cherry pick 254 284 (#286)

* add slimfacenet to paddleslim/models (#284)

* add slimfacenets (#254)

* add slimfacenet

demo/models/__init__.py

@@ -4,6 +4,7 @@ from .resnet import ResNet34, ResNet50
 from .resnet_vd import ResNet50_vd, ResNet101_vd
 from .mobilenet_v2 import MobileNetV2_x0_25, MobileNetV2
 from .pvanet import PVANet
+from .slimfacenet import SlimFaceNet_A_x0_60, SlimFaceNet_B_x0_75, SlimFaceNet_C_x0_75
 __all__ = [
     "model_list", "MobileNet", "ResNet34", "ResNet50", "MobileNetV2", "PVANet",
     "ResNet50_vd", "ResNet101_vd", "MobileNetV2_x0_25"

demo/models/slimfacenet.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 math
+import datetime
+import numpy as np
+
+import paddle
+import paddle.fluid as fluid
+from paddle.fluid.initializer import MSRA
+from paddle.fluid.param_attr import ParamAttr
+
+
+class SlimFaceNet():
+    def __init__(self, class_dim, scale=0.6, arch=None):
+
+        assert arch is not None
+        self.arch = arch
+        self.class_dim = class_dim
+        kernels = [3]
+        expansions = [2, 4, 6]
+        SE = [0, 1]
+        self.table = []
+        for k in kernels:
+            for e in expansions:
+                for se in SE:
+                    self.table.append((k, e, se))
+
+        if scale == 1.0:
+            # 100% - channel
+            self.Slimfacenet_bottleneck_setting = [
+                # t, c , n ,s
+                [2, 64, 5, 2],
+                [4, 128, 1, 2],
+                [2, 128, 6, 1],
+                [4, 128, 1, 2],
+                [2, 128, 2, 1]
+            ]
+        elif scale == 0.9:
+            # 90% - channel
+            self.Slimfacenet_bottleneck_setting = [
+                # t, c , n ,s
+                [2, 56, 5, 2],
+                [4, 116, 1, 2],
+                [2, 116, 6, 1],
+                [4, 116, 1, 2],
+                [2, 116, 2, 1]
+            ]
+        elif scale == 0.75:
+            # 75% - channel
+            self.Slimfacenet_bottleneck_setting = [
+                # t, c , n ,s
+                [2, 48, 5, 2],
+                [4, 96, 1, 2],
+                [2, 96, 6, 1],
+                [4, 96, 1, 2],
+                [2, 96, 2, 1]
+            ]
+        elif scale == 0.6:
+            # 60% - channel
+            self.Slimfacenet_bottleneck_setting = [
+                # t, c , n ,s
+                [2, 40, 5, 2],
+                [4, 76, 1, 2],
+                [2, 76, 6, 1],
+                [4, 76, 1, 2],
+                [2, 76, 2, 1]
+            ]
+        else:
+            print('WRONG scale')
+            exit()
+        self.extract_feature = True
+
+    def set_extract_feature_flag(self, flag):
+        self.extract_feature = flag
+
+    def net(self, input, label=None):
+        x = self.conv_bn_layer(
+            input,
+            filter_size=3,
+            num_filters=64,
+            stride=2,
+            padding=1,
+            num_groups=1,
+            if_act=True,
+            name='conv3x3')
+        x = self.conv_bn_layer(
+            x,
+            filter_size=3,
+            num_filters=64,
+            stride=1,
+            padding=1,
+            num_groups=64,
+            if_act=True,
+            name='dw_conv3x3')
+
+        in_c = 64
+        cnt = 0
+        for _exp, out_c, times, _stride in self.Slimfacenet_bottleneck_setting:
+            for i in range(times):
+                stride = _stride if i == 0 else 1
+                filter_size, exp, se = self.table[self.arch[cnt]]
+                se = False if se == 0 else True
+                x = self.residual_unit(
+                    x,
+                    num_in_filter=in_c,
+                    num_out_filter=out_c,
+                    stride=stride,
+                    filter_size=filter_size,
+                    expansion_factor=exp,
+                    use_se=se,
+                    name='residual_unit' + str(cnt + 1))
+                cnt += 1
+                in_c = out_c
+
+        out_c = 512
+        x = self.conv_bn_layer(
+            x,
+            filter_size=1,
+            num_filters=out_c,
+            stride=1,
+            padding=0,
+            num_groups=1,
+            if_act=True,
+            name='conv1x1')
+        x = self.conv_bn_layer(
+            x,
+            filter_size=(7, 6),
+            num_filters=out_c,
+            stride=1,
+            padding=0,
+            num_groups=out_c,
+            if_act=False,
+            name='global_dw_conv7x7')
+        x = fluid.layers.conv2d(
+            x,
+            num_filters=128,
+            filter_size=1,
+            stride=1,
+            padding=0,
+            groups=1,
+            act=None,
+            use_cudnn=True,
+            param_attr=ParamAttr(
+                name='linear_conv1x1_weights',
+                initializer=MSRA(),
+                regularizer=fluid.regularizer.L2Decay(4e-4)),
+            bias_attr=False)
+        bn_name = 'linear_conv1x1_bn'
+        x = fluid.layers.batch_norm(
+            x,
+            param_attr=ParamAttr(name=bn_name + "_scale"),
+            bias_attr=ParamAttr(name=bn_name + "_offset"),
+            moving_mean_name=bn_name + '_mean',
+            moving_variance_name=bn_name + '_variance')
+
+        x = fluid.layers.reshape(x, shape=[x.shape[0], x.shape[1]])
+
+        if self.extract_feature:
+            return x
+
+        out = self.arc_margin_product(
+            x, label, self.class_dim, s=32.0, m=0.50, mode=2)
+        softmax = fluid.layers.softmax(input=out)
+        cost = fluid.layers.cross_entropy(input=softmax, label=label)
+        loss = fluid.layers.mean(x=cost)
+        acc = fluid.layers.accuracy(input=out, label=label, k=1)
+        return loss, acc
+
+    def residual_unit(self,
+                      input,
+                      num_in_filter,
+                      num_out_filter,
+                      stride,
+                      filter_size,
+                      expansion_factor,
+                      use_se=False,
+                      name=None):
+
+        num_expfilter = int(round(num_in_filter * expansion_factor))
+        input_data = input
+
+        expand_conv = self.conv_bn_layer(
+            input=input,
+            filter_size=1,
+            num_filters=num_expfilter,
+            stride=1,
+            padding=0,
+            if_act=True,
+            name=name + '_expand')
