microsoft/onnxruntime-extensions
Publicmirrored fromhttps://github.com/microsoft/onnxruntime-extensionsAvailable
test/test_string_ops.py
845lines · modecode
| 1 | # coding: utf-8 |
| 2 | import unittest |
| 3 | import re |
| 4 | from binascii import crc32 |
| 5 | import numpy as np |
| 6 | from onnx import helper, onnx_pb as onnx_proto |
| 7 | import onnxruntime as _ort |
| 8 | from onnxruntime_customops import ( |
| 9 | onnx_op, PyCustomOpDef, |
| 10 | get_library_path as _get_library_path, |
| 11 | hash_64) |
| 12 | |
| 13 | NUM_BUCKETS = 23 |
| 14 | |
| 15 | |
| 16 | def _create_test_model_string_upper(prefix, domain='ai.onnx.contrib'): |
| 17 | nodes = [] |
| 18 | nodes[0:] = [helper.make_node('Identity', ['input_1'], ['identity1'])] |
| 19 | nodes[1:] = [helper.make_node('%sStringUpper' % prefix, |
| 20 | ['identity1'], ['customout'], |
| 21 | domain=domain)] |
| 22 | |
| 23 | input0 = helper.make_tensor_value_info( |
| 24 | 'input_1', onnx_proto.TensorProto.STRING, [None, None]) |
| 25 | output0 = helper.make_tensor_value_info( |
| 26 | 'customout', onnx_proto.TensorProto.STRING, [None, None]) |
| 27 | |
| 28 | graph = helper.make_graph(nodes, 'test0', [input0], [output0]) |
| 29 | model = helper.make_model( |
| 30 | graph, opset_imports=[helper.make_operatorsetid(domain, 1)]) |
| 31 | return model |
| 32 | |
| 33 | |
| 34 | def _create_test_model_string_join(prefix, domain='ai.onnx.contrib'): |
| 35 | nodes = [] |
| 36 | nodes.append( |
| 37 | helper.make_node('Identity', ['text'], ['identity1'])) |
| 38 | nodes.append( |
| 39 | helper.make_node('Identity', ['sep'], ['identity2'])) |
| 40 | nodes.append( |
| 41 | helper.make_node('Identity', ['axis'], ['identity3'])) |
| 42 | nodes.append( |
| 43 | helper.make_node( |
| 44 | '%sStringJoin' % prefix, ['identity1', 'identity2', 'identity3'], |
| 45 | ['customout'], domain=domain)) |
| 46 | |
| 47 | input0 = helper.make_tensor_value_info( |
| 48 | 'text', onnx_proto.TensorProto.STRING, None) |
| 49 | input1 = helper.make_tensor_value_info( |
| 50 | 'sep', onnx_proto.TensorProto.STRING, [1]) |
| 51 | input2 = helper.make_tensor_value_info( |
| 52 | 'axis', onnx_proto.TensorProto.INT64, [1]) |
| 53 | output0 = helper.make_tensor_value_info( |
| 54 | 'customout', onnx_proto.TensorProto.STRING, None) |
| 55 | |
| 56 | graph = helper.make_graph( |
| 57 | nodes, 'test0', [input0, input1, input2], [output0]) |
| 58 | model = helper.make_model( |
| 59 | graph, opset_imports=[helper.make_operatorsetid(domain, 1)]) |
| 60 | return model |
| 61 | |
| 62 | |
| 63 | def _create_test_model_string_replace(prefix, domain='ai.onnx.contrib'): |
| 64 | nodes = [] |
| 65 | nodes.append( |
| 66 | helper.make_node('Identity', ['text'], ['id1'])) |
| 67 | nodes.append( |
| 68 | helper.make_node('Identity', ['pattern'], ['id2'])) |
| 69 | nodes.append( |
| 70 | helper.make_node('Identity', ['rewrite'], ['id3'])) |
| 71 | nodes.append( |
| 72 | helper.make_node( |
| 73 | '%sStringRegexReplace' % prefix, ['id1', 'id2', 'id3'], |
| 74 | ['customout'], domain=domain)) |
| 75 | |
| 76 | input0 = helper.make_tensor_value_info( |
| 77 | 'text', onnx_proto.TensorProto.STRING, [None, 1]) |
| 78 | input1 = helper.make_tensor_value_info( |
| 79 | 'pattern', onnx_proto.TensorProto.STRING, [1]) |
| 80 | input2 = helper.make_tensor_value_info( |
| 81 | 'rewrite', onnx_proto.TensorProto.STRING, [1]) |
| 82 | output0 = helper.make_tensor_value_info( |
| 83 | 'customout', onnx_proto.TensorProto.STRING, [None, 1]) |
| 84 | |
| 85 | graph = helper.make_graph( |
| 86 | nodes, 'test0', [input0, input1, input2], [output0]) |
| 87 | model = helper.make_model( |
| 88 | graph, opset_imports=[helper.make_operatorsetid(domain, 1)]) |
| 89 | return model |
| 90 | |
| 91 | |
| 92 | def _create_test_model_string_to_hash( |
| 93 | prefix, domain='ai.onnx.contrib', kind=None): |
| 94 | if kind == 'crc32': |
| 95 | op_type = 'StringToCRC32' |
| 96 | out_type = onnx_proto.TensorProto.UINT32 |
| 97 | in_type = out_type |
| 98 | elif kind == 'hash_bucket': |
| 99 | op_type = 'StringToHashBucket' |
| 100 | out_type = onnx_proto.TensorProto.INT64 |
| 101 | in_type = out_type |
| 102 | elif kind == 'hash_bucket_fast': |
| 103 | op_type = 'StringToHashBucketFast' |
| 104 | out_type = onnx_proto.TensorProto.INT64 |
| 105 | in_type = out_type |
| 106 | else: |
| 107 | raise ValueError('Unknown value %r.' % kind) |
| 108 | nodes = [] |
| 109 | nodes.append( |
| 110 | helper.make_node('Identity', ['text'], ['id1'])) |
| 111 | nodes.append( |
| 112 | helper.make_node('Identity', ['num_buckets'], ['id2'])) |
| 113 | nodes.append( |
| 114 | helper.make_node( |
| 115 | '%s%s' % (prefix, op_type), ['id1', 'id2'], |
