TensorWaves currently enables TensorFlow's NumPy behavior with its default 64-bit inference and hard-codes tf.float64 in TFUniformRealNumberGenerator. Users should be able to select 32-bit precision consistently, particularly for autodiff workloads.
Centralize TensorFlow initialization and make the precision configurable, with the following behavior:
- TensorWaves defaults to 64-bit precision when nothing is specified.
- Users can explicitly select 32-bit precision through the TensorWaves configuration API.
- The setting controls both TensorFlow NumPy dtype inference and
TFUniformRealNumberGenerator output.
- Configuration is applied before TensorFlow arrays or compiled functions are created.
- Tests cover both precision modes, and the behavior is documented.
TensorWaves currently enables TensorFlow's NumPy behavior with its default 64-bit inference and hard-codes
tf.float64inTFUniformRealNumberGenerator. Users should be able to select 32-bit precision consistently, particularly for autodiff workloads.Centralize TensorFlow initialization and make the precision configurable, with the following behavior:
TFUniformRealNumberGeneratoroutput.