Merge branch 'master' of https://github.com/PaddlePaddle/hapi into sequence_tagging

cyclegan/README.md

+# Cycle GAN
+---
+## 内容
+
+- [安装](#安装)
+- [简介](#简介)
+- [代码结构](#代码结构)
+- [数据准备](#数据准备)
+- [模型训练与预测](#模型训练与预测)
+
+## 安装
+
+运行本目录下的程序示例需要使用PaddlePaddle develop最新版本。如果您的PaddlePaddle安装版本低于此要求，请按照[安装文档](http://www.paddlepaddle.org/docs/develop/documentation/zh/build_and_install/pip_install_cn.html)中的说明更新PaddlePaddle安装版本。
+
+## 简介
+Cycle GAN 是一种image to image 的图像生成网络，实现了非对称图像数据集的生成和风格迁移。模型结构如下图所示，我们的模型包含两个生成网络 G: X → Y 和 F: Y → X，以及相关的判别器 DY 和 DX 。通过训练DY,使G将X图尽量转换为Y图，反之亦然。同时引入两个“周期一致性损失”，它们保证:如果我们从一个领域转换到另一个领域，它还可以被转换回去:(b)正向循环一致性损失:x→G(x)→F(G(x))≈x， (c)反向循环一致性损失:y→F(y)→G(F(y))≈y
+
+<p align="center">
+<img src="image/net.png" hspace='10'/> <br />
+图1.网络结构
+</p>
+
+
+## 代码结构
+```
+├── data.py      # 读取、处理数据。
+├── layers.py    # 封装定义基础的layers。
+├── cyclegan.py  # 定义基础生成网络和判别网络。
+├── train.py     # 训练脚本。
+└── infer.py     # 预测脚本。
+```
+
+
+## 数据准备
+
+CycleGAN 支持的数据集可以参考download.py中的`cycle_pix_dataset`，可以通过指定`python download.py --dataset xxx` 下载得到。
+
+由于版权问题，cityscapes 数据集无法通过脚本直接获得，需要从[官方](https://www.cityscapes-dataset.com/)下载数据，
+下载完之后执行`python prepare_cityscapes_dataset.py --gtFine_dir ./gtFine/ --leftImg8bit_dir ./leftImg8bit --output_dir ./data/cityscapes/`处理，
+将数据存放在`data/cityscapes`。
+
+数据下载处理完毕后，需要您将数据组织为以下路径结构：
+```
+data
+|-- cityscapes
+|   |-- testA
+|   |-- testB
+|   |-- trainA
+|   |-- trainB
+
+```
+
+然后运行txt生成脚本：`python generate_txt.py`，最终数据组织如下所示:
+```
+data
+|-- cityscapes
+|   |-- testA
+|   |-- testA.txt
+|   |-- testB
+|   |-- testB.txt
+|   |-- trainA
+|   |-- trainA.txt
+|   |-- trainB
+|   `-- trainB.txt
+
+```
+
+以上数据文件中，`data`文件夹需要放在训练脚本`train.py`同级目录下。`testA`为存放真实街景图片的文件夹，`testB`为存放语义分割图片的文件夹，`testA.txt`和`testB.txt`分别为测试图片路径列表文件，格式如下：
+
+```
+data/cityscapes/testA/234_A.jpg
+data/cityscapes/testA/292_A.jpg
+data/cityscapes/testA/412_A.jpg
+```
+
+训练数据组织方式与测试数据相同。
+
+
+## 模型训练与预测
+
+### 训练
+
+在GPU单卡上训练:
+
+```
+env CUDA_VISIBLE_DEVICES=0 python train.py
+```
+
+执行`python train.py --help`可查看更多使用方式和参数详细说明。
+
+图1为训练152轮的训练损失示意图，其中横坐标轴为训练轮数，纵轴为在训练集上的损失。其中，'g_loss','da_loss'和'db_loss'分别为生成器、判别器A和判别器B的训练损失。
+
+
+### 测试
+
+执行以下命令可以选择已保存的训练权重，对测试集进行测试，通过 `--epoch` 制定权重轮次：
+
+```
+env CUDA_VISIBLE_DEVICES=0 python test.py --init_model=checkpoint/199
+```
+生成结果在 `output/eval`中
+
+
+### 预测
+
+执行以下命令读取单张或多张图片进行预测：
+
+真实街景生成分割图像：
+
+```
+env CUDA_VISIBLE_DEVICES=0 python infer.py \
+    --init_model="./checkpoints/199" --input="./image/testA/123_A.jpg" \
+    --input_style=A
+```
+
+分割图像生成真实街景：
+
+```
+env CUDA_VISIBLE_DEVICES=0 python infer.py \
+    --init_model="checkpoints/199" --input="./image/testB/78_B.jpg" \
+    --input_style=B
+```
+生成结果在 `output/single`中
+
+训练180轮的模型预测效果如fakeA和fakeB所示：
+
+
+<p align="center">
+<img src="image/A2B.png" width="620" hspace='10'/> <br/>
+<strong>A2B</strong>
+</p>
+
+
+<p align="center">
+<img src="image/B2A.png" width="620" hspace='10'/> <br/>
+<strong>B2A</strong>
+</p>
+
+>在本文示例中，均可通过修改`CUDA_VISIBLE_DEVICES`改变使用的显卡号。

cyclegan/check.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.
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import sys
+
+import paddle.fluid as fluid
+
+__all__ = ['check_gpu', 'check_version']
+
+
+def check_gpu(use_gpu):
+    """
+     Log error and exit when set use_gpu=true in paddlepaddle
+     cpu version.
+     """
+    err = "Config use_gpu cannot be set as true while you are " \
+          "using paddlepaddle cpu version ! \nPlease try: \n" \
+          "\t1. Install paddlepaddle-gpu to run model on GPU \n" \
+          "\t2. Set use_gpu as false in config file to run " \
+          "model on CPU"
+
+    try:
+        if use_gpu and not fluid.is_compiled_with_cuda():
+            print(err)
+            sys.exit(1)
+    except Exception as e:
+        pass
+
+
+def check_version():
+    """
+    Log error and exit when the installed version of paddlepaddle is
+    not satisfied.
+    """
+    err = "PaddlePaddle version 1.6 or higher is required, " \
+          "or a suitable develop version is satisfied as well. \n" \
+          "Please make sure the version is good with your code." \
+
+    try:
+        fluid.require_version('1.7.0')
+    except Exception as e:
+        print(err)
+        sys.exit(1)