| 116 | ['customout'], domain=domain)) |
| 117 | |
| 118 | input0 = helper.make_tensor_value_info( |
| 119 | 'text', onnx_proto.TensorProto.STRING, [None, None]) |
| 120 | input1 = helper.make_tensor_value_info( |
| 121 | 'num_buckets', in_type, [1]) |
| 122 | output0 = helper.make_tensor_value_info( |
| 123 | 'customout', out_type, [None, None]) |
| 124 | |
| 125 | graph = helper.make_graph( |
| 126 | nodes, 'test0', [input0, input1], [output0]) |
| 127 | model = helper.make_model( |
| 128 | graph, opset_imports=[helper.make_operatorsetid(domain, 1)]) |
| 129 | return model |
| 130 | |
| 131 | |
| 132 | def _create_test_model_string_equal(prefix, domain='ai.onnx.contrib'): |
| 133 | nodes = [] |
| 134 | nodes.append(helper.make_node('Identity', ['x'], ['id1'])) |
| 135 | nodes.append(helper.make_node('Identity', ['y'], ['id2'])) |
| 136 | nodes.append( |
| 137 | helper.make_node( |
| 138 | '%sStringEqual' % prefix, ['id1', 'id2'], ['z'], domain=domain)) |
| 139 | |
| 140 | input0 = helper.make_tensor_value_info( |
| 141 | 'x', onnx_proto.TensorProto.STRING, []) |
| 142 | input1 = helper.make_tensor_value_info( |
| 143 | 'y', onnx_proto.TensorProto.STRING, []) |
| 144 | output0 = helper.make_tensor_value_info( |
| 145 | 'z', onnx_proto.TensorProto.BOOL, []) |
| 146 | |
| 147 | graph = helper.make_graph(nodes, 'test0', [input0, input1], [output0]) |
| 148 | model = helper.make_model( |
| 149 | graph, opset_imports=[helper.make_operatorsetid(domain, 1)]) |
| 150 | return model |
| 151 | |
| 152 | |
| 153 | def _create_test_model_string_split(prefix, domain='ai.onnx.contrib'): |
| 154 | nodes = [] |
| 155 | nodes.append(helper.make_node('Identity', ['input'], ['id1'])) |
| 156 | nodes.append(helper.make_node('Identity', ['delimiter'], ['id2'])) |
| 157 | nodes.append(helper.make_node('Identity', ['skip_empty'], ['id3'])) |
| 158 | nodes.append( |
| 159 | helper.make_node( |
| 160 | '%sStringSplit' % prefix, ['id1', 'id2', 'id3'], |
| 161 | ['indices', 'values', 'shape'], domain=domain)) |
| 162 | |
| 163 | input0 = helper.make_tensor_value_info( |
| 164 | 'input', onnx_proto.TensorProto.STRING, []) |
| 165 | input1 = helper.make_tensor_value_info( |
| 166 | 'delimiter', onnx_proto.TensorProto.STRING, []) |
| 167 | input2 = helper.make_tensor_value_info( |
| 168 | 'skip_empty', onnx_proto.TensorProto.BOOL, []) |
| 169 | output0 = helper.make_tensor_value_info( |
| 170 | 'indices', onnx_proto.TensorProto.INT64, []) |
| 171 | output1 = helper.make_tensor_value_info( |
| 172 | 'values', onnx_proto.TensorProto.STRING, []) |
| 173 | output2 = helper.make_tensor_value_info( |
| 174 | 'shape', onnx_proto.TensorProto.INT64, []) |
| 175 | |
| 176 | graph = helper.make_graph(nodes, 'test0', [input0, input1, input2], |
| 177 | [output0, output1, output2]) |
| 178 | model = helper.make_model( |
| 179 | graph, opset_imports=[helper.make_operatorsetid(domain, 1)]) |
| 180 | return model |
| 181 | |
| 182 | |
| 183 | class TestPythonOpString(unittest.TestCase): |
| 184 | |
| 185 | _string_join = None |
| 186 | _string_to_crc32 = None |
| 187 | |
| 188 | @classmethod |
| 189 | def setUpClass(cls): |
| 190 | |
| 191 | @onnx_op(op_type="PyStringUpper", |
| 192 | inputs=[PyCustomOpDef.dt_string], |
| 193 | outputs=[PyCustomOpDef.dt_string]) |
| 194 | def string_upper(x): |
| 195 | # The user custom op implementation here. |
| 196 | return np.array([s.upper() for s in x.ravel()]).reshape(x.shape) |
| 197 | |
| 198 | @onnx_op(op_type="PyStringJoin", |
| 199 | inputs=[PyCustomOpDef.dt_string, PyCustomOpDef.dt_string, |
| 200 | PyCustomOpDef.dt_int64], |
| 201 | outputs=[PyCustomOpDef.dt_string]) |
| 202 | def string_join(x, sep, axis): |
| 203 | # The user custom op implementation here. |
| 204 | if sep.shape != (1, ): |
| 205 | raise RuntimeError( |
| 206 | "Unexpected shape {} for 'sep'.".format(sep.shape)) |
| 207 | if axis.shape != (1, ): |
| 208 | raise RuntimeError( |
| 209 | "Unexpected shape {} for 'axis'.".format(axis.shape)) |
| 210 | sp = sep[0] |
| 211 | ax = axis[0] |
| 212 | if ax < 0 or ax >= len(x.shape): |
| 213 | raise RuntimeError( |
| 214 | "axis must be in [%r,%r] but is %r" % ( |
| 215 | 0, len(x.shape), ax)) |
| 216 | if len(x.shape) == 1: |
| 217 | return np.array([sp.join(x)]) |
| 218 | dims = np.arange(len(x.shape)) |
| 219 | dims[ax], dims[-1] = dims[-1], dims[ax] |
| 220 | x2 = np.transpose(x, dims) |
| 221 | res_shape = x2.shape[:-1] |
| 222 | x2 = x2.reshape((-1, x2.shape[-1])) |
| 223 | res = np.empty(x2.shape[0], dtype=x.dtype) |
| 224 | for i in range(x2.shape[0]): |
| 225 | res[i] = sp.join(x2[i, :]) |
| 226 | return res.reshape(res_shape) |
| 227 | |
| 228 | @onnx_op(op_type="PyStringRegexReplace", |
