onnx2tf 1.28.8__py3-none-any.whl → 1.29.0__py3-none-any.whl

onnx2tf/__init__.py

@@ -1,3 +1,3 @@
 from onnx2tf.onnx2tf import convert, main
 
-__version__ = '1.28.8'
+__version__ = '1.29.0'

onnx2tf/onnx2tf.py

@@ -99,6 +99,8 @@ def convert(
     enable_rnn_unroll: Optional[bool] = False,
     disable_suppression_flextranspose: Optional[bool] = False,
     disable_strict_mode: Optional[bool] = False,
+    onnxruntime_output_memmap: Optional[bool] = True,
+    onnxruntime_output_memmap_dir: Optional[str] = None,
     number_of_dimensions_after_flextranspose_compression: Optional[int] = 6,
     disable_suppression_flexstridedslice: Optional[bool] = False,
     number_of_dimensions_after_flexstridedslice_compression: Optional[int] = 5,
@@ -372,6 +374,15 @@ def convert(
         correction process is skipped, but the frequency of transposition errors increases\n
         and accuracy errors are more likely to occur. Strict mode is enabled by default.
 
+    onnxruntime_output_memmap: Optional[bool]
+        Use onnxruntime IOBinding with np.memmap for dummy inference outputs when\n
+        the estimated output tensor size exceeds available RAM. This avoids OOM\n
+        but increases disk I/O and may slow down validation.
+
+    onnxruntime_output_memmap_dir: Optional[str]
+        Directory for memmap files used by onnxruntime_output_memmap.\n
+        If omitted, a temporary directory is created and removed on exit.
+
     number_of_dimensions_after_flextranspose_compression: Optional[int]
         Number of Transpose OP dimensions generated after avoiding FlexTranspose generation.\n
         Also suppress the creation of the Transpose itself by specifying 2.\n
@@ -1118,6 +1129,8 @@ def convert(
                     tf_layers_dict=tf_layers_dict,
                     use_cuda=use_cuda,
                     disable_strict_mode=disable_strict_mode,
+                    enable_ort_output_memmap=onnxruntime_output_memmap,
+                    ort_output_memmap_dir=onnxruntime_output_memmap_dir,
                     shape_hints=shape_hints if (check_onnx_tf_outputs_elementwise_close or check_onnx_tf_outputs_elementwise_close_full) else None,
                 )
             """
@@ -2041,6 +2054,8 @@ def convert(
                         custom_input_op_name_np_data_path=custom_input_op_name_np_data_path,
                         tf_layers_dict=tf_layers_dict,
                         use_cuda=use_cuda,
+                        enable_ort_output_memmap=onnxruntime_output_memmap,
+                        ort_output_memmap_dir=onnxruntime_output_memmap_dir,
                         shape_hints=shape_hints,
                     )
             except Exception as ex:
@@ -2304,6 +2319,8 @@ def convert(
                             custom_input_op_name_np_data_path=custom_input_op_name_np_data_path,
                             tf_layers_dict=tf_layers_dict,
                             use_cuda=use_cuda,
+                            enable_ort_output_memmap=onnxruntime_output_memmap,
+                            ort_output_memmap_dir=onnxruntime_output_memmap_dir,
                             shape_hints=shape_hints,
                         )
 