cyclegan/cyclegan.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.
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import numpy as np
+
+from layers import ConvBN, DeConvBN
+import paddle.fluid as fluid
+from model import Model, Loss
+
+
+class ResnetBlock(fluid.dygraph.Layer):
+    def __init__(self, dim, dropout=False):
+        super(ResnetBlock, self).__init__()
+        self.dropout = dropout
+        self.conv0 = ConvBN(dim, dim, 3, 1)
+        self.conv1 = ConvBN(dim, dim, 3, 1, act=None)
+
+    def forward(self, inputs):
+        out_res = fluid.layers.pad2d(inputs, [1, 1, 1, 1], mode="reflect")
+        out_res = self.conv0(out_res)
+        if self.dropout:
+            out_res = fluid.layers.dropout(out_res, dropout_prob=0.5)
+        out_res = fluid.layers.pad2d(out_res, [1, 1, 1, 1], mode="reflect")
+        out_res = self.conv1(out_res)
+        return out_res + inputs
+
+
+class ResnetGenerator(fluid.dygraph.Layer):
+    def __init__(self, input_channel, n_blocks=9, dropout=False):
+        super(ResnetGenerator, self).__init__()
+
+        self.conv0 = ConvBN(input_channel, 32, 7, 1)
+        self.conv1 = ConvBN(32, 64, 3, 2, padding=1)
+        self.conv2 = ConvBN(64, 128, 3, 2, padding=1)
+
+        dim = 128
+        self.resnet_blocks = []
+        for i in range(n_blocks):
+            block = self.add_sublayer("generator_%d" % (i + 1),
+                                      ResnetBlock(dim, dropout))
+            self.resnet_blocks.append(block)
+
+        self.deconv0 = DeConvBN(
+            dim, 32 * 2, 3, 2, padding=[1, 1], outpadding=[0, 1, 0, 1])
+        self.deconv1 = DeConvBN(
+            32 * 2, 32, 3, 2, padding=[1, 1], outpadding=[0, 1, 0, 1])
+
+        self.conv3 = ConvBN(
+            32, input_channel, 7, 1, norm=False, act=False, use_bias=True)
+
+    def forward(self, inputs):
+        pad_input = fluid.layers.pad2d(inputs, [3, 3, 3, 3], mode="reflect")
+        y = self.conv0(pad_input)
+        y = self.conv1(y)
+        y = self.conv2(y)
+        for resnet_block in self.resnet_blocks:
+            y = resnet_block(y)
+        y = self.deconv0(y)
+        y = self.deconv1(y)
+        y = fluid.layers.pad2d(y, [3, 3, 3, 3], mode="reflect")
+        y = self.conv3(y)
+        y = fluid.layers.tanh(y)
+        return y
+
+
+class NLayerDiscriminator(fluid.dygraph.Layer):
+    def __init__(self, input_channel, d_dims=64, d_nlayers=3):
+        super(NLayerDiscriminator, self).__init__()
+        self.conv0 = ConvBN(
+            input_channel,
+            d_dims,
+            4,
+            2,
+            1,
+            norm=False,
+            use_bias=True,
+            relufactor=0.2)
+
+        nf_mult, nf_mult_prev = 1, 1
+        self.conv_layers = []
+        for n in range(1, d_nlayers):
+            nf_mult_prev = nf_mult
+            nf_mult = min(2**n, 8)
+            conv = self.add_sublayer(
+                'discriminator_%d' % (n),
+                ConvBN(
+                    d_dims * nf_mult_prev,
+                    d_dims * nf_mult,
+                    4,
+                    2,
+                    1,
+                    relufactor=0.2))
+            self.conv_layers.append(conv)
+
+        nf_mult_prev = nf_mult
+        nf_mult = min(2**d_nlayers, 8)
+        self.conv4 = ConvBN(
+            d_dims * nf_mult_prev, d_dims * nf_mult, 4, 1, 1, relufactor=0.2)
+        self.conv5 = ConvBN(
+            d_dims * nf_mult,
+            1,
+            4,
+            1,
+            1,
+            norm=False,
+            act=None,
+            use_bias=True,
+            relufactor=0.2)
+
+    def forward(self, inputs):
+        y = self.conv0(inputs)
+        for conv in self.conv_layers:
+            y = conv(y)
+        y = self.conv4(y)
+        y = self.conv5(y)
+        return y
+
+
+class Generator(Model):
+    def __init__(self, input_channel=3):
+        super(Generator, self).__init__()
+        self.g = ResnetGenerator(input_channel)
+
+    def forward(self, input):
+        fake = self.g(input)
+        return fake
+
+
+class GeneratorCombine(Model):
+    def __init__(self, g_AB=None, g_BA=None, d_A=None, d_B=None,
+                 is_train=True):
+        super(GeneratorCombine, self).__init__()
+        self.g_AB = g_AB
+        self.g_BA = g_BA
+        self.is_train = is_train
+        if self.is_train:
+            self.d_A = d_A
+            self.d_B = d_B
+
+    def forward(self, input_A, input_B):
+        # Translate images to the other domain
+        fake_B = self.g_AB(input_A)
+        fake_A = self.g_BA(input_B)
+
+        # Translate images back to original domain
+        cyc_A = self.g_BA(fake_B)
+        cyc_B = self.g_AB(fake_A)
+        if not self.is_train:
+            return fake_A, fake_B, cyc_A, cyc_B
+
+        # Identity mapping of images
+        idt_A = self.g_AB(input_B)
+        idt_B = self.g_BA(input_A)
+
+        # Discriminators determines validity of translated images
+        # d_A(g_AB(A))
+        valid_A = self.d_A.d(fake_B)
+        # d_B(g_BA(A))
+        valid_B = self.d_B.d(fake_A)
+        return input_A, input_B, fake_A, fake_B, cyc_A, cyc_B, idt_A, idt_B, valid_A, valid_B
+
+
+class GLoss(Loss):
+    def __init__(self, lambda_A=10., lambda_B=10., lambda_identity=0.5):
+        super(GLoss, self).__init__()
+        self.lambda_A = lambda_A
+        self.lambda_B = lambda_B
+        self.lambda_identity = lambda_identity
+
+    def forward(self, outputs, labels=None):
+        input_A, input_B, fake_A, fake_B, cyc_A, cyc_B, idt_A, idt_B, valid_A, valid_B = outputs
+
+        def mse(a, b):
+            return fluid.layers.reduce_mean(fluid.layers.square(a - b))
+
+        def mae(a, b):  # L1Loss
+            return fluid.layers.reduce_mean(fluid.layers.abs(a - b))
+
+        g_A_loss = mse(valid_A, 1.)
+        g_B_loss = mse(valid_B, 1.)
+        g_loss = g_A_loss + g_B_loss
+
+        cyc_A_loss = mae(input_A, cyc_A) * self.lambda_A
+        cyc_B_loss = mae(input_B, cyc_B) * self.lambda_B
+        cyc_loss = cyc_A_loss + cyc_B_loss
+
+        idt_loss_A = mae(input_B, idt_A) * (self.lambda_B *
+                                            self.lambda_identity)
+        idt_loss_B = mae(input_A, idt_B) * (self.lambda_A *
+                                            self.lambda_identity)
+        idt_loss = idt_loss_A + idt_loss_B
+
+        loss = cyc_loss + g_loss + idt_loss
+        return loss
+
+
+class Discriminator(Model):
+    def __init__(self, input_channel=3):
+        super(Discriminator, self).__init__()
+        self.d = NLayerDiscriminator(input_channel)
+
+    def forward(self, real, fake):
+        pred_real = self.d(real)
+        pred_fake = self.d(fake)
+        return pred_real, pred_fake
+
+
+class DLoss(Loss):
+    def __init__(self):
+        super(DLoss, self).__init__()
+
+    def forward(self, inputs, labels=None):
+        pred_real, pred_fake = inputs
+        loss = fluid.layers.square(pred_fake) + fluid.layers.square(pred_real -
+                                                                    1.)
+        loss = fluid.layers.reduce_mean(loss / 2.0)
+        return loss

cyclegan/data.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.
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+import os
+import random
+import numpy as np
+from PIL import Image, ImageOps
+
+DATASET = "cityscapes"
+A_LIST_FILE = "./data/" + DATASET + "/trainA.txt"
+B_LIST_FILE = "./data/" + DATASET + "/trainB.txt"
+A_TEST_LIST_FILE = "./data/" + DATASET + "/testA.txt"
+B_TEST_LIST_FILE = "./data/" + DATASET + "/testB.txt"
+IMAGES_ROOT = "./data/" + DATASET + "/"
+
+import paddle.fluid as fluid
+
+
+class Cityscapes(fluid.io.Dataset):
+    def __init__(self, root_path, file_path, mode='train', return_name=False):
+        self.root_path = root_path
+        self.file_path = file_path
+        self.mode = mode
+        self.return_name = return_name
+        self.images = [root_path + l for l in open(file_path, 'r').readlines()]
+
+    def _train(self, image):
+        ## Resize
+        image = image.resize((286, 286), Image.BICUBIC)
+        ## RandomCrop
+        i = np.random.randint(0, 30)
+        j = np.random.randint(0, 30)
+        image = image.crop((i, j, i + 256, j + 256))
+        # RandomHorizontalFlip
+        if np.random.rand() > 0.5:
+            image = ImageOps.mirror(image)
+        return image
+
+    def __getitem__(self, idx):
+        f = self.images[idx].strip("\n\r\t ")
+        image = Image.open(f)
+        if self.mode == 'train':
+            image = self._train(image)
+        else:
+            image = image.resize((256, 256), Image.BICUBIC)
+        # ToTensor
+        image = np.array(image).transpose([2, 0, 1]).astype('float32')
+        image = image / 255.0
+        # Normalize, mean=[0.5,0.5,0.5], std=[0.5,0.5,0.5]
+        image = (image - 0.5) / 0.5
+        if self.return_name:
+            return [image], os.path.basename(f)
+        else:
+            return [image]
+
+    def __len__(self):
+        return len(self.images)
+
+
+def DataA(root=IMAGES_ROOT, fpath=A_LIST_FILE):
+    """
+    Reader of images with A style for training.
+    """
+    return Cityscapes(root, fpath)
+
+
+def DataB(root=IMAGES_ROOT, fpath=B_LIST_FILE):
+    """
+    Reader of images with B style for training.
+    """
+    return Cityscapes(root, fpath)
+
+
+def TestDataA(root=IMAGES_ROOT, fpath=A_TEST_LIST_FILE):
+    """
+    Reader of images with A style for training.
+    """
+    return Cityscapes(root, fpath, mode='test', return_name=True)
+
+
+def TestDataB(root=IMAGES_ROOT, fpath=B_TEST_LIST_FILE):
+    """
+    Reader of images with B style for training.
+    """
+    return Cityscapes(root, fpath, mode='test', return_name=True)
+
+
+class ImagePool(object):
+    def __init__(self, pool_size=50):
+        self.pool = []
+        self.count = 0
+        self.pool_size = pool_size
+
+    def get(self, image):
+        if self.count < self.pool_size:
+            self.pool.append(image)
+            self.count += 1
+            return image
+        else:
+            p = random.random()
+            if p > 0.5:
+                random_id = random.randint(0, self.pool_size - 1)
+                temp = self.pool[random_id]
+                self.pool[random_id] = image
+                return temp
+            else:
+                return image

