多分类损失函数
- label.shape:
[batch_size]; pred.shape:[batch_size, num_classes]
使用 tf.keras.losses.sparse_categorical_crossentropy(y_true, y_pred, from_logits=False, axis=-1)
- y_true 真实值， y_pred 预测值
- from_logits,我的理解是，如果预测结果经过了softmax(单次预测结果满足和为1)就使用设为`False`,
如果预测结果未经过softmax就设为`True`.pred = tf.convert_to_tensor([[0.9, 0.05, 0.05], [0.5, 0.89, 0.6], [2.05, 0.01, 0.94]])
label = tf.convert_to_tensor([0, 1, 2])
loss = tf.keras.losses.sparse_categorical_crossentropy(label, pred)
print(loss.numpy())
# 包含 reduction 参数, 用于对一个批次的损失函数求平均值,求和等
# loss = tf.keras.losses.SparseCategoricalCrossentropy()(label, pred)- label.shape:
[batch_size, num_classes](one_hot);pred.shape:[batch_size, num_classes]
使用 tf.keras.losses.categorical_crossentropy(y_true, y_pred, from_logits=False, axis=-1)
- y_true 真实值， y_pred 预测值
- from_logits 同上pred = tf.convert_to_tensor([[0.9, 0.05, 0.05], [0.5, 0.89, 0.6], [0.05, 0.01, 0.94]])
label = tf.convert_to_tensor([[1, 0, 0], [0, 1, 0], [0, 0, 1]])
loss = tf.keras.losses.categorical_crossentropy(label, pred)
print(loss.numpy())二分类损失损失函数
label = tf.convert_to_tensor([0, 0, 1, 1], dtype=tf.float32)
pred = tf.convert_to_tensor([1, 1, 1, 0], dtype=tf.float32)
loss = tf.keras.losses.BinaryCrossentropy()(label, pred)
print(loss.numpy())多分类与二分类
通常 categorical_crossentropy与 softmax激活函数搭配使用; binary_crossentropy 与 sigmoid搭配使用;