# import packages
import copy
# import packages
import math
import torch
import torch.nn.functional as F
from torch import nn
from torch.autograd import Variable
class SingleAttention(nn.Module):
def __init__(
self,
attention_input_dim,
attention_hidden_dim,
attention_type="add",
demographic_dim=12,
time_aware=False,
use_demographic=False,
):
super(SingleAttention, self).__init__()
self.attention_type = attention_type
self.attention_hidden_dim = attention_hidden_dim
self.attention_input_dim = attention_input_dim
self.use_demographic = use_demographic
self.demographic_dim = demographic_dim
self.time_aware = time_aware
# batch_time = torch.arange(0, batch_mask.size()[1], dtype=torch.float32).reshape(1, batch_mask.size()[1], 1)
# batch_time = batch_time.repeat(batch_mask.size()[0], 1, 1)
if attention_type == "add":
if self.time_aware:
# self.Wx = nn.Parameter(torch.randn(attention_input_dim+1, attention_hidden_dim))
self.Wx = nn.Parameter(
torch.randn(attention_input_dim, attention_hidden_dim)
)
self.Wtime_aware = nn.Parameter(torch.randn(1, attention_hidden_dim))
nn.init.kaiming_uniform_(self.Wtime_aware, a=math.sqrt(5))
else:
self.Wx = nn.Parameter(
torch.randn(attention_input_dim, attention_hidden_dim)
)
self.Wt = nn.Parameter(
torch.randn(attention_input_dim, attention_hidden_dim)
)
self.Wd = nn.Parameter(torch.randn(demographic_dim, attention_hidden_dim))
self.bh = nn.Parameter(
torch.zeros(
attention_hidden_dim,
)
)
self.Wa = nn.Parameter(torch.randn(attention_hidden_dim, 1))
self.ba = nn.Parameter(
torch.zeros(
1,
)
)
nn.init.kaiming_uniform_(self.Wd, a=math.sqrt(5))
nn.init.kaiming_uniform_(self.Wx, a=math.sqrt(5))
nn.init.kaiming_uniform_(self.Wt, a=math.sqrt(5))
nn.init.kaiming_uniform_(self.Wa, a=math.sqrt(5))
elif attention_type == "mul":
self.Wa = nn.Parameter(
torch.randn(attention_input_dim, attention_input_dim)
)
self.ba = nn.Parameter(
torch.zeros(
1,
)
)
nn.init.kaiming_uniform_(self.Wa, a=math.sqrt(5))
elif attention_type == "concat":
if self.time_aware:
self.Wh = nn.Parameter(
torch.randn(2 * attention_input_dim + 1, attention_hidden_dim)
)
else:
self.Wh = nn.Parameter(
torch.randn(2 * attention_input_dim, attention_hidden_dim)
)
self.Wa = nn.Parameter(torch.randn(attention_hidden_dim, 1))
self.ba = nn.Parameter(
torch.zeros(
1,
)
)
nn.init.kaiming_uniform_(self.Wh, a=math.sqrt(5))
nn.init.kaiming_uniform_(self.Wa, a=math.sqrt(5))
elif attention_type == "new":
self.Wt = nn.Parameter(
torch.randn(attention_input_dim, attention_hidden_dim)
)
self.Wx = nn.Parameter(
torch.randn(attention_input_dim, attention_hidden_dim)
)
self.rate = nn.Parameter(torch.zeros(1) + 0.8)
nn.init.kaiming_uniform_(self.Wx, a=math.sqrt(5))
nn.init.kaiming_uniform_(self.Wt, a=math.sqrt(5))
else:
raise RuntimeError("Wrong attention type.")
