Python与算法之美
阿里妈妈在CTR预估领域有3篇比较有名的文章。
2017年的深度兴趣网络, DIN(DeepInterestNetwork)。
2018年的深度兴趣演化网络, DIEN(DeepInterestEvolutionNetWork)。
2019年的深度会话兴趣网络, DSIN(DeepSessionInterestNetWork)。
这3篇文章的主要思想和相互关系用一句话分别概括如下：
第1篇DIN说，用户的行为日志中只有一部分和当前候选广告有关。可以利用Attention机制从用户行为日志中建模出和当前候选广告相关的用户兴趣表示。我们试过涨点了嘻嘻嘻。
第2篇DIEN说，用户最近的行为可能比较远的行为更加重要。可以用循环神经网络GRU建模用户兴趣随时间的演化。我们试过也涨点了嘿嘿嘿。
第3篇DSIN说，用户在同一次会话中的行为高度相关，在不同会话间的行为则相对独立。可以把用户行为日志按照时间间隔分割成会话并用SelfAttention机制建模它们之间的相互作用。我们试过又涨点了哈哈哈。
参考材料：
DIEN论文：https://arxiv.org/pdf/1809.03672.pdf
DIN+DIEN，机器学习唯一指定涨点技Attention：https://zhuanlan.zhihu.com/p/431131396
从DIN到DIEN看阿里CTR算法的进化脉络：https://zhuanlan.zhihu.com/p/78365283
代码实现参考：https://github.com/GitHub-HongweiZhang/prediction-flow
上一篇文章我们介绍了DIN, 本篇文章我们介绍DIEN。
DIEN这篇文章的主要创新之处有3点：
一是引入GRU来从用户行为日志序列中自然地抽取每个行为日志对应的用户兴趣表示(兴趣抽取层)。
二是设计了一个辅助loss层，通过做一个辅助任务(区分真实的用户历史点击行为和负采样的非用户点击行为)来强化用户兴趣表示的学习。
三是将注意力机制和GRU结构结合起来(AUGRU: Attention UPdate GRU)，来建模用户兴趣的时间演化得到最终的用户表示(兴趣演化层)。
其中引入辅助Loss的技巧是神经网络涨点非常通用的一种高级技巧，值得我们学习。
公众号后台回复关键词：DIEN，获取本文全部源码和Movielens数据集下载链接。
一，DIEN原理解析
DIEN的主要出发点是，用户最近的行为可能比较远的行为更加重要。可以用循环神经网络GRU建模用户兴趣随时间的演化。
DIEN选择的是不容易梯度消失且较快的GRU。
1, 兴趣抽取层
图中的 是用户的行为序列，而 是对应的embedding。随着自然发生的顺序， 被输入GRU中，这就是兴趣抽取层。
也是DIEN的第一条创新：引入GRU来从用户行为日志序列中自然地抽取每个行为日志对应的用户兴趣表示(兴趣抽取层)。
2，辅助loss
如果忽略上面的AUGRU环节，GRU中的隐状态 就应该成为用户的行为序列最后的表示。
如果直接就这样做，也不是不可以，但是学习到的东西可能不是我们想要的用户兴趣表示，或者说很难学习到有意义的信息。
因为 的迭代经过了很多步，然后还要和其他特征做拼接，然后还要经过MLP，最后才得到输出去计算Loss。
这样的结果就是最后来了一个正样本或负样本，反向传播很难归因到 上。
基于此DIEN给出了第二个要点：使用辅助Loss来强化的学习。
我们来看看这个辅助Loss是怎么做的？这里设计了一个辅助任务，使用来区分真实的用户历史点击行为和负采样的非用户点击行为。
由于 代表着 t 时刻的用户兴趣表示，我们可以用它来预测 t+1时刻的广告用户是否点击。
因为用户行为日志中都是用户点击过的广告(正样本, )，所以我们可以从全部的广告中给用户采样同样数量的用户没有点击过的广告作为负样本。
结合和 , 作为输入, 我们可以做一个二分类的辅助任务。
这个辅助任务给在每个t时刻都提供了一个监督信号，使得能够更好地成为用户兴趣的抽取表示。
真实应用场合下，你把开始的输入和最后的要求告诉网络，它就能给你一个好的结果的情况非常少。
大多数时候是需要你去控制每一步的输入输出，每一步的loss才能防止网络各种偷懒作弊。
辅助loss能够使得网络更受控制，向我们需要的方向发展，非常建议大家在实际业务中多试试辅助loss。
3，兴趣演化层
通过兴趣抽取层和辅助loss，我们得到了每个t时刻用户的一般兴趣表示。
注意这个兴趣表示是一般性的，还没有和我们的候选广告做Attention关联。
在DIN中，我们通过Attention机制构建了和候选广告相关的用户兴趣表示。
而在DIEN中，我们希望建立的是和和候选广告相关，并且和时间演化相关的用户兴趣表示。
DIEN通过结合Attention机制和GRU结构来做到这一点，这就是第三点创新AUGRU : Attention UPdate Gate GRU。
下面我们进行详细讲解。
一般地，各种RNN序列模型层(SimpleRNN,GRU,LSTM等)可以用函数表示如下:
这个公式的含义是：t时刻循环神经网络的输出向量由t-1时刻的输出向量和t时刻的输入变换而来。
为了结合Attention机制和GRU结构，我们需要设计这样的一个有三种输入的序列模型
这里的是 t时刻的用户兴趣表示输入 和候选广告计算出的attention 得分，是个标量。
我们先看看 GRU的 具体函数形式：
公式中的小圈表示哈达玛积，也就是两个向量逐位相乘。
其中(1)式和(2)式计算的是更新门和重置门，是两个长度和相同的向量。
更新门用于控制每一步被更新的比例，更新门越大，更新幅度越大。
重置门用于控制更新候选向量中前一步的状态被重新放入的比例，重置门越大，更新候选向量中被重新放进来的比例越大。
注意到(4)式 实际上和ResNet的残差结构是相似的，都是 f(x) = x + g(x) 的形式，可以有效地防止长序列学习反向传播过程中梯度消失问题。
如何在GRU的基础上把attention得分融入进来呢？有以下一些非常自然的想法：
1， 用缩放输入, 这就是AIGRU: Attention Input GRU。其含义是相关性高的在输入端进行放大。
2， 用代替GRU的更新门，这就是AGRU: Attention based GRU。其含义是用直接用相关性作为更新幅度。
3， 用缩放GRU的更新门，这就是AUGRU: Attention Update Gate GRU。其含义是用用相关性缩放更新幅度。
AIGRU实际上并没有改变GRU的结构，只是改变了其输入，这种方式对Attention的使用比较含蓄，我把每个历史广告的相关性强弱通过输入告诉GRU，GRU你就给我好好学吧，希望你把相关性强的广告多长点到脑子里。但是这种方式效果不是很理想，即使是相关性为0的历史广告，也会对进行更新。
AGRU是改变了GRU的结构的，并且对Attention的使用非常激进，完全删掉了GRU原有的的更新门，GRU你的脑子归Attention管了，遇到相关性高的广告，一定大大地记上一笔。不过AGRU也有一个缺陷，那就是Attention得分实际上是个标量，无法反应不同维度的差异。
AUGRU也是改变了GRU的结构的，并且对Attention的使用比较折衷，让Attention缩放GRU原有的更新幅度。GRU我给你找了个搭档Attention，你更新前先问问它，你两一起决定该迈多大的步子吧。
