2.0.12

Summary

• Extend flatbuffer_direct lowering/optimization pipeline for QAT INT8 ONNX models, focusing on reducing redundant transpose and quantize/dequantize bridges.

• Add ONNX frontend support for QLinearGlobalAveragePool and implement a dedicated flatbuffer-direct lowering path with quantized AVERAGE_POOL_2D preference (fallback to DEQUANTIZE -> MEAN -> QUANTIZE when constraints are not met).

• Improve NHWC propagation and graph sanitization passes to avoid unnecessary transpose insertion and remove dead tensors/operators safely.

• https://huggingface.co/opencv/face_detection_yunet/blob/main/face_detection_yunet_2023mar_int8.onnx

  onnx2tf -i face_detection_yunet_2023mar_int8.onnx \
  -cotof \
  -tb flatbuffer_direct \
  --report_op_coverage
  

• https://huggingface.co/opencv/person_reid_youtureid/blob/main/person_reid_youtu_2021nov_int8.onnx

  onnx2tf -i person_reid_youtu_2021nov_int8.onnx \
  -cotof \
  -tb flatbuffer_direct \
  --report_op_coverage
  

INT8 ONNX	INT8 TFLite(LiteRT)
<img width="300" alt="Image" src="https://github.com/user-attachments/assets/c1411cb7-35aa-489d-ad87-291d64b766ec" />	<img width="300" alt="Image" src="https://github.com/user-attachments/assets/41aefe26-f243-45fb-9d21-9e16c7ef30b3" />

Main changes

• flatbuffer_direct graph optimizer enhancements in onnx2tf/tflite_builder/lower_from_onnx2tf.py:
  • inverse transpose-chain elimination
  • transpose bridge optimizations around Quantize/Dequantize
  • transpose folding around binary ops (ADD/SUB/MUL/DIV)
  • concat/conv NHWC propagation optimization
  • terminal QUANTIZE -> DEQUANTIZE cleanup for float outputs
  • terminal transpose-before-dequantize sanitization
  • dead-op/dead-tensor pruning improvements
  • dynamic reshape static resolution pass integration
• New/updated quantized op builders:
  • add build_qlinear_global_average_pool_op
  • improve QLinearConv, Conv, Pool, and Reshape lowering behavior
  • preserve dynamic batch shape_signature across intermediate tensors
• Input layout handling improvements:
  • default input-side channel-last exposure (NWC/NHWC/NDHWC) in Input op path
  • respect keep options (-kat / keep-shape flags) to opt out of automatic transpose
• Registry/export plumbing updates:
  • wire additional keep-input-shape/layout options through flatbuffer-direct entry points
  • expand op registry coverage/metadata for newly supported quantized ops and paths
• Documentation/version updates:
  • update flatbuffer-direct support status section in README
  • bump version to 2.0.12

Validation

• pytest -q tests/test_tflite_builder_direct.py
  • Result: 109 passed
• Model-level verification used during development:
  • person_reid_youtu_2021nov_int8.onnx converted with -tb flatbuffer_direct --report_op_coverage
  • confirmed QLinearGlobalAveragePool path emits quantized AVERAGE_POOL_2D (and no MEAN) in generated TFLite artifacts.

What's Changed

• Enhance flatbuffer_direct INT8 lowering, NHWC propagation, and transpose optimizations by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/874

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/2.0.11...2.0.12

2.0.11

Summary

This PR extends flatbuffer_direct for quantized ONNX graphs and documents the updated support matrix.

face_recognition_sface_2021dec_int8.onnx.zip

ONNX	LiteRT
<img width="375" height="612" alt="image" src="https://github.com/user-attachments/assets/5f06c5e7-392a-4bd1-8330-1fd7b4b10268" />	<img width="383" height="761" alt="image" src="https://github.com/user-attachments/assets/923dbe31-3434-4e5f-9156-cfe523a991d7" />

What was added

1. Added direct lowering for quantized operators in flatbuffer_direct:

   • QuantizeLinear -> QUANTIZE
   • DequantizeLinear -> DEQUANTIZE
   • QLinearAdd -> ADD
   • QLinearMul -> MUL
   • QLinearConv -> CONV_2D / DEPTHWISE_CONV_2D
   • QLinearMatMul -> FULLY_CONNECTED

2. Added flatbuffer_direct-only preprocess fusion:

   • DequantizeLinear -> BatchNormalization -> PRelu -> QuantizeLinear
   • BN params are folded and rewritten into Mul + Add form before lowering.

3. Added supporting direct builders and registry coverage:

   • BatchNormalization direct builder (MUL + ADD form)
   • Flatten direct builder (RESHAPE form)
   • op registry validators for quantized op constraints (const inputs, rank/shape checks, axis checks)

4. Updated README support status section for flatbuffer_direct.

Implementation details

Quantized builder layer

• New file: onnx2tf/tflite_builder/op_builders/quantized.py
• Implements quant parameter handling (scale, zero_point, quantized_dimension) for both:
  • raw quantized path (QLinear*)
  • dequant/quant bridge path (DequantizeLinear / QuantizeLinear)
• Handles:
  • dtype derivation from zero_point
  • quant metadata propagation across intermediate tensors
  • bias scale derivation for int32 bias (input_scale * weight_scale)

Dispatch/validation layer

• Updated onnx2tf/tflite_builder/op_registry.py with new DispatchEntrys and validators:
  • _validate_quantize_dequantize_linear
  • _validate_qlinear_binary
  • _validate_qlinear_conv
  • _validate_qlinear_matmul
  • _validate_batch_norm
  • _validate_flatten

Preprocess rules

• New rule file: onnx2tf/tflite_builder/preprocess/rules/quant_chain_fusion.py
• Registered via:
  • onnx2tf/tflite_builder/preprocess/rules/__init__.py
  • onnx2tf/tflite_builder/preprocess/__init__.py
• Rule ID: quant_chain_fusion_wave3

Tests

Added/updated tests

• tests/test_tflite_builder_direct.py
  • Added qlinear conv chain model
  • Added qlinear matmul/fc chain model
  • Included both in direct operator smoke coverage
• tests/test_tflite_builder_preprocess.py
  • Added test_quant_chain_fusion_wave3_rewrites_dq_bn_prelu_q_chain

Test results

• pytest -q tests/test_tflite_builder_preprocess.py tests/test_tflite_builder_direct.py
  • 72 passed

Real model verification

Validated with:

• face_recognition_sface_2021dec_int8.onnx
• command: python -m onnx2tf -i face_recognition_sface_2021dec_int8.onnx -o /tmp/onnx2tf_sface_pr_check -tb flatbuffer_direct --report_op_coverage -n

Coverage report highlights:

• conversion_error = null
• graph_summary.supported_nodes = 278 / 278
• graph_summary.unsupported_nodes = 0
• graph_summary.coverage_ratio = 1
• quant_chain_fusion_wave3 applied (matched_patterns=26, rewritten_patterns=26)

Additional updates

• Version bump:
  • pyproject.toml: 2.0.11
  • onnx2tf/__init__.py: 2.0.11
• README Docker tag examples aligned to 2.0.11.

Supplement for additional commit 1090987.

Additional implementation (more INT8-oriented)

• Added direct lowering of PRelu to PRELU in flatbuffer_direct
  • onnx2tf/tflite_builder/op_builders/elementwise.py:215
• For quantized input (INT8/UINT8), alpha (slope) is converted to a quantized constant with quantization metadata
  • onnx2tf/tflite_builder/op_builders/elementwise.py:185
• Added PRelu dispatch and validator in the registry
  • onnx2tf/tflite_builder/op_registry.py:664
  • onnx2tf/tflite_builder/op_registry.py:940
• Removed PRelu expansion from pseudo_ops_wave1; PRelu is now handled by direct builder without decomposition
  • onnx2tf/tflite_builder/preprocess/rules/pseudo_ops.py:489
• Added PRELU builtin options mapping in model writer
  • onnx2tf/tflite_builder/model_writer.py:202
• Updated README support matrix
  • README.md:280 (builtin count 38)
  • README.md:315 (PRelu row)

Tests

• Existing + newly added tests:
  • tests/test_tflite_builder_direct.py:881 (verifies PRELU builtin emission)
• Result:
  • pytest -q tests/test_tflite_builder_preprocess.py tests/test_tflite_builder_direct.py tests/test_tflite_builder_op_coverage.py
  • 76 passed

Real model verification (face_recognition_sface_2021dec_int8.onnx)

• Conversion succeeded (--report_op_coverage): unsupported 0 / coverage 1.0
• All 27 PRelu nodes are dispatch_mode=builtin
• Number of float-output ops in generated TFLite:
  • Before: 223 / 335
  • After: 115 / 227
  • Main effect: decomposed PRelu (Neg/Relu/Mul/Sub) was consolidated into a single PRELU op

What's Changed

• feat(flatbuffer_direct): add QLinear ops and quant-chain fusion by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/873

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/2.0.10...2.0.11

2.0.10

Summary

• improve rewrite_tflite_inout_opname I/O tensor index resolution for -coion
• add name-based matching with normalization (serving_default_, :0, /, __ -> :)
• add staged fallback order: existing SignatureDef -> existing tensor names -> unambiguous shape-product -> raw tflite order
• validate SignatureDef candidate indices and avoid duplicate tensor index assignments
• fix SignatureDef tensorIndex assignment to use real tensor indices (instead of buffer - 1)
• bump package version to 2.0.10 (pyproject.toml, onnx2tf/__init__.py) and sync Docker tag examples in README

Why

When multiple inputs/outputs have identical shapes, shape-only matching cannot reliably preserve ONNX I/O name/order. This change reduces reliance on shape heuristics and uses stable metadata first.