self.tanh = nn.Tanh()
self.softmax = nn.Softmax(dim=1)
self.sigmoid = nn.Sigmoid()
self.relu = nn.ReLU()
def forward(self, input, device, demo=None):
(
batch_size,
time_step,
input_dim,
) = input.size() # batch_size * time_step * hidden_dim(i)
time_decays = (
torch.tensor(range(time_step - 1, -1, -1), dtype=torch.float32)
.unsqueeze(-1).unsqueeze(0).to(device=device)
) # 1*t*1
b_time_decays = time_decays.repeat(batch_size, 1, 1) + 1 # b t 1
if self.attention_type == "add": # B*T*I @ H*I
q = torch.matmul(input[:, -1, :], self.Wt) # b h
q = torch.reshape(q, (batch_size, 1, self.attention_hidden_dim)) # B*1*H
if self.time_aware == True:
k = torch.matmul(input, self.Wx) # b t h
time_hidden = torch.matmul(b_time_decays, self.Wtime_aware) # b t h
else:
k = torch.matmul(input, self.Wx) # b t h
if self.use_demographic:
d = torch.matmul(demo, self.Wd) # B*H
d = torch.reshape(
d, (batch_size, 1, self.attention_hidden_dim)
) # b 1 h
h = q + k + self.bh # b t h
if self.time_aware:
h += time_hidden
h = self.tanh(h) # B*T*H
e = torch.matmul(h, self.Wa) + self.ba # B*T*1
e = torch.reshape(e, (batch_size, time_step)) # b t
elif self.attention_type == "mul":
e = torch.matmul(input[:, -1, :], self.Wa) # b i
e = (
torch.matmul(e.unsqueeze(1), input.permute(0, 2, 1)).squeeze() + self.ba
) # b t
elif self.attention_type == "concat":
q = input[:, -1, :].unsqueeze(1).repeat(1, time_step, 1) # b t i
k = input
c = torch.cat((q, k), dim=-1) # B*T*2I
if self.time_aware:
c = torch.cat((c, b_time_decays), dim=-1) # B*T*2I+1
h = torch.matmul(c, self.Wh)
h = self.tanh(h)
e = torch.matmul(h, self.Wa) + self.ba # B*T*1
e = torch.reshape(e, (batch_size, time_step)) # b t
elif self.attention_type == "new":
q = torch.matmul(input[:, -1, :], self.Wt) # b h
q = torch.reshape(q, (batch_size, 1, self.attention_hidden_dim)) # B*1*H
k = torch.matmul(input, self.Wx) # b t h
dot_product = torch.matmul(q, k.transpose(1, 2)).squeeze() # b t
denominator = self.sigmoid(self.rate) * (
torch.log(2.72 + (1 - self.sigmoid(dot_product)))
* (b_time_decays.squeeze())
)
e = self.relu(self.sigmoid(dot_product) / (denominator)) # b * t
a = self.softmax(e) # B*T
v = torch.matmul(a.unsqueeze(1), input).squeeze() # B*I
return v, a
class FinalAttentionQKV(nn.Module):
def __init__(
self,
attention_input_dim,
attention_hidden_dim,
attention_type="add",
dropout=None,
):
super(FinalAttentionQKV, self).__init__()
self.attention_type = attention_type
self.attention_hidden_dim = attention_hidden_dim
self.attention_input_dim = attention_input_dim
self.W_q = nn.Linear(attention_input_dim, attention_hidden_dim)
self.W_k = nn.Linear(attention_input_dim, attention_hidden_dim)
self.W_v = nn.Linear(attention_input_dim, attention_hidden_dim)
self.W_out = nn.Linear(attention_hidden_dim, 1)
self.b_in = nn.Parameter(
torch.zeros(
1,
)
)
self.b_out = nn.Parameter(
torch.zeros(
1,
)
)
nn.init.kaiming_uniform_(self.W_q.weight, a=math.sqrt(5))
nn.init.kaiming_uniform_(self.W_k.weight, a=math.sqrt(5))
nn.init.kaiming_uniform_(self.W_v.weight, a=math.sqrt(5))
nn.init.kaiming_uniform_(self.W_out.weight, a=math.sqrt(5))
self.Wh = nn.Parameter(
torch.randn(2 * attention_input_dim, attention_hidden_dim)
)
self.Wa = nn.Parameter(torch.randn(attention_hidden_dim, 1))
self.ba = nn.Parameter(
torch.zeros(
1,
)
)
nn.init.kaiming_uniform_(self.Wh, a=math.sqrt(5))
nn.init.kaiming_uniform_(self.Wa, a=math.sqrt(5))
self.dropout = nn.Dropout(p=dropout)