+
+        depthwise_conv = self.conv_bn_layer(
+            input=expand_conv,
+            filter_size=filter_size,
+            num_filters=num_expfilter,
+            stride=stride,
+            padding=int((filter_size - 1) // 2),
+            if_act=True,
+            num_groups=num_expfilter,
+            use_cudnn=True,
+            name=name + '_depthwise')
+
+        if use_se:
+            depthwise_conv = self.se_block(
+                input=depthwise_conv,
+                num_out_filter=num_expfilter,
+                name=name + '_se')
+
+        linear_conv = self.conv_bn_layer(
+            input=depthwise_conv,
+            filter_size=1,
+            num_filters=num_out_filter,
+            stride=1,
+            padding=0,
+            if_act=False,
+            name=name + '_linear')
+        if num_in_filter != num_out_filter or stride != 1:
+            return linear_conv
+        else:
+            return fluid.layers.elementwise_add(
+                x=input_data, y=linear_conv, act=None)
+
+    def se_block(self, input, num_out_filter, ratio=4, name=None):
+        num_mid_filter = int(num_out_filter // ratio)
+        pool = fluid.layers.pool2d(
+            input=input, pool_type='avg', global_pooling=True, use_cudnn=False)
+        conv1 = fluid.layers.conv2d(
+            input=pool,
+            filter_size=1,
+            num_filters=num_mid_filter,
+            act=None,
+            param_attr=ParamAttr(name=name + '_1_weights'),
+            bias_attr=ParamAttr(name=name + '_1_offset'))
+        conv1 = fluid.layers.prelu(
+            conv1,
+            mode='channel',
+            param_attr=ParamAttr(
+                name=name + '_prelu',
+                regularizer=fluid.regularizer.L2Decay(0.0)))
+        conv2 = fluid.layers.conv2d(
+            input=conv1,
+            filter_size=1,
+            num_filters=num_out_filter,
+            act='hard_sigmoid',
+            param_attr=ParamAttr(name=name + '_2_weights'),
+            bias_attr=ParamAttr(name=name + '_2_offset'))
+        scale = fluid.layers.elementwise_mul(x=input, y=conv2, axis=0)
+        return scale
+
+    def conv_bn_layer(self,
+                      input,
+                      filter_size,
+                      num_filters,
+                      stride,
+                      padding,
+                      num_groups=1,
+                      if_act=True,
+                      name=None,
+                      use_cudnn=True):
+        conv = fluid.layers.conv2d(
+            input=input,
+            num_filters=num_filters,
+            filter_size=filter_size,
+            stride=stride,
+            padding=padding,
+            groups=num_groups,
+            act=None,
+            use_cudnn=use_cudnn,
+            param_attr=ParamAttr(
+                name=name + '_weights', initializer=MSRA()),
+            bias_attr=False)
+        bn_name = name + '_bn'
+        bn = fluid.layers.batch_norm(
+            input=conv,
+            param_attr=ParamAttr(name=bn_name + "_scale"),
+            bias_attr=ParamAttr(name=bn_name + "_offset"),
+            moving_mean_name=bn_name + '_mean',
+            moving_variance_name=bn_name + '_variance')
+        if if_act:
+            return fluid.layers.prelu(
+                bn,
+                mode='channel',
+                param_attr=ParamAttr(
+                    name=name + '_prelu',
+                    regularizer=fluid.regularizer.L2Decay(0.0)))
+        else:
+            return bn
+
+    def arc_margin_product(self, input, label, out_dim, s=32.0, m=0.50,
+                           mode=2):
+        input_norm = fluid.layers.sqrt(
+            fluid.layers.reduce_sum(
+                fluid.layers.square(input), dim=1))
+        input = fluid.layers.elementwise_div(input, input_norm, axis=0)
+
+        weight = fluid.layers.create_parameter(
+            shape=[out_dim, input.shape[1]],
+            dtype='float32',
+            name='weight_norm',
+            attr=fluid.param_attr.ParamAttr(
+                initializer=fluid.initializer.Xavier(),
+                regularizer=fluid.regularizer.L2Decay(4e-4)))
+
+        weight_norm = fluid.layers.sqrt(
+            fluid.layers.reduce_sum(
+                fluid.layers.square(weight), dim=1))
+        weight = fluid.layers.elementwise_div(weight, weight_norm, axis=0)
+        weight = fluid.layers.transpose(weight, perm=[1, 0])
+        cosine = fluid.layers.mul(input, weight)
+        sine = fluid.layers.sqrt(1.0 - fluid.layers.square(cosine))
+
+        cos_m = math.cos(m)
+        sin_m = math.sin(m)
+        phi = cosine * cos_m - sine * sin_m
+
+        th = math.cos(math.pi - m)
+        mm = math.sin(math.pi - m) * m
+
+        if mode == 1:
+            phi = self.paddle_where_more_than(cosine, 0, phi, cosine)
+        elif mode == 2:
+            phi = self.paddle_where_more_than(cosine, th, phi, cosine - mm)
+        else:
+            pass
+
+        one_hot = fluid.layers.one_hot(input=label, depth=out_dim)
+        output = fluid.layers.elementwise_mul(
+            one_hot, phi) + fluid.layers.elementwise_mul(
+                (1.0 - one_hot), cosine)
+        output = output * s
+        return output
+
+    def paddle_where_more_than(self, target, limit, x, y):
+        mask = fluid.layers.cast(x=(target > limit), dtype='float32')
+        output = fluid.layers.elementwise_mul(
+            mask, x) + fluid.layers.elementwise_mul((1.0 - mask), y)
+        return output
+
+
+def SlimFaceNet_A_x0_60(class_dim=None, scale=0.6, arch=None):
+    scale = 0.6
+    arch = [0, 1, 5, 1, 0, 2, 1, 2, 0, 1, 2, 1, 1, 0, 1]
+    return SlimFaceNet(class_dim=class_dim, scale=scale, arch=arch)
+
+
+def SlimFaceNet_B_x0_75(class_dim=None, scale=0.6, arch=None):
+    scale = 0.75
+    arch = [1, 1, 0, 1, 1, 1, 1, 0, 1, 0, 1, 3, 2, 2, 3]
+    return SlimFaceNet(class_dim=class_dim, scale=scale, arch=arch)
+
+
+def SlimFaceNet_C_x0_75(class_dim=None, scale=0.6, arch=None):
+    scale = 0.75
+    arch = [1, 1, 2, 1, 0, 2, 1, 0, 1, 0, 1, 1, 2, 2, 3]
+    return SlimFaceNet(class_dim=class_dim, scale=scale, arch=arch)
+
+
+if __name__ == "__main__":
+    x = fluid.data(name='x', shape=[-1, 3, 112, 112], dtype='float32')
+    print(x.shape)
+    model = SlimFaceNet(10000, [1, 3, 3, 1, 1, 0, 0, 1, 0, 1, 1, 0, 5, 5, 3])
+    y = model.net(x)