@@ -2836,6 +2853,28 @@ def main():
             'correction process is skipped, but the frequency of transposition errors increases \n' +
             'and accuracy errors are more likely to occur. Strict mode is enabled by default.'
     )
+    parser.add_argument(
+        '-doem',
+        '--disable_onnxruntime_output_memmap',
+        dest='disable_onnxruntime_output_memmap',
+        action='store_true',
+        help=\
+            'Disable onnxruntime output memmap. \n' +
+            'By default, onnx2tf uses onnxruntime IOBinding with np.memmap for dummy inference \n' +
+            'outputs only when the estimated output tensor size exceeds available RAM. \n' +
+            'Use this flag to force the standard in-memory output path instead. \n' +
+            'Default: disabled (memmap enabled when needed).'
+    )
+    parser.set_defaults(disable_onnxruntime_output_memmap=False)
+    parser.add_argument(
+        '-oemd',
+        '--onnxruntime_output_memmap_dir',
+        type=str,
+        help=\
+            'Directory for memmap files used by onnxruntime output memmap. \n' +
+            'If omitted, a temporary directory is created and removed on exit. \n' +
+            'This setting is used only when memmap is actually enabled.'
+    )
     parser.add_argument(
         '-nodafsc',
         '--number_of_dimensions_after_flexstridedslice_compression',
@@ -3133,6 +3172,8 @@ def main():
         enable_rnn_unroll=args.enable_rnn_unroll,
         disable_suppression_flextranspose=args.disable_suppression_flextranspose,
         disable_strict_mode=args.disable_strict_mode,
+        onnxruntime_output_memmap=not args.disable_onnxruntime_output_memmap,
+        onnxruntime_output_memmap_dir=args.onnxruntime_output_memmap_dir,
         number_of_dimensions_after_flextranspose_compression=args.number_of_dimensions_after_flextranspose_compression,
         disable_suppression_flexstridedslice=args.disable_suppression_flexstridedslice,
         number_of_dimensions_after_flexstridedslice_compression=args.number_of_dimensions_after_flexstridedslice_compression,

onnx2tf/utils/common_functions.py

@@ -7,6 +7,9 @@ import copy
 import json
 import psutil
 import random
+import atexit
+import tempfile
+import shutil
 random.seed(0)
 import requests
 import flatbuffers
@@ -3634,6 +3637,8 @@ def dummy_onnx_inference(
     tf_layers_dict: Optional[Dict] = None,
     use_cuda: bool = False,
     disable_strict_mode: bool = False,
+    enable_ort_output_memmap: bool = False,
+    ort_output_memmap_dir: Optional[str] = None,
     shape_hints: Optional[List[str]] = None,
 ) -> List[np.ndarray]:
     """Perform inference on ONNX subgraphs with an all-1 dummy tensor.
@@ -3663,6 +3668,14 @@ def dummy_onnx_inference(
     disable_strict_mode: Optional[bool]
         True to disable strict inference mode, False to enable it.
 
+    enable_ort_output_memmap: bool
+        True to use onnxruntime IOBinding with np.memmap for outputs when
+        output tensors are too large for available RAM.
+
+    ort_output_memmap_dir: Optional[str]
+        Directory to store memmap files. If not specified, a temporary
+        directory is created and removed on exit.
+
     Returns
     ----------
     outputs: List[np.ndarray]
@@ -3880,7 +3893,7 @@ def dummy_onnx_inference(
         op_output_size: int = 1
         if gs_graph_output.shape is not None:
             for s in gs_graph_output.shape:
-                if isinstance(s, int):
+                if isinstance(s, (int, np.integer)):
                     op_output_size *= s
             # Total bytes
             total_output_size += op_output_size * dtype_sizes.get(gs_graph_output.dtype, 4)
@@ -3888,7 +3901,8 @@ def dummy_onnx_inference(
     # When exact inference mode is enabled and the total size of the tensor of inference results exceeds approximately 80% of available RAM
     mem_available = psutil.virtual_memory().available * 0.80 // 1024 // 1024 //1024
     total_output_size_gb = (total_output_size // 1024 // 1024 //1024)
-    if (not disable_strict_mode and total_output_size_gb > mem_available):
+    use_memmap_outputs = enable_ort_output_memmap and total_output_size_gb > mem_available
+    if (not disable_strict_mode and total_output_size_gb > mem_available and not use_memmap_outputs):
         if tmp_onnx_path:
             os.remove(tmp_onnx_path)
             os.remove(tmp_onnx_external_weights_path)
@@ -3896,7 +3910,94 @@ def dummy_onnx_inference(
             f'The tool skipped dummy inference to avoid SWAP processing because the total size of the tensor of inference results exceeded about {mem_available} GB. (results: {total_output_size_gb} GB)'
         )
 