self.tanh = nn.Tanh()
self.softmax = nn.Softmax(dim=1)
self.sigmoid = nn.Sigmoid()
def forward(self, input):
(
batch_size,
time_step,
input_dim,
) = input.size() # batch_size * input_dim + 1 * hidden_dim(i)
input_q = self.W_q(input[:, -1, :]) # b h
input_k = self.W_k(input) # b t h
input_v = self.W_v(input) # b t h
if self.attention_type == "add": # B*T*I @ H*I
q = torch.reshape(
input_q, (batch_size, 1, self.attention_hidden_dim)
) # B*1*H
h = q + input_k + self.b_in # b t h
h = self.tanh(h) # B*T*H
e = self.W_out(h) # b t 1
e = torch.reshape(e, (batch_size, time_step)) # b t
elif self.attention_type == "mul":
q = torch.reshape(
input_q, (batch_size, self.attention_hidden_dim, 1)
) # B*h 1
e = torch.matmul(input_k, q).squeeze() # b t
elif self.attention_type == "concat":
q = input_q.unsqueeze(1).repeat(1, time_step, 1) # b t h
k = input_k
c = torch.cat((q, k), dim=-1) # B*T*2I
h = torch.matmul(c, self.Wh)
h = self.tanh(h)
e = torch.matmul(h, self.Wa) + self.ba # B*T*1
e = torch.reshape(e, (batch_size, time_step)) # b t
a = self.softmax(e) # B*T
if self.dropout is not None:
a = self.dropout(a)
v = torch.matmul(a.unsqueeze(1), input_v).squeeze() # B*I
return v, a
class PositionwiseFeedForward(nn.Module): # new added
"Implements FFN equation."
def __init__(self, d_model, d_ff, dropout=0.1):
super(PositionwiseFeedForward, self).__init__()
self.w_1 = nn.Linear(d_model, d_ff)
self.w_2 = nn.Linear(d_ff, d_model)
self.dropout = nn.Dropout(dropout)
def forward(self, x):
return self.w_2(self.dropout(F.relu(self.w_1(x)))), None
class PositionalEncoding(nn.Module): # new added / not use anymore
"Implement the PE function."
def __init__(self, d_model, dropout, max_len=400):
super(PositionalEncoding, self).__init__()
self.dropout = nn.Dropout(p=dropout)
# Compute the positional encodings once in log space.
pe = torch.zeros(max_len, d_model)
position = torch.arange(0.0, max_len).unsqueeze(1)
div_term = torch.exp(
torch.arange(0.0, d_model, 2) * -(math.log(10000.0) / d_model)
)
pe[:, 0::2] = torch.sin(position * div_term)
pe[:, 1::2] = torch.cos(position * div_term)
pe = pe.unsqueeze(0)
self.register_buffer("pe", pe)
def forward(self, x):
x = x + Variable(self.pe[:, : x.size(1)], requires_grad=False)
return self.dropout(x)
class MultiHeadedAttention(nn.Module):
def __init__(self, h, d_model, dropout=0):
"Take in model size and number of heads."
super(MultiHeadedAttention, self).__init__()
assert d_model % h == 0
# We assume d_v always equals d_k
self.d_k = d_model // h
self.h = h
self.linears = nn.ModuleList(
[nn.Linear(d_model, self.d_k * self.h) for _ in range(3)]
)
self.final_linear = nn.Linear(d_model, d_model)
self.attn = None
self.dropout = nn.Dropout(p=dropout)
def attention(self, query, key, value, mask=None, dropout=None):
"Compute 'Scaled Dot Product Attention'"
d_k = query.size(-1) # b h t d_k
scores = torch.matmul(query, key.transpose(-2, -1)) / math.sqrt(d_k) # b h t t
if mask is not None: # 1 1 t t
scores = scores.masked_fill(mask == 0, -1e9) # b h t t 下三角
p_attn = F.softmax(scores, dim=-1) # b h t t
if dropout is not None:
p_attn = dropout(p_attn)
return torch.matmul(p_attn, value), p_attn # b h t v (d_k)
def cov(self, m, y=None):
if y is not None:
m = torch.cat((m, y), dim=0)
m_exp = torch.mean(m, dim=1)
x = m - m_exp[:, None]
cov = 1 / (x.size(1) - 1) * x.mm(x.t())
return cov
def forward(self, query, key, value, mask=None):
if mask is not None:
