[JIT] JIT frontend uses len(op_token) before checking for None #25360
Description
🐛 Bug
If you try to script a module with an unsupported unary op (in this case, Invert applied to a tensor like ~X), build_UnaryOp() tries to use len(op_token) before checking if the token is None. Instead of printing the actual error message, it then fails with a ValueError:
...
", line 504, in build_UnaryOp
r = ctx.make_range(expr.lineno, expr.col_offset, expr.col_offset + len(op_token))
TypeError: object of type 'NoneType' has no len()
If I add a check for None and use 0 instead of len(op_token) in that case, I get the real error which says that Invert is not supported :torch.jit.frontend.NotSupportedError: unsupported unary operator: Invert
Followup question: is that expected? If I replace ~ with th.bitwise_not() then it works just fine.
To Reproduce
import torch as th
class TestMod(th.nn.Module):
def __init__(self):
super().__init__()
self.register_buffer("bool_tensor", th.zeros(3,).bool())
def forward(self, x):
return x[~self.bool_tensor]
mod = TestMod()
smod = th.jit.script(mod)Run the above script.
Traceback (most recent call last):
File "test.py", line 12, in <module>
smod = th.jit.script(mod)
File "/mnt/home/gbschwartz/anaconda/envs/py3_newpytorch_cuda10/lib/python3.7/site-packages/torch/jit/__init__.py", line 1161, in script
return torch.jit.torch.jit._recursive.recursive_script(obj)
File "/mnt/home/gbschwartz/anaconda/envs/py3_newpytorch_cuda10/lib/python3.7/site-packages/torch/jit/_recursive.py", line 133, in recursive_script
stubs = list(map(make_stub, methods))
File "/mnt/home/gbschwartz/anaconda/envs/py3_newpytorch_cuda10/lib/python3.7/site-packages/torch/jit/_recursive.py", line 131, in make_stub
return torch.jit.script_method(func, _jit_internal.createResolutionCallbackFromClosure(func))
File "/mnt/home/gbschwartz/anaconda/envs/py3_newpytorch_cuda10/lib/python3.7/site-packages/torch/jit/__init__.py", line 1226, in script_method
ast = get_jit_def(fn, self_name="ScriptModule")
File "/mnt/home/gbschwartz/anaconda/envs/py3_newpytorch_cuda10/lib/python3.7/site-packages/torch/jit/frontend.py", line 166, in get_jit_def
return build_def(ctx, py_ast.body[0], type_line, self_name)
File "/mnt/home/gbschwartz/anaconda/envs/py3_newpytorch_cuda10/lib/python3.7/site-packages/torch/jit/frontend.py", line 206, in build_def
build_stmts(ctx, body))
File "/mnt/home/gbschwartz/anaconda/envs/py3_newpytorch_cuda10/lib/python3.7/site-packages/torch/jit/frontend.py", line 122, in build_stmts
stmts = [build_stmt(ctx, s) for s in stmts]
File "/mnt/home/gbschwartz/anaconda/envs/py3_newpytorch_cuda10/lib/python3.7/site-packages/torch/jit/frontend.py", line 122, in <listcomp>
stmts = [build_stmt(ctx, s) for s in stmts]
File "/mnt/home/gbschwartz/anaconda/envs/py3_newpytorch_cuda10/lib/python3.7/site-packages/torch/jit/frontend.py", line 182, in __call__
return method(ctx, node)
File "/mnt/home/gbschwartz/anaconda/envs/py3_newpytorch_cuda10/lib/python3.7/site-packages/torch/jit/frontend.py", line 298, in build_Return
return Return(r, None if stmt.value is None else build_expr(ctx, stmt.value))
File "/mnt/home/gbschwartz/anaconda/envs/py3_newpytorch_cuda10/lib/python3.7/site-packages/torch/jit/frontend.py", line 182, in __call__
return method(ctx, node)
File "/mnt/home/gbschwartz/anaconda/envs/py3_newpytorch_cuda10/lib/python3.7/site-packages/torch/jit/frontend.py", line 596, in build_Subscript
return Subscript(base, [build_expr(ctx, expr.slice.value)])
File "/mnt/home/gbschwartz/anaconda/envs/py3_newpytorch_cuda10/lib/python3.7/site-packages/torch/jit/frontend.py", line 182, in __call__
return method(ctx, node)
File "/mnt/home/gbschwartz/anaconda/envs/py3_newpytorch_cuda10/lib/python3.7/site-packages/torch/jit/frontend.py", line 503, in build_UnaryOp
r = ctx.make_range(expr.lineno, expr.col_offset, expr.col_offset + len(op_token))
TypeError: object of type 'NoneType' has no len()
Expected behavior
", line 507, in build_UnaryOp
raise NotSupportedError(err_range, "unsupported unary operator: " + op.__name__)
torch.jit.frontend.NotSupportedError: unsupported unary operator: Invert
:
at test.py:9:18
def forward(self, x):
return x[~self.bool_tensor]
~~~~~~~~~~ <--- HERE
Is what I get if I do the hack I described above, this seems correct.
Environment
- PyTorch Version (e.g., 1.0): '1.3.0.dev20190816'
- OS (e.g., Linux): linux
- How you installed PyTorch (
conda,pip, source): conda (nightly) - Python version: 3.7.1
- CUDA/cuDNN version: 10 / 7.6
- GPU models and configuration: Titan RTX
cc @suo