Merge pull request #22 from PaddlePaddle/develop

update

Author: zhengya01

fluid/PaddleCV/gan/cycle_gan/train.py

@@ -147,19 +147,22 @@ def init_model():
         init_model()
     losses = [[], []]
     t_time = 0
+    build_strategy = fluid.BuildStrategy()
+    build_strategy.enable_inplace = False
+    build_strategy.memory_optimize = False
 
     g_A_trainer_program = fluid.CompiledProgram(
         g_A_trainer.program).with_data_parallel(
-            loss_name=g_A_trainer.g_loss_A.name)
+            loss_name=g_A_trainer.g_loss_A.name, build_strategy=build_strategy)
     g_B_trainer_program = fluid.CompiledProgram(
         g_B_trainer.program).with_data_parallel(
-            loss_name=g_B_trainer.g_loss_B.name)
+            loss_name=g_B_trainer.g_loss_B.name, build_strategy=build_strategy)
     d_B_trainer_program = fluid.CompiledProgram(
         d_B_trainer.program).with_data_parallel(
-            loss_name=d_B_trainer.d_loss_B.name)
+            loss_name=d_B_trainer.d_loss_B.name, build_strategy=build_strategy)
     d_A_trainer_program = fluid.CompiledProgram(
         d_A_trainer.program).with_data_parallel(
-            loss_name=d_A_trainer.d_loss_A.name)
+            loss_name=d_A_trainer.d_loss_A.name, build_strategy=build_strategy)
     for epoch in range(args.epoch):
         batch_id = 0
         for i in range(max_images_num):

fluid/PaddleCV/gan/cycle_gan/trainer.py

@@ -13,8 +13,6 @@ def __init__(self, input_A, input_B):
         self.program = fluid.default_main_program().clone()
         with fluid.program_guard(self.program):
             self.fake_B = build_generator_resnet_9blocks(input_A, name="g_A")
-            #FIXME set persistable explicitly to pass CE
-            self.fake_B.persistable = True
             self.fake_A = build_generator_resnet_9blocks(input_B, name="g_B")
             self.cyc_A = build_generator_resnet_9blocks(self.fake_B, "g_B")
             self.cyc_B = build_generator_resnet_9blocks(self.fake_A, "g_A")
@@ -60,8 +58,6 @@ def __init__(self, input_A, input_B):
         with fluid.program_guard(self.program):
             self.fake_B = build_generator_resnet_9blocks(input_A, name="g_A")
             self.fake_A = build_generator_resnet_9blocks(input_B, name="g_B")
-            #FIXME set persistable explicitly to pass CE
-            self.fake_A.persistable = True
             self.cyc_A = build_generator_resnet_9blocks(self.fake_B, "g_B")
             self.cyc_B = build_generator_resnet_9blocks(self.fake_A, "g_A")
             self.infer_program = self.program.clone()

fluid/PaddleCV/rcnn/.run_ce.sh

@@ -7,11 +7,11 @@ export OMP_NUM_THREADS=1
 cudaid=${face_detection:=0} # use 0-th card as default
 export CUDA_VISIBLE_DEVICES=$cudaid
 
-FLAGS_benchmark=true  python train.py --model_save_dir=output/ --data_dir=dataset/coco/ --max_iter=10 --enable_ce --pretrained_model=./imagenet_resnet50_fusebn | python _ce.py
+FLAGS_benchmark=true  python train.py --model_save_dir=output/ --data_dir=dataset/coco/ --max_iter=100 --enable_ce --pretrained_model=./imagenet_resnet50_fusebn | python _ce.py
 
 
 cudaid=${face_detection_m:=0,1,2,3} # use 0,1,2,3 card as default
 export CUDA_VISIBLE_DEVICES=$cudaid
 
