Tensorflow.Keras：自定义约束不起作用

我正在尝试实现权重正交约束所示here https://towardsdatascience.com/build-the-right-autoencoder-tune-and-optimize-using-pca-principles-part-ii-24b9cca69bd6，在第 2.0 节中。当我尝试在 Keras 密集层上使用它时，会出现值错误。

当尝试实现同一篇文章第 3.0 部分中的自定义不相关特征约束时，也会发生这种情况。

import tensorflow as tf
import numpy as np

class WeightsOrthogonalityConstraint(tf.keras.constraints.Constraint):
    def __init__(self, encoding_dim, weightage = 1.0, axis = 0):
        self.encoding_dim = encoding_dim
        self.weightage = weightage
        self.axis = axis

    def weights_orthogonality(self, w):
        if(self.axis==1):
            w = tf.keras.backend.transpose(w)
        if(self.encoding_dim > 1):
            m = tf.keras.backend.dot(tf.keras.backend.transpose(w), w) - tf.keras.backend.eye(self.encoding_dim)
            return self.weightage * tf.keras.backend.sqrt(tf.keras.backend.sum(tf.keras.backend.square(m)))
        else:
            m = tf.keras.backend.sum(w ** 2) - 1.
            return m

    def __call__(self, w):
        return self.weights_orthogonality(w)

rand_samples = np.random.rand(16, 4)
dummy_ds = tf.data.Dataset.from_tensor_slices((rand_samples, rand_samples)).shuffle(16).batch(16)

encoder = tf.keras.layers.Dense(2, "relu", input_shape=(4,), kernel_regularizer=WeightsOrthogonalityConstraint(2))
decoder = tf.keras.layers.Dense(4, "relu")

autoencoder = tf.keras.models.Sequential()
autoencoder.add(encoder)
autoencoder.add(decoder)

autoencoder.compile(metrics=['accuracy'],
                    loss='mean_squared_error',
                    optimizer='sgd')

autoencoder.summary()

autoencoder.fit(dummy_ds, epochs=1)

