How to pass custom weights to scikit_learn wrappers (e.g. KerasClassifier) in a multilabel classification problem

Question

I'm building a chain classifier for a multiclass problem that uses Keras binary Classifier model in a chain. I have 17 labels as classification target and dataset is an imbalanced dataset for these classes. I want to customize weights and train my chain classifier model based on these weights. Normally, when not using scikit_learn wrappers, I pass the custom weights to the fit function.

This is the code to generate weights for these classes:

from sklearn.utils import class_weight
y_ints = [y.argmax() for y in y_train]
class_weights = class_weight.compute_class_weight(custom_weight_dict,
                                                 np.unique(y_ints),
                                                y_ints)

and here is my model that takes keras model as input and have a chain of binary classifiers.

def create_model():
  input_size=length_long_sentence
  embedding_size=128
  lstm_size=64
  output_size=len(unique_tag_set)
    #----------------------------Model--------------------------------
  current_input=Input(shape=(input_size,)) 
  emb_current = Embedding(vocab_size, embedding_size, input_length=input_size)(current_input)
  out_current=Bidirectional(LSTM(units=lstm_size))(emb_current )
  #out_current = Reshape((1,2*lstm_size))(out_current)
  output = Dense(units=1, activation=  'sigmoid')(out_current)
  #output = Dense(units=1, activation='softmax')(out_current)
  model = Model(inputs=current_input, outputs=output)
  #-------------------------------compile-------------
  model.compile(optimizer='Adam', loss='binary_crossentropy', metrics=['accuracy'])
  return model
model = KerasClassifier(build_fn=create_model, epochs=1,batch_size=256, shuffle = True, verbose = 1,validation_split=0.2)
chain=ClassifierChain(model, order=multi_label_order, random_state=42)
history=chain.fit(X_train, y_train)

The fit method of chain classifier only take Train features and train labels as input. Is their anyway that I can pass my class weights so that it can be used during training so accuracy of rare classes can be improved?

Answer 1

Keras Model.fit routine takes class_weight argument the same way as sklearn fit routines do.