Post Review of Multi-Graph Support about nnsmith HOT 2 CLOSED

That's a TODO I haven't yet implemented. Feel free to implement it :-)

from nnsmith.

The first question I have is the dtype_comb parameter in pick_shape_var_idx (grpah_gen.py).

I guess it is trying to allow data type matching which makes sense to me. But I barely understand how it works:

          ishape_indices = self.pick_shape_var_idx(
              dim_spec_list, random.choice(op.in_dtypes))

Why do we need a random sampling here? Is it the case that we first determine the plausible data types and then assume it is the data type of selected nodes? If so, I would suggest first looking at what kind of types we have in the self.alive_shapes and then do sampling in the intersection of op.in_dtypes and types we have in the self.alive_shapes. Because according to my experience we don't have many "float64", "intxx" and "bool" types in self.alive_shape which will quickly run out of the trying budgets (max_shape_var_pick_time) we gave (i.e., 3 times).

from nnsmith.