DIEN论文中通过对比实验发现AUGRU的效果最好。
我们看看AUGRU的核心实现代码。基本上和公式是一致的，应用了F.linear函数来实现矩阵乘法和加偏置。
import torch
from torch import nn
class AttentionUpdateGateGRUCell(nn.Module):
def __init__(self, input_size, hidden_size, bias=True):
super().__init__()
self.input_size = input_size
self.hidden_size = hidden_size
self.bias = bias
# (Wu|Wr|Wn)
self.weight_ih = nn.Parameter(
torch.Tensor(3 * hidden_size, input_size))
# (Uu|Ur|Un)
self.weight_hh = nn.Parameter(
torch.Tensor(3 * hidden_size, hidden_size))
if bias:
# (b_iu|b_ir|b_in)
self.bias_ih = nn.Parameter(torch.Tensor(3 * hidden_size))
# (b_hu|b_hr|b_hn)
self.bias_hh = nn.Parameter(torch.Tensor(3 * hidden_size))
else:
self.register_parameter('bias_ih', None)
self.register_parameter('bias_hh', None)
self.reset_parameters()
def reset_parameters(self):
stdv = 1.0 / (self.hidden_size)**0.5
for weight in self.parameters():
nn.init.uniform_(weight, -stdv, stdv)
def forward(self, x, hx, att_score):
gi = F.linear(x, self.weight_ih, self.bias_ih)
gh = F.linear(hx, self.weight_hh, self.bias_hh)
i_r, i_u, i_n = gi.chunk(3, 1)
h_r, h_u, h_n = gh.chunk(3, 1)
resetgate = torch.sigmoid(i_r + h_r)
updategate = torch.sigmoid(i_u + h_u)
newgate = torch.tanh(i_n + resetgate * h_n)
updategate = att_score.view(-1, 1) * updategate
hy = (1-updategate)*hx + updategate*newgate
return hy
二，DIEN的pytorch实现
下面是一个DIEN模型的完整pytorch实现。许多代码和DIN的实现是一样的。
这里的AttentionGroup类用来建立候选广告属性，历史广告属性，以及负采样的广告属性的pair关系。
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.nn.utils.rnn import pack_padded_sequence, pad_packed_sequence
from collections import OrderedDict
class MaxPooling(nn.Module):
def __init__(self, dim):
super(MaxPooling, self).__init__()
self.dim = dim
def forward(self, input):
return torch.max(input, self.dim)[0]
class SumPooling(nn.Module):
def __init__(self, dim):
super(SumPooling, self).__init__()
self.dim = dim
def forward(self, input):
return torch.sum(input, self.dim)
class Dice(nn.Module):
"""
The Data Adaptive Activation Function in DIN, a generalization of PReLu.
"""
def __init__(self, emb_size, dim=2, epsilon=1e-8):
super(Dice, self).__init__()
assert dim == 2 or dim == 3
self.bn = nn.BatchNorm1d(emb_size, eps=epsilon)
self.sigmoid = nn.Sigmoid()
self.dim = dim
# wrap alpha in nn.Parameter to make it trainable
self.alpha = nn.Parameter(torch.zeros((emb_size,))) if self.dim == 2 else nn.Parameter(
torch.zeros((emb_size, 1)))
def forward(self, x):
assert x.dim() == self.dim
if self.dim == 2:
x_p = self.sigmoid(self.bn(x))
out = self.alpha * (1 - x_p) * x + x_p * x
else:
x = torch.transpose(x, 1, 2)
x_p = self.sigmoid(self.bn(x))
out = self.alpha * (1 - x_p) * x + x_p * x
out = torch.transpose(out, 1, 2)
return out
class Identity(nn.Module):
def __init__(self):
super().__init__()
def forward(self, x):
return x
def get_activation_layer(name, hidden_size=None, dice_dim=2):
name = name.lower()
name_dict = {x.lower():x for x in dir(nn) if '__' not in x and 'Z'>=x[0]>='A'}
if name=="linear":
return Identity()
elif name=="dice":
assert dice_dim
return Dice(hidden_size, dice_dim)
else:
assert name in name_dict, f'activation type {name} not supported!'
