repository·main·Indexed 21 days ago
https://github.com/pinto0309/onnx2tfA conversion tool for transforming ONNX files into LiteRT, TFLite, TensorFlow, and PyTorch formats. It supports generating PyTorch native code (nn.Module), TorchScript (.pt), state_dict (.pt), Exported Program (.pt2), and Dynamo ONNX, as well as direct conversion from LiteRT to PyTorch. The tool features a high-performance 'flatbuffer_direct' backend for fast conversion and a 'tf_converter' for a wider range of ONNX operator support.
onnx2tf is a conversion tool designed to transform ONNX files into various formats and codebases, including:
nn.Module).pt).pt).pt2)It also supports direct conversion from LiteRT to PyTorch.
Note: For certain use cases, the maintainers recommend using LiteRT Torch or AI Edge Quantizer instead of onnx2tf.
The refactoring of the flatbuffer_direct backend in version 2.6.6 maintains full backward compatibility for end-users.
flatbuffer_direct remains the default backend.-cotof flag do not require or import TensorFlow.The fb-refactor7 update focuses on improving diagnostic bookkeeping efficiency, stabilizing reconciliation and convergence fast paths, and ensuring explicit mutation evidence and result propagation. Key improvements include:
duplicate-fanout/quantized-PReLU and boundary-BatchMatMul/input-unary) now propagate nested evidence instead of None slots, preserving parent arity and child order.run_indexed_prune_reconcile_cleanup() owner shares a single ModelIRGraphIndex across dead-op pruning and static-shape reconciliation boundaries.ModelIRGraphIndex and using indexed candidate/mutation APIs.The flatbuffer-direct execution path in onnx2tf has undergone a significant refactor to improve terminal shape handling and binary layout adaptation. This refactor introduces a series of specialized optimization passes designed to handle complex operator chains (like Concat, Mul, Add, Reshape, Transpose, and Attention) while maintaining strict topology, rank-four metadata, and QDIM (Quantized Dimension) remapping.
Key improvements across the new passes include:
ModelIRGraphIndex and remove redundant nodes safely.The attention_gather_cleanup_layout pass (located in passes/attention_gather_cleanup_layout.py) optimizes the attention tail by rewriting Gather/Transpose/Reshape sequences.
Key behaviors and constraints:
Transpose operations and preserves dynamic axes by remapping per-axis QDIM.Reshape boundaries.INT32 TensorIR/buffer contract.The ModelIRGraphIndex is a central abstraction used to avoid quadratic complexity during graph transformations. Instead of rescanning the operator list for every boundary or consumer, the index provides:
ModelIRGraphIndex.has_consumer_at_or_after(), which only needs to check the last index rather than iterating through a generator.boundary_count * edge_count.The project uses GraphIndex (for ONNX) and ModelIRGraphIndex (for ModelIR) to provide differential mutation contracts. This allows rewriters to update the graph without rescanning the entire operator list.
producer, consumer, duplicate-producer, operator-position, and operator-type indices.The flatbuffer_direct architecture is a high-performance execution path designed to optimize ONNX to TensorFlow conversion by minimizing redundant graph scans and deep copies. It relies on several key architectural principles:
ModelIRGraphIndex to perform producer/consumer discovery and reachability analysis in a single pass, rather than rescanning the entire operator list for every operation._ModelIRRewriteBuilder that emits operators directly into a new stream, avoiding the overhead of cloning the entire source graph first.WHILE expansion, the system uses a 'preflight' check. If no changes are needed, it returns a borrowed reference to the input. If expansion is required, it performs a deep clone before mutation.Following the optional late-binary layout recovery, the process executes a composite cleanup for InstanceNorm layout repairs. This involves three specific passes executed in a fixed order using the same ModelIR and conversion-local LayoutState:
This composite operation is managed by run_pre_terminal_instancenorm_layout_cleanup(shared_model_ir_pass_context). It is designed to perform these repairs without summarizing counters or recording phase evidence into the main phase store.
The Conv1D-shim unary fan-out bypass (in passes/conv1d_unary_layout.py) handles cases where the rank-three unary output is used by both a standard NHWC reconstruction branch AND a side branch (e.g., a public graph output or a later NCHW consumer).
Instead of leaving the legacy helper's non-topological order, the pass reorders the operators to ensure a topological flow:
This ensures that the NCHW side consumer receives the correct values while maintaining the original unary object's identity and provenance.
The quantized-PReLU cleanup process follows these rules:
_layout_pass_set_1_quantized_prelu_stats.The recurrent orphan-step alias repair process (managed by passes/recurrent_alias.py::repair_orphan_recurrent_step_tensors) fixes tensor naming issues in recurrent models.
It works by:
*_h_step_N/*_c_step_N grammar.ModelIRGraphIndex to find the first valid Reshape among indexed consumers of the corresponding *_step_shape_N tensor.This process is shared by both direct lowering and PyTorch export paths.