Enable ParallelExecutor in SSD-MobileNet and Refine code. (#832)

* Enable ParallelExecutor in SSD-MobileNet and Refine code.

* Change the pre-trained model URL.

* Follow comments.

fluid/object_detection/.gitignore

+./data/pascalvoc/VOCdevkit/
+data/pascalvoc/test.txt
+data/pascalvoc/trainval.txt
+pretrained/ssd_mobilenet_v1_coco.tar.gz
+pretrained/ssd_mobilenet_v1_coco
+pretrained/mobilenet_v1_imagenet.tar.gz
+pretrained/mobilenet_v1_imagenet
+log*

fluid/object_detection/data/prepare_voc_data.py → fluid/object_detection/data/pascalvoc/create_list.py

@@ -60,4 +60,5 @@ def prepare_filelist(devkit_dir, years, output_dir):
             ftest.write(item[0] + ' ' + item[1] + '\n')
 
 
-prepare_filelist(devkit_dir, years, '.')
+if __name__ == '__main__':
+    prepare_filelist(devkit_dir, years, '.')

fluid/object_detection/data/pascalvoc/download.sh

+DIR="$( cd "$(dirname "$0")" ; pwd -P )"
+cd "$DIR"
+
+# Download the data.
+echo "Downloading..."
+wget http://host.robots.ox.ac.uk/pascal/VOC/voc2012/VOCtrainval_11-May-2012.tar
+wget http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCtrainval_06-Nov-2007.tar
+wget http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCtest_06-Nov-2007.tar
+# Extract the data.
+echo "Extractint..."
+tar -xf VOCtrainval_11-May-2012.tar
+tar -xf VOCtrainval_06-Nov-2007.tar
+tar -xf VOCtest_06-Nov-2007.tar
+
+echo "Creating data lists..."
+python create_list.py

fluid/object_detection/load_model.py

-import paddle.v2 as paddle
-import paddle.fluid as fluid
-import numpy as np
-
-
-# From npy
-def load_vars():
-    vars = {}
-    name_map = {}
-    with open('./ssd_mobilenet_v1_coco/names.map', 'r') as map_file:
-        for param in map_file:
-            fd_name, tf_name = param.strip().split('\t')
-            name_map[fd_name] = tf_name
-
-    tf_vars = np.load(
-        './ssd_mobilenet_v1_coco/ssd_mobilenet_v1_coco_2017_11_17.npy').item()
-    for fd_name in name_map:
-        tf_name = name_map[fd_name]
-        tf_var = tf_vars[tf_name]
-        if len(tf_var.shape) == 4 and 'depthwise' in tf_name:
-            vars[fd_name] = np.transpose(tf_var, (2, 3, 0, 1))
-        elif len(tf_var.shape) == 4:
-            vars[fd_name] = np.transpose(tf_var, (3, 2, 0, 1))
-        else:
-            vars[fd_name] = tf_var
-
-    return vars
-
-
-def load_and_set_vars(place):
-    vars = load_vars()
-    for k, v in vars.items():
-        t = fluid.global_scope().find_var(k).get_tensor()
-        #print(np.array(t).shape, v.shape, k)
-        assert np.array(t).shape == v.shape
-        t.set(v, place)
-
-
-# From Paddle V1
-def load_paddlev1_vars(place):
-    vars = {}
-    name_map = {}
-    with open('./caffe2paddle/names.map', 'r') as map_file:
-        for param in map_file:
-            fd_name, tf_name = param.strip().split('\t')
-            name_map[fd_name] = tf_name
-
-    from operator import mul
-
-    def load(file_name, shape):
-        with open(file_name, 'rb') as f:
-            f.read(16)
-            arr = np.fromfile(f, dtype=np.float32)
-            #print(arr.size, reduce(mul, shape), file_name)
-            assert arr.size == reduce(mul, shape)
-            return arr.reshape(shape)
-
-    for fd_name in name_map:
-        v1_name = name_map[fd_name]
-        t = fluid.global_scope().find_var(fd_name).get_tensor()
-        shape = np.array(t).shape
-        v1_var = load('./caffe2paddle/' + v1_name, shape)
-        t.set(v1_var, place)
-
-
-if __name__ == "__main__":
-    load_vars()

