当使用PaddlePaddle训练大量深层的 神经网络 时，开发者希望跟踪整个神经训练过程的信息，例如loss的变化趋势，迭代过程中 参数 的变化、分布及背后的原因、模型的网络结构等。PaddlePaddle推出的可视化工具VisualDL，只需要以下两个简单步骤，即可更好的调试和优化 神经网络 的训练过程：

• 调用Paddle Fluid的模型接口save_inference_model进行模型保存

• 使用VisualDL提供的命令行工具加载模型，在浏览器中访问

具体应用方式，请听小编为你详细解读：

# 安装VisualDL，建议使用anaconda或者virtualenv等环境下
  


pip install --upgrade visualdl
   
  


# vdl_create_scratch_log将会自动生成一份测试的log
  


vdl_create_scratch_log
  


# 在localhost:8080上启动一个visualdl服务，该服务加载scratch_log目录中的日志进行展示
 


visualdl --logdir=scratch_log --model_pb scratch_log/mnist_model.onnx --port=8080
   


# 访问 http://127.0.0.1:8080

import paddle
  


import paddle.fluid as fluid
 


from visualdl import LogWriter

  # 构建一个LeNet
  -5网络
  


  def lenet_5(img, label):
      


      conv1 = fluid.nets.simple_img_conv_pool(
          


          input=img,
                       


          filter_size=5,
          


          num_filters=20,
          


          pool_size=2,
          


          pool_stride=2,
          


          act="relu")
 
    


      conv1_bn = fluid.layers.batch_norm(input=conv1)
 
    


      conv2 = fluid.nets.simple_img_conv_pool(
        


          input=conv1_bn,
        


          filter_size=5,
        


          num_filters=50,
        


          pool_size=2,
        


          pool_stride=2,
        


          act="relu")
 
    


      predition = fluid.layers.fc(input=conv2, size=10, act="softmax", param_attr = "fc_w")




      cost = fluid.layers.cross_entropy(input=predition, label=label)
    


      avg_cost = fluid.layers.mean(cost)
    


      acc = fluid.layers.accuracy(input=predition, label=label)
 
    


      return avg_cost, acc

  img = fluid.layers.data(name="img", shape=[1, 28, 28], dtype="float32")
  


  label = fluid.layers.data(name="label", shape=[1], dtype="int64")
  


  avg_cost, acc = lenet_5(img, label)

  train_reader = paddle.batch(paddle.dataset.mnist.train(), batch_size=64)

  optimizer = fluid.optimizer.Adam(learning_rate=0.001)
  


  optimizer.minimize(avg_cost)

  place = fluid.CPUPlace()  feeder = fluid.DataFeeder(feed_list=[img, label], place=place)  exe = fluid.Executor(place)

log_writter = LogWriter("./vdl_log", sync_cycle=10)
  


  with log_writter.mode("train") as logger:
      


      # 分别用两个scalar来记录loss和accuracy的趋势
      


      scalar_loss = logger.scalar(tag="loss")
      


      scalar_accuracy = logger.scalar(tag="accuracy")
      


      # 用histogram来记录fc层的参数
  分布变化趋势
      


      histogram = logger.histogram(tag="histogram", num_buckets=50)

  exe.run(fluid.default_startup_program())
  


  step = 0
  


  epochs = 5
  


  param_name = "fc_w"

  for i in range(epochs):
      


      for batch in train_reader():
          


          cost, accuracy, param = exe.run(
              


              feed=feeder.feed(batch),
              


              fetch_list=[avg_cost.name, acc.name, param_name])
          


          step += 1
   
          


          # 将数据写入记录中
          


          scalar_loss.add_record(step, cost)
        


        scalar_accuracy.add_record(step, accuracy)
        


        histogram.add_record(step, param.flatten())
 
        


        print("epoch %d: step %d, acc is %.2f%% and loss is %.2f" % (i, step, accuracy * 100, cost))
  


  # 保存模型到model_save目录中，接着可以用visualdl --model_pb model_save来加载该模型      
fluid.io.save_inference_model(dirname = "model_save", feeded_var_names = [img.name], target_vars = [avg_cost, acc], executor = exe)

 visualdl --logdir "vdl_log" --model_pb "model_save"

--host 设定IP 


--port 设定端口 


-m / --model_pb 指定 ONNX 格式的模型文件或者paddle模型文件 


-L / --language 指定界面使用的语言，目前支持en和zh两种，默认为en