python - 自定义损失函数 : logits and targets must have the same shape ((? , 1) vs (45000,))

Question

我的模型将二元分类器中的所有内容预测为 0。我们总共有 4000 个 true 和 41000 个 false。因此，我们正在尝试制作一个自定义损失函数。

我收到的错误是:

(logits.get_shape()，targets.get_shape()))

ValueError:logits 和目标必须具有相同的形状 ((?, 1) vs (45000,))

代码如下所示:

combined = tf.keras.layers.concatenate([modelRNN.output, modelCNN.output])

final_dense = tf.keras.layers.Dense(10, activation='relu')(combined) #ff kijken of dit slim is
final_dense = tf.keras.layers.Dense(1, activation='sigmoid')(final_dense)

final_model = tf.keras.Model(inputs=[modelCNN.input, modelRNN.input], outputs=final_dense)

targets = match_train
logits = final_dense
pos_weight = (45000 - 4539) / 4539


custom_loss = tf.nn.weighted_cross_entropy_with_logits(
    targets,
    logits,
    pos_weight,
    )


final_model.compile(optimizer='adam',
                    loss=custom_loss,
                    metrics=['accuracy'])

初始数组的形状是:

modelCNN = (45000, 28, 28, 1) float64
modelRNN = (45000, 93, 13) float64
labels = (45000,1) boolean

通过注释中的代码部分解决了问题。我现在收到一个以前没有过的错误。它说:

TypeError: Using a `tf.Tensor` as a Python `bool` is not allowed. Use `if t is not None:` instead of `if t:` to test if a tensor is defined, and use TensorFlow ops such as tf.cond to execute subgraphs conditioned on the value of a tensor.

  File "<ipython-input-6-42327e5a4b50>", line 3, in <module>
    metrics=['accuracy'])

  File "C:\Users\Tijev\Anaconda3\envs\tfp3.6\lib\site-packages\tensorflow\python\training\checkpointable\base.py", line 442, in _method_wrapper
    method(self, *args, **kwargs)

  File "C:\Users\Tijev\Anaconda3\envs\tfp3.6\lib\site-packages\tensorflow\python\keras\engine\training.py", line 215, in compile
    loss = loss or {}

Answer 1

将标签 reshape 为 2D 张量。

targets = np.asarray(match_train).astype('float32').reshape((-1,1))

来源:Tensorflow estimator ValueError: logits and labels must have the same shape ((?, 1) vs (?,))