demo/slimfacenet/README.md

+# slimfacenet使用示例
+
+本示例将演示如何训练`slimfacenet`及评测`slimfacenet`量化模型。
+
+当前示例支持以下人脸识别模型：
+
+- `SlimFaceNet_A_x0_60`
+- `SlimFaceNet_B_x0_75`
+
+## 1. 数据准备
+
+本示例支持`CASIA`和`lfw`两种公开数据集默认情况：
+
+1). 训练数据集位置`./CASIA`
+2). 测试数据集位置`./lfw`
+
+## 2. 下载预训练模型
+
+如果使用预先训练并量化好的`slimfacenet`模型，可以从以下地址下载
+
+
+## 3. 启动`slimfacenet`训练任务
+
+通过以下命令启动训练任务：
+
+```
+sh slim_train.sh
+或者
+export CUDA_VISIBLE_DEVICES=0
+python -u train_eval.py \
+    --action train \
+    --model=SlimFaceNet_B_x0_75
+```
+
+其中，SlimFaceNet_A_x0_60是`slimfacenet`搜索空间中的一个模型结构，通道数的缩放系数为0.6，
+在每个缩放系数下搜索空间中都共有6**15(约4700亿)种不同的模型结构。模型训练好之后会保存在`./out_inference/`
+
+
+## 4. 将float32模型量化为int8模型
+
+通过以下命令启动训练任务：
+
+```
+sh slim_quant.sh
+或者
+export CUDA_VISIBLE_DEVICES=0
+python -u train_eval.py --action quant
+```
+执行完之后量化模型会保存在`./quant_model/`, 注当前阶段量化模型还是是按float32保存的，转paddlelite后会变为int8
+
+## 4. 加载和评估量化模型
+
+本节介绍如何加载并评测预先训练好并量化后的模型。
+
+执行以下代码加载模型并评估模型在测试集上的指标。
+
+```
+将量化模型默认地址在`./quant_model/`
+sh slim_eval.sh
+或者
+export CUDA_VISIBLE_DEVICES=0
+python train_eval.py --action test
+```

demo/slimfacenet/dataloader/__init__.py

+from .casia import CASIA_Face

demo/slimfacenet/dataloader/casia.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 scipy.misc
+import os
+import paddle
+from paddle import fluid
+
+
+class CASIA_Face(object):
+    def __init__(self, root):
+        self.root = root
+
+        img_txt_dir = os.path.join(root, 'CASIA-WebFace-112X96.txt')
+        image_list = []
+        label_list = []
+        with open(img_txt_dir) as f:
+            img_label_list = f.read().splitlines()
+        for info in img_label_list:
+            image_dir, label_name = info.split(' ')
+            image_list.append(
+                os.path.join(root, 'CASIA-WebFace-112X96', image_dir))
+            label_list.append(int(label_name))
+
+        self.image_list = image_list
+        self.label_list = label_list
+        self.class_nums = len(np.unique(self.label_list))
+        self.shuffle_idx = list(
+            np.random.choice(
+                len(self.image_list), len(self.image_list), False))
+
+    def reader(self):
+        while True:
+            if len(self.shuffle_idx) == 0:
+                self.shuffle_idx = list(
+                    np.random.choice(
+                        len(self.image_list), len(self.image_list), False))
+                return
+            index = self.shuffle_idx.pop()
+
+            img_path = self.image_list[index]
+            target = self.label_list[index]
+
+            try:
+                img = scipy.misc.imread(img_path)
+            except:
+                continue
+
+            if len(img.shape) == 2:
+                img = np.stack([img] * 3, 2)
+            flip = np.random.choice(2) * 2 - 1
+            img = img[:, ::flip, :]
+            img = (img - 127.5) / 128.0
+            img = img.transpose(2, 0, 1)
+
+            yield img, target
+
+    def __len__(self):
+        return len(self.image_list)
+
+
+if __name__ == '__main__':
+    data_dir = 'PATH to CASIA dataset'
+
+    place = fluid.CPUPlace()
+    with fluid.dygraph.guard(place):
+        dataset = CASIA_Face(root=data_dir)
+        print(len(dataset))
+        print(dataset.class_nums)
+        trainloader = paddle.batch(
+            dataset.reader, batch_size=1, drop_last=False)
+        for i in range(10):
+            for data in trainloader():
+                img = np.array([x[0] for x in data]).astype('float32')
+                img = fluid.dygraph.to_variable(img)
+                print(img.shape)
+                label = np.array([x[1] for x in data]).astype('int64').reshape(
+                    -1, 1)
+                label = fluid.dygraph.to_variable(label)
+                print(label.shape)
+        print(len(dataset))

demo/slimfacenet/dataloader/lfw.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 scipy.misc
+
+import paddle
+from paddle import fluid
+
+
+class LFW(object):
+    def __init__(self, imgl, imgr):
+
+        self.imgl_list = imgl
+        self.imgr_list = imgr
+        self.shuffle_idx = [i for i in range(len(self.imgl_list))]
+
+    def reader(self):
+        while True:
+            if len(self.shuffle_idx) == 0:
+                self.shuffle_idx = [i for i in range(len(self.imgl_list))]
+                return
+            index = self.shuffle_idx.pop(0)
+
+            imgl = scipy.misc.imread(self.imgl_list[index])
+            if len(imgl.shape) == 2:
+                imgl = np.stack([imgl] * 3, 2)
+            imgr = scipy.misc.imread(self.imgr_list[index])
+            if len(imgr.shape) == 2:
+                imgr = np.stack([imgr] * 3, 2)
+
+            imglist = [imgl, imgl[:, ::-1, :], imgr, imgr[:, ::-1, :]]
+            for i in range(len(imglist)):
+                imglist[i] = (imglist[i] - 127.5) / 128.0
+                imglist[i] = imglist[i].transpose(2, 0, 1)
+
+            imgs = [img.astype('float32') for img in imglist]
+            yield imgs
+
+    def __len__(self):
+        return len(self.imgl_list)
+
+
+if __name__ == '__main__':
+    pass

