|
import math |
|
import torch |
|
import torch.nn as nn |
|
import torch.nn.functional as F |
|
from networks.layers.basic import DropOutLogit, ScaleOffset, DWConv2d |
|
|
|
|
|
def multiply_by_ychunks(x, y, chunks=1): |
|
if chunks <= 1: |
|
return x @ y |
|
else: |
|
return torch.cat([x @ _y for _y in y.chunk(chunks, dim=-1)], dim=-1) |
|
|
|
|
|
def multiply_by_xchunks(x, y, chunks=1): |
|
if chunks <= 1: |
|
return x @ y |
|
else: |
|
return torch.cat([_x @ y for _x in x.chunk(chunks, dim=-2)], dim=-2) |
|
|
|
|
|
|
|
class MultiheadAttention(nn.Module): |
|
def __init__(self, |
|
d_model, |
|
num_head=8, |
|
dropout=0., |
|
use_linear=True, |
|
d_att=None, |
|
use_dis=False, |
|
qk_chunks=1, |
|
max_mem_len_ratio=-1, |
|
top_k=-1): |
|
super().__init__() |
|
self.d_model = d_model |
|
self.num_head = num_head |
|
self.use_dis = use_dis |
|
self.qk_chunks = qk_chunks |
|
self.max_mem_len_ratio = float(max_mem_len_ratio) |
|
self.top_k = top_k |
|
|
|
self.hidden_dim = d_model // num_head |
|
self.d_att = self.hidden_dim if d_att is None else d_att |
|
self.T = self.d_att**0.5 |
|
self.use_linear = use_linear |
|
|
|
if use_linear: |
|
self.linear_Q = nn.Linear(d_model, d_model) |
|
self.linear_K = nn.Linear(d_model, d_model) |
|
self.linear_V = nn.Linear(d_model, d_model) |
|
|
|
self.dropout = nn.Dropout(dropout) |
|
self.drop_prob = dropout |
|
self.projection = nn.Linear(d_model, d_model) |
|
self._init_weight() |
|
|
|
def forward(self, Q, K, V): |
|
""" |
|
:param Q: A 3d tensor with shape of [T_q, bs, C_q] |
|
:param K: A 3d tensor with shape of [T_k, bs, C_k] |
|
:param V: A 3d tensor with shape of [T_v, bs, C_v] |
|
""" |
|
num_head = self.num_head |
|
hidden_dim = self.hidden_dim |
|
|
|
bs = Q.size()[1] |
|
|
|
|
|
if self.use_linear: |
|
Q = self.linear_Q(Q) |
|
K = self.linear_K(K) |
|
V = self.linear_V(V) |
|
|
|
|
|
Q = Q / self.T |
|
|
|
if not self.training and self.max_mem_len_ratio > 0: |
|
mem_len_ratio = float(K.size(0)) / Q.size(0) |
|
if mem_len_ratio > self.max_mem_len_ratio: |
|
scaling_ratio = math.log(mem_len_ratio) / math.log( |
|
self.max_mem_len_ratio) |
|
Q = Q * scaling_ratio |
|
|
|
|
|
Q = Q.view(-1, bs, num_head, self.d_att).permute(1, 2, 0, 3) |
|
K = K.view(-1, bs, num_head, self.d_att).permute(1, 2, 3, 0) |
|
V = V.view(-1, bs, num_head, hidden_dim).permute(1, 2, 0, 3) |
|
|
|
|
|
QK = multiply_by_ychunks(Q, K, self.qk_chunks) |
|
if self.use_dis: |
|
QK = 2 * QK - K.pow(2).sum(dim=-2, keepdim=True) |
|
|
|
|
|
if not self.training and self.top_k > 0 and self.top_k < QK.size()[-1]: |
|
top_QK, indices = torch.topk(QK, k=self.top_k, dim=-1) |
|
top_attn = torch.softmax(top_QK, dim=-1) |
|
attn = torch.zeros_like(QK).scatter_(-1, indices, top_attn) |
|
else: |
|
attn = torch.softmax(QK, dim=-1) |
|
|
|
|
|
attn = self.dropout(attn) |
|
|
|
|
|
outputs = multiply_by_xchunks(attn, V, |
|
self.qk_chunks).permute(2, 0, 1, 3) |
|
|
|
|
|
outputs = outputs.reshape(-1, bs, self.d_model) |
|
|
|
outputs = self.projection(outputs) |
|
|
|
return outputs, attn |
|
|
|
def _init_weight(self): |
|
for p in self.parameters(): |
|
if p.dim() > 1: |
|
nn.init.xavier_uniform_(p) |
|
|
|
|
|
|
|
class MultiheadLocalAttentionV1(nn.Module): |
|
def __init__(self, |
|
d_model, |