cyclegan/infer.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.
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import os
+import glob
+import numpy as np
+import argparse
+
+from PIL import Image
+from scipy.misc import imsave
+
+import paddle.fluid as fluid
+from check import check_gpu, check_version
+
+from model import Model, Input, set_device
+from cyclegan import Generator, GeneratorCombine
+
+
+def main():
+    place = set_device(FLAGS.device)
+    fluid.enable_dygraph(place) if FLAGS.dynamic else None
+
+    # Generators
+    g_AB = Generator()
+    g_BA = Generator()
+    g = GeneratorCombine(g_AB, g_BA, is_train=False)
+
+    im_shape = [-1, 3, 256, 256]
+    input_A = Input(im_shape, 'float32', 'input_A')
+    input_B = Input(im_shape, 'float32', 'input_B')
+    g.prepare(inputs=[input_A, input_B])
+    g.load(FLAGS.init_model, skip_mismatch=True, reset_optimizer=True)
+
+    out_path = FLAGS.output + "/single"
+    if not os.path.exists(out_path):
+        os.makedirs(out_path)
+    for f in glob.glob(FLAGS.input):
+        image_name = os.path.basename(f)
+        image = Image.open(f).convert('RGB')
+        image = image.resize((256, 256), Image.BICUBIC)
+        image = np.array(image) / 127.5 - 1
+
+        image = image[:, :, 0:3].astype("float32")
+        data = image.transpose([2, 0, 1])[np.newaxis, :]
+
+        if FLAGS.input_style == "A":
+            _, fake, _, _ = g.test([data, data])
+
+        if FLAGS.input_style == "B":
+            fake, _, _, _ = g.test([data, data])
+
+        fake = np.squeeze(fake[0]).transpose([1, 2, 0])
+
+        opath = "{}/fake{}{}".format(out_path, FLAGS.input_style, image_name)
+        imsave(opath, ((fake + 1) * 127.5).astype(np.uint8))
+        print("transfer {} to {}".format(f, opath))
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser("CycleGAN inference")
+    parser.add_argument(
+        "-d", "--dynamic", action='store_false', help="Enable dygraph mode")
+    parser.add_argument(
+        "-p",
+        "--device",
+        type=str,
+        default='gpu',
+        help="device to use, gpu or cpu")
+    parser.add_argument(
+        "-i",
+        "--input",
+        type=str,
+        default='./image/testA/123_A.jpg',
+        help="input image")
+    parser.add_argument(
+        "-o",
+        '--output',
+        type=str,
+        default='output',
+        help="The test result to be saved to.")
+    parser.add_argument(
+        "-m",
+        "--init_model",
+        type=str,
+        default='checkpoint/199',
+        help="The init model file of directory.")
+    parser.add_argument(
+        "-s", "--input_style", type=str, default='A', help="A or B")
+    FLAGS = parser.parse_args()
+    print(FLAGS)
+    check_gpu(str.lower(FLAGS.device) == 'gpu')
+    check_version()
+    main()

cyclegan/layers.py

+# Copyright (c) 2019 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.
+
+from __future__ import division
+import paddle.fluid as fluid
+from paddle.fluid.dygraph.nn import Conv2D, Conv2DTranspose, BatchNorm
+
+# cudnn is not better when batch size is 1.
+use_cudnn = False
+import numpy as np
+
+
+class ConvBN(fluid.dygraph.Layer):
+    """docstring for Conv2D"""
+
+    def __init__(self,
+                 num_channels,
+                 num_filters,
+                 filter_size,
+                 stride=1,
+                 padding=0,
+                 stddev=0.02,
+                 norm=True,
+                 is_test=False,
+                 act='leaky_relu',
+                 relufactor=0.0,
+                 use_bias=False):
+        super(ConvBN, self).__init__()
+
+        pattr = fluid.ParamAttr(
+            initializer=fluid.initializer.NormalInitializer(
+                loc=0.0, scale=stddev))
+        self.conv = Conv2D(
+            num_channels=num_channels,
+            num_filters=num_filters,
+            filter_size=filter_size,
+            stride=stride,
+            padding=padding,
+            use_cudnn=use_cudnn,
+            param_attr=pattr,
+            bias_attr=use_bias)
+        if norm:
+            self.bn = BatchNorm(
+                num_filters,
+                param_attr=fluid.ParamAttr(
+                    initializer=fluid.initializer.NormalInitializer(1.0,
+                                                                    0.02)),
+                bias_attr=fluid.ParamAttr(
+                    initializer=fluid.initializer.Constant(0.0)),
+                is_test=False,
+                trainable_statistics=True)
+        self.relufactor = relufactor
+        self.norm = norm
+        self.act = act
+
+    def forward(self, inputs):
+        conv = self.conv(inputs)
+        if self.norm:
+            conv = self.bn(conv)
+
+        if self.act == 'leaky_relu':
+            conv = fluid.layers.leaky_relu(conv, alpha=self.relufactor)
+        elif self.act == 'relu':
+            conv = fluid.layers.relu(conv)
+        else:
+            conv = conv
+
+        return conv
+
+
+class DeConvBN(fluid.dygraph.Layer):
+    def __init__(self,
+                 num_channels,
+                 num_filters,
+                 filter_size,
+                 stride=1,
+                 padding=[0, 0],
+                 outpadding=[0, 0, 0, 0],
+                 stddev=0.02,
+                 act='leaky_relu',
+                 norm=True,
+                 is_test=False,
+                 relufactor=0.0,
+                 use_bias=False):
+        super(DeConvBN, self).__init__()
+
+        pattr = fluid.ParamAttr(
+            initializer=fluid.initializer.NormalInitializer(
+                loc=0.0, scale=stddev))
+        self._deconv = Conv2DTranspose(
+            num_channels,
+            num_filters,
+            filter_size=filter_size,
+            stride=stride,
+            padding=padding,
+            param_attr=pattr,
+            bias_attr=use_bias)
+        if norm:
+            self.bn = BatchNorm(
+                num_filters,
+                param_attr=fluid.ParamAttr(
+                    initializer=fluid.initializer.NormalInitializer(1.0,
+                                                                    0.02)),
+                bias_attr=fluid.ParamAttr(
+                    initializer=fluid.initializer.Constant(0.0)),
+                is_test=False,
+                trainable_statistics=True)
+        self.outpadding = outpadding
+        self.relufactor = relufactor
+        self.use_bias = use_bias
+        self.norm = norm
+        self.act = act
+
+    def forward(self, inputs):
+        conv = self._deconv(inputs)
+        conv = fluid.layers.pad2d(
+            conv, paddings=self.outpadding, mode='constant', pad_value=0.0)
+
+        if self.norm:
+            conv = self.bn(conv)
+
+        if self.act == 'leaky_relu':
+            conv = fluid.layers.leaky_relu(conv, alpha=self.relufactor)
+        elif self.act == 'relu':
+            conv = fluid.layers.relu(conv)
+        else:
+            conv = conv
+
+        return conv

