Pytorch中使用样本权重(sample_weight)的正确方式

step:

1.将标签转换为one-hot形式。
2.将每一个one-hot标签中的1改为预设样本权重的值
即可在Pytorch中使用样本权重。

eg:

对于单个样本：loss = - Q * log(P),如下：

P = [0.1,0.2,0.4,0.3]
Q = [0,0,1,0]
loss = -Q * np.log(P)

增加样本权重则为loss = - Q * log(P) *sample_weight

P = [0.1,0.2,0.4,0.3]
Q = [0,0,sample_weight,0]
loss_samle_weight = -Q * np.log(P)

在pytorch中示例程序

train_data = np.load(open('train_data.npy','rb'))
train_labels = []
for i in range(8):
    train_labels += [i] *100
train_labels = np.array(train_labels)
train_labels = to_categorical(train_labels).astype("float32")
sample_1 = [random.random()  for i in range(len(train_data))]
for i in range(len(train_data)):
    floor = i / 100
    train_labels[i][floor] = sample_1[i]
train_data = torch.from_numpy(train_data)  
train_labels = torch.from_numpy(train_labels) 
dataset = dataf.TensorDataset(train_data,train_labels)  
trainloader = dataf.DataLoader(dataset, batch_size=batch_size, shuffle=True)

对应one-target的多分类交叉熵损失函数如下：

def my_loss(outputs, targets):
    
    output2 = outputs - torch.max(outputs, 1, True)[0]


    P = torch.exp(output2) / torch.sum(torch.exp(output2), 1,True) + 1e-10


    loss = -torch.mean(targets * torch.log(P))


    return loss