| 229 | inputs=[PyCustomOpDef.dt_string, PyCustomOpDef.dt_string, |
| 230 | PyCustomOpDef.dt_string], |
| 231 | outputs=[PyCustomOpDef.dt_string]) |
| 232 | def string_replace(x, pattern, rewrite): |
| 233 | # The user custom op implementation here. |
| 234 | if pattern.shape != (1, ): |
| 235 | raise RuntimeError( |
| 236 | "Unexpected shape {} for 'pattern'.".format(pattern.shape)) |
| 237 | if rewrite.shape != (1, ): |
| 238 | raise RuntimeError( |
| 239 | "Unexpected shape {} for 'rewrite'.".format(rewrite.shape)) |
| 240 | reg = re.compile(pattern[0]) |
| 241 | res = np.array( |
| 242 | list(map(lambda t: reg.sub(rewrite[0], t), x.ravel()))) |
| 243 | return res.reshape(x.shape) |
| 244 | |
| 245 | @onnx_op(op_type="PyStringToCRC32", |
| 246 | inputs=[PyCustomOpDef.dt_string, PyCustomOpDef.dt_uint32], |
| 247 | outputs=[PyCustomOpDef.dt_uint32]) |
| 248 | def string_to_crc32(x, num_buckets): |
| 249 | if num_buckets.shape != (1, ): |
| 250 | raise RuntimeError( |
| 251 | "Unexpected shape {} for 'num_buckets'.".format( |
| 252 | num_buckets.shape)) |
| 253 | nb = num_buckets[0] |
| 254 | res = np.array( |
| 255 | list(map( |
| 256 | lambda x: crc32(x.encode('iso-8859-15')) % nb, |
| 257 | x.ravel()))) |
| 258 | return res.reshape(x.shape) |
| 259 | |
| 260 | @onnx_op(op_type="PyStringToHashBucket", |
| 261 | inputs=[PyCustomOpDef.dt_string, PyCustomOpDef.dt_int64], |
| 262 | outputs=[PyCustomOpDef.dt_int64]) |
| 263 | def string_to_hash_bucket(x, num_buckets): |
| 264 | if num_buckets.shape != (1, ): |
| 265 | raise RuntimeError( |
| 266 | "Unexpected shape {} for 'num_buckets'.".format( |
| 267 | num_buckets.shape)) |
| 268 | nb = num_buckets[0] |
| 269 | res = np.array( |
| 270 | list(map(lambda x: hash_64(x, nb, True), x.ravel()))) |
| 271 | return res.reshape(x.shape).astype(np.int64) |
| 272 | |
| 273 | @onnx_op(op_type="PyStringEqual", |
| 274 | inputs=[PyCustomOpDef.dt_string, PyCustomOpDef.dt_string], |
| 275 | outputs=[PyCustomOpDef.dt_bool]) |
| 276 | def string_equal(x, y): |
| 277 | return x == y |
| 278 | |
| 279 | @onnx_op(op_type="PyStringSplit", |
| 280 | inputs=[PyCustomOpDef.dt_string, PyCustomOpDef.dt_string, |
| 281 | PyCustomOpDef.dt_bool], |
| 282 | outputs=[PyCustomOpDef.dt_int64, PyCustomOpDef.dt_string, |
| 283 | PyCustomOpDef.dt_int64]) |
| 284 | def string_split(input, delimiter, skip_empty): |
| 285 | if delimiter.shape != (1, ): |
| 286 | raise RuntimeError("demiliter must a single element tensor.") |
| 287 | if skip_empty.shape != (1, ): |
| 288 | raise RuntimeError("skip_empty must a single element tensor.") |
| 289 | if len(input.shape) != 1: |
| 290 | raise RuntimeError("input must a one dimension tensor.") |
| 291 | delimiter = delimiter[0] |
| 292 | skip_empty = skip_empty[0] |
| 293 | texts = [] |
| 294 | indices = [] |
| 295 | max_split = 0 |
| 296 | for row, text in enumerate(input): |
| 297 | if not text: |
| 298 | continue |
| 299 | res = text.split(delimiter) |
| 300 | if skip_empty: |
| 301 | res = [t for t in res if t] |
| 302 | texts.extend(res) |
| 303 | max_split = max(max_split, len(res)) |
| 304 | indices.extend((row, i) for i in range(len(res))) |
| 305 | return (np.array(indices, dtype=np.int64), |
| 306 | np.array(texts), |
| 307 | np.array([len(input), max_split], dtype=np.int64)) |
| 308 | |
| 309 | cls._string_join = string_join |
| 310 | cls._string_to_crc32 = string_to_crc32 |
| 311 | |
| 312 | def test_check_types(self): |
| 313 | def_list = set(dir(PyCustomOpDef)) |
| 314 | type_list = [ |
| 315 | # 'dt_bfloat16', |
| 316 | 'dt_bool', |
| 317 | 'dt_complex128', |
| 318 | 'dt_complex64', |
| 319 | 'dt_double', |
| 320 | 'dt_float', |
| 321 | 'dt_float16', |
| 322 | 'dt_int16', |
| 323 | 'dt_int32', |
| 324 | 'dt_int64', |
| 325 | 'dt_int8', |
| 326 | 'dt_string', |
| 327 | 'dt_uint16', |
| 328 | 'dt_uint32', |
| 329 | 'dt_uint64', |
| 330 | 'dt_uint8'] |
| 331 | for t in type_list: |
| 332 | self.assertIn(t, def_list) |
| 333 | |
| 334 | def test_string_upper_cc(self): |
| 335 | so = _ort.SessionOptions() |
| 336 | so.register_custom_ops_library(_get_library_path()) |
| 337 | onnx_model = _create_test_model_string_upper('') |
| 338 | self.assertIn('op_type: "StringUpper"', str(onnx_model)) |
| 339 | sess = _ort.InferenceSession(onnx_model.SerializeToString(), so) |
| 340 | input_1 = np.array([["Abc"]]) |
| 341 | txout = sess.run(None, {'input_1': input_1}) |
| 342 | self.assertEqual(txout[0].tolist(), np.array([["ABC"]]).tolist()) |
| 343 | |
| 344 | def test_string_upper_cc_accent(self): |
| 345 | so = _ort.SessionOptions() |
| 346 | so.register_custom_ops_library(_get_library_path()) |