demo/slimfacenet/lfw_eval.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 argparse
+import time
+import scipy.io
+import numpy as np
+
+import paddle
+from paddle import fluid
+
+from dataloader.casia import CASIA_Face
+from dataloader.lfw import LFW
+from paddleslim import models
+
+
+def parse_filelist(root):
+    with open(os.path.join(root, 'pairs.txt')) as f:
+        pairs = f.read().splitlines()[1:]
+    folder_name = 'lfw-112X96'
+    nameLs = []
+    nameRs = []
+    folds = []
+    flags = []
+    for i, p in enumerate(pairs):
+        p = p.split('\t')
+        if len(p) == 3:
+            nameL = os.path.join(root, folder_name, p[0],
+                                 p[0] + '_' + '{:04}.jpg'.format(int(p[1])))
+            nameR = os.path.join(root, folder_name, p[0],
+                                 p[0] + '_' + '{:04}.jpg'.format(int(p[2])))
+            fold = i // 600
+            flag = 1
+        elif len(p) == 4:
+            nameL = os.path.join(root, folder_name, p[0],
+                                 p[0] + '_' + '{:04}.jpg'.format(int(p[1])))
+            nameR = os.path.join(root, folder_name, p[2],
+                                 p[2] + '_' + '{:04}.jpg'.format(int(p[3])))
+            fold = i // 600
+            flag = -1
+        nameLs.append(nameL)
+        nameRs.append(nameR)
+        folds.append(fold)
+        flags.append(flag)
+    return [nameLs, nameRs, folds, flags]
+
+
+def get_accuracy(scores, flags, threshold):
+    p = np.sum(scores[flags == 1] > threshold)
+    n = np.sum(scores[flags == -1] < threshold)
+    return 1.0 * (p + n) / len(scores)
+
+
+def get_threshold(scores, flags, thrNum):
+    accuracys = np.zeros((2 * thrNum + 1, 1))
+    thresholds = np.arange(-thrNum, thrNum + 1) * 1.0 / thrNum
+    for i in range(2 * thrNum + 1):
+        accuracys[i] = get_accuracy(scores, flags, thresholds[i])
+
+    max_index = np.squeeze(accuracys == np.max(accuracys))
+    bestThreshold = np.mean(thresholds[max_index])
+    return bestThreshold
+
+
+def evaluation_10_fold(root='result.mat'):
+    ACCs = np.zeros(10)
+    result = scipy.io.loadmat(root)
+    for i in range(10):
+        fold = result['fold']
+        flags = result['flag']
+        featureLs = result['fl']
+        featureRs = result['fr']
+
+        valFold = fold != i
+        testFold = fold == i
+        flags = np.squeeze(flags)
+
+        mu = np.mean(
+            np.concatenate(
+                (featureLs[valFold[0], :], featureRs[valFold[0], :]), 0), 0)
+        mu = np.expand_dims(mu, 0)
+        featureLs = featureLs - mu
+        featureRs = featureRs - mu
+        featureLs = featureLs / np.expand_dims(
+            np.sqrt(np.sum(np.power(featureLs, 2), 1)), 1)
+        featureRs = featureRs / np.expand_dims(
+            np.sqrt(np.sum(np.power(featureRs, 2), 1)), 1)
+
+        scores = np.sum(np.multiply(featureLs, featureRs), 1)
+        threshold = get_threshold(scores[valFold[0]], flags[valFold[0]], 10000)
+        ACCs[i] = get_accuracy(scores[testFold[0]], flags[testFold[0]],
+                               threshold)
+    return ACCs
+
+
+def test(test_reader, flods, flags, net, args):
+    net.eval()
+    featureLs = None
+    featureRs = None
+    for idx, data in enumerate(test_reader()):
+        data_list = [[] for _ in range(4)]
+        for _ in range(len(data)):
+            data_list[0].append(data[_][0])
+            data_list[1].append(data[_][1])
+            data_list[2].append(data[_][2])
+            data_list[3].append(data[_][3])
+        res = [
+            net(fluid.dygraph.to_variable(np.array(d))).numpy()
+            for d in data_list
+        ]
+        featureL = np.concatenate((res[0], res[1]), 1)
+        featureR = np.concatenate((res[2], res[3]), 1)
+        if featureLs is None:
+            featureLs = featureL
+        else:
+            featureLs = np.concatenate((featureLs, featureL), 0)
+        if featureRs is None:
+            featureRs = featureR
+        else:
+            featureRs = np.concatenate((featureRs, featureR), 0)
+    result = {'fl': featureLs, 'fr': featureRs, 'fold': flods, 'flag': flags}
+    scipy.io.savemat(args.feature_save_dir, result)
+    ACCs = evaluation_10_fold(args.feature_save_dir)
+    for i in range(len(ACCs)):
+        print('{}    {:.2f}'.format(i + 1, ACCs[i] * 100))
+    print('--------')
+    print('AVE    {:.2f}'.format(np.mean(ACCs) * 100))
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description='PaddlePaddle SlimFaceNet')
+    parser.add_argument(
+        '--use_gpu', default=0, type=int, help='Use GPU or not, 0 is not used')
+    parser.add_argument(
+        '--test_data_dir', default='./lfw', type=str, help='lfw_data_dir')
+    parser.add_argument(
+        '--resume', default='output/0', type=str, help='resume')
+    parser.add_argument(
+        '--feature_save_dir',
+        default='result.mat',
+        type=str,
+        help='The path of the extract features save, must be .mat file')
+    args = parser.parse_args()
+
+    place = fluid.CPUPlace() if args.use_gpu == 0 else fluid.CUDAPlace(0)
+    with fluid.dygraph.guard(place):
+        train_dataset = CASIA_Face(root=args.train_data_dir)
+        nl, nr, flods, flags = parse_filelist(args.test_data_dir)
+        test_dataset = LFW(nl, nr)
+        test_reader = paddle.batch(
+            test_dataset.reader,
+            batch_size=args.test_batchsize,
+            drop_last=False)
+
+        net = models.__dict__[args.model](class_dim=train_dataset.class_nums)
+        if args.resume:
+            assert os.path.exists(args.resume + ".pdparams"
+                                  ), "Given dir {}.pdparams not exist.".format(
+                                      args.resume)
+            para_dict, opti_dict = fluid.dygraph.load_dygraph(args.resume)
+            net.set_dict(para_dict)
+
+        test(test_reader, flods, flags, net, args)

demo/slimfacenet/slim_eval.sh

+# ================================================================
+#   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.
+#!/bin/bash
+export CUDA_VISIBLE_DEVICES=0
+python train_eval.py --action test \
+    --train_data_dir=/PATH_TO_CASIA_Dataset \
+    --test_data_dir=/PATH_TO_lfw \

demo/slimfacenet/slim_quant.sh

+# ================================================================
+#   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.
+#!/bin/bash
+export CUDA_VISIBLE_DEVICES=0
+python train_eval.py --action quant \
+    --train_data_dir=/PATH_TO_CASIA_Dataset \
+    --test_data_dir=/PATH_TO_lfw \

demo/slimfacenet/slim_train.sh

+# ================================================================
+#   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.
+#!/bin/bash
+export CUDA_VISIBLE_DEVICES=0
+python -u train_eval.py \
+    --train_data_dir=/PATH_TO_CASIA_Dataset \
+    --test_data_dir=/PATH_TO_LFW \
+    --action train \
+    --model=SlimFaceNet_B_x0_75