fluid/object_detection/mobilenet_ssd.py

@@ -13,7 +13,7 @@ def conv_bn(input,
             num_groups=1,
             act='relu',
             use_cudnn=True):
-    parameter_attr = ParamAttr(initializer=MSRA())
+    parameter_attr = ParamAttr(learning_rate=0.1, initializer=MSRA())
     conv = fluid.layers.conv2d(
         input=input,
         num_filters=num_filters,
@@ -25,11 +25,9 @@ def conv_bn(input,
         use_cudnn=use_cudnn,
         param_attr=parameter_attr,
         bias_attr=False)
-    #parameter_attr = ParamAttr(learning_rate=0.1, initializer=MSRA())
-    #bias_attr = ParamAttr(learning_rate=0.2)
-    return fluid.layers.batch_norm(input=conv, act=act, epsilon=0.00001)
-    #param_attr=parameter_attr,
-    #bias_attr=bias_attr)
+    parameter_attr = ParamAttr(learning_rate=0.1, initializer=MSRA())
+    bias_attr = ParamAttr(learning_rate=0.2)
+    return fluid.layers.batch_norm(input=conv, act=act)
 
 
 def depthwise_separable(input, num_filters1, num_filters2, num_groups, stride,

fluid/object_detection/pretrained/download_coco.sh

+DIR="$( cd "$(dirname "$0")" ; pwd -P )"
+cd "$DIR"
+
+# Download the data.
+echo "Downloading..."
+wget http://paddlemodels.bj.bcebos.com/ssd_mobilenet_v1_coco.tar.gz
+echo "Extractint..."
+tar -xf ssd_mobilenet_v1_coco.tar.gz

fluid/object_detection/pretrained/download_imagenet.sh

+DIR="$( cd "$(dirname "$0")" ; pwd -P )"
+cd "$DIR"
+
+# Download the data.
+echo "Downloading..."
+wget http://paddlemodels.bj.bcebos.com/mobilenet_v1_imagenet.tar.gz
+echo "Extractint..."
+tar -xf mobilenet_v1_imagenet.tar.gz

fluid/object_detection/reader.py

@@ -23,10 +23,6 @@ import os
 import time
 import copy
 
-# cocoapi 
-from pycocotools.coco import COCO
-from pycocotools.cocoeval import COCOeval
-
 
 class Settings(object):
     def __init__(self, dataset, toy, data_dir, label_file, resize_h, resize_w,
@@ -101,6 +97,10 @@ class Settings(object):
 def _reader_creator(settings, file_list, mode, shuffle):
     def reader():
         if settings.dataset == 'coco':
+            # cocoapi 
+            from pycocotools.coco import COCO
+            from pycocotools.cocoeval import COCOeval
+
             coco = COCO(file_list)
             image_ids = coco.getImgIds()
             images = coco.loadImgs(image_ids)
@@ -295,6 +295,7 @@ def draw_bounding_box_on_image(image,
 
 
 def train(settings, file_list, shuffle=True):
+    file_list = os.path.join(settings.data_dir, file_list)
     if settings.dataset == 'coco':
         train_settings = copy.copy(settings)
         if '2014' in file_list:
@@ -302,13 +303,13 @@ def train(settings, file_list, shuffle=True):
         elif '2017' in file_list:
             sub_dir = "train2017"
         train_settings.data_dir = os.path.join(settings.data_dir, sub_dir)
-        file_list = os.path.join(settings.data_dir, file_list)
         return _reader_creator(train_settings, file_list, 'train', shuffle)
     elif settings.dataset == 'pascalvoc':
         return _reader_creator(settings, file_list, 'train', shuffle)
 