| 347 | onnx_model = _create_test_model_string_upper('') |
| 348 | self.assertIn('op_type: "StringUpper"', str(onnx_model)) |
| 349 | sess = _ort.InferenceSession(onnx_model.SerializeToString(), so) |
| 350 | input_1 = np.array([["Abcé"]]) |
| 351 | txout = sess.run(None, {'input_1': input_1}) |
| 352 | self.assertEqual(txout[0].tolist(), np.array([["ABCé"]]).tolist()) |
| 353 | |
| 354 | def test_string_upper_python(self): |
| 355 | so = _ort.SessionOptions() |
| 356 | so.register_custom_ops_library(_get_library_path()) |
| 357 | onnx_model = _create_test_model_string_upper('Py') |
| 358 | self.assertIn('op_type: "PyStringUpper"', str(onnx_model)) |
| 359 | sess = _ort.InferenceSession(onnx_model.SerializeToString(), so) |
| 360 | input_1 = np.array([["Abc"]]) |
| 361 | txout = sess.run(None, {'input_1': input_1}) |
| 362 | self.assertEqual(txout[0].tolist(), np.array([["ABC"]]).tolist()) |
| 363 | |
| 364 | def test_string_upper_python_accent(self): |
| 365 | so = _ort.SessionOptions() |
| 366 | so.register_custom_ops_library(_get_library_path()) |
| 367 | onnx_model = _create_test_model_string_upper('Py') |
| 368 | self.assertIn('op_type: "PyStringUpper"', str(onnx_model)) |
| 369 | sess = _ort.InferenceSession(onnx_model.SerializeToString(), so) |
| 370 | input_1 = np.array([["Abcé"]]) |
| 371 | txout = sess.run(None, {'input_1': input_1}) |
| 372 | self.assertEqual(txout[0].tolist(), |
| 373 | np.array([["ABCé".upper()]]).tolist()) |
| 374 | |
| 375 | def test_string_join_python(self): |
| 376 | so = _ort.SessionOptions() |
| 377 | so.register_custom_ops_library(_get_library_path()) |
| 378 | onnx_model = _create_test_model_string_join('Py') |
| 379 | self.assertIn('op_type: "PyStringJoin"', str(onnx_model)) |
| 380 | sess = _ort.InferenceSession(onnx_model.SerializeToString(), so) |
| 381 | text = np.vstack([np.array([["a", "b", "c"]]), |
| 382 | np.array([["aa", "bb", ""]])]) |
| 383 | self.assertEqual(text.shape, (2, 3)) |
| 384 | sep = np.array([";"]) |
| 385 | axis = np.array([1], dtype=np.int64) |
| 386 | TestPythonOpString._string_join(text, sep, axis) |
| 387 | txout = sess.run(None, {'text': text, 'sep': sep, 'axis': axis}) |
| 388 | self.assertEqual( |
| 389 | txout[0].tolist(), np.array(["a;b;c", "aa;bb;"]).tolist()) |
| 390 | axis = np.array([0], dtype=np.int64) |
| 391 | TestPythonOpString._string_join(text, sep, axis) |
| 392 | txout = sess.run(None, {'text': text, 'sep': sep, 'axis': axis}) |
| 393 | self.assertEqual( |
| 394 | txout[0].tolist(), np.array(['a;aa', 'b;bb', 'c;']).tolist()) |
| 395 | |
| 396 | def test_string_join_python_3d(self): |
| 397 | so = _ort.SessionOptions() |
| 398 | so.register_custom_ops_library(_get_library_path()) |
| 399 | onnx_model = _create_test_model_string_join('Py') |
| 400 | self.assertIn('op_type: "PyStringJoin"', str(onnx_model)) |
| 401 | sess = _ort.InferenceSession(onnx_model.SerializeToString(), so) |
| 402 | text = np.vstack([np.array([["a", "b", "c"]]), |
| 403 | np.array([["aa", "bb", ""]])]).reshape((2, 3, 1)) |
| 404 | sep = np.array([";"]) |
| 405 | axis = np.array([1], dtype=np.int64) |
| 406 | TestPythonOpString._string_join(text, sep, axis) |
| 407 | txout = sess.run(None, {'text': text, 'sep': sep, 'axis': axis}) |
| 408 | self.assertEqual( |
| 409 | txout[0].tolist(), np.array([['a;b;c'], ['aa;bb;']]).tolist()) |
| 410 | |
| 411 | def test_string_join_python_1d(self): |
| 412 | so = _ort.SessionOptions() |
| 413 | so.register_custom_ops_library(_get_library_path()) |
| 414 | onnx_model = _create_test_model_string_join('Py') |
| 415 | self.assertIn('op_type: "PyStringJoin"', str(onnx_model)) |
| 416 | sess = _ort.InferenceSession(onnx_model.SerializeToString(), so) |
| 417 | text = np.array(["a", "b", "cc"]) |
| 418 | sep = np.array([";"]) |
| 419 | axis = np.array([0], dtype=np.int64) |
| 420 | txout = sess.run(None, {'text': text, 'sep': sep, 'axis': axis}) |
| 421 | self.assertEqual(txout[0].shape, (1, )) |
| 422 | self.assertEqual( |
| 423 | txout[0].tolist(), np.array(["a;b;cc"]).tolist()) |
| 424 | |
| 425 | def test_string_join_cc(self): |
| 426 | so = _ort.SessionOptions() |
| 427 | so.register_custom_ops_library(_get_library_path()) |
| 428 | onnx_model = _create_test_model_string_join('') |
| 429 | self.assertIn('op_type: "StringJoin"', str(onnx_model)) |
| 430 | sess = _ort.InferenceSession(onnx_model.SerializeToString(), so) |
| 431 | text = np.vstack([np.array([["a", "b", "c"]]), |
| 432 | np.array([["aa", "bb", ""]])]) |
| 433 | sep = np.array([";"]) |
| 434 | axis = np.array([1], dtype=np.int64) |
| 435 | txout = sess.run(None, {'text': text, 'sep': sep, 'axis': axis}) |
| 436 | self.assertEqual( |
| 437 | txout[0].tolist(), np.array(["a;b;c", "aa;bb;"]).tolist()) |
