onnxruntime_extensions 0.11.0__cp39-cp39-win_amd64.whl → 0.13.0__cp39-cp39-win_amd64.whl

onnxruntime_extensions/__init__.py

@@ -10,7 +10,6 @@ This enables more flexibility and control over model execution, thus expanding t
 
 __author__ = "Microsoft"
 
-
 from ._version import __version__
 from ._ocos import get_library_path
 from ._ocos import Opdef, PyCustomOpDef
@@ -66,6 +65,10 @@ if _lib_only:
     gen_processing_models = _unimplemented
     OrtPyFunction = _unimplemented
     ort_inference = _unimplemented
+    PyOrtFunction = _unimplemented
+    optimize_model = _unimplemented
+    make_onnx_model = _unimplemented
+    ONNXRuntimeError = _unimplemented
 
 else:
     __all__ += _offline_api

onnxruntime_extensions/_hf_cvt.py

@@ -48,8 +48,9 @@ class HFTokenizerConverter(CustomOpConverter):
             model_dir = hf_tokenizer.name_or_path
         else:
             model_dir = os.path.dirname(vocab_file)
-        tokenizer_json = json.load(
-            open(os.path.join(model_dir, tokenizer_file), "r", encoding="utf-8"))
+        f = open(os.path.join(model_dir, tokenizer_file), "r", encoding="utf-8")
+        tokenizer_json = json.load(f)
+        f.close()
         # get vocab object from json file
         vocab = tokenizer_json.get("model", {}).get("vocab", {})
         sorted_merges = tokenizer_json.get("model", {}).get("merges", [])

onnxruntime_extensions/_torch_cvt.py

@@ -17,7 +17,7 @@ from onnx import numpy_helper
 from ._ortapi2 import make_onnx_model
 from ._cuops import SingleOpGraph
 from ._hf_cvt import HFTokenizerConverter
-from .util import remove_unused_initializers
+from .util import remove_unused_initializers, mel_filterbank
 
 
 class _WhisperHParams:
@@ -30,53 +30,15 @@ class _WhisperHParams:
     N_FRAMES = N_SAMPLES // HOP_LENGTH
 
 
-def _mel_filterbank(
-        n_fft: int, n_mels: int = 80, sr=16000, min_mel=0, max_mel=45.245640471924965, dtype=np.float32):
-    """
-    Compute a Mel-filterbank. The filters are stored in the rows, the columns,
-    and it is Slaney normalized mel-scale filterbank.
-    """
-    fbank = np.zeros((n_mels, n_fft // 2 + 1), dtype=dtype)
-
-    # the centers of the frequency bins for the DFT
-    freq_bins = np.fft.rfftfreq(n=n_fft, d=1.0 / sr)
-
-    mel = np.linspace(min_mel, max_mel, n_mels + 2)
-    # Fill in the linear scale
-    f_min = 0.0
-    f_sp = 200.0 / 3
-    freqs = f_min + f_sp * mel
-
-    # And now the nonlinear scale
-    min_log_hz = 1000.0  # beginning of log region (Hz)
-    min_log_mel = (min_log_hz - f_min) / f_sp  # same (Mels)
-    logstep = np.log(6.4) / 27.0  # step size for log region
-
-    log_t = mel >= min_log_mel
-    freqs[log_t] = min_log_hz * np.exp(logstep * (mel[log_t] - min_log_mel))
-    mel_bins = freqs
-
-    mel_spacing = np.diff(mel_bins)
-
-    ramps = mel_bins.reshape(-1, 1) - freq_bins.reshape(1, -1)
-    for i in range(n_mels):
-        left = -ramps[i] / mel_spacing[i]
-        right = ramps[i + 2] / mel_spacing[i + 1]
-
-        # intersect them with each other and zero
-        fbank[i] = np.maximum(0, np.minimum(left, right))
-
-    energy_norm = 2.0 / (mel_bins[2: n_mels + 2] - mel_bins[:n_mels])
-    fbank *= energy_norm[:, np.newaxis]
-    return fbank
-
-
 class CustomOpStftNorm(torch.autograd.Function):
     @staticmethod
     def symbolic(g, self, n_fft, hop_length, window):
-        t_n_fft = g.op('Constant', value_t=torch.tensor(n_fft, dtype=torch.int64))
-        t_hop_length = g.op('Constant', value_t=torch.tensor(hop_length, dtype=torch.int64))
-        t_frame_size = g.op('Constant', value_t=torch.tensor(n_fft, dtype=torch.int64))
+        t_n_fft = g.op('Constant', value_t=torch.tensor(
+            n_fft, dtype=torch.int64))
+        t_hop_length = g.op('Constant', value_t=torch.tensor(
+            hop_length, dtype=torch.int64))
+        t_frame_size = g.op(
+            'Constant', value_t=torch.tensor(n_fft, dtype=torch.int64))
         return g.op("ai.onnx.contrib::StftNorm", self, t_n_fft, t_hop_length, window, t_frame_size)
 
