add ttfnet enhance (#2609)

* add ttfnet enhance

* add doc

* fix pafnet training

configs/ttfnet/README.md

-# TTFNet
+# 1. TTFNet
 
 ## 简介
 
@@ -15,6 +15,43 @@ TTFNet是一种用于实时目标检测且对训练时间友好的网络，对Ce
 | :-------------- | :------------- | :-----: | :-----: | :------------: | :-----: | :-----------------------------------------------------: | :-----: |
 | DarkNet53    | TTFNet           |    12    |   1x      |     ----     |  33.5  | [下载链接](https://paddledet.bj.bcebos.com/models/ttfnet_darknet53_1x_coco.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ttfnet/ttfnet_darknet53_1x_coco.yml) |
 
+
+
+
+
+# 2. PAFNet
+
+## 简介
+
+PAFNet（Paddle Anchor Free）是PaddleDetection基于TTFNet的优化模型，精度达到anchor free领域SOTA水平，同时产出移动端轻量级模型PAFNet-Lite
+
+PAFNet系列模型从如下方面优化TTFNet模型：
+
+- [CutMix](https://arxiv.org/abs/1905.04899)
+- 更优的骨干网络: ResNet50vd-DCN
+- 更大的训练batch size: 8 GPUs，每GPU batch_size=18
+- Synchronized Batch Normalization
+- [Deformable Convolution](https://arxiv.org/abs/1703.06211)
+- [Exponential Moving Average](https://www.investopedia.com/terms/e/ema.asp)
+- 更优的预训练模型
+
+
+## 模型库
+
+| 骨架网络        | 网络类型       | 每张GPU图片个数 | 学习率策略 |推理时间(fps) | Box AP |                           下载                          | 配置文件 |
+| :-------------- | :------------- | :-----: | :-----: | :------------: | :-----: | :-----------------------------------------------------: | :-----: |
+| ResNet50vd   | PAFNet           |    18    |   10x      |     ----     |  42.2  | [下载链接](https://paddledet.bj.bcebos.com/models/pafnet_10x_coco.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ttfnet/pafnet_10x_coco.yml) |
+
+
+
+### PAFNet-Lite
+
+| 骨架网络        | 网络类型       | 每张GPU图片个数 | 学习率策略 | Box AP | 麒麟990延时（ms） | 体积（M）                          | 下载                          | 配置文件 |
+| :-------------- | :------------- | :-----: | :-----: | :-----: | :------------: | :-----: | :-----------------------------------------------------: | :-----: |
+| MobileNetv3   |  PAFNet-Lite          |    12    |   20x     |     23.9    |  26.00   | 14 | [下载链接](https://paddledet.bj.bcebos.com/models/pafnet_lite_mobilenet_v3_20x_coco.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ttfnet/pafnet_lite_mobilenet_v3_20x_coco.yml) |
+
+
+
 ## Citations
 ```
 @article{liu2019training,

configs/ttfnet/_base_/optimizer_10x.yml

+epoch: 120
+
+LearningRate:
+  base_lr: 0.015
+  schedulers:
+  - !PiecewiseDecay
+    gamma: 0.1
+    milestones: [80, 110]
+  - !LinearWarmup
+    start_factor: 0.2
+    steps: 500
+
+OptimizerBuilder:
+  optimizer:
+    momentum: 0.9
+    type: Momentum
+  regularizer:
+    factor: 0.0004
+    type: L2

configs/ttfnet/_base_/optimizer_20x.yml

+epoch: 240
+
+LearningRate:
+  base_lr: 0.015
+  schedulers:
+  - !PiecewiseDecay
+    gamma: 0.1
+    milestones: [160, 220]
+  - !LinearWarmup
+    start_factor: 0.2
+    steps: 1000
+
+OptimizerBuilder:
+  clip_grad_by_norm: 35
+  optimizer:
+    momentum: 0.9
+    type: Momentum
+  regularizer:
+    factor: 0.0004
+    type: L2

configs/ttfnet/_base_/pafnet.yml

+architecture: TTFNet
+pretrain_weights: https://paddledet.bj.bcebos.com/models/pretrained/ResNet50_vd_ssld_pretrained.pdparams
+norm_type: sync_bn
+use_ema: true
+ema_decay: 0.9998
+
+TTFNet:
+  backbone: ResNet
+  neck: TTFFPN
+  ttf_head: TTFHead
+  post_process: BBoxPostProcess
+
+ResNet:
+  depth: 50
+  variant: d
+  return_idx: [0, 1, 2, 3]
+  freeze_at: -1
+  norm_decay: 0.
+  variant: d
+  dcn_v2_stages: [1, 2, 3]
+
+TTFFPN:
+  planes: [256, 128, 64]
+  shortcut_num: [3, 2, 1]
+
+TTFHead:
+  dcn_head: true
+  hm_loss:
+    name: CTFocalLoss
+    loss_weight: 1.
+  wh_loss:
+    name: GIoULoss
+    loss_weight: 5.
+    reduction: sum
+
+BBoxPostProcess:
+  decode:
+    name: TTFBox
+    max_per_img: 100
+    score_thresh: 0.01
+    down_ratio: 4

configs/ttfnet/_base_/pafnet_lite.yml

+architecture: TTFNet
+pretrain_weights: https://paddledet.bj.bcebos.com/models/pretrained/MobileNetV3_large_x1_0_ssld_pretrained.pdparams
+norm_type: sync_bn
+
+TTFNet:
+  backbone: MobileNetV3
+  neck: TTFFPN
+  ttf_head: TTFHead
+  post_process: BBoxPostProcess
+
+MobileNetV3:
+  scale: 1.0
+  model_name: large
+  feature_maps: [5, 8, 14, 17]
+  with_extra_blocks: true
+  lr_mult_list: [0.25, 0.25, 0.5, 0.5, 0.75]
+  conv_decay: 0.00001
+  norm_decay: 0.0
+  extra_block_filters: []
+
+TTFFPN:
+  planes: [96, 48, 24]
+  shortcut_num: [2, 2, 1]
+  lite_neck: true
+  fusion_method: concat
+
+TTFHead:
+  hm_head_planes: 48
+  wh_head_planes: 24
+  lite_head: true
+  hm_loss:
+    name: CTFocalLoss
+    loss_weight: 1.
+  wh_loss:
+    name: GIoULoss
+    loss_weight: 5.