| 438 | axis = np.array([0], dtype=np.int64) |
| 439 | txout = sess.run(None, {'text': text, 'sep': sep, 'axis': axis}) |
| 440 | self.assertEqual( |
| 441 | txout[0].tolist(), np.array(['a;aa', 'b;bb', 'c;']).tolist()) |
| 442 | |
| 443 | def test_string_join_cc_1d(self): |
| 444 | so = _ort.SessionOptions() |
| 445 | so.register_custom_ops_library(_get_library_path()) |
| 446 | onnx_model = _create_test_model_string_join('') |
| 447 | self.assertIn('op_type: "StringJoin"', str(onnx_model)) |
| 448 | sess = _ort.InferenceSession(onnx_model.SerializeToString(), so) |
| 449 | text = np.array(["a", "b", "cc"]) |
| 450 | sep = np.array([";"]) |
| 451 | axis = np.array([0], dtype=np.int64) |
| 452 | txout = sess.run(None, {'text': text, 'sep': sep, 'axis': axis}) |
| 453 | self.assertEqual( |
| 454 | txout[0].tolist(), np.array(["a;b;cc"]).tolist()) |
| 455 | |
| 456 | def test_string_join_cc_3d(self): |
| 457 | so = _ort.SessionOptions() |
| 458 | so.register_custom_ops_library(_get_library_path()) |
| 459 | onnx_model = _create_test_model_string_join('') |
| 460 | self.assertIn('op_type: "StringJoin"', str(onnx_model)) |
| 461 | sess = _ort.InferenceSession(onnx_model.SerializeToString(), so) |
| 462 | text = np.array(["a", "b", "c", "d", "e", "f", "g", "h"]).reshape(( |
| 463 | 2, 2, 2)) |
| 464 | sep = np.array([";"]) |
| 465 | axis = np.array([2], dtype=np.int64) |
| 466 | txout = sess.run(None, {'text': text, 'sep': sep, 'axis': axis}) |
| 467 | self.assertEqual( |
| 468 | txout[0].tolist(), |
| 469 | np.array([['a;b', 'c;d'], ['e;f', 'g;h']]).tolist()) |
| 470 | axis = np.array([1], dtype=np.int64) |
| 471 | txout = sess.run(None, {'text': text, 'sep': sep, 'axis': axis}) |
| 472 | self.assertEqual( |
| 473 | txout[0].tolist(), |
| 474 | np.array([['a;c', 'b;d'], ['e;g', 'f;h']]).tolist()) |
| 475 | axis = np.array([0], dtype=np.int64) |
| 476 | txout = sess.run(None, {'text': text, 'sep': sep, 'axis': axis}) |
| 477 | self.assertEqual( |
| 478 | txout[0].tolist(), |
| 479 | np.array([['a;e', 'b;f'], ['c;g', 'd;h']]).tolist()) |
| 480 | |
| 481 | def test_string_replace_cc(self): |
| 482 | so = _ort.SessionOptions() |
| 483 | so.register_custom_ops_library(_get_library_path()) |
| 484 | onnx_model = _create_test_model_string_replace('') |
| 485 | self.assertIn('op_type: "StringRegexReplace"', str(onnx_model)) |
| 486 | sess = _ort.InferenceSession(onnx_model.SerializeToString(), so) |
| 487 | pattern = np.array([r'def\s+([a-zA-Z_][a-zA-Z_0-9]*)\s*\(\s*\):']) |
| 488 | rewrite = np.array([r'static PyObject* py_\1(void) {']) |
| 489 | text = np.array([['def myfunc():'], ['def dummy():']]) |
| 490 | txout = sess.run( |
| 491 | None, {'text': text, 'pattern': pattern, 'rewrite': rewrite}) |
| 492 | exp = [['static PyObject* py_myfunc(void) {'], |
| 493 | ['static PyObject* py_dummy(void) {']] |
| 494 | self.assertEqual(exp, txout[0].tolist()) |
| 495 | |
| 496 | def test_string_replace_cc_x2(self): |
| 497 | so = _ort.SessionOptions() |
| 498 | so.register_custom_ops_library(_get_library_path()) |
| 499 | onnx_model = _create_test_model_string_replace('') |
| 500 | self.assertIn('op_type: "StringRegexReplace"', str(onnx_model)) |
| 501 | sess = _ort.InferenceSession(onnx_model.SerializeToString(), so) |
| 502 | pattern = np.array([r'def\s+([a-zA-Z_][a-zA-Z_0-9]*)\s*\(\s*\):']) |
| 503 | rewrite = np.array([r'static PyObject* py_\1(void) {']) |
| 504 | text = np.array([['def myfunc():'], ['def dummy():' * 2]]) |
| 505 | txout = sess.run( |
| 506 | None, {'text': text, 'pattern': pattern, 'rewrite': rewrite}) |
| 507 | exp = [['static PyObject* py_myfunc(void) {'], |
| 508 | ['static PyObject* py_dummy(void) {' * 2]] |
| 509 | self.assertEqual(exp, txout[0].tolist()) |
| 510 | |
| 511 | def test_string_replace_python(self): |
| 512 | so = _ort.SessionOptions() |
| 513 | so.register_custom_ops_library(_get_library_path()) |
| 514 | onnx_model = _create_test_model_string_replace('Py') |
| 515 | self.assertIn('op_type: "PyStringRegexReplace"', str(onnx_model)) |
| 516 | sess = _ort.InferenceSession(onnx_model.SerializeToString(), so) |
| 517 | pattern = np.array([r'def\s+([a-zA-Z_][a-zA-Z_0-9]*)\s*\(\s*\):']) |
| 518 | rewrite = np.array([r'static PyObject*\npy_\1(void)\n{']) |
| 519 | text = np.array([['def myfunc():'], ['def dummy():']]) |
| 520 | txout = sess.run( |
| 521 | None, {'text': text, 'pattern': pattern, 'rewrite': rewrite}) |
| 522 | exp = [['static PyObject*\npy_myfunc(void)\n{'], |
| 523 | ['static PyObject*\npy_dummy(void)\n{']] |
| 524 | self.assertEqual(exp, txout[0].tolist()) |
| 525 | |
| 526 | def test_string_replace_python_x2(self): |