 
 def test(settings, file_list):
+    file_list = os.path.join(settings.data_dir, file_list)
     if settings.dataset == 'coco':
         test_settings = copy.copy(settings)
         if '2014' in file_list:
@@ -316,7 +317,6 @@ def test(settings, file_list):
         elif '2017' in file_list:
             sub_dir = "val2017"
         test_settings.data_dir = os.path.join(settings.data_dir, sub_dir)
-        file_list = os.path.join(settings.data_dir, file_list)
         return _reader_creator(test_settings, file_list, 'test', False)
     elif settings.dataset == 'pascalvoc':
         return _reader_creator(settings, file_list, 'test', False)

fluid/object_detection/train.py

@@ -12,46 +12,36 @@ import functools
 
 parser = argparse.ArgumentParser(description=__doc__)
 add_arg = functools.partial(add_arguments, argparser=parser)
-add_arg('learning_rate', float, 0.001, "Learning rate.")
-add_arg('batch_size', int, 32, "Minibatch size.")
-add_arg('num_passes', int, 25, "Epoch number.")
-add_arg('parallel', bool, True, "Whether use parallel training.")
-add_arg('use_gpu', bool, True, "Whether use GPU.")
-add_arg('data_dir', str, './data/COCO17', "Root path of data")
-add_arg('train_file_list', str, 'annotations/instances_train2017.json',
-        "train file list")
-add_arg('val_file_list', str, 'annotations/instances_val2017.json',
-        "vaild file list")
-add_arg('model_save_dir', str, 'model_COCO17', "where to save model")
-
-add_arg('dataset', str, 'coco', "coco or pascalvoc")
-add_arg(
-    'is_toy', int, 0,
-    "Is Toy for quick debug, 0 means using all data, while n means using only n sample"
-)
-add_arg('label_file', str, 'label_list',
-        "Lable file which lists all label name")
-add_arg('apply_distort', bool, True, "Whether apply distort")
-add_arg('apply_expand', bool, False, "Whether appley expand")
-add_arg('resize_h', int, 300, "resize image size")
-add_arg('resize_w', int, 300, "resize image size")
-add_arg('mean_value_B', float, 127.5,
-        "mean value which will be subtracted")  #123.68
-add_arg('mean_value_G', float, 127.5,
-        "mean value which will be subtracted")  #116.78
-add_arg('mean_value_R', float, 127.5,
-        "mean value which will be subtracted")  #103.94
-
-
-def train(args,
-          train_file_list,
-          val_file_list,
-          data_args,
-          learning_rate,
-          batch_size,
-          num_passes,
-          model_save_dir='model',
-          init_model_path=None):
+# yapf: disable
+add_arg('learning_rate',    float, 0.001,     "Learning rate.")
+add_arg('batch_size',       int,   32,        "Minibatch size.")
+add_arg('num_passes',       int,   25,        "Epoch number.")
+add_arg('parallel',         bool,  True,      "Whether use parallel training.")
+add_arg('use_gpu',          bool,  True,      "Whether use GPU.")
+add_arg('use_nccl',         bool,  False,     "Whether use NCCL.")
+add_arg('dataset',          str, 'pascalvoc', "coco or pascalvoc.")
+add_arg('model_save_dir',   str, 'model',     "The path to save model.")
+add_arg('pretrained_model', str, 'pretrained/ssd_mobilenet_v1_coco/', "The init model path.")
+add_arg('apply_distort',    bool, True,   "Whether apply distort")
+add_arg('apply_expand',     bool, False,  "Whether appley expand")
+add_arg('resize_h',         int,  300,    "resize image size")
+add_arg('resize_w',         int,  300,    "resize image size")
+add_arg('mean_value_B',     float, 127.5, "mean value which will be subtracted")  #123.68
+add_arg('mean_value_G',     float, 127.5, "mean value which will be subtracted")  #116.78
+add_arg('mean_value_R',     float, 127.5, "mean value which will be subtracted")  #103.94
+add_arg('is_toy',           int, 0, "Toy for quick debug, 0 means using all data, while n means using only n sample")
+# yapf: disable
+
+
+def parallel_do(args,
+                train_file_list,
+                val_file_list,
+                data_args,
+                learning_rate,
+                batch_size,
+                num_passes,
+                model_save_dir,
+                pretrained_model=None):
     image_shape = [3, data_args.resize_h, data_args.resize_w]
     if data_args.dataset == 'coco':
         num_classes = 81
@@ -68,7 +58,7 @@ def train(args,
 
     if args.parallel:
         places = fluid.layers.get_places()
-        pd = fluid.layers.ParallelDo(places)
+        pd = fluid.layers.ParallelDo(places, use_nccl=args.use_nccl)
         with pd.do():
             image_ = pd.read_input(image)
             gt_box_ = pd.read_input(gt_box)
@@ -125,8 +115,11 @@ def train(args,
     exe = fluid.Executor(place)
     exe.run(fluid.default_startup_program())
 
