[JIT] Cannot index into tuple with non constant integer

[sidazhang]

🐛 Bug

[JIT] Cannot index into tuple with non constant integer. Constant array turns it the array into tuple

To Reproduce

 import torch
 
 
 class Normalize(torch.jit.ScriptModule):
     __constants__ = ['mean', 'std']
 
     def __init__(self):
         super(Normalize, self).__init__()
         self.mean = [1, 1, 1]
         self.std = [1, 1, 1]
 
     @torch.jit.script_method
     def forward(self, image):
         # type: (torch.Tensor) -> torch.Tensor
         for i in range(3):
             image[:, i] = image[:, i].sub_(self.mean[i]).div_(self.   std[i])
 
         return image
 
 
 Normalize()

tuple indices must be integer constants:
@torch.jit.script_method
def forward(self, image):
    # type: (torch.Tensor) -> torch.Tensor
    for i in range(3):
        image[:, i] = image[:, i].sub_(self.mean[i]).div_(self.std[i])
                                       ~~~~~~~~~~~ <--- HERE

    return image

[driazati]

You can workaround it by hijacking our Tuple -> List argument coercion by passing self.mean through something like:

@torch.jit.script
def make_list(x):
    # type: (Tuple[int, int, int]) -> List[int]
    return x

We should add support for list(a_tuple) though which should be used in place of the hack above, so leaving this open to track that

[sidazhang]

@driazati hahah this is actually kind of brilliant. Thank you! :P

[eellison]

This would only work with a Tuple of size 3; to hijack Tuple -> List coercion for any size make the input type of make_list a List[int].

@torch.jit.script
def make_list(x):
    # type: (List[int]) -> List[int]
    return x

But as David said, we should support list(a_tuple) / other better solution...

[sidazhang]

This would only work with a Tuple of size 3; to hijack Tuple -> List coercion for any size make the input type of make_list a List[int].

@torch.jit.script
def make_list(x):
    # type: (List[int]) -> List[int]
    return x

But as David said, we should support list(a_tuple) / other better solution...

lol Even more brilliant.

[eellison]

Closed with #20081