Question

需要重头开始训练的话把保存的权重文件删除

每训练 1000 次输出一次特征图

训练完一轮就保存一次权重

训练 40 个 epoch

def draw_train_process(iters,accs,loss):
    '''
    训练可视化
    '''
    plt.title('cat to classes12 ',fontsize=24)
    plt.xlabel('iters',fontsize=20)
    plt.ylabel('acc/loss',fontsize=20)
    plt.plot(iters,accs,color='red',label='accuracy')
    plt.plot(iters,loss,color='green',label='loss')
    plt.legend()
    plt.grid()
    plt.show()

with fluid.dygraph.guard():
    train_loader=data_load(train_paramters['train_image_list'],train_paramters['batch_size'])
    model=VGG16net()  #实列化模型
    if os.path.exists(train_paramters['save_model_name']+'.pdparams'):#存在模型参数则继续训练
        print('continue training')
        param_dict,_=fluid.dygraph.load_dygraph(train_paramters['save_model_name'])
        model.load_dict(param_dict)
    model.train()
    is_display_feature=False
    all_iter=0
    all_loss=[]
    all_iters=[]
    all_accs=[]
    opt=fluid.optimizer.SGDOptimizer(learning_rate=0.01,parameter_list=model.parameters())
    for pass_num in range(train_paramters['epoch_num']):
        for pass_id,data in enumerate(train_loader()):
            images,labels = data
            images = fluid.dygraph.to_variable(images)
            labels = fluid.dygraph.to_variable(labels)
            if(all_iter%893==0):
                is_display_feature=True
                out = model(images,is_display_feature)
                is_display_feature=False
            else:
                out = model(images,is_display_feature)
            loss=fluid.layers.cross_entropy(label=labels,input=out)
            avg_loss=fluid.layers.mean(loss)
            avg_loss.backward()
            opt.minimize(avg_loss)
            opt.clear_gradients()
            all_iter+=1
        #每训练完一个 epoch（reader()）
        acc=fluid.layers.accuracy(input=out,label=labels)
        all_loss.append(avg_loss.numpy()[0])
        all_iters.append(all_iter)
        all_accs.append(acc.numpy()[0])
        print('pass_epoch:{},iters:{},loss：{},acc：{}'.format(pass_num,all_iter,avg_loss.numpy()[0],acc.numpy()[0]))
        fluid.save_dygraph(model.state_dict(),train_paramters['save_model_name']) #保存模型参数
    draw_train_process(all_iters,all_accs,all_loss)
    print('finished training')

<Figure size 432x288 with 1 Axes>

<Figure size 432x288 with 1 Axes>

<Figure size 432x288 with 1 Axes>

<Figure size 432x288 with 1 Axes>

<Figure size 432x288 with 1 Axes>

<Figure size 432x288 with 1 Axes>

<Figure size 432x288 with 1 Axes>

<Figure size 432x288 with 1 Axes>

<Figure size 432x288 with 1 Axes>

<Figure size 432x288 with 1 Axes>

<Figure size 432x288 with 1 Axes>

<Figure size 432x288 with 1 Axes>

<Figure size 432x288 with 1 Axes>

pass_epoch:0,iters:67,loss：2.4799280166625977,acc：0.125
pass_epoch:1,iters:134,loss：2.517878532409668,acc：0.0625
pass_epoch:2,iters:201,loss：2.3678321838378906,acc：0.125
pass_epoch:3,iters:268,loss：2.5074377059936523,acc：0.09375
pass_epoch:4,iters:335,loss：2.395510196685791,acc：0.09375
pass_epoch:5,iters:402,loss：2.3040196895599365,acc：0.3125
pass_epoch:6,iters:469,loss：2.318796157836914,acc：0.15625
pass_epoch:7,iters:536,loss：2.408168077468872,acc：0.1875
pass_epoch:8,iters:603,loss：2.1314830780029297,acc：0.25
pass_epoch:9,iters:670,loss：2.3354663848876953,acc：0.21875
pass_epoch:10,iters:737,loss：2.0989861488342285,acc：0.21875
pass_epoch:11,iters:804,loss：2.01800799369812,acc：0.34375
pass_epoch:12,iters:871,loss：2.064736843109131,acc：0.1875