# Same mask applied to all h heads.
mask = mask.unsqueeze(1) # 1 1 t t
nbatches = query.size(0) # b
input_dim = query.size(1) # i+1
feature_dim = query.size(1) # i+1
# input size -> # batch_size * d_input * hidden_dim
# d_model => h * d_k
query, key, value = [
l(x).view(nbatches, -1, self.h, self.d_k).transpose(1, 2)
for l, x in zip(self.linears, (query, key, value))
] # b num_head d_input d_k
x, self.attn = self.attention(
query, key, value, mask=mask, dropout=self.dropout
) # b num_head d_input d_v (d_k)
x = (
x.transpose(1, 2).contiguous().view(nbatches, -1, self.h * self.d_k)
) # batch_size * d_input * hidden_dim
# DeCov
DeCov_contexts = x.transpose(0, 1).transpose(1, 2) # I+1 H B
Covs = self.cov(DeCov_contexts[0, :, :])
DeCov_loss = 0.5 * (
torch.norm(Covs, p="fro") ** 2 - torch.norm(torch.diag(Covs)) ** 2
)
for i in range(feature_dim - 1):
Covs = self.cov(DeCov_contexts[i + 1, :, :])
DeCov_loss += 0.5 * (
torch.norm(Covs, p="fro") ** 2 - torch.norm(torch.diag(Covs)) ** 2
)
return self.final_linear(x), DeCov_loss
class LayerNorm(nn.Module):
def __init__(self, features, eps=1e-7):
super(LayerNorm, self).__init__()
self.a_2 = nn.Parameter(torch.ones(features))
self.b_2 = nn.Parameter(torch.zeros(features))
self.eps = eps
def forward(self, x):
mean = x.mean(-1, keepdim=True)
std = x.std(-1, keepdim=True)
return self.a_2 * (x - mean) / (std + self.eps) + self.b_2
class SublayerConnection(nn.Module):
"""
A residual connection followed by a layer norm.
Note for code simplicity the norm is first as opposed to last.
"""
def __init__(self, size, dropout):
super(SublayerConnection, self).__init__()
self.norm = LayerNorm(size)
self.dropout = nn.Dropout(dropout)
def forward(self, x, sublayer):
"Apply residual connection to any sublayer with the same size."
