secsilm / understaing-datasets-estimators-tfrecords Goto Github PK

how to update lr with training epochs in tf.estimator?

请问estimator的模型训练好之后，怎么使用呢，能否提供一个demo作为参考，谢谢

Hi,
I have a question about the shape from iterator.get_next().
I convert an image to TFRecord, when I get the image through iterator.get_next(), I find that the shape of the image is [1, (an big number)].
And I find that in your file 'save-and-read-cifar10-tfrecords.ipynb' has the same question.
Is that ok to train the network or need other operation?

I got the same results with you:
Out[49]: array([ 59, 43, 50, ..., 140, 84, 72], dtype=uint8)
Out[61]: array([ 159., 150., 153., ..., 14., 17., 19.], dtype=float32)

but I do not understand：
they are both the first image of cifar10, and dataset is not shuffled,
why they are different?