cyclegan/test.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.
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import os
+import argparse
+import numpy as np
+from scipy.misc import imsave
+
+import paddle.fluid as fluid
+from check import check_gpu, check_version
+
+from model import Model, Input, set_device
+from cyclegan import Generator, GeneratorCombine
+import data as data
+
+
+def main():
+    place = set_device(FLAGS.device)
+    fluid.enable_dygraph(place) if FLAGS.dynamic else None
+
+    # Generators
+    g_AB = Generator()
+    g_BA = Generator()
+    g = GeneratorCombine(g_AB, g_BA, is_train=False)
+
+    im_shape = [-1, 3, 256, 256]
+    input_A = Input(im_shape, 'float32', 'input_A')
+    input_B = Input(im_shape, 'float32', 'input_B')
+    g.prepare(inputs=[input_A, input_B])
+    g.load(FLAGS.init_model, skip_mismatch=True, reset_optimizer=True)
+
+    if not os.path.exists(FLAGS.output):
+        os.makedirs(FLAGS.output)
+
+    test_data_A = data.TestDataA()
+    test_data_B = data.TestDataB()
+
+    for i in range(len(test_data_A)):
+        data_A, A_name = test_data_A[i]
+        data_B, B_name = test_data_B[i]
+        data_A = np.array(data_A).astype("float32")
+        data_B = np.array(data_B).astype("float32")
+
+        fake_A, fake_B, cyc_A, cyc_B = g.test([data_A, data_B])
+
+        datas = [fake_A, fake_B, cyc_A, cyc_B, data_A, data_B]
+        odatas = []
+        for o in datas:
+            d = np.squeeze(o[0]).transpose([1, 2, 0])
+            im = ((d + 1) * 127.5).astype(np.uint8)
+            odatas.append(im)
+        imsave(FLAGS.output + "/fakeA_" + B_name, odatas[0])
+        imsave(FLAGS.output + "/fakeB_" + A_name, odatas[1])
+        imsave(FLAGS.output + "/cycA_" + A_name, odatas[2])
+        imsave(FLAGS.output + "/cycB_" + B_name, odatas[3])
+        imsave(FLAGS.output + "/inputA_" + A_name, odatas[4])
+        imsave(FLAGS.output + "/inputB_" + B_name, odatas[5])
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser("CycleGAN test")
+    parser.add_argument(
+        "-d", "--dynamic", action='store_false', help="Enable dygraph mode")
+    parser.add_argument(
+        "-p",
+        "--device",
+        type=str,
+        default='gpu',
+        help="device to use, gpu or cpu")
+    parser.add_argument(
+        "-b", "--batch_size", default=1, type=int, help="batch size")
+    parser.add_argument(
+        "-o",
+        '--output',
+        type=str,
+        default='output/eval',
+        help="The test result to be saved to.")
+    parser.add_argument(
+        "-m",
+        "--init_model",
+        type=str,
+        default='checkpoint/199',
+        help="The init model file of directory.")
+    FLAGS = parser.parse_args()
+    print(FLAGS)
+    check_gpu(str.lower(FLAGS.device) == 'gpu')
+    check_version()
+    main()

cyclegan/train.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.
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import numpy as np
+import random
+import argparse
+import contextlib
+import time
+
+import paddle
+import paddle.fluid as fluid
+from check import check_gpu, check_version
+
+from model import Model, Input, set_device
+
+import data as data
+from cyclegan import Generator, Discriminator, GeneratorCombine, GLoss, DLoss
+
+step_per_epoch = 2974
+
+
+def opt(parameters):
+    lr_base = 0.0002
+    bounds = [100, 120, 140, 160, 180]
+    lr = [1., 0.8, 0.6, 0.4, 0.2, 0.1]
+    bounds = [i * step_per_epoch for i in bounds]
+    lr = [i * lr_base for i in lr]
+    optimizer = fluid.optimizer.Adam(
+        learning_rate=fluid.layers.piecewise_decay(
+            boundaries=bounds, values=lr),
+        parameter_list=parameters,
+        beta1=0.5)
+    return optimizer
+
+
+def main():
+    place = set_device(FLAGS.device)
+    fluid.enable_dygraph(place) if FLAGS.dynamic else None
+
+    # Generators
+    g_AB = Generator()
+    g_BA = Generator()
+
+    # Discriminators
+    d_A = Discriminator()
+    d_B = Discriminator()
+
+    g = GeneratorCombine(g_AB, g_BA, d_A, d_B)
+
+    da_params = d_A.parameters()
+    db_params = d_B.parameters()
+    g_params = g_AB.parameters() + g_BA.parameters()
+
+    da_optimizer = opt(da_params)
+    db_optimizer = opt(db_params)
+    g_optimizer = opt(g_params)
+
+    im_shape = [None, 3, 256, 256]
+    input_A = Input(im_shape, 'float32', 'input_A')
+    input_B = Input(im_shape, 'float32', 'input_B')
+    fake_A = Input(im_shape, 'float32', 'fake_A')
+    fake_B = Input(im_shape, 'float32', 'fake_B')
+
+    g_AB.prepare(inputs=[input_A])
+    g_BA.prepare(inputs=[input_B])
+
+    g.prepare(g_optimizer, GLoss(), inputs=[input_A, input_B])
+    d_A.prepare(da_optimizer, DLoss(), inputs=[input_B, fake_B])
+    d_B.prepare(db_optimizer, DLoss(), inputs=[input_A, fake_A])
+
+    if FLAGS.resume:
+        g.load(FLAGS.resume)
+
+    loader_A = fluid.io.DataLoader(
+        data.DataA(),
+        places=place,
+        shuffle=True,
+        return_list=True,
+        batch_size=FLAGS.batch_size)
+    loader_B = fluid.io.DataLoader(
+        data.DataB(),
+        places=place,
+        shuffle=True,
+        return_list=True,
+        batch_size=FLAGS.batch_size)
+
+    A_pool = data.ImagePool()
+    B_pool = data.ImagePool()
+
+    for epoch in range(FLAGS.epoch):
+        for i, (data_A, data_B) in enumerate(zip(loader_A, loader_B)):
+            data_A = data_A[0][0] if not FLAGS.dynamic else data_A[0]
+            data_B = data_B[0][0] if not FLAGS.dynamic else data_B[0]
+            start = time.time()
+
+            fake_B = g_AB.test(data_A)[0]
+            fake_A = g_BA.test(data_B)[0]
+            g_loss = g.train([data_A, data_B])[0]
+            fake_pb = B_pool.get(fake_B)
+            da_loss = d_A.train([data_B, fake_pb])[0]
+
+            fake_pa = A_pool.get(fake_A)
+            db_loss = d_B.train([data_A, fake_pa])[0]
+
+            t = time.time() - start
+            if i % 20 == 0:
+                print("epoch: {} | step: {:3d} | g_loss: {:.4f} | " \
+                      "da_loss: {:.4f} | db_loss: {:.4f} | s/step {:.4f}".
+                      format(epoch, i, g_loss[0], da_loss[0], db_loss[0], t))
+        g.save('{}/{}'.format(FLAGS.checkpoint_path, epoch))
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser("CycleGAN Training on Cityscapes")
+    parser.add_argument(
+        "-d", "--dynamic", action='store_false', help="Enable dygraph mode")
+    parser.add_argument(
+        "-p",
+        "--device",
+        type=str,
+        default='gpu',
+        help="device to use, gpu or cpu")
+    parser.add_argument(
+        "-e", "--epoch", default=200, type=int, help="Epoch number")
+    parser.add_argument(
+        "-b", "--batch_size", default=1, type=int, help="batch size")
+    parser.add_argument(
+        "-o",
+        "--checkpoint_path",
+        type=str,
+        default='checkpoint',
+        help="path to save checkpoint")
+    parser.add_argument(
+        "-r",
+        "--resume",
+        default=None,
+        type=str,
+        help="checkpoint path to resume")
+    FLAGS = parser.parse_args()
+    print(FLAGS)
+    check_gpu(str.lower(FLAGS.device) == 'gpu')
+    check_version()
+    main()

datasets/folder.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 sys
 import cv2
@@ -6,11 +20,11 @@ from paddle.fluid.io import Dataset
 
 
 def has_valid_extension(filename, extensions):
-    """Checks if a file is an allowed extension.
+    """Checks if a file is a vilid extension.
 
     Args:
-        filename (string): path to a file
-        extensions (tuple of strings): extensions to consider (lowercase)
+        filename (str): path to a file
+        extensions (tuple of str): extensions to consider (lowercase)
 
     Returns:
         bool: True if the filename ends with one of given extensions

image_classification/README.MD

@@ -30,6 +30,7 @@
 ```bash
 python -u main.py --arch resnet50 /path/to/imagenet -d
 ```
+-d 是使用动态模式训练，默认为静态图模式。
 
 ### 多卡训练
 执行如下命令进行训练
@@ -64,11 +65,28 @@ CUDA_VISIBLE_DEVICES=0,1,2,3 python -m paddle.distributed.launch main.py --arch
 * **output-dir**: 模型文件保存的文件夹，默认值：'output'
 * **num-workers**: dataloader的进程数，默认值：4
 * **resume**: 恢复训练的模型路径，默认值：None
-* **eval-only**: 仅仅进行预测，默认值：False
+* **eval-only**: 是否仅仅进行预测
+* **lr-scheduler**: 学习率衰减策略，默认值：piecewise
+* **milestones**: piecewise学习率衰减策略的边界，默认值：[30, 60, 80]
+* **weight-decay**: 模型权重正则化系数，默认值：1e-4
+* **momentum**: SGD优化器的动量，默认值：0.9
 