+    reduction: sum
+
+BBoxPostProcess:
+  decode:
+    name: TTFBox
+    max_per_img: 100
+    score_thresh: 0.01
+    down_ratio: 4

configs/ttfnet/_base_/pafnet_lite_reader.yml

+worker_num: 2
+TrainReader:
+  sample_transforms:
+    - Decode: {}
+    - Cutmix: {alpha: 1.5, beta: 1.5}
+    - RandomDistort: {}
+    - RandomExpand: {fill_value: [123.675, 116.28, 103.53]}
+    - RandomCrop: {aspect_ratio: NULL, cover_all_box: True}
+    - RandomFlip: {}
+    - GridMask: {upper_iter: 300000}
+  batch_transforms:
+    - BatchRandomResize: {target_size: [320, 352, 384, 416, 448, 480, 512], random_interp: True, keep_ratio: False}
+    - NormalizeImage: {mean: [123.675, 116.28, 103.53], std: [58.395, 57.12, 57.375], is_scale: false}
+    - Permute: {}
+    - Gt2TTFTarget: {down_ratio: 4}
+    - PadBatch: {pad_to_stride: 32}
+  batch_size: 12
+  shuffle: true
+  drop_last: true
+  use_shared_memory: true
+
+EvalReader:
+  sample_transforms:
+  - Decode: {}
+  - Resize: {interp: 1, target_size: [320, 320], keep_ratio: False}
+  - NormalizeImage: {is_scale: false, mean: [123.675, 116.28, 103.53], std: [58.395, 57.12, 57.375]}
+  - Permute: {}
+  batch_size: 1
+  drop_last: false
+  drop_empty: false
+
+TestReader:
+  sample_transforms:
+  - Decode: {}
+  - Resize: {interp: 1, target_size: [320, 320], keep_ratio: False}
+  - NormalizeImage: {is_scale: false, mean: [123.675, 116.28, 103.53], std: [58.395, 57.12, 57.375]}
+  - Permute: {}
+  batch_size: 1
+  drop_last: false
+  drop_empty: false

configs/ttfnet/_base_/pafnet_reader.yml

+worker_num: 2
+TrainReader:
+  sample_transforms:
+  - Decode: {}
+  - Cutmix: {alpha: 1.5, beta: 1.5}
+  - RandomDistort: {random_apply: false, random_channel: true}
+  - RandomExpand: {fill_value: [123.675, 116.28, 103.53]}
+  - RandomCrop: {aspect_ratio: NULL, cover_all_box: True}
+  - RandomFlip: {prob: 0.5}
+  batch_transforms:
+  - BatchRandomResize: {target_size: [416, 448, 480, 512, 544, 576, 608, 640, 672], keep_ratio: false}
+  - NormalizeImage: {mean: [123.675, 116.28, 103.53], std: [58.395, 57.12, 57.375], is_scale: false}
+  - Permute: {}
+  - Gt2TTFTarget: {down_ratio: 4}
+  - PadBatch: {pad_to_stride: 32}
+  batch_size: 18
+  shuffle: true
+  drop_last: true
+  use_shared_memory: true
+  mixup_epoch: 100
+
+EvalReader:
+  sample_transforms:
+  - Decode: {}
+  - Resize: {interp: 1, target_size: [512, 512], keep_ratio: False}
+  - NormalizeImage: {is_scale: false, mean: [123.675, 116.28, 103.53], std: [58.395, 57.12, 57.375]}
+  - Permute: {}
+  batch_size: 1
+  drop_last: false
+  drop_empty: false
+
+TestReader:
+  sample_transforms:
+  - Decode: {}
+  - Resize: {interp: 1, target_size: [512, 512], keep_ratio: False}
+  - NormalizeImage: {is_scale: false, mean: [123.675, 116.28, 103.53], std: [58.395, 57.12, 57.375]}
+  - Permute: {}
+  batch_size: 1
+  drop_last: false
+  drop_empty: false

configs/ttfnet/_base_/ttfnet_darknet53.yml

@@ -14,8 +14,9 @@ DarkNet:
   norm_type: bn
   norm_decay: 0.0004
 
-# use default config
-# TTFFPN:
+TTFFPN:
+  planes: [256, 128, 64]
+  shortcut_num: [3, 2, 1]
 
 TTFHead:
   hm_loss:

configs/ttfnet/pafnet_10x_coco.yml

+_BASE_: [
+  '../datasets/coco_detection.yml',
+  '../runtime.yml',
+  '_base_/optimizer_10x.yml',
+  '_base_/pafnet.yml',
+  '_base_/pafnet_reader.yml',
+]
+weights: output/pafnet_10x_coco/model_final

configs/ttfnet/pafnet_lite_mobilenet_v3_20x_coco.yml

+_BASE_: [
+  '../datasets/coco_detection.yml',
+  '../runtime.yml',
+  '_base_/optimizer_20x.yml',
+  '_base_/pafnet_lite.yml',
+  '_base_/pafnet_lite_reader.yml',
+]
+weights: output/pafnet_lite_mobilenet_v3_10x_coco/model_final

ppdet/data/source/dataset.py

@@ -55,6 +55,7 @@ class DetDataset(Dataset):
         self.sample_num = sample_num
         self.use_default_label = use_default_label
         self._epoch = 0
+        self._curr_iter = 0
 
     def __len__(self, ):
         return len(self.roidbs)
@@ -76,6 +77,8 @@ class DetDataset(Dataset):
                 copy.deepcopy(self.roidbs[np.random.randint(n)])
                 for _ in range(3)
             ]
+        roidb['curr_iter'] = self._curr_iter
+        self._curr_iter += 1
 
         return self.transform(roidb)
 

ppdet/data/transform/batch_operators.py

@@ -533,6 +533,8 @@ class Gt2TTFTarget(BaseOperator):
             sample.pop('is_crowd')
             sample.pop('gt_class')
             sample.pop('gt_bbox')
+            if 'gt_score' in sample:
+                sample.pop('gt_score')
         return samples
 
     def draw_truncate_gaussian(self, heatmap, center, h_radius, w_radius):

ppdet/data/transform/gridmask_utils.py

@@ -20,7 +20,7 @@ import numpy as np
 from PIL import Image
 
 
-class GridMask(object):
+class Gridmask(object):
     def __init__(self,
                  use_h=True,
                  use_w=True,
@@ -30,7 +30,7 @@ class GridMask(object):
                  mode=1,
                  prob=0.7,
                  upper_iter=360000):
-        super(GridMask, self).__init__()
+        super(Gridmask, self).__init__()
         self.use_h = use_h
         self.use_w = use_w
         self.rotate = rotate

ppdet/data/transform/operators.py

@@ -308,8 +308,8 @@ class GridMask(BaseOperator):
         self.prob = prob
         self.upper_iter = upper_iter
 
-        from .gridmask_utils import GridMask
-        self.gridmask_op = GridMask(
+        from .gridmask_utils import Gridmask
+        self.gridmask_op = Gridmask(
             use_h,
             use_w,
             rotate=rotate,
@@ -1516,14 +1516,14 @@ class Cutmix(BaseOperator):
         bbx2 = np.clip(cx + cut_w // 2, 0, w - 1)
         bby2 = np.clip(cy + cut_h // 2, 0, h - 1)
 
-        img_1 = np.zeros((h, w, img1.shape[2]), 'float32')
-        img_1[:img1.shape[0], :img1.shape[1], :] = \
+        img_1_pad = np.zeros((h, w, img1.shape[2]), 'float32')
+        img_1_pad[:img1.shape[0], :img1.shape[1], :] = \
             img1.astype('float32')
-        img_2 = np.zeros((h, w, img2.shape[2]), 'float32')
-        img_2[:img2.shape[0], :img2.shape[1], :] = \
+        img_2_pad = np.zeros((h, w, img2.shape[2]), 'float32')
+        img_2_pad[:img2.shape[0], :img2.shape[1], :] = \
             img2.astype('float32')
-        img_1[bby1:bby2, bbx1:bbx2, :] = img2[bby1:bby2, bbx1:bbx2, :]
-        return img_1
+        img_1_pad[bby1:bby2, bbx1:bbx2, :] = img_2_pad[bby1:bby2, bbx1:bbx2, :]
+        return img_1_pad
 
     def __call__(self, sample, context=None):
         if not isinstance(sample, Sequence):
