update ce (#4106)

* modify ce for deeplabv3+

modify ce for auto_dialogue_evaluation

add ce for human_pose_estimation

* modify gru4rec ce

PaddleCV/deeplabv3+/reader.py

@@ -28,6 +28,10 @@ default_config = {
     "crop_size": 769,
 }
 
+# used for ce
+if 'ce_mode' in os.environ:
+    np.random.seed(0)
+
 
 def slice_with_pad(a, s, value=0):
     pads = []

PaddleCV/deeplabv3+/train.py

@@ -145,12 +145,6 @@ deeplabv3p = models.deeplabv3p
 sp = fluid.Program()
 tp = fluid.Program()
 
-# only for ce
-if args.enable_ce:
-    SEED = 102
-    sp.random_seed = SEED
-    tp.random_seed = SEED
-
 crop_size = args.train_crop_size
 batch_size = args.batch_size
 image_shape = [crop_size, crop_size]
@@ -162,6 +156,13 @@ weight_decay = 0.00004
 base_lr = args.base_lr
 total_step = args.total_step
 
+# only for ce
+if args.enable_ce:
+    SEED = 102
+    sp.random_seed = SEED
+    tp.random_seed = SEED
+    reader.default_config['shuffle'] = False
+
 with fluid.program_guard(tp, sp):
     if args.use_py_reader:
         batch_size_each = batch_size // utility.get_device_count()

PaddleCV/human_pose_estimation/lib/coco_reader.py

@@ -12,7 +12,6 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 ##############################################################################
-
 """Data reader for COCO dataset."""
 
 from __future__ import absolute_import
@@ -60,6 +59,7 @@ from pycocotools.coco import COCO
 #   [7,9],[8,10],[9,11],[2,3],[1,2],[1,3],[2,4],[3,5],[4,6],[5,7]
 # ]
 
+
 class Config:
     """Configurations for COCO dataset.
     """
@@ -74,7 +74,8 @@ class Config:
     DATAROOT = 'data/coco'
     IMAGEDIR = 'images'
     NUM_JOINTS = 17
-    FLIP_PAIRS = [[1, 2], [3, 4], [5, 6], [7, 8], [9, 10], [11, 12], [13, 14], [15, 16]]
+    FLIP_PAIRS = [[1, 2], [3, 4], [5, 6], [7, 8], [9, 10], [11, 12], [13, 14],
+                  [15, 16]]
     PARENT_IDS = None
 
     # CFGS
@@ -90,12 +91,15 @@ class Config:
     STD = [0.229, 0.224, 0.225]
     PIXEL_STD = 200
 
+
 cfg = Config()
 
+
 def _box2cs(box):
     x, y, w, h = box[:4]
     return _xywh2cs(x, y, w, h)
 
+
 def _xywh2cs(x, y, w, h):
     center = np.zeros((2), dtype=np.float32)
     center[0] = x + w * 0.5
@@ -106,21 +110,20 @@ def _xywh2cs(x, y, w, h):
     elif w < cfg.ASPECT_RATIO * h:
         w = h * cfg.ASPECT_RATIO
     scale = np.array(
-        [w * 1.0 / cfg.PIXEL_STD, h * 1.0 / cfg.PIXEL_STD],
-        dtype=np.float32)
+        [w * 1.0 / cfg.PIXEL_STD, h * 1.0 / cfg.PIXEL_STD], dtype=np.float32)
     if center[0] != -1:
         scale = scale * 1.25
 
