update new reader for resnet, mobilenet; test=develop (#4685)

dygraph/mobilenet/README.md

@@ -2,6 +2,10 @@
 
 图像分类是计算机视觉的重要领域，它的目标是将图像分类到预定义的标签。CNN模型在图像分类领域取得了突破的成果，同时模型复杂度也在不断增加。MobileNet是一种小巧而高效CNN模型，本文介绍如何使PaddlePaddle的动态图MobileNet进行图像分类。
 
+**版本要求**
+该模型在develop版本必须使用develop版本或者paddle 2.0alpha的版本才能够运行
+如果期望在paddle 1.8 上执行，请切换到 release/1.8 分支
+
 **代码结构**
 
     ├── run_mul_v1.sh                 # 多卡训练启动脚本_v1
@@ -55,9 +59,9 @@
 **模型性能**
 
     Model          Top-1(单卡/4卡)    Top-5(单卡/4卡)    收敛时间(单卡/4卡)
-    
+
     MobileNetV1    0.707/0.711        0.897/0.899        116小时/30.9小时
-    
+
     MobileNetV2    0.708/0.724        0.899/0.906        227.8小时/60.8小时
 
 **参考论文**

dygraph/mobilenet/imagenet_dataset.py

+# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import os
+import cv2
+import math
+import random
+import numpy as np
+
+from paddle.incubate.hapi.datasets import DatasetFolder
+from paddle.incubate.hapi.vision.transforms import transforms
+from paddle import fluid
+
+
+class ImageNetDataset(DatasetFolder):
+    def __init__(self,
+                 path,
+                 mode='train',
+                 image_size=224,
+                 resize_short_size=256):
+        super(ImageNetDataset, self).__init__(path)
+        self.mode = mode
+
+        normalize = transforms.Normalize(
+            mean=[123.675, 116.28, 103.53], std=[58.395, 57.120, 57.375])
+        if self.mode == 'train':
+            self.transform = transforms.Compose([
+                transforms.RandomResizedCrop(image_size),
+                transforms.RandomHorizontalFlip(),
+                transforms.Permute(mode='CHW'), normalize
+            ])
+        else:
+            self.transform = transforms.Compose([
+                transforms.Resize(resize_short_size),
+                transforms.CenterCrop(image_size),
+                transforms.Permute(mode='CHW'), normalize
+            ])
+
+    def __getitem__(self, idx):
+        img_path, label = self.samples[idx]
+        img = cv2.imread(img_path).astype(np.float32)
+        label = np.array([label]).astype(np.int64)
+        return self.transform(img), label
+
+    def __len__(self):
+        return len(self.samples)

dygraph/mobilenet/train.py

@@ -31,6 +31,9 @@ from utils import *
 from mobilenet_v1 import *
 from mobilenet_v2 import *
 
+from imagenet_dataset import ImageNetDataset
+from paddle.io import DataLoader
+
 args = parse_args()
 if int(os.getenv("PADDLE_TRAINER_ID", 0)) == 0:
     print_arguments(args)
@@ -116,14 +119,31 @@ def train_mobilenet():
             optimizer.set_dict(opti_dict)
 
         # 3. reader
-        train_data_loader = utility.create_data_loader(is_train=True, args=args)
         test_data_loader = utility.create_data_loader(is_train=False, args=args)
         num_trainers = int(os.environ.get('PADDLE_TRAINERS_NUM', 1))
         imagenet_reader = reader.ImageNetReader(seed=0, place_num=place_num)
-        train_reader = imagenet_reader.train(settings=args)
-        test_reader = imagenet_reader.val(settings=args)
-        train_data_loader.set_sample_list_generator(train_reader, place)
-        test_data_loader.set_sample_list_generator(test_reader, place)
+
+        train_dataset = ImageNetDataset(
+            os.path.join(args.data_dir, "train"), mode='train')
+
+        train_data_loader = DataLoader(
+            train_dataset,
+            batch_size=args.batch_size,
+            places=place,
+            shuffle=True,
+            drop_last=True,
+            num_workers=10)
+
+        test_dataset = ImageNetDataset(
+            os.path.join(args.data_dir, "val"), mode='val')
+
+        test_data_loader = DataLoader(
+            test_dataset,
+            batch_size=args.batch_size,
+            places=place,
+            shuffle=True,
+            drop_last=True,
+            num_workers=1)
 
