1# Copyright 2015 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"""Script Language Operators.""" 16 17# pylint: disable=g-bad-name 18from __future__ import absolute_import 19from __future__ import division 20from __future__ import print_function 21 22import threading 23 24# Used by py_util.cc to get tracebacks. 25import traceback # pylint: disable=unused-import 26import weakref 27 28import numpy as np 29import six 30 31from tensorflow.python.eager import backprop 32from tensorflow.python.eager import context 33from tensorflow.python.framework import constant_op 34from tensorflow.python.framework import func_graph 35from tensorflow.python.framework import function 36from tensorflow.python.framework import ops 37from tensorflow.python.lib.core import _pywrap_py_func 38from tensorflow.python.ops import gen_script_ops 39from tensorflow.python.ops import resource_variable_ops 40from tensorflow.python.util import compat 41from tensorflow.python.util import deprecation 42from tensorflow.python.util import dispatch 43from tensorflow.python.util import lazy_loader 44from tensorflow.python.util import nest 45from tensorflow.python.util import tf_inspect 46from tensorflow.python.util.tf_export import tf_export 47 48autograph = lazy_loader.LazyLoader( 49 "autograph", globals(), 50 "tensorflow.python.autograph.impl.api") 51 52 53# Map from EagerPyFunc token to tuple (tape, eager args, eager outputs); 54# used for differentiation. 55tape_cache = {} 56 57 58def _maybe_copy_to_context_device(tensor, device_name): 59 """Copy an EagerTensor to the current device if it's not on `device_name`.""" 60 in_device = tensor.backing_device 61 if device_name == in_device: 62 return tensor 63 else: 64 # Note that EagerTensor._copy bypasses the placer and copies to the context 65 # device, which means e.g. int32 Tensors which would normally be forced onto 66 # the CPU can instead be placed on the GPU. This is necessary so that the 67 # PyFunc kernel always returns Tensors on the device it's executing on. 68 return tensor._copy() # pylint: disable=protected-access 69 70 71class EagerFunc(object): 72 """A wrapper for a function owned by an EagerPyFunc.""" 73 74 def __init__(self, func, Tout, is_grad_func, use_tape_cache=True): 75 """Constructs an EagerFunc. 76 77 Args: 78 func: The function to wrap. 79 Tout: A list of datatypes for the output; an empty list if the output is 80 None. 81 is_grad_func: Whether this EagerFunc is the gradient of another 82 EagerPyFunc. 83 use_tape_cache: (Optional.) Whether to cache `func` in the `tape_cache`. 84 For additional information, see description of `_eager_py_func`. 85 This parameter should be removed once the #35084 issue is fixed. 86 """ 87 self._func = func 88 self._out_dtypes = Tout 89 self._is_grad_func = is_grad_func 90 self._use_tape_cache = use_tape_cache 91 92 def _convert(self, value, dtype): 93 """Converts `value` to a tensor of type `dtype`, with error checking. 94 95 Args: 96 value: The tensor to convert. 97 dtype: The desired dtype. 98 99 Returns: 100 A tensor of type `dtype`, or a zeros tensor if value is None and 101 this function is in fact a gradient function. 102 103 Raises: 104 RuntimeError: if `value` is a variable. 105 """ 106 107 if isinstance(value, resource_variable_ops.ResourceVariable): 108 raise RuntimeError( 109 "Attempting to return a variable from an eagerly executed py_func. " 110 "Only numeric data structures like Tensors or NumPy arrays should " 111 "be returned; to return the value of a variable, make sure to obtain " 112 "the Tensor backing it by calling `.read_value()` on the variable in " 113 "question: %s" % value) 114 if value is None and self._is_grad_func: 115 # Gradient functions may legitimately return a list that contains 116 # both Tensors and Python Nones. Unfortunately this breaks the 117 # OpKernel, so for now we replace None objects with zeros, which is 118 # mathematically correct but will prevent short-circuiting gradient 119 # computations. 