【多任务学习】Multi-Task Learning Using Uncertainty to Weigh Losses for Scene Geometry and Semantics

【多任务学习】Multi-Task Learning Using Uncertainty to Weigh Losses for Scene Geometry and Semantics,第1张

·阅读摘要:
  本文提出针对CV领域的多任务模型,设置一个可以学习损失权重的损失层,可以提高模型精度。
·参考文献:
  [1] Multi-Task Learning Using Uncertainty to Weigh Losses for Scene Geometry and Semantics

  个人理解:我们使用传统的多任务时,损失函数一般都是各个任务的损失相加,最多会为每个任务的损失前添加权重系数。但是这样的超参数是很难去调参的,代价大,而且很难去调到一个最好的状态。最好的方式应该是交给深度学习。

  论文最重要的部分在损失函数的设置与推导。 这对我们优化自己的多任务学习模型有指导意义。

[1] Homoscedastic uncertainty as task-dependent uncertainty (同方差不确定性)

  作者的数学模型通过贝叶斯模型建立。作者首先提出贝叶斯建模中存在两类不确定性:

  · 认知不确定性(Epistemic uncertainty):由于缺少训练数据而引起的不确定性
  · 偶然不确定性(Aleatoric uncertainty):由于训练数据无法解释信息而引起的不确定性

  而对于偶然不确定性,又分为如下两个子类:

  · 数据依赖地(Data-dependant)或异方差(Heteroscedastic)不确定性
  · 任务依赖地(Task-dependant)或同方差(Homoscedastic)不确定性

  多任务中,任务不确定性捕获任务间相关置信度,反应回归或分类任务的内在不确定性。

【注】本篇论文的假设,是基于同方差不确定性的。关于同方差不确定性和异方差不确定性的通俗解释,可以参考知乎问题:https://www.zhihu.com/question/278182454/answer/398539763

[2] Multi-task likelihoods (多任务似然)

  基于极大似然估计,假设 f W \mathbf{f}^{\mathbf{W}} fW为网络输出, W W W为该项输出的权重,则对于回归任务有:
p ( y ∣ f W ( x ) ) = N ( f W ( x ) , σ 2 ) p(y∣\mathbf{f}^{\mathbf{W}}(x))=N(\mathbf{f}^{\mathbf{W}}(x),σ^2 ) p(yfW(x))=N(fW(x),σ2)
  对于分类任务有:
p ( y ∣ f W ( x ) ) = S o f t m a x ( f W ( x ) ) p(y∣\mathbf{f}^{\mathbf{W}} (x))=Softmax(\mathbf{f}^{\mathbf{W}} (x)) p(yfW(x))=Softmax(fW(x))
  多任务的概率:
p ( y 1 , … , y K ​ ∣ f W ( x ) ) = p ( y 1 ​ ∣ f W ( x ) ) … p ( y K ​ ∣ f W ( x ) ) p(y_1,…,y_K​ ∣\mathbf{f}^{\mathbf{W}}(x))=p(y_1​ ∣\mathbf{f}^{\mathbf{W}} (x))…p(y_K​ ∣\mathbf{f}^{\mathbf{W}} (x)) p(y1,,yKfW(x))=p(y1fW(x))p(yKfW(x))
  例如对于回归任务来说,极大似然估计转化为最小化负对数:
l o g p ( y ∣ f W ( x ) ) ∝ − 1 2 σ 2 ​ ∥ ∥ ​ y − f W ( x ) ∥ ∥ ​ 2 − l o g σ log{p(y∣\mathbf{f}^{\mathbf{W}}(x))}∝− \frac{1}{2σ^2}​ {∥∥​ y−\mathbf{f}^{\mathbf{W}} (x) ∥∥​}^2 −log\sigma logp(yfW(x))2σ21yfW(x)2logσ
  其中 σ \sigma σ 表示测量噪声的方差。

  双任务:

