Keras LSTM时间序列二分类中class_weights引发图执行错误排查

• 任务：使用Keras训练时间序列LSTM模型解决二分类问题，标签分布极不平衡（0类占比约75%），目标是校正不平衡问题，重点减少假阳性。
• 输入输出尺寸：

  X_train.shape  --> (8000, 100, 4)
  X_test.shape  --> (2000, 100, 4)
  y_train.shape  --> (8000, 100)
  y_test.shape  --> (2000, 100)
  

  注：每个样本对应100个时间点的预测任务

model = Sequential()
model.add(LSTM(64, input_shape=(X_train.shape[1], X_train.shape[2]), use_bias=True, unroll=True, kernel_initializer='glorot_normal', return_sequences=True))
model.add(BatchNormalization())
model.add(Dropout(.25))
model.add(LSTM(32, return_sequences=False, use_bias=True, unroll=True))
model.add(Dense(num_points_per_inp, activation='sigmoid'))
model.compile(optimizer=Adam(beta_1=.8, beta_2=.9), loss=tf.keras.losses.CategoricalCrossentropy(from_logits=True), metrics=tf.keras.metrics.BinaryAccuracy())  # loss=weighted_binary_crossentropy, metrics='accuracy')
model.summary()

已尝试调整BatchNormalization（BN）和Dropout（DO）的不同组合，若模型本身存在问题请指出。

不使用class_weights时模型可正常编译训练，但添加后触发以下错误：

---------------------------------------------------------------------------
InvalidArgumentError                      Traceback (most recent call last)
<ipython-input-113-2ebc95d5f9f5> in <module>()
      7                     validation_data=(X_test, y_test),
      8                     class_weight=weights,
----> 9                     verbose=1)

1 frames
/usr/local/lib/python3.7/dist-packages/tensorflow/python/eager/execute.py in quick_execute(op_name, num_outputs, inputs, attrs, ctx, name)
     53     ctx.ensure_initialized()
     54     tensors = pywrap_tfe.TFE_Py_Execute(ctx._handle, device_name, op_name,
---> 55                                         inputs, attrs, num_outputs)
     56   except core._NotOkStatusException as e:
     57     if name is not None:

InvalidArgumentError: Graph execution error:

2 root error(s) found.
  (0) INVALID_ARGUMENT:  indices[2] = 52 is not in [0, 2)
     [[{{node GatherV2}}]]
     [[IteratorGetNext]]
     [[IteratorGetNext/_2]]
  (1) INVALID_ARGUMENT:  indices[2] = 52 is not in [0, 2)
     [[{{node GatherV2}}]]
     [[IteratorGetNext]]
0 successful operations.
0 derived errors ignored. [Op:__inference_train_function_769866]

方案1：用sklearn生成class_weights #

from sklearn.utils import class_weight

w = class_weight.compute_class_weight(
          class_weight = 'balanced', 
          classes = np.unique(y_train), 
          y = y_train.flatten())
weights = dict(enumerate(np.round(w,2)))

方案2：手动计算class_weights #

counts = np.bincount(y_train.flatten().astype(int))
w0 = 1 / counts[0]
w1 = 1 / counts[1]
weights = {0: w0, 1: w1}

方案3：自定义加权损失函数（效果不符合预期） #

import keras.backend as K
def weighted_binary_crossentropy(y_true, y_pred):
  weights =(tf.math.abs(y_true) * 59.) + 1.
  bce = K.binary_crossentropy(y_true, y_pred)
  weighted_bce = K.mean(bce * weights)
  return weighted_bce

猜测问题可能与每个样本包含100个时间点的预测任务有关，但不确定；尝试使用CategoricalCrossentropy也无济于事。

请问该如何解决这个问题？

内容的提问来源于stack exchange，提问作者ArieAI