7

tf.math.square 行または @tf.function 装飾を省略した場合、次のコードが実行されます。

四角い線がグラデーションの問題を引き起こしているのはなぜですか?

import tensorflow as tf
import tensorflow.keras as keras
import tensorflow.keras.backend as K
import numpy as np

m = 1000
n = 1
X = np.random.randn(m, n).astype(np.float32)
y = (3 + 0 * np.random.randn(m)).astype(np.float32)

def create_model():
    a_input = keras.layers.Input(shape=(n,), dtype=np.float32)
    a = K.expand_dims(a_input, axis=2)
    q = keras.layers.Conv1D(1, 1)(a)
    q = - tf.math.square(q) # this breaks things, but only when using tf.function
    model = keras.models.Model(inputs=a_input, outputs=q)
    return model

model = create_model()
model.predict(X)

class Trainer():
    def __init__(self, epochs=10):
        self.epochs = epochs
        self.model = create_model()
        self.optimizer = tf.optimizers.Adam()
        self.step = 0
    def train(self, X, y, epochs=10):
        X = tf.convert_to_tensor(X, dtype=tf.float32)
        y = tf.convert_to_tensor(y, dtype=tf.float32)
        for epoch in range(epochs):
            l = self._train_one_step(X, y)
        return l
    @tf.function
    def _train_one_step(self, X, y):
        with tf.GradientTape() as tape:
            yp = self.model(X)
            loss = tf.reduce_mean(tf.math.square(y - yp))
        gradients = tape.gradient(loss, self.model.trainable_variables)
        l = self.optimizer.apply_gradients(zip(gradients, self.model.trainable_variables))
        d = dict(loss=loss)
        tf.print(yp[0], loss)
        self.step += 1

trainer = Trainer()
l = trainer.train(X, y, epochs=100)

要求に応じてトレースバック:

2019-04-18 14:23:03.776224: I tensorflow/core/platform/cpu_feature_guard.cc:142] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA
---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<ipython-input-1-8a9114e8a51b> in <module>
----> 1 import example_keras_tf2 as e

~/toplevelrepo/something-train/something/train/usr/HOME/example_keras_tf2.py in <module>
     44
     45 trainer = Trainer()
---> 46 l = trainer.train(X, y, epochs=100)

~/toplevelrepo/something-train/something/train/usr/HOME/example_keras_tf2.py in train(self, X, y, epochs)
     30         y = tf.convert_to_tensor(y, dtype=tf.float32)
     31         for epoch in range(epochs):
---> 32             l = self._train_one_step(X, y)
     33         return l
     34     @tf.function

~/anaconda3/envs/myenv/lib/python3.7/site-packages/tensorflow/python/eager/def_function.py in __call__(self, *args, **kwds)
    424     # This is the first call of __call__, so we have to initialize.
    425     initializer_map = {}
--> 426     self._initialize(args, kwds, add_initializers_to=initializer_map)
    427     if self._created_variables:
    428       try:

~/anaconda3/envs/myenv/lib/python3.7/site-packages/tensorflow/python/eager/def_function.py in _initialize(self, args, kwds, add_initializers_to)
    368     self._concrete_stateful_fn = (
    369         self._stateful_fn._get_concrete_function_internal_garbage_collected(  # pylint: disable=protected-access
--> 370             *args, **kwds))
    371
    372     def invalid_creator_scope(*unused_args, **unused_kwds):

~/anaconda3/envs/myenv/lib/python3.7/site-packages/tensorflow/python/eager/function.py in _get_concrete_function_internal_garbage_collected(self, *args, **kwargs)
   1311     if self._input_signature:
   1312       args, kwargs = None, None
-> 1313     graph_function, _, _ = self._maybe_define_function(args, kwargs)
   1314     return graph_function
   1315

~/anaconda3/envs/myenv/lib/python3.7/site-packages/tensorflow/python/eager/function.py in _maybe_define_function(self, args, kwargs)
   1578           or call_context_key not in self._function_cache.missed):
   1579         self._function_cache.missed.add(call_context_key)
-> 1580         graph_function = self._create_graph_function(args, kwargs)
   1581         self._function_cache.primary[cache_key] = graph_function
   1582         return graph_function, args, kwargs

~/anaconda3/envs/myenv/lib/python3.7/site-packages/tensorflow/python/eager/function.py in _create_graph_function(self, args, kwargs, override_flat_arg_shapes)
   1510             arg_names=arg_names,
   1511             override_flat_arg_shapes=override_flat_arg_shapes,
-> 1512             capture_by_value=self._capture_by_value),
   1513         self._function_attributes)
   1514

