Coverage for /pythoncovmergedfiles/medio/medio/usr/local/lib/python3.8/site-packages/keras/src/export/export_lib.py: 17%

161 statements  

« prev     ^ index     » next       coverage.py v7.4.0, created at 2024-01-03 07:57 +0000

1# Copyright 2023 The TensorFlow Authors. All Rights Reserved. 

2# 

3# Licensed under the Apache License, Version 2.0 (the "License"); 

4# you may not use this file except in compliance with the License. 

5# You may obtain a copy of the License at 

6# 

7# http://www.apache.org/licenses/LICENSE-2.0 

8# 

9# Unless required by applicable law or agreed to in writing, software 

10# distributed under the License is distributed on an "AS IS" BASIS, 

11# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 

12# See the License for the specific language governing permissions and 

13# limitations under the License. 

14# ============================================================================== 

15"""Library for exporting inference-only Keras models/layers.""" 

16 

17import tensorflow.compat.v2 as tf 

18from tensorflow.python.util.tf_export import keras_export 

19 

20from keras.src.engine import base_layer 

21from keras.src.engine import functional 

22from keras.src.engine import sequential 

23from keras.src.utils import io_utils 

24 

25 

26@keras_export("keras.export.ExportArchive") 

27class ExportArchive(tf.__internal__.tracking.AutoTrackable): 

28 """ExportArchive is used to write SavedModel artifacts (e.g. for inference). 

29 

30 If you have a Keras model or layer that you want to export as SavedModel for 

31 serving (e.g. via TensorFlow-Serving), you can use `ExportArchive` 

32 to configure the different serving endpoints you need to make available, 

33 as well as their signatures. Simply instantiate an `ExportArchive`, 

34 use `track()` to register the layer(s) or model(s) to be used, 

35 then use the `add_endpoint()` method to register a new serving endpoint. 

36 When done, use the `write_out()` method to save the artifact. 

37 

38 The resulting artifact is a SavedModel and can be reloaded via 

39 `tf.saved_model.load`. 

40 

41 Examples: 

42 

43 Here's how to export a model for inference. 

44 

45 ```python 

46 export_archive = ExportArchive() 

47 export_archive.track(model) 

48 export_archive.add_endpoint( 

49 name="serve", 

50 fn=model.call, 

51 input_signature=[tf.TensorSpec(shape=(None, 3), dtype=tf.float32)], 

52 ) 

53 export_archive.write_out("path/to/location") 

54 

55 # Elsewhere, we can reload the artifact and serve it. 

56 # The endpoint we added is available as a method: 

57 serving_model = tf.saved_model.load("path/to/location") 

58 outputs = serving_model.serve(inputs) 

59 ``` 

60 

61 Here's how to export a model with one endpoint for inference and one 

62 endpoint for a training-mode forward pass (e.g. with dropout on). 

63 

64 ```python 

65 export_archive = ExportArchive() 

66 export_archive.track(model) 

67 export_archive.add_endpoint( 

68 name="call_inference", 

69 fn=lambda x: model.call(x, training=False), 

70 input_signature=[tf.TensorSpec(shape=(None, 3), dtype=tf.float32)], 

71 ) 

72 export_archive.add_endpoint( 

73 name="call_training", 

74 fn=lambda x: model.call(x, training=True), 

75 input_signature=[tf.TensorSpec(shape=(None, 3), dtype=tf.float32)], 

76 ) 

77 export_archive.write_out("path/to/location") 

78 ``` 

79 

80 **Note on resource tracking:** 

81 

82 `ExportArchive` is able to automatically track all `tf.Variables` used 

83 by its endpoints, so most of the time calling `.track(model)` 

84 is not strictly required. However, if your model uses lookup layers such 

85 as `IntegerLookup`, `StringLookup`, or `TextVectorization`, 

86 it will need to be tracked explicitly via `.track(model)`. 