|
num_head, |
|
dropout=0., |
|
max_dis=7, |
|
dilation=1, |
|
use_linear=True, |
|
enable_corr=True): |
|
super().__init__() |
|
self.dilation = dilation |
|
self.window_size = 2 * max_dis + 1 |
|
self.max_dis = max_dis |
|
self.num_head = num_head |
|
self.T = ((d_model / num_head)**0.5) |
|
|
|
self.use_linear = use_linear |
|
if use_linear: |
|
self.linear_Q = nn.Conv2d(d_model, d_model, kernel_size=1) |
|
self.linear_K = nn.Conv2d(d_model, d_model, kernel_size=1) |
|
self.linear_V = nn.Conv2d(d_model, d_model, kernel_size=1) |
|
|
|
self.relative_emb_k = nn.Conv2d(d_model, |
|
num_head * self.window_size * |
|
self.window_size, |
|
kernel_size=1, |
|
groups=num_head) |
|
self.relative_emb_v = nn.Parameter( |
|
torch.zeros([ |
|
self.num_head, d_model // self.num_head, |
|
self.window_size * self.window_size |
|
])) |
|
|
|
self.enable_corr = enable_corr |
|
|
|
if enable_corr: |
|
from spatial_correlation_sampler import SpatialCorrelationSampler |
|
self.correlation_sampler = SpatialCorrelationSampler( |
|
kernel_size=1, |
|
patch_size=self.window_size, |
|
stride=1, |
|
padding=0, |
|
dilation=1, |
|
dilation_patch=self.dilation) |
|
|
|
self.projection = nn.Linear(d_model, d_model) |
|
|
|
self.dropout = nn.Dropout(dropout) |
|
self.drop_prob = dropout |
|
|
|
def forward(self, q, k, v): |
|
n, c, h, w = v.size() |
|
|
|
if self.use_linear: |
|
q = self.linear_Q(q) |
|
k = self.linear_K(k) |
|
v = self.linear_V(v) |
|
|
|
hidden_dim = c // self.num_head |
|
|
|
relative_emb = self.relative_emb_k(q) |
|
memory_mask = torch.ones((1, 1, h, w), device=v.device).float() |
|
|
|
|
|
q = q / self.T |
|
|
|
q = q.view(-1, hidden_dim, h, w) |
|
k = k.reshape(-1, hidden_dim, h, w).contiguous() |
|
unfolded_vu = self.pad_and_unfold(v).view( |
|
n, self.num_head, hidden_dim, self.window_size * self.window_size, |
|
h * w) + self.relative_emb_v.unsqueeze(0).unsqueeze(-1) |
|
|
|
relative_emb = relative_emb.view(n, self.num_head, |
|
self.window_size * self.window_size, |
|
h * w) |
|
unfolded_k_mask = self.pad_and_unfold(memory_mask).bool().view( |
|
1, 1, self.window_size * self.window_size, |
|
h * w).expand(n, self.num_head, -1, -1) |
|
|
|
if self.enable_corr: |
|
qk = self.correlation_sampler(q, k).view( |
|
n, self.num_head, self.window_size * self.window_size, |
|
h * w) + relative_emb |
|
else: |
|
unfolded_k = self.pad_and_unfold(k).view( |
|
n * self.num_head, hidden_dim, |
|
self.window_size * self.window_size, h, w) |
|
qk = (q.unsqueeze(2) * unfolded_k).sum(dim=1).view( |
|
n, self.num_head, self.window_size * self.window_size, |
|
h * w) + relative_emb |
|
|
|
qk_mask = 1 - unfolded_k_mask |
|
|
|
qk -= qk_mask * 1e+8 if qk.dtype == torch.float32 else qk_mask * 1e+4 |
|
|
|
local_attn = torch.softmax(qk, dim=2) |
|
|
|
local_attn = self.dropout(local_attn) |
|
|
|
output = (local_attn.unsqueeze(2) * unfolded_vu).sum(dim=3).permute( |
|
3, 0, 1, 2).view(h * w, n, c) |
|
|
|
output = self.projection(output) |
|
|
|
return output, local_attn |
|
|
|
def pad_and_unfold(self, x): |
|
pad_pixel = self.max_dis * self.dilation |
|
x = F.pad(x, (pad_pixel, pad_pixel, pad_pixel, pad_pixel), |
|
mode='constant', |
|
value=0) |
|
x = F.unfold(x, |
|
kernel_size=(self.window_size, self.window_size), |
|