     @staticmethod
@@ -97,7 +59,7 @@ class WhisperPrePipeline(torch.nn.Module):
         self.n_fft = n_fft
         self.window = torch.hann_window(n_fft)
         self.mel_filters = torch.from_numpy(
-            _mel_filterbank(sr=sr, n_fft=n_fft, n_mels=n_mels))
+            mel_filterbank(sr=sr, n_fft=n_fft, n_mels=n_mels))
 
     def forward(self, audio_pcm: torch.Tensor):
         stft_norm = CustomOpStftNorm.apply(audio_pcm,
@@ -112,7 +74,8 @@ class WhisperPrePipeline(torch.nn.Module):
         spec_shape = log_spec.shape
         padding_spec = torch.ones(spec_shape[0],
                                   spec_shape[1],
-                                  self.n_samples // self.hop_length - spec_shape[2],
+                                  self.n_samples // self.hop_length -
+                                  spec_shape[2],
                                   dtype=torch.float)
         padding_spec *= spec_min
         log_spec = torch.cat((log_spec, padding_spec), dim=2)
@@ -165,15 +128,20 @@ def _to_onnx_stft(onnx_model, n_fft):
         make_node('Slice', inputs=['transpose_1_output_0', 'const_18_output_0', 'const_minus_1_output_0',
                                    'const_17_output_0', 'const_20_output_0'], outputs=['slice_1_output_0'],
                   name='slice_1'),
-        make_node('Constant', inputs=[], outputs=['const0_output_0'], name='const0', value_int=0),
-        make_node('Constant', inputs=[], outputs=['const1_output_0'], name='const1', value_int=1),
+        make_node('Constant', inputs=[], outputs=[
+                  'const0_output_0'], name='const0', value_int=0),
+        make_node('Constant', inputs=[], outputs=[
+                  'const1_output_0'], name='const1', value_int=1),
         make_node('Gather', inputs=['slice_1_output_0', 'const0_output_0'], outputs=['gather_4_output_0'],
                   name='gather_4', axis=3),
         make_node('Gather', inputs=['slice_1_output_0', 'const1_output_0'], outputs=['gather_5_output_0'],
                   name='gather_5', axis=3),
-        make_node('Mul', inputs=['gather_4_output_0', 'gather_4_output_0'], outputs=['mul_output_0'], name='mul0'),
-        make_node('Mul', inputs=['gather_5_output_0', 'gather_5_output_0'], outputs=['mul_1_output_0'], name='mul1'),
-        make_node('Add', inputs=['mul_output_0', 'mul_1_output_0'], outputs=[stft_norm_node.output[0]], name='add0'),
+        make_node('Mul', inputs=['gather_4_output_0', 'gather_4_output_0'], outputs=[
+                  'mul_output_0'], name='mul0'),
+        make_node('Mul', inputs=['gather_5_output_0', 'gather_5_output_0'], outputs=[
+                  'mul_1_output_0'], name='mul1'),
+        make_node('Add', inputs=['mul_output_0', 'mul_1_output_0'], outputs=[
+                  stft_norm_node.output[0]], name='add0'),
     ]
     new_stft_nodes.extend(onnx_model.graph.node[:node_idx])
     new_stft_nodes.extend(replaced_nodes)
@@ -253,9 +221,11 @@ class WhisperDataProcGraph:
         del g.node[:]
         g.node.extend(nodes)
 