87 

88 Explicit tracking is also required if you need to be able to access 

89 the properties `variables`, `trainable_variables`, or 

90 `non_trainable_variables` on the revived archive. 

91 """ 

92 

93 def __init__(self): 

94 self._endpoint_names = [] 

95 self._endpoint_signatures = {} 

96 self.tensorflow_version = tf.__version__ 

97 self.variables = [] 

98 self.trainable_variables = [] 

99 self.non_trainable_variables = [] 

100 

101 @tf.__internal__.tracking.no_automatic_dependency_tracking 

102 def track(self, layer): 

103 """Track the variables (and other resources) of a layer or model.""" 

104 if not isinstance(layer, base_layer.Layer): 

105 raise ValueError( 

106 "Invalid layer type. Expected an instance of " 

107 "`keras.layers.Layer` or `keras.Model`. " 

108 f"Received instead an object of type '{type(layer)}'. " 

109 f"Object received: {layer}" 

110 ) 

111 if not layer.built: 

112 raise ValueError( 

113 "The layer provided has not yet been built. " 

114 "It must be built before export." 

115 ) 

116 

117 # Layers in `_tracked` are not part of the trackables that get saved, 

118 # because we're creating the attribute in a 

119 # no_automatic_dependency_tracking scope. 

120 if not hasattr(self, "_tracked"): 

121 self._tracked = [] 

122 self._tracked.append(layer) 

123 

124 # Variables in the lists below are actually part of the trackables 

125 # that get saved, because the lists are created in __init__. 

126 self.variables += layer.variables 

127 self.trainable_variables += layer.trainable_variables 

128 self.non_trainable_variables += layer.non_trainable_variables 

129 

130 def add_endpoint(self, name, fn, input_signature=None): 

131 """Register a new serving endpoint. 

132 

133 Arguments: 

134 name: Str, name of the endpoint. 

135 fn: A function. It should only leverage resources 

136 (e.g. `tf.Variable` objects or `tf.lookup.StaticHashTable` 

137 objects) that are available on the models/layers 

138 tracked by the `ExportArchive` (you can call `.track(model)` 

139 to track a new model). 

140 The shape and dtype of the inputs to the function must be 

141 known. For that purpose, you can either 1) make sure that 

142 `fn` is a `tf.function` that has been called at least once, or 

143 2) provide an `input_signature` argument that specifies the 

144 shape and dtype of the inputs (see below). 

145 input_signature: Used to specify the shape and dtype of the 

146 inputs to `fn`. List of `tf.TensorSpec` objects (one 

147 per positional input argument of `fn`). Nested arguments are 

148 allowed (see below for an example showing a Functional model 

149 with 2 input arguments). 

150 

151 Example: 

152 

153 Adding an endpoint using the `input_signature` argument when the 

154 model has a single input argument: 

155 

156 ```python 

157 export_archive = ExportArchive() 

158 export_archive.track(model) 

159 export_archive.add_endpoint( 

160 name="serve", 

161 fn=model.call, 

162 input_signature=[tf.TensorSpec(shape=(None, 3), dtype=tf.float32)], 

163 ) 

164 ``` 

165 

166 Adding an endpoint using the `input_signature` argument when the 

167 model has two positional input arguments: 

168 

169 ```python 

170 export_archive = ExportArchive() 

171 export_archive.track(model) 

172 export_archive.add_endpoint( 

173 name="serve", 

174 fn=model.call, 

175 input_signature=[ 

176 tf.TensorSpec(shape=(None, 3), dtype=tf.float32), 

177 tf.TensorSpec(shape=(None, 4), dtype=tf.float32), 

178 ], 

179 ) 

180 ``` 

181 

182 Adding an endpoint using the `input_signature` argument when the 

183 model has one input argument that is a list of 2 tensors (e.g. 