-FLAGS_benchmark=true  python train.py --model_save_dir=output/ --data_dir=dataset/coco/ --max_iter=10 --enable_ce --pretrained_model=./imagenet_resnet50_fusebn | python _ce.py
+FLAGS_benchmark=true  python train.py --model_save_dir=output/ --data_dir=dataset/coco/ --max_iter=100 --enable_ce --pretrained_model=./imagenet_resnet50_fusebn | python _ce.py
 

fluid/PaddleCV/rcnn/README.md

@@ -12,7 +12,7 @@
 
 ## Installation
 
-Running sample code in this directory requires PaddelPaddle Fluid v.1.0.0 and later. If the PaddlePaddle on your device is lower than this version, please follow the instructions in [installation document](http://www.paddlepaddle.org/documentation/docs/zh/0.15.0/beginners_guide/install/install_doc.html#paddlepaddle) and make an update.
+Running sample code in this directory requires PaddelPaddle Fluid v.1.3.0 and later. If the PaddlePaddle on your device is lower than this version, please follow the instructions in [installation document](http://paddlepaddle.org/documentation/docs/en/1.3/beginners_guide/install/index_en.html) and make an update.
 
 ## Introduction
 
@@ -37,6 +37,25 @@ Train the model on [MS-COCO dataset](http://cocodataset.org/#download), download
     cd dataset/coco
     ./download.sh
 
+The data catalog structure is as follows:
+
+  ```
+  data/coco/
+  ├── annotations
+  │   ├── instances_train2014.json
+  │   ├── instances_train2017.json
+  │   ├── instances_val2014.json
+  │   ├── instances_val2017.json
+  |   ...
+  ├── train2017
+  │   ├── 000000000009.jpg
+  │   ├── 000000580008.jpg
+  |   ...
+  ├── val2017
+  │   ├── 000000000139.jpg
+  │   ├── 000000000285.jpg
+  |   ...
+  ```
 
 ## Training
 
@@ -51,9 +70,8 @@ Please make sure that pretrained_model is downloaded and loaded correctly, other
 
 To train the model, [cocoapi](https://github.com/cocodataset/cocoapi) is needed. Install the cocoapi:
 
-    # COCOAPI=/path/to/clone/cocoapi
-    git clone https://github.com/cocodataset/cocoapi.git $COCOAPI
-    cd $COCOAPI/PythonAPI
+    git clone https://github.com/cocodataset/cocoapi.git
+    cd cocoapi/PythonAPI
     # if cython is not installed
     pip install Cython
     # Install into global site-packages
@@ -66,25 +84,29 @@ After data preparation, one can start the training step by:
 
 - Faster RCNN
 
+    ```
     python train.py \
        --model_save_dir=output/ \
        --pretrained_model=${path_to_pretrain_model} \
        --data_dir=${path_to_data} \
        --MASK_ON=False
+    ```
 
 - Mask RCNN
 
+    ```
     python train.py \
        --model_save_dir=output/ \
        --pretrained_model=${path_to_pretrain_model} \
        --data_dir=${path_to_data} \
        --MASK_ON=True
+    ```
 
-- Set ```export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7``` to specifiy 8 GPU to train.
-- Set ```MASK_ON``` to choose Faster RCNN or Mask RCNN model.
-- For more help on arguments:
+    - Set ```export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7``` to specifiy 8 GPU to train.
+    - Set ```MASK_ON``` to choose Faster RCNN or Mask RCNN model.
+    - For more help on arguments:
 
-    python train.py --help
+        python train.py --help
 
 **data reader introduction:**
 
@@ -116,20 +138,25 @@ Evaluation is to evaluate the performance of a trained model. This sample provid
 
 - Faster RCNN
 
+    ```
     python eval_coco_map.py \
         --dataset=coco2017 \
-        --pretrained_model=${path_to_pretrain_model} \
+        --pretrained_model=${path_to_trained_model} \
         --MASK_ON=False
+    ```
 
 - Mask RCNN
 
+    ```
     python eval_coco_map.py \
         --dataset=coco2017 \
-        --pretrained_model=${path_to_pretrain_model} \
+        --pretrained_model=${path_to_trainde_model} \
         --MASK_ON=True
+    ```
 
-- Set ```export CUDA_VISIBLE_DEVICES=0``` to specifiy one GPU to eval.
-- Set ```MASK_ON``` to choose Faster RCNN or Mask RCNN model.
+    - Set ```--pretrained_model=${path_to_trained_model}``` to specifiy the trained model, not the initialized model.
+    - Set ```export CUDA_VISIBLE_DEVICES=0``` to specifiy one GPU to eval.
+    - Set ```MASK_ON``` to choose Faster RCNN or Mask RCNN model.
 