stride=(1, 1), |
|
dilation=self.dilation) |
|
return x |
|
|
|
|
|
class MultiheadLocalAttentionV2(nn.Module): |
|
def __init__(self, |
|
d_model, |
|
num_head, |
|
dropout=0., |
|
max_dis=7, |
|
dilation=1, |
|
use_linear=True, |
|
enable_corr=True, |
|
d_att=None, |
|
use_dis=False): |
|
super().__init__() |
|
self.dilation = dilation |
|
self.window_size = 2 * max_dis + 1 |
|
self.max_dis = max_dis |
|
self.num_head = num_head |
|
self.hidden_dim = d_model // num_head |
|
self.d_att = self.hidden_dim if d_att is None else d_att |
|
self.T = self.d_att**0.5 |
|
self.use_dis = use_dis |
|
|
|
self.use_linear = use_linear |
|
if use_linear: |
|
self.linear_Q = nn.Conv2d(d_model, d_model, kernel_size=1) |
|
self.linear_K = nn.Conv2d(d_model, d_model, kernel_size=1) |
|
self.linear_V = nn.Conv2d(d_model, d_model, kernel_size=1) |
|
|
|
self.relative_emb_k = nn.Conv2d(self.d_att * self.num_head, |
|
num_head * self.window_size * |
|
self.window_size, |
|
kernel_size=1, |
|
groups=num_head) |
|
self.relative_emb_v = nn.Parameter( |
|
torch.zeros([ |
|
self.num_head, d_model // self.num_head, |
|
self.window_size * self.window_size |
|
])) |
|
|
|
self.enable_corr = enable_corr |
|
|
|
if enable_corr: |
|
from spatial_correlation_sampler import SpatialCorrelationSampler |
|
self.correlation_sampler = SpatialCorrelationSampler( |
|
kernel_size=1, |
|
patch_size=self.window_size, |
|
stride=1, |
|
padding=0, |
|
dilation=1, |
|
dilation_patch=self.dilation) |
|
|
|
self.projection = nn.Linear(d_model, d_model) |
|
|
|
self.dropout = nn.Dropout(dropout) |
|
|
|
self.drop_prob = dropout |
|
|
|
self.local_mask = None |
|
self.last_size_2d = None |
|
self.qk_mask = None |
|
|
|
def forward(self, q, k, v): |
|
n, c, h, w = v.size() |
|
|
|
if self.use_linear: |
|
q = self.linear_Q(q) |
|
k = self.linear_K(k) |
|
v = self.linear_V(v) |
|
|
|
hidden_dim = self.hidden_dim |
|
|
|
if self.qk_mask is not None and (h, w) == self.last_size_2d: |
|
qk_mask = self.qk_mask |
|
else: |
|
memory_mask = torch.ones((1, 1, h, w), device=v.device).float() |
|
unfolded_k_mask = self.pad_and_unfold(memory_mask).view( |
|
1, 1, self.window_size * self.window_size, h * w) |
|
qk_mask = 1 - unfolded_k_mask |
|
self.qk_mask = qk_mask |
|
|
|
relative_emb = self.relative_emb_k(q) |
|
|
|
|
|
q = q / self.T |
|
|
|
q = q.view(-1, self.d_att, h, w) |
|
k = k.view(-1, self.d_att, h, w) |
|
v = v.view(-1, self.num_head, hidden_dim, h * w) |
|
|
|
relative_emb = relative_emb.view(n, self.num_head, |
|
self.window_size * self.window_size, |
|
h * w) |
|
|
|
if self.enable_corr: |
|
qk = self.correlation_sampler(q, k).view( |
|
n, self.num_head, self.window_size * self.window_size, h * w) |
|
else: |
|
unfolded_k = self.pad_and_unfold(k).view( |
|
n * self.num_head, hidden_dim, |
|
self.window_size * self.window_size, h, w) |
|
qk = (q.unsqueeze(2) * unfolded_k).sum(dim=1).view( |
|
n, self.num_head, self.window_size * self.window_size, h * w) |
|
if self.use_dis: |
|
qk = 2 * qk - self.pad_and_unfold( |
|
k.pow(2).sum(dim=1, keepdim=True)).view( |
|
n, self.num_head, self.window_size * self.window_size, |
|
h * w) |
|
|
|
qk = qk + relative_emb |
|
|
|
qk -= qk_mask * 1e+8 if qk.dtype == torch.float32 else qk_mask * 1e+4 |
|
|
|
