CPGGraph.to_torch() generated class does not return correct shape for constant functions about hal-cgp HOT 4 CLOSED

I think that the constant float / torch.nn.Parameter needs to be unsqueezed and then the 0-th dimension needs to be blown up to the batch dimension. Something like this:
self._output_str = 'torch.cat(x.shape[0] * [self._p{}.unsqueeze(0)])'.format(self._idx)

from hal-cgp.

i'm not even sure what the CGPConstantFloat.format_output_str_torch method does, is it even used? it does not make use of the actual value of the node (self._output) but only of the node index to define a new member variable. maybe that function should never be called in the first place (add raise ...?), but only the function format_parameter_str that should actually be called format_parameter_str_torch? in this function, one could just resize the tensor to include a batch dimension, correct?

from hal-cgp.

As far as I understand, the CGPConstantFloat.format_output_str_torch is called in CGPGraph.compile_torch_class() when the forward method of the torch class is written:
https://github.com/jakobj/python-gp/blob/3bdff24be10ef503652b44b0be1606a639555627/gp/cgp_graph.py#L218

This function calls format_output_str_torch is called on the output nodes here which call this method on their input nodes leading to a recursive call.

On the other hand, the format_parameter_str method is used to construct the __init__() method of the torch class:
https://github.com/jakobj/python-gp/blob/3bdff24be10ef503652b44b0be1606a639555627/gp/cgp_graph.py#L200

This makes sense because in the __init__() method, the constant float has to be defined (for example set to 1.0) and then in the forward method it is used.

from hal-cgp.

ah i see, thanks for explaining! who the hell wrote that code? ;)

from hal-cgp.