TensorFlow：使用批处理或在线推断CPU时的结果不同

发布于 2025-02-08 17:38:06 字数 885 浏览 2 评论 0原文

我正在对简单的Keras模型进行一些推理测试。

在面临一些奇怪的错误之后，我意识到，有时，当我同时推断多个数据或仅一个数据时，我的结果略有不同。

这是我确定的特定行的一个小例子，向您展示问题：

predicted_all = np.round(self.network.predict(warm_up,         verbose=0).flatten() * raw_data_std + raw_data_mean).astype(np.int32)
predicted_121 = np.round(self.network.predict(warm_up[121, :], verbose=0).flatten() * raw_data_std + raw_data_mean).astype(np.int32)

print('Batch prediction: ',  predicted_all[121])
print('Single prediction: ', predicted_121[0])

基本上，我应该获得相同的结果，因为我推断了harm_up矩阵的相同行。结果：

Batch prediction:  264
Single prediction:  265

我确切地说，这两条线都在同一CPU上运行，并带有选项os.environ [“ cuda_visible_devices”] =“ -1”。所有以前的推论都是相等的。运行测试时，错误似乎随机出现。

我正在使用TensorFlow 2.9.1，Cudnn 8.1，Python 3.9.12在Windows 11上。

此处发生的事情有什么想法吗？

原文

I was running some inference tests on simple Keras model.

After facing some strange errors, I realize that sometimes, my results were slightly different when I was performing inference of multiple data at the same time or just one data.

This is a little example with a specific line I identified, to show you the issue:

predicted_all = np.round(self.network.predict(warm_up,         verbose=0).flatten() * raw_data_std + raw_data_mean).astype(np.int32)
predicted_121 = np.round(self.network.predict(warm_up[121, :], verbose=0).flatten() * raw_data_std + raw_data_mean).astype(np.int32)

print('Batch prediction: ',  predicted_all[121])
print('Single prediction: ', predicted_121[0])

Basically, I should obtain the same results since I'm infering the same lines of my warm_up matrix.
Result :

Batch prediction:  264
Single prediction:  265

I precise that both lines are running on the same CPU, with the option os.environ["CUDA_VISIBLE_DEVICES"] = "-1". All previous inferences were equal. When running my test, the error seems to appear randomly.

I'm using tensorflow 2.9.1, CudNN 8.1, Python 3.9.12 on Windows 11.

Any idea of what's happening here ?

分享到QQ

分享到微博