-    outputs = onnx_session.run(None, input_datas)
+    if use_memmap_outputs:
+        output_shapes = []
+        output_names_order = [out.name for out in gs_graph.outputs]
+        for out in gs_graph.outputs:
+            shape = out.shape
+            if shape is None or any(not isinstance(s, (int, np.integer)) for s in shape):
+                if tmp_onnx_path:
+                    os.remove(tmp_onnx_path)
+                    os.remove(tmp_onnx_external_weights_path)
+                raise Exception(
+                    'onnxruntime output memmap requires static output shapes. ' +
+                    'Provide --shape_hints or reduce validation outputs.'
+                )
+            output_shapes.append([int(s) for s in shape])
+
+        memmap_dir = ort_output_memmap_dir
+        cleanup_memmap_dir = False
+        if memmap_dir is None:
+            memmap_dir = tempfile.mkdtemp(prefix='onnx2tf_ort_mm_')
+            cleanup_memmap_dir = True
+        os.makedirs(memmap_dir, exist_ok=True)
+
+        try:
+            disk_free = psutil.disk_usage(memmap_dir).free
+            if total_output_size > disk_free:
+                raise Exception(
+                    f'Not enough disk space for memmap outputs. ' +
+                    f'Required: {total_output_size} bytes, Free: {disk_free} bytes.'
+                )
+        except Exception as ex:
+            if 'Not enough disk space' in str(ex):
+                if tmp_onnx_path:
+                    os.remove(tmp_onnx_path)
+                    os.remove(tmp_onnx_external_weights_path)
+                raise
+
+        if cleanup_memmap_dir:
+            atexit.register(shutil.rmtree, memmap_dir, ignore_errors=True)
+
+        info(
+            f'onnxruntime output memmap enabled. ' +
+            f'Outputs: {len(output_names_order)}, Path: {memmap_dir}'
+        )
+
+        io_binding = onnx_session.io_binding()
+
+        for input_name, input_data in input_datas.items():
+            if not input_data.flags['C_CONTIGUOUS']:
+                input_data = np.ascontiguousarray(input_data)
+                input_datas[input_name] = input_data
+            io_binding.bind_input(
+                input_name,
+                'cpu',
+                0,
+                input_data.dtype,
+                input_data.shape,
+                input_data.__array_interface__['data'][0],
+            )
+
+        memmap_outputs = {}
+        for idx, (output_name, output_shape) in enumerate(zip(output_names_order, output_shapes)):
+            safe_output_name = re.sub(r'[^0-9A-Za-z._-]+', '_', output_name)
+            memmap_path = os.path.join(memmap_dir, f'ort_output_{idx}_{safe_output_name}.mmap')
+            output_dtype = np.dtype(gs_graph.outputs[idx].dtype)
+            memmap_array = np.memmap(
+                memmap_path,
+                dtype=output_dtype,
+                mode='w+',
+                shape=tuple(output_shape),
+            )
+            memmap_outputs[output_name] = memmap_array
+            io_binding.bind_output(
+                output_name,
+                'cpu',
+                0,
+                output_dtype,
+                output_shape,
+                memmap_array.__array_interface__['data'][0],
+            )
+
+        onnx_session.run_with_iobinding(io_binding)
+        io_binding.synchronize_outputs()
+        for memmap_array in memmap_outputs.values():
+            memmap_array.flush()
+
+        outputs = [memmap_outputs[name] for name in output_names_order]
+    else:
+        outputs = onnx_session.run(None, input_datas)
     if tmp_onnx_path:
         os.remove(tmp_onnx_path)
         os.remove(tmp_onnx_external_weights_path)

