为什么使用Pearsonr作为损失函数时，为什么第一个时期的损失大于张量

Question

我想在DNN网络中使用Pearsonr作为损失功能，我遵循Scipy.stats.pearsonr来编写自己的损失功能，并且结果与Scipy相同。这是我的功能，结果被

def correlationLoss2(x, y):
    x = tf.convert_to_tensor(x)
    y = math_ops.cast(y, x.dtype)
    xmean = tf.reduce_mean(x)
    ymean = tf.reduce_mean(y)
    xsqsum = tf.reduce_sum( tf.math.squared_difference(x, xmean))
    ysqsum = tf.reduce_sum( tf.math.squared_difference(y, ymean))
    cov = tf.reduce_sum( (x - xmean) * (y - ymean))
    corr = cov / tf.sqrt(xsqsum * ysqsum)
    return - corr

x = data[:, 22]
y = data[:, -1]

print(- float(correlationLoss(x, y)))
print(stats.pearsonr(x, y)[0])

忽略了x和y，它们是我项目的数据集，结果是：

-0.017431042527298025
-0.017431042527298084

事情似乎还不错，但是当涉及到我的网络时，第一个时代说（我的指标func是（1- corr））：

Epoch 1/50
196/196 [==============================] - 35s 172ms/step - loss: 1.0501 - correlationMetric: 0.8766 - val_loss: -0.0295 - val_correlationMetric: 0.9123
Epoch 2/50
196/196 [==============================] - 11s 55ms/step - loss: -0.1150 - correlationMetric: 0.8465 - val_loss: -0.0598 - val_correlationMetric: 0.9115
Epoch 3/50
196/196 [==============================] - 11s 56ms/step - loss: -0.1299 - correlationMetric: 0.8352 - val_loss: -0.0674 - val_correlationMetric: 0.9061
Epoch 4/50
196/196 [==============================] - 10s 53ms/step - loss: -0.1383 - correlationMetric: 0.8270 - val_loss: -0.0734 - val_correlationMetric: 0.9013

网络的结果很棒，但是我仍然想知道为什么第一个时期的损失大于一个时代，这是没有意义的。这是我的网络：

def get_model():
    features_inputs = tf.keras.Input((Config.input_size, ), dtype=tf.float32)

    gn = tf.keras.layers.GaussianNoise(0.035)(features_inputs)
    feature_x = layers.Dense(256, activation='swish')(gn)
    feature_x = layers.Dropout(0.5)(feature_x)

    feature_x = layers.Dense(512, activation='swish', kernel_regularizer="l2")(feature_x)
    feature_x = layers.Dropout(0.44)(feature_x)
    feature_x = layers.Dense(128, activation='swish', kernel_regularizer="l2")(feature_x)
    feature_x = layers.Dropout(0.33)(feature_x)
    feature_x = layers.Dense(32, activation='swish', kernel_regularizer="l2")(feature_x)
    feature_x = layers.Dropout(0.3)(feature_x)
    output = layers.Dense(1)(feature_x)

    model = tf.keras.Model(inputs=[features_inputs], outputs=[output])
    model.compile(optimizer=tf.optimizers.Adam(0.001), loss=correlationLoss2,
                  metrics=[correlationMetric])

    return model