     return center, scale
 
+
 def _select_data(db):
     db_selected = []
     for rec in db:
         num_vis = 0
         joints_x = 0.0
         joints_y = 0.0
-        for joint, joint_vis in zip(
-                rec['joints_3d'], rec['joints_3d_vis']):
+        for joint, joint_vis in zip(rec['joints_3d'], rec['joints_3d_vis']):
             if joint_vis[0] <= 0:
                 continue
             num_vis += 1
@@ -135,8 +138,8 @@ def _select_data(db):
         area = rec['scale'][0] * rec['scale'][1] * (cfg.PIXEL_STD**2)
         joints_center = np.array([joints_x, joints_y])
         bbox_center = np.array(rec['center'])
-        diff_norm2 = np.linalg.norm((joints_center-bbox_center), 2)
-        ks = np.exp(-1.0*(diff_norm2**2) / ((0.2)**2*2.0*area))
+        diff_norm2 = np.linalg.norm((joints_center - bbox_center), 2)
+        ks = np.exp(-1.0 * (diff_norm2**2) / ((0.2)**2 * 2.0 * area))
 
         metric = (0.2 / 16) * num_vis + 0.45 - 0.2 / 16
         if ks > metric:
@@ -146,7 +149,9 @@ def _select_data(db):
     print('=> num selected db: {}'.format(len(db_selected)))
     return db_selected
 
-def _load_coco_keypoint_annotation(image_set_index, coco, _coco_ind_to_class_ind, image_set):
+
+def _load_coco_keypoint_annotation(image_set_index, coco,
+                                   _coco_ind_to_class_ind, image_set):
     """Ground truth bbox and keypoints.
     """
     print('generating coco gt_db...')
@@ -168,7 +173,7 @@ def _load_coco_keypoint_annotation(image_set_index, coco, _coco_ind_to_class_ind
             x2 = np.min((width - 1, x1 + np.max((0, w - 1))))
             y2 = np.min((height - 1, y1 + np.max((0, h - 1))))
             if obj['area'] > 0 and x2 >= x1 and y2 >= y1:
-                obj['clean_bbox'] = [x1, y1, x2-x1, y2-y1]
+                obj['clean_bbox'] = [x1, y1, x2 - x1, y2 - y1]
                 valid_objs.append(obj)
         objs = valid_objs
 
@@ -197,7 +202,8 @@ def _load_coco_keypoint_annotation(image_set_index, coco, _coco_ind_to_class_ind
 
             center, scale = _box2cs(obj['clean_bbox'][:4])
             rec.append({
-                'image': os.path.join(cfg.DATAROOT, cfg.IMAGEDIR, image_set+'2017', '%012d.jpg' % index),
+                'image': os.path.join(cfg.DATAROOT, cfg.IMAGEDIR,
+                                      image_set + '2017', '%012d.jpg' % index),
                 'center': center,
                 'scale': scale,
                 'joints_3d': joints_3d,
@@ -209,6 +215,7 @@ def _load_coco_keypoint_annotation(image_set_index, coco, _coco_ind_to_class_ind
         gt_db.extend(rec)
     return gt_db
 
+
 def data_augmentation(sample, is_train):
     image_file = sample['image']
     filename = sample['filename'] if 'filename' in sample else ''
@@ -220,13 +227,18 @@ def data_augmentation(sample, is_train):
     # imgnum = sample['imgnum'] if 'imgnum' in sample else ''
     r = 0
 
-    data_numpy = cv2.imread(
-        image_file, cv2.IMREAD_COLOR | cv2.IMREAD_IGNORE_ORIENTATION)
+    # used for ce
+    if 'ce_mode' in os.environ:
+        random.seed(0)
+        np.random.seed(0)
+
+    data_numpy = cv2.imread(image_file, cv2.IMREAD_COLOR |
+                            cv2.IMREAD_IGNORE_ORIENTATION)
 
     if is_train:
         sf = cfg.SCALE_FACTOR
         rf = cfg.ROT_FACTOR
-        s = s * np.clip(np.random.randn()*sf + 1, 1 - sf, 1 + sf)
+        s = s * np.clip(np.random.randn() * sf + 1, 1 - sf, 1 + sf)
         r = np.clip(np.random.randn()*rf, -rf*2, rf*2) \
                 if random.random() <= 0.6 else 0
 
