Paddle的单机训练的python api

[reyoung]

实现使用python来驱动训练流程，达到类似下面的效果

import paddle

# Context是使用设备的上下文。Paddle究竟使用多少设备，在这里指定
context = paddle.Context(devices=[paddle.cpu_all, paddle.gpu_all])  # use all device in one node.

# 定义一个网络。前面的注解说明这个函数是一个网络定义。
@context.network()
def simple_network(network):
      # network参数是一个网络定义的函数集合，包括了我们支持的layers
      ipt = network.data_layer(name="input", size=784)
      hidden = network.fc_layer(input=ipt, size=200)
      predict = network.fc_layer(input=hidden, size=10, act=SoftmaxActivation())
      cost = network.classification_cost(input=predict, label=network.data_layer(name="input", size=10))
      return cost  # 返回优化的目标。相当于现在paddle的outputs

# define a data provider, same as current Paddle process.
@paddle.provider()
def process_data(settings, filename):
    for sample in read_from_file(filename):
         yield sample

# train all networks in current context.
context.with_train_data(train_files=["a.txt", "b.txt"], method=process_data)  # set train data, and data provider
     .with_test_data(test_files=["c.txt"], test_period=Batch(1000), method=process_data) # set test data
     .use_optimizer(SgdOptimizer())  # set optimizer.
     .standard_sgd_train(num_passes=100)  # set use standard sgd strategy to train 100 pass.


context.exit(0)

这个事情需要通过以下几个步骤完成:

• 清除掉Paddle运行时依赖的全局变量。否则，我们只能使用多进程来加载多个网络了。
• 添加需要的python api，来完成这项功能。

[Yancey1989]

We have developed v2 API, this issue is too old and close it.