ALinrunrun opened a new issue, #19558:
URL: https://github.com/apache/tvm/issues/19558
### Expected behavior
TVM Relax should execute ONNX `ArgMax` consistently with ONNX Runtime when
the input contains `NaN`.
For these inputs, ONNX Runtime matches NumPy's `argmax` behavior:
```
x = [2, NaN, 7, 4, 1] -> index 1
x = [NaN, 2, 7, 4, 1] -> index 0
x = [2, 4, 7, 1, NaN] -> index 4
```
TVM Relax should return the same indices after importing the ONNX model
through the Relax ONNX frontend.
### Actual behavior
TVM Relax returns different indices for cases where NaN appears before the
finite maximum:
```
> x=[2, NaN, 7, 4, 1] np=1 ORT=1 TVM=2
> x=[NaN, 2, 7, 4, 1] np=0 ORT=0 TVM=2
> x=[2, 4, 7, 1, NaN] np=4 ORT=4 TVM=4
```
The discrepancy appears when importing an ONNX ArgMax model through the
Relax ONNX frontend and compiling it for the llvm target.
### Environment
TVM: 0.14 environment / Relax ONNX frontend
ONNX Runtime: 1.23
Python: 3.11
Target: llvm
OS: Linux
### Steps to reproduce
```
import warnings
warnings.filterwarnings("ignore")
import numpy as np
import onnx
import onnxruntime as ort
import tvm
from onnx import TensorProto, helper
from tvm import relax
from tvm.relax.frontend.onnx import from_onnx
node = helper.make_node("ArgMax", ["x"], ["y"], axis=0, keepdims=0)
graph = helper.make_graph(
[node],
"g",
[helper.make_tensor_value_info("x", TensorProto.FLOAT, [5])],
[helper.make_tensor_value_info("y", TensorProto.INT64, [])],
)
model = helper.make_model(graph, opset_imports=[helper.make_opsetid("", 18)])
model.ir_version = 9
sess = ort.InferenceSession(
model.SerializeToString(),
providers=["CPUExecutionProvider"],
)
mod = from_onnx(model)
with tvm.transform.PassContext(opt_level=3):
ex = tvm.compile(mod, target=tvm.target.Target("llvm"))
vm = relax.VirtualMachine(ex, tvm.cpu())
def run_both(x):
ort_out = int(sess.run(None, {"x": x})[0])
out = vm["main"](tvm.runtime.tensor(x, tvm.cpu()))
tvm_out = int((out[0] if isinstance(out, (list, tuple)) else
out).numpy())
numpy_out = int(np.argmax(x))
return numpy_out, ort_out, tvm_out
cases = [
np.array([2.0, np.nan, 7.0, 4.0, 1.0], dtype=np.float32),
np.array([np.nan, 2.0, 7.0, 4.0, 1.0], dtype=np.float32),
np.array([2.0, 4.0, 7.0, 1.0, np.nan], dtype=np.float32),
]
for x in cases:
numpy_out, ort_out, tvm_out = run_both(x)
print(f"x={x.tolist()} np={numpy_out} ORT={ort_out} TVM={tvm_out}")
```
### Triage
* needs-triage
--
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]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
