TensorFlow中损失非零但梯度为0的问题排查求助
TensorFlow损失值非零但梯度为0问题分析
问题描述
遇到TensorFlow中损失值非零但梯度为0的问题,最小复现代码如下:
import tensorflow as tf import numpy as np def setup_mat(): np_data = np.asarray( [ [-1, -2, -3], [-0.1, -0.2, -0.3], [0.1, 0.2, 0.3], [1, 2, 3] ] ) return tf.convert_to_tensor( np_data, dtype_hint=tf.float32) def report_binary_xent(): Xs = setup_mat() weights = tf.Variable([[1.0], [2.0], [3.0]], shape=(3, 1), trainable=True) bce = tf.keras.losses.BinaryCrossentropy(from_logits=False) labels = np.asarray([0.0, 1.0, 0.0, 0.0]) with tf.GradientTape(persistent=True) as g: g.watch([weights]) logits = tf.matmul(Xs, weights) preds = tf.math.softmax(logits) bce_res = bce(labels, preds) dydx = g.gradient(bce_res, [preds, logits, weights]) print(dydx)
认为梯度应该非零的依据:
- weights已标记为可训练(trainable=True)
- 已通过GradientTape监视weights
- 损失值(bce_res)非零
问题原因分析
核心问题是二分类场景错误使用了softmax激活函数,具体逻辑如下:
- 代码中
logits的形状是(4,1),即每个样本对应一个输出值;tf.math.softmax是对张量的最后一维做归一化,单元素维度归一化后结果恒为1.0,因此preds的所有值都是1.0。 - 二元交叉熵的梯度公式为:
(pred - label) / (pred * (1 - pred)),当pred=1.0时,分母1-pred=0,TensorFlow为避免数值异常会将梯度置为0。 - 损失值非零是因为TensorFlow对
log(0)做了数值截断处理(避免直接出现无穷大),但这无法改变梯度为0的结果。
补充:GradientTape默认会自动监视tf.Variable,因此g.watch([weights])属于冗余代码,但不是导致梯度为0的原因。
修正方案
针对二分类场景,有两种正确的处理方式:
方案1:使用sigmoid激活函数 + from_logits=False
将softmax替换为sigmoid,对应二分类场景输出单个样本的正类概率:
import tensorflow as tf import numpy as np def setup_mat(): np_data = np.asarray( [ [-1, -2, -3], [-0.1, -0.2, -0.3], [0.1, 0.2, 0.3], [1, 2, 3] ] ) return tf.convert_to_tensor( np_data, dtype_hint=tf.float32) def report_binary_xent(): Xs = setup_mat() weights = tf.Variable([[1.0], [2.0], [3.0]], shape=(3, 1), trainable=True) bce = tf.keras.losses.BinaryCrossentropy(from_logits=False) labels = np.asarray([0.0, 1.0, 0.0, 0.0]) with tf.GradientTape(persistent=True) as g: logits = tf.matmul(Xs, weights) preds = tf.math.sigmoid(logits) # 替换为sigmoid bce_res = bce(labels, preds) dydx = g.gradient(bce_res, [preds, logits, weights]) print("损失值:", bce_res.numpy()) print("梯度:", [d.numpy() if d is not None else d for d in dydx])
方案2:设置from_logits=True,直接传入logits
让损失函数内部处理数值稳定的sigmoid计算,无需手动添加激活函数:
import tensorflow as tf import numpy as np def setup_mat(): np_data = np.asarray( [ [-1, -2, -3], [-0.1, -0.2, -0.3], [0.1, 0.2, 0.3], [1, 2, 3] ] ) return tf.convert_to_tensor( np_data, dtype_hint=tf.float32) def report_binary_xent(): Xs = setup_mat() weights = tf.Variable([[1.0], [2.0], [3.0]], shape=(3, 1), trainable=True) bce = tf.keras.losses.BinaryCrossentropy(from_logits=True) # 开启from_logits labels = np.asarray([0.0, 1.0, 0.0, 0.0]) with tf.GradientTape(persistent=True) as g: logits = tf.matmul(Xs, weights) bce_res = bce(labels, logits) # 直接传入logits dydx = g.gradient(bce_res, [logits, weights]) print("损失值:", bce_res.numpy()) print("梯度:", [d.numpy() if d is not None else d for d in dydx])
内容的提问来源于stack exchange,提问作者IanQ
相关产品推荐
相关产品推荐