local_attn = torch.softmax(qk, dim=2) |
|
|
|
local_attn = self.dropout(local_attn) |
|
|
|
agg_bias = torch.einsum('bhwn,hcw->bhnc', local_attn, |
|
self.relative_emb_v) |
|
|
|
global_attn = self.local2global(local_attn, h, w) |
|
|
|
agg_value = (global_attn @ v.transpose(-2, -1)) |
|
|
|
output = (agg_value + agg_bias).permute(2, 0, 1, |
|
3).reshape(h * w, n, c) |
|
|
|
output = self.projection(output) |
|
|
|
self.last_size_2d = (h, w) |
|
return output, local_attn |
|
|
|
def local2global(self, local_attn, height, width): |
|
batch_size = local_attn.size()[0] |
|
|
|
pad_height = height + 2 * self.max_dis |
|
pad_width = width + 2 * self.max_dis |
|
|
|
if self.local_mask is not None and (height, |
|
width) == self.last_size_2d: |
|
local_mask = self.local_mask |
|
else: |
|
ky, kx = torch.meshgrid([ |
|
torch.arange(0, pad_height, device=local_attn.device), |
|
torch.arange(0, pad_width, device=local_attn.device) |
|
]) |
|
qy, qx = torch.meshgrid([ |
|
torch.arange(0, height, device=local_attn.device), |
|
torch.arange(0, width, device=local_attn.device) |
|
]) |
|
|
|
offset_y = qy.reshape(-1, 1) - ky.reshape(1, -1) + self.max_dis |
|
offset_x = qx.reshape(-1, 1) - kx.reshape(1, -1) + self.max_dis |
|
|
|
local_mask = (offset_y.abs() <= self.max_dis) & (offset_x.abs() <= |
|
self.max_dis) |
|
local_mask = local_mask.view(1, 1, height * width, pad_height, |
|
pad_width) |
|
self.local_mask = local_mask |
|
|
|
global_attn = torch.zeros( |
|
(batch_size, self.num_head, height * width, pad_height, pad_width), |
|
device=local_attn.device) |
|
global_attn[local_mask.expand(batch_size, self.num_head, |
|
-1, -1, -1)] = local_attn.transpose( |
|
-1, -2).reshape(-1) |
|
global_attn = global_attn[:, :, :, self.max_dis:-self.max_dis, |
|
self.max_dis:-self.max_dis].reshape( |
|
batch_size, self.num_head, |
|
height * width, height * width) |
|
|
|
return global_attn |
|
|
|
def pad_and_unfold(self, x): |
|
pad_pixel = self.max_dis * self.dilation |
|
x = F.pad(x, (pad_pixel, pad_pixel, pad_pixel, pad_pixel), |
|
mode='constant', |
|
value=0) |
|
x = F.unfold(x, |
|
kernel_size=(self.window_size, self.window_size), |
|
stride=(1, 1), |
|
dilation=self.dilation) |
|
return x |
|
|
|
|
|
class MultiheadLocalAttentionV3(nn.Module): |
|
def __init__(self, |
|
d_model, |
|
num_head, |
|
dropout=0., |
|
max_dis=7, |
|
dilation=1, |
|
use_linear=True): |
|
super().__init__() |
|
self.dilation = dilation |
|
self.window_size = 2 * max_dis + 1 |
|
self.max_dis = max_dis |
|
self.num_head = num_head |
|
self.T = ((d_model / num_head)**0.5) |
|
|
|
self.use_linear = use_linear |
|
if use_linear: |
|
self.linear_Q = nn.Conv2d(d_model, d_model, kernel_size=1) |
|
self.linear_K = nn.Conv2d(d_model, d_model, kernel_size=1) |
|
self.linear_V = nn.Conv2d(d_model, d_model, kernel_size=1) |
|
|
|
self.relative_emb_k = nn.Conv2d(d_model, |
|
num_head * self.window_size * |
|
self.window_size, |
|
kernel_size=1, |
|
groups=num_head) |
|
self.relative_emb_v = nn.Parameter( |
|
torch.zeros([ |
|
self.num_head, d_model // self.num_head, |
|
self.window_size * self.window_size |
|
])) |
|
|
|
self.projection = nn.Linear(d_model, d_model) |
|
self.dropout = DropOutLogit(dropout) |
|
|
|
self.padded_local_mask = None |
|