@@ -1546,16 +1546,27 @@ class Cutmix(BaseOperator):
         gt_class1 = sample[0]['gt_class']
         gt_class2 = sample[1]['gt_class']
         gt_class = np.concatenate((gt_class1, gt_class2), axis=0)
-        gt_score1 = sample[0]['gt_score']
-        gt_score2 = sample[1]['gt_score']
+        gt_score1 = np.ones_like(sample[0]['gt_class'])
+        gt_score2 = np.ones_like(sample[1]['gt_class'])
         gt_score = np.concatenate(
             (gt_score1 * factor, gt_score2 * (1. - factor)), axis=0)
-        sample = sample[0]
-        sample['image'] = img
-        sample['gt_bbox'] = gt_bbox
-        sample['gt_score'] = gt_score
-        sample['gt_class'] = gt_class
-        return sample
+        result = copy.deepcopy(sample[0])
+        result['image'] = img
+        result['gt_bbox'] = gt_bbox
+        result['gt_score'] = gt_score
+        result['gt_class'] = gt_class
+        if 'is_crowd' in sample[0]:
+            is_crowd1 = sample[0]['is_crowd']
+            is_crowd2 = sample[1]['is_crowd']
+            is_crowd = np.concatenate((is_crowd1, is_crowd2), axis=0)
+            result['is_crowd'] = is_crowd
+        if 'difficult' in sample[0]:
+            is_difficult1 = sample[0]['difficult']
+            is_difficult2 = sample[1]['difficult']
+            is_difficult = np.concatenate(
+                (is_difficult1, is_difficult2), axis=0)
+            result['difficult'] = is_difficult
+        return result
 
 
 @register_op

ppdet/modeling/backbones/mobilenet_v3.py

@@ -330,16 +330,16 @@ class MobileNetV3(nn.Layer):
                 [3, 16, 16, False, "relu", 1],
                 [3, 64, 24, False, "relu", 2],
                 [3, 72, 24, False, "relu", 1],
-                [5, 72, 40, True, "relu", 2],
+                [5, 72, 40, True, "relu", 2],  # RCNN output
                 [5, 120, 40, True, "relu", 1],
                 [5, 120, 40, True, "relu", 1],  # YOLOv3 output
-                [3, 240, 80, False, "hard_swish", 2],
+                [3, 240, 80, False, "hard_swish", 2],  # RCNN output
                 [3, 200, 80, False, "hard_swish", 1],
                 [3, 184, 80, False, "hard_swish", 1],
                 [3, 184, 80, False, "hard_swish", 1],
                 [3, 480, 112, True, "hard_swish", 1],
                 [3, 672, 112, True, "hard_swish", 1],  # YOLOv3 output
-                [5, 672, 160, True, "hard_swish", 2],  # SSD/SSDLite output
+                [5, 672, 160, True, "hard_swish", 2],  # SSD/SSDLite/RCNN output
                 [5, 960, 160, True, "hard_swish", 1],
                 [5, 960, 160, True, "hard_swish", 1],  # YOLOv3 output
             ]
@@ -347,14 +347,14 @@ class MobileNetV3(nn.Layer):
             self.cfg = [
                 # k, exp, c,  se,     nl,  s,
                 [3, 16, 16, True, "relu", 2],
-                [3, 72, 24, False, "relu", 2],
+                [3, 72, 24, False, "relu", 2],  # RCNN output
                 [3, 88, 24, False, "relu", 1],  # YOLOv3 output
-                [5, 96, 40, True, "hard_swish", 2],
+                [5, 96, 40, True, "hard_swish", 2],  # RCNN output
                 [5, 240, 40, True, "hard_swish", 1],
                 [5, 240, 40, True, "hard_swish", 1],
                 [5, 120, 48, True, "hard_swish", 1],
                 [5, 144, 48, True, "hard_swish", 1],  # YOLOv3 output
-                [5, 288, 96, True, "hard_swish", 2],  # SSD/SSDLite output
+                [5, 288, 96, True, "hard_swish", 2],  # SSD/SSDLite/RCNN output
                 [5, 576, 96, True, "hard_swish", 1],
                 [5, 576, 96, True, "hard_swish", 1],  # YOLOv3 output
             ]

ppdet/modeling/heads/ttf_head.py

@@ -19,6 +19,7 @@ from paddle import ParamAttr
 from paddle.nn.initializer import Constant, Uniform, Normal
 from paddle.regularizer import L2Decay
 from ppdet.core.workspace import register
+from ppdet.modeling.layers import DeformableConvV2, LiteConv
 import numpy as np
 
 
@@ -30,16 +31,50 @@ class HMHead(nn.Layer):
         ch_out (int): The channel number of output Tensor.
         num_classes (int): Number of classes.
         conv_num (int): The convolution number of hm_feat.
+        dcn_head(bool): whether use dcn in head. False by default. 
+        lite_head(bool): whether use lite version. False by default.
+        norm_type (string): norm type, 'sync_bn', 'bn', 'gn' are optional.
+            bn by default
+
     Return:
         Heatmap head output
     """
-    __shared__ = ['num_classes']
+    __shared__ = ['num_classes', 'norm_type']
 
-    def __init__(self, ch_in, ch_out=128, num_classes=80, conv_num=2):
+    def __init__(
+            self,
+            ch_in,
+            ch_out=128,
+            num_classes=80,
+            conv_num=2,
+            dcn_head=False,
+            lite_head=False,
+            norm_type='bn', ):
         super(HMHead, self).__init__()
         head_conv = nn.Sequential()
         for i in range(conv_num):
             name = 'conv.{}'.format(i)
+            if lite_head:
+                lite_name = 'hm.' + name
+                head_conv.add_sublayer(
+                    lite_name,
+                    LiteConv(
+                        in_channels=ch_in if i == 0 else ch_out,
+                        out_channels=ch_out,
+                        norm_type=norm_type,
+                        name=lite_name))
+                head_conv.add_sublayer(lite_name + '.act', nn.ReLU6())
+            else:
+                if dcn_head:
+                    head_conv.add_sublayer(
+                        name,
+                        DeformableConvV2(
+                            in_channels=ch_in if i == 0 else ch_out,
+                            out_channels=ch_out,
+                            kernel_size=3,
+                            weight_attr=ParamAttr(initializer=Normal(0, 0.01)),
+                            name='hm.' + name))
+                else:
                     head_conv.add_sublayer(
                         name,
                         nn.Conv2D(
@@ -78,15 +113,47 @@ class WHHead(nn.Layer):
         ch_in (int): The channel number of input Tensor.
         ch_out (int): The channel number of output Tensor.
         conv_num (int): The convolution number of wh_feat.
+        dcn_head(bool): whether use dcn in head. False by default.
+        lite_head(bool): whether use lite version. False by default.
+        norm_type (string): norm type, 'sync_bn', 'bn', 'gn' are optional.
