Merge pull request #1363 from MissPenguin/dygraph

add east & sast

configs/det/det_mv3_east.yml

+Global:
+  use_gpu: true
+  epoch_num: 10000
+  log_smooth_window: 20
+  print_batch_step: 2
+  save_model_dir: ./output/east_mv3/
+  save_epoch_step: 1000
+  # evaluation is run every 5000 iterations after the 4000th iteration
+  eval_batch_step: [4000, 5000]
+  # if pretrained_model is saved in static mode, load_static_weights must set to True
+  load_static_weights: True
+  cal_metric_during_train: False
+  pretrained_model: ./pretrain_models/MobileNetV3_large_x0_5_pretrained
+  checkpoints: 
+  save_inference_dir:
+  use_visualdl: False
+  infer_img: 
+  save_res_path: ./output/det_east/predicts_east.txt
+
+Architecture:
+  model_type: det
+  algorithm: EAST
+  Transform:
+  Backbone:
+    name: MobileNetV3
+    scale: 0.5
+    model_name: large
+  Neck:
+    name: EASTFPN
+    model_name: small
+  Head:
+    name: EASTHead
+    model_name: small
+
+Loss:
+  name: EASTLoss
+  
+Optimizer:
+  name: Adam
+  beta1: 0.9
+  beta2: 0.999
+  lr:
+  #  name: Cosine
+    learning_rate: 0.001
+  #  warmup_epoch: 0
+  regularizer:
+    name: 'L2'
+    factor: 0
+
+PostProcess:
+  name: EASTPostProcess
+  score_thresh: 0.8
+  cover_thresh: 0.1
+  nms_thresh: 0.2
+
+Metric:
+  name: DetMetric
+  main_indicator: hmean
+
+Train:
+  dataset:
+    name: SimpleDataSet
+    data_dir: ./train_data/icdar2015/text_localization/
+    label_file_list:
+      - ./train_data/icdar2015/text_localization/train_icdar2015_label.txt
+    ratio_list: [1.0]
+    transforms:
+      - DecodeImage: # load image
+          img_mode: BGR
+          channel_first: False
+      - DetLabelEncode: # Class handling label
+      - EASTProcessTrain:
+          image_shape: [512, 512]
+          background_ratio: 0.125
+          min_crop_side_ratio: 0.1
+          min_text_size: 10
+      - KeepKeys:
+          keep_keys: ['image', 'score_map', 'geo_map', 'training_mask'] # dataloader will return list in this order
+  loader:
+    shuffle: True
+    drop_last: False
+    batch_size_per_card: 16
+    num_workers: 8
+
+Eval:
+  dataset:
+    name: SimpleDataSet
+    data_dir: ./train_data/icdar2015/text_localization/
+    label_file_list:
+      - ./train_data/icdar2015/text_localization/test_icdar2015_label.txt
+    transforms:
+      - DecodeImage: # load image
+          img_mode: BGR
+          channel_first: False
+      - DetLabelEncode: # Class handling label
+      - DetResizeForTest:
+          limit_side_len: 2400
+          limit_type: max
+      - NormalizeImage:
+          scale: 1./255.
+          mean: [0.485, 0.456, 0.406]
+          std: [0.229, 0.224, 0.225]
+          order: 'hwc'
+      - ToCHWImage:
+      - KeepKeys:
+          keep_keys: ['image', 'shape', 'polys', 'ignore_tags']
+  loader:
+    shuffle: False
+    drop_last: False
+    batch_size_per_card: 1 # must be 1
+    num_workers: 2
\ No newline at end of file

configs/det/det_r50_vd_east.yml

+Global:
+  use_gpu: true
+  epoch_num: 10000
+  log_smooth_window: 20
+  print_batch_step: 2
+  save_model_dir: ./output/east_r50_vd/
+  save_epoch_step: 1000
+  # evaluation is run every 5000 iterations after the 4000th iteration
+  eval_batch_step: [4000, 5000]
+  # if pretrained_model is saved in static mode, load_static_weights must set to True
+  load_static_weights: True
+  cal_metric_during_train: False
+  pretrained_model: ./pretrain_models/ResNet50_vd_pretrained/
+  checkpoints: 
+  save_inference_dir:
+  use_visualdl: False
+  infer_img: 
+  save_res_path: ./output/det_east/predicts_east.txt
+
+Architecture:
+  model_type: det
+  algorithm: EAST
+  Transform:
+  Backbone:
+    name: ResNet
+    layers: 50
+  Neck:
+    name: EASTFPN
+    model_name: large
+  Head:
+    name: EASTHead
+    model_name: large
+
+Loss:
+  name: EASTLoss
+  
+Optimizer:
+  name: Adam
+  beta1: 0.9
+  beta2: 0.999
+  lr:
+  #  name: Cosine
+    learning_rate: 0.001
+  #  warmup_epoch: 0
+  regularizer:
+    name: 'L2'
+    factor: 0
+
+PostProcess:
+  name: EASTPostProcess
+  score_thresh: 0.8
+  cover_thresh: 0.1
+  nms_thresh: 0.2
+
+Metric:
+  name: DetMetric
+  main_indicator: hmean
+
+Train:
+  dataset:
+    name: SimpleDataSet
+    data_dir: ./train_data/icdar2015/text_localization/
+    label_file_list:
+      - ./train_data/icdar2015/text_localization/train_icdar2015_label.txt
+    ratio_list: [1.0]
+    transforms:
+      - DecodeImage: # load image
+          img_mode: BGR
+          channel_first: False
+      - DetLabelEncode: # Class handling label
+      - EASTProcessTrain:
+          image_shape: [512, 512]
+          background_ratio: 0.125
+          min_crop_side_ratio: 0.1
+          min_text_size: 10
+      - KeepKeys:
+          keep_keys: ['image', 'score_map', 'geo_map', 'training_mask'] # dataloader will return list in this order
+  loader:
+    shuffle: True
+    drop_last: False
+    batch_size_per_card: 8
+    num_workers: 8
+
+Eval:
+  dataset:
+    name: SimpleDataSet
+    data_dir: ./train_data/icdar2015/text_localization/
+    label_file_list:
+      - ./train_data/icdar2015/text_localization/test_icdar2015_label.txt
+    transforms:
+      - DecodeImage: # load image
+          img_mode: BGR
+          channel_first: False
+      - DetLabelEncode: # Class handling label
+      - DetResizeForTest:
+          limit_side_len: 2400
+          limit_type: max
+      - NormalizeImage:
+          scale: 1./255.
+          mean: [0.485, 0.456, 0.406]
+          std: [0.229, 0.224, 0.225]
+          order: 'hwc'
+      - ToCHWImage:
+      - KeepKeys:
+          keep_keys: ['image', 'shape', 'polys', 'ignore_tags']
+  loader:
+    shuffle: False
+    drop_last: False
+    batch_size_per_card: 1 # must be 1
+    num_workers: 2
\ No newline at end of file