returned_value = sublayer(self.norm(x))
return x + self.dropout(returned_value[0]), returned_value[1]
class ConCare(nn.Module):
def __init__(
self,
lab_dim, # lab_dim
hidden_dim,
demo_dim,
d_model,
MHD_num_head,
d_ff,
# output_dim,
# device,
drop=0.5,
):
super(ConCare, self).__init__()
# hyperparameters
self.lab_dim = lab_dim
self.hidden_dim = hidden_dim # d_model
self.d_model = d_model
self.MHD_num_head = MHD_num_head
self.d_ff = d_ff
# self.output_dim = output_dim
self.drop = drop
self.demo_dim = demo_dim
# layers
self.PositionalEncoding = PositionalEncoding(
self.d_model, dropout=0, max_len=400
)
self.GRUs = nn.ModuleList(
[
copy.deepcopy(nn.GRU(1, self.hidden_dim, batch_first=True))
for _ in range(self.lab_dim)
]
)
self.LastStepAttentions = nn.ModuleList(
[
copy.deepcopy(
SingleAttention(
self.hidden_dim,
8,
attention_type="new",
demographic_dim=12,
time_aware=True,
use_demographic=False,
)
)
for _ in range(self.lab_dim)
]
)
self.FinalAttentionQKV = FinalAttentionQKV(
self.hidden_dim,
self.hidden_dim,
attention_type="mul",
dropout=self.drop,
)
self.MultiHeadedAttention = MultiHeadedAttention(
self.MHD_num_head, self.d_model, dropout=self.drop
)
self.SublayerConnection = SublayerConnection(self.d_model, dropout=self.drop)
self.PositionwiseFeedForward = PositionwiseFeedForward(
self.d_model, self.d_ff, dropout=0.1
)
self.demo_lab_proj = nn.Linear(self.demo_dim + self.lab_dim, self.hidden_dim)
self.demo_proj_main = nn.Linear(self.demo_dim, self.hidden_dim)
self.demo_proj = nn.Linear(self.demo_dim, self.hidden_dim)
self.output0 = nn.Linear(self.hidden_dim, self.hidden_dim)
# self.output1 = nn.Linear(self.hidden_dim, self.output_dim)
self.dropout = nn.Dropout(p=self.drop)
self.tanh = nn.Tanh()
self.softmax = nn.Softmax()
self.sigmoid = nn.Sigmoid()
self.relu = nn.ReLU()
def concare_encoder(self, input, demo_input, device):
# input shape [batch_size, timestep, feature_dim]
demo_main = self.tanh(self.demo_proj_main(demo_input)).unsqueeze(
1
) # b hidden_dim
batch_size = input.size(0)
time_step = input.size(1)
feature_dim = input.size(2)
assert feature_dim == self.lab_dim # input Tensor : 256 * 48 * 76
assert self.d_model % self.MHD_num_head == 0
# forward
GRU_embeded_input = self.GRUs[0](
input[:, :, 0].unsqueeze(-1).to(device=device),
Variable(
torch.zeros(batch_size, self.hidden_dim)
.to(device=device)
.unsqueeze(0)
),
)[
0
] # b t h
Attention_embeded_input = self.LastStepAttentions[0](GRU_embeded_input, device)[
0
].unsqueeze(
1
) # b 1 h
for i in range(feature_dim - 1):
embeded_input = self.GRUs[i + 1](
input[:, :, i + 1].unsqueeze(-1),
Variable(
torch.zeros(batch_size, self.hidden_dim)
.to(device=device)
.unsqueeze(0)
),
)[
0
] # b 1 h
embeded_input = self.LastStepAttentions[i + 1](embeded_input, device)[0].unsqueeze(
1
) # b 1 h
Attention_embeded_input = torch.cat(
(Attention_embeded_input, embeded_input), 1
) # b i h
Attention_embeded_input = torch.cat(
(Attention_embeded_input, demo_main), 1
) # b i+1 h
posi_input = self.dropout(
Attention_embeded_input
) # batch_size * d_input+1 * hidden_dim
contexts = self.SublayerConnection(
posi_input,
lambda x: self.MultiHeadedAttention(
posi_input, posi_input, posi_input, None
),
) # # batch_size * d_input * hidden_dim
DeCov_loss = contexts[1]
contexts = contexts[0]
contexts = self.SublayerConnection(
contexts, lambda x: self.PositionwiseFeedForward(contexts)
)[0]
weighted_contexts = self.FinalAttentionQKV(contexts)[0]
return weighted_contexts
def forward(self, x, device, info=None):
"""extra info is not used here"""
batch_size, time_steps, _ = x.size()
demo_input = x[:, 0, : self.demo_dim]
lab_input = x[:, :, self.demo_dim :]
out = torch.zeros((batch_size, time_steps, self.hidden_dim))
for cur_time in range(time_steps):
# print(cur_time, end=" ")
cur_lab = lab_input[:, : cur_time + 1, :]
# print("cur_lab", cur_lab.shape)
if cur_time == 0:
out[:, cur_time, :] = self.demo_lab_proj(x[:, 0, :])
else:
out[:, cur_time, :] = self.concare_encoder(cur_lab, demo_input, device)
# print()
return out