~/anaconda3/envs/myenv/lib/python3.7/site-packages/tensorflow/python/framework/func_graph.py in func_graph_from_py_func(name, python_func, args, kwargs, signature, func_graph, autograph, autograph_options, add_control_dependencies, arg_names, op_return_value, collections, capture_by_value, override_flat_arg_shapes)
    692                                           converted_func)
    693
--> 694       func_outputs = python_func(*func_args, **func_kwargs)
    695
    696       # invariant: `func_outputs` contains only Tensors, IndexedSlices,

~/anaconda3/envs/myenv/lib/python3.7/site-packages/tensorflow/python/eager/def_function.py in wrapped_fn(*args, **kwds)
    315         # __wrapped__ allows AutoGraph to swap in a converted function. We give
    316         # the function a weak reference to itself to avoid a reference cycle.
--> 317         return weak_wrapped_fn().__wrapped__(*args, **kwds)
    318     weak_wrapped_fn = weakref.ref(wrapped_fn)
    319

~/anaconda3/envs/myenv/lib/python3.7/site-packages/tensorflow/python/eager/function.py in bound_method_wrapper(*args, **kwargs)
   2106     # If __wrapped__ was replaced, then it is always an unbound function
   2107     # that takes self as first argument.
-> 2108     return wrapped_fn(weak_instance(), *args, **kwargs)
   2109   weak_bound_method_wrapper = weakref.ref(bound_method_wrapper)
   2110

~/anaconda3/envs/myenv/lib/python3.7/site-packages/tensorflow/python/framework/func_graph.py in wrapper(*args, **kwargs)
    684                   optional_features=autograph_options,
    685                   force_conversion=True,
--> 686               ), args, kwargs)
    687
    688         # Wrapping around a decorator allows checks like tf_inspect.getargspec

~/anaconda3/envs/myenv/lib/python3.7/site-packages/tensorflow/python/autograph/impl/api.py in converted_call(f, owner, options, args, kwargs)
    390     return _call_unconverted(f, args, kwargs)
    391
--> 392   result = converted_f(*effective_args, **kwargs)
    393
    394   # The converted function's closure is simply inserted into the function's

/var/folders/8s/kcb8_98n1pqbnwrv2h0b14rm0000gp/T/tmp9aj7e32f.py in tf___train_one_step(self, X, y)
      5     loss = ag__.converted_call('reduce_mean', tf, ag__.ConversionOptions(recursive=True, verbose=0, strip_decorators=(tf.function, defun_1, ag__.convert, ag__.do_not_convert, ag__.converted_call), force_conversion=False, optional_features=(), internal_convert_user_code=True), (tf.math.square(y - yp),), {})
      6   gradients = ag__.converted_call('gradient', tape, ag__.ConversionOptions(recursive=True, verbose=0, strip_decorators=(tf.function, defun_2, ag__.convert, ag__.do_not_convert, ag__.converted_call), force_conversion=False, optional_features=(), internal_convert_user_code=True), (loss, self.model.trainable_variables), {})
----> 7   l = ag__.converted_call('apply_gradients', self.optimizer, ag__.ConversionOptions(recursive=True, verbose=0, strip_decorators=(tf.function, defun_3, ag__.convert, ag__.do_not_convert, ag__.converted_call), force_conversion=False, optional_features=(), internal_convert_user_code=True), (zip(gradients, self.model.trainable_variables),), {})
      8   d = ag__.converted_call(dict, None, ag__.ConversionOptions(recursive=True, verbose=0, strip_decorators=(tf.function, defun_4, ag__.convert, ag__.do_not_convert, ag__.converted_call), force_conversion=False, optional_features=(), internal_convert_user_code=True), (), {'loss': loss})
      9   ag__.converted_call('print', tf, ag__.ConversionOptions(recursive=True, verbose=0, strip_decorators=(tf.function, defun_5, ag__.convert, ag__.do_not_convert, ag__.converted_call), force_conversion=False, optional_features=(), internal_convert_user_code=True), (yp[0], loss), {})

~/anaconda3/envs/myenv/lib/python3.7/site-packages/tensorflow/python/autograph/impl/api.py in converted_call(f, owner, options, args, kwargs)
    265
    266   if not options.force_conversion and conversion.is_whitelisted_for_graph(f):
--> 267     return _call_unconverted(f, args, kwargs)
    268
    269   # internal_convert_user_code is for example turned off when issuing a dynamic