         # 4. train loop
         total_batch_num = 0  #this is for benchmark
@@ -184,7 +204,7 @@ def train_mobilenet():
                 total_sample += 1
                 batch_id += 1
                 t_last = time.time()
-               
+
                 # NOTE: used for benchmark
                 total_batch_num = total_batch_num + 1
 

dygraph/resnet/reader.py

+#copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
+#
+#Licensed under the Apache License, Version 2.0 (the "License");
+#you may not use this file except in compliance with the License.
+#You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+#Unless required by applicable law or agreed to in writing, software
+#distributed under the License is distributed on an "AS IS" BASIS,
+#WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+#See the License for the specific language governing permissions and
+#limitations under the License.
+
+import sys
+import os
+import math
+import random
+import functools
+import numpy as np
+import cv2
+
+import paddle
+from paddle import fluid
+from PIL import Image
+
+policy = None
+
+random.seed(0)
+np.random.seed(0)
+
+
+def random_crop(img, size, interpolation=Image.BILINEAR):
+    """random crop image
+        
+    Args:
+        img: image data
+        size: crop size
+
+    Returns:
+        random cropped image data
+    """
+    lower_scale = 0.08
+    lower_ratio = 3. / 4.
+    upper_ratio = 4. / 3.
+    scale = [lower_scale, 1.0]
+    ratio = [lower_ratio, upper_ratio]
+
+    aspect_ratio = math.sqrt(np.random.uniform(*ratio))
+    w = 1. * aspect_ratio
+    h = 1. / aspect_ratio
+
+    bound = min((float(img.shape[0]) / img.shape[1]) / (h**2),
+                (float(img.shape[1]) / img.shape[0]) / (w**2))
+
+    scale_max = min(scale[1], bound)
+    scale_min = min(scale[0], bound)
+
+    target_area = img.shape[0] * img.shape[1] * np.random.uniform(scale_min,
+                                                                  scale_max)
+    target_size = math.sqrt(target_area)
+    w = int(target_size * w)
+    h = int(target_size * h)
+
+    i = np.random.randint(0, img.shape[0] - h + 1)
+    j = np.random.randint(0, img.shape[1] - w + 1)
+    img = img[i:i + h, j:j + w, :]
+
+    if interpolation:
+        resized = cv2.resize(img, (size, size), interpolation=interpolation)
+    else:
+        resized = cv2.resize(img, (size, size))
+    return resized
+
+
+def resize_short(img, target_size, interpolation=None):
+    """resize image
+    
+    Args:
+        img: image data
+        target_size: resize short target size
+        interpolation: interpolation mode
+
+    Returns:
+        resized image data
+    """
+    percent = float(target_size) / min(img.shape[0], img.shape[1])
+    resized_width = int(round(img.shape[1] * percent))
+    resized_height = int(round(img.shape[0] * percent))
+    if interpolation:
+        resized = cv2.resize(
+            img, (resized_width, resized_height), interpolation=interpolation)
+    else:
+        resized = cv2.resize(img, (resized_width, resized_height))
+    return resized
+
+
+def crop_image(img, target_size, center):
+    """crop image 
+    
+    Args:
+        img: images data
+        target_size: crop target size
+        center: crop mode
+    
+    Returns:
+        img: cropped image data
+    """
+    height, width = img.shape[:2]
+    size = target_size
+    if center == True:
+        w_start = (width - size) // 2
+        h_start = (height - size) // 2
+    else:
+        w_start = np.random.randint(0, width - size + 1)
+        h_start = np.random.randint(0, height - size + 1)
+    w_end = w_start + size
+    h_end = h_start + size
+    img = img[h_start:h_end, w_start:w_end, :]
+    return img
+
+
+def process_image(sample, settings, mode):
+    """ process_image """
+
+    mean = [0.485, 0.456, 0.406]
+    std = [0.229, 0.224, 0.225]
+    crop_size = 224
+
+    img_path = sample[0]
+    img = cv2.imread(img_path)
+
+    if mode == 'train':
+        img = random_crop(img, crop_size)
+        if np.random.randint(0, 2) == 1:
+            img = img[:, ::-1, :]
+    else:
+        if crop_size > 0:
+            target_size = settings.resize_short_size
+            img = resize_short(
+                img, target_size, interpolation=settings.interpolation)
+            img = crop_image(img, target_size=crop_size, center=True)
+
+    img = img[:, :, ::-1]
+
+    img = img.astype('float32').transpose((2, 0, 1)) / 255
+    img_mean = np.array(mean).reshape((3, 1, 1))
+    img_std = np.array(std).reshape((3, 1, 1))
