model_example.py

class TFBertClassifier(TFBertPreTrainedModel):
    def __init__(self, dir_path, num_class=2):
        super().__init__(config)
        
        self.bert = TFBertModel(name="bert").from_pretrained('bert-base-multilingual-cased', cache_dir=dir_path)
        self.num_class = num_class
        self.dropout = tf.keras.layers.Dropout(self.bert.config.hidden_dropout_prob)
        self.classifier = tf.keras.layers.Dense(self.num_class, 
                                                kernel_initializer=tf.keras.initializers.TruncatedNormal(config.initializer_range), 
                                                name="classifier")
        
    def call(self, inputs, attention_mask=None, token_type_ids=None, training=False):
        
        #outputs 값: # sequence_output, pooled_output, (hidden_states), (attentions)
        outputs = self.bert(inputs, attention_mask=attention_mask, token_type_ids=token_type_ids)
        pooled_output = outputs[1] 
        pooled_output = self.dropout(pooled_output, training=training)
        logits = self.classifier(pooled_output)
        
        outputs = (logits,) + outputs[2:] 

        return outputs
    
    
class TFGPT2Classifier(tf.keras.Model):
    def __init__(self, dir_path, num_class=2):
        super(TFGPT2Classifier, self).__init__()
        
        self.gpt2 = TFGPT2Model.from_pretrained(dir_path)
        self.num_class = num_class
        
        self.dropout = tf.keras.layers.Dropout(self.gpt2.config.summary_first_dropout)
        self.classifier = tf.keras.layers.Dense(self.num_class, 
                                                kernel_initializer=tf.keras.initializers.TruncatedNormal(stddev=self.gpt2.config.initializer_range), 
                                                name="classifier")
        
    def call(self, inputs):
        outputs = self.gpt2(inputs)
        pooled_output = outputs[0][:, -1]

        pooled_output = self.dropout(pooled_output, training=training)
        logits = self.classifier(pooled_output)

        outputs = (logits,) + outputs[2:] 

        return outputs

class TFBertClassifier(tf.keras.Model):
    def __init__(self, model_name, dir_path, num_class):
        super(TFBertClassifier, self).__init__()

        self.bert = TFBertModel.from_pretrained(model_name, cache_dir=dir_path)
        self.num_class = num_class
        self.dropout = tf.keras.layers.Dropout(self.bert.config.hidden_dropout_prob)
        self.classifier = tf.keras.layers.Dense(self.num_class, 
                                                kernel_initializer=tf.keras.initializers.TruncatedNormal(self.bert.config.initializer_range),
                                                name="classifier")

    def call(self, inputs, attention_mask=None, token_type_ids=None, training=False):

        #outputs 값: # sequence_output, pooled_output, (hidden_states), (attentions)
        outputs = self.bert(inputs, attention_mask=attention_mask, token_type_ids=token_type_ids)
        pooled_output = outputs[1] 
        pooled_output = self.dropout(pooled_output, training=training)
        logits = self.classifier(pooled_output)

        return logits

class TFGPT2Classifier(tf.keras.Model):
    def __init__(self, dir_path, num_class):
        super(TFGPT2Classifier, self).__init__()
        
        self.gpt2 = TFGPT2Model.from_pretrained(dir_path)
        self.num_class = num_class
        self.dropout = tf.keras.layers.Dropout(self.gpt2.config.summary_first_dropout)
        self.classifier = tf.keras.layers.Dense(self.num_class, 
                                                kernel_initializer=tf.keras.initializers.TruncatedNormal(stddev=self.gpt2.config.initializer_range), 
                                                name="classifier")
        
    def call(self, inputs):
        outputs = self.gpt2(inputs)
        pooled_output = outputs[0][:, -1]

        pooled_output = self.dropout(pooled_output, training=training)
        logits = self.classifier(pooled_output)

        return logits

@reniew
@Taekyoon

안녕하세요 binary classification 문제에 모델 실험을 해보다 막힌 부분이 있어 연락드립니다.

model.predict(테스트데이터)를 실행하니 (len(테스트데이터),2) 형태의 logit이 return됩니다.

logit[0] 이 class1에 속할 logit, logit[1]이 class2에 속할 logit이 맞을까요?

logit을 binary result로 변환시키고 싶은데 def call 함수에서 return logits를 tf.round(tf.nn.sigmoid(logits))로 바꾸면 될까요?