configs/det/det_r50_vd_sast_icdar15.yml

+Global:
+  use_gpu: true
+  epoch_num: 5000
+  log_smooth_window: 20
+  print_batch_step: 2
+  save_model_dir: ./output/sast_r50_vd_ic15/
+  save_epoch_step: 1000
+  # evaluation is run every 5000 iterations after the 4000th iteration
+  eval_batch_step: [4000, 5000]
+  # if pretrained_model is saved in static mode, load_static_weights must set to True
+  load_static_weights: True
+  cal_metric_during_train: False
+  pretrained_model: ./pretrain_models/ResNet50_vd_ssld_pretrained/
+  checkpoints: 
+  save_inference_dir:
+  use_visualdl: False
+  infer_img: 
+  save_res_path: ./output/sast_r50_vd_ic15/predicts_sast.txt
+
+Architecture:
+  model_type: det
+  algorithm: SAST
+  Transform:
+  Backbone:
+    name: ResNet_SAST
+    layers: 50
+  Neck:
+    name: SASTFPN
+    with_cab: True
+  Head:
+    name: SASTHead
+
+Loss:
+  name: SASTLoss
+  
+Optimizer:
+  name: Adam
+  beta1: 0.9
+  beta2: 0.999
+  lr:
+  #  name: Cosine
+    learning_rate: 0.001
+  #  warmup_epoch: 0
+  regularizer:
+    name: 'L2'
+    factor: 0
+
+PostProcess:
+  name: SASTPostProcess
+  score_thresh: 0.5
+  sample_pts_num: 2
+  nms_thresh: 0.2
+  expand_scale: 1.0
+  shrink_ratio_of_width: 0.3
+
+Metric:
+  name: DetMetric
+  main_indicator: hmean
+
+Train:
+  dataset:
+    name: SimpleDataSet
+    data_dir: ./train_data/
+    label_file_path: [./train_data/art_latin_icdar_14pt/train_no_tt_test/train_label_json.txt, ./train_data/total_text_icdar_14pt/train_label_json.txt]
+    data_ratio_list: [0.5, 0.5]
+    transforms:
+      - DecodeImage: # load image
+          img_mode: BGR
+          channel_first: False
+      - DetLabelEncode: # Class handling label
+      - SASTProcessTrain:
+          image_shape: [512, 512]
+          min_crop_side_ratio: 0.3
+          min_crop_size: 24
+          min_text_size: 4
+          max_text_size: 512
+      - KeepKeys:
+          keep_keys: ['image', 'score_map', 'border_map', 'training_mask', 'tvo_map', 'tco_map'] # dataloader will return list in this order
+  loader:
+    shuffle: True
+    drop_last: False
+    batch_size_per_card: 4
+    num_workers: 4
+
+Eval:
+  dataset:
+    name: SimpleDataSet
+    data_dir: ./train_data/icdar2015/text_localization/
+    label_file_list:
+      - ./train_data/icdar2015/text_localization/test_icdar2015_label.txt
+    transforms:
+      - DecodeImage: # load image
+          img_mode: BGR
+          channel_first: False
+      - DetLabelEncode: # Class handling label
+      - DetResizeForTest:
+          resize_long: 1536
+      - NormalizeImage:
+          scale: 1./255.
+          mean: [0.485, 0.456, 0.406]
+          std: [0.229, 0.224, 0.225]
+          order: 'hwc'
+      - ToCHWImage:
+      - KeepKeys:
+          keep_keys: ['image', 'shape', 'polys', 'ignore_tags']
+  loader:
+    shuffle: False
+    drop_last: False
+    batch_size_per_card: 1 # must be 1
+    num_workers: 2
\ No newline at end of file

configs/det/det_r50_vd_sast_totaltext.yml

+Global:
+  use_gpu: true
+  epoch_num: 5000
+  log_smooth_window: 20
+  print_batch_step: 2
+  save_model_dir: ./output/sast_r50_vd_tt/
+  save_epoch_step: 1000
+  # evaluation is run every 5000 iterations after the 4000th iteration
+  eval_batch_step: [4000, 5000]
+  # if pretrained_model is saved in static mode, load_static_weights must set to True
+  load_static_weights: True
+  cal_metric_during_train: False
+  pretrained_model: ./pretrain_models/ResNet50_vd_ssld_pretrained/
+  checkpoints: 
+  save_inference_dir:
+  use_visualdl: False
+  infer_img: 
+  save_res_path: ./output/sast_r50_vd_tt/predicts_sast.txt
+
+Architecture:
+  model_type: det
+  algorithm: SAST
+  Transform:
+  Backbone:
+    name: ResNet_SAST
+    layers: 50
+  Neck:
+    name: SASTFPN
+    with_cab: True
+  Head:
+    name: SASTHead
+
+Loss:
+  name: SASTLoss
+  
+Optimizer:
+  name: Adam
+  beta1: 0.9
+  beta2: 0.999
+  lr:
+  #  name: Cosine
+    learning_rate: 0.001
+  #  warmup_epoch: 0
+  regularizer:
+    name: 'L2'
+    factor: 0
+
+PostProcess:
+  name: SASTPostProcess
+  score_thresh: 0.5
+  sample_pts_num: 6
+  nms_thresh: 0.2
+  expand_scale: 1.2
+  shrink_ratio_of_width: 0.2
+
+Metric:
+  name: DetMetric
+  main_indicator: hmean
+
+Train:
+  dataset:
+    name: SimpleDataSet
+    label_file_list: [./train_data/icdar2013/train_label_json.txt, ./train_data/icdar2015/train_label_json.txt, ./train_data/icdar17_mlt_latin/train_label_json.txt, ./train_data/coco_text_icdar_4pts/train_label_json.txt]
+    ratio_list: [0.1, 0.45, 0.3, 0.15]
+    transforms:
+      - DecodeImage: # load image
+          img_mode: BGR
+          channel_first: False
+      - DetLabelEncode: # Class handling label
+      - SASTProcessTrain:
+          image_shape: [512, 512]
+          min_crop_side_ratio: 0.3
+          min_crop_size: 24
+          min_text_size: 4
+          max_text_size: 512
+      - KeepKeys:
+          keep_keys: ['image', 'score_map', 'border_map', 'training_mask', 'tvo_map', 'tco_map'] # dataloader will return list in this order
+  loader:
+    shuffle: True
+    drop_last: False
+    batch_size_per_card: 4
+    num_workers: 4
+
+Eval:
+  dataset:
+    name: SimpleDataSet
+    data_dir: ./train_data/
+    label_file_list: 
+        - ./train_data/total_text_icdar_14pt/test_label_json.txt
+    transforms:
+      - DecodeImage: # load image
+          img_mode: BGR
+          channel_first: False
+      - DetLabelEncode: # Class handling label
+      - DetResizeForTest:
+          resize_long: 768
+      - NormalizeImage:
+          scale: 1./255.
+          mean: [0.485, 0.456, 0.406]
+          std: [0.229, 0.224, 0.225]
+          order: 'hwc'
+      - ToCHWImage:
+      - KeepKeys:
+          keep_keys: ['image', 'shape', 'polys', 'ignore_tags']
+  loader:
+    shuffle: False
+    drop_last: False
+    batch_size_per_card: 1 # must be 1
+    num_workers: 2
\ No newline at end of file

ppocr/data/imaug/__init__.py

@@ -26,6 +26,9 @@ from .randaugment import RandAugment
 from .operators import *
 from .label_ops import *
 