+            bn by default
     Return:
         Width & Height head output
     """
+    __shared__ = ['norm_type']
 
-    def __init__(self, ch_in, ch_out=64, conv_num=2):
+    def __init__(self,
+                 ch_in,
+                 ch_out=64,
+                 conv_num=2,
+                 dcn_head=False,
+                 lite_head=False,
+                 norm_type='bn'):
         super(WHHead, self).__init__()
         head_conv = nn.Sequential()
         for i in range(conv_num):
             name = 'conv.{}'.format(i)
+            if lite_head:
+                lite_name = 'wh.' + name
+                head_conv.add_sublayer(
+                    lite_name,
+                    LiteConv(
+                        in_channels=ch_in if i == 0 else ch_out,
+                        out_channels=ch_out,
+                        norm_type=norm_type,
+                        name=lite_name))
+                head_conv.add_sublayer(lite_name + '.act', nn.ReLU6())
+            else:
+                if dcn_head:
+                    head_conv.add_sublayer(
+                        name,
+                        DeformableConvV2(
+                            in_channels=ch_in if i == 0 else ch_out,
+                            out_channels=ch_out,
+                            kernel_size=3,
+                            weight_attr=ParamAttr(initializer=Normal(0, 0.01)),
+                            name='wh.' + name))
+                else:
                     head_conv.add_sublayer(
                         name,
                         nn.Conv2D(
@@ -94,10 +161,11 @@ class WHHead(nn.Layer):
                             out_channels=ch_out,
                             kernel_size=3,
                             padding=1,
-                    weight_attr=ParamAttr(initializer=Normal(0, 0.001)),
+                            weight_attr=ParamAttr(initializer=Normal(0, 0.01)),
                             bias_attr=ParamAttr(
                                 learning_rate=2., regularizer=L2Decay(0.))))
                 head_conv.add_sublayer(name + '.act', nn.ReLU())
+
         self.feat = self.add_sublayer('wh_feat', head_conv)
         self.head = self.add_sublayer(
             'wh_head',
@@ -137,9 +205,12 @@ class TTFHead(nn.Layer):
             16.0 by default.
         down_ratio (int): the actual down_ratio is calculated by base_down_ratio
             (default 16) and the number of upsample layers.
+        lite_head(bool): whether use lite version. False by default.
+        norm_type (string): norm type, 'sync_bn', 'bn', 'gn' are optional.
+            bn by default
     """
 
-    __shared__ = ['num_classes', 'down_ratio']
+    __shared__ = ['num_classes', 'down_ratio', 'norm_type']
     __inject__ = ['hm_loss', 'wh_loss']
 
     def __init__(self,
@@ -152,12 +223,16 @@ class TTFHead(nn.Layer):
                  hm_loss='CTFocalLoss',
                  wh_loss='GIoULoss',
                  wh_offset_base=16.,
-                 down_ratio=4):
+                 down_ratio=4,
+                 dcn_head=False,
+                 lite_head=False,
+                 norm_type='bn'):
         super(TTFHead, self).__init__()
         self.in_channels = in_channels
         self.hm_head = HMHead(in_channels, hm_head_planes, num_classes,
-                              hm_head_conv_num)
-        self.wh_head = WHHead(in_channels, wh_head_planes, wh_head_conv_num)
+                              hm_head_conv_num, dcn_head, lite_head, norm_type)
+        self.wh_head = WHHead(in_channels, wh_head_planes, wh_head_conv_num,
+                              dcn_head, lite_head, norm_type)
         self.hm_loss = hm_loss
         self.wh_loss = wh_loss
 

ppdet/modeling/layers.py

@@ -23,7 +23,7 @@ from paddle import ParamAttr
 from paddle import to_tensor
 from paddle.nn import Conv2D, BatchNorm2D, GroupNorm
 import paddle.nn.functional as F
-from paddle.nn.initializer import Normal, Constant
+from paddle.nn.initializer import Normal, Constant, XavierUniform
 from paddle.regularizer import L2Decay
 
 from ppdet.core.workspace import register, serializable
@@ -112,6 +112,7 @@ class ConvNormLayer(nn.Layer):
                  ch_out,
                  filter_size,
                  stride,
+                 groups=1,
                  norm_type='bn',
                  norm_decay=0.,
                  norm_groups=32,
@@ -142,7 +143,7 @@ class ConvNormLayer(nn.Layer):
                 kernel_size=filter_size,
                 stride=stride,
                 padding=(filter_size - 1) // 2,
-                groups=1,
+                groups=groups,
                 weight_attr=ParamAttr(
                     name=name + "_weight",
                     initializer=initializer,
@@ -158,7 +159,7 @@ class ConvNormLayer(nn.Layer):
                 kernel_size=filter_size,
                 stride=stride,
                 padding=(filter_size - 1) // 2,
-                groups=1,
+                groups=groups,
                 weight_attr=ParamAttr(
                     name=name + "_weight",
                     initializer=initializer,
@@ -197,6 +198,71 @@ class ConvNormLayer(nn.Layer):
         return out
 
 
+class LiteConv(nn.Layer):
+    def __init__(self,
+                 in_channels,
+                 out_channels,
+                 stride=1,
+                 with_act=True,
+                 norm_type='sync_bn',
+                 name=None):
+        super(LiteConv, self).__init__()
+        self.lite_conv = nn.Sequential()
+        conv1 = ConvNormLayer(
+            in_channels,
+            in_channels,
+            filter_size=5,
+            stride=stride,
+            groups=in_channels,
+            norm_type=norm_type,
+            initializer=XavierUniform(),
+            norm_name=name + '.conv1.norm',
+            name=name + '.conv1')
+        conv2 = ConvNormLayer(
+            in_channels,
+            out_channels,
+            filter_size=1,
+            stride=stride,
+            norm_type=norm_type,
+            initializer=XavierUniform(),
+            norm_name=name + '.conv2.norm',
+            name=name + '.conv2')
+        conv3 = ConvNormLayer(
+            out_channels,
+            out_channels,
+            filter_size=1,
+            stride=stride,
+            norm_type=norm_type,
+            initializer=XavierUniform(),
+            norm_name=name + '.conv3.norm',
+            name=name + '.conv3')
+        conv4 = ConvNormLayer(
+            out_channels,
+            out_channels,
+            filter_size=5,
+            stride=stride,
+            groups=out_channels,
+            norm_type=norm_type,
+            initializer=XavierUniform(),
+            norm_name=name + '.conv4.norm',
+            name=name + '.conv4')
+        conv_list = [conv1, conv2, conv3, conv4]
+        self.lite_conv.add_sublayer('conv1', conv1)
+        self.lite_conv.add_sublayer('relu6_1', nn.ReLU6())
+        self.lite_conv.add_sublayer('conv2', conv2)
+        if with_act:
+            self.lite_conv.add_sublayer('relu6_2', nn.ReLU6())
+        self.lite_conv.add_sublayer('conv3', conv3)
+        self.lite_conv.add_sublayer('relu6_3', nn.ReLU6())
+        self.lite_conv.add_sublayer('conv4', conv4)
+        if with_act:
+            self.lite_conv.add_sublayer('relu6_4', nn.ReLU6())
+
+    def forward(self, inputs):
+        out = self.lite_conv(inputs)
+        return out
+
+
 @register
 @serializable
 class AnchorGeneratorRPN(object):

ppdet/modeling/necks/ttf_fpn.py

@@ -16,11 +16,11 @@ import paddle
 import paddle.nn as nn
 import paddle.nn.functional as F
 from paddle import ParamAttr
-from paddle.nn.initializer import Constant, Uniform, Normal
+from paddle.nn.initializer import Constant, Uniform, Normal, XavierUniform
 from paddle import ParamAttr
 from ppdet.core.workspace import register, serializable
 from paddle.regularizer import L2Decay
-from ppdet.modeling.layers import DeformableConvV2
+from ppdet.modeling.layers import DeformableConvV2, ConvNormLayer, LiteConv
 import math
 from ppdet.modeling.ops import batch_norm
 from ..shape_spec import ShapeSpec
@@ -29,7 +29,7 @@ __all__ = ['TTFFPN']
 
 
 class Upsample(nn.Layer):
-    def __init__(self, ch_in, ch_out, name=None):
+    def __init__(self, ch_in, ch_out, norm_type='bn', name=None):
         super(Upsample, self).__init__()
         fan_in = ch_in * 3 * 3
         stdv = 1. / math.sqrt(fan_in)
@@ -46,7 +46,7 @@ class Upsample(nn.Layer):
             regularizer=L2Decay(0.))