return getattr(nn,name_dict[name])()
def init_weights(model):
if isinstance(model, nn.Linear):
if model.weight is not None:
nn.init.kaiming_uniform_(model.weight.data)
if model.bias is not None:
nn.init.normal_(model.bias.data)
elif isinstance(model, (nn.BatchNorm1d,nn.BatchNorm2d,nn.BatchNorm3d)):
if model.weight is not None:
nn.init.normal_(model.weight.data, mean=1, std=0.02)
if model.bias is not None:
nn.init.constant_(model.bias.data, 0)
else:
pass
class MLP(nn.Module):
def __init__(self, input_size, hidden_layers,
dropout=0.0, batchnorm=True, activation='relu'):
super(MLP, self).__init__()
modules = OrderedDict()
previous_size = input_size
for index, hidden_layer in enumerate(hidden_layers):
modules[f"dense{index}"] = nn.Linear(previous_size, hidden_layer)
if batchnorm:
modules[f"batchnorm{index}"] = nn.BatchNorm1d(hidden_layer)
if activation:
modules[f"activation{index}"] = get_activation_layer(activation,hidden_layer,2)
if dropout:
modules[f"dropout{index}"] = nn.Dropout(dropout)
previous_size = hidden_layer
self.mlp = nn.Sequential(modules)
def forward(self, x):
return self.mlp(x)
class AttentionGRUCell(nn.Module):
def __init__(self, input_size, hidden_size, bias=True):
super().__init__()
self.input_size = input_size
self.hidden_size = hidden_size
self.bias = bias
# (Wr|Wn)
self.weight_ih = nn.Parameter(
torch.Tensor(2 * hidden_size, input_size))
# (Ur|Un)
self.weight_hh = nn.Parameter(
torch.Tensor(2 * hidden_size, hidden_size))
if bias:
# (b_ir|b_in)
self.bias_ih = nn.Parameter(torch.Tensor(2 * hidden_size))
# (b_hr|b_hn)
self.bias_hh = nn.Parameter(torch.Tensor(2 * hidden_size))
else:
self.register_parameter('bias_ih', None)
self.register_parameter('bias_hh', None)
self.reset_parameters()
def reset_parameters(self):
stdv = 1.0 / (self.hidden_size)**0.5
for weight in self.parameters():
nn.init.uniform_(weight, -stdv, stdv)
def forward(self, x, hx, att_score):
gi = F.linear(x, self.weight_ih, self.bias_ih)
gh = F.linear(hx, self.weight_hh, self.bias_hh)
i_r, i_n = gi.chunk(2, 1)
h_r, h_n = gh.chunk(2, 1)
resetgate = torch.sigmoid(i_r + h_r)
newgate = torch.tanh(i_n + resetgate * h_n)
att_score = att_score.view(-1, 1)
hy = (1. - att_score) * hx + att_score * newgate
return hy
class AttentionUpdateGateGRUCell(nn.Module):
def __init__(self, input_size, hidden_size, bias=True):
super().__init__()
self.input_size = input_size
self.hidden_size = hidden_size
self.bias = bias
# (Wu|Wr|Wn)
self.weight_ih = nn.Parameter(
torch.Tensor(3 * hidden_size, input_size))
# (Uu|Ur|Un)
self.weight_hh = nn.Parameter(
torch.Tensor(3 * hidden_size, hidden_size))
if bias:
# (b_iu|b_ir|b_in)
self.bias_ih = nn.Parameter(torch.Tensor(3 * hidden_size))
# (b_hu|b_hr|b_hn)
self.bias_hh = nn.Parameter(torch.Tensor(3 * hidden_size))
else:
self.register_parameter('bias_ih', None)
self.register_parameter('bias_hh', None)
self.reset_parameters()
def reset_parameters(self):
stdv = 1.0 / (self.hidden_size)**0.5
for weight in self.parameters():
nn.init.uniform_(weight, -stdv, stdv)
def forward(self, x, hx, att_score):
gi = F.linear(x, self.weight_ih, self.bias_ih)
gh = F.linear(hx, self.weight_hh, self.bias_hh)
i_u,i_r, i_n = gi.chunk(3, 1)
h_u,h_r, h_n = gh.chunk(3, 1)
updategate = torch.sigmoid(i_u + h_u)
resetgate = torch.sigmoid(i_r + h_r)
newgate = torch.tanh(i_n + resetgate * h_n)
updategate = att_score.view(-1, 1) * updategate
hy = (1-updategate)*hx + updategate*newgate
return hy
class DynamicGRU(nn.Module):
def __init__(self, input_size, hidden_size, bias=True, gru_type='AGRU'):
super(DynamicGRU, self).__init__()
self.input_size = input_size
self.hidden_size = hidden_size
if gru_type == 'AGRU':
self.rnn = AttentionGRUCell(input_size, hidden_size, bias)
elif gru_type == 'AUGRU':
self.rnn = AttentionUpdateGateGRUCell(
input_size, hidden_size, bias)
def forward(self, x, att_scores, hx=None):
is_packed_input = isinstance(x, nn.utils.rnn.PackedSequence)
if not is_packed_input:
raise NotImplementedError(
"DynamicGRU only supports packed input")
is_packed_att_scores = isinstance(att_scores, nn.utils.rnn.PackedSequence)
if not is_packed_att_scores:
raise NotImplementedError(
"DynamicGRU only supports packed att_scores")
x, batch_sizes, sorted_indices, unsorted_indices = x
att_scores, _, _, _ = att_scores
max_batch_size = batch_sizes[0]
max_batch_size = int(max_batch_size)
if hx is None:
hx = torch.zeros(
max_batch_size, self.hidden_size,
dtype=x.dtype, device=x.device)
outputs = torch.zeros(
x.size(0), self.hidden_size,
dtype=x.dtype, device=x.device)
begin = 0
for batch in batch_sizes:
new_hx = self.rnn(
x[begin: begin + batch],
hx[0:batch],
att_scores[begin: begin + batch])
outputs[begin: begin + batch] = new_hx
hx = new_hx
begin += batch
return nn.utils.rnn.PackedSequence(
outputs, batch_sizes, sorted_indices, unsorted_indices)
class Attention(nn.Module):
def __init__(
self,
input_size,
hidden_layers,
dropout=0.0,
batchnorm=True,
activation='prelu',
return_scores=False):
super().__init__()
self.return_scores = return_scores
self.mlp = MLP(
input_size=input_size * 4,
hidden_layers=hidden_layers,
dropout=dropout,
batchnorm=batchnorm,
activation=activation)
self.fc = nn.Linear(hidden_layers[-1], 1)