demo/slimfacenet/train_eval.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 shutil
+import subprocess
+import argparse
+import time
+import scipy.io
+import numpy as np
+
+import paddle
+import paddle.fluid as fluid
+import paddle.fluid.compiler as compiler
+
+from dataloader.casia import CASIA_Face
+from dataloader.lfw import LFW
+from lfw_eval import parse_filelist, evaluation_10_fold
+from paddleslim import models
+from paddleslim.quant import quant_post
+
+
+def now():
+    return time.strftime('%Y-%m-%d %H:%M:%S', time.localtime(time.time()))
+
+
+def creat_optimizer(args, trainset_scale):
+    start_step = trainset_scale * args.start_epoch // args.train_batchsize
+
+    if args.lr_strategy == 'piecewise_decay':
+        bd = [
+            trainset_scale * int(e) // args.train_batchsize
+            for e in args.lr_steps.strip().split(',')
+        ]
+        lr = [float(e) for e in args.lr_list.strip().split(',')]
+        assert len(bd) == len(lr) - 1
+        optimizer = fluid.optimizer.Momentum(
+            learning_rate=fluid.layers.piecewise_decay(
+                boundaries=bd, values=lr),
+            momentum=0.9,
+            regularization=fluid.regularizer.L2Decay(args.l2_decay))
+    elif args.lr_strategy == 'cosine_decay':
+        lr = args.lr
+        step_each_epoch = trainset_scale // args.train_batchsize
+        optimizer = fluid.optimizer.Momentum(
+            learning_rate=fluid.layers.cosine_decay(lr, step_each_epoch,
+                                                    args.total_epoch),
+            momentum=0.9,
+            regularization=fluid.regularizer.L2Decay(args.l2_decay))
+    else:
+        print('Wrong learning rate strategy')
+        exit()
+    return optimizer
+
+
+def test(test_exe, test_program, test_out, args):
+    featureLs = None
+    featureRs = None
+    out_feature, test_reader, flods, flags = test_out
+    for idx, data in enumerate(test_reader()):
+        res = []
+        res.append(
+            test_exe.run(test_program,
+                         feed={u'image_test': data[0][u'image_test1']},
+                         fetch_list=out_feature))
+        res.append(
+            test_exe.run(test_program,
+                         feed={u'image_test': data[0][u'image_test2']},
+                         fetch_list=out_feature))
+        res.append(
+            test_exe.run(test_program,
+                         feed={u'image_test': data[0][u'image_test3']},
+                         fetch_list=out_feature))
+        res.append(
+            test_exe.run(test_program,
+                         feed={u'image_test': data[0][u'image_test4']},
+                         fetch_list=out_feature))
+        featureL = np.concatenate((res[0][0], res[1][0]), 1)
+        featureR = np.concatenate((res[2][0], res[3][0]), 1)
+        if featureLs is None:
+            featureLs = featureL
+        else:
+            featureLs = np.concatenate((featureLs, featureL), 0)
+        if featureRs is None:
+            featureRs = featureR
+        else:
+            featureRs = np.concatenate((featureRs, featureR), 0)
+    result = {'fl': featureLs, 'fr': featureRs, 'fold': flods, 'flag': flags}
+    scipy.io.savemat(args.feature_save_dir, result)
+    ACCs = evaluation_10_fold(args.feature_save_dir)
+    with open(os.path.join(args.save_ckpt, 'log.txt'), 'a+') as f:
+        f.writelines('eval model {}\n'.format(args.model))
+    for i in range(len(ACCs)):
+        print('{}    {}'.format(i + 1, ACCs[i] * 100))
+        with open(os.path.join(args.save_ckpt, 'log.txt'), 'a+') as f:
+            f.writelines('{}    {}\n'.format(i + 1, ACCs[i] * 100))
+    print('--------')
+    print('AVE {}'.format(np.mean(ACCs) * 100))
+    with open(os.path.join(args.save_ckpt, 'log.txt'), 'a+') as f:
+        f.writelines('--------\n')
+        f.writelines('AVE    {}\n'.format(np.mean(ACCs) * 100))
+    return np.mean(ACCs) * 100
+
+
+def train(exe, train_program, train_out, test_program, test_out, args):
+    loss, acc, global_lr, train_reader = train_out
+    fetch_list_train = [loss.name, acc.name, global_lr.name]
+    build_strategy = fluid.BuildStrategy()
+    build_strategy.fuse_all_optimizer_ops = True
+    compiled_prog = compiler.CompiledProgram(
+        train_program, build_strategy=build_strategy).with_data_parallel(
+            loss_name=loss.name, build_strategy=build_strategy)
+    for epoch_id in range(args.start_epoch, args.total_epoch):
+        for batch_id, data in enumerate(train_reader()):
+            loss, acc, global_lr = exe.run(compiled_prog,
+                                           feed=data,
+                                           fetch_list=fetch_list_train)
+            avg_loss = np.mean(np.array(loss))
+            avg_acc = np.mean(np.array(acc))
+            print(
+                '{}  Epoch: {:^4d} step: {:^4d} loss: {:.6f}, acc: {:.6f}, lr: {}'.
+                format(now(), epoch_id, batch_id, avg_loss, avg_acc,
+                       float(np.mean(np.array(global_lr)))))
+        if batch_id % args.save_frequency == 0:
+            model_path = os.path.join(args.save_ckpt, str(epoch_id))
+            fluid.io.save_persistables(
+                executor=exe, dirname=model_path, main_program=train_program)
+            test(exe, test_program, test_out, args)
+    out_feature, test_reader, flods, flags = test_out
+    fluid.io.save_inference_model(
+        executor=exe,
+        dirname='./out_inference',
+        feeded_var_names=['image_test'],
+        target_vars=[out_feature],
+        main_program=test_program)
+
+
+def build_program(program, startup, args, is_train=True):
+    num_trainers = len(os.getenv('CUDA_VISIBLE_DEVICES').split(','))
+    places = fluid.cuda_places() if args.use_gpu else fluid.CPUPlace()
+
+    train_dataset = CASIA_Face(root=args.train_data_dir)
+    trainset_scale = len(train_dataset)
+
+    with fluid.program_guard(main_program=program, startup_program=startup):
+        with fluid.unique_name.guard():
+            # Model construction
+            model = models.__dict__[args.model](
+                class_dim=train_dataset.class_nums)
+
+            if is_train:
+                image = fluid.data(
+                    name='image', shape=[-1, 3, 112, 96], dtype='float32')
+                label = fluid.data(name='label', shape=[-1, 1], dtype='int64')
+                train_reader = paddle.batch(
+                    train_dataset.reader,
+                    batch_size=args.train_batchsize // num_trainers,
+                    drop_last=False)
+                reader = fluid.io.DataLoader.from_generator(
+                    feed_list=[image, label],
+                    capacity=64,
+                    iterable=True,
+                    return_list=False)
+                reader.set_sample_list_generator(train_reader, places=places)
+
+                model.extract_feature = False
+                loss, acc = model.net(image, label)
+                optimizer = creat_optimizer(args, trainset_scale)
+                optimizer.minimize(loss)
+                global_lr = optimizer._global_learning_rate()
+                out = (loss, acc, global_lr, reader)
+
+            else:
+                nl, nr, flods, flags = parse_filelist(args.test_data_dir)
+                test_dataset = LFW(nl, nr)
+                test_reader = paddle.batch(
+                    test_dataset.reader,
+                    batch_size=args.test_batchsize,