184 a Functional model with 2 inputs): 

185 

186 ```python 

187 model = keras.Model(inputs=[x1, x2], outputs=outputs) 

188 

189 export_archive = ExportArchive() 

190 export_archive.track(model) 

191 export_archive.add_endpoint( 

192 name="serve", 

193 fn=model.call, 

194 input_signature=[ 

195 [ 

196 tf.TensorSpec(shape=(None, 3), dtype=tf.float32), 

197 tf.TensorSpec(shape=(None, 4), dtype=tf.float32), 

198 ], 

199 ], 

200 ) 

201 ``` 

202 

203 This also works with dictionary inputs: 

204 

205 ```python 

206 model = keras.Model(inputs={"x1": x1, "x2": x2}, outputs=outputs) 

207 

208 export_archive = ExportArchive() 

209 export_archive.track(model) 

210 export_archive.add_endpoint( 

211 name="serve", 

212 fn=model.call, 

213 input_signature=[ 

214 { 

215 "x1": tf.TensorSpec(shape=(None, 3), dtype=tf.float32), 

216 "x2": tf.TensorSpec(shape=(None, 4), dtype=tf.float32), 

217 }, 

218 ], 

219 ) 

220 ``` 

221 

222 Adding an endpoint that is a `tf.function`: 

223 

224 ```python 

225 @tf.function() 

226 def serving_fn(x): 

227 return model(x) 

228 

229 # The function must be traced, i.e. it must be called at least once. 

230 serving_fn(tf.random.normal(shape=(2, 3))) 

231 

232 export_archive = ExportArchive() 

233 export_archive.track(model) 

234 export_archive.add_endpoint(name="serve", fn=serving_fn) 

235 ``` 

236 """ 

237 if name in self._endpoint_names: 

238 raise ValueError(f"Endpoint name '{name}' is already taken.") 

239 

240 if input_signature: 

241 decorated_fn = tf.function(fn, input_signature=input_signature) 

242 self._endpoint_signatures[name] = input_signature 

243 else: 

244 if isinstance(fn, tf.types.experimental.GenericFunction): 

245 if not fn._list_all_concrete_functions(): 

246 raise ValueError( 

247 f"The provided tf.function '{fn}' " 

248 "has never been called. " 

249 "To specify the expected shape and dtype " 

250 "of the function's arguments, " 

251 "you must either provide a function that " 

252 "has been called at least once, or alternatively pass " 

253 "an `input_signature` argument in `add_endpoint()`." 

254 ) 

255 decorated_fn = fn 

256 else: 

257 raise ValueError( 

258 "If the `fn` argument provided is not a `tf.function`, " 

259 "you must provide an `input_signature` argument to " 

260 "specify the shape and dtype of the function arguments. " 

261 "Example:\n\n" 

262 "export_archive.add_endpoint(\n" 

263 " name='call',\n" 

264 " fn=model.call,\n" 

265 " input_signature=[\n" 

266 " tf.TensorSpec(\n" 

267 " shape=(None, 224, 224, 3),\n" 

268 " dtype=tf.float32,\n" 

269 " )\n" 

270 " ],\n" 

271 ")" 

272 ) 

273 setattr(self, name, decorated_fn) 

274 self._endpoint_names.append(name) 

275 

276 def add_variable_collection(self, name, variables): 

277 """Register a set of variables to be retrieved after reloading. 

278 

279 Arguments: 

280 name: The string name for the collection. 

281 variables: A tuple/list/set of `tf.Variable` instances. 

282 

283 Example: 

284 

285 ```python 

286 export_archive = ExportArchive() 

287 export_archive.track(model) 

288 # Register an endpoint 

289 export_archive.add_endpoint( 

290 name="serve", 

291 fn=model.call, 

292 input_signature=[tf.TensorSpec(shape=(None, 3), dtype=tf.float32)], 

293 ) 

294 # Save a variable collection 

295 export_archive.add_variable_collection( 

296 name="optimizer_variables", variables=model.optimizer.variables) 

297 export_archive.write_out("path/to/location") 

298 

299 # Reload the object 

300 revived_object = tf.saved_model.load("path/to/location") 

301 # Retrieve the variables 

302 optimizer_variables = revived_object.optimizer_variables 

303 ``` 

304 """ 

305 if not isinstance(variables, (list, tuple, set)): 

306 raise ValueError( 

307 "Expected `variables` to be a list/tuple/set. " 

308 f"Received instead object of type '{type(variables)}'." 