 
         self.bn = batch_norm(
-            ch_out, norm_type='bn', initializer=Constant(1.), name=name)
+            ch_out, norm_type=norm_type, initializer=Constant(1.), name=name)
 
     def forward(self, feat):
         dcn = self.dcn(feat)
@@ -56,16 +56,92 @@ class Upsample(nn.Layer):
         return out
 
 
+class DeConv(nn.Layer):
+    def __init__(self, ch_in, ch_out, norm_type='bn', name=None):
+        super(DeConv, self).__init__()
+        self.deconv = nn.Sequential()
+        conv1 = ConvNormLayer(
+            ch_in=ch_in,
+            ch_out=ch_out,
+            stride=1,
+            filter_size=1,
+            norm_type=norm_type,
+            initializer=XavierUniform(),
+            norm_name=name + '.conv1.norm',
+            name=name + '.conv1')
+        conv2 = nn.Conv2DTranspose(
+            in_channels=ch_out,
+            out_channels=ch_out,
+            kernel_size=4,
+            padding=1,
+            stride=2,
+            groups=ch_out,
+            weight_attr=ParamAttr(initializer=XavierUniform()),
+            bias_attr=False)
+        bn = batch_norm(
+            ch_out, norm_type=norm_type, norm_decay=0., name=name + '.bn')
+        conv3 = ConvNormLayer(
+            ch_in=ch_out,
+            ch_out=ch_out,
+            stride=1,
+            filter_size=1,
+            norm_type=norm_type,
+            initializer=XavierUniform(),
+            norm_name=name + '.conv3.norm',
+            name=name + '.conv3')
+
+        self.deconv.add_sublayer('conv1', conv1)
+        self.deconv.add_sublayer('relu6_1', nn.ReLU6())
+        self.deconv.add_sublayer('conv2', conv2)
+        self.deconv.add_sublayer('bn', bn)
+        self.deconv.add_sublayer('relu6_2', nn.ReLU6())
+        self.deconv.add_sublayer('conv3', conv3)
+        self.deconv.add_sublayer('relu6_3', nn.ReLU6())
+
+    def forward(self, inputs):
+        return self.deconv(inputs)
+
+
+class LiteUpsample(nn.Layer):
+    def __init__(self, ch_in, ch_out, norm_type='bn', name=None):
+        super(LiteUpsample, self).__init__()
+        self.deconv = DeConv(
+            ch_in, ch_out, norm_type=norm_type, name=name + '.deconv')
+        self.conv = LiteConv(
+            ch_in, ch_out, norm_type=norm_type, name=name + '.liteconv')
+
+    def forward(self, inputs):
+        deconv_up = self.deconv(inputs)
+        conv = self.conv(inputs)
+        interp_up = F.interpolate(conv, scale_factor=2., mode='bilinear')
+        return deconv_up + interp_up
+
+
 class ShortCut(nn.Layer):
-    def __init__(self, layer_num, ch_out, name=None):
+    def __init__(self,
+                 layer_num,
+                 ch_in,
+                 ch_out,
+                 norm_type='bn',
+                 lite_neck=False,
+                 name=None):
         super(ShortCut, self).__init__()
         shortcut_conv = nn.Sequential()
-        ch_in = ch_out * 2
         for i in range(layer_num):
             fan_out = 3 * 3 * ch_out
             std = math.sqrt(2. / fan_out)
             in_channels = ch_in if i == 0 else ch_out
             shortcut_name = name + '.conv.{}'.format(i)
+            if lite_neck:
+                shortcut_conv.add_sublayer(
+                    shortcut_name,
+                    LiteConv(
+                        in_channels=in_channels,
+                        out_channels=ch_out,
+                        with_act=i < layer_num - 1,
+                        norm_type=norm_type,
+                        name=shortcut_name))
+            else:
                 shortcut_conv.add_sublayer(
                     shortcut_name,
                     nn.Conv2D(
@@ -77,7 +153,8 @@ class ShortCut(nn.Layer):
                         bias_attr=ParamAttr(
                             learning_rate=2., regularizer=L2Decay(0.))))
                 if i < layer_num - 1:
-                shortcut_conv.add_sublayer(shortcut_name + '.act', nn.ReLU())
+                    shortcut_conv.add_sublayer(shortcut_name + '.act',
+                                               nn.ReLU())
         self.shortcut = self.add_sublayer('short', shortcut_conv)
 
     def forward(self, feat):
@@ -93,35 +170,68 @@ class TTFFPN(nn.Layer):
         in_channels (list): number of input feature channels from backbone.
             [128,256,512,1024] by default, means the channels of DarkNet53
             backbone return_idx [1,2,3,4].
+        planes (list): the number of output feature channels of FPN.
+            [256, 128, 64] by default
         shortcut_num (list): the number of convolution layers in each shortcut.
             [3,2,1] by default, means DarkNet53 backbone return_idx_1 has 3 convs
             in its shortcut, return_idx_2 has 2 convs and return_idx_3 has 1 conv.
+        norm_type (string): norm type, 'sync_bn', 'bn', 'gn' are optional. 
+            bn by default
+        lite_neck (bool): whether to use lite conv in TTFNet FPN, 
+            False by default
+        fusion_method (string): the method to fusion upsample and lateral layer.