+                    drop_last=False)
+                image_test = fluid.data(
+                    name='image_test', shape=[-1, 3, 112, 96], dtype='float32')
+                image_test1 = fluid.data(
+                    name='image_test1',
+                    shape=[-1, 3, 112, 96],
+                    dtype='float32')
+                image_test2 = fluid.data(
+                    name='image_test2',
+                    shape=[-1, 3, 112, 96],
+                    dtype='float32')
+                image_test3 = fluid.data(
+                    name='image_test3',
+                    shape=[-1, 3, 112, 96],
+                    dtype='float32')
+                image_test4 = fluid.data(
+                    name='image_test4',
+                    shape=[-1, 3, 112, 96],
+                    dtype='float32')
+                reader = fluid.io.DataLoader.from_generator(
+                    feed_list=[
+                        image_test1, image_test2, image_test3, image_test4
+                    ],
+                    capacity=64,
+                    iterable=True,
+                    return_list=False)
+                reader.set_sample_list_generator(
+                    test_reader,
+                    places=fluid.cuda_places()
+                    if args.use_gpu else fluid.CPUPlace())
+
+                model.extract_feature = True
+                feature = model.net(image_test)
+                out = (feature, reader, flods, flags)
+
+            return out
+
+
+def quant_val_reader_batch():
+    nl, nr, flods, flags = parse_filelist(args.test_data_dir)
+    test_dataset = LFW(nl, nr)
+    test_reader = paddle.batch(
+        test_dataset.reader, batch_size=1, drop_last=False)
+    shuffle_reader = fluid.io.shuffle(test_reader, 1)
+
+    def _reader():
+        while True:
+            for idx, data in enumerate(shuffle_reader()):
+                yield np.expand_dims(data[0][0], axis=0)
+
+    return _reader
+
+
+def main():
+    global args
+    parser = argparse.ArgumentParser(description='PaddlePaddle SlimFaceNet')
+    parser.add_argument(
+        '--action', default='train', type=str, help='train/test/quant')
+    parser.add_argument(
+        '--model',
+        default='SlimFaceNet_B_x0_75',
+        type=str,
+        help='SlimFaceNet_B_x0_75/SlimFaceNet_C_x0_75/SlimFaceNet_A_x0_60')
+    parser.add_argument(
+        '--use_gpu', default=1, type=int, help='Use GPU or not, 0 is not used')
+    parser.add_argument(
+        '--lr_strategy',
+        default='piecewise_decay',
+        type=str,
+        help='lr_strategy')
+    parser.add_argument('--lr', default=0.1, type=float, help='learning rate')
+    parser.add_argument(
+        '--lr_list',
+        default='0.1,0.01,0.001,0.0001',
+        type=str,
+        help='learning rate list (piecewise_decay)')
+    parser.add_argument(
+        '--lr_steps',
+        default='36,52,58',
+        type=str,
+        help='learning rate decay at which epochs')
+    parser.add_argument(
+        '--l2_decay', default=4e-5, type=float, help='base l2_decay')
+    parser.add_argument(
+        '--train_data_dir', default='./CASIA', type=str, help='train_data_dir')
+    parser.add_argument(
+        '--test_data_dir', default='./lfw', type=str, help='lfw_data_dir')
+    parser.add_argument(
+        '--train_batchsize', default=512, type=int, help='train_batchsize')
+    parser.add_argument(
+        '--test_batchsize', default=500, type=int, help='test_batchsize')
+    parser.add_argument(
+        '--img_shape', default='3,112,96', type=str, help='img_shape')
+    parser.add_argument(
+        '--start_epoch', default=0, type=int, help='start_epoch')
+    parser.add_argument(
+        '--total_epoch', default=80, type=int, help='total_epoch')
+    parser.add_argument(
+        '--save_frequency', default=1, type=int, help='save_frequency')
+    parser.add_argument(
+        '--save_ckpt', default='output', type=str, help='save_ckpt')
+    parser.add_argument(
+        '--feature_save_dir',
+        default='result.mat',
+        type=str,
+        help='The path of the extract features save, must be .mat file')
+    args = parser.parse_args()
+
+    num_trainers = len(os.getenv('CUDA_VISIBLE_DEVICES').split(','))
+    print(args)
+    print('num_trainers: {}'.format(num_trainers))
+    if args.save_ckpt == None:
+        args.save_ckpt = 'output'
+    if not os.path.exists(args.save_ckpt):
+        subprocess.call(['mkdir', '-p', args.save_ckpt])
+
+    with open(os.path.join(args.save_ckpt, 'log.txt'), 'w+') as f:
+        f.writelines(str(args) + '\n')
+        f.writelines('num_trainers: {}'.format(num_trainers) + '\n')
+
+    if args.action == 'train':
+        train_program = fluid.Program()
+    test_program = fluid.Program()
+    startup_program = fluid.Program()
+
+    if args.action == 'train':
+        train_out = build_program(train_program, startup_program, args, True)
+    test_out = build_program(test_program, startup_program, args, False)
+    test_program = test_program.clone(for_test=True)
+    place = fluid.CUDAPlace(0) if args.use_gpu else fluid.CPUPlace()
+    exe = fluid.Executor(place)
+    exe.run(startup_program)
+
+    if args.action == 'train':
+        train(exe, train_program, train_out, test_program, test_out, args)
+    elif args.action == 'quant':
+        quant_post(
+            executor=exe,
+            model_dir='./out_inference/',
+            quantize_model_path='./quant_model/',
+            sample_generator=quant_val_reader_batch(),
+            model_filename=None,  #'model',
+            params_filename=None,  #'params',
+            batch_size=100,
+            batch_nums=10)
+    elif args.action == 'test':
+        [inference_program, feed_target_names,
+         fetch_targets] = fluid.io.load_inference_model(
+             dirname='./quant_model/',
+             model_filename=None,
+             params_filename=None,
+             executor=exe)
+        nl, nr, flods, flags = parse_filelist(args.test_data_dir)
+        test_dataset = LFW(nl, nr)
+        test_reader = paddle.batch(
+            test_dataset.reader,
+            batch_size=args.test_batchsize,
+            drop_last=False)
+        image_test = fluid.data(
+            name='image_test', shape=[-1, 3, 112, 96], dtype='float32')
+        image_test1 = fluid.data(
+            name='image_test1', shape=[-1, 3, 112, 96], dtype='float32')
+        image_test2 = fluid.data(
+            name='image_test2', shape=[-1, 3, 112, 96], dtype='float32')
+        image_test3 = fluid.data(
+            name='image_test3', shape=[-1, 3, 112, 96], dtype='float32')
+        image_test4 = fluid.data(
+            name='image_test4', shape=[-1, 3, 112, 96], dtype='float32')
+        reader = fluid.io.DataLoader.from_generator(
+            feed_list=[image_test1, image_test2, image_test3, image_test4],
+            capacity=64,
+            iterable=True,
+            return_list=False)
+        reader.set_sample_list_generator(
+            test_reader,
+            places=fluid.cuda_places() if args.use_gpu else fluid.CPUPlace())
+        test_out = (fetch_targets, reader, flods, flags)
+        print('fetch_targets[0]: ', fetch_targets[0])
+        print('feed_target_names: ', feed_target_names)
+        test(exe, inference_program, test_out, args)
+    else:
+        print('WRONG ACTION')
+
+
+if __name__ == '__main__':
+    main()