self.local_mask = None |
|
self.last_size_2d = None |
|
self.qk_mask = None |
|
|
|
def forward(self, q, k, v): |
|
n, c, h, w = q.size() |
|
|
|
if self.use_linear: |
|
q = self.linear_Q(q) |
|
k = self.linear_K(k) |
|
v = self.linear_V(v) |
|
|
|
hidden_dim = c // self.num_head |
|
|
|
relative_emb = self.relative_emb_k(q) |
|
relative_emb = relative_emb.view(n, self.num_head, |
|
self.window_size * self.window_size, |
|
h * w) |
|
padded_local_mask, local_mask = self.compute_mask(h, |
|
w, |
|
device=q.device) |
|
qk_mask = (~padded_local_mask).float() |
|
|
|
|
|
q = q / self.T |
|
|
|
q = q.view(-1, self.num_head, hidden_dim, h * w) |
|
k = k.view(-1, self.num_head, hidden_dim, h * w) |
|
v = v.view(-1, self.num_head, hidden_dim, h * w) |
|
|
|
qk = q.transpose(-1, -2) @ k |
|
|
|
pad_pixel = self.max_dis * self.dilation |
|
|
|
padded_qk = F.pad(qk.view(-1, self.num_head, h * w, h, w), |
|
(pad_pixel, pad_pixel, pad_pixel, pad_pixel), |
|
mode='constant', |
|
value=-1e+8 if qk.dtype == torch.float32 else -1e+4) |
|
|
|
qk_mask = qk_mask * 1e+8 if (padded_qk.dtype |
|
== torch.float32) else qk_mask * 1e+4 |
|
padded_qk = padded_qk - qk_mask |
|
|
|
padded_qk[padded_local_mask.expand(n, self.num_head, -1, -1, |
|
-1)] += relative_emb.transpose( |
|
-1, -2).reshape(-1) |
|
padded_qk = self.dropout(padded_qk) |
|
|
|
local_qk = padded_qk[padded_local_mask.expand(n, self.num_head, -1, -1, |
|
-1)] |
|
|
|
global_qk = padded_qk[:, :, :, self.max_dis:-self.max_dis, |
|
self.max_dis:-self.max_dis].reshape( |
|
n, self.num_head, h * w, h * w) |
|
|
|
local_attn = torch.softmax(local_qk.reshape( |
|
n, self.num_head, h * w, self.window_size * self.window_size), |
|
dim=3) |
|
global_attn = torch.softmax(global_qk, dim=3) |
|
|
|
agg_bias = torch.einsum('bhnw,hcw->nbhc', local_attn, |
|
self.relative_emb_v).reshape(h * w, n, c) |
|
|
|
agg_value = (global_attn @ v.transpose(-2, -1)) |
|
|
|
output = agg_value + agg_bias |
|
|
|
output = self.projection(output) |
|
|
|
self.last_size_2d = (h, w) |
|
return output, local_attn |
|
|
|
def compute_mask(self, height, width, device=None): |
|
pad_height = height + 2 * self.max_dis |
|
pad_width = width + 2 * self.max_dis |
|
|
|
if self.padded_local_mask is not None and (height, |
|
width) == self.last_size_2d: |
|
padded_local_mask = self.padded_local_mask |
|
local_mask = self.local_mask |
|
|
|
else: |
|
ky, kx = torch.meshgrid([ |
|
torch.arange(0, pad_height, device=device), |
|
torch.arange(0, pad_width, device=device) |
|
]) |
|
qy, qx = torch.meshgrid([ |
|
torch.arange(0, height, device=device), |
|
torch.arange(0, width, device=device) |
|
]) |
|
|
|
qy = qy.reshape(-1, 1) |
|
qx = qx.reshape(-1, 1) |
|
offset_y = qy - ky.reshape(1, -1) + self.max_dis |
|
offset_x = qx - kx.reshape(1, -1) + self.max_dis |
|
padded_local_mask = (offset_y.abs() <= self.max_dis) & ( |
|
offset_x.abs() <= self.max_dis) |
|
padded_local_mask = padded_local_mask.view(1, 1, height * width, |
|
pad_height, pad_width) |
|
local_mask = padded_local_mask[:, :, :, self.max_dis:-self.max_dis, |
|
self.max_dis:-self.max_dis] |
|
pad_pixel = self.max_dis * self.dilation |
|
local_mask = F.pad(local_mask.float(), |
|
(pad_pixel, pad_pixel, pad_pixel, pad_pixel), |
|
mode='constant', |
|
