j-paulus opened a new issue #6414:
URL: https://github.com/apache/incubator-tvm/issues/6414
When a torch model contains strided slicing of an N-D array, the stride
argument is ignored and non-strided version is returned. Because the striding
is ignored, the size and the contents of the result are wrong.
Pseudocode example:
```
x = Tensor[1,4]
y = x[:, 0::2] # => shape:(1, 4), should be (1, 2)
```
A small script triggering the bug:
```
import torch
from tvm import relay
class TriggerBug(torch.nn.Module):
def __init__(self):
super(TriggerBug, self).__init__()
def forward(self, x):
return x[..., 0::2] + x[..., 1::2]
#x_in = torch.randn(4) # this would work
x_in = torch.randn(1, 4) # this doesn't work
torch_model = TriggerBug()
traced_model = torch.jit.trace(torch_model, (x_in,))
mod, params = relay.frontend.from_pytorch(traced_model, [('x_in',
x_in.shape)])
```
The output is:
> mod, params = relay.frontend.from_pytorch(traced_model, [('x_in',
x_in.shape)])
> File
"/Users/name/opt/anaconda3/envs/env/lib/python3.8/site-packages/tvm-0.7.dev1-py3.8-macosx-10.9-x86_64.egg/tvm/relay/frontend/pytorch.py",
line 2788, in from_pytorch
> mod["main"] = tvm.relay.Function(_analysis.free_vars(ret[0]), ret[0])
> File
"/Users/name/opt/anaconda3/envs/env/lib/python3.8/site-packages/tvm-0.7.dev1-py3.8-macosx-10.9-x86_64.egg/tvm/ir/module.py",
line 74, in __setitem__
> return self._add(var, val)
> File
"/Users/name/opt/anaconda3/envs/env/lib/python3.8/site-packages/tvm-0.7.dev1-py3.8-macosx-10.9-x86_64.egg/tvm/ir/module.py",
line 83, in _add
> _ffi_api.Module_Add(self, var, val, update)
> File
"/Users/name/opt/anaconda3/envs/env/lib/python3.8/site-packages/tvm-0.7.dev1-py3.8-macosx-10.9-x86_64.egg/tvm/_ffi/_ctypes/packed_func.py",
line 225, in __call__
> raise get_last_ffi_error()
> tvm._ffi.base.TVMError: Traceback (most recent call last):
> [bt] (8) 9 libffi.7.dylib 0x000000010ebd7ead
ffi_call_unix64 + 85
> [bt] (7) 8 libtvm.dylib 0x000000012801c3c8
TVMFuncCall + 72
> [bt] (6) 7 libtvm.dylib 0x000000012754260c
std::__1::__function::__func<tvm::$_3, std::__1::allocator<tvm::$_3>, void
(tvm::runtime::TVMArgs,
tvm::runtime::TVMRetValue*)>::operator()(tvm::runtime::TVMArgs&&,
tvm::runtime::TVMRetValue*&&) + 492
> [bt] (5) 6 libtvm.dylib 0x0000000127536437
tvm::IRModuleNode::Add(tvm::GlobalVar const&, tvm::BaseFunc const&, bool) + 183
> [bt] (4) 5 libtvm.dylib 0x0000000127535edf
tvm::RunTypeCheck(tvm::IRModule const&, tvm::GlobalVar const&,
tvm::relay::Function) + 1103
> [bt] (3) 4 libtvm.dylib 0x0000000127e773e0
tvm::relay::InferType(tvm::relay::Function const&, tvm::IRModule const&,
tvm::GlobalVar const&) + 544
> [bt] (2) 3 libtvm.dylib 0x0000000127e76587
tvm::relay::TypeInferencer::Infer(tvm::RelayExpr) + 119
> [bt] (1) 2 libtvm.dylib 0x000000012752714c
tvm::ErrorReporter::RenderErrors(tvm::IRModule const&, bool) + 5308
> [bt] (0) 1 libtvm.dylib 0x000000012735447f
dmlc::LogMessageFatal::~LogMessageFatal() + 111
> File "/Users/puu/code/python/tvm_fix/tvm/src/ir/error.cc", line 132
> TVMError:
> Error(s) have occurred. The program has been annotated with them:
>
> In `main`:
> #[version = "0.0.5"]
> fn (%x_in: Tensor[(1, 4), float32]) {
> %0 = strided_slice(%x_in, meta[relay.Constant][0],
meta[relay.Constant][1], meta[relay.Constant][2], begin=[0, 0], end=[1, 4],
strides=[2]);
> %1 = strided_slice(%x_in, meta[relay.Constant][3],
meta[relay.Constant][4], meta[relay.Constant][5], begin=[0, 1], end=[1, 4],
strides=[2]);
> add(%0, %1) Incompatible broadcast type TensorType([1, 4], float32) and
TensorType([1, 3], float32);
> }
This suggests that the slicing ignores the stride argument.
If the tensor to be sliced is 1D, the result is correct, but even a 2D fails.
#6316 seems somehow related, but unfortunately it doesn't fix this issue.
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
[email protected]