 
 ## 模型
 
 | 模型 | top1 acc | top5 acc |
 | --- | --- | --- |
-| ResNet50 | 76.28 | 93.04 | 
+| [ResNet50](https://paddle-hapi.bj.bcebos.com/models/resnet50.pdparams) | 76.28 | 93.04 |
+| [vgg16](https://paddle-hapi.bj.bcebos.com/models/vgg16.pdparams) | 71.84 | 90.71 | 
+| [mobilenet_v1](https://paddle-hapi.bj.bcebos.com/models/mobilenet_v1_x1.0.pdparams) | 71.25 | 89.92 | 
+| [mobilenet_v2](https://paddle-hapi.bj.bcebos.com/models/mobilenet_v2_x1.0.pdparams) | 72.27 | 90.66 | 
+
+上述模型的复现参数请参考scripts下的脚本。
+
+
+## 参考文献
+- ResNet: [Deep Residual Learning for Image Recognitio](https://arxiv.org/abs/1512.03385), Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun
+- MobileNetV1: [MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications](https://arxiv.org/abs/1704.04861), Andrew G. Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, Hartwig Adam
+- MobileNetV2: [MobileNetV2: Inverted Residuals and Linear Bottlenecks](https://arxiv.org/pdf/1801.04381v4.pdf), Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhmoginov, Liang-Chieh Chen
+- VGG: [Very Deep Convolutional Networks for Large-scale Image Recognition](https://arxiv.org/pdf/1409.1556), Karen Simonyan, Andrew Zisserman
+

image_classification/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

image_classification/main.py

-# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+# 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.
@@ -37,23 +37,36 @@ from paddle.fluid.io import BatchSampler, DataLoader
 
 def make_optimizer(step_per_epoch, parameter_list=None):
     base_lr = FLAGS.lr
-    momentum = 0.9
-    weight_decay = 1e-4
+    lr_scheduler = FLAGS.lr_scheduler
+    momentum = FLAGS.momentum
+    weight_decay = FLAGS.weight_decay
+
+    if lr_scheduler == 'piecewise':
+        milestones = FLAGS.milestones
+        boundaries = [step_per_epoch * e for e in milestones]
+        values = [base_lr * (0.1**i) for i in range(len(boundaries) + 1)]
+        learning_rate = fluid.layers.piecewise_decay(
+            boundaries=boundaries, values=values)
+    elif lr_scheduler == 'cosine':
+        learning_rate = fluid.layers.cosine_decay(base_lr, step_per_epoch,
+                                                  FLAGS.epoch)
+    else:
+        raise ValueError(
+            "Expected lr_scheduler in ['piecewise', 'cosine'], but got {}".
+            format(lr_scheduler))
 
-    boundaries = [step_per_epoch * e for e in [30, 60, 80]]
-    values = [base_lr * (0.1**i) for i in range(len(boundaries) + 1)]
-    learning_rate = fluid.layers.piecewise_decay(
-        boundaries=boundaries, values=values)
     learning_rate = fluid.layers.linear_lr_warmup(
         learning_rate=learning_rate,
         warmup_steps=5 * step_per_epoch,
         start_lr=0.,
         end_lr=base_lr)
+
     optimizer = fluid.optimizer.Momentum(
         learning_rate=learning_rate,
         momentum=momentum,
         regularization=fluid.regularizer.L2Decay(weight_decay),
         parameter_list=parameter_list)
+
     return optimizer
 
 
@@ -138,6 +151,20 @@ if __name__ == '__main__':
         help="checkpoint path to resume")
     parser.add_argument(
         "--eval-only", action='store_true', help="enable dygraph mode")
+    parser.add_argument(
+        "--lr-scheduler",
+        default='piecewise',
+        type=str,
+        help="learning rate scheduler")
+    parser.add_argument(
+        "--milestones",
+        nargs='+',
+        type=int,
+        default=[30, 60, 80],
+        help="piecewise decay milestones")
+    parser.add_argument(
+        "--weight-decay", default=1e-4, type=float, help="weight decay")
+    parser.add_argument("--momentum", default=0.9, type=float, help="momentum")
     FLAGS = parser.parse_args()
     assert FLAGS.data, "error: must provide data path"
     main()

model.py

@@ -42,6 +42,14 @@ __all__ = ['Model', 'Loss', 'CrossEntropy', 'Input', 'set_device']
 
 
 def set_device(device):
+    """
+    Args:
+        device (str): specify device type, 'cpu' or 'gpu'.
+        
+    Returns:
+        fluid.CUDAPlace or fluid.CPUPlace: Created GPU or CPU place.
+    """
+
     assert isinstance(device, six.string_types) and device.lower() in ['cpu', 'gpu'], \
     "Expected device in ['cpu', 'gpu'], but got {}".format(device)
 
@@ -114,9 +122,9 @@ class Loss(object):
     def forward(self, outputs, labels):
         raise NotImplementedError()
 
-    def __call__(self, outputs, labels):
+    def __call__(self, outputs, labels=None):
         labels = to_list(labels)
-        if in_dygraph_mode():
+        if in_dygraph_mode() and labels:
             labels = [to_variable(l) for l in labels]
         losses = to_list(self.forward(to_list(outputs), labels))
         if self.average:
@@ -853,8 +861,6 @@ class Model(fluid.dygraph.Layer):
             if not isinstance(inputs, (list, dict, Input)):
                 raise TypeError(
                     "'inputs' must be list or dict in static graph mode")
-            if loss_function and not isinstance(labels, (list, Input)):
-                raise TypeError("'labels' must be list in static graph mode")
 
         metrics = metrics or []
         for metric in to_list(metrics):
@@ -1084,7 +1090,11 @@ class Model(fluid.dygraph.Layer):
 
         return eval_result
 
-    def predict(self, test_data, batch_size=1, num_workers=0, stack_outputs=True):
+    def predict(self,
+                test_data,
+                batch_size=1,
+                num_workers=0,
+                stack_outputs=True):
         """
         FIXME: add more comments and usage
         Args:

models/__init__.py

@@ -13,13 +13,22 @@
 #limitations under the License.
 
 from . import resnet
+from . import vgg
+from . import mobilenetv1
+from . import mobilenetv2
 from . import darknet
 from . import yolov3
 
 from .resnet import *
+from .mobilenetv1 import *
+from .mobilenetv2 import *
+from .vgg import *
 from .darknet import *
 from .yolov3 import *
 
 __all__ = resnet.__all__ \
+        + vgg.__all__ \
+        + mobilenetv1.__all__ \
+        + mobilenetv2.__all__ \
         + darknet.__all__ \
         + yolov3.__all__

