TensorFlow.Keras代码在Jupyter Notebook中仅首次运行正常的问题求助

我最近碰到个棘手的问题：在Jupyter Notebook里用tensorflow.keras搭建神经网络拟合真实数据，第一次运行代码完全正常，能顺利生成符合要求的模型，但第二次再运行时，不管我把成功标准（代码里的NSE_cut）调得多低，都找不到合适的拟合模型了。更奇怪的是没有任何报错信息，代码看起来和第一次运行时一样正常执行，就是死活拟合不出结果。

现在我只能重启Jupyter的内核才能让代码重新生成模型，但这样会清空之前所有的数据和处理结果。更关键的是，我需要给不同处理程度的输入数据（比如无平滑、轻微平滑、重度平滑等）分别建模，对比哪个效果最好，总不能每次建模都重启内核吧？

我到底哪里操作错了？为什么代码只有第一次运行能生成有效模型？

附上我的代码：

### --- Calculate Neural Network fit for modes --- ###
def nn_fit(reof_ds, q, NSE_cut):
    indx_qual_mode = []
    best_model = []
    best_score = []
    best_nse = 0

    # ----- Train Tensorflow Hydro-to-TPC models mode-by-mode -----
    for mode in reof_ds.mode.values:  
        keras.backend.clear_session()    
        print('Building models for mode-'+str(mode).zfill(2))
        tpc = reof_ds.temporal_modes.sel(mode=int(mode))        

        X = q.values.reshape(len(q),1)
        Y = tpc.values.reshape(len(q),1)
    
        X_train = X
        Y_train = Y     
    
        # -- adapt function is to get the mean and STD used to normalize the input data of the model --
        normalizer.adapt(X_train)    
    
        # ----- Construct the model and get summary -----
        model = build_and_compile_model(normalizer)
        #if vis_tf_nn==0:
        #elif vis_tf_nn==1:
        #  plot_model(model, to_file='NOAA_WF\\hydro2rtpc_mdl\\'+data_src+'\\site-'+str(gaugeID_list[site])+'_tpc'+str(mode+1).zfill(2)+'.png', show_shapes=True, show_layer_names=True)
    
         # ----- Fit the model -----
        train_proc = model.fit(
            X_train, 
            Y_train, 
            callbacks=[callback],
            batch_size=32, 
            epochs=200, 
            verbose=0, 
            #validation_split=0.2
        )    

        # ----- Plot model estimation and original scatter plot -----
        X_sim = tf.linspace(np.amin(X), np.amax(X), X.size*10^10)
        Y_sim = model.predict(X_sim)
   

        # ----- The second-time REOF mode screening based on quality of regression models. If qualified, export trained model -----
        Y_mdl = model.predict(X[:,0])
        nse = 1 - ( (np.nansum(np.square( Y - Y_mdl ))) / (np.nansum(np.square( Y - np.nanmean(Y) ))) )

        if nse >= NSE_cut: # Moriasi et al., 2007. Consider NSE>0.5 as satisfactory
            
            model.summary()
            print(mode, nse)

            # ----- Plot training progress -----  
            plot_loss(train_proc)

            best_nse = nse
            best_score.append(best_nse)
            best_model.append(model)
            indx_qual_mode.append(mode)

            fig = plt.figure()
            ax = fig.add_axes([0,0,1,1])
            plt.scatter(X, Y)
            plt.plot(X_sim, Y_sim, color='r')
            plt.xlabel('discharge (m3/d)')
            plt.ylabel(f"mode {mode}")  
            plt.xticks(rotation=45, ha='right')
            plt.yticks(rotation=45)
            plt.legend(['NN-Model','Data'],loc='upper left')
            plt.text(0.2,0.5,'NSE: '+"{:.2f}".format(nse), transform=ax.transAxes)
            plt.savefig('test.png', dpi=300, bbox_inches='tight')
            plt.show()
    
            # ----- Export trained model -----
            #best_model.save('test.keras')  
    # ADDED by Knicely
        del Y_mdl, X_sim, Y_sim, X, Y, X_train, Y_train, train_proc, model
    keras.backend.clear_session()    
    
    return(best_model, indx_qual_mode)

注：reof_ds包含旋转经验正交函数分解得到的空间和时间模态。

备注：内容来源于stack exchange，提问作者Kas Knicely