Validation

• python -m py_compile onnx2tf/utils/common_functions.py onnx2tf/__init__.py

What's Changed

• Fix -coion input/output mapping ambiguity for same-shape tensors by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/872

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/2.0.9...2.0.10

2.0.9

Summary

This PR improves robustness in quantized conversion paths and adds a new NonMaxSuppression option for class-score shrinking.

It addresses three concrete conversion failures seen during real model conversion, and adds a user-facing CLI/API switch for NMS behavior.

Background / Motivation

While converting quantized and post-process-heavy models, the following failure patterns were observed:

1. QLinearConv aborted when output shape metadata was None in the auto_pad == 'NOTSET' path.
2. QLinearConv depthwise weight reshape failed with mismatched element counts (invalid reshape target).
3. PRelu failed broadcasting when slope was channel-first style (e.g. [C,1,1]) but runtime tensor layout was channel-last.

In addition, for some NMS post-processing models (e.g. DAMO-YOLO style layouts), users requested a mode to shrink scores class dimension by argmax before NMS.

Changes

1) QLinearConv robustness fixes

File: onnx2tf/ops/QLinearConv.py

• Guarded SAME-padding shape comparison against missing metadata:
  • Before: accessed output_tensor_shape[2:] directly
  • After: checks both input/output shapes are not None before comparison
• Fixed depthwise filter reshape target:
  • Before: [..., input_weights_shape[2], input_weights_shape[3] // group]
  • After: [..., -1, input_weights_shape[3] // group]
  • This aligns behavior with Conv.py and avoids invalid reshape size errors.

2) PRelu slope layout alignment

File: onnx2tf/ops/PRelu.py

• Reworked slope handling to align slope layout with runtime input tensor layout.
• Added reshape logic for patterns like:
  • input: [N,H,W,C]
  • slope: [C,1,1]
  • transformed slope: [1,1,C]
• Implemented in a layout-agnostic way based on runtime tensor shape inspection.

3) New NMS option: --output_nms_with_argmax (-onwa)

Files:

• onnx2tf/onnx2tf.py
• onnx2tf/ops/NonMaxSuppression.py
• README.md

Added new CLI/API option to shrink class dimension of NMS scores:

• Input scores: [B, C, N]
• With option: argmax/reduce_max over class axis -> [B, 1, N]

Implementation details in NonMaxSuppression.py:

• Compute per-box class ids with argmax(scores, axis=1)
• Use reduced scores for NMS (reduce_max(..., keepdims=True))
• Reconstruct output [batch_index, class_index, box_index] by gathering class ids for selected boxes
• If scores rank is not 3, warn and fallback to existing behavior

<img width="495" height="322" alt="image" src="https://github.com/user-attachments/assets/576a47f1-3f7e-4644-a38c-135124fb5a7a" />

<img width="1380" height="505" alt="image" src="https://github.com/user-attachments/assets/f9f7d18a-8af3-44c9-8894-0652e937fab2" /> <img width="826" height="600" alt="image" src="https://github.com/user-attachments/assets/5c7b05fc-db25-4f75-af51-9b6bd86e546c" /> <img width="1396" height="268" alt="image" src="https://github.com/user-attachments/assets/918af08a-bd8b-47f7-a364-b78634bcd39b" />

4) Documentation and version update

Files:

• README.md

• onnx2tf/__init__.py

• pyproject.toml

• Documented the new NMS argmax option in README (CLI and Python API sections).

• Updated version from 2.0.8 to 2.0.9.

Validation

• python -m py_compile onnx2tf/onnx2tf.py onnx2tf/ops/QLinearConv.py onnx2tf/ops/PRelu.py onnx2tf/ops/NonMaxSuppression.py onnx2tf/__init__.py
• Sanity-checked NMS output tuple construction for argmax mode and verified output format [batch, class, box].

Compatibility / Risk

• Default behavior is unchanged unless --output_nms_with_argmax is explicitly enabled.
• QLinearConv and PRelu changes are defensive and targeted at previously failing shape/layout edge cases.

Issue

• [TODO] Implementation of the ArgMax force-replacement option for NMS with associated class score refinement. #420

What's Changed

• Fix QLinearConv/PRelu edge cases and add NMS argmax option by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/871

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/2.0.8...2.0.9

• Gelu
  • Forced to avoid the problem of onnx normality check terminating abnormally due to the presence of Gelu, which should not exist in opset=17.
  • https://huggingface.co/qualcomm/ConvNext-Tiny/blob/5b8dbb4db0da2c36a4a05e1577cf5171a9395ec0/ConvNext-Tiny.onnx
• Added --test_data_nhwc_path (-tdnp) support for validation input loading.
• Clarified README difference between -tdnp and -cind.
• Clarified multi-input behavior:
  • -tdnp uses a single NHWC test array and reuses it across eligible inputs (after per-input resize/layout conversion).
  • Per-input mapping is not supported by -tdnp; use -cind when per-input control is needed.

What's Changed

• Support broken onnx/Gelu by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/868
• test_data_nhwc_path by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/870

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/2.0.7...2.0.8

• Docker
  • Build and install from source code.
  • x64/arm64

What's Changed

• Build and install from source code by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/867

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/2.0.6...2.0.7

2.0.6

Summary

flatbuffer introduces an opt-in direct TFLite generation backend and related tooling.

1) New backend and integration points

• Add tflite_backend switch with tf_converter (default) and flatbuffer_direct
• Keep existing TensorFlow-based path as default while enabling direct path incrementally
• Wire direct backend into CLI/API entry points

2) Direct FlatBuffer builder implementation

• Add new onnx2tf/tflite_builder/ module tree (IR, lowering, op dispatch/builders, model writer)
• Implement direct export path for float32/float16 .tflite
• Add schema loader with pinned LiteRT schema.fbs tag handling

3) Quantization, evaluation, and splitting features

• Add staged quantization support in direct path:
  • dynamic range quantization controls
  • integer/full-integer variants including int16-activation variants
• Add ONNX-vs-TFLite accuracy evaluation reports
• Add size-based auto split planning/output and split accuracy evaluation

4) Coverage diagnostics and custom-op policy

• Add op registry + coverage report generation (*_op_coverage_report.json)
• Add policy matrix and reason-code based unsupported diagnostics
• Add custom-op candidate flow with allowlist controls
• Add fallback option --flatbuffer_direct_fallback_to_tf_converter

5) Direct preprocess pipeline and migration from TF-path optimizations

• Add preprocess pipeline foundation and rules:
  • pattern fusion
  • pseudo-op rewrites
  • attribute normalization
  • limited constant folding
• Expand direct builtin support incrementally (including reductions/index/norm and selected activations)

6) Tests and docs

• Add focused test suites for direct builder, preprocess, split planner, and op coverage snapshots
• Update README and migration guide with:
  • supported/unsupported policy visibility
  • TF vs direct behavior differences
  • rollout and troubleshooting guidance
  • explicit note that flatbuffer_direct is experimental

Diff summary

• Commits ahead of main: 67
• Major additions include onnx2tf/tflite_builder/** and new direct-path test suites

Notes

• This PR does not change the default backend (tf_converter remains default).