+from .east_process import *
+from .sast_process import *
+
 
 def transform(data, ops=None):
     """ transform """

ppocr/data/imaug/east_process.py

+#copyright (c) 2020 PaddlePaddle Authors. All Rights Reserve.
+#
+#Licensed under the Apache License, Version 2.0 (the "License");
+#you may not use this file except in compliance with the License.
+#You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+#Unless required by applicable law or agreed to in writing, software
+#distributed under the License is distributed on an "AS IS" BASIS,
+#WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+#See the License for the specific language governing permissions and
+#limitations under the License.
+
+import math
+import cv2
+import numpy as np
+import json
+import sys
+import os
+
+__all__ = ['EASTProcessTrain']
+
+
+class EASTProcessTrain(object):
+    def __init__(self,
+                 image_shape = [512, 512],
+                 background_ratio = 0.125,
+                 min_crop_side_ratio = 0.1,
+                 min_text_size = 10,
+                 **kwargs):
+        self.input_size = image_shape[1]
+        self.random_scale = np.array([0.5, 1, 2.0, 3.0])
+        self.background_ratio = background_ratio
+        self.min_crop_side_ratio = min_crop_side_ratio
+        self.min_text_size = min_text_size
+
+    def preprocess(self, im):
+        input_size = self.input_size
+        im_shape = im.shape
+        im_size_min = np.min(im_shape[0:2])
+        im_size_max = np.max(im_shape[0:2])
+        im_scale = float(input_size) / float(im_size_max)
+        im = cv2.resize(im, None, None, fx=im_scale, fy=im_scale)
+        img_mean = [0.485, 0.456, 0.406]
+        img_std = [0.229, 0.224, 0.225]
+        # im = im[:, :, ::-1].astype(np.float32)
+        im = im / 255
+        im -= img_mean
+        im /= img_std
+        new_h, new_w, _ = im.shape
+        im_padded = np.zeros((input_size, input_size, 3), dtype=np.float32)
+        im_padded[:new_h, :new_w, :] = im
+        im_padded = im_padded.transpose((2, 0, 1))
+        im_padded = im_padded[np.newaxis, :]
+        return im_padded, im_scale
+
+    def rotate_im_poly(self, im, text_polys):
+        """
+        rotate image with 90 / 180 / 270 degre
+        """
+        im_w, im_h = im.shape[1], im.shape[0]
+        dst_im = im.copy()
+        dst_polys = []
+        rand_degree_ratio = np.random.rand()
+        rand_degree_cnt = 1
+        if 0.333 < rand_degree_ratio < 0.666:
+            rand_degree_cnt = 2
+        elif rand_degree_ratio > 0.666:
+            rand_degree_cnt = 3
+        for i in range(rand_degree_cnt):
+            dst_im = np.rot90(dst_im)
+        rot_degree = -90 * rand_degree_cnt
+        rot_angle = rot_degree * math.pi / 180.0
+        n_poly = text_polys.shape[0]
+        cx, cy = 0.5 * im_w, 0.5 * im_h
+        ncx, ncy = 0.5 * dst_im.shape[1], 0.5 * dst_im.shape[0]
+        for i in range(n_poly):
+            wordBB = text_polys[i]
+            poly = []
+            for j in range(4):
+                sx, sy = wordBB[j][0], wordBB[j][1]
+                dx = math.cos(rot_angle) * (sx - cx)\
+                    - math.sin(rot_angle) * (sy - cy) + ncx
+                dy = math.sin(rot_angle) * (sx - cx)\
+                    + math.cos(rot_angle) * (sy - cy) + ncy
+                poly.append([dx, dy])
+            dst_polys.append(poly)
+        dst_polys = np.array(dst_polys, dtype=np.float32)
+        return dst_im, dst_polys
+
+    def polygon_area(self, poly):
+        """
+        compute area of a polygon
+        :param poly:
+        :return:
+        """
+        edge = [(poly[1][0] - poly[0][0]) * (poly[1][1] + poly[0][1]),
+                (poly[2][0] - poly[1][0]) * (poly[2][1] + poly[1][1]),
+                (poly[3][0] - poly[2][0]) * (poly[3][1] + poly[2][1]),
+                (poly[0][0] - poly[3][0]) * (poly[0][1] + poly[3][1])]
+        return np.sum(edge) / 2.
+
+    def check_and_validate_polys(self, polys, tags, img_height, img_width):
+        """
+        check so that the text poly is in the same direction,
+        and also filter some invalid polygons
+        :param polys:
+        :param tags:
+        :return:
+        """
+        h, w = img_height, img_width
+        if polys.shape[0] == 0:
+            return polys
+        polys[:, :, 0] = np.clip(polys[:, :, 0], 0, w - 1)
+        polys[:, :, 1] = np.clip(polys[:, :, 1], 0, h - 1)
+
+        validated_polys = []
+        validated_tags = []
+        for poly, tag in zip(polys, tags):
+            p_area = self.polygon_area(poly)
+            #invalid poly
+            if abs(p_area) < 1:
+                continue
+            if p_area > 0:
+                #'poly in wrong direction'
+                if not tag:
+                    tag = True  #reversed cases should be ignore
+                poly = poly[(0, 3, 2, 1), :]
+            validated_polys.append(poly)
+            validated_tags.append(tag)
+        return np.array(validated_polys), np.array(validated_tags)
+
+    def draw_img_polys(self, img, polys):
+        if len(img.shape) == 4:
+            img = np.squeeze(img, axis=0)
+        if img.shape[0] == 3:
+            img = img.transpose((1, 2, 0))
+            img[:, :, 2] += 123.68
+            img[:, :, 1] += 116.78
+            img[:, :, 0] += 103.94
+        cv2.imwrite("tmp.jpg", img)
+        img = cv2.imread("tmp.jpg")
+        for box in polys:
+            box = box.astype(np.int32).reshape((-1, 1, 2))
+            cv2.polylines(img, [box], True, color=(255, 255, 0), thickness=2)
+        import random
+        ino = random.randint(0, 100)
+        cv2.imwrite("tmp_%d.jpg" % ino, img)
+        return
+
+    def shrink_poly(self, poly, r):
+        """
+        fit a poly inside the origin poly, maybe bugs here...
+        used for generate the score map
+        :param poly: the text poly
+        :param r: r in the paper
+        :return: the shrinked poly