def forward(self, query, keys, keys_length):
"""
Parameters
----------
query: 2D tensor, [Batch, Hidden]
keys: 3D tensor, [Batch, Time, Hidden]
keys_length: 1D tensor, [Batch]
Returns
-------
outputs: 2D tensor, [Batch, Hidden]
"""
batch_size, max_length, dim = keys.size()
query = query.unsqueeze(1).expand(-1, max_length, -1)
din_all = torch.cat(
[query, keys, query - keys, query * keys], dim=-1)
din_all = din_all.view(batch_size * max_length, -1)
outputs = self.mlp(din_all)
outputs = self.fc(outputs).view(batch_size, max_length) # [B, T]
# Scale
outputs = outputs / (dim ** 0.5)
# Mask
mask = (torch.arange(max_length, device=keys_length.device).repeat(
batch_size, 1) < keys_length.view(-1, 1))
outputs[~mask] = -np.inf
# Activation
outputs = F.softmax(outputs, dim=1) #DIN uses sigmoid,DIEN uses softmax; [B, T]
if not self.return_scores:
# Weighted sum
outputs = torch.matmul(
outputs.unsqueeze(1), keys).squeeze() # [B, H]
return outputs
class AuxiliaryNet(nn.Module):
def __init__(self, input_size, hidden_layers, activation='sigmoid'):
super().__init__()
modules = OrderedDict()
previous_size = input_size
for index, hidden_layer in enumerate(hidden_layers):
modules[f"dense{index}"] = nn.Linear(previous_size, hidden_layer)
if activation:
modules[f"activation{index}"] = get_activation_layer(activation)
previous_size = hidden_layer
modules["final_layer"] = nn.Linear(previous_size, 1)
self.mlp = nn.Sequential(modules)
def forward(self, x):
return torch.sigmoid(self.mlp(x))
class Interest(nn.Module):
SUPPORTED_GRU_TYPE = ['GRU', 'AIGRU', 'AGRU', 'AUGRU']
def __init__(
self,
input_size,
gru_type='AUGRU',
gru_dropout=0.0,
att_hidden_layers=[80, 40],
att_dropout=0.0,
att_batchnorm=True,
att_activation='prelu',
use_negsampling=False):
super(Interest, self).__init__()
if gru_type not in Interest.SUPPORTED_GRU_TYPE:
raise NotImplementedError(f"gru_type: {gru_type} is not supported")
self.gru_type = gru_type
self.use_negsampling = use_negsampling
self.interest_extractor = nn.GRU(
input_size=input_size,
hidden_size=input_size,
batch_first=True,
bidirectional=False)
if self.use_negsampling:
self.auxiliary_net = AuxiliaryNet(
input_size * 2, hidden_layers=[100, 50])
if gru_type == 'GRU':
self.attention = Attention(
input_size=input_size,
hidden_layers=att_hidden_layers,
dropout=att_dropout,
batchnorm=att_batchnorm,
activation=att_activation)
self.interest_evolution = nn.GRU(
input_size=input_size,
hidden_size=input_size,
batch_first=True,
bidirectional=False)
elif gru_type == 'AIGRU':
self.attention = Attention(
input_size=input_size,
hidden_layers=att_hidden_layers,
dropout=att_dropout,
batchnorm=att_batchnorm,
activation=att_activation,
return_scores=True)
self.interest_evolution = nn.GRU(
input_size=input_size,
hidden_size=input_size,
batch_first=True,
bidirectional=False)
elif gru_type == 'AGRU' or gru_type == 'AUGRU':
self.attention = Attention(
input_size=input_size,
hidden_layers=att_hidden_layers,
dropout=att_dropout,
batchnorm=att_batchnorm,
activation=att_activation,
return_scores=True)
self.interest_evolution = DynamicGRU(
input_size=input_size,
hidden_size=input_size,
gru_type=gru_type)
@staticmethod
def get_last_state(states, keys_length):
# states [B, T, H]
batch_size, max_seq_length, hidden_size = states.size()
mask = (torch.arange(max_seq_length, device=keys_length.device).repeat(
batch_size, 1) == (keys_length.view(-1, 1) - 1))
return states[mask]
def cal_auxiliary_loss(
self, states, click_seq, noclick_seq, keys_length):
# states [B, T, H]
# click_seq [B, T, H]
# noclick_seq [B, T, H]
# keys_length [B]
batch_size, max_seq_length, embedding_size = states.size()
mask = (torch.arange(max_seq_length, device=states.device).repeat(
batch_size, 1) < keys_length.view(-1, 1)).float()
click_input = torch.cat([states, click_seq], dim=-1)
noclick_input = torch.cat([states, noclick_seq], dim=-1)
embedding_size = embedding_size * 2
click_p = self.auxiliary_net(
click_input.view(
batch_size * max_seq_length, embedding_size)).view(
batch_size, max_seq_length)[mask > 0].view(-1, 1)
click_target = torch.ones(
click_p.size(), dtype=torch.float, device=click_p.device)
noclick_p = self.auxiliary_net(
noclick_input.view(
batch_size * max_seq_length, embedding_size)).view(
batch_size, max_seq_length)[mask > 0].view(-1, 1)
noclick_target = torch.zeros(
noclick_p.size(), dtype=torch.float, device=noclick_p.device)
loss = F.binary_cross_entropy(
torch.cat([click_p, noclick_p], dim=0),
torch.cat([click_target, noclick_target], dim=0))
return loss
def forward(self, query, keys, keys_length, neg_keys=None):
"""
Parameters
----------
query: 2D tensor, [Batch, Hidden]
keys: 3D tensor, [Batch, Time, Hidden]
keys_length: 1D tensor, [Batch]
neg_keys: 3D tensor, [Batch, Time, Hidden]
Returns
-------
outputs: 2D tensor, [Batch, Hidden]
"""
batch_size, max_length, dim = keys.size()
packed_keys = pack_padded_sequence(
keys,
lengths=keys_length.squeeze().cpu(),
batch_first=True,
enforce_sorted=False)
packed_interests, _ = self.interest_extractor(packed_keys)
aloss = None
if (self.gru_type != 'GRU') or self.use_negsampling:
interests, _ = pad_packed_sequence(
packed_interests,
batch_first=True,