What's Changed

• Plan by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/838
• Step 1: Add staged tflite_backend switch and flatbuffer_direct stub by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/839
• Step 10-12: Add limited dynamic-range quantization for flatbuffer_direct by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/840
• Step 13: Expand dynamic-range quantization targets in flatbuffer_direct by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/841
• Step 14: Wire quant_type into flatbuffer_direct dynamic-range quantization by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/842
• Step 15: Add limited integer quantization support for flatbuffer_direct by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/843
• Step 16-17: Improve quantization precision controls and add int16-act direct outputs by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/844
• Step 18: Finalize concrete specs for evaluation/splitting/full-op rollout by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/845
• Step 19: ONNX vs TFLite accuracy report foundation by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/846
• Step 20: Operationalize ONNX-TFLite evaluation for CLI/CI by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/847
• Step 21: Add TFLite split size estimation and partition planning by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/848
• Step 22: Implement split outputs and manifest generation by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/849
• Step 23: Add split-model accuracy evaluation and reporting by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/850
• Step 24: OPレジストリ化とカバレッジレポート導入 by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/851
• Step 25: Wave1 high-frequency unary op support by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/852
• Step 26: Wave2 reduce/index/norm support by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/853
• Step 27: establish custom-op policy and diagnostics by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/854
• Step 29: auto split size option with KB/MB/GB and split-target linkage by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/855
• Step 30: document flatbuffer_direct ONNX op support matrix by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/856
• Step 31: add direct optimization migration plan by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/857
• Step A1: 棚卸し結果とカバレッジ運用ルールを反映 by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/858
• Step A2: direct 共通プリパス基盤の導入 by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/859
• Step A3: 擬似OP置換ルール（Wave1）移植 by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/860
• Step A4: 複合パターン吸収（Wave2）移植 by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/861
• Step A5: 属性・shape 正規化の強化 by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/862
• Step A6: custom-op candidate 縮小フェーズ（Einsum builtin優先） by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/863
• Step A7: テスト・回帰固定（coverage key/reason_code snapshot） by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/864
• Step A8: 文書化・運用導線整備と experimental 明示 by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/865
• Merge flatbuffer: add direct TFLite FlatBuffer backend and tooling by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/866

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/2.0.5...2.0.6

• Update ai_edge_litert==2.1.2

What's Changed

• ai_edge_litert==2.1.2 by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/837

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/2.0.4...2.0.5

v2.0.4

Background

• Running onnx2tf.py -i densenet-12.onnx -cotof failed at Mul (sng_Mul_1) with a broadcast shape mismatch.
• Actual TF shapes at failure:
  • x: (1, 112, 112, 64)
  • y: (1, 64, 1, 1) (channel-first style)

Root Cause

• In shape_unmatched_special_avoidance_workaround, some transpose paths converted constant np.ndarray inputs into tf.Tensor.
• After this conversion, downstream explicit_broadcast behavior changed and the constant was no longer normalized to NHWC-compatible shape.
• As a result, the constant remained (1, 64, 1, 1) instead of (1, 1, 1, 64), causing tf.math.multiply to fail.

Implemented Fix

• File: onnx2tf/utils/common_functions.py
• Function: shape_unmatched_special_avoidance_workaround
• Added helper:
  • _transpose_preserve_array(tensor, perm)
  • Uses np.transpose when tensor is np.ndarray
  • Uses existing transpose_with_flexing_deterrence for tensor-like inputs
• Replaced internal transpose call sites in this function with _transpose_preserve_array so constants keep np.ndarray type through alignment.

Effect

• Constant layout alignment now preserves array semantics and allows explicit_broadcast to correctly reshape constants to channel-last form when needed.
• For the failing case, y is now normalized to (1, 1, 1, 64), and Mul executes successfully.

Validation

• Reproduced and re-ran:
  • python onnx2tf/onnx2tf.py -i densenet-12.onnx -cotof
• Result:
  • Conversion completed successfully (exit code 0)
  • Validation logs reported matching outputs for the model conversion path.

What's Changed

• fix-shape_unmatched_special_avoidance_workaround by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/836

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/2.0.3...2.0.4

v2.0.3

Key Changes

1. Removed onnx-graphsurgeon dependency (breaking change)

• Introduced onnx2tf/gs.py and migrated away from onnx_graphsurgeon.
• Switched imports in almost all OP implementations from onnx_graphsurgeon as gs to onnx2tf.gs as gs.
• Core commit:
  • 3dae8d2 Breaking away from dependency on onnx-graphsurgeon

2. Improved GS compatibility

• Separated Constant from Variable inheritance and introduced a new Tensor base class.
• Improved GS-compatible behavior for constant handling and I/O resolution.
• Core commit:
  • 4ad5d63 Constant has been separated from Variable inheritance and a new Tensor base class has been created

3. Conversion stability improvements (layout/shape handling)

• Fixed issues around ScatterND after GS migration.
• Added mismatch mitigation such as shape normalization (absorbing list/tuple differences).
• Core commits:
  • 2552377 Bug fixes for ScatterND after gs migration
  • d0be8d8 normalization

4. Dependency and runtime environment updates

• Updated onnx / onnxruntime.
• Updated CI runtime to Ubuntu 24.04 / Python 3.12.
• Replaced onnxsim with onnxsim-prebuilt.
• Updated sng / sne versions.
• Related commits:
  • 7c21a22, 1528092, 3f386df, 14e0048

5. CI / publishing / documentation maintenance

• Updated workflows (test-models.yml, python-publish.yml, codeql-analysis.yml).
• Updated README.md.
• Added GS compatibility tests (tests/test_gs_compat.py).
• Related commits:
  • afa0423, 3dd8e01, c65fea5

6. Dependencies

• requests==2.32.5
• numpy==1.26.4
• onnx==1.20.1
• onnxruntime==1.24.1
• opencv-python==4.11.0.86
• onnxsim-prebuilt==0.4.39.post2
• onnxoptimizer==0.4.2
• ai-edge-litert==2.1.0
• tensorflow==2.19.0
• tf-keras==2.19.0
• sne4onnx==2.0.0
• sng4onnx==2.0.0
• psutil==5.9.5
• protobuf==4.25.5
• h5py==3.12.1
• ml_dtypes==0.5.1
• flatbuffers==25.12.19

7. Docker

• Ubuntu 24.04 linux/amd64
• Ubuntu 24.04 linux/arm64

What's Changed

• Breaking away from dependency on onnx-graphsurgeon by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/829
• onnx,onnxruntime latest,ubuntu24.04,python3.12 by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/830
• Update deps by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/831
• Constant has been separated from Variable inheritance and a new Tensor base class has been created by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/832
• Bug fixes for ScatterND after gs migration by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/833
• normalization by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/834
• V2 by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/835

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.29.24...2.0.2

• 100% coverage of ONNX operations :bear:
  • Support for CenterCropPad, Optional, TfIdfVectorizer, GroupNormalization
  • Bugs in each implemented operation will be dealt with individually as issues arise.

What's Changed

• ops full support by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/828

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.29.23...1.29.24

PReLU fallback handling

• Try native PReLU first; if it raises an exception, fall back to a pseudo‑PReLU implementation only for that failing op.
• This limits the pseudo‑PReLU path to problematic nodes instead of applying it globally.

Affected files

• onnx2tf/ops/PRelu.py

Model

• something.onnx.zip

What's Changed

• Addressing edge cases of PRelu by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/827

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.29.22...1.29.23

Summary

• Added a MaxPool guard to brute-force NCHW/NHWC axis alignment when batch-dimension mismatch is detected with fully static shapes.
• Implemented sorted Unique output handling (lexicographic for axis-based uniques), axis conversion, and safer output handling for optional outputs.
• Improved ScatterElements to handle rank-mismatched indices/updates via padding + broadcast before scatter update.
• Bumped version to 1.29.22 across package metadata and documentation.
• conv_tasnet.onnx
• model_hawp.onnx

Details

• onnx2tf/ops/MaxPool.py
  • Added a conservative, shape-gated permutation search to align input layout when batch dimensions are swapped/misaligned.
• onnx2tf/ops/Unique.py
  • Added axis conversion for layout-transposed tensors.
  • Implemented sorted=1 behavior (stable lexicographic sort across subtensors) with warnings on unsupported dynamic/dtype cases.
  • Wrote outputs conditionally based on how many outputs the ONNX node defines.
• onnx2tf/ops/ScatterElements.py
  • Added rank padding/broadcast so indices and updates can be aligned before scatter.
• Versioning
  • onnx2tf/__init__.py, pyproject.toml, README.md: version bumped to 1.29.22.