{onnx2tf-1.28.8.dist-info → onnx2tf-1.29.0.dist-info}/METADATA

@@ -1,27 +1,38 @@
 Metadata-Version: 2.4
 Name: onnx2tf
-Version: 1.28.8
-Summary: Self-Created Tools to convert ONNX files (NCHW) to TensorFlow/TFLite/Keras format (NHWC). The purpose of this tool is to solve the massive Transpose extrapolation problem in onnx-tensorflow (onnx-tf).
+Version: 1.29.0
+Summary: Self-Created Tools to convert ONNX files (NCHW) to TensorFlow/TFLite/Keras format (NHWC).
 Home-page: https://github.com/PINTO0309/onnx2tf
 Author: Katsuya Hyodo
 Author-email: rmsdh122@yahoo.co.jp
-License: MIT License
 Platform: linux
 Platform: unix
 Requires-Python: >=3.10
 Description-Content-Type: text/markdown
 License-File: LICENSE
 License-File: LICENSE_onnx-tensorflow
+Requires-Dist: requests==2.32.5
+Requires-Dist: numpy==1.26.4
+Requires-Dist: onnx==1.19.0
+Requires-Dist: onnxruntime==1.23.0
+Requires-Dist: opencv-python==4.11.0.86
+Requires-Dist: onnxsim==0.4.30
+Requires-Dist: ai-edge-litert==2.1.0
+Requires-Dist: tensorflow==2.19.0
+Requires-Dist: tf-keras==2.19.0
+Requires-Dist: onnx-graphsurgeon==0.5.8
+Requires-Dist: simple-onnx-processing-tools==1.1.32
+Requires-Dist: psutil==5.9.5
+Requires-Dist: protobuf==4.25.5
+Requires-Dist: h5py==3.11.0
+Requires-Dist: ml_dtypes==0.5.1
+Requires-Dist: flatbuffers==25.12.19
 Dynamic: author
 Dynamic: author-email
-Dynamic: description
-Dynamic: description-content-type
 Dynamic: home-page
-Dynamic: license
 Dynamic: license-file
 Dynamic: platform
 Dynamic: requires-python
-Dynamic: summary
 
 # onnx2tf
 Self-Created Tools to convert ONNX files (NCHW) to TensorFlow/TFLite/Keras format (NHWC). The purpose of this tool is to solve the massive Transpose extrapolation problem in [onnx-tensorflow](https://github.com/onnx/onnx-tensorflow) ([onnx-tf](https://pypi.org/project/onnx-tf/)). I don't need a Star, but give me a pull request. Since I am adding challenging model optimizations and fixing bugs almost daily, I frequently embed potential bugs that would otherwise break through CI's regression testing. Therefore, if you encounter new problems, I recommend that you try a package that is a few versions older, or try the latest package that will be released in a few days.
@@ -334,7 +345,7 @@ Video speed is adjusted approximately 50 times slower than actual speed.
   docker run --rm -it \
   -v `pwd`:/workdir \
   -w /workdir \
-  ghcr.io/pinto0309/onnx2tf:1.28.8
+  ghcr.io/pinto0309/onnx2tf:1.29.0
 
   or
 
@@ -342,7 +353,7 @@ Video speed is adjusted approximately 50 times slower than actual speed.
   docker run --rm -it \
   -v `pwd`:/workdir \
   -w /workdir \
-  docker.io/pinto0309/onnx2tf:1.28.8
+  docker.io/pinto0309/onnx2tf:1.29.0
 