-    #load_model.load_and_set_vars(place)
-    load_model.load_paddlev1_vars(place)
+    if pretrained_model:
+        def if_exist(var):
+            return os.path.exists(os.path.join(pretrained_model, var.name))
+        fluid.io.load_vars(exe, pretrained_model, predicate=if_exist)
+
     train_reader = paddle.batch(
         reader.train(data_args, train_file_list), batch_size=batch_size)
     test_reader = paddle.batch(
@@ -151,7 +144,6 @@ def train(args,
         for batch_id, data in enumerate(train_reader()):
             prev_start_time = start_time
             start_time = time.time()
-            #print("Batch {} start at {:.2f}".format(batch_id, start_time))
             loss_v = exe.run(fluid.default_main_program(),
                              feed=feeder.feed(data),
                              fetch_list=[loss])
@@ -164,29 +156,149 @@ def train(args,
         if pass_id % 10 == 0 or pass_id == num_passes - 1:
             model_path = os.path.join(model_save_dir, str(pass_id))
             print 'save models to %s' % (model_path)
-            fluid.io.save_inference_model(model_path, ['image'], [nmsed_out],
-                                          exe)
+            fluid.io.save_persistables(exe, model_path)
+
+
+def parallel_exe(args,
+                 train_file_list,
+                 val_file_list,
+                 data_args,
+                 learning_rate,
+                 batch_size,
+                 num_passes,
+                 model_save_dir='model',
+                 pretrained_model=None):
+    image_shape = [3, data_args.resize_h, data_args.resize_w]
+    if data_args.dataset == 'coco':
+        num_classes = 81
+    elif data_args.dataset == 'pascalvoc':
+        num_classes = 21
+
+    image = fluid.layers.data(name='image', shape=image_shape, dtype='float32')
+    gt_box = fluid.layers.data(
+        name='gt_box', shape=[4], dtype='float32', lod_level=1)
+    gt_label = fluid.layers.data(
+        name='gt_label', shape=[1], dtype='int32', lod_level=1)
+    difficult = fluid.layers.data(
+        name='gt_difficult', shape=[1], dtype='int32', lod_level=1)
+
+    locs, confs, box, box_var = mobile_net(num_classes, image, image_shape)
+    nmsed_out = fluid.layers.detection_output(
+        locs, confs, box, box_var, nms_threshold=0.45)
+    loss = fluid.layers.ssd_loss(locs, confs, gt_box, gt_label, box,
+                                 box_var)
+    loss = fluid.layers.reduce_sum(loss)
+
+    test_program = fluid.default_main_program().clone(for_test=True)
+    with fluid.program_guard(test_program):
+        map_eval = fluid.evaluator.DetectionMAP(
+            nmsed_out,
+            gt_label,
+            gt_box,
+            difficult,
+            num_classes,
+            overlap_threshold=0.5,
+            evaluate_difficult=False,
+            ap_version='integral')
+
+    if data_args.dataset == 'coco':
+        # learning rate decay in 12, 19 pass, respectively
+        if '2014' in train_file_list:
+            boundaries = [82783 / batch_size * 12, 82783 / batch_size * 19]
+        elif '2017' in train_file_list:
+            boundaries = [118287 / batch_size * 12, 118287 / batch_size * 19]
+    elif data_args.dataset == 'pascalvoc':
+        boundaries = [40000, 60000]
+    values = [learning_rate, learning_rate * 0.5, learning_rate * 0.25]
+    optimizer = fluid.optimizer.RMSProp(
+        learning_rate=fluid.layers.piecewise_decay(boundaries, values),
+        regularization=fluid.regularizer.L2Decay(0.00005), )
+
+    optimizer.minimize(loss)
 