What's Changed

• MaxPool, ScatterElements, Unique by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/826

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.29.21...1.29.22

Added

• New CLI option --value_hints (-vh) to set dummy inference input values by input name, with * as a default fallback.
• Parsing for scalar value hints (ints/floats/bool) and default fallback in dummy inference.
• Global defaults for shape_hints and value_hints so OP-level TF dummy inference can inherit them without explicit plumbing.

Changed

• Dummy inference now uses value_hints to fill inputs via np.full when custom inputs are not provided.
• convert() now propagates value_hints to auto-split sub-conversions and initializes dummy hint defaults at startup.

Fixed

• has_external_data is now initialized before sng4onnx runs to avoid UnboundLocalError when onnx_graph is provided.

Docs

• README updated: CLI parameter list, accuracy-check usage notes, and in-script convert() signature/argument descriptions now include value_hints with examples.

What's Changed

• Support --value_hints by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/825

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.29.20...1.29.21

Auto-split & partitioning

• sne4onnx==1.0.15 https://github.com/PINTO0309/sne4onnx
• sng4onnx=1.0.5 https://github.com/PINTO0309/sng4onnx
• Enable automatic splitting when ONNX file size or estimated initializer size exceeds 2GB.
• Add forced auto-split mode via --enable_auto_split_model (short -easm), using --auto_split_max_size_mb as target size.
• Add CLI option --auto_split_max_size_mb (short -asmsm) to control target partition size (MB).
• Save each split ONNX directly via sne4onnx.extraction(..., output_onnx_file_path=...) into per-part folders.
• Avoid overwriting the original ONNX by explicitly setting each part's input_onnx_file_path to the split ONNX path.
• Switch split folder/file numbering to 4-digit zero-padded format (part_0001, *_part_0001.onnx).
• Auto-split uses temporary memmap-backed .npy files for input/output propagation and cleans them up after conversion.
• Partitioning respects reachable nodes (excludes constant-only islands as partition boundaries) and merges invalid ranges.
• ONNX with more than 2GB of External Data <img width="477" height="83" alt="image" src="https://github.com/user-attachments/assets/ff3a109b-a3ff-4966-bfde-401b366255ab" />
• Command sample

  pip install sne4onnx==1.0.15 sng4onnx==1.0.5
  onnx2tf -i vit_h_encoder.onnx -asmsm 50 -nuo -nuonag -cotof
  

  <img width="877" height="566" alt="image" src="https://github.com/user-attachments/assets/313666e1-b14f-402f-85bf-891bba3725da" />
• Protocol Buffer split output and conversion <img width="583" height="179" alt="image" src="https://github.com/user-attachments/assets/006db626-c331-46b7-900a-4a7ba7b1e306" />
• Results

  Split.1	Split.2	Split.3
  <img width="462" height="253" alt="image" src="https://github.com/user-attachments/assets/0b694adc-3191-4e3e-bb46-e12a58c1ea10" />	<img width="471" height="253" alt="image" src="https://github.com/user-attachments/assets/cfcf991f-93d8-4b50-86f1-5544ef47a1ca" />	<img width="470" height="217" alt="image" src="https://github.com/user-attachments/assets/f51c954a-ee0c-4570-8679-beabc89ec246" />

ONNX/TF output propagation across parts

• Cache ONNX dummy outputs per partition to feed later partitions.
• Run TF dummy inference after each part conversion and cache TF outputs for the next part.
• Build a strict ONNX output name ↔ TF output tensor mapping:
  • Prefer name-based matches.
  • Fall back to value/shape matching (including common layout transpose checks).
  • Use mapping to populate TF cache and split_output_layouts reliably.
• Infer keep_shape_absolutely_input_names for the next part based on ONNX/TF output shape comparison.
• Store TF cache under both full (name:0) and base (name) keys to maximize cache hits.

Auto-split dummy input handling

• Generate per-part dummy inputs with correct shapes and dtypes using numpy.memmap.
• Preserve user-provided custom_input data for part1; auto-generate for later parts.
• Improve prefilled TF input handling so cached TF outputs override dummy generation.

Diagnostics & robustness

• Improve fallback logging for auto-split TF input mismatch: input name, TF shape, numpy shape, and file path.
• Add clearer log message when auto-split is forced.
• Fix dummy_tf_inference to avoid input_dtype unbound error and ensure dtype casting.

Operator fixes (indices/transpose behavior)

• Correct handling of indices-like inputs to avoid invalid layout transposes:
  • GatherND, GatherElements, ScatterElements, ScatterND, TensorScatter.
• Remap indices when params are transposed (where applicable), instead of transposing indices directly.

Documentation

• Document --auto_split_max_size_mb and --enable_auto_split_model (including -easm shortcut) in README.

• Update API signature sections to include new parameters.

• [TODO] Implementation of Proto Splitter (breakthrough of 2GB contract for Protocol Buffers) #455

What's Changed

• Split FlatBuffer by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/824

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.29.19...1.29.20

• Support for DeformConv, DFT
• DeformConv operation from Onnx to TFLite #469

What's Changed

• DeformConv, DFT by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/823

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.29.18...1.29.19

• Support for RotaryEmbedding, Scan, StringConcat, StringSplit, TensorScatter, ImageDecoder, RMSNormalization, NegativeLogLikelihoodLoss, SoftmaxCrossEntropyLoss, RegexFullMatch

What's Changed

• Ops by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/822

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.29.17...1.29.18

• Concat

  • Improved Concat conversion stability.
  • In 4D, the maximum number of attempts introduced by the new search logic is roughly bounded as:

    4 * P^N
    

  • 4 is the number of candidate axes (0–3).
  • P is the number of permutation candidates per input.
  • N is the number of input tensors.

  Permutation count P:

  • Full search: P = 24 (all 4D permutations).
  • If 24^N exceeds the guard threshold (20,000), the search switches to a limited set, so P = 3.

  Worst-case attempts:

  • N=2: 4 * 24^2 = 2,304
  • N=3: 4 * 24^3 = 55,296
  • N=4: 24^4 exceeds the threshold, so limited mode: 4 * 3^4 = 324

These are hard upper bounds; in practice the actual attempts are typically much fewer because shape-incompatible candidates are pruned before tf.concat, and the search stops early once the ONNX output shape matches.

• e2pose <img width="1556" height="478" alt="image" src="https://github.com/user-attachments/assets/2b92266c-bbc5-4178-b56f-2542f8386669" />

What's Changed

• Improved Concat conversion stability by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/821

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.29.16...1.29.17

• Experimental
• Dequantized Float32 unification for quant ops
  • QuantizeLinear / DequantizeLinear / DynamicQuantizeLinear / QLinearConv / QLinearMatMul now operate in Float32.
  • Broadcast handling for per‑tensor / per‑axis / blocked quantization was corrected to match spec.
  • Added is_dequantized to avoid double dequantization and redundant quant→dequant paths.
• Removal of QDQ expansion (Mul/Round/Min/Relu/Mul)
  • When QuantizeLinear output is consumed only by Cast → DequantizeLinear (or direct DequantizeLinear), fake‑quant is bypassed and Float32 passes through.
  • This removes Mul/Round/Minimum/Relu from the generated TFLite.
• QLinearConcat fixes and stabilization
  • QLinearConcat now mirrors Concat behaviors, including axis correction, NHWC handling, and shape‑mismatch avoidance.
  • Fixed the dequantization formula inside QLinearConcat to (x - zp) * scale.
• Dynamic axis inference for Concat
  • Added runtime axis inference for Concat when all shapes are dynamic.
    • The only mismatched dimension is selected as the concat axis.
• Split axis inference and FlexSplitV avoidance
  • Added runtime axis inference in Split based on sum(split).
  • For non‑uniform split lists, tf.split is replaced with tf.strided_slice decomposition to avoid FlexSplitV in TFLite.

ONNX	TFLite
<img width="935" height="594" alt="image" src="https://github.com/user-attachments/assets/10528382-26db-4a12-8285-5894bea49165" />	<img width="759" height="531" alt="image" src="https://github.com/user-attachments/assets/25c36f8c-b356-4b9b-bed9-b40caa4d7d98" />

What's Changed

• [experimental] Supports inverse quantization of INT8 models by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/820

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.29.15...1.29.16

• Shape
  • Fixed an infinite loop bug.
• Unable to convert zipformer to tflite #785

What's Changed

• Fixed an infinite loop bug by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/819

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.29.14...1.29.15

• Transformation stability has been significantly improved for models with many dynamic tensors containing NonZero OPs.