-        inputs = [onnx.helper.make_tensor_value_info("sequences", onnx.TensorProto.INT32, ['N', 'seq_len', 'ids'])]
+        inputs = [onnx.helper.make_tensor_value_info(
+            "sequences", onnx.TensorProto.INT32, ['N', 'seq_len', 'ids'])]
         del g.input[:]
         g.input.extend(inputs)
-        g.output[0].type.CopyFrom(onnx.helper.make_tensor_type_proto(onnx.TensorProto.STRING, ['N', 'text']))
+        g.output[0].type.CopyFrom(onnx.helper.make_tensor_type_proto(
+            onnx.TensorProto.STRING, ['N', 'text']))
 
         return make_onnx_model(g, opset_version=self.opset_version)

onnxruntime_extensions/_version.py

@@ -1,2 +1,2 @@
 # Generated by setup.py, DON'T MANUALLY UPDATE IT!
-__version__ = "0.11.0"
+__version__ = "0.13.0"

onnxruntime_extensions/pp_api.py

@@ -0,0 +1,81 @@
+# Copyright (c) Microsoft Corporation. All rights reserved.
+# Licensed under the MIT License. See License.txt in the project root for
+# license information.
+###############################################################################
+
+import os
+from . import _extensions_pydll as _C
+if not hasattr(_C, "delete_object"):
+    raise ImportError(
+        "onnxruntime_extensions is not built with pre-processing C API\n"
+        "To enable it, please build the package with --ortx-user-option=pp_api")
+
+create_processor = _C.create_processor
+load_images = _C.load_images
+image_pre_process = _C.image_pre_process
+tensor_result_get_at = _C.tensor_result_get_at
+
+create_tokenizer = _C.create_tokenizer
+batch_tokenize = _C.batch_tokenize
+batch_detokenize = _C.batch_detokenize
+
+delete_object = _C.delete_object
+
+
+class Tokenizer:
+    def __init__(self, tokenizer_dir):
+        self.tokenizer = None
+        if os.path.isdir(tokenizer_dir):
+            self.tokenizer = create_tokenizer(tokenizer_dir)
+        else:
+            try:
+                from transformers.utils import cached_file
+                resolved_full_file = cached_file(
+                    tokenizer_dir, "tokenizer.json")
+                resolved_config_file = cached_file(
+                    tokenizer_dir, "tokenizer_config.json")
+            except ImportError:
+                raise ValueError(
+                    f"Directory '{tokenizer_dir}' not found and transformers is not available")
+            if not os.path.exists(resolved_full_file):
+                raise FileNotFoundError(
+                    f"Downloaded HF file '{resolved_full_file}' cannot be found")
+            if (os.path.dirname(resolved_full_file) != os.path.dirname(resolved_config_file)):
+                raise FileNotFoundError(
+                    f"Downloaded HF files '{resolved_full_file}' "
+                    f"and '{resolved_config_file}' are not in the same directory")
+
+            tokenizer_dir = os.path.dirname(resolved_full_file)
+            self.tokenizer = create_tokenizer(tokenizer_dir)
+
+    def tokenize(self, text):
+        return batch_tokenize(self.tokenizer, [text])[0]
+
+    def detokenize(self, tokens):
+        return batch_detokenize(self.tokenizer, [tokens])[0]
+
+    def __del__(self):
+        if delete_object and self.tokenizer:
+            delete_object(self.tokenizer)
+        self.tokenizer = None
+
+
+class ImageProcessor:
+    def __init__(self, processor_json):
+        self.processor = create_processor(processor_json)
+
+    def pre_process(self, images):
+        if isinstance(images, str):
+            images = [images]
+        if isinstance(images, list):
+            images = load_images(images)
+        return image_pre_process(self.processor, images)
+
+    @staticmethod
+    def to_numpy(result):
+        return tensor_result_get_at(result, 0)
+
+    def __del__(self):
+        if delete_object and self.processor:
+            delete_object(self.processor)
+        self.processor = None

{onnxruntime_extensions-0.11.0.dist-info → onnxruntime_extensions-0.13.0.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: onnxruntime_extensions
-Version: 0.11.0
+Version: 0.13.0
 Summary: ONNXRuntime Extensions
 Home-page: https://github.com/microsoft/onnxruntime-extensions
 Author: Microsoft Corporation
@@ -25,29 +25,17 @@ License-File: LICENSE
 