+            'add' and 'concat' are optional, add by default
     """
 
+    __shared__ = ['norm_type']
+
     def __init__(self,
-                 in_channels=[128, 256, 512, 1024],
-                 shortcut_num=[3, 2, 1]):
+                 in_channels,
+                 planes=[256, 128, 64],
+                 shortcut_num=[3, 2, 1],
+                 norm_type='bn',
+                 lite_neck=False,
+                 fusion_method='add'):
         super(TTFFPN, self).__init__()
-        self.planes = [c // 2 for c in in_channels[:-1]][::-1]
+        self.planes = planes
         self.shortcut_num = shortcut_num[::-1]
         self.shortcut_len = len(shortcut_num)
         self.ch_in = in_channels[::-1]
+        self.fusion_method = fusion_method
 
         self.upsample_list = []
         self.shortcut_list = []
+        self.upper_list = []
         for i, out_c in enumerate(self.planes):
-            in_c = self.ch_in[i] if i == 0 else self.ch_in[i] // 2
+            in_c = self.ch_in[i] if i == 0 else self.upper_list[-1]
+            upsample_module = LiteUpsample if lite_neck else Upsample
             upsample = self.add_sublayer(
                 'upsample.' + str(i),
-                Upsample(
-                    in_c, out_c, name='upsample.' + str(i)))
+                upsample_module(
+                    in_c,
+                    out_c,
+                    norm_type=norm_type,
+                    name='deconv_layers.' + str(i)))
             self.upsample_list.append(upsample)
             if i < self.shortcut_len:
                 shortcut = self.add_sublayer(
                     'shortcut.' + str(i),
                     ShortCut(
-                        self.shortcut_num[i], out_c, name='shortcut.' + str(i)))
+                        self.shortcut_num[i],
+                        self.ch_in[i + 1],
+                        out_c,
+                        norm_type=norm_type,
+                        lite_neck=lite_neck,
+                        name='shortcut.' + str(i)))
                 self.shortcut_list.append(shortcut)
+                if self.fusion_method == 'add':
+                    upper_c = out_c
+                elif self.fusion_method == 'concat':
+                    upper_c = out_c * 2
+                else:
+                    raise ValueError('Illegal fusion method. Expected add or\
+                        concat, but received {}'.format(self.fusion_method))
+                self.upper_list.append(upper_c)
 
     def forward(self, inputs):
         feat = inputs[-1]
@@ -129,7 +239,10 @@ class TTFFPN(nn.Layer):
             feat = self.upsample_list[i](feat)
             if i < self.shortcut_len:
                 shortcut = self.shortcut_list[i](inputs[-i - 2])
+                if self.fusion_method == 'add':
                     feat = feat + shortcut
+                else:
+                    feat = paddle.concat([feat, shortcut], axis=1)
         return feat
 
     @classmethod
@@ -138,4 +251,4 @@ class TTFFPN(nn.Layer):
 
     @property
     def out_shape(self):
-        return [ShapeSpec(channels=self.planes[-1], )]
+        return [ShapeSpec(channels=self.upper_list[-1], )]

static/configs/anchor_free/pafnet_10x_coco.yml

+architecture: TTFNet
+use_gpu: true
+max_iters: 150000
+log_smooth_window: 20
+save_dir: output
+snapshot_iter: 10000
+metric: COCO
+pretrain_weights: https://paddle-imagenet-models-name.bj.bcebos.com/ResNet50_vd_ssld_pretrained.tar
+weights: output/pafnet_10x_coco/model_final
+num_classes: 80
+use_ema: true
+ema_decay: 0.9998
+
+TTFNet:
+  backbone: ResNet
+  ttf_head: TTFHead
+
+ResNet:
+  norm_type: sync_bn
+  freeze_at: 0
+  freeze_norm: false
+  norm_decay: 0.
+  depth: 50
+  feature_maps: [2, 3, 4, 5]
+  variant: d
+  dcn_v2_stages: [3, 4, 5]
+
+TTFHead:
+  head_conv: 128
+  wh_conv: 64
+  hm_head_conv_num: 2
+  wh_head_conv_num: 2
+  wh_offset_base: 16
+  wh_loss: GiouLoss
+  dcn_head: True
+
+GiouLoss:
+  loss_weight: 5.
+  do_average: false
+  use_class_weight: false
+
+LearningRate:
+  base_lr: 0.015
+  schedulers:
+  - !PiecewiseDecay
+    gamma: 0.1
+    milestones:
+    - 112500
+    - 137500
+  - !LinearWarmup
+    start_factor: 0.2
+    steps: 500
+
+OptimizerBuilder:
+  optimizer:
+    momentum: 0.9
+    type: Momentum
+  regularizer:
+    factor: 0.0004
+    type: L2
+
+TrainReader:
+  inputs_def:
+    fields: ['image', 'ttf_heatmap', 'ttf_box_target', 'ttf_reg_weight']
+  dataset:
+    !COCODataSet
+    image_dir: train2017
+    anno_path: annotations/instances_train2017.json
+    dataset_dir: dataset/coco
+    with_background: false
+  sample_transforms:
+  - !DecodeImage
+    to_rgb: true
+    with_cutmix: True
+  - !CutmixImage
+    alpha: 1.5
+    beta: 1.5
+  - !ColorDistort
+    hue: [-18., 18., 0.5]
+    saturation: [0.5, 1.5, 0.5]
+    contrast: [0.5, 1.5, 0.5]
+    brightness: [-32., 32., 0.5]
+    random_apply: False
+    hsv_format: True
+    random_channel: True
+  - !RandomExpand
+    ratio: 4
+    prob: 0.5
+    fill_value: [123.675, 116.28, 103.53]
+  - !RandomCrop
+    aspect_ratio: NULL
+    cover_all_box: True
+  - !RandomFlipImage
+    prob: 0.5
+  batch_transforms:
+  - !RandomShape
+    sizes: [416, 448, 480, 512, 544, 576, 608, 640, 672]
+    random_inter: True
+    resize_box: True
+  - !NormalizeImage
+    is_channel_first: false
+    is_scale: false
+    mean: [123.675, 116.28, 103.53]
+    std: [58.395, 57.12, 57.375]
+  - !Permute
+    to_bgr: false
+    channel_first: true
+  - !Gt2TTFTarget
+    num_classes: 80
+    down_ratio: 4
+  - !PadBatch
+    pad_to_stride: 32
+  batch_size: 12
+  shuffle: true
+  worker_num: 8
+  bufsize: 2
+  use_process: false
+  cutmix_epoch: 100
+
+EvalReader:
+  inputs_def:
+    image_shape: [3, 512, 512]
+    fields: ['image', 'im_id', 'scale_factor']
+  dataset:
+    !COCODataSet
+      image_dir: val2017
+      anno_path: annotations/instances_val2017.json
+      dataset_dir: dataset/coco
+      with_background: false
+  sample_transforms:
+    - !DecodeImage
+      to_rgb: True
+    - !Resize
+      target_dim: 512
+    - !NormalizeImage
+      mean: [123.675, 116.28, 103.53]
+      std: [58.395, 57.12, 57.375]
+      is_scale: false
+      is_channel_first: false
+    - !Permute
+      to_bgr: false
+      channel_first: True
+  batch_size: 1
+  drop_empty: false
+  worker_num: 8
+  bufsize: 16
+
+TestReader:
+  inputs_def:
+    image_shape: [3, 512, 512]
+    fields: ['image', 'im_id', 'scale_factor']
+  dataset:
+    !ImageFolder
+      anno_path: annotations/instances_val2017.json
+      with_background: false
+  sample_transforms:
+    - !DecodeImage
+      to_rgb: True
+    - !Resize
+      interp: 1
+      target_dim: 512
+    - !NormalizeImage
+      mean: [123.675, 116.28, 103.53]
+      std: [58.395, 57.12, 57.375]
+      is_scale: false
+      is_channel_first: false
+    - !Permute
+      to_bgr: false
+      channel_first: True
+  batch_size: 1

static/configs/anchor_free/pafnet_lite_mobilenet_v3_20x_coco.yml

+architecture: TTFNet
+use_gpu: true
+max_iters: 300000
+log_smooth_window: 20
+save_dir: output
+snapshot_iter: 50000
+metric: COCO
+pretrain_weights: https://paddle-imagenet-models-name.bj.bcebos.com/MobileNetV3_large_x1_0_ssld_pretrained.tar
+weights: output/pafnet_lite_mobilenet_v3_20x_coco/model_final
+num_classes: 80
+
+TTFNet:
+  backbone: MobileNetV3RCNN
+  ttf_head: TTFLiteHead
+
+MobileNetV3RCNN:
+  norm_type: sync_bn
+  norm_decay: 0.0
+  model_name: large
+  scale: 1.0
+  conv_decay: 0.00001
+  lr_mult_list: [0.25, 0.25, 0.5, 0.5, 0.75]
+  freeze_norm: false
+
+TTFLiteHead:
+  head_conv: 48
+
+GiouLoss:
+  loss_weight: 5.