• VLM Encoder

• NonZero

  • Added input data capture for ONNX/TF dummy inference to reuse actual input tensors.
  • Implemented apply_nonzero_passthrough to store NonZero outputs as their input tensors in ONNX dummy outputs (optionally updating graph output shapes).
  • Implemented apply_nonzero_passthrough_tf to mirror the same passthrough behavior on TF dummy outputs.
  • Wired both passthroughs into validation (-coto/-cotof) and auto-generate-json validation flows.
  • dummy_onnx_inference: added input_datas_for_validation output hook.
  • dummy_tf_inference: added input_datas_for_validation output hook.

• Flatten

  • brute-force alignment
  • Added brute-force permutation + axis search in onnx2tf/ops/Flatten.py to match ONNX dummy outputs exactly when available.
  • Falls back to ONNX input tensor values if TF-side validation input is not obtainable.

• explicit_broadcast now reshapes 1D operands to the last axis when it matches the other operand’s last dimension, avoiding unintended transposes.

• The input shape broadcasting error of the ADD operator and replace.json cannot resolve the issue #784

• model: https://huggingface.co/HuggingFaceTB/SmolVLM-256M-Instruct/blob/main/onnx/vision_encoder.onnx

  onnx2tf \
  -i vision_encoder.onnx \
  -sh pixel_values:1,1,3,512,512 pixel_attention_mask:1,1,512,512
  

  <img width="1575" height="506" alt="image" src="https://github.com/user-attachments/assets/595b7249-14fa-4b7d-b613-15916ad8ecec" /> <img width="807" height="532" alt="image" src="https://github.com/user-attachments/assets/4308a7b9-5f94-4730-8ee2-e93c0ed63f4b" />

• The input shape broadcasting error of the ADD operator and replace.json cannot resolve the issue #784

What's Changed

• Transformation stability has been significantly improved for models with many dynamic tensors containing NonZero OPs by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/818

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.29.13...1.29.14

• AveragePool, Expand, Slice
  • Improved stability of dynamic tensor transformations.
  • Improved shape estimation algorithm with -sh and -kat combination.

onnx2tf -i campp_vin.onnx -kat input -sh input:1,100,80

<img width="1464" height="573" alt="image" src="https://github.com/user-attachments/assets/66fb3b33-abe4-4b49-9447-c107e67d78d9" />

<img width="642" height="537" alt="image" src="https://github.com/user-attachments/assets/61732ef8-dea5-4f3b-8f88-cbc8f6e92dc5" />

• Unable to convert a model with 3d input shape of dynamic length into tflite int8 format #673

What's Changed

• Improved stability of dynamic tensor transformations by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/817

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.29.12...1.29.13

• MatMul, BatchMatMul

  • I adjusted the MatMul post‑processing to handle cases where the ONNX output rank is smaller than the TF matmul result by squeezing unit dims before attempting a transpose. This prevents the invalid perm=['0'] call that caused the Dimension must be 2 but is 1 error.

  • Added a small shape‑match helper and a squeeze step for trailing or leading 1 dims when ONNX output rank is smaller. Recomputed shape after squeeze and only transpose when the perm length matches the tensor rank.

    onnx2tf -i rtdetrv4_s.onnx -cotof
    

    <img width="1171" height="424" alt="image" src="https://github.com/user-attachments/assets/f3420029-af0a-426b-bbdb-29b331458f33" />

• Transpose dimension error in RT-DETRv4 onnx #815

What's Changed

• Improved MatMul/BatchMatMul conversion stability by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/816

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.29.11...1.29.12

• uv package build
• requires-python = ">=3.10.12"

What's Changed

• Use uv build backend by @Boulaouaney in https://github.com/PINTO0309/onnx2tf/pull/813
• Version fix by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/814

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.29.10...1.29.11

• Support for Loop, LpPool, MaxRoiPool.
  • Stable Loop conversion and TFLite export
    • Avoided the SymbolicTensor error during TFLite conversion by preferring from_keras_model and falling back to from_concrete_functions only if needed.
      • File: onnx2tf/onnx2tf.py
  • Loop implementation fixes (conversion + inference stability)
    • Fixed list/tuple nesting mismatch in tf.while_loop returns.
    • Forced scalar handling for cond and loop-carried scalars.
    • Prevented constants from being created inside the while-body by registering subgraph constants outside the loop.
    • Recursively registers constants from nested subgraphs (e.g., If branches inside Loop).
      • File: onnx2tf/ops/Loop.py
  • Dummy inference for 1D inputs
    • Adjusted dummy_tf_inference to avoid forcing batch-size alignment for 1D inputs, which caused cond to become vector-shaped.
      • File: onnx2tf/utils/common_functions.py
  • Suppress repeated subgraph logs
    • Added a suppress_log flag to print_node_info, and disabled logging during subgraph execution for Loop and If.
      • Files: onnx2tf/utils/common_functions.py, onnx2tf/ops/Loop.py, onnx2tf/ops/If.py

What's Changed

• Support for Loop, LpPool, MaxRoiPool by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/812

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.29.9...1.29.10

• Unsqueeze
  • Improved stability when handling undefined shapes.
• Docker
  • Add git

What's Changed

• unsqueeze none by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/811

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.29.8...1.29.9

• onnxsim ir_version=10
  • onnx==1.19.0
  • onnxruntime==1.23.0
  • onnxsim==0.4.36
  • onnxoptimizer==0.4.2
• Not fix: https://github.com/PINTO0309/onnx2tf/issues/436

What's Changed

• verup by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/810

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.29.7...1.29.8

• GridSample
  • Modes: linear, nearest, and cubic (the legacy bilinear string is mapped to linear).
  • Padding modes: zeros, border, and reflection.
  • Dimensionality: 1D, 2D, and 3D inputs are supported with N-linear, N-nearest, and N-cubic behavior.
  • Cubic interpolation uses a Catmull-Rom kernel (a = -0.75). 2D uses 4x4 samples, 3D uses 4x4x4 samples.
  • Performance: linear-index gather (tf.gather) is used for 2D/3D to reduce gather_nd overhead.
  • Dtype: interpolation runs in float, then casts back to the original input dtype.
    • string inputs are allowed only with nearest mode.
• Improved grid_sample #544

What's Changed

• Improve GridSample by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/809

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.29.6...1.29.7

• Support for Attention, AffineGrid, BlackmanWindow

What's Changed

• Attention, AffineGrid, BlackmanWindow by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/808

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.29.5...1.29.6

• Support for CumProd

What's Changed

• Support for CumProd by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/807

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.29.4...1.29.5

• Support for BitwiseAnd, BitwiseNot, BitwiseOr, BitwiseXor

What's Changed

• Support for Bitwise OP by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/806

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.29.3...1.29.4

• AvgPool
  • ceil_mode=1 + Improved conversion stability for dynamic tensor

<img width="936" height="546" alt="image" src="https://github.com/user-attachments/assets/36df94aa-60e6-4062-a28c-32f56759feec" />

<img width="886" height="575" alt="image" src="https://github.com/user-attachments/assets/2b86293a-1475-4033-91ff-df62342334c5" />

onnx2tf -i campp_vin.onnx -kat input -sh input:1,100,80

<img width="1135" height="567" alt="image" src="https://github.com/user-attachments/assets/6868432d-9cea-4132-9c69-a94ec6a8dd78" />

• Unable to convert a model with 3d input shape of dynamic length into tflite int8 format #673

What's Changed

• Avgpool ceilmode=1 + dynamic by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/805

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.29.2...1.29.3

• Conv

  • Support for dilations conv 3x3 - strides 2x2

  • sgscsh.onnx.tar.gz

    <img width="801" height="585" alt="image" src="https://github.com/user-attachments/assets/11c30a2c-2921-4e7a-b05e-8eb18522d459" />

    onnx2tf -i sgscsh.onnx -cotof
    

    <img width="1040" height="459" alt="image" src="https://github.com/user-attachments/assets/75832561-5168-4b37-aaf6-206ca9ed80f6" />

    <img width="955" height="611" alt="image" src="https://github.com/user-attachments/assets/5d08a605-c62f-4f0d-bdc4-e6aace8b3660" />

• Large Dilated Convolution is working strange #251

What's Changed

• fix dilations by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/804

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.29.1...1.29.2

• SequenceEmpty
  • From:

    sequence_dtype = ONNX_DTYPES_TO_TF_DTYPES(graph_node.attrs.get('dtype', 1)) # Float32 
    

  • To:

    sequence_dtype = ONNX_DTYPES_TO_TF_DTYPES[graph_node.attrs.get('dtype', 1)] # Float32 
    

• Bug: TypeError: 'dict' object is not callable #786

What's Changed

• bug fix SequenceEmpty by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/803

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.29.0...1.29.1

• Support for uv environment

• Large model conversion

  • Improved to convert models with a very large number of operations that exceed the combined size of RAM and SWAP.