 ## What's ONNXRuntime-Extensions
 
-Introduction: ONNXRuntime-Extensions is a library that extends the capability of the ONNX models and inference with ONNX Runtime, via ONNX Runtime Custom Operator ABIs. It includes a set of [ONNX Runtime Custom Operator](https://onnxruntime.ai/docs/reference/operators/add-custom-op.html) to support the common pre- and post-processing operators for vision, text, and nlp models. And it supports multiple languages and platforms, like Python on Windows/Linux/macOS, some mobile platforms like Android and iOS, and Web-Assembly etc. The basic workflow is to enhance a ONNX model firstly and then do the model inference with ONNX Runtime and ONNXRuntime-Extensions package.
+Introduction: ONNXRuntime-Extensions is a C/C++ library that extends the capability of the ONNX models and inference with ONNX Runtime, via ONNX Runtime Custom Operator ABIs. It includes a set of [ONNX Runtime Custom Operator](https://onnxruntime.ai/docs/reference/operators/add-custom-op.html) to support the common pre- and post-processing operators for vision, text, and nlp models. And it supports multiple languages and platforms, like Python on Windows/Linux/macOS, some mobile platforms like Android and iOS, and Web-Assembly etc. The basic workflow is to enhance a ONNX model firstly and then do the model inference with ONNX Runtime and ONNXRuntime-Extensions package.
 
 
 ## Quickstart
+The library can be utilized as either a C/C++ library or other advance language packages like Python, Java, C#, etc. To build it as a shared library, you can use the `build.bat` or `build.sh` scripts located in the root folder. The CMake build definition is available in the `CMakeLists.txt` file and can be modified by appending options to `build.bat` or `build.sh`, such as `build.bat -DOCOS_BUILD_SHARED_LIB=OFF`. For more details, please refer to the [C API documentation](./docs/c_api.md).
 
 ### **Python installation**
 ```bash
 pip install onnxruntime-extensions
 ````
-
-
-### **Nightly Build**
-
-#### <strong>on Windows</strong>
-```cmd
-pip install --index-url https://aiinfra.pkgs.visualstudio.com/PublicPackages/_packaging/ORT-Nightly/pypi/simple/ onnxruntime-extensions
-```
-Please ensure that you have met the prerequisites of onnxruntime-extensions (e.g., onnx and onnxruntime) in your Python environment.
-#### <strong>on Linux/macOS</strong>
-Please make sure the compiler toolkit like gcc(later than g++ 8.0) or clang are installed before the following command
-```bash
-python -m pip install git+https://github.com/microsoft/onnxruntime-extensions.git
-```
+The nightly build is also available for the latest features, please refer to [nightly build](./docs/development.md#nightly-build)
 
 
 ## Usage

{onnxruntime_extensions-0.11.0.dist-info → onnxruntime_extensions-0.13.0.dist-info}/RECORD