 Evalutaion result is shown as below:
 
@@ -159,12 +186,14 @@ Mask RCNN:
 
 Inference is used to get prediction score or image features based on trained models. `infer.py`  is the main executor for inference, one can start infer step by:
 
-    python infer.py \
-       --dataset=coco2017 \
-        --pretrained_model=${path_to_pretrain_model}  \
-        --image_path=dataset/coco/val2017/  \
-        --image_name=000000000139.jpg \
-        --draw_threshold=0.6
+```
+python infer.py \
+    --pretrained_model=${path_to_trained_model}  \
+    --image_path=dataset/coco/val2017/000000000139.jpg  \
+    --draw_threshold=0.6
+```
+
+Please set the model path and image path correctly. GPU device is used by default, you can set `--use_gpu=False` to switch to CPU device. And you can set `draw_threshold` to tune score threshold to control the number of output detection boxes.
 
 Visualization of infer result is shown as below:
 <p align="center">

fluid/PaddleCV/rcnn/README_cn.md

@@ -12,7 +12,7 @@
 
 ## 安装
 
-在当前目录下运行样例代码需要PadddlePaddle Fluid的v.1.0.0或以上的版本。如果你的运行环境中的PaddlePaddle低于此版本，请根据[安装文档](http://www.paddlepaddle.org/documentation/docs/zh/0.15.0/beginners_guide/install/install_doc.html#paddlepaddle)中的说明来更新PaddlePaddle。
+在当前目录下运行样例代码需要PadddlePaddle Fluid的v.1.3.0或以上的版本。如果你的运行环境中的PaddlePaddle低于此版本，请根据[安装文档](http://www.paddlepaddle.org/)中的说明来更新PaddlePaddle。
 
 ## 简介
 区域卷积神经网络（RCNN）系列模型为两阶段目标检测器。通过对图像生成候选区域，提取特征，判别特征类别并修正候选框位置。
@@ -37,6 +37,27 @@ Mask RCNN同样为两阶段框架，第一阶段扫描图像生成候选框；
     cd dataset/coco
     ./download.sh
 
+数据目录结构如下：
+
+```
+data/coco/
+├── annotations
+│   ├── instances_train2014.json
+│   ├── instances_train2017.json
+│   ├── instances_val2014.json
+│   ├── instances_val2017.json
+|   ...
+├── train2017
+│   ├── 000000000009.jpg
+│   ├── 000000580008.jpg
+|   ...
+├── val2017
+│   ├── 000000000139.jpg
+│   ├── 000000000285.jpg
+|   ...
+
+```
+
 ## 模型训练
 
 **下载预训练模型：** 本示例提供Resnet-50预训练模型，该模性转换自Caffe，并对批标准化层(Batch Normalization Layer)进行参数融合。采用如下命令下载预训练模型：
@@ -50,9 +71,8 @@ Mask RCNN同样为两阶段框架，第一阶段扫描图像生成候选框；
 
 训练前需要首先下载[cocoapi](https://github.com/cocodataset/cocoapi)：
 
-    # COCOAPI=/path/to/clone/cocoapi
-    git clone https://github.com/cocodataset/cocoapi.git $COCOAPI
-    cd $COCOAPI/PythonAPI
+    git clone https://github.com/cocodataset/cocoapi.git
+    cd cocoapi/PythonAPI
     # if cython is not installed
     pip install Cython
     # Install into global site-packages
@@ -65,25 +85,29 @@ Mask RCNN同样为两阶段框架，第一阶段扫描图像生成候选框；
 
 - Faster RCNN
 
+    ```
     python train.py \
        --model_save_dir=output/ \
        --pretrained_model=${path_to_pretrain_model} \
        --data_dir=${path_to_data} \
        --MASK_ON=False
+    ```
 
 - Mask RCNN
 
+    ```
     python train.py \
        --model_save_dir=output/ \
        --pretrained_model=${path_to_pretrain_model} \
        --data_dir=${path_to_data} \
        --MASK_ON=True
+    ```
 
