float32

Question

float32

Closed this issue 3 years ago · 4 comments

Hello, thank you for implementing this great application. However, for your train_denoiser.ipynb, I ran it on my Google Colab and it gave me a mismatched type of float32. If I change all float32 types to float64, it fixes the error. This can be something to keep in mind in the next update of your code.

Answer 1 · 2021-06-10T18:29:45.000Z

Thanks for reporting this. However, I did not find any such errors when I ran it. Could you share a notebook that has the errors?

Answer 2 · 2021-06-10T18:32:51.000Z

https://colab.research.google.com/drive/1dTiQridJDCLdK6D4wIMe0PWm4U-MGaLN?usp=sharing
this happens to the train_classifier also

Answer 3 · 2021-06-11T01:56:30.000Z

Sure. Will check and get back.

Answer 4 · 2021-06-11T02:27:44.000Z

This is likely because of the TensorFlow version mismatch. I used TensorFlow 2.4.1. Besides if you change the dtype of the noise to float32 it should not happen.

noise = tf.experimental.numpy.random.randn(batch_size, 32, 32, 3) * self.sigma
noise = tf.cast(noise, tf.float32)

I have made a note about it in the README.