@@ -1,14 +1,15 @@
-onnxruntime_extensions/__init__.py,sha256=V21JqGUbsnBjatRb9Z83x7F8kjlDtw9pXlIrt4wEUbw,2239
+onnxruntime_extensions/__init__.py,sha256=GMnMIHJ-uqvJGPn5fpCZOi7OG16kFVpfOTTO88kYJWY,2387
 onnxruntime_extensions/_cuops.py,sha256=SUD2NhEWHeMem8ylCtCGBKutSuZQs4WMj1ke65-52vA,16193
-onnxruntime_extensions/_extensions_pydll.cp39-win_amd64.pyd,sha256=yDFhYAzduNnCSw2fOYuHGs0xVjzSzSlr61C-H_cCRxE,5211136
+onnxruntime_extensions/_extensions_pydll.cp39-win_amd64.pyd,sha256=E54SOjYcuHaJ9UBjtlYQi4QhDtHlLVIC7r4F4WiAJcQ,3374080
 onnxruntime_extensions/_extensions_pydll.pyi,sha256=mYXkqNaCgAbs161RDKgDjxIX9vWdYdVPDC-0X9cieco,1070
-onnxruntime_extensions/_hf_cvt.py,sha256=HJwpcdc02aYV9qgAYkrtSYbkargYi0xTqf7Ye60D84A,14062
+onnxruntime_extensions/_hf_cvt.py,sha256=3RDEr4uga_FYBReSDgqLqvj_2-7HgVOk073BRT8lK_E,14082
 onnxruntime_extensions/_ocos.py,sha256=OlDOlCH_vWFOBkjbp6Pujgw6rgk8Fd3_2Mi5ev1eeS0,4193
 onnxruntime_extensions/_ortapi2.py,sha256=Tfrf9fQMQ0e7Wa4R8s4SHdwMNBdmj33wH3y5vMkVVQE,9951
-onnxruntime_extensions/_torch_cvt.py,sha256=1EQI7t_Bz6MvGYwQ15O1bIanPImacKSS0A_JtZrAN58,11403
-onnxruntime_extensions/_version.py,sha256=NhimnAwvZjvrFikdniA7jaWFoc8VEjT3TZb9r4Rfhb0,76
+onnxruntime_extensions/_torch_cvt.py,sha256=hGOiw24QuFpK_3CLjg8Fs2GD_cCdM049xcJxkHVRbAk,10185
+onnxruntime_extensions/_version.py,sha256=uFV2Hf2pKZgA12fk8O9oCxhl_fUheVDYuvrK3A-Vjnc,76
 onnxruntime_extensions/cmd.py,sha256=eIiNNY0ohbUCPgmr9RwOfi0Gzw7nWL17i625L-ZKezI,2428
 onnxruntime_extensions/cvt.py,sha256=XMz0CZXBJQ9IwnixjzJwz-utKyu9HREIEUCviZg6v8A,3977
+onnxruntime_extensions/pp_api.py,sha256=MpW3frODcWXOmYaoTQiYWiM807rC8GjlWgpIYc-CDy8,3051
 onnxruntime_extensions/util.py,sha256=KxNFY0-5CG1i9HADcCc4V33PNukTO46Os_KIL8pj-l8,7394
 onnxruntime_extensions/onnxprocess/__init__.py,sha256=BnveHXnu2nTQNbCLeZujZgZwO9A3yWFbQGTDthCFbIc,534
 onnxruntime_extensions/onnxprocess/_builder.py,sha256=L_afKeE7Wc4mWJ47eVXQ2stvmal_37QVTQZgKmt0ZK8,1844
@@ -35,8 +36,8 @@ onnxruntime_extensions/tools/pre_post_processing/steps/__init__.py,sha256=pdVRZB
 onnxruntime_extensions/tools/pre_post_processing/steps/general.py,sha256=fF_XVFSKOCu482Sqjp-nVPbs-ZVGpPal2ekbO1gUO_4,13781
 onnxruntime_extensions/tools/pre_post_processing/steps/nlp.py,sha256=ZCxRNxqfANplxCe0I-6BfHziM1jDYJsNQKbHdM3Y1I0,15173
 onnxruntime_extensions/tools/pre_post_processing/steps/vision.py,sha256=BM6CGylOSu4l6UarPfW0I2tgkJDa1Q-gYz__CxZle-k,53183
-onnxruntime_extensions-0.11.0.dist-info/LICENSE,sha256=mQaUD2Gx8LUz-n2ZuvVReLKAj74RPqUd-_rYVyzNXys,1162
-onnxruntime_extensions-0.11.0.dist-info/METADATA,sha256=862ZX9u4FpIcC3g5P8oACkHj98sfgEBwq6H28mpmRb0,4452
-onnxruntime_extensions-0.11.0.dist-info/WHEEL,sha256=Z6c-bE0pUM47a70GvqO_SvH_XXU0lm62gEAKtoNJ08A,100
-onnxruntime_extensions-0.11.0.dist-info/top_level.txt,sha256=XyAgQDKyXsf6_0MJb58kRdHwigpTn7A7kl9diBEjs8M,23
-onnxruntime_extensions-0.11.0.dist-info/RECORD,,
+onnxruntime_extensions-0.13.0.dist-info/LICENSE,sha256=mQaUD2Gx8LUz-n2ZuvVReLKAj74RPqUd-_rYVyzNXys,1162
+onnxruntime_extensions-0.13.0.dist-info/METADATA,sha256=MmpoFbjh5Etph2BR7msz-Nrg6QNh56wtGqL46NUSiaE,4483
+onnxruntime_extensions-0.13.0.dist-info/WHEEL,sha256=UAgGEIlEKluxCX20ppULf3M0rfG_1DUMdncOI_parX8,99
+onnxruntime_extensions-0.13.0.dist-info/top_level.txt,sha256=XyAgQDKyXsf6_0MJb58kRdHwigpTn7A7kl9diBEjs8M,23
+onnxruntime_extensions-0.13.0.dist-info/RECORD,,

{onnxruntime_extensions-0.11.0.dist-info → onnxruntime_extensions-0.13.0.dist-info}/WHEEL

@@ -1,5 +1,5 @@
 Wheel-Version: 1.0
-Generator: bdist_wheel (0.43.0)
+Generator: setuptools (75.2.0)
 Root-Is-Purelib: false
 Tag: cp39-cp39-win_amd64