Merge branch 'release/v1.7' of https://github.com/PaddlePaddle/PaddleHub into release/v1.7

README.md

@@ -50,7 +50,8 @@ PaddleHub以预训练模型应用为核心具备以下特点：
 
 ### 安装命令
 
-PaddlePaddle框架的安装请查阅[飞桨快速安装](https://www.paddlepaddle.org.cn/install/quick)
+在安装paddlehub之前，请先安装PaddlePaddle深度学习框架。
+PaddlePaddle深度学习框架的安装请查阅[飞桨快速安装](https://www.paddlepaddle.org.cn/install/quick)
 
 ```shell
 pip install paddlehub

paddlehub/autodl/DELTA/README.md → autodl/DELTA/README.md

+# DELTA: DEep Learning Transfer using Feature Map with Attention for Convolutional Networks
 
-# Introduction
-This page implements the [DELTA](https://arxiv.org/abs/1901.09229) algorithm in [PaddlePaddle](https://www.paddlepaddle.org.cn/install/quick).
+## Introduction
+
+This page implements the [DELTA](https://arxiv.org/abs/1901.09229) algorithm in [PaddlePaddle](https://www.paddlepaddle.org.cn).
 
 > Li, Xingjian, et al. "DELTA: Deep learning transfer using feature map with attention for convolutional networks." ICLR 2019.
 
-# Preparation of Data and Pre-trained Model
+## Preparation of Data and Pre-trained Model
 
 - Download transfer learning target datasets, like [Caltech-256](http://www.vision.caltech.edu/Image_Datasets/Caltech256/), [CUB_200_2011](http://www.vision.caltech.edu/visipedia/CUB-200-2011.html) or others. Arrange the dataset in this way:
 ```
@@ -23,7 +25,7 @@ This page implements the [DELTA](https://arxiv.org/abs/1901.09229) algorithm in
 
 - Download [the pretrained models](https://github.com/PaddlePaddle/models/tree/release/1.7/PaddleCV/image_classification#resnet-series). We give the results of ResNet-101 below.
 
-# Running Scripts
+## Running Scripts
 
 Modify `global_data_path` in `datasets/data_path` to the path root where the dataset is.
 

paddlehub/finetune/task/base_task.py

@@ -832,7 +832,9 @@ class BaseTask(object):
             self.config.checkpoint_dir,
             self.exe,
             main_program=self.main_program)
-
+        # Revise max_train_steps when incremental training
+        self.max_train_steps = self.env.current_step + self.max_train_steps / self.config.num_epoch * (
+            self.config.num_epoch - self.env.current_epoch + 1)
         return is_load_successful
 
     def load_parameters(self, dirname):