   or
 

{onnx2tf-1.28.8.dist-info → onnx2tf-1.29.0.dist-info}/RECORD

@@ -1,6 +1,6 @@
-onnx2tf/__init__.py,sha256=PpTfjKEZ7PWIDHIvtSVU9Epc3rkkmE02EOIWvlHFnWU,66
+onnx2tf/__init__.py,sha256=DObbckS1XgiAp4svAwPuHhkeAswQOzsYLXEyNPN7vwY,66
 onnx2tf/__main__.py,sha256=2RSCQ7d4lc6CwD-rlGn9UicPFg-P5du7ZD_yh-kuBEU,57
-onnx2tf/onnx2tf.py,sha256=v6dx8ZQrj2h4-xv3lae4wkdts1NO3tEINS_sK0dykvw,148911
+onnx2tf/onnx2tf.py,sha256=XZBqVn1Q_qmPmrYw_Dz30vRWJ8uaJURkbUSo_8rZrjk,151116
 onnx2tf/ops/Abs.py,sha256=V7btmCG_ZvK_qJovUsguq0ZMJ349mhNQ4FHSgzP_Yuo,4029
 onnx2tf/ops/Acos.py,sha256=Fo8YkFKuWq8Fi2xUrBdKcAH1yJ8r5pjSD0wgLttTNdk,4003
 onnx2tf/ops/Acosh.py,sha256=ATQj2cT5JS_mTfXi0kXqJ1yzSZu5J0zHA5VjV3j7uKY,3588
@@ -185,15 +185,14 @@ onnx2tf/ops/_Loop.py,sha256=eo5sNfrfOnKV6_I737AWsM5LJTY9DVOxQEvhanxtP4g,11322
 onnx2tf/ops/__Loop.py,sha256=ClwMcbNS4hqUtW_pzwjMa9Cqg7ONvz9aplke55A0uJ0,19704
 onnx2tf/ops/__init__.py,sha256=jnmUWWa-3dHzBZV9bmPzXu6eoz2dumJTzO7i8JdcgSM,25
 onnx2tf/utils/__init__.py,sha256=E9FM9He68VIASDnYp-OrxvHFVn55GzWqw2OEkCqn1zg,27
-onnx2tf/utils/common_functions.py,sha256=_GPrqIz4Ueg3hriOulErfoNGhoMw_5OKiWqLTYZh6YI,245622
+onnx2tf/utils/common_functions.py,sha256=o9a4g56OdQKocODzBp2Uxesves_Tl-Iizh5r4Okmu6Q,249631
 onnx2tf/utils/enums.py,sha256=7c5TqetqB07VjyHoxJHfLgtqBqk9ZRyUF33fPOJR1IM,1649
 onnx2tf/utils/iterative_json_optimizer.py,sha256=qqeIxWGxrhcCYk8-ebWnblnOkzDCwi-nseipHzHR_bk,10436
 onnx2tf/utils/json_auto_generator.py,sha256=C-4P8sii7B2_LRwW6AFG13tI3R5DHPhNirgxfmBU1F8,61944
 onnx2tf/utils/logging.py,sha256=yUCmPuJ_XiUItM3sZMcaMO24JErkQy7zZwVTYWAuiKg,1982
-onnx2tf-1.28.8.dist-info/licenses/LICENSE,sha256=5v_Kxihy8i6mzHVl349ikSREaIdsl9YeUnX1KBDLD2w,1070
-onnx2tf-1.28.8.dist-info/licenses/LICENSE_onnx-tensorflow,sha256=gK4GtS9S5YcyINu6uuNNWdo-kBClyEM4MFLFGiNTeRM,11231
-onnx2tf-1.28.8.dist-info/METADATA,sha256=NKZj0tTAVIo80eNaKXJTxQDpRtl3-smFlcQCS3QpnMc,152862
-onnx2tf-1.28.8.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
-onnx2tf-1.28.8.dist-info/entry_points.txt,sha256=gDPK8ToCFPKMvm8jr9xrGOkXtORJJVh4736fBEKO5k0,41
-onnx2tf-1.28.8.dist-info/top_level.txt,sha256=WgfPiEy3f6vZ_FOpAIEA2CF3TCx1eYrhGw93Ih6b9Fw,8
-onnx2tf-1.28.8.dist-info/RECORD,,
+onnx2tf-1.29.0.dist-info/licenses/LICENSE,sha256=5v_Kxihy8i6mzHVl349ikSREaIdsl9YeUnX1KBDLD2w,1070
+onnx2tf-1.29.0.dist-info/licenses/LICENSE_onnx-tensorflow,sha256=gK4GtS9S5YcyINu6uuNNWdo-kBClyEM4MFLFGiNTeRM,11231
+onnx2tf-1.29.0.dist-info/METADATA,sha256=bqMWcqDQNEcmcnraz4IoNm7XGT3mARr-WvMnVlOdIaM,153189
+onnx2tf-1.29.0.dist-info/WHEEL,sha256=qELbo2s1Yzl39ZmrAibXA2jjPLUYfnVhUNTlyF1rq0Y,92
+onnx2tf-1.29.0.dist-info/top_level.txt,sha256=WgfPiEy3f6vZ_FOpAIEA2CF3TCx1eYrhGw93Ih6b9Fw,8
+onnx2tf-1.29.0.dist-info/RECORD,,

{onnx2tf-1.28.8.dist-info → onnx2tf-1.29.0.dist-info}/WHEEL

@@ -1,5 +1,5 @@
 Wheel-Version: 1.0
-Generator: setuptools (80.9.0)
+Generator: setuptools (80.10.1)
 Root-Is-Purelib: true
 Tag: py3-none-any
 

onnx2tf-1.28.8.dist-info/entry_points.txt

@@ -1,2 +0,0 @@
-[console_scripts]
-onnx2tf = onnx2tf:main