@@ -239,8 +251,7 @@ def data_augmentation(sample, is_train):
     trans = get_affine_transform(c, s, r, cfg.IMAGE_SIZE)
     input = cv2.warpAffine(
         data_numpy,
-            trans,
-            (int(cfg.IMAGE_SIZE[0]), int(cfg.IMAGE_SIZE[1])),
+        trans, (int(cfg.IMAGE_SIZE[0]), int(cfg.IMAGE_SIZE[1])),
         flags=cv2.INTER_LINEAR)
 
     for i in range(cfg.NUM_JOINTS):
@@ -263,23 +274,30 @@ def data_augmentation(sample, is_train):
     else:
         return input, target, target_weight, c, s, score, image_file
 
-# Create a reader
-def _reader_creator(root, image_set, shuffle=False, is_train=False, use_gt_bbox=False):
 
+# Create a reader
+def _reader_creator(root,
+                    image_set,
+                    shuffle=False,
+                    is_train=False,
+                    use_gt_bbox=False):
     def reader():
         if image_set in ['train', 'val']:
-            file_name = os.path.join(root, 'annotations', 'person_keypoints_'+image_set+'2017.json')
+            file_name = os.path.join(
+                root, 'annotations',
+                'person_keypoints_' + image_set + '2017.json')
         elif image_set in ['test', 'test-dev']:
-            file_name = os.path.join(root, 'annotations', 'image_info_'+image_set+'2017.json')
+            file_name = os.path.join(root, 'annotations',
+                                     'image_info_' + image_set + '2017.json')
         else:
-            raise ValueError("The dataset '{}' is not supported".format(image_set))
+            raise ValueError("The dataset '{}' is not supported".format(
+                image_set))
 
         # Load annotations
         coco = COCO(file_name)
 
         # Deal with class names
-        cats = [cat['name']
-                for cat in coco.loadCats(coco.getCatIds())]
+        cats = [cat['name'] for cat in coco.loadCats(coco.getCatIds())]
         classes = ['__background__'] + cats
         print('=> classes: {}'.format(classes))
         num_classes = len(classes)
@@ -308,23 +326,40 @@ def _reader_creator(root, image_set, shuffle=False, is_train=False, use_gt_bbox=
     mapper = functools.partial(data_augmentation, is_train=is_train)
     return reader, mapper
 
+
 def train():
-    reader, mapper = _reader_creator(cfg.DATAROOT, 'train', shuffle=True, is_train=True)
+    reader, mapper = _reader_creator(
+        cfg.DATAROOT, 'train', shuffle=True, is_train=True)
+
+    # used for ce
+    if 'ce_mode' in os.environ:
+        reader, mapper = _reader_creator(
+            cfg.DATAROOT, 'train', shuffle=False, is_train=True)
+
     def pop():
         for i, x in enumerate(reader()):
             yield mapper(x)
+
     return pop
 
+
 def valid():
-    reader, mapper = _reader_creator(cfg.DATAROOT, 'val', shuffle=False, is_train=False, use_gt_bbox=True)
+    reader, mapper = _reader_creator(
+        cfg.DATAROOT, 'val', shuffle=False, is_train=False, use_gt_bbox=True)
+
     def pop():
         for i, x in enumerate(reader()):
             yield mapper(x)
+
     return pop
 
+
 def test():
-    reader, mapper = _reader_creator(cfg.DATAROOT, 'test', shuffle=False, is_train=False, use_gt_bbox=True)
+    reader, mapper = _reader_creator(
+        cfg.DATAROOT, 'test', shuffle=False, is_train=False, use_gt_bbox=True)
+
     def pop():
         for i, x in enumerate(reader()):
             yield mapper(x)
+
     return pop

PaddleCV/human_pose_estimation/train.py

@@ -12,7 +12,6 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 ##############################################################################
-
 """Functions for training."""
 
