从检查点加载权重在 keras 模型中不起作用

Question

我快要疯了。

我使用 tensorflow keras 定义了一个顺序模型：

model = tf.keras.Sequential([tf.keras.layer.Dense(128,input_shape(784,),activation="relu"),
                             tf.keras.layer.Dense(10,activation="softmax"])
model.compile(optimizer="adam",loss="mse")
keras.experimental.export_saved_model(model,"keras_model")

我使用 c_api.h在C 程序中训练所述模型

C 程序将权重保存在检查点文件中。

尝试从检查点文件中恢复 python 中的权重时：

keras.experimental.load_from_saved_model("keras_model/")
#OR
model = tf.keras.Sequential([tf.keras.layer.Dense(128,input_shape(784,),activation="relu"),
                             tf.keras.layer.Dense(10,activation="softmax"])
model.load_weights("keras_model/variables/variables")
#OR
checkpoint = tf.train.Checkpoint(model=model)
status = checkpoint.restore("keras_model/variables/variables")

我最终得到一个错误并且没有恢复权重。

我能够恢复重量并继续在我的 C 程序中训练

keras.experimental.load_from_saved_model("keras_model/")
WARNING: Logging before flag parsing goes to stderr.
W0918 15:18:04.350199 140418474760000 deprecation.py:323] From <ipython-input-2-06ea110fdc8e>:1: load_from_saved_model (from tensorflow.python.keras.saving.saved_model_experimental) is deprecated and will be removed in a future version.
Instructions for updating:
The experimental save and load functions have been  deprecated. Please switch to `tf.keras.models.load_model`.
2019-09-18 15:18:04.390271: I tensorflow/core/platform/profile_utils/cpu_utils.cc:94] CPU Frequency: 1696040000 Hz
2019-09-18 15:18:04.390913: I tensorflow/compiler/xla/service/service.cc:168] XLA service 0x4bf4790 executing computations on platform Host. Devices:
2019-09-18 15:18:04.390961: I tensorflow/compiler/xla/service/service.cc:175]   StreamExecutor device (0): Host, Default Version
W0918 15:18:04.436281 140418474760000 deprecation.py:323] From /home/jregalado/Projects/tensorflow/env/lib/python3.6/site-packages/tensorflow_core/python/training/tracking/util.py:1249: NameBasedSaverStatus.__init__ (from tensorflow.python.training.tracking.util) is deprecated and will be removed in a future version.
Instructions for updating:
Restoring a name-based tf.train.Saver checkpoint using the object-based restore API. This mode uses global names to match variables, and so is somewhat fragile. It also adds new restore ops to the graph each time it is called when graph building. Prefer re-encoding training checkpoints in the object-based format: run save() on the object-based saver (the same one this message is coming from) and use that checkpoint in the future.
---------------------------------------------------------------------------
AssertionError                            Traceback (most recent call last)
<ipython-input-2-06ea110fdc8e> in <module>
----> 1 keras.experimental.load_from_saved_model("keras_model/")

~/Projects/tensorflow/env/lib/python3.6/site-packages/tensorflow_core/python/util/deprecation.py in new_func(*args, **kwargs)
322               'in a future version' if date is None else ('after %s' % date),
323               instructions)
--> 324       return func(*args, **kwargs)
325     return tf_decorator.make_decorator(
326         func, new_func, 'deprecated',

~/Projects/tensorflow/env/lib/python3.6/site-packages/tensorflow_core/python/keras/saving/saved_model_experimental.py in load_from_saved_model(saved_model_path, custom_objects)
425       compat.as_text(constants.VARIABLES_DIRECTORY),
426       compat.as_text(constants.VARIABLES_FILENAME))
--> 427   model.load_weights(checkpoint_prefix)
428   return model

~/Projects/tensorflow/env/lib/python3.6/site-packages/tensorflow_core/python/keras/engine/training.py in load_weights(self, filepath, by_name)
179         raise ValueError('Load weights is not yet supported with TPUStrategy '
180                          'with steps_per_run greater than 1.')
--> 181     return super(Model, self).load_weights(filepath, by_name)
182
183   @trackable.no_automatic_dependency_tracking