+        """
+        # shrink ratio
+        R = 0.3
+        # find the longer pair
+        dist0 = np.linalg.norm(poly[0] - poly[1])
+        dist1 = np.linalg.norm(poly[2] - poly[3])
+        dist2 = np.linalg.norm(poly[0] - poly[3])
+        dist3 = np.linalg.norm(poly[1] - poly[2])
+        if dist0 + dist1 > dist2 + dist3:
+            # first move (p0, p1), (p2, p3), then (p0, p3), (p1, p2)
+            ## p0, p1
+            theta = np.arctan2((poly[1][1] - poly[0][1]),
+                               (poly[1][0] - poly[0][0]))
+            poly[0][0] += R * r[0] * np.cos(theta)
+            poly[0][1] += R * r[0] * np.sin(theta)
+            poly[1][0] -= R * r[1] * np.cos(theta)
+            poly[1][1] -= R * r[1] * np.sin(theta)
+            ## p2, p3
+            theta = np.arctan2((poly[2][1] - poly[3][1]),
+                               (poly[2][0] - poly[3][0]))
+            poly[3][0] += R * r[3] * np.cos(theta)
+            poly[3][1] += R * r[3] * np.sin(theta)
+            poly[2][0] -= R * r[2] * np.cos(theta)
+            poly[2][1] -= R * r[2] * np.sin(theta)
+            ## p0, p3
+            theta = np.arctan2((poly[3][0] - poly[0][0]),
+                               (poly[3][1] - poly[0][1]))
+            poly[0][0] += R * r[0] * np.sin(theta)
+            poly[0][1] += R * r[0] * np.cos(theta)
+            poly[3][0] -= R * r[3] * np.sin(theta)
+            poly[3][1] -= R * r[3] * np.cos(theta)
+            ## p1, p2
+            theta = np.arctan2((poly[2][0] - poly[1][0]),
+                               (poly[2][1] - poly[1][1]))
+            poly[1][0] += R * r[1] * np.sin(theta)
+            poly[1][1] += R * r[1] * np.cos(theta)
+            poly[2][0] -= R * r[2] * np.sin(theta)
+            poly[2][1] -= R * r[2] * np.cos(theta)
+        else:
+            ## p0, p3
+            # print poly
+            theta = np.arctan2((poly[3][0] - poly[0][0]),
+                               (poly[3][1] - poly[0][1]))
+            poly[0][0] += R * r[0] * np.sin(theta)
+            poly[0][1] += R * r[0] * np.cos(theta)
+            poly[3][0] -= R * r[3] * np.sin(theta)
+            poly[3][1] -= R * r[3] * np.cos(theta)
+            ## p1, p2
+            theta = np.arctan2((poly[2][0] - poly[1][0]),
+                               (poly[2][1] - poly[1][1]))
+            poly[1][0] += R * r[1] * np.sin(theta)
+            poly[1][1] += R * r[1] * np.cos(theta)
+            poly[2][0] -= R * r[2] * np.sin(theta)
+            poly[2][1] -= R * r[2] * np.cos(theta)
+            ## p0, p1
+            theta = np.arctan2((poly[1][1] - poly[0][1]),
+                               (poly[1][0] - poly[0][0]))
+            poly[0][0] += R * r[0] * np.cos(theta)
+            poly[0][1] += R * r[0] * np.sin(theta)
+            poly[1][0] -= R * r[1] * np.cos(theta)
+            poly[1][1] -= R * r[1] * np.sin(theta)
+            ## p2, p3
+            theta = np.arctan2((poly[2][1] - poly[3][1]),
+                               (poly[2][0] - poly[3][0]))
+            poly[3][0] += R * r[3] * np.cos(theta)
+            poly[3][1] += R * r[3] * np.sin(theta)
+            poly[2][0] -= R * r[2] * np.cos(theta)
+            poly[2][1] -= R * r[2] * np.sin(theta)
+        return poly
+
+    def generate_quad(self, im_size, polys, tags):
+        """
+        Generate quadrangle.
+        """
+        h, w = im_size
+        poly_mask = np.zeros((h, w), dtype=np.uint8)
+        score_map = np.zeros((h, w), dtype=np.uint8)
+        # (x1, y1, ..., x4, y4, short_edge_norm)
+        geo_map = np.zeros((h, w, 9), dtype=np.float32)
+        # mask used during traning, to ignore some hard areas
+        training_mask = np.ones((h, w), dtype=np.uint8)
+        for poly_idx, poly_tag in enumerate(zip(polys, tags)):
+            poly = poly_tag[0]
+            tag = poly_tag[1]
+
+            r = [None, None, None, None]
+            for i in range(4):
+                dist1 = np.linalg.norm(poly[i] - poly[(i + 1) % 4])
+                dist2 = np.linalg.norm(poly[i] - poly[(i - 1) % 4])
+                r[i] = min(dist1, dist2)
+            # score map
+            shrinked_poly = self.shrink_poly(
+                poly.copy(), r).astype(np.int32)[np.newaxis, :, :]
+            cv2.fillPoly(score_map, shrinked_poly, 1)
+            cv2.fillPoly(poly_mask, shrinked_poly, poly_idx + 1)
+            # if the poly is too small, then ignore it during training
+            poly_h = min(
+                np.linalg.norm(poly[0] - poly[3]),
+                np.linalg.norm(poly[1] - poly[2]))
+            poly_w = min(
+                np.linalg.norm(poly[0] - poly[1]),
+                np.linalg.norm(poly[2] - poly[3]))
+            if min(poly_h, poly_w) < self.min_text_size:
+                cv2.fillPoly(training_mask,
+                             poly.astype(np.int32)[np.newaxis, :, :], 0)
+
+            if tag:
+                cv2.fillPoly(training_mask,
+                             poly.astype(np.int32)[np.newaxis, :, :], 0)
+
+            xy_in_poly = np.argwhere(poly_mask == (poly_idx + 1))
+            # geo map.
+            y_in_poly = xy_in_poly[:, 0]
+            x_in_poly = xy_in_poly[:, 1]
+            poly[:, 0] = np.minimum(np.maximum(poly[:, 0], 0), w)
+            poly[:, 1] = np.minimum(np.maximum(poly[:, 1], 0), h)
+            for pno in range(4):
+                geo_channel_beg = pno * 2
+                geo_map[y_in_poly, x_in_poly, geo_channel_beg] =\
+                    x_in_poly - poly[pno, 0]
+                geo_map[y_in_poly, x_in_poly, geo_channel_beg+1] =\
+                    y_in_poly - poly[pno, 1]
+            geo_map[y_in_poly, x_in_poly, 8] = \
+                1.0 / max(min(poly_h, poly_w), 1.0)
+        return score_map, geo_map, training_mask
+
+    def crop_area(self,
+                  im,
+                  polys,
+                  tags,
+                  crop_background=False,