-- 通过设置export CUDA\_VISIBLE\_DEVICES=0,1,2,3,4,5,6,7指定8卡GPU训练。
-- 通过设置```MASK_ON```选择Faster RCNN和Mask RCNN模型。
-- 可选参数见：
+    - 通过设置export CUDA\_VISIBLE\_DEVICES=0,1,2,3,4,5,6,7指定8卡GPU训练。
+    - 通过设置```MASK_ON```选择Faster RCNN和Mask RCNN模型。
+    - 可选参数见：
 
-    python train.py --help
+        python train.py --help
 
 **数据读取器说明：** 数据读取器定义在reader.py中。所有图像将短边等比例缩放至`scales`，若长边大于`max_size`, 则再次将长边等比例缩放至`max_size`。在训练阶段，对图像采用水平翻转。支持将同一个batch内的图像padding为相同尺寸。
 
@@ -110,20 +134,25 @@ Mask RCNN同样为两阶段框架，第一阶段扫描图像生成候选框；
 
 - Faster RCNN
 
+    ```
     python eval_coco_map.py \
         --dataset=coco2017 \
-        --pretrained_model=${path_to_pretrain_model} \
+        --pretrained_model=${path_to_trained_model} \
         --MASK_ON=False
+    ```
 
 - Mask RCNN
 
+    ```
     python eval_coco_map.py \
         --dataset=coco2017 \
-        --pretrained_model=${path_to_pretrain_model} \
+        --pretrained_model=${path_to_trained_model} \
         --MASK_ON=True
+    ```
 
-- 通过设置export CUDA\_VISIBLE\_DEVICES=0指定单卡GPU评估。
-- 通过设置```MASK_ON```选择Faster RCNN和Mask RCNN模型。
+    - 通过设置`--pretrained_model=${path_to_trained_model}`指定训练好的模型，注意不是初始化的模型。
+    - 通过设置`export CUDA\_VISIBLE\_DEVICES=0`指定单卡GPU评估。
+    - 通过设置```MASK_ON```选择Faster RCNN和Mask RCNN模型。
 
 下表为模型评估结果：
 
@@ -155,12 +184,14 @@ Mask RCNN:
 
 模型推断可以获取图像中的物体及其对应的类别，`infer.py`是主要执行程序，调用示例如下：
 
-    python infer.py \
-       --dataset=coco2017 \
-        --pretrained_model=${path_to_pretrain_model}  \
-        --image_path=dataset/coco/val2017/  \
-        --image_name=000000000139.jpg \
-        --draw_threshold=0.6
+```
+python infer.py \
+    --pretrained_model=${path_to_trained_model}  \
+    --image_path=dataset/coco/val2017/000000000139.jpg  \
+    --draw_threshold=0.6
+```
+
+注意，请正确设置模型路径`${path_to_trained_model}`和预测图片路径。默认使用GPU设备，也可通过设置`--use_gpu=False`使用CPU设备。可通过设置`draw_threshold`调节得分阈值控制检测框的个数。
 
 下图为模型可视化预测结果：
 <p align="center">

fluid/PaddleCV/rcnn/data_utils.py

@@ -31,6 +31,23 @@
 import os
 
 
+class DatasetPath(object):
+    def __init__(self, mode):
+        self.mode = mode
+        mode_name = 'train' if mode == 'train' else 'val'
+        if cfg.dataset != 'coco2014' and cfg.dataset != 'coco2017':
+            raise NotImplementedError('Dataset {} not supported'.format(
+                cfg.dataset))
+        self.sub_name = mode_name + cfg.dataset[-4:]
+
+    def get_data_dir(self):
+        return os.path.join(cfg.data_dir, self.sub_name)
+
+    def get_file_list(self):
+        sfile_list = 'annotations/instances_' + self.sub_name + '.json'
+        return os.path.join(cfg.data_dir, sfile_list)
+
+
 def get_image_blob(roidb, mode):
     """Builds an input blob from the images in the roidb at the specified
     scales.

fluid/PaddleCV/rcnn/eval_coco_map.py

@@ -29,7 +29,7 @@
 from pycocotools.coco import COCO
 from pycocotools.cocoeval import COCOeval, Params
 from config import cfg
-from roidbs import DatasetPath
+from data_utils import DatasetPath
 
 
 def eval():