  假设是两个回归任务,那么概率如下:

   p ( y 1 , y 2 ​ ∣ f W ( x ) ) ​ = p ( y 1 ​ ∣ f W ( x ) ) ⋅ p ( y 2 ​ ∣ f W ( x ) ) = N ( y 1 ​ ; f W ( x ) , σ 1 2 ​ ) ⋅ N ( y 2 ​ ; f W ( x ) , σ 2 2 ​ ) p(y_1 ,y_2​ ∣\mathbf{f}^{\mathbf{W}} (x))​ =p(y_1​ ∣\mathbf{f}^{\mathbf{W}} (x))⋅p(y_2​ ∣\mathbf{f}^{\mathbf{W}} (x))=N(y_1​;\mathbf{f}^{\mathbf{W}} (x),σ_1^2​ )⋅N(y_2​ ;\mathbf{f}^{\mathbf{W}} (x),σ_2^2​ ) p(y1,y2fW(x))=p(y1fW(x))p(y2fW(x))=N(y1;fW(x),σ12)N(y2;fW(x),σ22)

  为了优化我们的 Loss 函数 L ( W , σ 1 , σ 2 ) \mathcal{L}\left(\mathbf{W}, \sigma_{1}, \sigma_{2}\right) L(W,σ1,σ2),取最小化负对数:

   = − l o g p ( y 1 ​ , y 2 ​ ∣ f W ( x ) ) =−logp(y_1​,y_2​ ∣\mathbf{f}^{\mathbf{W}} (x)) =logp(y1,y2fW(x))

   ∝ 1 2 σ 1 2 ​ ∥ y 1 ​ − f W ( x ) ∥ ​ 2 + 1 2 σ 2 2 ​ ∥ ​ y 2 ​ − f W ( x ) ∥ ​ 2 + l o g σ 1 ​ σ 2 ∝ {\frac{1}{2σ_1^2}​ }∥y_1​−\mathbf{f}^{\mathbf{W}} (x) ∥^​2 + \frac{1}{2σ_2^2}​ ∥​ y_2​ −\mathbf{f}^{\mathbf{W}} (x) ∥​^2 +logσ_1​ σ_2 2σ121y1fW(x)2+2σ221y2fW(x)2+logσ1σ2

   = 1 2 σ 1 2 ​ L 1 ​ ( W ) + 1 2 σ 2 2 ​ L 2 ​ ( W ) + l o g σ 1 ​ σ 2 ​ = \frac{1}{2σ_1^2}​ L_1​ (W)+ \frac{1}{2σ_2^2}​ L_2​ (W)+logσ_1​σ_2​ =2σ121L1(W)+2σ221L2(W)+logσ1σ2

  假设是一个回归任务和一个分类任务,仍然可以推导出上述结论:

   = 1 2 σ 1 2 ​ L 1 ​ ( W ) + 1 2 σ 2 2 ​ L 2 ​ ( W ) + l o g σ 1 ​ σ 2 ​ = \frac{1}{2σ_1^2}​ L_1​ (W)+ \frac{1}{2σ_2^2}​ L_2​ (W)+logσ_1​σ_2​ =2σ121L1(W)+2σ221L2(W)+logσ1σ2

  所以,根据整个多任务问题的联合 Loss 形式,那么我们需要优化的参数不仅有 W W W 还有 σ 1 \sigma_{1} σ1 σ 2 \sigma_{2} σ2

pytorch版代码实现

  首先要明确,损失函数中的参数不仅有 W W W 还有 σ 1 \sigma_{1} σ1 σ 2 \sigma_{2} σ2 都是需要经过反向传播来学习的。

  代码如下:


import math

import pylab
import numpy as np

import torch
import torch.nn as nn
from torch.utils.data import Dataset, DataLoader


def gen_data(N):
    X = np.random.randn(N, 1)
    w1 = 2.
    b1 = 8.
    sigma1 = 1e1  # ground truth
    Y1 = X.dot(w1) + b1 + sigma1 * np.random.randn(N, 1)
    w2 = 3
    b2 = 3.
    sigma2 = 1e0  # ground truth
    Y2 = X.dot(w2) + b2 + sigma2 * np.random.randn(N, 1)
    return X, Y1, Y2