120 # 121 # TODO(akshayka): Make it possible to return a list of both Tensors and 122 # Nones from an EagerPyFunc. 123 return constant_op.constant(0.0, dtype=dtype) 124 return ops.convert_to_tensor(value, dtype=dtype) 125 126 def __call__(self, device, token, args): 127 """Passes `args` to `self._func`, which is executed eagerly.""" 128 129 with context.eager_mode(), backprop.GradientTape() as tape: 130 # Only watch tensors with a floating or complex dtype. 131 for tensor in args: 132 for t in nest.flatten(tensor): 133 if t.dtype.is_floating or t.dtype.is_complex: 134 tape.watch(t) 135 ret = self._func(*args) 136 # copy the returned tensors to the PyFunc op's device if necessary. 137 device_name = device 138 if device_name is None: 139 # "None" here means "CPU", from the nullptr convention with C++ device 140 # pointers. 141 device_name = "/job:localhost/replica:0/task:0/device:CPU:0" 142 with ops.device(device): 143 if isinstance(ret, (tuple, list)): 144 outputs = [ 145 _maybe_copy_to_context_device(self._convert(x, dtype=dtype), 146 device_name) 147 for (x, dtype) in zip(ret, self._out_dtypes) 148 ] 149 elif ret is None: 150 outputs = None 151 else: 152 outputs = _maybe_copy_to_context_device( 153 self._convert(ret, dtype=self._out_dtypes[0]), device_name) 154 if self._use_tape_cache: 155 tape_cache[compat.as_bytes(token)] = (tape, args, outputs) 156 return outputs 157 158 159class FuncRegistry(object): 160 """A helper class to keep track of registered py functions. 161 162 FuncRegistry keeps a map from unique tokens (string) to python 163 functions, which takes numpy arrays and outputs numpy arrays. 164 """ 165 166 def __init__(self): 167 self._lock = threading.Lock() 168 self._unique_id = 0 # GUARDED_BY(self._lock) 169 # Only store weakrefs to the functions. The strong reference is stored in 170 # the graph. 171 self._funcs = weakref.WeakValueDictionary() 172 173 @property 174 def _ctx(self): 175 # N.B. This is needed to support calling py_func with GPU tensors, 176 # which must be transferred to CPU if used in any of the NumPy APIs. 177 context.ensure_initialized() 178 return context.context()._handle # pylint: disable=protected-access 179 180 def insert(self, func): 181 """Registers `func` and returns a unique token for this entry.""" 182 token = self._next_unique_token() 183 # Store a weakref to the function 184 self._funcs[token] = func 185 return token 186 187 def remove(self, token): 188 """Removes the registered function corresponding to `token`.""" 189 self._funcs.pop(token, None) 190 191 @staticmethod 192 def _convert(value, dtype=None): 193 """Converts an arg to numpy, avoiding dangerous string and unicode dtypes. 194 195 Numpy pads with zeros when using string and unicode dtypes if different 196 components of a tensor have different lengths. This is bad: ignoring the 197 padding is wrong for text data, and removing the padding is wrong for binary 198 data. To avoid this bug, we redo the conversion using an object dtype. 199 Additionally, we convert unicode strings to (byte-)strings for 200 compatibility. 201 202 Args: 203 value: Value to convert to a numpy array. 204 dtype: (Optional.) Desired NumPy type for the returned value. 205 206 Returns: 207 A numpy array. 208 """ 209 result = np.asarray(value, dtype=dtype, order="C") 210 if result.dtype.char == "S" and result is not value: 211 return np.asarray(value, order="C", dtype=object) 212 elif result.dtype.char == "U" and result is not value: 213 value = np.vectorize(lambda x: x.encode("utf8"))(value) 214 return np.asarray(value, order="C", dtype=object) 215 elif result.dtype.char == "U": 216 return result.astype(np.bytes_) 217 else: 218 return result 219 220 def __call__(self, token, device, args): 221 """Calls the registered function for `token` with args. 222 223 Args: 224 token: A key into this `FuncRegistry` identifying which function to call. 225 device: Name of the device on which outputs of `token`'s corresponding 226 operation should be placed. Used iff the function registered for `token` 227 is an EagerPyFunc. 