[tmva][sofie] ParseBatchNormalization accepts invalid input count and fails late #21759
Description
Check duplicate issues.
- Checked for duplicates
Description
BatchNormalization parser currently accepts invalid input count and fails later during code generation.
I was going through tmva/sofie_parsers/src/ParseBatchNormalization.cxx and the code only constructs the operator when nodeproto.input_size() == 5, but does not throw otherwise:
case ETensorType::FLOAT:
if (nodeproto.input_size() == 5) {
op.reset(new ROperator_BatchNormalization<float>(...));
}
break;
This means that nvalid ONNX nodes(e.g. 4 inputs) pass Parse() and fail later in Generate() with an unrelated/less clear error. Expected behavior should have been that Parser rejects immediately with a clear message (BatchNormalization requires exactly 5 inputs).
Reproducer
On running the following script-
import os, tempfile
import onnx
from onnx import helper, TensorProto
import ROOT
X = helper.make_tensor_value_info('X', TensorProto.FLOAT, [1,1,1,1])
S = helper.make_tensor_value_info('S', TensorProto.FLOAT, [1])
B = helper.make_tensor_value_info('B', TensorProto.FLOAT, [1])
M = helper.make_tensor_value_info('M', TensorProto.FLOAT, [1])
Y = helper.make_tensor_value_info('Y', TensorProto.FLOAT, [1,1,1,1])
# Invalid BN: only 4 inputs
node = helper.make_node('BatchNormalization', ['X','S','B','M'], ['Y'])
graph = helper.make_graph([node], 'bn_bad_inputs', [X,S,B,M], [Y])
model = helper.make_model(graph, opset_imports=[helper.make_opsetid('', 13)])
model.ir_version = 10
fd, p = tempfile.mkstemp(suffix='.onnx'); os.close(fd)
onnx.save(model, p)
ROOT.gSystem.Load('libROOTTMVASofieParser')
parser = ROOT.TMVA.Experimental.SOFIE.RModelParser_ONNX()
m = parser.Parse(p) # currently succeeds (unexpected)
m.Generate() # fails laterObserved behavior:
- SOFIE: Parse succeeded, Generate failed later
- ONNX Runtime: input size 4 not in range [5,5]
ROOT version
6.39.01 (built from source)
Installation method
Built from source with SOFIE enabled.
Operating system
Linux (Ubuntu 22.04)
Additional context
No response