自定义损失函数返回 - InvalidArgumentError：第二个输入必须是标量，但它具有形状 [64]

Posted 2023-02-16

【中文标题】自定义损失函数返回 - InvalidArgumentError：第二个输入必须是标量，但它具有形状 [64]

【英文标题】：Custom Loss Function returning - InvalidArgumentError: The second input must be a scalar, but it has shape [64]

【发布时间】：2021-12-16 22:45:30

【问题描述】：

我正在尝试使用 this custom loss 的修改版本，但出现以下错误

InvalidArgumentError:  The second input must be a scalar, but it has shape [64] [[node gradient_tape/custom_loss/cond_1/StatelessIf/gradient_tape/custom_loss/weighted_loss/Mul/_30]] [Op:__inference_train_function_147002]
    
    Function call stack:
    train_function

这是代码

import time
import numpy as np
import tensorflow as tf
from tensorflow.keras.losses import Loss
from tensorflow.keras.models import Sequential, load_model
from tensorflow.keras.layers import Dense, Dropout, LSTM, BatchNormalization, Flatten
from tensorflow.compat.v1.keras.layers import CuDNNLSTM
from tensorflow.keras.callbacks import TensorBoard, ModelCheckpoint

def custom_loss(y_true, y_pred):

  mse = tf.keras.losses.MeanSquaredError()
  penalty = 10

  # penalize the loss heavily if the actual and the prediction are on different sides of zero
  loss = tf.cond( tf.logical_or( 
                    (tf.logical_and(tf.greater(y_true, 0.0), tf.less(y_pred, 0.0))), 
                    (tf.logical_and(tf.less(y_true, 0.0), tf.greater(y_pred, 0.0))) 
                    ),
                   lambda: mse(y_true, y_pred) * penalty,
                   lambda: mse(y_true, y_pred) * penalty / 4)
  
  print("starting second condition")
  # add slightly more penalty if prediction overshoots actual in any direction
  loss = tf.cond( tf.logical_or( 
                    (tf.logical_and(tf.greater(y_true, 0.0), tf.greater(y_pred, y_true))), 
                    (tf.logical_and(tf.less(y_true, 0.0), tf.less(y_pred, y_true))) 
                    ),
                   lambda: loss * penalty / 5,
                   lambda: loss * penalty / 10)
  return loss

EPOCHS = 25
BATCH_SIZE = 64
MODEL_NAME = f"MODEL 01-str(int(time.time())"

model = Sequential()
model.add(LSTM(128, input_shape=(train_x.shape[1:]), return_sequences=True))
model.add(Dropout(0.2))
model.add(BatchNormalization())

model.add(LSTM(128, input_shape=(train_x.shape[1:]), return_sequences=True))
model.add(Dropout(0.2))
model.add(BatchNormalization())

model.add(LSTM(128, input_shape=(train_x.shape[1:])))
model.add(Dropout(0.2))
model.add(BatchNormalization())

model.add(Flatten())

model.add(Dense(32, activation='relu'))
model.add(Dropout(0.2))
model.add(BatchNormalization())

model.add(Dense(1))

opt = tf.keras.optimizers.Adam(learning_rate=1e-3, decay=1e-6)

metric= tf.keras.metrics.MeanSquaredError()
model.compile(loss=custom_loss, optimizer=opt, metrics=[metric]) 
val_metric = 'val_'+metric.name

tensorboard = TensorBoard(log_dir=f'logs/MODEL_NAME')
filepath = base_path+"cryptodata/models/RNN_Final-epoch:02d-val_mean_squared_error:.3f-"+str(int(time.time()))+".hd5"
checkpoint = ModelCheckpoint(filepath=filepath, monitor=val_metric, verbose=0, mode='max',metric=metric)

train_x = np.random.randn(1588, 60, 34)  
train_y = np.random.rand(1588,)
val_x = np.random.randn(85, 60, 34)
val_y = np.random.randn(85,)

history = model.fit(train_x, train_y, 
                    batch_size=BATCH_SIZE,
                    epochs=100, 
                    validation_data=(val_x, val_y),
                    callbacks=[checkpoint, tensorboard])

我已经尝试在自定义损失函数中投射y_true 和y_pred，就像y_pred=tf.convert_to_tensor(y_pred); y_true = tf.cast(y_true, y_pred.dtype 一样，但这不起作用。还添加了 print 函数显示该函数被调用了两次成功但之后失败了。

我在使用内置损失函数时没有收到错误消息。

【参考方案1】：

问题在于您的custom_loss 返回的是函数而不是标量值。如果您将tf.cond 替换为tf.where，您的代码将起作用。

import numpy as np
import tensorflow as tf
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Dense, Dropout, LSTM, BatchNormalization, Flatten

def custom_loss(y_true, y_pred):

    mse = tf.keras.losses.MeanSquaredError()

    penalty = 10

    # penalize the loss heavily if the actual and the prediction are on different sides of zero
    loss = tf.where(
        condition=tf.logical_or((tf.logical_and(tf.greater(y_true, 0.0), tf.less(y_pred, 0.0))), (tf.logical_and(tf.less(y_true, 0.0), tf.greater(y_pred, 0.0)))),
        x=mse(y_true, y_pred) * penalty,
        y=mse(y_true, y_pred) * penalty / 4
    )

    # add slightly more penalty if prediction overshoots actual in any direction
    loss = tf.where(
        condition=tf.logical_or((tf.logical_and(tf.greater(y_true, 0.0), tf.greater(y_pred, y_true))), (tf.logical_and(tf.less(y_true, 0.0), tf.less(y_pred, y_true)))),
        x=loss * penalty / 5,
        y=loss * penalty / 10
    )

    return loss

train_x = np.random.randn(1588, 60, 34)
train_y = np.random.rand(1588, )
val_x = np.random.randn(85, 60, 34)
val_y = np.random.randn(85, )

model = Sequential()
model.add(LSTM(128, input_shape=(train_x.shape[1:]), return_sequences=True))
model.add(Dropout(0.2))
model.add(BatchNormalization())

model.add(LSTM(128, input_shape=(train_x.shape[1:]), return_sequences=True))
model.add(Dropout(0.2))
model.add(BatchNormalization())

model.add(LSTM(128, input_shape=(train_x.shape[1:])))
model.add(Dropout(0.2))
model.add(BatchNormalization())

model.add(Flatten())

model.add(Dense(32, activation='relu'))
model.add(Dropout(0.2))
model.add(BatchNormalization())

model.add(Dense(1))

opt = tf.keras.optimizers.Adam(learning_rate=1e-3, decay=1e-6)

model.compile(loss=custom_loss, optimizer=opt, metrics=['mse'])

model.fit(train_x, train_y, batch_size=128, epochs=3, validation_data=(val_x, val_y))
# Epoch 1/3
# 13/13 [==============================] - 8s 321ms/step - loss: 11.3129 - mse: 1.6341 - val_loss: 6.9313 - val_mse: 1.1116
# Epoch 2/3
# 13/13 [==============================] - 3s 234ms/step - loss: 7.3409 - mse: 1.0789 - val_loss: 7.2055 - val_mse: 1.1238
# Epoch 3/3
# 13/13 [==============================] - 3s 231ms/step - loss: 5.3962 - mse: 0.8513 - val_loss: 7.4492 - val_mse: 1.1512

model.predict(train_x)
# array([[0.25150445],
#        [0.2647993 ],
#        [0.2405027 ],
#        ...,
#        [0.31251353],
#        [0.29376918],
#        [0.21620636]], dtype=float32)