228 args: The arguments to pass to the function registered for `token`. 229 230 Returns: 231 The output of the function registered for `token`. 232 233 Raises: 234 ValueError: if no function is registered for `token`. 235 """ 236 func = self._funcs.get(token, None) 237 if func is None: 238 raise ValueError("callback %s is not found" % token) 239 if isinstance(func, EagerFunc): 240 # NB: Different invocations of the same py_func will share the same 241 # token, and the entries they stash in the tape_cache will collide. 242 # In practice, when executing a graph, this should only happen if 243 # the py_func is in a while_loop whose iterations are run in parallel 244 # or if the graph is being driven by concurrent session.run() calls. 245 # 246 # TODO(akshayka): Key the tape cache in a thread-safe way. 247 return func(device, token, args) 248 else: 249 ret = func(*args) 250 # Strings seem to lead to a memory leak here if they're not wrapped in a 251 # list. 252 if isinstance(ret, six.binary_type): 253 ret = [ret] 254 # Ensures that we return either a single numpy array or a list of numpy 255 # arrays. 256 if isinstance(ret, (tuple, list)): 257 return [self._convert(x) for x in ret] 258 else: 259 return self._convert(ret) 260 261 def size(self): 262 """Returns how many functions are currently registered.""" 263 return len(self._funcs) 264 265 def _next_unique_token(self): 266 """Returns a unique token.""" 267 with self._lock: 268 uid = self._unique_id 269 self._unique_id += 1 270 return "pyfunc_%d" % uid 271 272 273# Global registry for py functions. 274_py_funcs = FuncRegistry() 275 276_pywrap_py_func.initialize_py_trampoline(_py_funcs) 277 278 279def _internal_py_func(func, 280 inp, 281 Tout, 282 stateful=None, 283 eager=False, 284 is_grad_func=False, 285 name=None, 286 use_tape_cache=True): 287 """See documentation for py_func and eager_py_func.""" 288 if not callable(func): 289 raise ValueError("Expected func to be callable, got func of type {}".format( 290 type(func))) 291 292 original_func = func 293 func = autograph.do_not_convert(func) 294 295 is_list_or_tuple = False 296 if isinstance(Tout, (list, tuple)): 297 is_list_or_tuple = True 298 else: 299 Tout = [Tout] 300 301 if eager: 302 func = EagerFunc(func, Tout, is_grad_func, use_tape_cache=use_tape_cache) 303 304 # Tying the registered function's lifetime with the current default graph is 305 # not reliable. For example, Estimator-based binaries may switch graphs in 306 # between model training end evaluation, via saved_model. Those binaries work 307 # because the original function is global, and break once the registered 308 # function is an anonymous lambda, like the one produced by do_not_convert. 309 # To avoid breaking those cases, we attach the wrapper to the original 310 # function so that their lifetime is connected. 311 # TODO(b/144286616): Remove this. 312 if tf_inspect.isfunction(original_func): 313 # Note: this check is needed because original_func may be a descriptor 314 # (https://docs.python.org/3/howto/descriptor.html) 315 # and we can't attach attributes to those. 316 original_func.ag_dnc_wrapper__ = func 317 318 token = _py_funcs.insert(func) 319 # We tie the registered function's lifetime with the current default graph, 320 # i.e., when the current graph is destroyed, we remove its py funcs. 321 graph = ops.get_default_graph() 322 323 while True: 324 current_graph = graph 325 if isinstance(graph, function._FuncGraph): # pylint: disable=protected-access 326 graph = graph._outer_graph # pylint: disable=protected-access 327 elif isinstance(graph, func_graph.FuncGraph): 328 graph = graph.outer_graph 329 if graph is current_graph: 330 break 331 332 # TODO(zhifengc): Consider adding a Graph method to collect 333 # `cleanup` objects in one of its member. 