models/mobilenetv1.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 numpy as np
+import paddle
+import paddle.fluid as fluid
+from paddle.fluid.initializer import MSRA
+from paddle.fluid.param_attr import ParamAttr
+from paddle.fluid.dygraph.nn import Conv2D, Pool2D, BatchNorm, Linear
+
+from model import Model
+from .download import get_weights_path
+
+__all__ = ['MobileNetV1', 'mobilenet_v1']
+
+model_urls = {
+    'mobilenetv1_1.0':
+    ('https://paddle-hapi.bj.bcebos.com/models/mobilenet_v1_x1.0.pdparams',
+     'bf0d25cb0bed1114d9dac9384ce2b4a6')
+}
+
+
+class ConvBNLayer(fluid.dygraph.Layer):
+    def __init__(self,
+                 num_channels,
+                 filter_size,
+                 num_filters,
+                 stride,
+                 padding,
+                 channels=None,
+                 num_groups=1,
+                 act='relu',
+                 use_cudnn=True,
+                 name=None):
+        super(ConvBNLayer, self).__init__()
+
+        self._conv = Conv2D(
+            num_channels=num_channels,
+            num_filters=num_filters,
+            filter_size=filter_size,
+            stride=stride,
+            padding=padding,
+            groups=num_groups,
+            act=None,
+            use_cudnn=use_cudnn,
+            param_attr=ParamAttr(
+                initializer=MSRA(), name=self.full_name() + "_weights"),
+            bias_attr=False)
+
+        self._batch_norm = BatchNorm(
+            num_filters,
+            act=act,
+            param_attr=ParamAttr(name=self.full_name() + "_bn" + "_scale"),
+            bias_attr=ParamAttr(name=self.full_name() + "_bn" + "_offset"),
+            moving_mean_name=self.full_name() + "_bn" + '_mean',
+            moving_variance_name=self.full_name() + "_bn" + '_variance')
+
+    def forward(self, inputs):
+        y = self._conv(inputs)
+        y = self._batch_norm(y)
+        return y
+
+
+class DepthwiseSeparable(fluid.dygraph.Layer):
+    def __init__(self,
+                 num_channels,
+                 num_filters1,
+                 num_filters2,
+                 num_groups,
+                 stride,
+                 scale,
+                 name=None):
+        super(DepthwiseSeparable, self).__init__()
+
+        self._depthwise_conv = ConvBNLayer(
+            num_channels=num_channels,
+            num_filters=int(num_filters1 * scale),
+            filter_size=3,
+            stride=stride,
+            padding=1,
+            num_groups=int(num_groups * scale),
+            use_cudnn=False)
+
+        self._pointwise_conv = ConvBNLayer(
+            num_channels=int(num_filters1 * scale),
+            filter_size=1,
+            num_filters=int(num_filters2 * scale),
+            stride=1,
+            padding=0)
+
+    def forward(self, inputs):
+        y = self._depthwise_conv(inputs)
+        y = self._pointwise_conv(y)
+        return y
+
+
+class MobileNetV1(Model):
+    """MobileNetV1 model from
+    `"MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications" <https://arxiv.org/abs/1704.04861>`_.
+
+    Args:
+        scale (float): scale of channels in each layer. Default: 1.0.
+        class_dim (int): output dim of last fc layer. Default: 1000.
+    """
+
+    def __init__(self, scale=1.0, class_dim=1000):
+        super(MobileNetV1, self).__init__()
+        self.scale = scale
+        self.dwsl = []
+
+        self.conv1 = ConvBNLayer(
+            num_channels=3,
+            filter_size=3,
+            channels=3,
+            num_filters=int(32 * scale),
+            stride=2,
+            padding=1)
+
+        dws21 = self.add_sublayer(
+            sublayer=DepthwiseSeparable(
+                num_channels=int(32 * scale),
+                num_filters1=32,
+                num_filters2=64,
+                num_groups=32,
+                stride=1,
+                scale=scale),
+            name="conv2_1")
+        self.dwsl.append(dws21)
+
+        dws22 = self.add_sublayer(
+            sublayer=DepthwiseSeparable(
+                num_channels=int(64 * scale),
+                num_filters1=64,
+                num_filters2=128,
+                num_groups=64,
+                stride=2,
+                scale=scale),
+            name="conv2_2")
+        self.dwsl.append(dws22)
+
+        dws31 = self.add_sublayer(
+            sublayer=DepthwiseSeparable(
+                num_channels=int(128 * scale),
+                num_filters1=128,
+                num_filters2=128,
+                num_groups=128,
+                stride=1,
+                scale=scale),
+            name="conv3_1")
+        self.dwsl.append(dws31)
+
+        dws32 = self.add_sublayer(
+            sublayer=DepthwiseSeparable(
+                num_channels=int(128 * scale),
+                num_filters1=128,
+                num_filters2=256,
+                num_groups=128,
+                stride=2,
+                scale=scale),
+            name="conv3_2")
+        self.dwsl.append(dws32)
+
+        dws41 = self.add_sublayer(
+            sublayer=DepthwiseSeparable(
+                num_channels=int(256 * scale),
+                num_filters1=256,
+                num_filters2=256,
+                num_groups=256,
+                stride=1,
+                scale=scale),
+            name="conv4_1")
+        self.dwsl.append(dws41)
+
+        dws42 = self.add_sublayer(
+            sublayer=DepthwiseSeparable(
+                num_channels=int(256 * scale),
+                num_filters1=256,
+                num_filters2=512,
+                num_groups=256,
+                stride=2,
+                scale=scale),
+            name="conv4_2")
+        self.dwsl.append(dws42)
+
+        for i in range(5):
+            tmp = self.add_sublayer(
+                sublayer=DepthwiseSeparable(
+                    num_channels=int(512 * scale),
+                    num_filters1=512,
+                    num_filters2=512,
+                    num_groups=512,
+                    stride=1,
+                    scale=scale),
+                name="conv5_" + str(i + 1))
+            self.dwsl.append(tmp)
+
+        dws56 = self.add_sublayer(
+            sublayer=DepthwiseSeparable(
+                num_channels=int(512 * scale),
+                num_filters1=512,
+                num_filters2=1024,
+                num_groups=512,
+                stride=2,
+                scale=scale),
+            name="conv5_6")
+        self.dwsl.append(dws56)
+
+        dws6 = self.add_sublayer(
+            sublayer=DepthwiseSeparable(
+                num_channels=int(1024 * scale),
+                num_filters1=1024,
+                num_filters2=1024,
+                num_groups=1024,
+                stride=1,
+                scale=scale),
+            name="conv6")
+        self.dwsl.append(dws6)
+
+        self.pool2d_avg = Pool2D(pool_type='avg', global_pooling=True)
+
+        self.out = Linear(
+            int(1024 * scale),
+            class_dim,
+            act='softmax',
+            param_attr=ParamAttr(
+                initializer=MSRA(), name=self.full_name() + "fc7_weights"),
+            bias_attr=ParamAttr(name="fc7_offset"))
+
+    def forward(self, inputs):
+        y = self.conv1(inputs)
+        for dws in self.dwsl:
+            y = dws(y)
+        y = self.pool2d_avg(y)
+        y = fluid.layers.reshape(y, shape=[-1, 1024])
+        y = self.out(y)
+        return y
+
+
+def _mobilenet(arch, pretrained=False, **kwargs):
+    model = MobileNetV1(**kwargs)
+    if pretrained:
+        assert arch in model_urls, "{} model do not have a pretrained model now, you should set pretrained=False".format(
+            arch)
+        weight_path = get_weights_path(model_urls[arch][0],
+                                       model_urls[arch][1])
+        assert weight_path.endswith(
+            '.pdparams'), "suffix of weight must be .pdparams"
+        model.load(weight_path[:-9])
+
+    return model
+
+
+def mobilenet_v1(pretrained=False, scale=1.0):
+    model = _mobilenet('mobilenetv1_' + str(scale), pretrained, scale=scale)
+    return model

