange(w_, 'b i j -> b j i')
|
|
h_ = torch.einsum('bij,bjk->bik', v, w_)
|
|
h_ = rearrange(h_, 'b c (h w) -> b c h w', h=h)
|
|
h_ = self.proj_out(h_)
|
|
|
|
return x + h_
|