value=0).view(1, 1, height * width, pad_height, |
|
pad_width) |
|
self.padded_local_mask = padded_local_mask |
|
self.local_mask = local_mask |
|
|
|
return padded_local_mask, local_mask |
|
|
|
|
|
def linear_gate(x, dim=-1): |
|
|
|
return torch.softmax(x, dim=dim) |
|
|
|
|
|
def silu(x): |
|
return x * torch.sigmoid(x) |
|
|
|
|
|
class GatedPropagation(nn.Module): |
|
def __init__(self, |
|
d_qk, |
|
d_vu, |
|
num_head=8, |
|
dropout=0., |
|
use_linear=True, |
|
d_att=None, |
|
use_dis=False, |
|
qk_chunks=1, |
|
max_mem_len_ratio=-1, |
|
top_k=-1, |
|
expand_ratio=2.): |
|
super().__init__() |
|
expand_ratio = expand_ratio |
|
self.expand_d_vu = int(d_vu * expand_ratio) |
|
self.d_vu = d_vu |
|
self.d_qk = d_qk |
|
self.num_head = num_head |
|
self.use_dis = use_dis |
|
self.qk_chunks = qk_chunks |
|
self.max_mem_len_ratio = float(max_mem_len_ratio) |
|
self.top_k = top_k |
|
|
|
self.hidden_dim = self.expand_d_vu // num_head |
|
self.d_att = d_qk // num_head if d_att is None else d_att |
|
self.T = self.d_att**0.5 |
|
self.use_linear = use_linear |
|
self.d_middle = self.d_att * self.num_head |
|
|
|
if use_linear: |
|
self.linear_QK = nn.Linear(d_qk, self.d_middle) |
|
half_d_vu = self.hidden_dim * num_head // 2 |
|
self.linear_V1 = nn.Linear(d_vu // 2, half_d_vu) |
|
self.linear_V2 = nn.Linear(d_vu // 2, half_d_vu) |
|
self.linear_U1 = nn.Linear(d_vu // 2, half_d_vu) |
|
self.linear_U2 = nn.Linear(d_vu // 2, half_d_vu) |
|
|
|
self.dropout = nn.Dropout(dropout) |
|
self.drop_prob = dropout |
|
|
|
self.dw_conv = DWConv2d(self.expand_d_vu) |
|
self.projection = nn.Linear(self.expand_d_vu, d_vu) |
|
|
|
self._init_weight() |
|
|
|
def forward(self, Q, K, V, U, size_2d): |
|
""" |
|
:param Q: A 3d tensor with shape of [T_q, bs, C_q] |
|
:param K: A 3d tensor with shape of [T_k, bs, C_k] |
|
:param V: A 3d tensor with shape of [T_v, bs, C_v] |
|
""" |
|
num_head = self.num_head |
|
hidden_dim = self.hidden_dim |
|
|
|
l, bs, _ = Q.size() |
|
|
|
|
|
if self.use_linear: |
|
Q = K = self.linear_QK(Q) |
|
|
|
def cat(X1, X2): |
|
if num_head > 1: |
|
X1 = X1.view(-1, bs, num_head, hidden_dim // 2) |
|
X2 = X2.view(-1, bs, num_head, hidden_dim // 2) |
|
X = torch.cat([X1, X2], |
|
dim=-1).view(-1, bs, num_head * hidden_dim) |
|
else: |
|
X = torch.cat([X1, X2], dim=-1) |
|
return X |
|
|
|
V1, V2 = torch.split(V, self.d_vu // 2, dim=-1) |
|
V1 = self.linear_V1(V1) |
|
V2 = self.linear_V2(V2) |
|
V = silu(cat(V1, V2)) |
|
|
|
U1, U2 = torch.split(U, self.d_vu // 2, dim=-1) |
|
U1 = self.linear_U1(U1) |
|
U2 = self.linear_U2(U2) |
|
U = silu(cat(U1, U2)) |
|
|
|
|
|
Q = Q / self.T |
|
|
|
if not self.training and self.max_mem_len_ratio > 0: |
|
mem_len_ratio = float(K.size(0)) / Q.size(0) |
|
if mem_len_ratio > self.max_mem_len_ratio: |
|
scaling_ratio = math.log(mem_len_ratio) / math.log( |
|
self.max_mem_len_ratio) |
|
Q = Q * scaling_ratio |
|
|
|
|
|
Q = Q.view(-1, bs, num_head, self.d_att).permute(1, 2, 0, 3) |
|
K = K.view(-1, bs, num_head, self.d_att).permute(1, 2, 3, 0) |
|
V = V.view(-1, bs, num_head, hidden_dim).permute(1, 2, 0, 3) |
|
|
|
|
|
QK = multiply_by_ychunks(Q, K, self.qk_chunks) |
|
if self.use_dis: |
|
QK = 2 * QK - K.pow(2).sum(dim=-2, keepdim=True) |
|
|
|
|
|