class TrainData(Dataset):

    def __init__(self, feature_num, X, Y1, Y2):

        self.feature_num = feature_num

        self.X = torch.tensor(X, dtype=torch.float32)
        self.Y1 = torch.tensor(Y1, dtype=torch.float32)
        self.Y2 = torch.tensor(Y2, dtype=torch.float32)

    def __len__(self):
        return self.feature_num

    def __getitem__(self, idx):
        return self.X[idx,:], self.Y1[idx,:], self.Y2[idx,:]


class MultiTaskLossWrapper(nn.Module):
    def __init__(self, task_num, model):
        super(MultiTaskLossWrapper, self).__init__()
        self.model = model
        self.task_num = task_num
        self.log_vars = nn.Parameter(torch.zeros((task_num)))

    def forward(self, input, targets):

        outputs = self.model(input)

        precision1 = torch.exp(-self.log_vars[0])
        loss = torch.sum(precision1 * (targets[0] - outputs[0]) ** 2. + self.log_vars[0], -1)

        precision2 = torch.exp(-self.log_vars[1])
        loss += torch.sum(precision2 * (targets[1] - outputs[1]) ** 2. + self.log_vars[1], -1)

        loss = torch.mean(loss)

        return loss, self.log_vars.data.tolist()


class MTLModel(torch.nn.Module):
    def __init__(self, n_hidden, n_output):
        super(MTLModel, self).__init__()

        self.net1 = nn.Sequential(nn.Linear(1, n_hidden), nn.ReLU(), nn.Linear(n_hidden, n_output))
        self.net2 = nn.Sequential(nn.Linear(1, n_hidden), nn.ReLU(), nn.Linear(n_hidden, n_output))

    def forward(self, x):
        return [self.net1(x), self.net2(x)]
np.random.seed(0)

feature_num = 100
nb_epoch = 2000
batch_size = 20
hidden_dim = 1024

X, Y1, Y2 = gen_data(feature_num)
pylab.figure(figsize=(3, 1.5))
pylab.scatter(X[:, 0], Y1[:, 0])
pylab.scatter(X[:, 0], Y2[:, 0])
pylab.show()

train_data = TrainData(feature_num, X, Y1, Y2)
train_data_loader = DataLoader(train_data, shuffle=True, batch_size=batch_size)

model = MTLModel(hidden_dim, 1)

mtl = MultiTaskLossWrapper(2, model)
mtl

# https://github.com/keras-team/keras/blob/master/keras/optimizers.py
# k.epsilon() = keras.backend.epsilon()
optimizer = torch.optim.Adam(mtl.parameters(), lr=0.001, eps=1e-07)

loss_list = []
for t in range(nb_epoch):
    cumulative_loss = 0
    
    for X, Y1, Y2 in train_data_loader:

        loss, log_vars = mtl(X, [Y1, Y2])

        optimizer.zero_grad()
        loss.backward()
        optimizer.step()

        cumulative_loss += loss.item()

    loss_list.append(cumulative_loss/batch_size)
pylab.plot(loss_list)
pylab.show()

print(log_vars)
[4.2984442710876465, -0.2037072628736496]
# Found standard deviations (ground truth is 10 and 1):
print([math.exp(log_var) ** 0.5 for log_var in log_vars])
[8.578183137529612, 0.9031617364804738]

【注】假设 L o s s = a ∗ L 1 ( W ) + b ∗ L 2 ( W ) Loss = a * L_1(W) + b * L_2(W) Loss=aL1(W)+bL2(W),我们这样简单的学习参数 a a a b b b是不行的,这样Loss会越来越小,永无止境。我们需要设置 L o s s Loss Loss有正则项,使得它有最小值,这样学习才对。

欢迎分享,转载请注明来源:内存溢出

原文地址: http://outofmemory.cn/langs/919017.html

(0)
打赏 微信扫一扫 微信扫一扫 支付宝扫一扫 支付宝扫一扫
上一篇 2022-05-16
下一篇 2022-05-16

发表评论

登录后才能评论

评论列表(0条)

保存