+  do_average: false
+  use_class_weight: false
+
+LearningRate:
+  base_lr: 0.015
+  schedulers:
+  - !PiecewiseDecay
+    gamma: 0.1
+    milestones:
+    - 225000
+    - 275000
+  - !LinearWarmup
+    start_factor: 0.2
+    steps: 1000
+
+OptimizerBuilder:
+  clip_grad_by_norm: 35
+  optimizer:
+    momentum: 0.9
+    type: Momentum
+  regularizer:
+    factor: 0.0004
+    type: L2
+
+TrainReader:
+  inputs_def:
+    fields: ['image', 'ttf_heatmap', 'ttf_box_target', 'ttf_reg_weight']
+  dataset:
+    !COCODataSet
+    image_dir: train2017
+    anno_path: annotations/instances_train2017.json
+    dataset_dir: dataset/coco
+    with_background: false
+  sample_transforms:
+  - !DecodeImage
+    to_rgb: true
+    with_cutmix: True
+  - !ColorDistort
+    hue: [-18., 18., 0.5]
+    saturation: [0.5, 1.5, 0.5]
+    contrast: [0.5, 1.5, 0.5]
+    brightness: [-32., 32., 0.5]
+    random_apply: False
+    hsv_format: False
+    random_channel: True
+  - !RandomExpand
+    ratio: 4
+    prob: 0.5
+    fill_value: [123.675, 116.28, 103.53]
+  - !RandomCrop
+    aspect_ratio: NULL
+    cover_all_box: True
+  - !CutmixImage
+    alpha: 1.5
+    beta: 1.5
+  - !RandomFlipImage
+    prob: 0.5
+  - !GridMaskOp
+    use_h: true
+    use_w: true
+    rotate: 1
+    offset: false
+    ratio: 0.5
+    mode: 1
+    prob: 0.7
+    upper_iter: 300000
+  batch_transforms:
+  - !RandomShape
+    sizes: [320, 352, 384, 416, 448, 480, 512]
+    random_inter: True
+    resize_box: True
+  - !NormalizeImage
+    is_channel_first: false
+    is_scale: false
+    mean: [123.675, 116.28, 103.53]
+    std: [58.395, 57.12, 57.375]
+  - !Permute
+    to_bgr: false
+    channel_first: true
+  - !Gt2TTFTarget
+    num_classes: 80
+    down_ratio: 4
+  - !PadBatch
+    pad_to_stride: 32
+  batch_size: 12
+  shuffle: true
+  worker_num: 8
+  bufsize: 2
+  use_process: false
+  cutmix_epoch: 200
+
+EvalReader:
+  inputs_def:
+    image_shape: [3, 320, 320]
+    fields: ['image', 'im_id', 'scale_factor']
+  dataset:
+    !COCODataSet
+      image_dir: val2017
+      anno_path: annotations/instances_val2017.json
+      dataset_dir: dataset/coco
+      with_background: false
+  sample_transforms:
+    - !DecodeImage
+      to_rgb: True
+    - !Resize
+      target_dim: 320
+    - !NormalizeImage
+      mean: [123.675, 116.28, 103.53]
+      std: [58.395, 57.12, 57.375]
+      is_scale: false
+      is_channel_first: false
+    - !Permute
+      to_bgr: false
+      channel_first: True
+  batch_size: 1
+  drop_empty: false
+  worker_num: 2
+  bufsize: 2
+
+TestReader:
+  inputs_def:
+    image_shape: [3, 320, 320]
+    fields: ['image', 'im_id', 'scale_factor']
+  dataset:
+    !ImageFolder
+      anno_path: annotations/instances_val2017.json
+      with_background: false
+  sample_transforms:
+    - !DecodeImage
+      to_rgb: True
+    - !Resize
+      interp: 1
+      target_dim: 320
+    - !NormalizeImage
+      mean: [123.675, 116.28, 103.53]
+      std: [58.395, 57.12, 57.375]
+      is_scale: false
+      is_channel_first: false
+    - !Permute
+      to_bgr: false
+      channel_first: True
+  batch_size: 1

static/ppdet/modeling/anchor_heads/ttf_head.py

@@ -24,10 +24,10 @@ from paddle.fluid.param_attr import ParamAttr
 from paddle.fluid.initializer import Normal, Constant, Uniform, Xavier
 from paddle.fluid.regularizer import L2Decay
 from ppdet.core.workspace import register
-from ppdet.modeling.ops import DeformConv, DropBlock
+from ppdet.modeling.ops import DeformConv, DropBlock, ConvNorm
 from ppdet.modeling.losses import GiouLoss
 
-__all__ = ['TTFHead']
+__all__ = ['TTFHead', 'TTFLiteHead']
 
 
 @register
@@ -65,6 +65,8 @@ class TTFHead(object):
         drop_block(bool): whether use dropblock. False by default.
         block_size(int): block_size parameter for drop_block. 3 by default.
         keep_prob(float): keep_prob parameter for drop_block. 0.9 by default.
+        fusion_method (string): Method to fusion upsample and lateral branch.
+            'add' and 'concat' are optional, add by default
     """
 
     __inject__ = ['wh_loss']
@@ -90,7 +92,8 @@ class TTFHead(object):
                  dcn_head=False,
                  drop_block=False,
                  block_size=3,
-                 keep_prob=0.9):
+                 keep_prob=0.9,
+                 fusion_method='add'):
         super(TTFHead, self).__init__()
         self.head_conv = head_conv
         self.num_classes = num_classes
@@ -115,6 +118,7 @@ class TTFHead(object):
         self.drop_block = drop_block
         self.block_size = block_size
         self.keep_prob = keep_prob
+        self.fusion_method = fusion_method
 
     def shortcut(self, x, out_c, layer_num, kernel_size=3, padding=1,
                  name=None):
@@ -255,7 +259,14 @@ class TTFHead(object):
                     out_c,
                     self.shortcut_num[i],
                     name=name + '.shortcut_layers.' + str(i))
+                if self.fusion_method == 'add':
                     feat = fluid.layers.elementwise_add(feat, shortcut)
+                elif self.fusion_method == 'concat':
+                    feat = fluid.layers.concat([feat, shortcut], axis=1)
+                else:
+                    raise ValueError(
+                        "Illegal fusion method, expected 'add' or 'concat', but received {}".