if not self.training and self.top_k > 0 and self.top_k < QK.size()[-1]: |
|
top_QK, indices = torch.topk(QK, k=self.top_k, dim=-1) |
|
top_attn = linear_gate(top_QK, dim=-1) |
|
attn = torch.zeros_like(QK).scatter_(-1, indices, top_attn) |
|
else: |
|
attn = linear_gate(QK, dim=-1) |
|
|
|
|
|
attn = self.dropout(attn) |
|
|
|
|
|
outputs = multiply_by_xchunks(attn, V, |
|
self.qk_chunks).permute(2, 0, 1, 3) |
|
|
|
|
|
outputs = outputs.reshape(l, bs, -1) * U |
|
|
|
outputs = self.dw_conv(outputs, size_2d) |
|
outputs = self.projection(outputs) |
|
|
|
return outputs, attn |
|
|
|
def _init_weight(self): |
|
for p in self.parameters(): |
|
if p.dim() > 1: |
|
nn.init.xavier_uniform_(p) |
|
|
|
|
|
class LocalGatedPropagation(nn.Module): |
|
def __init__(self, |
|
d_qk, |
|
d_vu, |
|
num_head, |
|
dropout=0., |
|
max_dis=7, |
|
dilation=1, |
|
use_linear=True, |
|
enable_corr=True, |
|
d_att=None, |
|
use_dis=False, |
|
expand_ratio=2.): |
|
super().__init__() |
|
expand_ratio = expand_ratio |
|
self.expand_d_vu = int(d_vu * expand_ratio) |
|
self.d_qk = d_qk |
|
self.d_vu = d_vu |
|
self.dilation = dilation |
|
self.window_size = 2 * max_dis + 1 |
|
self.max_dis = max_dis |
|
self.num_head = num_head |
|
self.hidden_dim = self.expand_d_vu // num_head |
|
self.d_att = d_qk // num_head if d_att is None else d_att |
|
self.T = self.d_att**0.5 |
|
self.use_dis = use_dis |
|
|
|
self.d_middle = self.d_att * self.num_head |
|
self.use_linear = use_linear |
|
if use_linear: |
|
self.linear_QK = nn.Conv2d(d_qk, self.d_middle, kernel_size=1) |
|
self.linear_V = nn.Conv2d(d_vu, |
|
self.expand_d_vu, |
|
kernel_size=1, |
|
groups=2) |
|
self.linear_U = nn.Conv2d(d_vu, |
|
self.expand_d_vu, |
|
kernel_size=1, |
|
groups=2) |
|
|
|
self.relative_emb_k = nn.Conv2d(self.d_middle, |
|
num_head * self.window_size * |
|
self.window_size, |
|
kernel_size=1, |
|
groups=num_head) |
|
|
|
self.enable_corr = enable_corr |
|
|
|
if enable_corr: |
|
from spatial_correlation_sampler import SpatialCorrelationSampler |
|
self.correlation_sampler = SpatialCorrelationSampler( |
|
kernel_size=1, |
|
patch_size=self.window_size, |
|
stride=1, |
|
padding=0, |
|
dilation=1, |
|
dilation_patch=self.dilation) |
|
|
|
self.dw_conv = DWConv2d(self.expand_d_vu) |
|
self.projection = nn.Linear(self.expand_d_vu, d_vu) |
|
|
|
self.dropout = nn.Dropout(dropout) |
|
|
|
self.drop_prob = dropout |
|
|
|
self.local_mask = None |
|
self.last_size_2d = None |
|
self.qk_mask = None |
|
|
|
def forward(self, q, k, v, u, size_2d): |
|
n, c, h, w = v.size() |
|
hidden_dim = self.hidden_dim |
|
|
|
if self.use_linear: |
|
q = k = self.linear_QK(q) |
|
v = silu(self.linear_V(v)) |
|
u = silu(self.linear_U(u)) |
|
if self.num_head > 1: |
|
v = v.view(-1, 2, self.num_head, hidden_dim // 2, |
|
h * w).permute(0, 2, 1, 3, 4).reshape(n, -1, h, w) |
|
u = u.view(-1, 2, self.num_head, hidden_dim // 2, |
|
h * w).permute(4, 0, 2, 1, 3).reshape(h * w, n, -1) |
|
else: |
|
u = u.permute(2, 3, 0, 1).reshape(h * w, n, -1) |
|
|
|
if self.qk_mask is not None and (h, w) == self.last_size_2d: |
|
qk_mask = self.qk_mask |
|
else: |
|
memory_mask = torch.ones((1, 1, h, w), device=v.device).float() |
|
unfolded_k_mask = self.pad_and_unfold(memory_mask).view( |
|