  • I offloaded all the dummy inference tensors needed for the massive amount of shape estimation to storage via np.memmap .

  • Essentially, this means that any large model can be converted, as long as there is enough storage space.

  • Conversion of dehaze_maxim_2022aug_opt_sim.onnx requires 145GB of storage instead of RAM.

    ssc4onnx -if dehaze_maxim_2022aug_opt_sim.onnx
    ┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━┳━━━━━━━━━━━━┓
    ┃ OP Type                ┃ OPs        ┃ Sizes      ┃
    ┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━╇━━━━━━━━━━━━┩
    │ Add                    │ 1,523      │ 301.4KiB   │
    │ Cast                   │ 48         │ 1.3KiB     │
    │ Concat                 │ 74         │ 2.0KiB     │
    │ Conv                   │ 212        │ 18.9MiB    │
    │ ConvTranspose          │ 12         │ 1.6MiB     │
    │ Einsum                 │ 14         │ 4.9MiB     │
    │ Exp                    │ 35         │ 0.0B       │
    │ Gather                 │ 4          │ 3.5KiB     │
    │ Gemm                   │ 34         │ 13.0MiB    │
    │ GreaterOrEqual         │ 24         │ 96.0B      │
    │ MatMul                 │ 416        │ 33.2MiB    │
    │ Max                    │ 272        │ 1.1KiB     │
    │ Mul                    │ 2,579      │ 83.9KiB    │
    │ Neg                    │ 35         │ 0.0B       │
    │ Not                    │ 24         │ 0.0B       │
    │ Reciprocal             │ 275        │ 0.0B       │
    │ ReduceSum              │ 512        │ 30.2KiB    │
    │ Reshape                │ 1,060      │ 36.5KiB    │
    │ Slice                  │ 240        │ 22.5KiB    │
    │ Sqrt                   │ 240        │ 0.0B       │
    │ Squeeze                │ 34         │ 272.0B     │
    │ Sub                    │ 480        │ 0.0B       │
    │ Tanh                   │ 152        │ 0.0B       │
    │ Transpose              │ 952        │ 85.7KiB    │
    │ Unsqueeze              │ 22         │ 176.0B     │
    │ ---------------------- │ ---------- │ ---------- │
    │ Total number of OPs    │ 9,273      │            │
    │ ---------------------- │ ---------- │ ---------- │
    │ Total params           │ 18.0M      │            │
    │ ====================== │ ========== │ ========== │
    │ Model Size             │ 65.6MiB    │ 72.3MiB    │
    └────────────────────────┴────────────┴────────────┘
    INFO: file: dehaze_maxim_2022aug_opt_sim.onnx
    INFO: opset: 13
    INFO: input_name.1: input_image shape: [1, 3, 512, 640] dtype: float32
    INFO: output_name.1: Identity_6__0 shape: [1, 128, 160, 3] dtype: float32
    INFO: output_name.2: Identity_7__0 shape: [1, 256, 320, 3] dtype: float32
    INFO: output_name.3: Identity_8__0 shape: [1, 512, 640, 3] dtype: float32
    INFO: output_name.4: Identity_9__0 shape: [1, 128, 160, 3] dtype: float32
    INFO: output_name.5: Identity_10__0 shape: [1, 256, 320, 3] dtype: float32
    INFO: output_name.6: Identity_11__0 shape: [1, 512, 640, 3] dtype: float32
    INFO: Finish!
    

    <img width="1045" height="184" alt="image" src="https://github.com/user-attachments/assets/04cdde58-a2ed-466b-b821-5b4d33b20826" />

    <img width="428" height="264" alt="image" src="https://github.com/user-attachments/assets/7548b928-7e0b-47e7-82b8-86ff9bb402f7" />

    <img width="1450" height="470" alt="image" src="https://github.com/user-attachments/assets/0d20384c-8409-43a4-bf67-894e2db843a5" />

• [TODO] Switch to a mode that performs inference in slow delay shape estimation mode during very large size transformations such as diffusion models #430

• Improvement of huge model transformation (implementation of Out of Memory workaround) #679

What's Changed

• np.memmap by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/802
• uv by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/800

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.28.8...1.29.0

• Quantization option

  quant_norm_mean: Optional[str]
      Normalized average value during quantization.\n
      Only valid when the "-cind" option is not used.\n
      Default: "[[[[0.485, 0.456, 0.406]]]]"
  
  quant_norm_std: Optional[str]
      Normalized standard deviation during quantization.\n
      Only valid when the "-cind" option is not used.\n
      Default: "[[[[0.229, 0.224, 0.225]]]]"
  

What's Changed

• -qmean, -qstd by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/799

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.28.7...1.28.8

• ir_version=10, kernel_shape handling

• Conv

  • Support for ir_version=10, where kernel_shape was deleted

    <img width="1504" height="367" alt="image" src="https://github.com/user-attachments/assets/f775555c-d202-488f-afa8-9710b7181121" />

    <img width="1079" height="502" alt="image" src="https://github.com/user-attachments/assets/4a7b466b-7173-495e-a7cf-458ff50309e9" />

• axes don't match array, input Conv : [ Torch -> ONNX -> TFLite ] #797

What's Changed

• ir_version=10, kernel_shape handling by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/798

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.28.6...1.28.7

• common_functions.py In the original code, when downloading test image data from GitHub releases, the program switches to Wasabi Storage if the request times out. However, the request to Wasabi Storage did not include a timeout parameter, which could cause the program to hang for a long time on unstable networks.

  Added a timeout=(1.0, 5.0) parameter to the request when downloading from Wasabi Storage, so that in case of unstable network, the program will properly notify the user of a download failure instead of hanging indefinitely.

What's Changed

• Add timeout to Wasabi Storage download to prevent hanging on unstable networks by @shyhyawJou in https://github.com/PINTO0309/onnx2tf/pull/796

New Contributors

• @shyhyawJou made their first contribution in https://github.com/PINTO0309/onnx2tf/pull/796

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.28.5...1.28.6

• Fix Dockerfile

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.28.4...1.28.5

What's Changed

• Fix combine_fixes arguments by @sn2234 in https://github.com/PINTO0309/onnx2tf/pull/793
• Fix CI test by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/794

New Contributors

• @sn2234 made their first contribution in https://github.com/PINTO0309/onnx2tf/pull/793

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.28.3...1.28.4

  -agje, --auto_generate_json_on_error
    Attempts to generate a parameter replacement JSON when accuracy validation finds errors
    greater than 1e-2. Useful for quickly capturing fixes during -cotof runs.
    Disabled by default to avoid unexpected file generation.

What's Changed

• --auto_generate_json_on_error by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/792

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.28.2...1.28.3

• MatMul Op MatMul improved resource usage for matrix shape compatibility check.

  1. Content and background The MatMul op converter tries to numerically optimize operation M=A*B. To only optimize compatible shapes, a check is present. The current check uses np.matmul on dummy matrices to verify the output shape is compatible with expected shape. For some of the compatible shapes (due to broadcasting), this can consume significant resources both in compute time and memory.

  2. Summary of corrections Add a function to compute shape returned via np.matmul without actually running the np.matmul. In case the function fails (incompatible shapes) and returns None, use current implementation with np.matmul to verify the result (based on my internal unit tests it seems np.matmul can be safely removed, but I left it as it no longer poses threat of high resource usage).

What's Changed

• Op MatMul improved resource usage for matrix shape compatibility check by @radekzc in https://github.com/PINTO0309/onnx2tf/pull/780

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.28.1...1.28.2

• Fix boolean parameter replacement

  The replace_parameter function converts parameter values from JSON replacement files into appropriate Python types during ONNX model conversion. A case-sensitivity bug prevented proper conversion of the capitalized boolean string "True", causing it to remain as a string instead of being converted to boolean True, which affects both the parameter replacement feature and the new auto-JSON generation feature.

  The bug occurred because the code called .lower() on the input string (converting "True" → "true") but then compared against "True" (capital T), so "true" == "True" returned False and conversion failed. The fix was to change the comparison from replace_value.lower() == "True" to replace_value.lower() == "true". Note that "False" was not affected as it correctly compared against "false".

