from graphCodeBert import GraphCodeBert from keras.models import load_model, Model import numpy as np, json class Predict: def __generate_code_embedding(self,code_snippet): embedding = np.array(GraphCodeBert().generate_individual_embedding(code_snippet)).reshape((1,768)) return embedding def __calculate_loss(self,code_embedding,model_name): model:Model = load_model(f'results/{model_name}.hdf5') return model.evaluate(code_embedding,code_embedding) def predict(self,code_snippet): model_name="autoencoder_25" code_embedding = self.__generate_code_embedding(code_snippet) print("Input code snippet shape: ",code_embedding.shape) loss = self.__calculate_loss(code_embedding,model_name) print("Reconstruction Loss: ",loss) with open('./results/metrics.json',"r") as fp: metric_json = json.loads(fp.read()) threshold = metric_json["Threshold"] return "Not a candidate for refactoring" if loss>threshold else "Is a candidate for refactoring", threshold, loss if __name__=="__main__": Predict().predict(""" public void sleep(){ int s1 = 1; int s2 = 2; int s3 = 3; int s4 = 4; int s5 = 5; int s6 = 6; int s7 = 7; int s8 = 8; }""")