历史词典的损失与训练期间在屏幕上印刷的内容不同

发布于 2025-02-06 07:57:53 字数 917 浏览 1 评论 0原文

系统信息 使用Tensorflow 2.8.0

描述当前行为 在调用“历史= model.fit（...）”之后，保存在历史上的值['损失']和历史['准确性']与训练期间在屏幕上打印的值不同。

我有一个自定义培训循环，其中train_step（）函数返回2个损失：d_loss和g_loss。当我调用model.fit（） -

Epoch 1/2
110/110 [==============================] - 33s 294ms/step - d_loss: -1.5941 - g_loss: 2.2890
Epoch 2/2
110/110 [==============================] - 32s 288ms/step - d_loss: -2.5873 - g_loss: 2.8013

print([round(x, 4) for x in history_init.history['d_loss']])
print([round(x, 4) for x in history_init.history['g_loss']])

输出： D_LOSS：[-2.7844，-2.2764] G_LOSS：[2.4266，3.3268]

描述预期行为 值应该与先前的TF版本和其他参数相同。

先前已经对此进行了讨论，并提出了一个错误，但似乎从TF2.4.1回来了： https://github.com/tensorflow/tensorflow/issues/48408

原文

System information
Using TensorFlow 2.8.0

Describe the current behavior
After calling "history = model.fit(...)", values saved in history.history['loss'] and history.history['accuracy'] are different from the ones printed on-screen at every epoch during training.

I have a custom training loop with the train_step() function returning 2 losses: d_loss and g_loss.
The values turn out to be different when I call model.fit() -

Epoch 1/2
110/110 [==============================] - 33s 294ms/step - d_loss: -1.5941 - g_loss: 2.2890
Epoch 2/2
110/110 [==============================] - 32s 288ms/step - d_loss: -2.5873 - g_loss: 2.8013

print([round(x, 4) for x in history_init.history['d_loss']])
print([round(x, 4) for x in history_init.history['g_loss']])

OUTPUT:
d_loss: [-2.7844, -2.2764]
g_loss: [2.4266, 3.3268]

Describe the expected behavior
Values should be the same, as is with previous tf versions and with other parameters.

There has been a previous discussion on this and a bug raised, but it seems to be back from tf2.4.1: https://github.com/tensorflow/tensorflow/issues/48408

分享到QQ

分享到微博