<Figure size 432x288 with 1 Axes>

<Figure size 432x288 with 1 Axes>

<Figure size 432x288 with 1 Axes>

<Figure size 432x288 with 1 Axes>

<Figure size 432x288 with 1 Axes>

<Figure size 432x288 with 1 Axes>

<Figure size 432x288 with 1 Axes>

<Figure size 432x288 with 1 Axes>

<Figure size 432x288 with 1 Axes>

<Figure size 432x288 with 1 Axes>

<Figure size 432x288 with 1 Axes>

<Figure size 432x288 with 1 Axes>

<Figure size 432x288 with 1 Axes>

pass_epoch:13,iters:938,loss：1.6188513040542603,acc：0.46875
pass_epoch:14,iters:1005,loss：1.7495455741882324,acc：0.375
pass_epoch:15,iters:1072,loss：2.243290901184082,acc：0.15625
pass_epoch:16,iters:1139,loss：2.2192680835723877,acc：0.25
pass_epoch:17,iters:1206,loss：2.190603733062744,acc：0.25
pass_epoch:18,iters:1273,loss：1.7756171226501465,acc：0.375
pass_epoch:19,iters:1340,loss：2.0200839042663574,acc：0.34375
pass_epoch:20,iters:1407,loss：1.5426690578460693,acc：0.40625
pass_epoch:21,iters:1474,loss：1.4749078750610352,acc：0.5
pass_epoch:22,iters:1541,loss：0.8402727246284485,acc：0.75
pass_epoch:23,iters:1608,loss：1.150139331817627,acc：0.5625
pass_epoch:24,iters:1675,loss：1.0551210641860962,acc：0.53125
pass_epoch:25,iters:1742,loss：1.583927035331726,acc：0.40625

<Figure size 432x288 with 1 Axes>

<Figure size 432x288 with 1 Axes>

<Figure size 432x288 with 1 Axes>

<Figure size 432x288 with 1 Axes>

<Figure size 432x288 with 1 Axes>

<Figure size 432x288 with 1 Axes>

<Figure size 432x288 with 1 Axes>

<Figure size 432x288 with 1 Axes>

<Figure size 432x288 with 1 Axes>

<Figure size 432x288 with 1 Axes>

<Figure size 432x288 with 1 Axes>

<Figure size 432x288 with 1 Axes>

<Figure size 432x288 with 1 Axes>

pass_epoch:26,iters:1809,loss：1.2691419124603271,acc：0.5
pass_epoch:27,iters:1876,loss：1.7214078903198242,acc：0.5
pass_epoch:28,iters:1943,loss：0.85252845287323,acc：0.71875
pass_epoch:29,iters:2010,loss：1.1482943296432495,acc：0.5625
pass_epoch:30,iters:2077,loss：0.7138460874557495,acc：0.84375
pass_epoch:31,iters:2144,loss：0.30972999334335327,acc：0.875
pass_epoch:32,iters:2211,loss：0.8071508407592773,acc：0.6875
pass_epoch:33,iters:2278,loss：1.2095224857330322,acc：0.71875
pass_epoch:34,iters:2345,loss：0.25210511684417725,acc：0.96875
pass_epoch:35,iters:2412,loss：0.707565188407898,acc：0.6875
pass_epoch:36,iters:2479,loss：0.47325584292411804,acc：0.90625
pass_epoch:37,iters:2546,loss：0.025728696957230568,acc：1.0
pass_epoch:38,iters:2613,loss：0.0027548810467123985,acc：1.0

<Figure size 432x288 with 1 Axes>

<Figure size 432x288 with 1 Axes>

<Figure size 432x288 with 1 Axes>

<Figure size 432x288 with 1 Axes>

<Figure size 432x288 with 1 Axes>

<Figure size 432x288 with 1 Axes>

<Figure size 432x288 with 1 Axes>

<Figure size 432x288 with 1 Axes>

<Figure size 432x288 with 1 Axes>

<Figure size 432x288 with 1 Axes>

<Figure size 432x288 with 1 Axes>

<Figure size 432x288 with 1 Axes>

<Figure size 432x288 with 1 Axes>

pass_epoch:39,iters:2680,loss：0.0015298365615308285,acc：1.0

<Figure size 432x288 with 1 Axes>

finished training