 import os
@@ -42,8 +41,10 @@ add_arg('pretrained_model', str,   "pretrained/resnet_50/115",   "Whether to use
 add_arg('checkpoint',       str,   None,                         "Whether to resume checkpoint.")
 add_arg('lr',               float, 0.001,                        "Set learning rate.")
 add_arg('lr_strategy',      str,   "piecewise_decay",            "Set the learning rate decay strategy.")
+add_arg('enable_ce',        bool,  False,                        "If set True, enable continuous evaluation job.")
 # yapf: enable
 
+
 def optimizer_setting(args, params):
     lr_drop_ratio = 0.1
 
@@ -95,20 +96,33 @@ def train(args):
         args.kp_dim = 16
         args.total_images = 22246
     else:
-        raise ValueError('The dataset {} is not supported yet.'.format(args.dataset))
+        raise ValueError('The dataset {} is not supported yet.'.format(
+            args.dataset))
 
     print_arguments(args)
 
     # Image and target
-    image = layers.data(name='image', shape=[3, IMAGE_SIZE[1], IMAGE_SIZE[0]], dtype='float32')
-    target = layers.data(name='target', shape=[args.kp_dim, HEATMAP_SIZE[1], HEATMAP_SIZE[0]], dtype='float32')
-    target_weight = layers.data(name='target_weight', shape=[args.kp_dim, 1], dtype='float32')
+    image = layers.data(
+        name='image', shape=[3, IMAGE_SIZE[1], IMAGE_SIZE[0]], dtype='float32')
+    target = layers.data(
+        name='target',
+        shape=[args.kp_dim, HEATMAP_SIZE[1], HEATMAP_SIZE[0]],
+        dtype='float32')
+    target_weight = layers.data(
+        name='target_weight', shape=[args.kp_dim, 1], dtype='float32')
+
+    # used for ce
+    if args.enable_ce:
+        fluid.default_startup_program().random_seed = 90
+        fluid.default_main_program().random_seed = 90
 
     # Build model
     model = pose_resnet.ResNet(layers=50, kps_num=args.kp_dim)
 
     # Output
-    loss, output = model.net(input=image, target=target, target_weight=target_weight)
+    loss, output = model.net(input=image,
+                             target=target,
+                             target_weight=target_weight)
 
     # Parameters from model and arguments
     params = {}
@@ -127,11 +141,13 @@ def train(args):
     exe = fluid.Executor(place)
     exe.run(fluid.default_startup_program())
 
-
     if args.pretrained_model:
+
         def if_exist(var):
-            exist_flag = os.path.exists(os.path.join(args.pretrained_model, var.name))
+            exist_flag = os.path.exists(
+                os.path.join(args.pretrained_model, var.name))
             return exist_flag
+
         fluid.io.load_vars(exe, args.pretrained_model, predicate=if_exist)
 
     if args.checkpoint is not None:
@@ -139,7 +155,8 @@ def train(args):
 
     # Dataloader
     train_reader = paddle.batch(reader.train(), batch_size=args.batch_size)
-    feeder = fluid.DataFeeder(place=place, feed_list=[image, target, target_weight])
+    feeder = fluid.DataFeeder(
+        place=place, feed_list=[image, target, target_weight])
 
     train_exe = fluid.ParallelExecutor(
         use_cuda=True if args.use_gpu else False, loss_name=loss.name)
@@ -147,7 +164,8 @@ def train(args):
 
     for pass_id in range(params["num_epochs"]):
         for batch_id, data in enumerate(train_reader()):
-            current_lr = np.array(paddle.fluid.global_scope().find_var('learning_rate').get_tensor())
+            current_lr = np.array(paddle.fluid.global_scope().find_var(
+                'learning_rate').get_tensor())
 
             input_image, loss, out_heatmaps = train_exe.run(
                 fetch_list, feed=feeder.feed(data))
@@ -155,21 +173,31 @@ def train(args):
             loss = np.mean(np.array(loss))
 
