MNN/docs/pymnn/loss.md

loss

module loss

loss模块是模型训练使用的模块，提供了多个损失函数

---

cross_entropy(predicts, onehot_targets)

求交叉熵损失

参数：

• predicts:Var 输出层的预测值，dtype=float，shape=(batch_size, num_classes)
• onehot_targets:Var onehot编码的标签，dtype=float，shape=(batch_size, num_classes)

返回：交叉熵损失

返回类型：Var

示例

>>> predict = np.random.random([2,3])
>>> onehot = np.array([[1., 0., 0.], [0., 1., 0.]])
>>> nn.loss.cross_entropy(predict, onehot)
array(4.9752955, dtype=float32)

---

kl(predicts, onehot_targets)

求KL损失

参数：

• predicts:Var 输出层的预测值，dtype=float，shape=(batch_size, num_classes)
• onehot_targets:Var onehot编码的标签，dtype=float，shape=(batch_size, num_classes)

返回：KL损失

返回类型：Var

示例

>>> predict = np.random.random([2,3])
>>> onehot = np.array([[1., 0., 0.], [0., 1., 0.]])
>>> nn.loss.kl(predict, onehot)
array(inf, dtype=float32)

---

mse(predicts, onehot_targets)

求MSE损失

参数：

• predicts:Var 输出层的预测值，dtype=float，shape=(batch_size, num_classes)
• onehot_targets:Var onehot编码的标签，dtype=float，shape=(batch_size, num_classes)

返回：MSE损失

返回类型：Var

示例

>>> predict = np.random.random([2,3])
>>> onehot = np.array([[1., 0., 0.], [0., 1., 0.]])
>>> nn.loss.mse(predict, onehot)
array(1.8694793, dtype=float32)

---

mae(predicts, onehot_targets)

求MAE损失

参数：

• predicts:Var 输出层的预测值，dtype=float，shape=(batch_size, num_classes)
• onehot_targets:Var onehot编码的标签，dtype=float，shape=(batch_size, num_classes)

返回：MAE损失

返回类型：Var

示例

>>> predict = np.random.random([2,3])
>>> onehot = np.array([[1., 0., 0.], [0., 1., 0.]])
>>> nn.loss.mae(predict, onehot)
array(2.1805272, dtype=float32)

---

hinge(predicts, onehot_targets)

求Hinge损失

参数：

• predicts:Var 输出层的预测值，dtype=float，shape=(batch_size, num_classes)
• onehot_targets:Var onehot编码的标签，dtype=float，shape=(batch_size, num_classes)

返回：Hinge损失

返回类型：Var

示例

>>> predict = np.random.random([2,3])
>>> onehot = np.array([[1., 0., 0.], [0., 1., 0.]])
>>> nn.loss.hinge(predict, onehot)
array(2.791432, dtype=float32)