+                  max_tries=50):
+        """
+        make random crop from the input image
+        :param im:
+        :param polys:
+        :param tags:
+        :param crop_background:
+        :param max_tries:
+        :return:
+        """
+        h, w, _ = im.shape
+        pad_h = h // 10
+        pad_w = w // 10
+        h_array = np.zeros((h + pad_h * 2), dtype=np.int32)
+        w_array = np.zeros((w + pad_w * 2), dtype=np.int32)
+        for poly in polys:
+            poly = np.round(poly, decimals=0).astype(np.int32)
+            minx = np.min(poly[:, 0])
+            maxx = np.max(poly[:, 0])
+            w_array[minx + pad_w:maxx + pad_w] = 1
+            miny = np.min(poly[:, 1])
+            maxy = np.max(poly[:, 1])
+            h_array[miny + pad_h:maxy + pad_h] = 1
+        # ensure the cropped area not across a text
+        h_axis = np.where(h_array == 0)[0]
+        w_axis = np.where(w_array == 0)[0]
+        if len(h_axis) == 0 or len(w_axis) == 0:
+            return im, polys, tags
+
+        for i in range(max_tries):
+            xx = np.random.choice(w_axis, size=2)
+            xmin = np.min(xx) - pad_w
+            xmax = np.max(xx) - pad_w
+            xmin = np.clip(xmin, 0, w - 1)
+            xmax = np.clip(xmax, 0, w - 1)
+            yy = np.random.choice(h_axis, size=2)
+            ymin = np.min(yy) - pad_h
+            ymax = np.max(yy) - pad_h
+            ymin = np.clip(ymin, 0, h - 1)
+            ymax = np.clip(ymax, 0, h - 1)
+            if xmax - xmin < self.min_crop_side_ratio * w or \
+               ymax - ymin < self.min_crop_side_ratio * h:
+                # area too small
+                continue
+            if polys.shape[0] != 0:
+                poly_axis_in_area = (polys[:, :, 0] >= xmin)\
+                    & (polys[:, :, 0] <= xmax)\
+                    & (polys[:, :, 1] >= ymin)\
+                    & (polys[:, :, 1] <= ymax)
+                selected_polys = np.where(
+                    np.sum(poly_axis_in_area, axis=1) == 4)[0]
+            else:
+                selected_polys = []
+
+            if len(selected_polys) == 0:
+                # no text in this area
+                if crop_background:
+                    im = im[ymin:ymax + 1, xmin:xmax + 1, :]
+                    polys = []
+                    tags = []
+                    return im, polys, tags
+                else:
+                    continue
+
+            im = im[ymin:ymax + 1, xmin:xmax + 1, :]
+            polys = polys[selected_polys]
+            tags = tags[selected_polys]
+            polys[:, :, 0] -= xmin
+            polys[:, :, 1] -= ymin
+            return im, polys, tags
+        return im, polys, tags
+
+    def crop_background_infor(self, im, text_polys, text_tags):
+        im, text_polys, text_tags = self.crop_area(
+            im, text_polys, text_tags, crop_background=True)
+
+        if len(text_polys) > 0:
+            return None
+        # pad and resize image
+        input_size = self.input_size
+        im, ratio = self.preprocess(im)
+        score_map = np.zeros((input_size, input_size), dtype=np.float32)
+        geo_map = np.zeros((input_size, input_size, 9), dtype=np.float32)
+        training_mask = np.ones((input_size, input_size), dtype=np.float32)
+        return im, score_map, geo_map, training_mask
+
+    def crop_foreground_infor(self, im, text_polys, text_tags):
+        im, text_polys, text_tags = self.crop_area(
+            im, text_polys, text_tags, crop_background=False)
+
+        if text_polys.shape[0] == 0:
+            return None
+        #continue for all ignore case
+        if np.sum((text_tags * 1.0)) >= text_tags.size:
+            return None
+        # pad and resize image
+        input_size = self.input_size
+        im, ratio = self.preprocess(im)
+        text_polys[:, :, 0] *= ratio
+        text_polys[:, :, 1] *= ratio
+        _, _, new_h, new_w = im.shape
+        #         print(im.shape)
+        #         self.draw_img_polys(im, text_polys)
+        score_map, geo_map, training_mask = self.generate_quad(
+            (new_h, new_w), text_polys, text_tags)
+        return im, score_map, geo_map, training_mask
+
+    def __call__(self, data):
+        im = data['image']
+        text_polys = data['polys']
+        text_tags = data['ignore_tags']
+        if im is None:
+            return None
+        if text_polys.shape[0] == 0:
+            return None
+
+        #add rotate cases
+        if np.random.rand() < 0.5:
+            im, text_polys = self.rotate_im_poly(im, text_polys)
+        h, w, _ = im.shape
+        text_polys, text_tags = self.check_and_validate_polys(text_polys,
+                                                              text_tags, h, w)
+        if text_polys.shape[0] == 0:
+            return None
+
+        # random scale this image
+        rd_scale = np.random.choice(self.random_scale)
+        im = cv2.resize(im, dsize=None, fx=rd_scale, fy=rd_scale)
+        text_polys *= rd_scale
+        if np.random.rand() < self.background_ratio:
+            outs = self.crop_background_infor(im, text_polys, text_tags)
+        else:
+            outs = self.crop_foreground_infor(im, text_polys, text_tags)
+
+        if outs is None:
+            return None
+        im, score_map, geo_map, training_mask = outs
+        score_map = score_map[np.newaxis, ::4, ::4].astype(np.float32)
+        geo_map = np.swapaxes(geo_map, 1, 2)
+        geo_map = np.swapaxes(geo_map, 1, 0)
+        geo_map = geo_map[:, ::4, ::4].astype(np.float32)
+        training_mask = training_mask[np.newaxis, ::4, ::4]
+        training_mask = training_mask.astype(np.float32)
+
+        data['image'] = im[0]
+        data['score_map'] = score_map
+        data['geo_map'] = geo_map
+        data['training_mask'] = training_mask
+        # print(im.shape, score_map.shape, geo_map.shape, training_mask.shape)
+        return data
\ No newline at end of file