~/anaconda3/envs/myenv/lib/python3.7/site-packages/tensorflow/python/autograph/impl/api.py in _call_unconverted(f, args, kwargs)
    186     return f.__self__.call(args, kwargs)
    187
--> 188   return f(*args, **kwargs)
    189
    190

~/anaconda3/envs/myenv/lib/python3.7/site-packages/tensorflow/python/keras/optimizer_v2/optimizer_v2.py in apply_gradients(self, grads_and_vars, name)
    394       ValueError: If none of the variables have gradients.
    395     """
--> 396     grads_and_vars = _filter_grads(grads_and_vars)
    397     var_list = [v for (_, v) in grads_and_vars]
    398

~/anaconda3/envs/myenv/lib/python3.7/site-packages/tensorflow/python/keras/optimizer_v2/optimizer_v2.py in _filter_grads(grads_and_vars)
    922   if not filtered:
    923     raise ValueError("No gradients provided for any variable: %s." %
--> 924                      ([v.name for _, v in grads_and_vars],))
    925   if vars_with_empty_grads:
    926     logging.warning(

ValueError: No gradients provided for any variable: ['conv1d_1/kernel:0', 'conv1d_1/bias:0'].

また、まだ読み方がわからないので、これが役立つかどうかはわかりませんが、python_function ダンプは次のとおりです。

In [6]: print(tf.autograph.to_code(e.trainer._train_one_step.python_function))
from __future__ import print_function

def tf___train_one_step(*args, **kwargs):
  try:
    """Wraps either a dummy MethodType or a converted AutoGraph function."""
    with ag__.function_scope('bound_method_wrapper'):
      do_return = False
      retval_ = None
      strong_bound_method_wrapper = ag__.converted_call(weak_bound_method_wrapper, None, ag__.ConversionOptions(recursive=True, verbose=0, strip_decorators=(ag__.convert, ag__.do_not_convert, ag__.converted_call), force_conversion=False, optional_features=ag__.Feature.ALL, internal_convert_user_code=True), (), {})
      wrapped_fn = strong_bound_method_wrapper.__wrapped__
      cond_1 = ag__.is_(wrapped_fn, strong_bound_method_wrapper.__original_wrapped__)

      def if_true_1():
        with ag__.function_scope('if_true_1'):
          wrapped_fn = original_function.python_function
          cond = ag__.converted_call('ismethod', tf_inspect, ag__.ConversionOptions(recursive=True, verbose=0, strip_decorators=(ag__.convert, ag__.do_not_convert, ag__.converted_call), force_conversion=False, optional_features=ag__.Feature.ALL, internal_convert_user_code=True), (wrapped_fn,), {})

          def if_true():
            with ag__.function_scope('if_true'):
              wrapped_fn_1, = wrapped_fn,
              wrapped_fn_1 = ag__.converted_call('get_unbound_function', six, ag__.ConversionOptions(recursive=True, verbose=0, strip_decorators=(ag__.convert, ag__.do_not_convert, ag__.converted_call), force_conversion=False, optional_features=ag__.Feature.ALL, internal_convert_user_code=True), (wrapped_fn_1,), {})
              return wrapped_fn_1

          def if_false():
            with ag__.function_scope('if_false'):
              return wrapped_fn
          wrapped_fn = ag__.if_stmt(cond, if_true, if_false)
          do_return = True
          retval_ = ag__.converted_call(wrapped_fn, None, ag__.ConversionOptions(recursive=True, verbose=0, strip_decorators=(ag__.convert, ag__.do_not_convert, ag__.converted_call), force_conversion=False, optional_features=ag__.Feature.ALL, internal_convert_user_code=True), (weak_instance(),) + tuple(args), dict(kwargs, **{}))
          return retval_

      def if_false_1():
        with ag__.function_scope('if_false_1'):
          do_return = True
          retval_ = ag__.converted_call(wrapped_fn, None, ag__.ConversionOptions(recursive=True, verbose=0, strip_decorators=(ag__.convert, ag__.do_not_convert, ag__.converted_call), force_conversion=False, optional_features=ag__.Feature.ALL, internal_convert_user_code=True), (weak_instance(),) + tuple(args), dict(kwargs, **{}))
          return retval_
      retval_ = ag__.if_stmt(cond_1, if_true_1, if_false_1)
      return retval_
  except:
    ag__.rewrite_graph_construction_error(ag_source_map__)



tf___train_one_step.autograph_info__ = {}
4

0 に答える 0