             print_immediately('Epoch [{:4d}/{:3d}] LR: {:.10f} '
-                  'Loss = {:.5f}'.format(
-                  batch_id, pass_id, current_lr[0], loss))
+                              'Loss = {:.5f}'.format(batch_id, pass_id,
+                                                     current_lr[0], loss))
 
             if batch_id % 10 == 0:
-                save_batch_heatmaps(input_image, out_heatmaps, file_name='visualization@train.jpg', normalize=True)
-
-        model_path = os.path.join(args.model_save_dir + '/' + 'simplebase-{}'.format(args.dataset),
+                save_batch_heatmaps(
+                    input_image,
+                    out_heatmaps,
+                    file_name='visualization@train.jpg',
+                    normalize=True)
+
+        model_path = os.path.join(
+            args.model_save_dir + '/' + 'simplebase-{}'.format(args.dataset),
             str(pass_id))
         if not os.path.isdir(model_path):
             os.makedirs(model_path)
         fluid.io.save_persistables(exe, model_path)
 
+    # used for ce
+    if args.enable_ce:
+        device_num = fluid.core.get_cuda_device_count() if args.use_gpu else 1
+        print("kpis\t{}_train_cost_card{}\t{:.5f}".format(args.dataset,
+                                                          device_num, loss))
+
 
 if __name__ == '__main__':
     args = parser.parse_args()
     check_cuda(args.use_gpu)
     train(args)
-

PaddleNLP/PaddleDialogue/auto_dialogue_evaluation/ade/reader.py

@@ -19,6 +19,7 @@ import sys
 import time
 import random
 import numpy as np
+import os
 
 import paddle
 import paddle.fluid as fluid
@@ -38,7 +39,7 @@ class DataProcessor(object):
         index = 0
         fr = io.open(self.data_file, 'r', encoding="utf8")
         for line in fr:
-            if index !=0 and index % 100 == 0: 
+            if index != 0 and index % 100 == 0:
                 print("processing data: %d" % index)
             index += 1
             examples.append(line.strip())
@@ -53,11 +54,7 @@ class DataProcessor(object):
         self.num_examples[phase] = count
         return self.num_examples[phase]
 
-    def data_generator(self,
-                       place,
-                       phase="train",
-                       shuffle=True,
-                       sample_pro=1):
+    def data_generator(self, place, phase="train", shuffle=True, sample_pro=1):
         """
         Generate data for train, dev or test.
 
@@ -67,6 +64,13 @@ class DataProcessor(object):
             sample_pro: sample data ratio
         """
         examples = self.get_examples()
+
+        # used for ce
+        if 'ce_mode' in os.environ:
+            np.random.seed(0)
+            random.seed(0)
+            shuffle = False
+
         if shuffle:
             np.random.shuffle(examples)
 
@@ -84,8 +88,10 @@ class DataProcessor(object):
                     print("please input data: context \t response \t label")
                     continue
 
-                context = [int(x) for x in tokens[0].split()[: self.max_seq_len]]
-                response = [int(x) for x in tokens[1].split()[: self.max_seq_len]]
+                context = [int(x) for x in tokens[0].split()[:self.max_seq_len]]
+                response = [
+                    int(x) for x in tokens[1].split()[:self.max_seq_len]
+                ]
                 label = [int(tokens[2])]
                 instance = (context, response, label)
 
@@ -127,4 +133,3 @@ class DataProcessor(object):
                 yield input_batch
 
         return wrapper
-

PaddleRec/gru4rec/utils.py

@@ -100,7 +100,7 @@ def prepare_data(file_dir,
                  is_train=True):
     """ prepare the English Pann Treebank (PTB) data """
     print("start constuct word dict")
-    if is_train:
+    if is_train and 'ce_mode' not in os.environ:
         vocab_size = get_vocab_size(vocab_path)
         reader = sort_batch(
             paddle.reader.shuffle(