1, 1, self.window_size * self.window_size, h * w) |
|
qk_mask = 1 - unfolded_k_mask |
|
self.qk_mask = qk_mask |
|
|
|
relative_emb = self.relative_emb_k(q) |
|
|
|
|
|
q = q / self.T |
|
|
|
q = q.view(-1, self.d_att, h, w) |
|
k = k.view(-1, self.d_att, h, w) |
|
v = v.view(-1, self.num_head, hidden_dim, h * w) |
|
|
|
relative_emb = relative_emb.view(n, self.num_head, |
|
self.window_size * self.window_size, |
|
h * w) |
|
|
|
if self.enable_corr: |
|
qk = self.correlation_sampler(q, k).view( |
|
n, self.num_head, self.window_size * self.window_size, h * w) |
|
else: |
|
unfolded_k = self.pad_and_unfold(k).view( |
|
n * self.num_head, self.d_att, |
|
self.window_size * self.window_size, h, w) |
|
qk = (q.unsqueeze(2) * unfolded_k).sum(dim=1).view( |
|
n, self.num_head, self.window_size * self.window_size, h * w) |
|
if self.use_dis: |
|
qk = 2 * qk - self.pad_and_unfold( |
|
k.pow(2).sum(dim=1, keepdim=True)).view( |
|
n, self.num_head, self.window_size * self.window_size, |
|
h * w) |
|
|
|
qk = qk + relative_emb |
|
|
|
qk -= qk_mask * 1e+8 if qk.dtype == torch.float32 else qk_mask * 1e+4 |
|
|
|
local_attn = linear_gate(qk, dim=2) |
|
|
|
local_attn = self.dropout(local_attn) |
|
|
|
global_attn = self.local2global(local_attn, h, w) |
|
|
|
agg_value = (global_attn @ v.transpose(-2, -1)).permute( |
|
2, 0, 1, 3).reshape(h * w, n, -1) |
|
|
|
output = agg_value * u |
|
|
|
output = self.dw_conv(output, size_2d) |
|
output = self.projection(output) |
|
|
|
self.last_size_2d = (h, w) |
|
return output, local_attn |
|
|
|
def local2global(self, local_attn, height, width): |
|
batch_size = local_attn.size()[0] |
|
|
|
pad_height = height + 2 * self.max_dis |
|
pad_width = width + 2 * self.max_dis |
|
|
|
if self.local_mask is not None and (height, |
|
width) == self.last_size_2d: |
|
local_mask = self.local_mask |
|
else: |
|
ky, kx = torch.meshgrid([ |
|
torch.arange(0, pad_height, device=local_attn.device), |
|
torch.arange(0, pad_width, device=local_attn.device) |
|
]) |
|
qy, qx = torch.meshgrid([ |
|
torch.arange(0, height, device=local_attn.device), |
|
torch.arange(0, width, device=local_attn.device) |
|
]) |
|
|
|
offset_y = qy.reshape(-1, 1) - ky.reshape(1, -1) + self.max_dis |
|
offset_x = qx.reshape(-1, 1) - kx.reshape(1, -1) + self.max_dis |
|
|
|
local_mask = (offset_y.abs() <= self.max_dis) & (offset_x.abs() <= |
|
self.max_dis) |
|
local_mask = local_mask.view(1, 1, height * width, pad_height, |
|
pad_width) |
|
self.local_mask = local_mask |
|
|
|
global_attn = torch.zeros( |
|
(batch_size, self.num_head, height * width, pad_height, pad_width), |
|
device=local_attn.device) |
|
global_attn[local_mask.expand(batch_size, self.num_head, |
|
-1, -1, -1)] = local_attn.transpose( |
|
-1, -2).reshape(-1) |
|
global_attn = global_attn[:, :, :, self.max_dis:-self.max_dis, |
|
self.max_dis:-self.max_dis].reshape( |
|
batch_size, self.num_head, |
|
height * width, height * width) |
|
|
|
return global_attn |
|
|
|
def pad_and_unfold(self, x): |
|
pad_pixel = self.max_dis * self.dilation |
|
x = F.pad(x, (pad_pixel, pad_pixel, pad_pixel, pad_pixel), |
|
mode='constant', |
|
value=0) |
|
x = F.unfold(x, |
|
kernel_size=(self.window_size, self.window_size), |
|
stride=(1, 1), |
|
dilation=self.dilation) |
|
return x |
|
|