如果我停止使用约束，则不会出现错误，但是在使用时，会引发下一个错误：

2019-09-07 14:20:25.962610: I tensorflow/stream_executor/platform/default/dso_loader.cc:42] Successfully opened dynamic library nvcuda.dll
2019-09-07 14:20:26.997957: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1640] Found device 0 with properties: 
name: GeForce GTX 1060 major: 6 minor: 1 memoryClockRate(GHz): 1.733
pciBusID: 0000:01:00.0
2019-09-07 14:20:27.043016: I tensorflow/stream_executor/platform/default/dlopen_checker_stub.cc:25] GPU libraries are statically linked, skip dlopen check.        
2019-09-07 14:20:27.050749: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1763] Adding visible gpu devices: 0
2019-09-07 14:20:27.081369: I tensorflow/core/platform/cpu_feature_guard.cc:142] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2
2019-09-07 14:20:27.113598: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1640] Found device 0 with properties: 
name: GeForce GTX 1060 major: 6 minor: 1 memoryClockRate(GHz): 1.733
pciBusID: 0000:01:00.0
2019-09-07 14:20:27.144194: I tensorflow/stream_executor/platform/default/dlopen_checker_stub.cc:25] GPU libraries are statically linked, skip dlopen check.        
2019-09-07 14:20:27.151802: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1763] Adding visible gpu devices: 0
2019-09-07 14:20:27.800616: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1181] Device interconnect StreamExecutor with strength 1 edge matrix:
2019-09-07 14:20:27.817323: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1187]      0 
2019-09-07 14:20:27.840635: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1200] 0:   N 
2019-09-07 14:20:27.848536: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1326] Created TensorFlow device (/job:localhost/replica:0/task:0/device:GPU:0 with 4712 MB memory) -> physical GPU (device: 0, name: GeForce GTX 1060, pci bus id: 0000:01:00.0, compute capability: 6.1)
Traceback (most recent call last):
  File "c:\Users\whitm\.vscode\extensions\ms-python.python-2019.9.34911\pythonFiles\ptvsd_launcher.py", line 43, in <module>
    main(ptvsdArgs)
  File "c:\Users\whitm\.vscode\extensions\ms-python.python-2019.9.34911\pythonFiles\lib\python\ptvsd\__main__.py", line 432, in main
    run()
  File "c:\Users\whitm\.vscode\extensions\ms-python.python-2019.9.34911\pythonFiles\lib\python\ptvsd\__main__.py", line 316, in run_file
    runpy.run_path(target, run_name='__main__')
  File "C:\ProgramData\Anaconda3\envs\tensorflow-gpu\lib\runpy.py", line 263, in run_path
    pkg_name=pkg_name, script_name=fname)
  File "C:\ProgramData\Anaconda3\envs\tensorflow-gpu\lib\runpy.py", line 96, in _run_module_code
    mod_name, mod_spec, pkg_name, script_name)
  File "C:\ProgramData\Anaconda3\envs\tensorflow-gpu\lib\runpy.py", line 85, in _run_code
    exec(code, run_globals)
  File "c:\Users\whitm\Desktop\CodeProjects\ForestClassifier-DEC\Test.py", line 35, in <module>
    optimizer='sgd')
  File "C:\ProgramData\Anaconda3\envs\tensorflow-gpu\lib\site-packages\tensorflow\python\training\tracking\base.py", line 458, in _method_wrapper
    result = method(self, *args, **kwargs)
  File "C:\ProgramData\Anaconda3\envs\tensorflow-gpu\lib\site-packages\tensorflow\python\keras\engine\training.py", line 337, in compile
    self._compile_weights_loss_and_weighted_metrics()
  File "C:\ProgramData\Anaconda3\envs\tensorflow-gpu\lib\site-packages\tensorflow\python\training\tracking\base.py", line 458, in _method_wrapper
    result = method(self, *args, **kwargs)
  File "C:\ProgramData\Anaconda3\envs\tensorflow-gpu\lib\site-packages\tensorflow\python\keras\engine\training.py", line 1494, in _compile_weights_loss_and_weighted_metrics
    self.total_loss = self._prepare_total_loss(masks)
  File "C:\ProgramData\Anaconda3\envs\tensorflow-gpu\lib\site-packages\tensorflow\python\keras\engine\training.py", line 1601, in _prepare_total_loss
    custom_losses = self.get_losses_for(None) + self.get_losses_for(
  File "C:\ProgramData\Anaconda3\envs\tensorflow-gpu\lib\site-packages\tensorflow\python\keras\engine\base_layer.py", line 1209, in get_losses_for
    return [l for l in self.losses if l._unconditional_loss]
  File "C:\ProgramData\Anaconda3\envs\tensorflow-gpu\lib\site-packages\tensorflow\python\keras\engine\base_layer.py", line 835, in losses
    return collected_losses + self._gather_children_attribute('losses')
  File "C:\ProgramData\Anaconda3\envs\tensorflow-gpu\lib\site-packages\tensorflow\python\keras\engine\base_layer.py", line 2129, in _gather_children_attribute      
    getattr(layer, attribute) for layer in nested_layers))
  File "C:\ProgramData\Anaconda3\envs\tensorflow-gpu\lib\site-packages\tensorflow\python\keras\engine\base_layer.py", line 2129, in <genexpr>
    getattr(layer, attribute) for layer in nested_layers))
  File "C:\ProgramData\Anaconda3\envs\tensorflow-gpu\lib\site-packages\tensorflow\python\keras\engine\base_layer.py", line 832, in losses
    loss_tensor = regularizer()
  File "C:\ProgramData\Anaconda3\envs\tensorflow-gpu\lib\site-packages\tensorflow\python\keras\engine\base_layer.py", line 907, in _tag_unconditional
    loss = loss()
  File "C:\ProgramData\Anaconda3\envs\tensorflow-gpu\lib\site-packages\tensorflow\python\keras\engine\base_layer.py", line 1659, in _loss_for_variable
    regularization = regularizer(v)
  File "c:\Users\whitm\Desktop\CodeProjects\ForestClassifier-DEC\Test.py", line 21, in __call__
    return self.weights_orthogonality(w)
  File "c:\Users\whitm\Desktop\CodeProjects\ForestClassifier-DEC\Test.py", line 14, in weights_orthogonality
    m = tf.keras.backend.dot(tf.keras.backend.transpose(w), w) - tf.keras.backend.eye(self.encoding_dim)
  File "C:\ProgramData\Anaconda3\envs\tensorflow-gpu\lib\site-packages\tensorflow\python\keras\backend.py", line 1310, in eye
    return variable(linalg_ops.eye(size, dtype=tf_dtype), dtype, name)
  File "C:\ProgramData\Anaconda3\envs\tensorflow-gpu\lib\site-packages\tensorflow\python\keras\backend.py", line 785, in variable
    constraint=constraint)
  File "C:\ProgramData\Anaconda3\envs\tensorflow-gpu\lib\site-packages\tensorflow\python\ops\variables.py", line 264, in __call__
    return super(VariableMetaclass, cls).__call__(*args, **kwargs)
  File "C:\ProgramData\Anaconda3\envs\tensorflow-gpu\lib\site-packages\tensorflow\python\ops\resource_variable_ops.py", line 464, in __init__
    shape=shape)
  File "C:\ProgramData\Anaconda3\envs\tensorflow-gpu\lib\site-packages\tensorflow\python\ops\resource_variable_ops.py", line 550, in _init_from_args
    raise ValueError("Tensor-typed variable initializers must either be "
ValueError: Tensor-typed variable initializers must either be wrapped in an init_scope or callable (e.g., `tf.Variable(lambda : tf.truncated_normal([10, 40]))`) when building functions. Please file a feature request if this restriction inconveniences you.

提前致谢！

PD: Here https://colab.research.google.com/drive/19jTi5jRaDKFey0QZ1FQgOUqiz3UqS3xm是一个显示错误的 Colab Notebook

PD2：我设法找到导致问题的线路，是这个：

m = tf.keras.backend.dot(tf.keras.backend.transpose(w), w) - tf.keras.backend.eye(self.encoding_dim)

特别是 keras 后端 eye() 函数导致了问题

我设法解决这个问题：

导致错误的函数是第 14 行的 tf.keras.backed.eye()。我在那里读到，该函数的 keras 后端中的实现使用 numpy 数组作为单位矩阵，但张量流和其他后端已经有了它们的实现使用张量来实现此函数。 tf2.0 上缺少张量导致错误，只需将 tf.keras.backed.eye() 更改为 tf.eye() 即可解决问题。