ppocr/data/imaug/label_ops.py

@@ -52,6 +52,7 @@ class DetLabelEncode(object):
                 txt_tags.append(True)
             else:
                 txt_tags.append(False)
+        boxes = self.expand_points_num(boxes)
         boxes = np.array(boxes, dtype=np.float32)
         txt_tags = np.array(txt_tags, dtype=np.bool)
 
@@ -70,6 +71,17 @@ class DetLabelEncode(object):
         rect[3] = pts[np.argmax(diff)]
         return rect
 
+    def expand_points_num(self, boxes):
+        max_points_num = 0
+        for box in boxes:
+            if len(box) > max_points_num:
+                max_points_num = len(box)
+        ex_boxes = []
+        for box in boxes:
+            ex_box = box + [box[-1]] * (max_points_num - len(box))
+            ex_boxes.append(ex_box)
+        return ex_boxes
+
 
 class BaseRecLabelEncode(object):
     """ Convert between text-label and text-index """
@@ -83,7 +95,7 @@ class BaseRecLabelEncode(object):
             'ch', 'en', 'en_sensitive', 'french', 'german', 'japan', 'korean'
         ]
         assert character_type in support_character_type, "Only {} are supported now but get {}".format(
-            support_character_type, self.character_str)
+            support_character_type, character_type)
 
         self.max_text_len = max_text_length
         if character_type == "en":

ppocr/data/imaug/operators.py

@@ -120,26 +120,37 @@ class DetResizeForTest(object):
         if 'limit_side_len' in kwargs:
             self.limit_side_len = kwargs['limit_side_len']
             self.limit_type = kwargs.get('limit_type', 'min')
+        if 'resize_long' in kwargs:
+            self.resize_type = 2
+            self.resize_long = kwargs.get('resize_long', 960)
         else:
             self.limit_side_len = 736
             self.limit_type = 'min'
 
     def __call__(self, data):
         img = data['image']
+        src_h, src_w, _ = img.shape
 
         if self.resize_type == 0:
-            img, shape = self.resize_image_type0(img)
+            # img, shape = self.resize_image_type0(img)
+            img, [ratio_h, ratio_w] = self.resize_image_type0(img)
+        elif self.resize_type == 2:
+            img, [ratio_h, ratio_w] = self.resize_image_type2(img)
         else:
-            img, shape = self.resize_image_type1(img)
+            # img, shape = self.resize_image_type1(img)
+            img, [ratio_h, ratio_w] = self.resize_image_type1(img)
         data['image'] = img
-        data['shape'] = shape
+        data['shape'] = np.array([src_h, src_w, ratio_h, ratio_w])
         return data
 