334 if not hasattr(graph, "_py_funcs_used_in_graph"): 335 graph._py_funcs_used_in_graph = [] # pylint: disable=protected-access 336 337 # Store a reference to the function in the graph to ensure it stays alive 338 # as long as the graph lives. When the graph is destroyed, the function 339 # is left to the garbage collector for destruction as well. 340 graph._py_funcs_used_in_graph.append(func) # pylint: disable=protected-access 341 342 if eager: 343 result = gen_script_ops.eager_py_func( 344 input=inp, 345 token=token, 346 is_async=context.is_async(), 347 Tout=Tout, 348 name=name) 349 else: 350 if stateful: 351 result = gen_script_ops.py_func( 352 input=inp, token=token, Tout=Tout, name=name) 353 else: 354 result = gen_script_ops.py_func_stateless( 355 input=inp, token=token, Tout=Tout, name=name) 356 return result if is_list_or_tuple else result[0] 357 358 359# TODO(akshayka): Implement higher-order derivatives. 360@ops.RegisterGradient("EagerPyFunc") 361def _EagerPyFuncGrad(op, *dy): 362 """Computes the gradient of an EagerPyFunc.""" 363 364 token = op.get_attr("token") 365 366 def eagerly_executed_grad(*dy): 367 tape, eager_inputs, eager_outputs = tape_cache.pop(compat.as_bytes(token)) 368 return tape.gradient(eager_outputs, eager_inputs, output_gradients=dy) 369 370 with ops.control_dependencies(op.outputs): 371 return _internal_py_func( 372 func=eagerly_executed_grad, 373 inp=dy, 374 Tout=[tensor.dtype for tensor in op.inputs], 375 eager=True, 376 is_grad_func=True) 377 378 379def _eager_py_func(func, inp, Tout, name=None, use_tape_cache=True): 380 """Wraps a python function into a TensorFlow op that executes it eagerly. 381 382 This function is the internal implementation for `eager_py_func`, see the 383 `eager_py_func` docstring for the full description. 384 385 Note: this function as a layer of indirection was added with one 386 specific purpose: as a workaround for github issue #35084. 387 It does all the same as `eager_py_func` used to do with one difference: 388 it can be used to instruct underlying EagerFunc not to use `tape_cache` 389 to avoid memory leak. When the issue #35084 is fixed - this function should 390 be removed, its body should be moved back to become the body of 391 `eager_py_func` and all the call sites should be reverted to 392 using `eager_py_func` without `use_tape_cache` argument of any value. 393 394 Args: 395 func: A Python function which accepts a list of `Tensor` objects having 396 element types that match the corresponding `tf.Tensor` objects in `inp` 397 and returns a list of `Tensor` objects (or a single `Tensor`, or `None`) 398 having element types that match the corresponding values in `Tout`. 399 inp: A list of `Tensor` objects. 400 Tout: A list or tuple of tensorflow data types or a single tensorflow data 401 type if there is only one, indicating what `func` returns; an empty list 402 if no value is returned (i.e., if the return value is `None`). 403 name: A name for the operation (optional). 404 use_tape_cache: (Optional.) Whether to cache `func` in the `tape_cache`. 405 For additional information, see description of `_eager_py_func`. 406 This parameter should be removed once the #35084 issue is fixed. 407 408 Returns: 409 A list of `Tensor` or a single `Tensor` which `func` computes; an empty list 410 if `func` returns None. 411 """ 412 if ops.executing_eagerly_outside_functions(): 413 with ops.device(context.context().host_address_space()): 414 return _internal_py_func( 415 func=func, 416 inp=inp, 417 Tout=Tout, 418 eager=True, 419 name=name, 420 use_tape_cache=use_tape_cache) 421 422 return _internal_py_func( 423 func=func, 424 inp=inp, 425 Tout=Tout, 426 eager=True, 427 name=name, 428 use_tape_cache=use_tape_cache) 429 430 431@tf_export("py_function") 432@dispatch.add_dispatch_support 433def eager_py_func(func, inp, Tout, name=None): 434 """Wraps a python function into a TensorFlow op that executes it eagerly. 