padding_value=0.0,
total_length=max_length)
if self.use_negsampling:
aloss = self.cal_auxiliary_loss(
interests[:, :-1, :],
keys[:, 1:, :],
neg_keys[:, 1:, :],
keys_length - 1)
if self.gru_type == 'GRU':
packed_interests, _ = self.interest_evolution(packed_interests)
interests, _ = pad_packed_sequence(
packed_interests,
batch_first=True,
padding_value=0.0,
total_length=max_length)
outputs = self.attention(query, interests, keys_length)
elif self.gru_type == 'AIGRU':
# attention
scores = self.attention(query, interests, keys_length)
interests = interests * scores.unsqueeze(-1)
packed_interests = pack_padded_sequence(
interests,
lengths=keys_length.squeeze().cpu(),
batch_first=True,
enforce_sorted=False)
_, outputs = self.interest_evolution(packed_interests)
outputs = outputs.squeeze()
elif self.gru_type == 'AGRU' or self.gru_type == 'AUGRU':
# attention
scores = self.attention(query, interests, keys_length)
packed_interests = pack_padded_sequence(
interests,
lengths=keys_length.squeeze().cpu(),
batch_first=True,
enforce_sorted=False)
packed_scores = pack_padded_sequence(
scores,
lengths=keys_length.squeeze().cpu(),
batch_first=True,
enforce_sorted=False)
outputs, _ = pad_packed_sequence(
self.interest_evolution(
packed_interests, packed_scores), batch_first=True)
# pick last state
outputs = Interest.get_last_state(
outputs, keys_length.squeeze())
return outputs, aloss
class AttentionGroup(object):
def __init__(self, name, pairs,
hidden_layers, activation='dice', att_dropout=0.0,
gru_type='AUGRU', gru_dropout=0.0):
self.name = name
self.pairs = pairs
self.hidden_layers = hidden_layers
self.activation = activation
self.att_dropout = att_dropout
self.gru_type = gru_type
self.gru_dropout = gru_dropout
self.related_feature_names = set()
self.neg_feature_names = set()
for pair in pairs:
self.related_feature_names.add(pair['ad'])
self.related_feature_names.add(pair['pos_hist'])
if 'neg_hist' in pair:
self.related_feature_names.add(pair['neg_hist'])
self.neg_feature_names.add(pair['neg_hist'])
def is_attention_feature(self, feature_name):
if feature_name in self.related_feature_names:
return True
return False
def is_neg_sampling_feature(self, feature_name):
if feature_name in self.neg_feature_names:
return True
return False
@property
def pairs_count(self):
return len(self.pairs)
class DIEN(nn.Module):
def __init__(self, num_features,cat_features,seq_features,
cat_nums,embedding_size, attention_groups,
mlp_hidden_layers, mlp_activation='prelu', mlp_dropout=0.0,
use_negsampling = False,
d_out = 1
):
super().__init__()
self.num_features = num_features
self.cat_features = cat_features
self.seq_features = seq_features
self.cat_nums = cat_nums
self.embedding_size = embedding_size
self.attention_groups = attention_groups
self.mlp_hidden_layers = mlp_hidden_layers
self.mlp_activation = mlp_activation
self.mlp_dropout = mlp_dropout
self.d_out = d_out
self.use_negsampling = use_negsampling
#embedding
self.embeddings = OrderedDict()
for feature in self.cat_features+self.seq_features:
self.embeddings[feature] = nn.Embedding(
self.cat_nums[feature], self.embedding_size, padding_idx=0)
self.add_module(f"embedding:{feature}",self.embeddings[feature])
self.sequence_poolings = OrderedDict()
self.attention_poolings = OrderedDict()
total_embedding_sizes = 0
for feature in self.cat_features:
total_embedding_sizes += self.embedding_size
for feature in self.seq_features:
if not self.is_neg_sampling_feature(feature):
total_embedding_sizes += self.embedding_size
#sequence_pooling
for feature in self.seq_features:
if not self.is_attention_feature(feature):
self.sequence_poolings[feature] = MaxPooling(1)
self.add_module(f"pooling:{feature}",self.sequence_poolings[feature])
#attention_pooling
for attention_group in self.attention_groups:
self.attention_poolings[attention_group.name] = (
self.create_attention_fn(attention_group))
self.add_module(f"attention_pooling:{attention_group.name}",
self.attention_poolings[attention_group.name])
total_input_size = total_embedding_sizes+len(self.num_features)
self.mlp = MLP(
total_input_size,
mlp_hidden_layers,
dropout=mlp_dropout, batchnorm=True, activation=mlp_activation)
self.final_layer = nn.Linear(mlp_hidden_layers[-1], self.d_out)
self.apply(init_weights)
def forward(self, x):
final_layer_inputs = list()
# linear
number_inputs = list()
for feature in self.num_features:
number_inputs.append(x[feature].view(-1, 1))
embeddings = OrderedDict()
for feature in self.cat_features:
embeddings[feature] = self.embeddings[feature](x[feature])
for feature in self.seq_features:
if not self.is_attention_feature(feature):
embeddings[feature] = self.sequence_poolings[feature](
self.embeddings[feature](x[feature]))
auxiliary_losses = []
for attention_group in self.attention_groups:
query = torch.cat(
[embeddings[pair['ad']]
for pair in attention_group.pairs],
dim=-1)
pos_hist = torch.cat(
[self.embeddings[pair['pos_hist']](
x[pair['pos_hist']]) for pair in attention_group.pairs],
dim=-1)
#hist_length = torch.sum(hist>0,axis=1)
keys_length = torch.min(torch.cat(
[torch.sum(x[pair['pos_hist']]>0,axis=1).view(-1, 1)
for pair in attention_group.pairs],
dim=-1), dim=-1)[0]