| 527 | so = _ort.SessionOptions() |
| 528 | so.register_custom_ops_library(_get_library_path()) |
| 529 | onnx_model = _create_test_model_string_replace('Py') |
| 530 | self.assertIn('op_type: "PyStringRegexReplace"', str(onnx_model)) |
| 531 | sess = _ort.InferenceSession(onnx_model.SerializeToString(), so) |
| 532 | pattern = np.array([r'def\s+([a-zA-Z_][a-zA-Z_0-9]*)\s*\(\s*\):']) |
| 533 | rewrite = np.array([r'static PyObject*\npy_\1(void)\n{']) |
| 534 | text = np.array([['def myfunc():'], ['def dummy():' * 2]]) |
| 535 | txout = sess.run( |
| 536 | None, {'text': text, 'pattern': pattern, 'rewrite': rewrite}) |
| 537 | exp = [['static PyObject*\npy_myfunc(void)\n{'], |
| 538 | ['static PyObject*\npy_dummy(void)\n{' * 2]] |
| 539 | self.assertEqual(exp, txout[0].tolist()) |
| 540 | |
| 541 | def test_string_to_crc32_python(self): |
| 542 | so = _ort.SessionOptions() |
| 543 | so.register_custom_ops_library(_get_library_path()) |
| 544 | onnx_model = _create_test_model_string_to_hash('Py', kind='crc32') |
| 545 | self.assertIn('op_type: "PyStringToCRC32"', str(onnx_model)) |
| 546 | sess = _ort.InferenceSession(onnx_model.SerializeToString(), so) |
| 547 | text = np.array([["abc", "abcdé"], ["$$^l!%*ù", ""]]) |
| 548 | num_buckets = np.array([44], dtype=np.uint32) |
| 549 | res = self._string_to_crc32(text, num_buckets) |
| 550 | self.assertEqual(res.shape, text.shape) |
| 551 | exp = np.array([[10, 38], [29, 0]], dtype=np.uint32) |
| 552 | self.assertEqual(exp.tolist(), res.tolist()) |
| 553 | txout = sess.run( |
| 554 | None, {'text': text, 'num_buckets': num_buckets}) |
| 555 | self.assertEqual(exp.tolist(), txout[0].tolist()) |
| 556 | |
| 557 | def test_string_to_hash_bucket_cc(self): |
| 558 | so = _ort.SessionOptions() |
| 559 | so.register_custom_ops_library(_get_library_path()) |
| 560 | onnx_model = _create_test_model_string_to_hash( |
| 561 | '', kind='hash_bucket') |
| 562 | self.assertIn('op_type: "StringToHashBucket"', str(onnx_model)) |
| 563 | sess = _ort.InferenceSession(onnx_model.SerializeToString(), so) |
| 564 | raw = ["abc", "abcdé", "$$^l!%*ù", "", "a", "A"] |
| 565 | text = np.array(raw).reshape((3, 2)) |
| 566 | num_buckets = np.array([NUM_BUCKETS], dtype=np.int64) |
| 567 | txout = sess.run( |
| 568 | None, {'text': text, 'num_buckets': num_buckets}) |
| 569 | try: |
| 570 | from tensorflow.raw_ops import StringToHashBucket |
| 571 | dotf = True |
| 572 | except ImportError: |
| 573 | dotf = False |
| 574 | if dotf: |
| 575 | tfres = StringToHashBucket( |
| 576 | string_tensor=text, num_buckets=num_buckets[0]) |
| 577 | self.assertEqual(tfres.shape, txout[0].shape) |
| 578 | self.assertEqual(tfres.numpy().tolist(), txout[0].tolist()) |
| 579 | exp = np.array([[15, 11], [10, 21], [20, 21]], dtype=np.int64) |
| 580 | self.assertEqual(exp.shape, txout[0].shape) |
| 581 | self.assertEqual(exp.tolist(), txout[0].tolist()) |
| 582 | |
| 583 | def test_string_to_hash_bucket_fast_cc(self): |
| 584 | so = _ort.SessionOptions() |
| 585 | so.register_custom_ops_library(_get_library_path()) |
| 586 | onnx_model = _create_test_model_string_to_hash( |
| 587 | '', kind='hash_bucket_fast') |
| 588 | self.assertIn('op_type: "StringToHashBucketFast"', str(onnx_model)) |
| 589 | sess = _ort.InferenceSession(onnx_model.SerializeToString(), so) |
| 590 | raw = ["abc", "abcdé", "$$^l!%*ù", "", "a", "A"] |
| 591 | text = np.array(raw).reshape((3, 2)) |
| 592 | num_buckets = np.array([NUM_BUCKETS], dtype=np.int64) |
| 593 | txout = sess.run( |
| 594 | None, {'text': text, 'num_buckets': num_buckets}) |
| 595 | try: |
| 596 | from tensorflow.raw_ops import StringToHashBucketFast |
| 597 | dotf = True |
| 598 | except ImportError: |
| 599 | dotf = False |
| 600 | if dotf: |
| 601 | tfres = StringToHashBucketFast( |
| 602 | input=text, num_buckets=num_buckets[0]) |
| 603 | self.assertEqual(tfres.shape, txout[0].shape) |
| 604 | self.assertEqual(tfres.numpy().tolist(), txout[0].tolist()) |
| 605 | exp = np.array([[9, 17], [4, 21], [14, 12]], dtype=np.int64) |
| 606 | self.assertEqual(exp.shape, txout[0].shape) |
| 607 | self.assertEqual(exp.tolist(), txout[0].tolist()) |
| 608 | |
| 609 | def test_string_to_hash_bucket_python(self): |
| 610 | so = _ort.SessionOptions() |
| 611 | so.register_custom_ops_library(_get_library_path()) |
| 612 | onnx_model = _create_test_model_string_to_hash( |
| 613 | 'Py', kind='hash_bucket') |
| 614 | self.assertIn('op_type: "PyStringToHashBucket"', str(onnx_model)) |
| 615 | sess = _ort.InferenceSession(onnx_model.SerializeToString(), so) |
| 616 | raw = ["abc", "abcdé", "$$^l!%*ù", "", "a", "A"] |
| 617 | text = np.array(raw).reshape((3, 2)) |