models/mobilenetv2.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 numpy as np
+import paddle
+import paddle.fluid as fluid
+from paddle.fluid.param_attr import ParamAttr
+from paddle.fluid.dygraph.nn import Conv2D, Pool2D, BatchNorm, Linear
+
+from model import Model
+from .download import get_weights_path
+
+__all__ = ['MobileNetV2', 'mobilenet_v2']
+
+model_urls = {
+    'mobilenetv2_1.0':
+    ('https://paddle-hapi.bj.bcebos.com/models/mobilenet_v2_x1.0.pdparams',
+     '8ff74f291f72533f2a7956a4efff9d88')
+}
+
+
+class ConvBNLayer(fluid.dygraph.Layer):
+    def __init__(self,
+                 num_channels,
+                 filter_size,
+                 num_filters,
+                 stride,
+                 padding,
+                 channels=None,
+                 num_groups=1,
+                 use_cudnn=True):
+        super(ConvBNLayer, self).__init__()
+
+        tmp_param = ParamAttr(name=self.full_name() + "_weights")
+        self._conv = Conv2D(
+            num_channels=num_channels,
+            num_filters=num_filters,
+            filter_size=filter_size,
+            stride=stride,
+            padding=padding,
+            groups=num_groups,
+            act=None,
+            use_cudnn=use_cudnn,
+            param_attr=tmp_param,
+            bias_attr=False)
+
+        self._batch_norm = BatchNorm(
+            num_filters,
+            param_attr=ParamAttr(name=self.full_name() + "_bn" + "_scale"),
+            bias_attr=ParamAttr(name=self.full_name() + "_bn" + "_offset"),
+            moving_mean_name=self.full_name() + "_bn" + '_mean',
+            moving_variance_name=self.full_name() + "_bn" + '_variance')
+
+    def forward(self, inputs, if_act=True):
+        y = self._conv(inputs)
+        y = self._batch_norm(y)
+        if if_act:
+            y = fluid.layers.relu6(y)
+        return y
+
+
+class InvertedResidualUnit(fluid.dygraph.Layer):
+    def __init__(
+            self,
+            num_channels,
+            num_in_filter,
+            num_filters,
+            stride,
+            filter_size,
+            padding,
+            expansion_factor, ):
+        super(InvertedResidualUnit, self).__init__()
+        num_expfilter = int(round(num_in_filter * expansion_factor))
+        self._expand_conv = ConvBNLayer(
+            num_channels=num_channels,
+            num_filters=num_expfilter,
+            filter_size=1,
+            stride=1,
+            padding=0,
+            num_groups=1)
+
+        self._bottleneck_conv = ConvBNLayer(
+            num_channels=num_expfilter,
+            num_filters=num_expfilter,
+            filter_size=filter_size,
+            stride=stride,
+            padding=padding,
+            num_groups=num_expfilter,
+            use_cudnn=False)
+
+        self._linear_conv = ConvBNLayer(
+            num_channels=num_expfilter,
+            num_filters=num_filters,
+            filter_size=1,
+            stride=1,
+            padding=0,
+            num_groups=1)
+
+    def forward(self, inputs, ifshortcut):
+        y = self._expand_conv(inputs, if_act=True)
+        y = self._bottleneck_conv(y, if_act=True)
+        y = self._linear_conv(y, if_act=False)
+        if ifshortcut:
+            y = fluid.layers.elementwise_add(inputs, y)
+        return y
+
+
+class InvresiBlocks(fluid.dygraph.Layer):
+    def __init__(self, in_c, t, c, n, s):
+        super(InvresiBlocks, self).__init__()
+
+        self._first_block = InvertedResidualUnit(
+            num_channels=in_c,
+            num_in_filter=in_c,
+            num_filters=c,
+            stride=s,
+            filter_size=3,
+            padding=1,
+            expansion_factor=t)
+
+        self._inv_blocks = []
+        for i in range(1, n):
+            tmp = self.add_sublayer(
+                sublayer=InvertedResidualUnit(
+                    num_channels=c,
+                    num_in_filter=c,
+                    num_filters=c,
+                    stride=1,
+                    filter_size=3,
+                    padding=1,
+                    expansion_factor=t),
+                name=self.full_name() + "_" + str(i + 1))
+            self._inv_blocks.append(tmp)
+
+    def forward(self, inputs):
+        y = self._first_block(inputs, ifshortcut=False)
+        for inv_block in self._inv_blocks:
+            y = inv_block(y, ifshortcut=True)
+        return y
+
+
+class MobileNetV2(Model):
+    """MobileNetV2 model from
+    `"MobileNetV2: Inverted Residuals and Linear Bottlenecks" <https://arxiv.org/abs/1801.04381>`_.
+
+    Args:
+        scale (float): scale of channels in each layer. Default: 1.0.
+        class_dim (int): output dim of last fc layer. Default: 1000.
+    """
+
+    def __init__(self, scale=1.0, class_dim=1000):
+        super(MobileNetV2, self).__init__()
+        self.scale = scale
+        self.class_dim = class_dim
+
+        bottleneck_params_list = [
+            (1, 16, 1, 1),
+            (6, 24, 2, 2),
+            (6, 32, 3, 2),
+            (6, 64, 4, 2),
+            (6, 96, 3, 1),
+            (6, 160, 3, 2),
+            (6, 320, 1, 1),
+        ]
+
+        #1. conv1 
+        self._conv1 = ConvBNLayer(
+            num_channels=3,
+            num_filters=int(32 * scale),
+            filter_size=3,
+            stride=2,
+            padding=1)
+
+        #2. bottleneck sequences
+        self._invl = []
+        i = 1
+        in_c = int(32 * scale)
+        for layer_setting in bottleneck_params_list:
+            t, c, n, s = layer_setting
+            i += 1
+            tmp = self.add_sublayer(
+                sublayer=InvresiBlocks(
+                    in_c=in_c, t=t, c=int(c * scale), n=n, s=s),
+                name='conv' + str(i))
+            self._invl.append(tmp)
+            in_c = int(c * scale)
+
+        #3. last_conv
+        self._out_c = int(1280 * scale) if scale > 1.0 else 1280
+        self._conv9 = ConvBNLayer(
+            num_channels=in_c,
+            num_filters=self._out_c,
+            filter_size=1,
+            stride=1,
+            padding=0)
+
+        #4. pool
+        self._pool2d_avg = Pool2D(pool_type='avg', global_pooling=True)
+
+        #5. fc
+        tmp_param = ParamAttr(name=self.full_name() + "fc10_weights")
+        self._fc = Linear(
+            self._out_c,
+            class_dim,
+            act='softmax',
+            param_attr=tmp_param,
+            bias_attr=ParamAttr(name="fc10_offset"))
+
+    def forward(self, inputs):
+        y = self._conv1(inputs, if_act=True)
+        for inv in self._invl:
+            y = inv(y)
+        y = self._conv9(y, if_act=True)
+        y = self._pool2d_avg(y)
+        y = fluid.layers.reshape(y, shape=[-1, self._out_c])
+        y = self._fc(y)
+        return y
+
+
+def _mobilenet(arch, pretrained=False, **kwargs):
+    model = MobileNetV2(**kwargs)
+    if pretrained:
+        assert arch in model_urls, "{} model do not have a pretrained model now, you should set pretrained=False".format(
+            arch)
+        weight_path = get_weights_path(model_urls[arch][0],
+                                       model_urls[arch][1])
+        assert weight_path.endswith(
+            '.pdparams'), "suffix of weight must be .pdparams"
+        model.load(weight_path[:-9])
+
+    return model
+
+
+def mobilenet_v2(pretrained=False, scale=1.0):
+    """MobileNetV2
+    
+    Args:
+        pretrained (bool): If True, returns a model pre-trained on ImageNet
+    """
+    model = _mobilenet('mobilenetv2_' + str(scale), pretrained, scale=scale)
+    return model

models/resnet.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.
+
 from __future__ import division
 from __future__ import print_function
 
@@ -11,7 +25,9 @@ from paddle.fluid.dygraph.container import Sequential
 from model import Model
 from .download import get_weights_path
 
-__all__ = ['ResNet', 'resnet50', 'resnet101', 'resnet152']
+__all__ = [
+    'ResNet', 'resnet18', 'resnet34', 'resnet50', 'resnet101', 'resnet152'
+]
 
 model_urls = {
     'resnet50': ('https://paddle-hapi.bj.bcebos.com/models/resnet50.pdparams',
@@ -48,7 +64,52 @@ class ConvBNLayer(fluid.dygraph.Layer):
         return x
 
 
+class BasicBlock(fluid.dygraph.Layer):
+
+    expansion = 1
+
+    def __init__(self, num_channels, num_filters, stride, shortcut=True):
+        super(BasicBlock, self).__init__()
+
+        self.conv0 = ConvBNLayer(
+            num_channels=num_channels,
+            num_filters=num_filters,
+            filter_size=3,
+            act='relu')
+        self.conv1 = ConvBNLayer(
+            num_channels=num_filters,
+            num_filters=num_filters,
+            filter_size=3,
+            stride=stride,
+            act='relu')
+
+        if not shortcut:
+            self.short = ConvBNLayer(
+                num_channels=num_channels,
+                num_filters=num_filters,
+                filter_size=1,
+                stride=stride)
+
+        self.shortcut = shortcut
+
+    def forward(self, inputs):
+        y = self.conv0(inputs)
+        conv1 = self.conv1(y)
+
+        if self.shortcut:
+            short = inputs
+        else:
+            short = self.short(inputs)
+
+        y = short + conv1
+
+        return fluid.layers.relu(y)
+
+
 class BottleneckBlock(fluid.dygraph.Layer):
+
+    expansion = 4
+
     def __init__(self, num_channels, num_filters, stride, shortcut=True):
         super(BottleneckBlock, self).__init__()
 
@@ -65,20 +126,20 @@ class BottleneckBlock(fluid.dygraph.Layer):
             act='relu')
         self.conv2 = ConvBNLayer(
             num_channels=num_filters,
-            num_filters=num_filters * 4,
+            num_filters=num_filters * self.expansion,
             filter_size=1,
             act=None)
 
         if not shortcut:
             self.short = ConvBNLayer(
                 num_channels=num_channels,
-                num_filters=num_filters * 4,
+                num_filters=num_filters * self.expansion,
                 filter_size=1,
                 stride=stride)
 
         self.shortcut = shortcut
 
-        self._num_channels_out = num_filters * 4
+        self._num_channels_out = num_filters * self.expansion
 
     def forward(self, inputs):
         x = self.conv0(inputs)
@@ -92,16 +153,25 @@ class BottleneckBlock(fluid.dygraph.Layer):
 
         x = fluid.layers.elementwise_add(x=short, y=conv2)
 
-        layer_helper = LayerHelper(self.full_name(), act='relu')
-        return layer_helper.append_activation(x)
-        # return fluid.layers.relu(x)
+        return fluid.layers.relu(x)
 
 
 class ResNet(Model):
+    """ResNet model from
+    `"Deep Residual Learning for Image Recognition" <https://arxiv.org/pdf/1512.03385.pdf>`_
+
+    Args:
+        Block (BasicBlock|BottleneckBlock): block module of model.
+        depth (int): layers of resnet, default: 50.
+        num_classes (int): output dim of last fc layer, default: 1000.
+    """
+
     def __init__(self, Block, depth=50, num_classes=1000):
         super(ResNet, self).__init__()
 
         layer_config = {
+            18: [2, 2, 2, 2],
+            34: [3, 4, 6, 3],
             50: [3, 4, 6, 3],
             101: [3, 4, 23, 3],
             152: [3, 8, 36, 3],
@@ -111,8 +181,9 @@ class ResNet(Model):
                 layer_config.keys(), depth)
 
         layers = layer_config[depth]
-        num_in = [64, 256, 512, 1024]
-        num_out = [64, 128, 256, 512]
+
+        in_channels = 64
+        out_channels = [64, 128, 256, 512]
 
         self.conv = ConvBNLayer(
             num_channels=3,
@@ -128,9 +199,11 @@ class ResNet(Model):
             blocks = []
             shortcut = False
             for b in range(num_blocks):
+                if b == 1:
+                    in_channels = out_channels[idx] * Block.expansion
                 block = Block(
-                    num_channels=num_in[idx] if b == 0 else num_out[idx] * 4,
-                    num_filters=num_out[idx],
+                    num_channels=in_channels,
+                    num_filters=out_channels[idx],
                     stride=2 if b == 0 and idx != 0 else 1,
                     shortcut=shortcut)
                 blocks.append(block)
@@ -142,8 +215,8 @@ class ResNet(Model):
         self.global_pool = Pool2D(
             pool_size=7, pool_type='avg', global_pooling=True)
 