  Use the following program to test it:

  import sys
  from onnx2tf.utils.common_functions import replace_parameter
  
  result = replace_parameter(
      value_before_replacement=False,
      param_target='test',
      param_name='test', 
      op_rep_params=[{'param_target': 'test', 'param_name': 'test', 'values': 'True'}]
  )
  
  print(f'Output: {result} (type: {type(result).__name__})')
  

What's Changed

• Fix boolean parameter replacement by @emmanuel-ferdman in https://github.com/PINTO0309/onnx2tf/pull/775

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.28.0...1.28.1

• [Experimental] Added automatic JSON generation function

• Currently, this is a trial implementation and only supports a very limited number of OPs. If I feel like it, I will expand the OPs that are the target of JSON auto-generation.

• https://github.com/PINTO0309/onnx2tf/blob/main/AUTO_JSON_FEATURE_SUMMARY.md

  # Automatic JSON generation only
  # Generates an optimal parameter replacement JSON file for model conversion.
  # The JSON file is saved to {model_name}_auto.json when conversion errors occur
  # or accuracy issues are detected.
  onnx2tf -i model.onnx -agj
  
  # Accuracy validation only (no JSON generation)
  # Validates the accuracy between ONNX and TensorFlow outputs without generating
  # any parameter replacement JSON file.
  onnx2tf -i model.onnx -cotof
  
  # Accuracy validation + automatic JSON generation
  # First generates an optimal parameter replacement JSON file, then uses it
  # to validate the model accuracy. This ensures the best possible conversion accuracy.
  onnx2tf -i model.onnx -agj -cotof
  

  -agj, --auto_generate_json
    Automatically generates a parameter replacement JSON file that achieves minimal error
    when converting the model. This option explores various parameter combinations to find
    the best settings that result in successful conversion and highest accuracy.
    The search stops when the final output OP accuracy check shows "Matches".
    When used together with -cotof, the generated JSON is used to re-evaluate accuracy.
    WARNING: This option performs an exhaustive search to find the optimal conversion patterns,
    which can take a very long time depending on the model complexity.
  

• onnx2tf doesn't work after exporting PINTO0309/LightGlue-ONNX #772

What's Changed

• [Experimental] Added automatic JSON generation function by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/773

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.27.10...1.28.0

• common_functions.py
  • Fixed handling of dynamic tensors in --broadcast_for_gpu_delegate
• TFLite Conversion with GPU Delegate #768

What's Changed

• Fixed handling of dynamic tensors in broadcast_for_gpu_delegate by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/769

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.27.9...1.27.10

• onnx2tf.py, common_functions.py

• replace_yolov9t_dynamic.json

  • Support for shape_hints.

  • With the addition of this feature, conversion is now possible even when there are multiple dynamic dimensions other than batch size. It can be used mainly for debugging by arbitrarily changing the size of intermediate tensors.

  • You can now specify input tensors of any size for accuracy evaluation using the -cotof and -coto options.

  • The introduction of shape_hints is the first step toward implementing a feature that automatically corrects conversion errors in models with dynamic tensor inputs within the onnx2tf processing.

  • In future versions, I plan to make the --shape_hints option mandatory when specifying models that use dynamic tensors as input.

    • CLI

      -sh SHAPE_HINTS [SHAPE_HINTS ...], \
          --shape_hints SHAPE_HINTS [SHAPE_HINTS ...]
        Shape hints for input tensors containing dynamic dimensions.
        Specify input shapes for test inference with -cotof or -coto.
        Unlike `--overwrite_input_shape`, this operation does not overwrite
        the ONNX input shape with a static shape.
        The format is
        "i1:dim0,...,dimN" "i2:dim0,...,dimN" "i3:dim0,...,dimN"
        When there is only one input, for example,
        "data:1,3,224,224"
        When there are multiple inputs, for example,
        "data1:1,3,224,224" "data2:1,3,112" "data3:5"
        A value of 1 or more must be specified.
        Numerical values other than dynamic dimensions are ignored.
      

    • In-Script

      shape_hints: Optional[List[str]]
        Shape hints for input tensors containing dynamic dimensions.
        Specify input shapes for test inference with -cotof or -coto.
        Unlike `--overwrite_input_shape`, this operation does not overwrite
        the ONNX input shape with a static shape.
        The format is
        ['i1:dim0,...,dimN', 'i2:dim0,...,dimN', 'i3:dim0,...,dimN']
        When there is only one input, for example,
        ['data:1,3,224,224']
        When there are multiple inputs, for example,
        ['data1:1,3,224,224', 'data2:1,3,112', 'data3:5']
        A value of 1 or more must be specified.
        Numerical values other than dynamic dimensions are ignored.
      

  • Example of use with YOLOv9-T [N, 3, H, W]

    • yolov9_t_wholebody28_Nx3HxW.onnx.zip <details><summary>JSON <b>replace_yolov9t_dynamic.json</b></summary>

      {
        "format_version": 1,
        "operations": [
          {
            "op_name": "/model.22/Gather",
            "param_target": "inputs",
            "param_name": "/model.22/Constant_output_0",
            "values": 0
          },
      
      
      
          {
            "op_name": "/model.22/Gather_3",
            "param_target": "inputs",
            "param_name": "/model.22/Constant_6_output_0",
            "values": 1
          },
          {
            "op_name": "/model.22/Gather_4",
            "param_target": "inputs",
            "param_name": "/model.22/Constant_7_output_0",
            "values": 2
          },
          {
            "op_name": "/model.22/Gather_5",
            "param_target": "inputs",
            "param_name": "/model.22/Constant_6_output_0",
            "values": 1
          },
          {
            "op_name": "/model.22/Gather_6",
            "param_target": "inputs",
            "param_name": "/model.22/Constant_7_output_0",
            "values": 2
          },
          {
            "op_name": "/model.22/Gather_1",
            "param_target": "inputs",
            "param_name": "/model.22/Constant_6_output_0",
            "values": 1
          },
          {
            "op_name": "/model.22/Gather_2",
            "param_target": "inputs",
            "param_name": "/model.22/Constant_7_output_0",
            "values": 2
          },
      
      
      
          {
            "op_name": "/model.22/Concat_23",
            "param_target": "attributes",
            "param_name": "axis",
            "values": 2
          },
          {
            "op_name": "/model.22/Slice",
            "param_target": "op",
            "begin": [0,0,0],
            "end": [0,64,0],
            "end_mask": 5
          },
          {
            "op_name": "/model.22/dfl/Gather",
            "param_target": "inputs",
            "param_name": "/model.22/Constant_output_0",
            "values": 0
          },
          {
            "op_name": "/model.22/dfl/Gather_1",
            "param_target": "inputs",
            "param_name": "/model.22/Constant_6_output_0",
            "values": 2
          },
      
      
      
          {
            "op_name": "/model.22/dfl/Reshape",
            "param_target": "inputs",
            "param_name": "/model.22/dfl/Concat_output_0",
            "pre_process_transpose_perm": [0,2,1]
          },
          {
            "op_name": "/model.22/dfl/Reshape",
            "param_target": "inputs",
            "param_name": "/model.22/Slice_output_0",
            "pre_process_transpose_perm": [0,2,1]
          },
      
      
          {
            "op_name": "/model.22/dfl/Transpose",
            "param_target": "attributes",
            "param_name": "perm",
            "values": [0,2,1,3]
          },
      
          {
            "op_name": "/model.22/dfl/Softmax",
            "param_target": "attributes",
            "param_name": "axis",
            "values": 1
          },
          {
            "op_name": "/model.22/dfl/Softmax",
            "param_target": "outputs",
            "param_name": "/model.22/dfl/Softmax_output_0",
            "post_process_transpose_perm": [0,2,3,1]
          },
      
      
      
      
          {
            "op_name": "/model.22/Slice_1",
            "param_target": "op",
            "begin": [0,64,0],
            "end": [0,92,0],
            "end_mask": 5
          },
      
      
          {
            "op_name": "/model.22/Slice_2",
            "param_target": "op",
            "begin": [0,0,0],
            "end": [0,2,0],
            "end_mask": 5
          },
          {
            "op_name": "/model.22/Slice_3",
            "param_target": "op",
            "begin": [0,2,0],
            "end": [0,3,0],
            "end_mask": 5
          },
      