+    img -= img_mean
+    img /= img_std
+
+    if mode == 'train' or mode == 'val':
+        return (img, np.array([sample[1]]))
+    elif mode == 'test':
+        return (img, )
+
+
+def process_batch_data(input_data, settings, mode):
+    batch_data = []
+    for sample in input_data:
+        if os.path.isfile(sample[0]):
+            batch_data.append(process_image(sample, settings, mode))
+        else:
+            print("File not exist : %s" % sample[0])
+    return batch_data
+
+
+class ImageNetReader:
+    def __init__(self, seed=None):
+        self.shuffle_seed = seed
+
+    def set_shuffle_seed(self, seed):
+        assert isinstance(seed, int), "shuffle seed must be int"
+        self.shuffle_seed = seed
+
+    def _reader_creator(self,
+                        settings,
+                        file_list,
+                        mode,
+                        shuffle=False,
+                        data_dir=None):
+        if mode == 'test':
+            batch_size = 1
+        else:
+            batch_size = settings.batch_size
+
+        def reader():
+            def read_file_list():
+                with open(file_list) as flist:
+                    full_lines = [line.strip() for line in flist]
+                    if mode != "test" and len(full_lines) < settings.batch_size:
+                        print(
+                            "Warning: The number of the whole data ({}) is smaller than the batch_size ({}), and drop_last is turnning on, so nothing  will feed in program, Terminated now. Please reset batch_size to a smaller number or feed more data!"
+                            .format(len(full_lines), settings.batch_size))
+                        os._exit(1)
+                    if shuffle:
+                        assert self.shuffle_seed is not None, "multiprocess train, shuffle seed must be set!"
+                        np.random.RandomState(self.shuffle_seed).shuffle(
+                            full_lines)
+
+                batch_data = []
+                for line in full_lines:
+                    img_path, label = line.split()
+                    img_path = os.path.join(data_dir, img_path)
+                    batch_data.append([img_path, int(label)])
+                    if len(batch_data) == batch_size:
+                        if mode == 'train' or mode == 'val' or mde == 'test':
+                            yield batch_data
+                        batch_data = []
+
+            return read_file_list
+
+        data_reader = reader()
+        mapper = functools.partial(
+            process_batch_data, settings=settings, mode=mode)
+
+        ret = fluid.io.xmap_readers(
+            mapper,
+            data_reader,
+            settings.reader_thread,
+            settings.reader_buf_size,
+            order=False)
+
+        return ret
+
+    def train(self, settings):
+        """Create a reader for trainning
+
+        Args:
+            settings: arguments
+
+        Returns:
+            train reader
+        """
+        file_list = os.path.join(settings.data_dir, 'train_list.txt')
+        assert os.path.isfile(
+            file_list), "{} doesn't exist, please check data list path".format(
+                file_list)
+
+        reader = self._reader_creator(
+            settings,
+            file_list,
+            'train',
+            shuffle=True,
+            data_dir=settings.data_dir)
+
+        return reader
+
+    def val(self, settings):
+        """Create a reader for eval
+
+        Args:
+            settings: arguments
+
+        Returns:
+            eval reader
+        """
+        file_list = os.path.join(settings.data_dir, 'val_list.txt')
+
+        assert os.path.isfile(
+            file_list), "{} doesn't exist, please check data list path".format(
+                file_list)
+
+        return self._reader_creator(
+            settings,
+            file_list,
+            'val',
+            shuffle=False,
+            data_dir=settings.data_dir)
+
+    def test(self, settings):
+        """Create a reader for testing
+
+        Args:
+            settings: arguments
+
+        Returns:
+            test reader
+        """
+        file_list = os.path.join(settings.data_dir, 'val_list.txt')
+
+        assert os.path.isfile(
+            file_list), "{} doesn't exist, please check data list path".format(
+                file_list)
+        return self._reader_creator(
+            settings,
+            file_list,
+            'test',
+            shuffle=False,
+            data_dir=settings.data_dir)

