Merge pull request #424 from yinhaofeng/dygraph_save

add dygraph to static

Author: seemingwang

doc/inference.md

@@ -1,5 +1,5 @@
 # Paddle Inference的使用方法
-paddlerec目前提供在静态图训练时使用save_inference_model接口保存模型，以及将保存的模型使用Inference预测库进行服务端部署的功能。本教程将以wide_deep模型为例，说明如何使用这两项功能。  
+paddlerec目前提供在静态图训练时使用save_inference_model接口保存模型，动态图训练后将保存的模型转化为静态图的样式，以及将保存的模型使用Inference预测库进行服务端部署的功能。本教程将以wide_deep模型为例，说明如何使用这三项功能。  
 
 ## 使用save_inference_model接口保存模型
 在服务器端使用python部署需要先使用save_inference_model接口保存模型。  
@@ -12,8 +12,8 @@ runner:
   ...
   # use inference save model
   use_inference: True  # 静态图训练时保存为inference model
-  save_inference_feed_varnames: ["label","C1","C2","C3","C4","C5","C6","C7","C8","C9","C10","C11","C12","C13","C14","C15","C16","C17","C18","C19","C20","C21","C22","C23","C24","C25","C26","dense_input"] # inference model 的feed参数的名字
-  save_inference_fetch_varnames: ["cast_0.tmp_0"] # inference model 的fetch参数的名字
+  save_inference_feed_varnames: ["C1","C2","C3","C4","C5","C6","C7","C8","C9","C10","C11","C12","C13","C14","C15","C16","C17","C18","C19","C20","C21","C22","C23","C24","C25","C26","dense_input"] # inference model 的feed参数的名字
+  save_inference_fetch_varnames: ["sigmoid_0.tmp_0"] # inference model 的fetch参数的名字
 ```
 3. 启动静态图训练
 ```bash
@@ -23,6 +23,39 @@ runner:
 python -u ../../../tools/static_trainer.py -m config.yaml # 全量数据运行config_bigdata.yaml 
 ```
 
+## 使用to_static.py脚本转化动态图保存下来的模型
+若您在使用动态图训练完成,希望将保存下来的模型转化为静态图inference，那么可以参考我们提供的to_static.py脚本。
+1. 首先正常使用动态图训练保存参数
+```bash
+# 进入模型目录
+# cd models/rank/wide_deep # 在任意目录均可运行
+# 动态图训练
+python -u ../../../tools/trainer.py -m config.yaml # 全量数据运行config_bigdata.yaml 
+```
+2. 打开yaml配置，增加`model_init_path`选项  
+to_static.py脚本会先加载`model_init_path`地址处的模型，然后再转化为静态图保存。注意不要在一开始训练时就打开这个选项，不然会变成热启动训练。
+3. 更改to_static脚本，根据您的模型需求改写其中to_static语句。
+我们以wide_deep模型为例，在wide_deep模型的组网中，需要保存前向forward的部分,具体代码可参考[net.py](https://github.com/PaddlePaddle/PaddleRec/blob/master/models/rank/wide_deep/net.py)。其输入参数为26个离散特征组成的list，以及1个连续特征。离散特征的shape统一为（batchsize，1）类型为int64，连续特征的shape为（batchsize，13）类型为float32。
+所以我们在to_static脚本中的paddle.jit.to_static语句中指定input_spec如下所示。input_spec的详细用法：[InputSpec 功能介绍](https://www.paddlepaddle.org.cn/documentation/docs/zh/guides/04_dygraph_to_static/input_spec_cn.html)。
+```python
+# example dnn and wide_deep model forward
+dy_model = paddle.jit.to_static(dy_model,
+    input_spec=[[paddle.static.InputSpec(shape=[None, 1], dtype='int64') for jj in range(26)], paddle.static.InputSpec(shape=[None, 13], dtype='float32')])
+```
+4. 运行to_static脚本, 参数为您的yaml文件，即可保存成功。将您在yaml文件中指定的model_init_path路径下的参数，转换并保存到model_save_path/(infer_end_epoch-1)目录下。  
+注：infer_end_epoch-1是因为epoch从0开始计数，如运行3个epoch即0~2
+```bash
+python -u ../../../tools/to_static.py -m config.yaml
+```
+5. 我们在使用inference预测库预测时也需要根据输入和输出做出对应的调整。比如我们保存的模型为wide_deep模型的组网中，前向forward的部分。输入为26个离散特征组成的list以及1个连续特征，输出为prediction预测值。所以我们在使用inference预测库预测时也需要将输入和输出做出对应的调整。  
+将criteo_reader.py输入数据中的label部分去除：  
+```python
+# 无需改动部分不再赘述
+# 在最后输出的list中，去除第一个np.array，即label部分。
+  yield output_list[1:]
+```
+将inference预测得到的prediction预测值和数据集中的label对比，使用另外的脚本计算auc指标即可。
+
 ## 将保存的模型使用Inference预测库进行服务端部署
 paddlerec提供tools/paddle_infer.py脚本，供您方便的使用inference预测库高效的对模型进行预测。  
 

doc/serving.md

@@ -12,8 +12,8 @@ runner:
   ...
   # use inference save model
   use_inference: True  # 静态图训练时保存为inference model
-  save_inference_feed_varnames: ["label","C1","C2","C3","C4","C5","C6","C7","C8","C9","C10","C11","C12","C13","C14","C15","C16","C17","C18","C19","C20","C21","C22","C23","C24","C25","C26","dense_input"] # inference model 的feed参数的名字
-  save_inference_fetch_varnames: ["cast_0.tmp_0"] # inference model 的fetch参数的名字
+  save_inference_feed_varnames: ["C1","C2","C3","C4","C5","C6","C7","C8","C9","C10","C11","C12","C13","C14","C15","C16","C17","C18","C19","C20","C21","C22","C23","C24","C25","C26","dense_input"] # inference model 的feed参数的名字
+  save_inference_fetch_varnames: ["sigmoid_0.tmp_0"] # inference model 的fetch参数的名字
 ```
 3. 启动静态图训练
 ```bash
@@ -102,6 +102,15 @@ python ../../../tools/webserver.py gpu 9393
 # CPU
 python ../../../tools/webserver.py cpu 9393
 ```
+### 调整reader
+我们在服务端底层使用Inference预测库预测。和直接使用Inference预测库一样,需要在reader中将输入和输出做出对应的调整。比如我们保存的模型为wide_deep模型的组网中。输入为26个离散特征组成的list以及1个连续特征，输出为prediction预测值。  
+将criteo_reader.py输入数据中的label部分去除：  
+```python
+# 无需改动部分不再赘述
+# 在最后输出的list中，去除第一个np.array，即label部分。
+  yield output_list[1:]
+```
+将预测得到的prediction预测值和数据集中的label对比，使用另外的脚本计算auc指标即可。
 
