Clean_name() erases dots instead of replacing them, causing tensor name collisions

[AdityaDRathore]

Description

UTILITY::Clean_name() in SOFIE_common.cxx erases dot characters (.) from tensor names instead of replacing them with a safe character (e.g., _). This causes tensor name collisions when parsing models whose intermediate tensor names contain dots.

The Collision

ONNX graphs exported from frameworks like TorchScript generate intermediate tensor names using a sequential dotted pattern:

input.1 → input0 → input1 → input2 → ...

When Clean_name() processes these:

Original Name	After Clean_name()	Intended Role
input.1	input1	Model input tensor
input1	input1	Intermediate tensor (output of first op)

Both map to the same C++ variable name input1, causing the generated inference code to contain duplicate declarations or silently overwrite tensor data.

Root Cause

std::string UTILITY::Clean_name(std::string input_tensor_name){
   std::string s (input_tensor_name);
   std::replace( s.begin(), s.end(), '-', '_');
   s.erase(std::remove_if(s.begin(), s.end(), 
       []( char const& c ) -> bool { return !std::isalnum(c) && c != '_'; } ), s.end());
   return s;
}

The function replaces - → _ explicitly, but all other non-alphanumeric characters (including .) are erased rather than replaced. This is inconsistent: - gets a safe replacement while . gets deleted.

Expected behavior

input.1 → input_1 (distinct from input1)

Current behavior

input.1 → input1 (collides with input1)

Fix

Replace dots with underscores before erasing, consistent with the existing - → _ treatment:

std::string UTILITY::Clean_name(std::string input_tensor_name){
   std::string s (input_tensor_name);
   std::replace( s.begin(), s.end(), '-', '_');
   std::replace( s.begin(), s.end(), '.', '_');  // <-- add this line
   s.erase(std::remove_if(s.begin(), s.end(), 
       []( char const& c ) -> bool { return !std::isalnum(c) && c != '_'; } ), s.end());
   return s;
}

This makes input.1 → input_1 (distinct from input1), eliminating the collision.

Related

• Reported in [tmva][sofie] Clean_name() erases dots causing tensor name collisions in PyTorch parser root-project/root#21527
• Fix PR: Fix Clean_name() to replace dots with underscores instead of erasing #10