paddleslim/models/__init__.py

@@ -14,5 +14,6 @@
 
 from __future__ import absolute_import
 from .util import image_classification
+from .slimfacenet import SlimFaceNet_A_x0_60, SlimFaceNet_B_x0_75, SlimFaceNet_C_x0_75
 
 __all__ = ["image_classification"]

paddleslim/models/slimfacenet.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 math
+import datetime
+import numpy as np
+
+import paddle
+import paddle.fluid as fluid
+from paddle.fluid.initializer import MSRA
+from paddle.fluid.param_attr import ParamAttr
+
+
+class SlimFaceNet():
+    def __init__(self, class_dim, scale=0.6, arch=None):
+
+        assert arch is not None
+        self.arch = arch
+        self.class_dim = class_dim
+        kernels = [3]
+        expansions = [2, 4, 6]
+        SE = [0, 1]
+        self.table = []
+        for k in kernels:
+            for e in expansions:
+                for se in SE:
+                    self.table.append((k, e, se))
+
+        if scale == 1.0:
+            # 100% - channel
+            self.Slimfacenet_bottleneck_setting = [
+                # t, c , n ,s
+                [2, 64, 5, 2],
+                [4, 128, 1, 2],
+                [2, 128, 6, 1],
+                [4, 128, 1, 2],
+                [2, 128, 2, 1]
+            ]
+        elif scale == 0.9:
+            # 90% - channel
+            self.Slimfacenet_bottleneck_setting = [
+                # t, c , n ,s
+                [2, 56, 5, 2],
+                [4, 116, 1, 2],
+                [2, 116, 6, 1],
+                [4, 116, 1, 2],
+                [2, 116, 2, 1]
+            ]
+        elif scale == 0.75:
+            # 75% - channel
+            self.Slimfacenet_bottleneck_setting = [
+                # t, c , n ,s
+                [2, 48, 5, 2],
+                [4, 96, 1, 2],
+                [2, 96, 6, 1],
+                [4, 96, 1, 2],
+                [2, 96, 2, 1]
+            ]
+        elif scale == 0.6:
+            # 60% - channel
+            self.Slimfacenet_bottleneck_setting = [
+                # t, c , n ,s
+                [2, 40, 5, 2],
+                [4, 76, 1, 2],
+                [2, 76, 6, 1],
+                [4, 76, 1, 2],
+                [2, 76, 2, 1]
+            ]
+        else:
+            print('WRONG scale')
+            exit()
+        self.extract_feature = True
+
+    def set_extract_feature_flag(self, flag):
+        self.extract_feature = flag
+
+    def net(self, input, label=None):
+        x = self.conv_bn_layer(
+            input,
+            filter_size=3,
+            num_filters=64,
+            stride=2,
+            padding=1,
+            num_groups=1,
+            if_act=True,
+            name='conv3x3')
+        x = self.conv_bn_layer(
+            x,
+            filter_size=3,
+            num_filters=64,
+            stride=1,
+            padding=1,
+            num_groups=64,
+            if_act=True,
+            name='dw_conv3x3')
+
+        in_c = 64
+        cnt = 0
+        for _exp, out_c, times, _stride in self.Slimfacenet_bottleneck_setting:
+            for i in range(times):
+                stride = _stride if i == 0 else 1
+                filter_size, exp, se = self.table[self.arch[cnt]]
+                se = False if se == 0 else True
+                x = self.residual_unit(
+                    x,
+                    num_in_filter=in_c,
+                    num_out_filter=out_c,
+                    stride=stride,
+                    filter_size=filter_size,
+                    expansion_factor=exp,
+                    use_se=se,
+                    name='residual_unit' + str(cnt + 1))
+                cnt += 1
+                in_c = out_c
+
+        out_c = 512
+        x = self.conv_bn_layer(
+            x,
+            filter_size=1,
+            num_filters=out_c,
+            stride=1,
+            padding=0,
+            num_groups=1,
+            if_act=True,
+            name='conv1x1')
+        x = self.conv_bn_layer(
+            x,
+            filter_size=(7, 6),
+            num_filters=out_c,
+            stride=1,
+            padding=0,
+            num_groups=out_c,
+            if_act=False,
+            name='global_dw_conv7x7')
+        x = fluid.layers.conv2d(
+            x,
+            num_filters=128,
+            filter_size=1,
+            stride=1,
+            padding=0,
+            groups=1,
+            act=None,
+            use_cudnn=True,
+            param_attr=ParamAttr(
+                name='linear_conv1x1_weights',
+                initializer=MSRA(),
+                regularizer=fluid.regularizer.L2Decay(4e-4)),
+            bias_attr=False)
+        bn_name = 'linear_conv1x1_bn'
+        x = fluid.layers.batch_norm(
+            x,
+            param_attr=ParamAttr(name=bn_name + "_scale"),
+            bias_attr=ParamAttr(name=bn_name + "_offset"),
+            moving_mean_name=bn_name + '_mean',
+            moving_variance_name=bn_name + '_variance')
+
+        x = fluid.layers.reshape(x, shape=[x.shape[0], x.shape[1]])
+
+        if self.extract_feature:
+            return x
+
+        out = self.arc_margin_product(
+            x, label, self.class_dim, s=32.0, m=0.50, mode=2)
+        softmax = fluid.layers.softmax(input=out)
+        cost = fluid.layers.cross_entropy(input=softmax, label=label)
+        loss = fluid.layers.mean(x=cost)
+        acc = fluid.layers.accuracy(input=out, label=label, k=1)
+        return loss, acc
+
+    def residual_unit(self,
+                      input,
+                      num_in_filter,
+                      num_out_filter,
+                      stride,
+                      filter_size,
+                      expansion_factor,
+                      use_se=False,
+                      name=None):
+
+        num_expfilter = int(round(num_in_filter * expansion_factor))
+        input_data = input
+
+        expand_conv = self.conv_bn_layer(
+            input=input,
+            filter_size=1,
+            num_filters=num_expfilter,
+            stride=1,
+            padding=0,
+            if_act=True,
+            name=name + '_expand')
+
+        depthwise_conv = self.conv_bn_layer(
+            input=expand_conv,
+            filter_size=filter_size,
+            num_filters=num_expfilter,