309 ) 

310 if not all(isinstance(v, tf.Variable) for v in variables): 

311 raise ValueError( 

312 "Expected all elements in `variables` to be " 

313 "`tf.Variable` instances. Found instead the following types: " 

314 f"{list(set(type(v) for v in variables))}" 

315 ) 

316 setattr(self, name, list(variables)) 

317 

318 def write_out(self, filepath, options=None): 

319 """Write the corresponding SavedModel to disk. 

320 

321 Arguments: 

322 filepath: `str` or `pathlib.Path` object. 

323 Path where to save the artifact. 

324 options: `tf.saved_model.SaveOptions` object that specifies 

325 SavedModel saving options. 

326 

327 **Note on TF-Serving**: all endpoints registered via `add_endpoint()` 

328 are made visible for TF-Serving in the SavedModel artifact. In addition, 

329 the first endpoint registered is made visible under the alias 

330 `"serving_default"` (unless an endpoint with the name 

331 `"serving_default"` was already registered manually), 

332 since TF-Serving requires this endpoint to be set. 

333 """ 

334 if not self._endpoint_names: 

335 raise ValueError( 

336 "No endpoints have been set yet. Call add_endpoint()." 

337 ) 

338 self._filter_and_track_resources() 

339 

340 signatures = {} 

341 for name in self._endpoint_names: 

342 signatures[name] = self._get_concrete_fn(name) 

343 # Add "serving_default" signature key for TFServing 

344 if "serving_default" not in self._endpoint_names: 

345 signatures["serving_default"] = self._get_concrete_fn( 

346 self._endpoint_names[0] 

347 ) 

348 tf.saved_model.save( 

349 self, filepath, options=options, signatures=signatures 

350 ) 

351 # Print out available endpoints 

352 endpoints = "\n\n".join( 

353 _print_signature(getattr(self, name), name) 

354 for name in self._endpoint_names 

355 ) 

356 io_utils.print_msg( 

357 f"Saved artifact at '{filepath}'. " 

358 "The following endpoints are available:\n\n" 

359 f"{endpoints}" 

360 ) 

361 

362 def _get_concrete_fn(self, endpoint): 

363 """Workaround for some SavedModel quirks.""" 

364 if endpoint in self._endpoint_signatures: 

365 return getattr(self, endpoint) 

366 else: 

367 traces = getattr(self, endpoint)._trackable_children("saved_model") 

368 return list(traces.values())[0] 

369 

370 def _get_variables_used_by_endpoints(self): 

371 fns = [self._get_concrete_fn(name) for name in self._endpoint_names] 

372 return _list_variables_used_by_fns(fns) 

373 

374 def _filter_and_track_resources(self): 

375 """Track resources used by endpoints / referenced in `track()` calls.""" 

376 # Start by extracting variables from endpoints. 

377 fns = [self._get_concrete_fn(name) for name in self._endpoint_names] 

378 tvs, ntvs = _list_variables_used_by_fns(fns) 

379 self._all_variables = list(tvs + ntvs) 

380 

381 # Next, track lookup tables. 

382 # Hopefully, one day this will be automated at the tf.function level. 