dygraph/resnet/train.py

@@ -26,6 +26,7 @@ from paddle.fluid import framework
 import math
 import sys
 import time
+import reader
 
 IMAGENET1000 = 1281167
 base_lr = 0.1
@@ -46,9 +47,90 @@ def parse_args():
     parser.add_argument(
         "-b", "--batch_size", default=32, type=int, help="set epoch")
     parser.add_argument("--ce", action="store_true", help="run ce")
-   
+
     # NOTE:used in benchmark
-    parser.add_argument("--max_iter", default=0, type=int, help="the max iters to train, used in benchmark")
+    parser.add_argument(
+        "--max_iter",
+        default=0,
+        type=int,
+        help="the max iters to train, used in benchmark")
+    parser.add_argument(
+        "--class_dim",
+        default=102,
+        type=int,
+        help="the class number of flowers dataset")
+    parser.add_argument(
+        "--use_imagenet_data",
+        action="store_true",
+        help="Use imagenet dataset instead of the flowers dataset(small dataset)"
+    )
+    parser.add_argument(
+        '--data_dir',
+        default="./data/ILSVRC2012",
+        type=str,
+        help="The ImageNet dataset root directory.")
+    parser.add_argument(
+        '--lower_scale',
+        default=0.08,
+        type=float,
+        help="The value of lower_scale in ramdom_crop")
+    parser.add_argument(
+        '--lower_ratio',
+        default=3. / 4.,
+        type=float,
+        help="The value of lower_ratio in ramdom_crop")
+    parser.add_argument(
+        '--upper_ratio',
+        default=4. / 3.,
+        type=float,
+        help="The value of upper_ratio in ramdom_crop")
+    parser.add_argument(
+        '--resize_short_size',
+        default=256,
+        type=int,
+        help="The value of resize_short_size")
+    parser.add_argument(
+        '--crop_size', default=224, type=int, help="The value of crop size")
+    parser.add_argument(
+        '--use_mixup', default=False, type=bool, help="Whether to use mixup")
+    parser.add_argument(
+        '--mixup_alpha',
+        default=0.2,
+        type=float,
+        help="The value of mixup_alpha")
+    parser.add_argument(
+        '--reader_thread',
+        default=8,
+        type=int,
+        help="The number of multi thread reader")
+    parser.add_argument(
+        '--reader_buf_size',
+        default=16,
+        type=int,
+        help="The buf size of multi thread reader")
+    parser.add_argument(
+        '--interpolation',
+        default=None,
+        type=int,
+        help="The interpolation mode")
+    parser.add_argument(
+        '--use_aa',
+        default=False,
+        type=bool,
+        help="Whether to use auto augment")
+    parser.add_argument(
+        '--image_mean',
+        nargs='+',
+        type=float,
+        default=[0.485, 0.456, 0.406],
+        help="The mean of input image data")
+    parser.add_argument(
+        '--image_std',
+        nargs='+',
+        type=float,
+        default=[0.229, 0.224, 0.225],
+        help="The std of input image data")
+
     args = parser.parse_args()
     return args
 
@@ -288,30 +370,47 @@ def train_resnet():
         if args.use_data_parallel:
             strategy = fluid.dygraph.parallel.prepare_context()
 
-        resnet = ResNet()
+        resnet = ResNet(class_dim=args.class_dim)
         optimizer = optimizer_setting(parameter_list=resnet.parameters())
 
         if args.use_data_parallel:
             resnet = fluid.dygraph.parallel.DataParallel(resnet, strategy)
 
-        train_reader = paddle.batch(
-            reader_decorator(paddle.dataset.flowers.train(use_xmap=True)),
-            batch_size=batch_size,
-            drop_last=True)
+        if args.use_imagenet_data:
+            imagenet_reader = reader.ImageNetReader(0)
+            train_reader = imagenet_reader.train(settings=args)
+        else:
+            train_reader = paddle.batch(
+                reader_decorator(paddle.dataset.flowers.train(use_xmap=True)),
+                batch_size=batch_size,
+                drop_last=True)
 
         if args.use_data_parallel:
             train_reader = fluid.contrib.reader.distributed_batch_reader(
                 train_reader)
 
-        test_reader = paddle.batch(
-            reader_decorator(paddle.dataset.flowers.test(use_xmap=True)),
-            batch_size=batch_size,
-            drop_last=True)
-
-        train_loader = fluid.io.DataLoader.from_generator(capacity=10)
+        if args.use_imagenet_data:
+            test_reader = imagenet_reader.val(settings=args)
+        else:
+            test_reader = paddle.batch(
+                reader_decorator(paddle.dataset.flowers.test(use_xmap=True)),
+                batch_size=batch_size,
+                drop_last=True)
+
+        train_loader = fluid.io.DataLoader.from_generator(
+            capacity=32,
+            use_double_buffer=True,
+            iterable=True,
+            return_list=True,
+            use_multiprocess=True)
         train_loader.set_sample_list_generator(train_reader, places=place)
 
-        test_loader = fluid.io.DataLoader.from_generator(capacity=10)
+        test_loader = fluid.io.DataLoader.from_generator(
+            capacity=64,
+            use_double_buffer=True,
+            iterable=True,
+            return_list=True,
+            use_multiprocess=True)
         test_loader.set_sample_list_generator(test_reader, places=place)
 
         #file_name = './model/epoch_0.npz'
@@ -368,7 +467,7 @@ def train_resnet():
                 total_acc1 += acc_top1.numpy()
                 total_acc5 += acc_top5.numpy()
                 total_sample += 1
-                total_batch_num = total_batch_num + 1 #this is for benchmark
+                total_batch_num = total_batch_num + 1  #this is for benchmark
                 #print("epoch id: %d, batch step: %d, loss: %f" % (eop, batch_id, dy_out))
                 if batch_id % 10 == 0:
                     print( "epoch %d | batch step %d, loss %0.3f acc1 %0.3f acc5 %0.3f, batch cost: %.5f" % \