neg_hist = None
if self.use_negsampling:
neg_hist = torch.cat(
[self.embeddings[pair['neg_hist']](
x[pair['neg_hist']])
for pair in attention_group.pairs],
dim=-1)
embeddings[attention_group.name], tmp_loss = (
self.attention_poolings[attention_group.name](
query, pos_hist, keys_length, neg_hist))
if tmp_loss is not None:
auxiliary_losses.append(tmp_loss)
emb_concat = torch.cat(number_inputs + [
emb for emb in embeddings.values()], dim=-1)
final_layer_inputs = self.mlp(emb_concat)
output = self.final_layer(final_layer_inputs)
auxiliary_avg_loss = None
if auxiliary_losses:
auxiliary_avg_loss = auxiliary_losses[0]
size = len(auxiliary_losses)
for i in range(1, size):
auxiliary_avg_loss += auxiliary_losses[i]
auxiliary_avg_loss /= size
if self.d_out==1:
output = output.squeeze()
return output, auxiliary_avg_loss
def create_attention_fn(self, attention_group):
return Interest(
attention_group.pairs_count * self.embedding_size,
gru_type=attention_group.gru_type,
gru_dropout=attention_group.gru_dropout,
att_hidden_layers=attention_group.hidden_layers,
att_dropout=attention_group.att_dropout,
att_activation=attention_group.activation,
use_negsampling=self.use_negsampling)
def is_attention_feature(self, feature):
for group in self.attention_groups:
if group.is_attention_feature(feature):
return True
return False
def is_neg_sampling_feature(self, feature):
for group in self.attention_groups:
if group.is_neg_sampling_feature(feature):
return True
return False
三，Movielens数据集完整范例
下面是一个基于Movielens评价数据集的DIEN完整范例，根据用户过去对一些电影的评价结果，来预测用户对候选电影是否会给好评。
这个数据集不大，用CPU就能跑。😁
1，准备数据
import numpy as np
import pandas as pd
from sklearn.base import BaseEstimator, TransformerMixin
from sklearn.preprocessing import QuantileTransformer
from sklearn.pipeline import Pipeline, FeatureUnion
from sklearn.impute import SimpleImputer
from collections import Counter
class CategoryEncoder(BaseEstimator, TransformerMixin):
def __init__(self, min_cnt=5, word2idx=None, idx2word=None):
super().__init__()
self.min_cnt = min_cnt
self.word2idx = word2idx if word2idx else dict()
self.idx2word = idx2word if idx2word else dict()
def fit(self, x, y=None):
if not self.word2idx:
counter = Counter(np.asarray(x).ravel())
selected_terms = sorted(
list(filter(lambda x: counter[x] >= self.min_cnt, counter)))
self.word2idx = dict(
zip(selected_terms, range(1, len(selected_terms) + 1)))
self.word2idx['__PAD__'] = 0
if '__UNKNOWN__' not in self.word2idx:
self.word2idx['__UNKNOWN__'] = len(self.word2idx)
if not self.idx2word:
self.idx2word = {
index: word for word, index in self.word2idx.items()}
return self
def transform(self, x):
transformed_x = list()
for term in np.asarray(x).ravel():
try:
transformed_x.append(self.word2idx[term])
except KeyError:
transformed_x.append(self.word2idx['__UNKNOWN__'])
return np.asarray(transformed_x, dtype=np.int64)
def dimension(self):
return len(self.word2idx)
class SequenceEncoder(BaseEstimator, TransformerMixin):
def __init__(self, sep=' ', min_cnt=5, max_len=None,
word2idx=None, idx2word=None):
super().__init__()
self.sep = sep
self.min_cnt = min_cnt
self.max_len = max_len
self.word2idx = word2idx if word2idx else dict()
self.idx2word = idx2word if idx2word else dict()
def fit(self, x, y=None):
if not self.word2idx:
counter = Counter()
max_len = 0
for sequence in np.array(x).ravel():
words = sequence.split(self.sep)
counter.update(words)
max_len = max(max_len, len(words))
if self.max_len is None:
self.max_len = max_len
# drop rare words
words = sorted(
list(filter(lambda x: counter[x] >= self.min_cnt, counter)))
self.word2idx = dict(zip(words, range(1, len(words) + 1)))
self.word2idx['__PAD__'] = 0
if '__UNKNOWN__' not in self.word2idx:
self.word2idx['__UNKNOWN__'] = len(self.word2idx)
if not self.idx2word:
self.idx2word = {
index: word for word, index in self.word2idx.items()}
if not self.max_len:
max_len = 0
for sequence in np.array(x).ravel():
words = sequence.split(self.sep)
max_len = max(max_len, len(words))
self.max_len = max_len
return self
def transform(self, x):
transformed_x = list()
for sequence in np.asarray(x).ravel():
words = list()
for word in sequence.split(self.sep):
try:
words.append(self.word2idx[word])
except KeyError:
words.append(self.word2idx['__UNKNOWN__'])
transformed_x.append(
np.asarray(words[0:self.max_len], dtype=np.int64))
return np.asarray(transformed_x, dtype=object)
def dimension(self):
return len(self.word2idx)
def max_length(self):
return self.max_len
from sklearn.preprocessing import QuantileTransformer
from sklearn.pipeline import Pipeline
from sklearn.impute import SimpleImputer
from tqdm import tqdm
dftrain = pd.read_csv("../data/ml_1m/train.csv")
dfval = pd.read_csv("../data/ml_1m/test.csv")
for col in ["movieId","histHighRatedMovieIds","negHistMovieIds","genres"]:
dftrain[col] = dftrain[col].astype(str)
dfval[col] = dfval[col].astype(str)
num_features = ['age']
cat_features = ['gender', 'movieId', 'occupation', 'zipCode']
seq_features = ['genres', 'histHighRatedMovieIds', 'negHistMovieIds']
num_pipe = Pipeline(steps = [('impute',SimpleImputer()),('quantile',QuantileTransformer())])
encoders = {}
print("preprocess number features...")