383 self._misc_assets = [] 

384 from keras.src.layers.preprocessing.index_lookup import IndexLookup 

385 

386 if hasattr(self, "_tracked"): 

387 for root in self._tracked: 

388 descendants = tf.train.TrackableView(root).descendants() 

389 for trackable in descendants: 

390 if isinstance(trackable, IndexLookup): 

391 self._misc_assets.append(trackable) 

392 

393 

394def export_model(model, filepath): 

395 export_archive = ExportArchive() 

396 export_archive.track(model) 

397 if isinstance(model, (functional.Functional, sequential.Sequential)): 

398 input_signature = tf.nest.map_structure(_make_tensor_spec, model.inputs) 

399 if isinstance(input_signature, list) and len(input_signature) > 1: 

400 input_signature = [input_signature] 

401 export_archive.add_endpoint("serve", model.__call__, input_signature) 

402 else: 

403 save_spec = model._get_save_spec() 

404 if not save_spec: 

405 raise ValueError( 

406 "The model provided has never called. " 

407 "It must be called at least once before export." 

408 ) 

409 input_signature = [save_spec] 

410 export_archive.add_endpoint("serve", model.__call__, input_signature) 

411 export_archive.write_out(filepath) 

412 

413 

414class ReloadedLayer(base_layer.Layer): 

415 """Reload a Keras model/layer that was saved via SavedModel / ExportArchive. 

416 

417 Arguments: 

418 filepath: `str` or `pathlib.Path` object. The path to the SavedModel. 

419 call_endpoint: Name of the endpoint to use as the `call()` method 

420 of the reloaded layer. If the SavedModel was created 

421 via `model.export()`, 

422 then the default endpoint name is `'serve'`. In other cases 

423 it may be named `'serving_default'`. 

424 

425 Example: 

426 

427 ```python 

428 model.export("path/to/artifact") 

429 reloaded_layer = ReloadedLayer("path/to/artifact") 

430 outputs = reloaded_layer(inputs) 

431 ``` 

432 

433 The reloaded object can be used like a regular Keras layer, and supports 

434 training/fine-tuning of its trainable weights. Note that the reloaded 

435 object retains none of the internal structure or custom methods of the 

436 original object -- it's a brand new layer created around the saved 

437 function. 

438 

439 **Limitations:** 

440 

441 * Only call endpoints with a single `inputs` tensor argument 

442 (which may optionally be a dict/tuple/list of tensors) are supported. 

443 For endpoints with multiple separate input tensor arguments, consider 

444 subclassing `ReloadedLayer` and implementing a `call()` method with a 

445 custom signature. 

446 * If you need training-time behavior to differ from inference-time behavior 

447 (i.e. if you need the reloaded object to support a `training=True` argument 

448 in `__call__()`), make sure that the training-time call function is 

449 saved as a standalone endpoint in the artifact, and provide its name 

450 to the `ReloadedLayer` via the `call_training_endpoint` argument. 

451 """ 

452 

453 def __init__( 

454 self, 

455 filepath, 

456 call_endpoint="serve", 

457 call_training_endpoint=None, 

458 trainable=True, 

459 name=None, 

460 dtype=None, 

461 ): 

462 # Initialize an empty layer, then add_weight() etc. as needed. 

463 super().__init__(trainable=trainable, name=name, dtype=dtype) 

464 

465 self._reloaded_obj = tf.saved_model.load(filepath) 

466 

467 self.filepath = filepath 

468 self.call_endpoint = call_endpoint 

469 self.call_training_endpoint = call_training_endpoint 

470 

471 # Resolve the call function. 

472 if hasattr(self._reloaded_obj, call_endpoint): 

473 # Case 1: it's set as an attribute. 

474 self.call_endpoint_fn = getattr(self._reloaded_obj, call_endpoint) 

475 elif call_endpoint in self._reloaded_obj.signatures: 

476 # Case 2: it's listed in the `signatures` field. 

477 self.call_endpoint_fn = self._reloaded_obj.signatures[call_endpoint] 

478 else: 

479 raise ValueError( 

480 f"The endpoint '{call_endpoint}' is neither an " 

481 "attribute of the reloaded SavedModel, nor an entry " 

482 "in the `signatures` field of the reloaded SavedModel. " 

483 ) 

484 

485 # Resolving the training function. 