-        stdv = 1.0 / math.sqrt(2048 * 1.0)
-        self.fc_input_dim = num_out[-1] * 4 * 1 * 1
+        stdv = 1.0 / math.sqrt(out_channels[-1] * Block.expansion * 1.0)
+        self.fc_input_dim = out_channels[-1] * Block.expansion * 1 * 1
         self.fc = Linear(
             self.fc_input_dim,
             num_classes,
@@ -175,13 +248,46 @@ def _resnet(arch, Block, depth, pretrained):
     return model
 
 
+def resnet18(pretrained=False):
+    """ResNet 18-layer model
+    
+    Args:
+        pretrained (bool): If True, returns a model pre-trained on ImageNet
+    """
+    return _resnet('resnet18', BasicBlock, 18, pretrained)
+
+
+def resnet34(pretrained=False):
+    """ResNet 34-layer model
+    
+    Args:
+        pretrained (bool): If True, returns a model pre-trained on ImageNet
+    """
+    return _resnet('resnet34', BasicBlock, 34, pretrained)
+
+
 def resnet50(pretrained=False):
+    """ResNet 50-layer model
+    
+    Args:
+        pretrained (bool): If True, returns a model pre-trained on ImageNet
+    """
     return _resnet('resnet50', BottleneckBlock, 50, pretrained)
 
 
 def resnet101(pretrained=False):
+    """ResNet 101-layer model
+    
+    Args:
+        pretrained (bool): If True, returns a model pre-trained on ImageNet
+    """
     return _resnet('resnet101', BottleneckBlock, 101, pretrained)
 
 
 def resnet152(pretrained=False):
+    """ResNet 152-layer model
+    
+    Args:
+        pretrained (bool): If True, returns a model pre-trained on ImageNet
+    """
     return _resnet('resnet152', BottleneckBlock, 152, pretrained)

models/vgg.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 paddle
+import paddle.fluid as fluid
+from paddle.fluid.dygraph.nn import Conv2D, Pool2D, BatchNorm, Linear
+from paddle.fluid.dygraph.container import Sequential
+
+from model import Model
+from .download import get_weights_path
+
+__all__ = [
+    'VGG',
+    'vgg11',
+    'vgg11_bn',
+    'vgg13',
+    'vgg13_bn',
+    'vgg16',
+    'vgg16_bn',
+    'vgg19_bn',
+    'vgg19',
+]
+
+model_urls = {
+    'vgg16': ('https://paddle-hapi.bj.bcebos.com/models/vgg16.pdparams',
+              'c788f453a3b999063e8da043456281ee')
+}
+
+
+class Classifier(fluid.dygraph.Layer):
+    def __init__(self, num_classes):
+        super(Classifier, self).__init__()
+        self.linear1 = Linear(512 * 7 * 7, 4096)
+        self.linear2 = Linear(4096, 4096)
+        self.linear3 = Linear(4096, num_classes, act='softmax')
+
+    def forward(self, x):
+        x = self.linear1(x)
+        x = fluid.layers.relu(x)
+        x = fluid.layers.dropout(x, 0.5)
+        x = self.linear2(x)
+        x = fluid.layers.relu(x)
+        x = fluid.layers.dropout(x, 0.5)
+        out = self.linear3(x)
+        return out
+
+
+class VGG(Model):
+    """VGG model from
+    `"Very Deep Convolutional Networks For Large-Scale Image Recognition" <https://arxiv.org/pdf/1409.1556.pdf>`_
+
+    Args:
+        features (fluid.dygraph.Layer): vgg features create by function make_layers.
+        num_classes (int): output dim of last fc layer. Default: 1000.
+    """
+
+    def __init__(self, features, num_classes=1000):
+        super(VGG, self).__init__()
+        self.features = features
+        classifier = Classifier(num_classes)
+        self.classifier = self.add_sublayer("classifier",
+                                            Sequential(classifier))
+
+    def forward(self, x):
+        x = self.features(x)
+        x = fluid.layers.flatten(x, 1)
+        x = self.classifier(x)
+        return x
+
+
+def make_layers(cfg, batch_norm=False):
+    layers = []
+    in_channels = 3
+
+    for v in cfg:
+        if v == 'M':
+            layers += [Pool2D(pool_size=2, pool_stride=2)]
+        else:
+            if batch_norm:
+                conv2d = Conv2D(in_channels, v, filter_size=3, padding=1)
+                layers += [conv2d, BatchNorm(v, act='relu')]
+            else:
+                conv2d = Conv2D(
+                    in_channels, v, filter_size=3, padding=1, act='relu')
+                layers += [conv2d]
+            in_channels = v
+    return Sequential(*layers)
+
+
+cfgs = {
+    'A': [64, 'M', 128, 'M', 256, 256, 'M', 512, 512, 'M', 512, 512, 'M'],
+    'B':
+    [64, 64, 'M', 128, 128, 'M', 256, 256, 'M', 512, 512, 'M', 512, 512, 'M'],
+    'D': [
+        64, 64, 'M', 128, 128, 'M', 256, 256, 256, 'M', 512, 512, 512, 'M',
+        512, 512, 512, 'M'
+    ],
+    'E': [
+        64, 64, 'M', 128, 128, 'M', 256, 256, 256, 256, 'M', 512, 512, 512,
+        512, 'M', 512, 512, 512, 512, 'M'
+    ],
+}
+
+
+def _vgg(arch, cfg, batch_norm, pretrained, **kwargs):
+    model = VGG(make_layers(cfgs[cfg], batch_norm=batch_norm), **kwargs)
+
+    if pretrained:
+        assert arch in model_urls, "{} model do not have a pretrained model now, you should set pretrained=False".format(
+            arch)
+        weight_path = get_weights_path(model_urls[arch][0],
+                                       model_urls[arch][1])
+        assert weight_path.endswith(
+            '.pdparams'), "suffix of weight must be .pdparams"
+        model.load(weight_path[:-9])
+
+    return model
+
+
+def vgg11(pretrained=False, **kwargs):
+    """VGG 11-layer model
+    
+    Args:
+        pretrained (bool): If True, returns a model pre-trained on ImageNet
+    """
+    return _vgg('vgg11', 'A', False, pretrained, **kwargs)
+
+
+def vgg11_bn(pretrained=False, **kwargs):
+    """VGG 11-layer model with batch normalization
+    
+    Args:
+        pretrained (bool): If True, returns a model pre-trained on ImageNet
+    """
+    return _vgg('vgg11_bn', 'A', True, pretrained, **kwargs)
+
+
+def vgg13(pretrained=False, **kwargs):
+    """VGG 13-layer model
+    
+    Args:
+        pretrained (bool): If True, returns a model pre-trained on ImageNet
+    """
+    return _vgg('vgg13', 'B', False, pretrained, **kwargs)
+
+
+def vgg13_bn(pretrained=False, **kwargs):
+    """VGG 13-layer model with batch normalization
+    
+    Args:
+        pretrained (bool): If True, returns a model pre-trained on ImageNet
+    """
+    return _vgg('vgg13_bn', 'B', True, pretrained, **kwargs)
+
+
+def vgg16(pretrained=False, **kwargs):
+    """VGG 16-layer model 
+    
+    Args:
+        pretrained (bool): If True, returns a model pre-trained on ImageNet
+    """
+    return _vgg('vgg16', 'D', False, pretrained, **kwargs)
+
+
+def vgg16_bn(pretrained=False, **kwargs):
+    """VGG 16-layer with batch normalization
+    
+    Args:
+        pretrained (bool): If True, returns a model pre-trained on ImageNet
+    """
+    return _vgg('vgg16_bn', 'D', True, pretrained, **kwargs)
+
+
+def vgg19(pretrained=False, **kwargs):
+    """VGG 19-layer model 
+    
+    Args:
+        pretrained (bool): If True, returns a model pre-trained on ImageNet
+    """
+    return _vgg('vgg19', 'E', False, pretrained, **kwargs)
+
+
+def vgg19_bn(pretrained=False, **kwargs):
+    """VGG 19-layer model with batch normalization
+    
+    Args:
+        pretrained (bool): If True, returns a model pre-trained on ImageNet
+    """
+    return _vgg('vgg19_bn', 'E', True, pretrained, **kwargs)