dftrain[num_features] = num_pipe.fit_transform(dftrain[num_features]).astype(np.float32)
dfval[num_features] = num_pipe.transform(dfval[num_features]).astype(np.float32)
print("preprocess category features...")
for col in tqdm(cat_features):
encoders[col] = CategoryEncoder(min_cnt=5)
dftrain[col] = encoders[col].fit_transform(dftrain[col])
dfval[col] = encoders[col].transform(dfval[col])
print("preprocess sequence features...")
for col in tqdm(seq_features):
encoders[col] = SequenceEncoder(sep="|",min_cnt=5)
dftrain[col] = encoders[col].fit_transform(dftrain[col])
dfval[col] = encoders[col].transform(dfval[col])
from collections import OrderedDict
from itertools import chain
from torch.utils.data import Dataset,DataLoader
class Df2Dataset(Dataset):
def __init__(self, dfdata, num_features, cat_features,
seq_features, encoders, label_col="label"):
self.dfdata = dfdata
self.num_features = num_features
self.cat_features = cat_features
self.seq_features = seq_features
self.encoders = encoders
self.label_col = label_col
self.size = len(self.dfdata)
def __len__(self):
return self.size
@staticmethod
def pad_sequence(sequence,max_length):
#zero is special index for padding
padded_seq = np.zeros(max_length, np.int32)
padded_seq[0: sequence.shape[0]] = sequence
return padded_seq
def __getitem__(self, idx):
record = OrderedDict()
for col in self.num_features:
record[col] = self.dfdata[col].iloc[idx].astype(np.float32)
for col in self.cat_features:
record[col] = self.dfdata[col].iloc[idx].astype(np.int64)
for col in self.seq_features:
seq = self.dfdata[col].iloc[idx]
max_length = self.encoders[col].max_length()
record[col] = Df2Dataset.pad_sequence(seq,max_length)
if self.label_col is not None:
record['label'] = self.dfdata[self.label_col].iloc[idx].astype(np.float32)
return record
def get_num_batches(self, batch_size):
return np.ceil(self.size / batch_size)
ds_train = Df2Dataset(dftrain, num_features, cat_features, seq_features, encoders)
ds_val = Df2Dataset(dfval,num_features, cat_features, seq_features, encoders)
dl_train = DataLoader(ds_train, batch_size=128,shuffle=True)
dl_val = DataLoader(ds_val,batch_size=128,shuffle=False)
cat_nums = {k:v.dimension() for k,v in encoders.items()}
for batch in dl_train:
break
print(cat_nums)
2，定义模型
def create_net():
augru_attention_groups_with_neg = [
AttentionGroup(
name='group1',
pairs=[{'ad': 'movieId', 'pos_hist': 'histHighRatedMovieIds', 'neg_hist': 'negHistMovieIds'}],
hidden_layers=[16, 8], att_dropout=0.1, gru_type='AUGRU')
]
net = DIEN(num_features=num_features,
cat_features=cat_features,
seq_features=seq_features,
cat_nums = cat_nums,
embedding_size=16,
attention_groups=augru_attention_groups_with_neg,
mlp_hidden_layers=[32,16],
mlp_activation="prelu",
mlp_dropout=0.25,
use_negsampling=True,
d_out=1
)
return net
net = create_net()
out,aloss = net.forward(batch)
from torchkeras.summary import summary
summary(net,input_data=batch);
3，训练模型
import os,sys,time
import numpy as np
import pandas as pd
import datetime
from tqdm import tqdm
import torch
from torch import nn
from accelerate import Accelerator
from copy import deepcopy
def printlog(info):
nowtime = datetime.datetime.now().strftime('%Y-%m-%d %H:%M:%S')
print("\n"+"=========="*8 + "%s"%nowtime)
print(str(info)+"\n")
class StepRunner:
def __init__(self, net, loss_fn,stage = "train", metrics_dict = None,
optimizer = None, lr_scheduler = None,
accelerator = None
):
self.net,self.loss_fn,self.metrics_dict,self.stage = net,loss_fn,metrics_dict,stage
self.optimizer,self.lr_scheduler = optimizer,lr_scheduler
self.accelerator = accelerator
def __call__(self, batch):
#loss
preds,aloss = self.net(batch)
loss = self.loss_fn(preds,batch["label"])+aloss
#backward()
if self.optimizer is not None and self.stage=="train":
if self.accelerator is None:
loss.backward()
else:
self.accelerator.backward(loss)
self.optimizer.step()
if self.lr_scheduler is not None:
self.lr_scheduler.step()
self.optimizer.zero_grad()
#metrics
step_metrics = {self.stage+"_"+name:metric_fn(preds, batch["label"]).item()
for name,metric_fn in self.metrics_dict.items()}
return loss.item(),step_metrics
class EpochRunner:
def __init__(self,steprunner):
self.steprunner = steprunner
self.stage = steprunner.stage
self.steprunner.net.train() if self.stage=="train" else self.steprunner.net.eval()
def __call__(self,dataloader):
total_loss,step = 0,0
loop = tqdm(enumerate(dataloader), total =len(dataloader))
for i, batch in loop:
if self.stage=="train":
loss, step_metrics = self.steprunner(batch)
else:
with torch.no_grad():
loss, step_metrics = self.steprunner(batch)
step_log = dict({self.stage+"_loss":loss},**step_metrics)
total_loss += loss
step+=1
if i!=len(dataloader)-1:
loop.set_postfix(**step_log)
else:
epoch_loss = total_loss/step
epoch_metrics = {self.stage+"_"+name:metric_fn.compute().item()
for name,metric_fn in self.steprunner.metrics_dict.items()}
epoch_log = dict({self.stage+"_loss":epoch_loss},**epoch_metrics)
loop.set_postfix(**epoch_log)