| 618 | num_buckets = np.array([NUM_BUCKETS], dtype=np.int64) |
| 619 | exp = np.array([[9, 17], [4, 21], [14, 12]], dtype=np.int64) |
| 620 | txout = sess.run( |
| 621 | None, {'text': text, 'num_buckets': num_buckets}) |
| 622 | self.assertEqual(exp.shape, txout[0].shape) |
| 623 | self.assertEqual(exp.tolist(), txout[0].tolist()) |
| 624 | |
| 625 | def enumerate_matrix_couples(self): |
| 626 | for i in range(1, 5): |
| 627 | shape = (3,) * i |
| 628 | a = (np.random.rand(*shape) * 10).astype(np.int32).astype(np.str) |
| 629 | yield a, a |
| 630 | for j in range(i): |
| 631 | shape2 = list(shape) |
| 632 | shape2[j] = 1 |
| 633 | b = (np.random.rand(*shape2) * 10).astype( |
| 634 | np.int32).astype(np.str) |
| 635 | yield a, b |
| 636 | for k in range(j+1, i): |
| 637 | shape3 = list(shape2) |
| 638 | shape3[k] = 1 |
| 639 | b = (np.random.rand(*shape3) * 10).astype( |
| 640 | np.int32).astype(np.str) |
| 641 | yield a, b |
| 642 | |
| 643 | def test_string_equal_python(self): |
| 644 | so = _ort.SessionOptions() |
| 645 | so.register_custom_ops_library(_get_library_path()) |
| 646 | onnx_model = _create_test_model_string_equal('Py') |
| 647 | self.assertIn('op_type: "PyStringEqual"', str(onnx_model)) |
| 648 | sess = _ort.InferenceSession(onnx_model.SerializeToString(), so) |
| 649 | |
| 650 | for x, y in self.enumerate_matrix_couples(): |
| 651 | txout = sess.run(None, {'x': x, 'y': y}) |
| 652 | self.assertEqual(txout[0].tolist(), (x == y).tolist()) |
| 653 | txout = sess.run(None, {'x': y, 'y': x}) |
| 654 | self.assertEqual(txout[0].tolist(), (y == x).tolist()) |
| 655 | |
| 656 | def test_string_equal_cc(self): |
| 657 | so = _ort.SessionOptions() |
| 658 | so.register_custom_ops_library(_get_library_path()) |
| 659 | onnx_model = _create_test_model_string_equal('') |
| 660 | self.assertIn('op_type: "StringEqual"', str(onnx_model)) |
| 661 | sess = _ort.InferenceSession(onnx_model.SerializeToString(), so) |
| 662 | |
| 663 | for x, y in self.enumerate_matrix_couples(): |
| 664 | txout = sess.run(None, {'x': x, 'y': y}) |
| 665 | self.assertEqual(txout[0].tolist(), (x == y).tolist()) |
| 666 | txout = sess.run(None, {'x': y, 'y': x}) |
| 667 | self.assertEqual(txout[0].tolist(), (y == x).tolist()) |
| 668 | |
| 669 | def test_string_split_python(self): |
| 670 | so = _ort.SessionOptions() |
| 671 | so.register_custom_ops_library(_get_library_path()) |
| 672 | onnx_model = _create_test_model_string_split('Py') |
| 673 | self.assertIn('op_type: "PyStringSplit"', str(onnx_model)) |
| 674 | sess = _ort.InferenceSession(onnx_model.SerializeToString(), so) |
| 675 | input = np.array(["a,,b", "", "aa,b,c", "dddddd"]) |
| 676 | delimiter = np.array([","]) |
| 677 | |
| 678 | for skip in [True, False]: |
| 679 | with self.subTest(skip=skip): |
| 680 | skip_empty = np.array([skip]) |
| 681 | |
| 682 | txout = sess.run( |
| 683 | None, {'input': input, 'delimiter': delimiter, |
| 684 | 'skip_empty': skip_empty}) |
| 685 | |
| 686 | if skip_empty: |
| 687 | exp_indices = np.array( |
| 688 | [[0, 0], [0, 1], [2, 0], [2, 1], [2, 2], [3, 0]]) |
| 689 | exp_text = np.array(['a', 'b', 'aa', 'b', 'c', 'dddddd']) |
| 690 | else: |
| 691 | exp_indices = np.array( |
| 692 | [[0, 0], [0, 1], [0, 2], [2, 0], [2, 1], |
| 693 | [2, 2], [3, 0]]) |
| 694 | exp_text = np.array( |
| 695 | ['a', '', 'b', 'aa', 'b', 'c', 'dddddd']) |
| 696 | exp_shape = np.array([4, 3]) |
| 697 | self.assertEqual(exp_indices.tolist(), txout[0].tolist()) |
| 698 | self.assertEqual(exp_text.tolist(), txout[1].tolist()) |
| 699 | self.assertEqual(exp_shape.tolist(), txout[2].tolist()) |
| 700 | |
| 701 | def test_string_split_cc(self): |
| 702 | so = _ort.SessionOptions() |
| 703 | so.register_custom_ops_library(_get_library_path()) |
| 704 | onnx_model = _create_test_model_string_split('') |
| 705 | self.assertIn('op_type: "StringSplit"', str(onnx_model)) |
| 706 | sess = _ort.InferenceSession(onnx_model.SerializeToString(), so) |
| 707 | input = np.array(["a,,b", "", "aa,b,c", "dddddd"]) |
| 708 | delimiter = np.array([","]) |
| 709 | |
| 710 | for skip in [True, False]: |
| 711 | with self.subTest(skip=skip): |
| 712 | skip_empty = np.array([skip]) |
| 713 | |
| 714 | txout = sess.run( |
| 715 | None, {'input': input, 'delimiter': delimiter, |
| 716 | 'skip_empty': skip_empty}) |
| 717 | |
| 718 | try: |
| 719 | from tensorflow.raw_ops import StringSplit |
| 720 | dotf = True |
| 721 | except ImportError: |
| 722 | dotf = False |
| 723 | if dotf: |
| 724 | tfres = StringSplit( |
| 725 | input=input, delimiter=",,", skip_empty=skip) |
| 726 | self.assertEqual( |
| 727 | [_.decode() for _ in tfres[1].numpy().tolist()], |
| 728 | txout[1].tolist()) |