+    place = fluid.CUDAPlace(0) if args.use_gpu else fluid.CPUPlace()
+    exe = fluid.Executor(place)
+    exe.run(fluid.default_startup_program())
+
+    if pretrained_model:
+        def if_exist(var):
+            return os.path.exists(os.path.join(pretrained_model, var.name))
+        fluid.io.load_vars(exe, pretrained_model, predicate=if_exist)
+
+    train_exe = fluid.ParallelExecutor(use_cuda=args.use_gpu,
+                                       loss_name=loss.name)
+
+    train_reader = paddle.batch(
+        reader.train(data_args, train_file_list), batch_size=batch_size)
+    test_reader = paddle.batch(
+        reader.test(data_args, val_file_list), batch_size=batch_size)
+    feeder = fluid.DataFeeder(
+        place=place, feed_list=[image, gt_box, gt_label, difficult])
+
+    def test(pass_id):
+        _, accum_map = map_eval.get_map_var()
+        map_eval.reset(exe)
+        test_map = None
+        for _, data in enumerate(test_reader()):
+            test_map = exe.run(test_program,
+                               feed=feeder.feed(data),
+                               fetch_list=[accum_map])
+        print("Test {0}, map {1}".format(pass_id, test_map[0]))
+
+    for pass_id in range(num_passes):
+        start_time = time.time()
+        prev_start_time = start_time
+        end_time = 0
+        test(pass_id)
+        for batch_id, data in enumerate(train_reader()):
+            prev_start_time = start_time
+            start_time = time.time()
+            loss_v, = train_exe.run(fetch_list=[loss.name],
+                                   feed_dict=feeder.feed(data))
+            end_time = time.time()
+            loss_v = np.mean(np.array(loss_v))
+            if batch_id % 20 == 0:
+                print("Pass {0}, batch {1}, loss {2}, time {3}".format(
+                    pass_id, batch_id, loss_v, start_time - prev_start_time))
+
+        if pass_id % 10 == 0 or pass_id == num_passes - 1:
+            model_path = os.path.join(model_save_dir, str(pass_id))
+            print 'save models to %s' % (model_path)
+            fluid.io.save_persistables(exe, model_path)
 
 if __name__ == '__main__':
     args = parser.parse_args()
     print_arguments(args)
+
+    data_dir = 'data/pascalvoc'
+    train_file_list = 'trainval.txt'
+    val_file_list = 'test.txt'
+    label_file = 'label_list'
+    model_save_dir = args.model_save_dir
+    if args.dataset == 'coco':
+        data_dir = './data/COCO17'
+        train_file_list = 'annotations/instances_train2017.json'
+        val_file_list = 'annotations/instances_val2017.json'
+        label_file = 'label_list'
+
     data_args = reader.Settings(
-        dataset=args.dataset,  # coco or pascalvoc
+        dataset=args.dataset,
         toy=args.is_toy,
-        data_dir=args.data_dir,
-        label_file=args.label_file,
+        data_dir=data_dir,
+        label_file=label_file,
         apply_distort=args.apply_distort,
         apply_expand=args.apply_expand,
         resize_h=args.resize_h,
         resize_w=args.resize_w,
         mean_value=[args.mean_value_B, args.mean_value_G, args.mean_value_R])
-    train(
-        args,
-        train_file_list=args.train_file_list,
-        val_file_list=args.val_file_list,
-        data_args=data_args,
-        learning_rate=args.learning_rate,
-        batch_size=args.batch_size,
-        num_passes=args.num_passes,
-        model_save_dir=args.model_save_dir)
+    #method = parallel_do
+    method = parallel_exe
+    method(args,
+           train_file_list=train_file_list,
+           val_file_list=val_file_list,
+           data_args=data_args,
+           learning_rate=args.learning_rate,
+           batch_size=args.batch_size,
+           num_passes=args.num_passes,
+           model_save_dir=model_save_dir,
+           pretrained_model=args.pretrained_model)