for name,metric_fn in self.steprunner.metrics_dict.items():
metric_fn.reset()
return epoch_log
class KerasModel(torch.nn.Module):
def __init__(self,net,loss_fn,metrics_dict=None,optimizer=None,lr_scheduler = None):
super().__init__()
self.accelerator = Accelerator()
self.history = {}
self.net = net
self.loss_fn = loss_fn
self.metrics_dict = nn.ModuleDict(metrics_dict)
self.optimizer = optimizer if optimizer is not None else torch.optim.Adam(
self.parameters(), lr=1e-2)
self.lr_scheduler = lr_scheduler
self.net,self.loss_fn,self.metrics_dict,self.optimizer = self.accelerator.prepare(
self.net,self.loss_fn,self.metrics_dict,self.optimizer)
def forward(self, x):
if self.net:
return self.net.forward(x)[0]
else:
raise NotImplementedError
def fit(self, train_data, val_data=None, epochs=10, ckpt_path='checkpoint.pt',
patience=5, monitor="val_loss", mode="min"):
train_data = self.accelerator.prepare(train_data)
val_data = self.accelerator.prepare(val_data) if val_data else []
for epoch in range(1, epochs+1):
printlog("Epoch {0} / {1}".format(epoch, epochs))
# 1，train -------------------------------------------------
train_step_runner = StepRunner(net = self.net,stage="train",
loss_fn = self.loss_fn,metrics_dict=deepcopy(self.metrics_dict),
optimizer = self.optimizer, lr_scheduler = self.lr_scheduler,
accelerator = self.accelerator)
train_epoch_runner = EpochRunner(train_step_runner)
train_metrics = train_epoch_runner(train_data)
for name, metric in train_metrics.items():
self.history[name] = self.history.get(name, []) + [metric]
# 2，validate -------------------------------------------------
if val_data:
val_step_runner = StepRunner(net = self.net,stage="val",
loss_fn = self.loss_fn,metrics_dict=deepcopy(self.metrics_dict),
accelerator = self.accelerator)
val_epoch_runner = EpochRunner(val_step_runner)
with torch.no_grad():
val_metrics = val_epoch_runner(val_data)
val_metrics["epoch"] = epoch
for name, metric in val_metrics.items():
self.history[name] = self.history.get(name, []) + [metric]
# 3，early-stopping -------------------------------------------------
arr_scores = self.history[monitor]
best_score_idx = np.argmax(arr_scores) if mode=="max" else np.argmin(arr_scores)
if best_score_idx==len(arr_scores)-1:
torch.save(self.net.state_dict(),ckpt_path)
print("<<<<<< reach best {0} : {1} >>>>>>".format(monitor,
arr_scores[best_score_idx]),file=sys.stderr)
if len(arr_scores)-best_score_idx>patience:
print("<<<<<< {} without improvement in {} epoch, early stopping >>>>>>".format(
monitor,patience),file=sys.stderr)
self.net.load_state_dict(torch.load(ckpt_path))
break
return pd.DataFrame(self.history)
@torch.no_grad()
def evaluate(self, val_data):
val_data = self.accelerator.prepare(val_data)
val_step_runner = StepRunner(net = self.net,stage="val",
loss_fn = self.loss_fn,metrics_dict=deepcopy(self.metrics_dict),
accelerator = self.accelerator)
val_epoch_runner = EpochRunner(val_step_runner)
val_metrics = val_epoch_runner(val_data)
return val_metrics
@torch.no_grad()
def predict(self, dataloader):
dataloader = self.accelerator.prepare(dataloader)
self.net.eval()
result = torch.cat([self.forward(t) for t in dataloader])
return result.data
from torchkeras.metrics import AUC
loss_fn = nn.BCEWithLogitsLoss()
metrics_dict = {"auc":AUC()}
optimizer = torch.optim.Adam(net.parameters(), lr=0.002, weight_decay=0.001)
#lr_scheduler = torch.optim.lr_scheduler.CyclicLR(optimizer, base_lr=0.01, max_lr=0.1)
model = KerasModel(net,
loss_fn = loss_fn,
metrics_dict= metrics_dict,
optimizer = optimizer,
#lr_scheduler = lr_scheduler
)
dfhistory = model.fit(train_data=dl_train,val_data=dl_val,epochs=100, patience=10,
monitor = "val_auc",mode="max",ckpt_path='checkpoint.pt')
4，评估模型
%matplotlib inline
%config InlineBackend.figure_format = 'svg'
import matplotlib.pyplot as plt
def plot_metric(dfhistory, metric):
train_metrics = dfhistory["train_"+metric]
val_metrics = dfhistory['val_'+metric]
epochs = range(1, len(train_metrics) + 1)
plt.plot(epochs, train_metrics, 'bo--')
plt.plot(epochs, val_metrics, 'ro-')
plt.title('Training and validation '+ metric)
plt.xlabel("Epochs")
plt.ylabel(metric)
plt.legend(["train_"+metric, 'val_'+metric])
plt.show()
plot_metric(dfhistory,"loss")
plot_metric(dfhistory,"auc")
model.evaluate(dl_val)
{'val_loss': 0.7020544648170471, 'val_auc': 0.6469045281410217}
5，使用模型
from sklearn.metrics import roc_auc_score
labels = torch.tensor([x["label"] for x in ds_val])
preds = model.predict(dl_val)
val_auc = roc_auc_score(labels.cpu().numpy(),preds.cpu().numpy())
print(val_auc)
0.6469045283797497
6，保存模型
torch.save(model.net.state_dict(),"best_dien.pt")
net_clone = create_net()
net_clone.load_state_dict(torch.load("best_dien.pt"))
net_clone.eval()
labels = torch.tensor([x["label"] for x in ds_val])
preds = torch.cat([net_clone(x)[0].data for x in dl_val])
val_auc = roc_auc_score(labels.cpu().numpy(),preds.cpu().numpy())
print(val_auc)
0.6469045283797497
可以看到，相比DIN模型，验证集AUC从0.6392涨到了0.6469. 又是一波基操勿6！
公众号后台回复关键词：DIEN，获取本文全部源码和Movielens数据集下载链接。
以上。
万水千山总是情，点个在看行不行？