| 729 | self.assertEqual( |
| 730 | tfres[0].numpy().tolist(), txout[0].tolist()) |
| 731 | self.assertEqual( |
| 732 | tfres[2].numpy().tolist(), txout[2].tolist()) |
| 733 | |
| 734 | if skip_empty: |
| 735 | exp_indices = np.array( |
| 736 | [[0, 0], [0, 1], [2, 0], [2, 1], [2, 2], [3, 0]]) |
| 737 | exp_text = np.array(['a', 'b', 'aa', 'b', 'c', 'dddddd']) |
| 738 | else: |
| 739 | exp_indices = np.array( |
| 740 | [[0, 0], [0, 1], [0, 2], [2, 0], [2, 1], |
| 741 | [2, 2], [3, 0]]) |
| 742 | exp_text = np.array( |
| 743 | ['a', '', 'b', 'aa', 'b', 'c', 'dddddd']) |
| 744 | exp_shape = np.array([4, 3]) |
| 745 | self.assertEqual(exp_indices.tolist(), txout[0].tolist()) |
| 746 | self.assertEqual(exp_text.tolist(), txout[1].tolist()) |
| 747 | self.assertEqual(exp_shape.tolist(), txout[2].tolist()) |
| 748 | |
| 749 | def test_string_split_cc_sep2(self): |
| 750 | so = _ort.SessionOptions() |
| 751 | so.register_custom_ops_library(_get_library_path()) |
| 752 | onnx_model = _create_test_model_string_split('') |
| 753 | self.assertIn('op_type: "StringSplit"', str(onnx_model)) |
| 754 | sess = _ort.InferenceSession(onnx_model.SerializeToString(), so) |
| 755 | input = np.array(["a*b", "a,*b", "aa,b,,c", 'z', "dddddd,", "**"]) |
| 756 | delimiter = np.array([",*"]) |
| 757 | |
| 758 | for skip in [True, False]: |
| 759 | with self.subTest(skip=skip): |
| 760 | skip_empty = np.array([skip]) |
| 761 | |
| 762 | txout = sess.run( |
| 763 | None, {'input': input, 'delimiter': delimiter, |
| 764 | 'skip_empty': skip_empty}) |
| 765 | |
| 766 | try: |
| 767 | from tensorflow.raw_ops import StringSplit |
| 768 | dotf = True |
| 769 | except ImportError: |
| 770 | dotf = False |
| 771 | if dotf: |
| 772 | tfres = StringSplit( |
| 773 | input=input, delimiter=",*", skip_empty=skip) |
| 774 | self.assertEqual( |
| 775 | [_.decode() for _ in tfres[1].numpy().tolist()], |
| 776 | txout[1].tolist()) |
| 777 | self.assertEqual( |
| 778 | tfres[0].numpy().tolist(), txout[0].tolist()) |
| 779 | self.assertEqual( |
| 780 | tfres[2].numpy().tolist(), txout[2].tolist()) |
| 781 | |
| 782 | if skip_empty: |
| 783 | exp_indices = np.array( |
| 784 | [[0, 0], [0, 1], [1, 0], [1, 1], [2, 0], [2, 1], |
| 785 | [2, 2], [3, 0], [4, 0]]) |
| 786 | exp_text = np.array( |
| 787 | ['a', 'b', 'a', 'b', 'aa', 'b', 'c', 'z', 'dddddd']) |
| 788 | exp_shape = np.array([6, 3]) |
| 789 | else: |
| 790 | exp_indices = np.array( |
| 791 | [[0, 0], [0, 1], [1, 0], [1, 1], [1, 2], [2, 0], |
| 792 | [2, 1], [2, 2], [2, 3], [3, 0], [4, 0], [4, 1], |
| 793 | [5, 0], [5, 1], [5, 2]]) |
| 794 | exp_text = np.array( |
| 795 | ['a', 'b', 'a', '', 'b', 'aa', 'b', '', 'c', |
| 796 | 'z', 'dddddd', '', '', '', '']) |
| 797 | exp_shape = np.array([6, 4]) |
| 798 | self.assertEqual(exp_text.tolist(), txout[1].tolist()) |
| 799 | self.assertEqual(exp_indices.tolist(), txout[0].tolist()) |
| 800 | self.assertEqual(exp_shape.tolist(), txout[2].tolist()) |
| 801 | |
| 802 | def test_string_split_cc_sep0(self): |
| 803 | so = _ort.SessionOptions() |
| 804 | so.register_custom_ops_library(_get_library_path()) |
| 805 | onnx_model = _create_test_model_string_split('') |
| 806 | self.assertIn('op_type: "StringSplit"', str(onnx_model)) |
| 807 | sess = _ort.InferenceSession(onnx_model.SerializeToString(), so) |
| 808 | input = np.array(["a*b", "a,*b"]) |
| 809 | delimiter = np.array([""]) |
| 810 | |
| 811 | for skip in [True, False]: |
| 812 | with self.subTest(skip=skip): |
| 813 | skip_empty = np.array([skip]) |
| 814 | |
| 815 | txout = sess.run( |
| 816 | None, {'input': input, 'delimiter': delimiter, |
| 817 | 'skip_empty': skip_empty}) |
| 818 | |
| 819 | try: |
| 820 | from tensorflow.raw_ops import StringSplit |
| 821 | dotf = True |
| 822 | except ImportError: |
| 823 | dotf = False |
| 824 | if dotf: |
| 825 | tfres = StringSplit( |
| 826 | input=input, delimiter="", skip_empty=skip) |
| 827 | self.assertEqual( |
| 828 | [_.decode() for _ in tfres[1].numpy().tolist()], |
| 829 | txout[1].tolist()) |
| 830 | self.assertEqual( |
| 831 | tfres[0].numpy().tolist(), txout[0].tolist()) |
| 832 | self.assertEqual( |
| 833 | tfres[2].numpy().tolist(), txout[2].tolist()) |
| 834 | |
| 835 | exp_indices = np.array( |
| 836 | [[0, 0], [0, 1], [0, 2], [1, 0], [1, 1], [1, 2], [1, 3]]) |
| 837 | exp_text = np.array(['a', '*', 'b', 'a', ',', '*', 'b']) |
| 838 | exp_shape = np.array([2, 4]) |
| 839 | self.assertEqual(exp_text.tolist(), txout[1].tolist()) |
| 840 | self.assertEqual(exp_indices.tolist(), txout[0].tolist()) |
| 841 | self.assertEqual(exp_shape.tolist(), txout[2].tolist()) |
| 842 | |
| 843 | |
| 844 | if __name__ == "__main__": |
| 845 | unittest.main() |