435 436 This function allows expressing computations in a TensorFlow graph as 437 Python functions. In particular, it wraps a Python function `func` 438 in a once-differentiable TensorFlow operation that executes it with eager 439 execution enabled. As a consequence, `tf.py_function` makes it 440 possible to express control flow using Python constructs (`if`, `while`, 441 `for`, etc.), instead of TensorFlow control flow constructs (`tf.cond`, 442 `tf.while_loop`). For example, you might use `tf.py_function` to 443 implement the log huber function: 444 445 ```python 446 def log_huber(x, m): 447 if tf.abs(x) <= m: 448 return x**2 449 else: 450 return m**2 * (1 - 2 * tf.math.log(m) + tf.math.log(x**2)) 451 452 x = tf.compat.v1.placeholder(tf.float32) 453 m = tf.compat.v1.placeholder(tf.float32) 454 455 y = tf.py_function(func=log_huber, inp=[x, m], Tout=tf.float32) 456 dy_dx = tf.gradients(y, x)[0] 457 458 with tf.compat.v1.Session() as sess: 459 # The session executes `log_huber` eagerly. Given the feed values below, 460 # it will take the first branch, so `y` evaluates to 1.0 and 461 # `dy_dx` evaluates to 2.0. 462 y, dy_dx = sess.run([y, dy_dx], feed_dict={x: 1.0, m: 2.0}) 463 ``` 464 465 You can also use `tf.py_function` to debug your models at runtime 466 using Python tools, i.e., you can isolate portions of your code that 467 you want to debug, wrap them in Python functions and insert `pdb` tracepoints 468 or print statements as desired, and wrap those functions in 469 `tf.py_function`. 470 471 For more information on eager execution, see the 472 [Eager guide](https://tensorflow.org/guide/eager). 473 474 `tf.py_function` is similar in spirit to `tf.compat.v1.py_func`, but unlike 475 the latter, the former lets you use TensorFlow operations in the wrapped 476 Python function. In particular, while `tf.compat.v1.py_func` only runs on CPUs 477 and 478 wraps functions that take NumPy arrays as inputs and return NumPy arrays as 479 outputs, `tf.py_function` can be placed on GPUs and wraps functions 480 that take Tensors as inputs, execute TensorFlow operations in their bodies, 481 and return Tensors as outputs. 482 483 Like `tf.compat.v1.py_func`, `tf.py_function` has the following limitations 484 with respect to serialization and distribution: 485 486 * The body of the function (i.e. `func`) will not be serialized in a 487 `GraphDef`. Therefore, you should not use this function if you need to 488 serialize your model and restore it in a different environment. 489 490 * The operation must run in the same address space as the Python program 491 that calls `tf.py_function()`. If you are using distributed 492 TensorFlow, you must run a `tf.distribute.Server` in the same process as the 493 program that calls `tf.py_function()` and you must pin the created 494 operation to a device in that server (e.g. using `with tf.device():`). 495 496 497 Args: 498 func: A Python function which accepts a list of `Tensor` objects having 499 element types that match the corresponding `tf.Tensor` objects in `inp` 500 and returns a list of `Tensor` objects (or a single `Tensor`, or `None`) 501 having element types that match the corresponding values in `Tout`. 502 inp: A list of `Tensor` objects. 503 Tout: A list or tuple of tensorflow data types or a single tensorflow data 504 type if there is only one, indicating what `func` returns; an empty list 505 if no value is returned (i.e., if the return value is `None`). 506 name: A name for the operation (optional). 507 508 Returns: 509 A list of `Tensor` or a single `Tensor` which `func` computes; an empty list 510 if `func` returns None. 511 """ 512 return _eager_py_func( 513 func=func, inp=inp, Tout=Tout, name=name, use_tape_cache=True) 514 515 516def py_func_common(func, inp, Tout, stateful=True, name=None): 517 """Wraps a python function and uses it as a TensorFlow op. 518 519 Given a python function `func`, which takes numpy arrays as its 520 arguments and returns numpy arrays as its outputs, wrap this function as an 521 operation in a TensorFlow graph. The following snippet constructs a simple 522 TensorFlow graph that invokes the `np.sinh()` NumPy function as a operation 523 in the graph: 524 525 ```python 526 def my_func(x): 527 # x will be a numpy array with the contents of the placeholder below 528 return np.sinh(x) 529 input = tf.compat.v1.placeholder(tf.float32) 530 y = tf.compat.v1.py_func(my_func, [input], tf.float32) 531 ``` 532 533 **N.B.