      
          {
            "op_name": "/model.22/Mul_3",
            "param_target": "inputs",
            "param_name": "/model.22/Transpose_1_output_0",
            "pre_process_transpose_perm": [1,0]
          },
          {
            "op_name": "/model.22/Concat_24",
            "param_target": "attributes",
            "param_name": "axis",
            "values": 1
          }
        ]
      }
      

    </details>

  • convert

    onnx2tf \
    -i yolov9_t_wholebody28_Nx3HxW.onnx \
    -coion \
    -cotof \
    -sh "images:1,3,480,640" \
    -prf replace_yolov9t_dynamic.json \
    

  • results yolov9_t_wholebody28_Nx3HxW_float32.tflite.zip

• FlexConv2d generated due to dynamic input shape? #759

What's Changed

• Support for shape_hints by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/767

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.27.8...1.27.9

• AveragePool

  • [Experimental] Dynamic AveragePool support

• Tests https://github.com/PINTO0309/onnx2tf/releases/download/1.27.1/rec.onnx

  • replace_rec.json

    {
      "model": "https://github.com/PINTO0309/onnx2tf/releases/download/1.27.1/rec.onnx",
      "issue": "https://github.com/PINTO0309/onnx2tf/issues/747",
      "command": "onnx2tf -i rec.onnx -prf replace_rec.json",
      "format_version": 1,
      "operations": [
        {
          "op_name": "p2o.Transpose.2",
          "param_target": "attributes",
          "param_name": "perm",
          "values": [0,1,3,2]
        },
        {
          "op_name": "p2o.Transpose.5",
          "param_target": "attributes",
          "param_name": "perm",
          "values": [0,1,3,2]
        },
        {
          "op_name": "p2o.Transpose.7",
          "param_target": "attributes",
          "param_name": "perm",
          "values": [0,1,2,3]
        }
      ]
    }
    

  • Run

    onnx2tf -i rec.onnx -prf replace_rec.json -coion
    

• Results rec_float32.tflite.zip

• paddleocr model conversion to tflite fails #747

What's Changed

• Dynamic AveragePool support by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/766

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.27.7...1.27.8

• onnx2tf.py, common_functions.py
  • Improved log display during dummy inference
    1. Remove display of tensorflow output OP names
    2. Add display of onnx and tensorflow tensor shapes when Skipped (Deleted or Shape Unmatched) occurs.
    3. Remove wa/.
  • This is mainly for debugging purposes when converting dynamic tensors.

• FlexConv2d generated due to dynamic input shape? #759
• Unable to convert RTMDet-ins to tflite #761

What's Changed

• Improved log display during dummy inference by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/765

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.27.6...1.27.7

• onnx2tf.py
  • Bug fix for parameter substitution processing.
• Unable to convert RTMDet-ins to tflite #761

What's Changed

• Bug fix for parameter substitution processing by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/764

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.27.5...1.27.6

• Unsqueeze, Slice
  • Supports individual parameter replacement for Unsqueeze and Slice
  • e.g.

    {
      "format_version": 1,
      "operations": [
        {
          "op_name": "Concat_551",
          "param_target": "attributes",
          "param_name": "axis",
          "values": 2
        },
        {
          "op_name": "Gather_560",
          "param_target": "attributes",
          "param_name": "axis",
          "values": 2
        },
        {
          "op_name": "Unsqueeze_655",
          "param_target": "attributes",
          "param_name": "axes",
          "values": [1]
        },
        {
          "op_name": "Slice_657",
          "param_target": "inputs",
          "param_name": "2062",
          "values": [0]
        },
        {
          "op_name": "Slice_657",
          "param_target": "inputs",
          "param_name": "2029",
          "values": [2]
        },
        {
          "op_name": "Slice_657",
          "param_target": "inputs",
          "param_name": "1538",
          "values": [3]
        },
        {
          "op_name": "Slice_657",
          "param_target": "inputs",
          "param_name": "2015",
          "values": [1]
        }
      ]
    }
    

• Unable to convert RTMDet-ins to tflite #761

What's Changed

• Fix unsqueeze prf by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/763

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.27.4...1.27.5

• Reshape, Resize
  • Improved stability when handling tensors of any undefined shape.
  • However, for operations where the shape changes dynamically, full automatic conversion with onnx2tf is not successful, so manual adjustment of the conversion operation is required. NonMaxSuppression, NonZero, If, TopK https://github.com/PINTO0309/onnx2tf#parameter-replacement
  • First of all, the post-processing of mmdetection is catastrophically redundant, so it is highly discouraged to use the official implementation as is.

ValueError: Exception encountered when calling layer 'tf.math.subtract_2' (type TFOpLambda).

Dimensions must be equal, but are 4 and 2 for '{{node model_79/tf.math.subtract_2/Sub}} = Sub[T=DT_FLOAT](model_79/tf.strided_slice_84/StridedSlice, model_79/tf.tile/Tile)' with input shapes: [1,5,1,4], [1,1,6400,2].

Call arguments received by layer 'tf.math.subtract_2' (type TFOpLambda):
  • x=tf.Tensor(shape=(1, 5, 1, 4), dtype=float32)
  • y=tf.Tensor(shape=(1, 1, 6400, 2), dtype=float32)

• Unable to convert RTMDet-ins to tflite #761
• similar issues: https://github.com/PINTO0309/onnx2tf/issues?q=label%3A%22Dynamic%20batch%20%2F%20Dynamic%20shape%22

What's Changed

• Improved stability when handling tensors of any undefined shape. Reshape, Resize by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/762

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.27.3...1.27.4

• Conv, GroupConvolution
  • Delete warning message
  • Since Keras v2 supports output of saved_model containing GroupConvolution, the warning message is no longer necessary.
• FlexConv2d generated due to dynamic input shape? #759

What's Changed

• Delete GroupConv warning by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/760

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.27.2...1.27.3

• Split
  • Split now supports input of dynamically sized tensors, improving conversion stability
  • Fixed a problem that occurred when using a combination of dynamic batches and Split.
  • yolo11x-obb for oriented bounding boxes
  • yolo11x-obb-dynamic.onnx
• Cannot convert Yolo11x-obb to saved_model with dynamic batch size #751

What's Changed

• Split now supports input of dynamically sized tensors, improving conversion stability by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/753

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.27.1...1.27.2

• Add, Div, Mod, Mul, Sub, Cast, MatMul
  • Improved conversion stability for models exported with torch.autocast("cuda", dtype=torch.float16)
  • deim_hgnetv2_n_wholebody28_1250query_fp16.onnx.zip

What's Changed

• Improved conversion stability for models exported with torch.autocast("cuda", dtype=torch.float16) by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/746

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.27.0...1.27.1

• from tensorflow==2.17.0 to tensorflow==2.19.0
• from tf-keras~=2.16 to tf-keras==2.19.0
• from tf.lite.Interpreter() to ai_edge_litert.interpreter.Interpreter()
  • https://github.com/tensorflow/tensorflow/releases/tag/v2.19.0
• requirements
  • ai-edge-litert==1.2.0

What's Changed

• tf_2.19.0 by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/745

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.26.9...1.27.0

• ReduceL1, ReduceL2
  • Fixed the logic to avoid a TensorFlow bug that caused the output tensor of tf.norm() to be (None, None, None, None).
  • TensorFlow bugs when converting a model that includes ReduceL1 or ReduceL2.

    INFO: 20 / 819
    INFO: onnx_op_type: ReduceL2 onnx_op_name: wa/model/stages/stages.0/blocks/blocks.0/mlp/grn/ReduceL2
    INFO:  input_name.1: wa/model/stages/stages.0/blocks/blocks.0/mlp/act/Mul_1_output_0 shape: [1, 56, 56, 512] dtype: float32
    INFO:  output_name.1: wa/model/stages/stages.0/blocks/blocks.0/mlp/grn/ReduceL2_output_0 shape: [1, 1, 1, 512] dtype: float32
    INFO: tf_op_type: l2_normalize
    INFO:  input.1.x: name: tf.math.multiply_8/Mul:0 shape: (1, 56, 56, 512) dtype: <dtype: 'float32'> 
    INFO:  input.2.axis: val: [1, 2] 
    INFO:  output.1.output: name: tf.compat.v1.norm_6/norm/transpose_1:0 shape: (None, None, None, None) dtype: <dtype: 'float32'> 
    

• convnextv2_base timm model conversion fails #742

What's Changed

• Fixed the logic to avoid a TensorFlow bug that caused the output tensor of tf.norm() to be (None, None, None, None) by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/743

Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.26.8...1.26.9