在 Tensorflow 中实现自定义损失函数会导致“ValueError：必须在循环之前定义‘输出’。”

Question

我正在开发我的第一个 Python 机器学习项目 - 使用 TensorFlow 尝试使用 Moby Hyphenator II 数据集对单词进行音节化。

我将此视为多标签分类问题，其中单词及其音节按以下格式编码：

T e n - s o r - f l o w
0 0 1   0 0 1   0 0 0 0

在阅读 this guide 作为起点时，我看到作者使用了一个自定义函数——他们在 PyTorch 中用均方根误差对加权二元交叉熵进行平均：

def bce_rmse(pred, target, pos_weight = 1.3, epsilon = 1e-12):
    # Weighted binary cross entropy
    loss_pos = target * torch.log(pred + epsilon)
    loss_neg = (1 - target) * torch.log(1 - pred + epsilon)
    bce = torch.mean(torch.neg(pos_weight * loss_pos + loss_neg))

    # Root mean squared error
    mse = (torch.sum(pred, dim = 0) - torch.sum(target, dim = 0)) ** 2
    rmse = torch.mean(torch.sqrt(mse + epsilon))

    return (bce + rmse) / 2

我已经尝试通过以下方式在 TensorFlow 中实现这一点：

def weighted_bce_mse(y_true, y_prediction):
    # Binary crossentropy with weighting
    epsilon = 1e-12
    positive_weight = 4.108897148948174
    loss_positive = y_true * tf.math.log(y_prediction + epsilon)
    loss_negative = (1 - y_true) * tf.math.log(1 - y_prediction + epsilon)
    bce_loss = np.mean(tf.math.negative(positive_weight * loss_positive + loss_negative))
    
    # Mean squared error
    mse = tf.keras.losses.MeanSquaredError()
    mse_loss = mse(y_true, y_prediction)

    averaged_bce_mse = (bce_loss + mse_loss) / 2
    return averaged_bce_mse

这样做时，我收到错误 ValueError: 'outputs' must be defined before the loop.，我不确定为什么在构建和编译模型之前定义此函数。

我正在使用 Keras 功能 API，我的编译和适配阶段是：

model.compile(optimizer="adam", loss=weighted_bce_mse, metrics=["accuracy"], steps_per_execution=64)
history = model.fit(padded_inputs, padded_outputs, validation_data=(validation_inputs, validation_outputs), epochs=10, verbose=2)

Answer 1

如前所述，显示的错误与自定义损失函数无关。您显示的代码还有许多其他错误，例如未正确导入tf.keras.layers。修复这些错误后，请参阅下面的代码并在以下版本上进行测试（工作正常）：

tensorflow 2.4.1
numpy 1.19.5
python 3.9.6

import tensorflow as tf

# Custom loss function - mean of binary crossentropy and mean squared error
def mean_weighted_bce_mse(y_true, y_prediction):
    # Binary crossentropy with weighting
    epsilon = 1e-12
    positive_weight = 4.108897148948174
    loss_positive = y_true * tf.math.log(y_prediction + epsilon)
    loss_negative = (1 - y_true) * tf.math.log(1 - y_prediction + epsilon)
    bce_loss = np.mean(tf.math.negative(positive_weight * loss_positive + loss_negative))
    
    # Mean squared error
    mse = tf.keras.losses.MeanSquaredError()
    mse_loss = mse(y_true, y_prediction)
 
    averaged_bce_mse = (bce_loss + mse_loss) / 2
    return tf.math.reduce_mean(averaged_bce_mse, axis=-1)
 
 
inputs = tf.keras.Input(shape=(15,))
x = tf.keras.layers.Embedding(64, 64, mask_zero=True)(inputs)
 
x = tf.keras.layers.Bidirectional(tf.keras.layers.LSTM(128, return_sequences=True))(x)
x = tf.keras.layers.Dropout(0.2)(x)
x = tf.keras.layers.Bidirectional(tf.keras.layers.LSTM(128, return_sequences=True))(x)
x = tf.keras.layers.Dropout(0.2)(x)
x = tf.keras.layers.Bidirectional(tf.keras.layers.LSTM(128, return_sequences=True))(x)
x = tf.keras.layers.Dropout(0.3)(x)
 
 
x = tf.keras.layers.Conv1D(64, kernel_size=1)(x)
x = tf.keras.layers.Dropout(0.2)(x)
x = tf.keras.layers.Conv1D(64, kernel_size=1)(x)
x = tf.keras.layers.Dropout(0.2)(x)
x = tf.keras.layers.GlobalMaxPool1D()(x)
x = tf.keras.layers.Dropout(0.5)(x)
 
x = tf.keras.layers.Dense(32, activation="relu")(x)
x = tf.keras.layers.Dense(15, activation="sigmoid")(x)
 
 
model = tf.keras.models.Model(inputs=inputs, outputs=x)
model.compile(optimizer="adam", loss=mean_weighted_bce_mse, metrics=["accuracy"], steps_per_execution=64)
 
# history = model.fit(padded_inputs,
#                     padded_outputs,
#                     validation_data=(validation_inputs, validation_outputs),
#                     epochs=20,
#                     batch_size=8)

Answer 2

在下面的代码行中：

model.compile(optimizer="adam", loss=mean_weighted_bce_mse, metrics=["accuracy"], steps_per_execution=64)
history = model.fit(padded_inputs,
                padded_outputs,
                validation_data=(validation_inputs, validation_outputs),
                epochs=20,
                batch_size=8)

输入数据的长度是多少？ steps_per_execution 应该是len(input_data)/Batch_size。 删除steps_per_execution 并再次检查。

Answer 3

我发现错误源于我在自定义损失函数中使用的操作：

bce_loss = np.mean(tf.math.negative(positive_weight * loss_positive + loss_negative))

此行使用 np.mean 导致错误 - 将其替换为 tf.math.reduce_mean 以及通过 tf.cast 将 y_true 和 y_prediction 转换为 tf.float32 解决了问题：

# Custom loss function - mean of binary crossentropy and mean squared error
def mean_weighted_bce_mse(y_true, y_prediction):
    y_true = tf.cast(y_true, tf.float32)
    y_prediction = tf.cast(y_prediction, tf.float32)

    # Binary crossentropy with weighting
    epsilon = 1e-12
    positive_weight = 4.108897148948174
    loss_positive = y_true * tf.math.log(y_prediction + epsilon)
    loss_negative = (1 - y_true) * tf.math.log(1 - y_prediction + epsilon)
    bce_loss = tf.math.reduce_mean(tf.math.negative(positive_weight * loss_positive + loss_negative))
    
    # Mean squared error
    mse = tf.keras.losses.MeanSquaredError()
    mse_loss = mse(y_true, y_prediction)

    averaged_bce_mse = (bce_loss + mse_loss) / 2
    return averaged_bce_mse