** The `tf.compat.v1.py_func()` operation has the following known 534 limitations: 535 536 * The body of the function (i.e. `func`) will not be serialized in a 537 `GraphDef`. Therefore, you should not use this function if you need to 538 serialize your model and restore it in a different environment. 539 540 * The operation must run in the same address space as the Python program 541 that calls `tf.compat.v1.py_func()`. If you are using distributed 542 TensorFlow, you 543 must run a `tf.distribute.Server` in the same process as the program that 544 calls 545 `tf.compat.v1.py_func()` and you must pin the created operation to a device 546 in that 547 server (e.g. using `with tf.device():`). 548 549 Note: It produces tensors of unknown shape and rank as shape inference 550 does not work on arbitrary Python code. 551 If you need the shape, you need to set it based on statically 552 available information. 553 554 E.g. 555 ```python 556 import tensorflow as tf 557 import numpy as np 558 559 def make_synthetic_data(i): 560 return np.cast[np.uint8](i) * np.ones([20,256,256,3], 561 dtype=np.float32) / 10. 562 563 def preprocess_fn(i): 564 ones = tf.py_function(make_synthetic_data,[i],tf.float32) 565 ones.set_shape(tf.TensorShape([None, None, None, None])) 566 ones = tf.image.resize(ones, [224,224]) 567 return ones 568 569 ds = tf.data.Dataset.range(10) 570 ds = ds.map(preprocess_fn) 571 ``` 572 573 Args: 574 func: A Python function, which accepts `ndarray` objects as arguments and 575 returns a list of `ndarray` objects (or a single `ndarray`). This function 576 must accept as many arguments as there are tensors in `inp`, and these 577 argument types will match the corresponding `tf.Tensor` objects in `inp`. 578 The returns `ndarray`s must match the number and types defined `Tout`. 579 Important Note: Input and output numpy `ndarray`s of `func` are not 580 guaranteed to be copies. In some cases their underlying memory will be 581 shared with the corresponding TensorFlow tensors. In-place modification 582 or storing `func` input or return values in python datastructures 583 without explicit (np.)copy can have non-deterministic consequences. 584 inp: A list of `Tensor` objects. 585 Tout: A list or tuple of tensorflow data types or a single tensorflow data 586 type if there is only one, indicating what `func` returns. 587 stateful: (Boolean.) If True, the function should be considered stateful. If 588 a function is stateless, when given the same input it will return the same 589 output and have no observable side effects. Optimizations such as common 590 subexpression elimination are only performed on stateless operations. 591 name: A name for the operation (optional). 592 593 Returns: 594 A list of `Tensor` or a single `Tensor` which `func` computes. 595 """ 596 if context.executing_eagerly(): 597 result = func(*[np.array(x) for x in inp]) 598 result = nest.flatten(result) 599 600 result = [x if x is None else ops.convert_to_tensor(x) for x in result] 601 if len(result) == 1: 602 # Mimic the automatic unwrapping in graph-mode py_func 603 result, = result 604 return result 605 606 if ops.executing_eagerly_outside_functions(): 607 with ops.device(context.context().host_address_space()): 608 return _internal_py_func( 609 func=func, 610 inp=inp, 611 Tout=Tout, 612 stateful=stateful, 613 eager=False, 614 name=name) 615 616 return _internal_py_func( 617 func=func, inp=inp, Tout=Tout, stateful=stateful, eager=False, name=name) 618 619 620@deprecation.deprecated( 621 date=None, 622 instructions="""tf.py_func is deprecated in TF V2. Instead, there are two 623 options available in V2. 624 - tf.py_function takes a python function which manipulates tf eager 625 tensors instead of numpy arrays. It's easy to convert a tf eager tensor to 626 an ndarray (just call tensor.numpy()) but having access to eager tensors 627 means `tf.py_function`s can use accelerators such as GPUs as well as 628 being differentiable using a gradient tape. 629 - tf.numpy_function maintains the semantics of the deprecated tf.py_func 630 (it is not differentiable, and manipulates numpy arrays). It drops the 631 stateful argument making all functions stateful. 