     def resize_image_type1(self, img):
         resize_h, resize_w = self.image_shape
         ori_h, ori_w = img.shape[:2]  # (h, w, c)
+        ratio_h = float(resize_h) / ori_h
+        ratio_w = float(resize_w) / ori_w
         img = cv2.resize(img, (int(resize_w), int(resize_h)))
-        return img, np.array([ori_h, ori_w])
+        # return img, np.array([ori_h, ori_w])
+        return img, [ratio_h, ratio_w]
 
     def resize_image_type0(self, img):
         """
@@ -182,4 +193,31 @@ class DetResizeForTest(object):
         except:
             print(img.shape, resize_w, resize_h)
             sys.exit(0)
-        return img, np.array([h, w])
+        ratio_h = resize_h / float(h)
+        ratio_w = resize_w / float(w)
+        # return img, np.array([h, w])
+        return img, [ratio_h, ratio_w]
+    
+    def resize_image_type2(self, img):
+        h, w, _ = img.shape
+
+        resize_w = w
+        resize_h = h
+
+        # Fix the longer side
+        if resize_h > resize_w:
+            ratio = float(self.resize_long) / resize_h
+        else:
+            ratio = float(self.resize_long) / resize_w
+
+        resize_h = int(resize_h * ratio)
+        resize_w = int(resize_w * ratio)
+
+        max_stride = 128
+        resize_h = (resize_h + max_stride - 1) // max_stride * max_stride
+        resize_w = (resize_w + max_stride - 1) // max_stride * max_stride
+        img = cv2.resize(img, (int(resize_w), int(resize_h)))
+        ratio_h = resize_h / float(h)
+        ratio_w = resize_w / float(w)
+
+        return img, [ratio_h, ratio_w]

ppocr/losses/__init__.py

@@ -18,6 +18,8 @@ import copy
 def build_loss(config):
     # det loss
     from .det_db_loss import DBLoss
+    from .det_east_loss import EASTLoss
+    from .det_sast_loss import SASTLoss
 
     # rec loss
     from .rec_ctc_loss import CTCLoss
@@ -25,7 +27,7 @@ def build_loss(config):
     # cls loss
     from .cls_loss import ClsLoss
 
-    support_dict = ['DBLoss', 'CTCLoss', 'ClsLoss']
+    support_dict = ['DBLoss', 'EASTLoss', 'SASTLoss', 'CTCLoss', 'ClsLoss']
 
     config = copy.deepcopy(config)
     module_name = config.pop('name')

ppocr/losses/det_east_loss.py

+# copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import paddle
+from paddle import nn
+from .det_basic_loss import DiceLoss
+
+
+class EASTLoss(nn.Layer):
+    """
+    """
+
+    def __init__(self,
+                 eps=1e-6,
+                 **kwargs):
+        super(EASTLoss, self).__init__()
+        self.dice_loss = DiceLoss(eps=eps)
+
+    def forward(self, predicts, labels):
+        l_score, l_geo, l_mask = labels[1:]
+        f_score = predicts['f_score']
+        f_geo = predicts['f_geo']
+
+        dice_loss = self.dice_loss(f_score, l_score, l_mask)
+
+        #smoooth_l1_loss
+        channels = 8
+        l_geo_split = paddle.split(
+            l_geo, num_or_sections=channels + 1, axis=1)
+        f_geo_split = paddle.split(f_geo, num_or_sections=channels, axis=1)
+        smooth_l1 = 0
+        for i in range(0, channels):
+            geo_diff = l_geo_split[i] - f_geo_split[i]
+            abs_geo_diff = paddle.abs(geo_diff)
+            smooth_l1_sign = paddle.less_than(abs_geo_diff, l_score)
+            smooth_l1_sign = paddle.cast(smooth_l1_sign, dtype='float32')
+            in_loss = abs_geo_diff * abs_geo_diff * smooth_l1_sign + \
+                (abs_geo_diff - 0.5) * (1.0 - smooth_l1_sign)
+            out_loss = l_geo_split[-1] / channels * in_loss * l_score
+            smooth_l1 += out_loss
+        smooth_l1_loss = paddle.mean(smooth_l1 * l_score)
+
+        dice_loss = dice_loss * 0.01
+        total_loss = dice_loss + smooth_l1_loss
+        losses = {"loss":total_loss, \
+                  "dice_loss":dice_loss,\
+                  "smooth_l1_loss":smooth_l1_loss}
+        return losses

ppocr/losses/det_sast_loss.py

+# copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import paddle
+from paddle import nn
+from .det_basic_loss import DiceLoss
+import paddle.fluid as fluid
+import numpy as np
+
+
+class SASTLoss(nn.Layer):
+    """
+    """
+
+    def __init__(self,
+                 eps=1e-6,
+                 **kwargs):
+        super(SASTLoss, self).__init__()
+        self.dice_loss = DiceLoss(eps=eps)
+
+    def forward(self, predicts, labels):
+        """
+        tcl_pos: N x 128 x 3
+        tcl_mask: N x 128 x 1
+        tcl_label: N x X list or LoDTensor
+        """
+                
+        f_score = predicts['f_score']
+        f_border = predicts['f_border']
+        f_tvo = predicts['f_tvo']
+        f_tco = predicts['f_tco']
+
+        l_score, l_border, l_mask, l_tvo, l_tco = labels[1:]
+
+        #score_loss
+        intersection = paddle.sum(f_score * l_score * l_mask)
+        union = paddle.sum(f_score * l_mask) + paddle.sum(l_score * l_mask)
+        score_loss = 1.0 - 2 * intersection / (union + 1e-5)
+
+        #border loss
+        l_border_split, l_border_norm = paddle.split(l_border, num_or_sections=[4, 1], axis=1)
+        f_border_split = f_border
+        border_ex_shape = l_border_norm.shape * np.array([1, 4, 1, 1])
+        l_border_norm_split = paddle.expand(x=l_border_norm, shape=border_ex_shape)
+        l_border_score = paddle.expand(x=l_score, shape=border_ex_shape)   
+        l_border_mask = paddle.expand(x=l_mask, shape=border_ex_shape)  
+
+        border_diff = l_border_split - f_border_split
+        abs_border_diff = paddle.abs(border_diff) 
+        border_sign = abs_border_diff < 1.0
+        border_sign = paddle.cast(border_sign, dtype='float32')
+        border_sign.stop_gradient = True
+        border_in_loss = 0.5 * abs_border_diff * abs_border_diff * border_sign + \
+                    (abs_border_diff - 0.5) * (1.0 - border_sign)
+        border_out_loss = l_border_norm_split * border_in_loss
+        border_loss = paddle.sum(border_out_loss * l_border_score * l_border_mask) / \
+                    (paddle.sum(l_border_score * l_border_mask) + 1e-5)
+
+        #tvo_loss
+        l_tvo_split, l_tvo_norm = paddle.split(l_tvo, num_or_sections=[8, 1], axis=1)
+        f_tvo_split = f_tvo
+        tvo_ex_shape = l_tvo_norm.shape * np.array([1, 8, 1, 1])
+        l_tvo_norm_split = paddle.expand(x=l_tvo_norm, shape=tvo_ex_shape)
+        l_tvo_score = paddle.expand(x=l_score, shape=tvo_ex_shape)   
+        l_tvo_mask = paddle.expand(x=l_mask, shape=tvo_ex_shape)   
+        #
+        tvo_geo_diff = l_tvo_split - f_tvo_split
+        abs_tvo_geo_diff = paddle.abs(tvo_geo_diff) 
+        tvo_sign = abs_tvo_geo_diff < 1.0
+        tvo_sign = paddle.cast(tvo_sign, dtype='float32')
+        tvo_sign.stop_gradient = True
+        tvo_in_loss = 0.5 * abs_tvo_geo_diff * abs_tvo_geo_diff * tvo_sign + \
+                    (abs_tvo_geo_diff - 0.5) * (1.0 - tvo_sign)
+        tvo_out_loss = l_tvo_norm_split * tvo_in_loss
+        tvo_loss = paddle.sum(tvo_out_loss * l_tvo_score * l_tvo_mask) / \
+                    (paddle.sum(l_tvo_score * l_tvo_mask) + 1e-5)
+
+        #tco_loss
+        l_tco_split, l_tco_norm = paddle.split(l_tco, num_or_sections=[2, 1], axis=1)
+        f_tco_split = f_tco
+        tco_ex_shape = l_tco_norm.shape * np.array([1, 2, 1, 1])
+        l_tco_norm_split = paddle.expand(x=l_tco_norm, shape=tco_ex_shape)
+        l_tco_score = paddle.expand(x=l_score, shape=tco_ex_shape)   
+        l_tco_mask = paddle.expand(x=l_mask, shape=tco_ex_shape) 
+        
+        tco_geo_diff = l_tco_split - f_tco_split
+        abs_tco_geo_diff = paddle.abs(tco_geo_diff) 
+        tco_sign = abs_tco_geo_diff < 1.0
+        tco_sign = paddle.cast(tco_sign, dtype='float32')
+        tco_sign.stop_gradient = True
+        tco_in_loss = 0.5 * abs_tco_geo_diff * abs_tco_geo_diff * tco_sign + \
+                    (abs_tco_geo_diff - 0.5) * (1.0 - tco_sign)
+        tco_out_loss = l_tco_norm_split * tco_in_loss
+        tco_loss = paddle.sum(tco_out_loss * l_tco_score * l_tco_mask) / \
+                    (paddle.sum(l_tco_score * l_tco_mask) + 1e-5)
+
+
+        # total loss
+        tvo_lw, tco_lw = 1.5, 1.5
+        score_lw, border_lw = 1.0, 1.0
+        total_loss = score_loss * score_lw + border_loss * border_lw + \
+                    tvo_loss * tvo_lw + tco_loss * tco_lw
+                    
+        losses = {'loss':total_loss, "score_loss":score_loss,\
+            "border_loss":border_loss, 'tvo_loss':tvo_loss, 'tco_loss':tco_loss}
+        return losses
\ No newline at end of file