+            stride=stride,
+            padding=int((filter_size - 1) // 2),
+            if_act=True,
+            num_groups=num_expfilter,
+            use_cudnn=True,
+            name=name + '_depthwise')
+
+        if use_se:
+            depthwise_conv = self.se_block(
+                input=depthwise_conv,
+                num_out_filter=num_expfilter,
+                name=name + '_se')
+
+        linear_conv = self.conv_bn_layer(
+            input=depthwise_conv,
+            filter_size=1,
+            num_filters=num_out_filter,
+            stride=1,
+            padding=0,
+            if_act=False,
+            name=name + '_linear')
+        if num_in_filter != num_out_filter or stride != 1:
+            return linear_conv
+        else:
+            return fluid.layers.elementwise_add(
+                x=input_data, y=linear_conv, act=None)
+
+    def se_block(self, input, num_out_filter, ratio=4, name=None):
+        num_mid_filter = int(num_out_filter // ratio)
+        pool = fluid.layers.pool2d(
+            input=input, pool_type='avg', global_pooling=True, use_cudnn=False)
+        conv1 = fluid.layers.conv2d(
+            input=pool,
+            filter_size=1,
+            num_filters=num_mid_filter,
+            act=None,
+            param_attr=ParamAttr(name=name + '_1_weights'),
+            bias_attr=ParamAttr(name=name + '_1_offset'))
+        conv1 = fluid.layers.prelu(
+            conv1,
+            mode='channel',
+            param_attr=ParamAttr(
+                name=name + '_prelu',
+                regularizer=fluid.regularizer.L2Decay(0.0)))
+        conv2 = fluid.layers.conv2d(
+            input=conv1,
+            filter_size=1,
+            num_filters=num_out_filter,
+            act='hard_sigmoid',
+            param_attr=ParamAttr(name=name + '_2_weights'),
+            bias_attr=ParamAttr(name=name + '_2_offset'))
+        scale = fluid.layers.elementwise_mul(x=input, y=conv2, axis=0)
+        return scale
+
+    def conv_bn_layer(self,
+                      input,
+                      filter_size,
+                      num_filters,
+                      stride,
+                      padding,
+                      num_groups=1,
+                      if_act=True,
+                      name=None,
+                      use_cudnn=True):
+        conv = fluid.layers.conv2d(
+            input=input,
+            num_filters=num_filters,
+            filter_size=filter_size,
+            stride=stride,
+            padding=padding,
+            groups=num_groups,
+            act=None,
+            use_cudnn=use_cudnn,
+            param_attr=ParamAttr(
+                name=name + '_weights', initializer=MSRA()),
+            bias_attr=False)
+        bn_name = name + '_bn'
+        bn = fluid.layers.batch_norm(
+            input=conv,
+            param_attr=ParamAttr(name=bn_name + "_scale"),
+            bias_attr=ParamAttr(name=bn_name + "_offset"),
+            moving_mean_name=bn_name + '_mean',
+            moving_variance_name=bn_name + '_variance')
+        if if_act:
+            return fluid.layers.prelu(
+                bn,
+                mode='channel',
+                param_attr=ParamAttr(
+                    name=name + '_prelu',
+                    regularizer=fluid.regularizer.L2Decay(0.0)))
+        else:
+            return bn
+
+    def arc_margin_product(self, input, label, out_dim, s=32.0, m=0.50,
+                           mode=2):
+        input_norm = fluid.layers.sqrt(
+            fluid.layers.reduce_sum(
+                fluid.layers.square(input), dim=1))
+        input = fluid.layers.elementwise_div(input, input_norm, axis=0)
+
+        weight = fluid.layers.create_parameter(
+            shape=[out_dim, input.shape[1]],
+            dtype='float32',
+            name='weight_norm',
+            attr=fluid.param_attr.ParamAttr(
+                initializer=fluid.initializer.Xavier(),
+                regularizer=fluid.regularizer.L2Decay(4e-4)))
+
+        weight_norm = fluid.layers.sqrt(
+            fluid.layers.reduce_sum(
+                fluid.layers.square(weight), dim=1))
+        weight = fluid.layers.elementwise_div(weight, weight_norm, axis=0)
+        weight = fluid.layers.transpose(weight, perm=[1, 0])
+        cosine = fluid.layers.mul(input, weight)
+        sine = fluid.layers.sqrt(1.0 - fluid.layers.square(cosine))
+
+        cos_m = math.cos(m)
+        sin_m = math.sin(m)
+        phi = cosine * cos_m - sine * sin_m
+
+        th = math.cos(math.pi - m)
+        mm = math.sin(math.pi - m) * m
+
+        if mode == 1:
+            phi = self.paddle_where_more_than(cosine, 0, phi, cosine)
+        elif mode == 2:
+            phi = self.paddle_where_more_than(cosine, th, phi, cosine - mm)
+        else:
+            pass
+
+        one_hot = fluid.layers.one_hot(input=label, depth=out_dim)
+        output = fluid.layers.elementwise_mul(
+            one_hot, phi) + fluid.layers.elementwise_mul(
+                (1.0 - one_hot), cosine)
+        output = output * s
+        return output
+
+    def paddle_where_more_than(self, target, limit, x, y):
+        mask = fluid.layers.cast(x=(target > limit), dtype='float32')
+        output = fluid.layers.elementwise_mul(
+            mask, x) + fluid.layers.elementwise_mul((1.0 - mask), y)
+        return output
+
+
+def SlimFaceNet_A_x0_60(class_dim=None, scale=0.6, arch=None):
+    scale = 0.6
+    arch = [0, 1, 5, 1, 0, 2, 1, 2, 0, 1, 2, 1, 1, 0, 1]
+    return SlimFaceNet(class_dim=class_dim, scale=scale, arch=arch)
+
+
+def SlimFaceNet_B_x0_75(class_dim=None, scale=0.6, arch=None):
+    scale = 0.75
+    arch = [1, 1, 0, 1, 1, 1, 1, 0, 1, 0, 1, 3, 2, 2, 3]
+    return SlimFaceNet(class_dim=class_dim, scale=scale, arch=arch)
+
+
+def SlimFaceNet_C_x0_75(class_dim=None, scale=0.6, arch=None):
+    scale = 0.75
+    arch = [1, 1, 2, 1, 0, 2, 1, 0, 1, 0, 1, 1, 2, 2, 3]
+    return SlimFaceNet(class_dim=class_dim, scale=scale, arch=arch)
+
+
+if __name__ == "__main__":
+    x = fluid.data(name='x', shape=[-1, 3, 112, 112], dtype='float32')
+    print(x.shape)
+    model = SlimFaceNet(10000, [1, 3, 3, 1, 1, 0, 0, 1, 0, 1, 1, 0, 5, 5, 3])
+    y = model.net(x)