~/Projects/tensorflow/env/lib/python3.6/site-packages/tensorflow_core/python/keras/engine/network.py in load_weights(self, filepath, by_name)
1372         # streaming restore for any variables created in the future.
1373         trackable_utils.streaming_restore(status=status, session=session)
-> 1374       status.assert_nontrivial_match()
1375       return status
1376     if h5py is None:

~/Projects/tensorflow/env/lib/python3.6/site-packages/tensorflow_core/python/training/tracking/util.py in assert_nontrivial_match(self)
964     # assert_nontrivial_match and assert_consumed (and both are less
965     # useful since we don't touch Python objects or Python state).
--> 966     return self.assert_consumed()
967
968   def _gather_saveable_objects(self):

~/Projects/tensorflow/env/lib/python3.6/site-packages/tensorflow_core/python/training/tracking/util.py in assert_consumed(self)
941       raise AssertionError(
942           "Some objects had attributes which were not restored:{}".format(
--> 943               "".join(unused_attribute_strings)))
944     for trackable in self._graph_view.list_objects():
945       # pylint: disable=protected-access

AssertionError: Some objects had attributes which were not restored:
<tf.Variable 'a/kernel:0' shape=(784, 128) dtype=float32, numpy=
array([[-0.03716458, -0.04911711, -0.01023878, ...,  0.0636776 ,
0.02892563, -0.05542086],
[-0.02324755, -0.07362694, -0.0399951 , ...,  0.0680329 ,
0.05201877, -0.05149256],
[ 0.00954343,  0.05673491,  0.05108347, ...,  0.01994208,
-0.01107961,  0.06192174],
...,
[ 0.07091486, -0.07734856, -0.04417738, ...,  0.01921409,
-0.01908814, -0.05070668],
[ 0.01353646, -0.05189713, -0.01391671, ..., -0.05795977,
0.04801518,  0.00801209],
[-0.05304915,  0.01870193,  0.05657425, ..., -0.06819408,
-0.00760372, -0.0106293 ]], dtype=float32)>: ['a/kernel']
<tf.Variable 'a/bias:0' shape=(128,) dtype=float32, numpy=
array([0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0.], dtype=float32)>: ['a/bias']
<tf.Variable 'b/kernel:0' shape=(128, 10) dtype=float32, numpy=
array([[-0.1759212 , -0.09282549, -0.11045764, ..., -0.13727605,
-0.02849793,  0.14510198],
[ 0.06857841, -0.01459177,  0.08369003, ...,  0.05089156,
-0.05319159, -0.08594933],
[-0.180914  , -0.18932283,  0.20551099, ..., -0.17210156,
-0.10069884,  0.06433241],
...,
[ 0.09097584, -0.03930017, -0.15125516, ...,  0.02359283,
-0.16158347, -0.13176063],
[-0.04145582, -0.03205152,  0.20097663, ..., -0.15124482,
0.16874255, -0.15434337],
[-0.13188484,  0.04145408,  0.05036192, ..., -0.10489662,
0.12316228,  0.08794598]], dtype=float32)>: ['b/kernel']
<tf.Variable 'b/bias:0' shape=(10,) dtype=float32, numpy=array([0., 0., 0., 0., 0., 0., 0., 0., 0., 0.], dtype=float32)>: ['b/bias']

Answer 1

虽然当我遇到这个问题时，我只使用 Python 而不是 C/C++，实际问题是我将文件而不是主干传递给.load_weights()函数：.index

错误的：

model = make_some_model()
model.load_weights("output/20220801-pretrain_test/checkpoints/checkpoint_weights_e10.ckpt.index")

正确的：

model = make_some_model()
model.load_weights("output/20220801-pretrain_test/checkpoints/checkpoint_weights_e10.ckpt")