Use numpy in script class (torch.jit.script)

jit
1

I was wondering if I can use numpy APIs in a function which is going to be scripted by torch.jit.script. I have this simple function which does not work:

import torch
import torch.nn as nn

class MyModule(nn.Module):
def init(self):
super(MyModule, self).init()

@torch.jit.ignore
def call_np():
    return torch.jit.export(np.random.choice(2, p=[.95,.05]))

def forward(self):
    pass 

@torch.jit.export
def func(self):
    done = self.call_np()
    print (done)

scripted_module = torch.jit.script(MyModule())
scripted_module.func()

which results in:

---------------------------------------------------------------------------
RuntimeError                              Traceback (most recent call last)
<ipython-input-133-ab1ce37d6edc> in <module>()
     18         print (done)
     19 
---> 20 scripted_module = torch.jit.script(MyModule())
     21 scripted_module.func()

C:\ProgramData\Anaconda3\lib\site-packages\torch\jit\__init__.py in script(obj, optimize, _frames_up, _rcb)
   1201 
   1202     if isinstance(obj, torch.nn.Module):
-> 1203         return torch.jit.torch.jit._recursive.recursive_script(obj)
   1204 
   1205     qualified_name = _qualified_name(obj)

C:\ProgramData\Anaconda3\lib\site-packages\torch\jit\_recursive.py in recursive_script(mod, exclude_methods)
    171     filtered_methods = filter(ignore_overloaded, methods)
    172     stubs = list(map(make_stub, filtered_methods))
--> 173     return copy_to_script_module(mod, overload_stubs + stubs)
    174 
    175 

C:\ProgramData\Anaconda3\lib\site-packages\torch\jit\_recursive.py in copy_to_script_module(original, stubs)
     93             setattr(script_module, name, item)
     94 
---> 95     torch.jit._create_methods_from_stubs(script_module, stubs)
     96 
     97     # Now that methods have been compiled, take methods that have been compiled

C:\ProgramData\Anaconda3\lib\site-packages\torch\jit\__init__.py in _create_methods_from_stubs(self, stubs)
   1421     rcbs = [m.resolution_callback for m in stubs]
   1422     defaults = [get_default_args(m.original_method) for m in stubs]
-> 1423     self._c._create_methods(self, defs, rcbs, defaults)
   1424 
   1425 # For each user-defined class that subclasses ScriptModule this meta-class,

RuntimeError: Unable to cast Python instance of type <class 'int'> to C++ type 'unsigned __int64'

I appreciate any help or comment.

2

TorchScript only supports PyTorch and the math module, so numpy functions won’t work natively and can’t be exported. You can use torch.jit.ignore as you have done to leave a call to the Python interpreter. Modifying your example slightly and running with the latest version of PyTorch:

@torch.jit.ignore
def call_np() -> int:
    return np.random.choice(2, p=[.95,.05])

class MyModule(nn.Module):
    def forward(self):
        pass

    @torch.jit.export
    def func(self):
        done = call_np()
        print (done)

scripted_module = torch.jit.script(MyModule())
print(scripted_module.func.graph)
scripted_module.func()

prints

graph(%self : __torch__.MyModule):
  %3 : None = prim::Constant() # ../test.py:184:4
  %done.1 : int = ^call_np()() # ../test.py:185:15
   = prim::Print(%done.1) # ../test.py:186:8
  return (%3)

0

In which we can see ^call_np() which is an upcall to Python. You must also type annotate call_np since all arguments / returns are assumed to be Tensors unless otherwise specified. torch.jit.export is also intended to be used only as a decorator on a function (see here for details).

I couldn’t repro your exact error, could you file an issue on GitHub with a full repro to produce your issue? Even if what’s happening should throw an error the message could be made more clear. I also ran into this issue while trying to repro it which seems like another bug.

1 Like
3

Thanks for the answer. I’ll fine the the bug on github page.