632 """) 633@tf_export(v1=["py_func"]) 634@dispatch.add_dispatch_support 635def py_func(func, inp, Tout, stateful=True, name=None): 636 return py_func_common(func, inp, Tout, stateful, name=name) 637 638 639py_func.__doc__ = "%s" % py_func_common.__doc__ 640 641 642@tf_export("numpy_function") 643@dispatch.add_dispatch_support 644def numpy_function(func, inp, Tout, name=None): 645 """Wraps a python function and uses it as a TensorFlow op. 646 647 Given a python function `func` wrap this function as an operation in a 648 TensorFlow function. `func` must take numpy arrays as its arguments and 649 return numpy arrays as its outputs. 650 651 The following example creates a TensorFlow graph with `np.sinh()` as an 652 operation in the graph: 653 654 >>> def my_numpy_func(x): 655 ... # x will be a numpy array with the contents of the input to the 656 ... # tf.function 657 ... return np.sinh(x) 658 >>> @tf.function(input_signature=[tf.TensorSpec(None, tf.float32)]) 659 ... def tf_function(input): 660 ... y = tf.numpy_function(my_numpy_func, [input], tf.float32) 661 ... return y * y 662 >>> tf_function(tf.constant(1.)) 663 <tf.Tensor: shape=(), dtype=float32, numpy=1.3810978> 664 665 Comparison to `tf.py_function`: 666 `tf.py_function` and `tf.numpy_function` are very similar, except that 667 `tf.numpy_function` takes numpy arrays, and not `tf.Tensor`s. If you want the 668 function to contain `tf.Tensors`, and have any TensorFlow operations executed 669 in the function be differentiable, please use `tf.py_function`. 670 671 Note: The `tf.numpy_function` operation has the following known 672 limitations: 673 674 * The body of the function (i.e. `func`) will not be serialized in a 675 `tf.SavedModel`. Therefore, you should not use this function if you need to 676 serialize your model and restore it in a different environment. 677 678 * The operation must run in the same address space as the Python program 679 that calls `tf.numpy_function()`. If you are using distributed 680 TensorFlow, you must run a `tf.distribute.Server` in the same process as the 681 program that calls `tf.numpy_function` you must pin the created 682 operation to a device in that server (e.g. using `with tf.device():`). 683 684 * Since the function takes numpy arrays, you cannot take gradients 685 through a numpy_function. If you require something that is differentiable, 686 please consider using tf.py_function. 687 688 * The resulting function is assumed stateful and will never be optimized. 689 690 Args: 691 func: A Python function, which accepts `numpy.ndarray` objects as arguments 692 and returns a list of `numpy.ndarray` objects (or a single 693 `numpy.ndarray`). This function must accept as many arguments as there are 694 tensors in `inp`, and these argument types will match the corresponding 695 `tf.Tensor` objects in `inp`. The returns `numpy.ndarray`s must match the 696 number and types defined `Tout`. 697 Important Note: Input and output `numpy.ndarray`s of `func` are not 698 guaranteed to be copies. In some cases their underlying memory will be 699 shared with the corresponding TensorFlow tensors. In-place modification 700 or storing `func` input or return values in python datastructures 701 without explicit (np.)copy can have non-deterministic consequences. 702 inp: A list of `tf.Tensor` objects. 703 Tout: A list or tuple of tensorflow data types or a single tensorflow data 704 type if there is only one, indicating what `func` returns. 705 name: (Optional) A name for the operation. 706 707 Returns: 708 Single or list of `tf.Tensor` which `func` computes. 709 """ 710 return py_func_common(func, inp, Tout, stateful=True, name=name) 711 712 713ops.NotDifferentiable("PyFunc") 714ops.NotDifferentiable("PyFuncStateless") 715