We've found out when using FNO that for examples like Burgers (03) the model learns better if the weights of FourierLayer are initialized with Lux.glorot_uniform.
However, for examples 02.04 the a posteriori fitting fails unless the weights are intialized as Lux.zeros32.
This was an option for FourierLayer but not for create_fno_model. We've added this argument with default to Lux.glorot_uniform.
We've tested the examples using FNO and adapted the syntax when necessary. For example, in the latest implementation we need to add an extra element in the definitions of ch_fno. This was done for 02.04 and 03.01 in this PR.
We've found out when using FNO that for examples like Burgers (03) the model learns better if the weights of
FourierLayerare initialized withLux.glorot_uniform. However, for examples 02.04 the a posteriori fitting fails unless the weights are intialized asLux.zeros32.This was an option for
FourierLayerbut not forcreate_fno_model. We've added this argument with default toLux.glorot_uniform.We've tested the examples using FNO and adapted the syntax when necessary. For example, in the latest implementation we need to add an extra element in the definitions of
ch_fno. This was done for 02.04 and 03.01 in this PR.