486 if call_training_endpoint: 

487 if hasattr(self._reloaded_obj, call_training_endpoint): 

488 self.call_training_endpoint_fn = getattr( 

489 self._reloaded_obj, call_training_endpoint 

490 ) 

491 elif call_training_endpoint in self._reloaded_obj.signatures: 

492 self.call_training_endpoint_fn = self._reloaded_obj.signatures[ 

493 call_training_endpoint 

494 ] 

495 else: 

496 raise ValueError( 

497 f"The endpoint '{call_training_endpoint}' is " 

498 "neither an attribute of the reloaded SavedModel, " 

499 "nor an entry in the `signatures` field of " 

500 "the reloaded SavedModel. " 

501 ) 

502 

503 # Add trainable and non-trainable weights from the call_endpoint_fn. 

504 all_fns = [self.call_endpoint_fn] 

505 if call_training_endpoint: 

506 all_fns.append(self.call_training_endpoint_fn) 

507 tvs, ntvs = _list_variables_used_by_fns(all_fns) 

508 for v in tvs: 

509 self._add_existing_weight(v, trainable=True) 

510 for v in ntvs: 

511 self._add_existing_weight(v, trainable=False) 

512 self.built = True 

513 

514 def _add_existing_weight(self, weight, trainable): 

515 """Calls add_weight() to register but not create an existing weight.""" 

516 self.add_weight( 

517 name=weight.name, 

518 shape=weight.shape, 

519 dtype=weight.dtype, 

520 trainable=trainable, 

521 getter=lambda *_, **__: weight, 

522 ) 

523 

524 def call(self, inputs, training=False, **kwargs): 

525 if training: 

526 if self.call_training_endpoint: 

527 return self.call_training_endpoint_fn(inputs, **kwargs) 

528 return self.call_endpoint_fn(inputs, **kwargs) 

529 

530 def get_config(self): 

531 base_config = super().get_config() 

532 config = { 

533 # Note: this is not intended to be portable. 

534 "filepath": self.filepath, 

535 "call_endpoint": self.call_endpoint, 

536 "call_training_endpoint": self.call_training_endpoint, 

537 } 

538 return {**base_config, **config} 

539 

540 

541def _make_tensor_spec(x): 

542 return tf.TensorSpec(x.shape, dtype=x.dtype, name=x.name) 

543 

544 

545def _print_signature(fn, name): 

546 concrete_fn = fn._list_all_concrete_functions()[0] 

547 pprinted_signature = concrete_fn.pretty_printed_signature(verbose=True) 

548 lines = pprinted_signature.split("\n") 

549 lines = [f"* Endpoint '{name}'"] + lines[1:] 

550 endpoint = "\n".join(lines) 

551 return endpoint 

552 

553 

554def _list_variables_used_by_fns(fns): 

555 trainable_variables = [] 

556 non_trainable_variables = [] 

557 trainable_variables_ids = set() 

558 non_trainable_variables_ids = set() 

559 for fn in fns: 

560 if hasattr(fn, "concrete_functions"): 

561 concrete_functions = fn.concrete_functions 

562 elif hasattr(fn, "get_concrete_function"): 

563 concrete_functions = [fn.get_concrete_function()] 

564 else: 

565 concrete_functions = [fn] 

566 for concrete_fn in concrete_functions: 

567 for v in concrete_fn.trainable_variables: 

568 if id(v) not in trainable_variables_ids: 

569 trainable_variables.append(v) 

570 trainable_variables_ids.add(id(v)) 

571 

572 for v in concrete_fn.variables: 

573 if ( 

574 id(v) not in trainable_variables_ids 

575 and id(v) not in non_trainable_variables_ids 

576 ): 

577 non_trainable_variables.append(v) 

578 non_trainable_variables_ids.add(id(v)) 

579 return trainable_variables, non_trainable_variables 

580