wzh99 opened a new issue, #11687:
URL: https://github.com/apache/tvm/issues/11687
### Expected behavior
`nn.pad` in the following Relay program should output the same result as its
counterpart in NumPy, given input `[[1., 1.]]`:
```
def @main(%x: Tensor[(1, 2), float32]) {
nn.pad(%x, 0, pad_width=[[1, 1], [1, 1]], pad_mode="reflect")
}
```
### Actual behavior
`nn.pad` in TVM outputs different result from `pad` in NumPy:
```
TVM:
[[ 4.5864499e-41 -7.9918004e-28 4.5864499e-41 -7.9918004e-28]
[ 1.0000000e+00 1.0000000e+00 1.0000000e+00 1.0000000e+00]
[ 2.2953549e-40 -7.9916771e-28 2.2953549e-40 -7.9916771e-28]]
NumPy:
[[1. 1. 1. 1.]
[1. 1. 1. 1.]
[1. 1. 1. 1.]]
```
Padded values in the result given by `nn.pad` in TVM seem randomized and
differ between two runs.
### Environment
macOS 12.4. Compiled using Clang 13.1.6 with LLVM support. TVM commit
[`8341e33`](https://github.com/apache/tvm/commit/8341e33d05868b7bb8496c913679b7951836f3b9).
### Steps to reproduce
```python
import numpy as np
from tvm import relay, transform, cpu, IRModule
from tvm.contrib.graph_executor import GraphModule
x_shape = (1, 2)
x_input = np.ones(x_shape, dtype='float32')
y = relay.nn.pad(relay.var('x', shape=x_shape), [[1, 1], [1, 1]],
pad_mode='reflect')
mod = IRModule.from_expr(y)
with transform.PassContext(opt_level=0):
lib = relay.build(mod, target='llvm')
gmod = GraphModule(lib['default'](cpu()))
gmod.run(x=x_input)
print('TVM:\n', gmod.get_output(0))
print('NumPy:\n', np.pad(x_input, [[1, 1], [1, 1]], mode='reflect'))
```
--
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.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