 ## 测试部署的服务
 在服务器端启动serving服务成功后，部署客户端需要您打开新的终端页面。

models/rank/wide_deep/config.yaml

@@ -22,7 +22,7 @@ runner:
   train_batch_size: 2
   epochs: 3
   print_interval: 2
-  #model_init_path: "output_model/0" # init model
+  # model_init_path: "output_model_wide_deep/2" # init model
   model_save_path: "output_model_wide_deep"
   test_data_dir: "data/sample_data/train"
   infer_reader_path: "criteo_reader" # importlib format
@@ -32,8 +32,8 @@ runner:
   infer_end_epoch: 3
   #use inference save model
   use_inference: False
-  save_inference_feed_varnames: ["label","C1","C2","C3","C4","C5","C6","C7","C8","C9","C10","C11","C12","C13","C14","C15","C16","C17","C18","C19","C20","C21","C22","C23","C24","C25","C26","dense_input"]
-  save_inference_fetch_varnames: ["cast_0.tmp_0"]
+  save_inference_feed_varnames: ["C1","C2","C3","C4","C5","C6","C7","C8","C9","C10","C11","C12","C13","C14","C15","C16","C17","C18","C19","C20","C21","C22","C23","C24","C25","C26","dense_input"]
+  save_inference_fetch_varnames: ["sigmoid_0.tmp_0"]
 
 # hyper parameters of user-defined network
 hyper_parameters:

tools/static_trainer.py

@@ -118,8 +118,8 @@ def main(args):
                 config, use_visual, log_visual, step_num)
             metric_str = ""
             for var_idx, var_name in enumerate(fetch_vars):
-                metric_str += "{}: {}, ".format(var_name,
-                                                fetch_batch_var[var_idx])
+                metric_str += "{}: {}, ".format(
+                    var_name, str(fetch_batch_var[var_idx]).strip("[]"))
             logger.info("epoch: {} done, ".format(epoch_id) + metric_str +
                         "epoch time: {:.2f} s".format(time.time() -
                                                       epoch_begin))

tools/to_static.py

@@ -69,6 +69,7 @@ def main(args):
     print_interval = config.get("runner.print_interval", None)
     model_save_path = config.get("runner.model_save_path", "model_output")
     model_init_path = config.get("runner.model_init_path", None)
+    end_epoch = config.get("runner.infer_end_epoch", 0)
 
     logger.info("**************common.configs**********")
     logger.info(
@@ -80,6 +81,7 @@ def main(args):
     place = paddle.set_device('gpu' if use_gpu else 'cpu')
 
     dy_model = dy_model_class.create_model(config)
+    model_save_path = os.path.join(model_save_path, str(end_epoch - 1))
 
     load_model(model_init_path, dy_model)
     # example dnn model forward