Rounding errors when using set_weight

[Koehlibert]

Hi,
I'm working on a project in which I have to manually set weights for some layers of my network. However when doing so, the resulting weights are changed (not even rounded) at around 8 digits, as seen in the picture.
This breaks my algorithm and I am curious if someone knows how this can happen or how this can be prevented. Since this is part of a very large project, I'm afraid that giving a "minimal" working example would need several thousands of lines, so I will refrain from doing so. Maybe someone had a similar problem and has some insights.

[t-kalinowski]

You're losing precision by converting from float64 to float32. Keras layer weights default to float32 typically. You can customize that by passing dtype = 'float64' when creating the layer.

[t-kalinowski]

You can also try setting a global default with keras::k_set_floatx("float64"), but be aware that it doesn't always propagate to all the tensorflow operations, so you may still have to search out stray 'float32' conversions in a large codebase.