+                        format(self.fusion_method))
 
         hm = self.hm_head(feat, name=name + '.hm', is_test=is_test)
         wh = self.wh_head(feat, name=name + '.wh') * self.wh_offset_base
@@ -273,12 +284,13 @@ class TTFHead(object):
         # batch size is 1
         scores_r = fluid.layers.reshape(scores, [cat, -1])
         topk_scores, topk_inds = fluid.layers.topk(scores_r, k)
-        topk_ys = topk_inds / width
+        topk_ys = topk_inds // width
         topk_xs = topk_inds % width
 
         topk_score_r = fluid.layers.reshape(topk_scores, [-1])
         topk_score, topk_ind = fluid.layers.topk(topk_score_r, k)
-        topk_clses = fluid.layers.cast(topk_ind / k, 'float32')
+        k_t = fluid.layers.assign(np.array([k], dtype='int64'))
+        topk_clses = fluid.layers.cast(topk_ind / k_t, 'float32')
 
         topk_inds = fluid.layers.reshape(topk_inds, [-1])
         topk_ys = fluid.layers.reshape(topk_ys, [-1, 1])
@@ -384,3 +396,172 @@ class TTFHead(object):
 
         ttf_loss = {'hm_loss': hm_loss, 'wh_loss': wh_loss}
         return ttf_loss
+
+
+@register
+class TTFLiteHead(TTFHead):
+    """
+    TTFLiteHead
+
+    Lite version for TTFNet
+    Args:
+        head_conv(int): the default channel number of convolution in head.
+            32 by default.
+        num_classes(int): the number of classes, 80 by default.
+        planes(tuple): the channel number of convolution in each upsample.
+            (96, 48, 24) by default.
+        wh_conv(int): the channel number of convolution in wh head.
+            24 by default.
+        wh_loss(object): `GiouLoss` instance.
+        shortcut_num(tuple): the number of convolution layers in each shortcut.
+            (1, 2, 2) by default.
+        fusion_method (string): Method to fusion upsample and lateral branch.
+            'add' and 'concat' are optional, add by default
+    """
+    __inject__ = ['wh_loss']
+    __shared__ = ['num_classes']
+
+    def __init__(self,
+                 head_conv=32,
+                 num_classes=80,
+                 planes=(96, 48, 24),
+                 wh_conv=24,
+                 wh_loss='GiouLoss',
+                 shortcut_num=(1, 2, 2),
+                 fusion_method='concat'):
+        super(TTFLiteHead, self).__init__(
+            head_conv=head_conv,
+            num_classes=num_classes,
+            planes=planes,
+            wh_conv=wh_conv,
+            wh_loss=wh_loss,
+            shortcut_num=shortcut_num,
+            fusion_method=fusion_method)
+
+    def _lite_conv(self, x, out_c, act=None, name=None):
+        conv1 = ConvNorm(
+            input=x,
+            num_filters=x.shape[1],
+            filter_size=5,
+            groups=x.shape[1],
+            norm_type='bn',
+            act='relu6',
+            initializer=Xavier(),
+            name=name + '.depthwise',
+            norm_name=name + '.depthwise.bn')
+
+        conv2 = ConvNorm(
+            input=conv1,
+            num_filters=out_c,
+            filter_size=1,
+            norm_type='bn',
+            act=act,
+            initializer=Xavier(),
+            name=name + '.pointwise_linear',
+            norm_name=name + '.pointwise_linear.bn')
+
+        conv3 = ConvNorm(
+            input=conv2,
+            num_filters=out_c,
+            filter_size=1,
+            norm_type='bn',
+            act='relu6',
+            initializer=Xavier(),
+            name=name + '.pointwise',
+            norm_name=name + '.pointwise.bn')
+
+        conv4 = ConvNorm(
+            input=conv3,
+            num_filters=out_c,
+            filter_size=5,
+            groups=out_c,
+            norm_type='bn',
+            act=act,
+            initializer=Xavier(),
+            name=name + '.depthwise_linear',
+            norm_name=name + '.depthwise_linear.bn')
+
+        return conv4
+
+    def shortcut(self, x, out_c, layer_num, name=None):
+        assert layer_num > 0
+        for i in range(layer_num):
+            param_name = name + '.layers.' + str(i * 2)
+            act = 'relu6' if i < layer_num - 1 else None
+            x = self._lite_conv(x, out_c, act, param_name)
+        return x
+
+    def _deconv_upsample(self, x, out_c, name=None):
+        conv1 = ConvNorm(
+            input=x,
+            num_filters=out_c,
+            filter_size=1,
+            norm_type='bn',
+            act='relu6',
+            name=name + '.pointwise',
+            initializer=Xavier(),
+            norm_name=name + '.pointwise.bn')
+        conv2 = fluid.layers.conv2d_transpose(
+            input=conv1,
+            num_filters=out_c,
+            filter_size=4,
+            padding=1,
+            stride=2,
+            groups=out_c,
+            param_attr=ParamAttr(
+                name=name + '.deconv.weights', initializer=Xavier()),
+            bias_attr=False)
+        bn = fluid.layers.batch_norm(
+            input=conv2,
+            act='relu6',
+            param_attr=ParamAttr(
+                name=name + '.deconv.bn.scale', regularizer=L2Decay(0.)),
+            bias_attr=ParamAttr(
+                name=name + '.deconv.bn.offset', regularizer=L2Decay(0.)),
+            moving_mean_name=name + '.deconv.bn.mean',
+            moving_variance_name=name + '.deconv.bn.variance')
+        conv3 = ConvNorm(
+            input=bn,
+            num_filters=out_c,
+            filter_size=1,
+            norm_type='bn',
+            act='relu6',
+            name=name + '.normal',
+            initializer=Xavier(),
+            norm_name=name + '.normal.bn')
+        return conv3
+
+    def _interp_upsample(self, x, out_c, name=None):
+        conv = self._lite_conv(x, out_c, 'relu6', name)
+        up = fluid.layers.resize_bilinear(conv, scale=2)
+        return up
+
+    def upsample(self, x, out_c, name=None):
+        deconv_up = self._deconv_upsample(x, out_c, name=name + '.dilation_up')
+        interp_up = self._interp_upsample(x, out_c, name=name + '.interp_up')
+        return deconv_up + interp_up
+
+    def _head(self,
+              x,
+              out_c,
+              conv_num=1,
+              head_out_c=None,
+              name=None,
+              is_test=False):
+        head_out_c = self.head_conv if not head_out_c else head_out_c
+        for i in range(conv_num):
+            conv_name = '{}.{}.conv'.format(name, i)
+            x = self._lite_conv(x, head_out_c, 'relu6', conv_name)
+        bias_init = float(-np.log((1 - 0.01) / 0.01)) if '.hm' in name else 0.
+        conv_b_init = Constant(bias_init)
+        x = fluid.layers.conv2d(
+            x,
+            out_c,
+            1,
+            param_attr=ParamAttr(name='{}.{}.weight'.format(name, conv_num)),
+            bias_attr=ParamAttr(
+                learning_rate=2.,
+                regularizer=L2Decay(0.),
+                name='{}.{}.bias'.format(name, conv_num),
+                initializer=conv_b_init))
+        return x