ppocr/modeling/backbones/__init__.py

@@ -19,6 +19,7 @@ def build_backbone(config, model_type):
     if model_type == 'det':
         from .det_mobilenet_v3 import MobileNetV3
         from .det_resnet_vd import ResNet
+        from .det_resnet_vd_sast import ResNet_SAST
         support_dict = ['MobileNetV3', 'ResNet', 'ResNet_SAST']
     elif model_type == 'rec' or model_type == 'cls':
         from .rec_mobilenet_v3 import MobileNetV3

ppocr/modeling/backbones/det_resnet_vd_sast.py

+# copyright (c) 2020 PaddlePaddle Authors. All Rights Reserve.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import paddle
+from paddle import ParamAttr
+import paddle.nn as nn
+import paddle.nn.functional as F
+
+__all__ = ["ResNet_SAST"]
+
+
+class ConvBNLayer(nn.Layer):
+    def __init__(
+            self,
+            in_channels,
+            out_channels,
+            kernel_size,
+            stride=1,
+            groups=1,
+            is_vd_mode=False,
+            act=None,
+            name=None, ):
+        super(ConvBNLayer, self).__init__()
+
+        self.is_vd_mode = is_vd_mode
+        self._pool2d_avg = nn.AvgPool2D(
+            kernel_size=2, stride=2, padding=0, ceil_mode=True)
+        self._conv = nn.Conv2D(
+            in_channels=in_channels,
+            out_channels=out_channels,
+            kernel_size=kernel_size,
+            stride=stride,
+            padding=(kernel_size - 1) // 2,
+            groups=groups,
+            weight_attr=ParamAttr(name=name + "_weights"),
+            bias_attr=False)
+        if name == "conv1":
+            bn_name = "bn_" + name
+        else:
+            bn_name = "bn" + name[3:]
+        self._batch_norm = nn.BatchNorm(
+            out_channels,
+            act=act,
+            param_attr=ParamAttr(name=bn_name + '_scale'),
+            bias_attr=ParamAttr(bn_name + '_offset'),
+            moving_mean_name=bn_name + '_mean',
+            moving_variance_name=bn_name + '_variance')
+
+    def forward(self, inputs):
+        if self.is_vd_mode:
+            inputs = self._pool2d_avg(inputs)
+        y = self._conv(inputs)
+        y = self._batch_norm(y)
+        return y
+
+
+class BottleneckBlock(nn.Layer):
+    def __init__(self,
+                 in_channels,
+                 out_channels,
+                 stride,
+                 shortcut=True,
+                 if_first=False,
+                 name=None):
+        super(BottleneckBlock, self).__init__()
+
+        self.conv0 = ConvBNLayer(
+            in_channels=in_channels,
+            out_channels=out_channels,
+            kernel_size=1,
+            act='relu',
+            name=name + "_branch2a")
+        self.conv1 = ConvBNLayer(
+            in_channels=out_channels,
+            out_channels=out_channels,
+            kernel_size=3,
+            stride=stride,
+            act='relu',
+            name=name + "_branch2b")
+        self.conv2 = ConvBNLayer(
+            in_channels=out_channels,
+            out_channels=out_channels * 4,
+            kernel_size=1,
+            act=None,
+            name=name + "_branch2c")
+
+        if not shortcut:
+            self.short = ConvBNLayer(
+                in_channels=in_channels,
+                out_channels=out_channels * 4,
+                kernel_size=1,
+                stride=1,
+                is_vd_mode=False if if_first else True,
+                name=name + "_branch1")
+
+        self.shortcut = shortcut
+
+    def forward(self, inputs):
+        y = self.conv0(inputs)
+        conv1 = self.conv1(y)
+        conv2 = self.conv2(conv1)
+
+        if self.shortcut:
+            short = inputs
+        else:
+            short = self.short(inputs)
+        y = paddle.add(x=short, y=conv2)
+        y = F.relu(y)
+        return y
+
+
+class BasicBlock(nn.Layer):
+    def __init__(self,
+                 in_channels,
+                 out_channels,
+                 stride,
+                 shortcut=True,
+                 if_first=False,
+                 name=None):
+        super(BasicBlock, self).__init__()
+        self.stride = stride
+        self.conv0 = ConvBNLayer(
+            in_channels=in_channels,
+            out_channels=out_channels,
+            kernel_size=3,
+            stride=stride,
+            act='relu',
+            name=name + "_branch2a")
+        self.conv1 = ConvBNLayer(
+            in_channels=out_channels,
+            out_channels=out_channels,
+            kernel_size=3,
+            act=None,
+            name=name + "_branch2b")
+
+        if not shortcut:
+            self.short = ConvBNLayer(
+                in_channels=in_channels,
+                out_channels=out_channels,
+                kernel_size=1,
+                stride=1,
+                is_vd_mode=False if if_first else True,
+                name=name + "_branch1")
+
+        self.shortcut = shortcut
+
+    def forward(self, inputs):
+        y = self.conv0(inputs)
+        conv1 = self.conv1(y)
+
+        if self.shortcut:
+            short = inputs
+        else:
+            short = self.short(inputs)
+        y = paddle.add(x=short, y=conv1)
+        y = F.relu(y)
+        return y
+
+
+class ResNet_SAST(nn.Layer):
+    def __init__(self, in_channels=3, layers=50, **kwargs):
+        super(ResNet_SAST, self).__init__()
+
+        self.layers = layers
+        supported_layers = [18, 34, 50, 101, 152, 200]
+        assert layers in supported_layers, \
+            "supported layers are {} but input layer is {}".format(
+                supported_layers, layers)
+
+        if layers == 18:
+            depth = [2, 2, 2, 2]
+        elif layers == 34 or layers == 50:
+            # depth = [3, 4, 6, 3]
+            depth = [3, 4, 6, 3, 3]
+        elif layers == 101:
+            depth = [3, 4, 23, 3]
+        elif layers == 152:
+            depth = [3, 8, 36, 3]
+        elif layers == 200:
+            depth = [3, 12, 48, 3]
+        # num_channels = [64, 256, 512,
+        #                 1024] if layers >= 50 else [64, 64, 128, 256]
+        # num_filters = [64, 128, 256, 512]
+        num_channels = [64, 256, 512,
+                        1024, 2048] if layers >= 50 else [64, 64, 128, 256]
+        num_filters = [64, 128, 256, 512, 512]
+
+        self.conv1_1 = ConvBNLayer(
+            in_channels=in_channels,
+            out_channels=32,
+            kernel_size=3,
+            stride=2,
+            act='relu',
+            name="conv1_1")
+        self.conv1_2 = ConvBNLayer(
+            in_channels=32,
+            out_channels=32,
+            kernel_size=3,
+            stride=1,
+            act='relu',
+            name="conv1_2")
+        self.conv1_3 = ConvBNLayer(
+            in_channels=32,
+            out_channels=64,
+            kernel_size=3,
+            stride=1,
+            act='relu',
+            name="conv1_3")
+        self.pool2d_max = nn.MaxPool2D(kernel_size=3, stride=2, padding=1)
+
+        self.stages = []
+        self.out_channels = [3, 64]
+        if layers >= 50:
+            for block in range(len(depth)):
+                block_list = []
+                shortcut = False
+                for i in range(depth[block]):
+                    if layers in [101, 152] and block == 2:
+                        if i == 0:
+                            conv_name = "res" + str(block + 2) + "a"
+                        else:
+                            conv_name = "res" + str(block + 2) + "b" + str(i)
+                    else:
+                        conv_name = "res" + str(block + 2) + chr(97 + i)
+                    bottleneck_block = self.add_sublayer(
+                        'bb_%d_%d' % (block, i),
+                        BottleneckBlock(
+                            in_channels=num_channels[block]
+                            if i == 0 else num_filters[block] * 4,
+                            out_channels=num_filters[block],
+                            stride=2 if i == 0 and block != 0 else 1,
+                            shortcut=shortcut,
+                            if_first=block == i == 0,
+                            name=conv_name))
+                    shortcut = True
+                    block_list.append(bottleneck_block)
+                self.out_channels.append(num_filters[block] * 4)
+                self.stages.append(nn.Sequential(*block_list))
+        else:
+            for block in range(len(depth)):
+                block_list = []
+                shortcut = False
+                for i in range(depth[block]):
+                    conv_name = "res" + str(block + 2) + chr(97 + i)
+                    basic_block = self.add_sublayer(
+                        'bb_%d_%d' % (block, i),
+                        BasicBlock(
+                            in_channels=num_channels[block]
+                            if i == 0 else num_filters[block],
+                            out_channels=num_filters[block],
+                            stride=2 if i == 0 and block != 0 else 1,
+                            shortcut=shortcut,
+                            if_first=block == i == 0,
+                            name=conv_name))
+                    shortcut = True
+                    block_list.append(basic_block)
+                self.out_channels.append(num_filters[block])
+                self.stages.append(nn.Sequential(*block_list))
+
+    def forward(self, inputs):
+        out = [inputs]
+        y = self.conv1_1(inputs)
+        y = self.conv1_2(y)
+        y = self.conv1_3(y)
+        out.append(y)
+        y = self.pool2d_max(y)
+        for block in self.stages:
+            y = block(y)
+            out.append(y)
+        return out
\ No newline at end of file

ppocr/modeling/heads/__init__.py

@@ -18,13 +18,15 @@ __all__ = ['build_head']
 def build_head(config):
     # det head
     from .det_db_head import DBHead
+    from .det_east_head import EASTHead
+    from .det_sast_head import SASTHead
 
     # rec head
     from .rec_ctc_head import CTCHead
 
     # cls head
     from .cls_head import ClsHead
-    support_dict = ['DBHead', 'CTCHead', 'ClsHead']
+    support_dict = ['DBHead', 'EASTHead', 'SASTHead', 'CTCHead', 'ClsHead']
 
     module_name = config.pop('name')
     assert module_name in support_dict, Exception('head only support {}'.format(

ppocr/modeling/heads/det_east_head.py

+# copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import math
+import paddle
+from paddle import nn
+import paddle.nn.functional as F
+from paddle import ParamAttr
+
+
+class ConvBNLayer(nn.Layer):
+    def __init__(self,
+                 in_channels,
+                 out_channels,
+                 kernel_size,
+                 stride,
+                 padding,
+                 groups=1,
+                 if_act=True,
+                 act=None,
+                 name=None):
+        super(ConvBNLayer, self).__init__()
+        self.if_act = if_act
+        self.act = act
+        self.conv = nn.Conv2D(
+            in_channels=in_channels,
+            out_channels=out_channels,
+            kernel_size=kernel_size,
+            stride=stride,
+            padding=padding,
+            groups=groups,
+            weight_attr=ParamAttr(name=name + '_weights'),
+            bias_attr=False)
+
+        self.bn = nn.BatchNorm(
+            num_channels=out_channels,
+            act=act,
+            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")
+
+    def forward(self, x):
+        x = self.conv(x)
+        x = self.bn(x)
+        return x
+
+
+class EASTHead(nn.Layer):
+    """
+    """
+    def __init__(self, in_channels, model_name, **kwargs):
+        super(EASTHead, self).__init__()
+        self.model_name = model_name
+        if self.model_name == "large":
+            num_outputs = [128, 64, 1, 8]
+        else:
+            num_outputs = [64, 32, 1, 8]
+
+        self.det_conv1 = ConvBNLayer(
+            in_channels=in_channels,
+            out_channels=num_outputs[0],
+            kernel_size=3,
+            stride=1,
+            padding=1,
+            if_act=True,
+            act='relu',
+            name="det_head1")
+        self.det_conv2 = ConvBNLayer(
+            in_channels=num_outputs[0],
+            out_channels=num_outputs[1],
+            kernel_size=3,
+            stride=1,
+            padding=1,
+            if_act=True,
+            act='relu',
+            name="det_head2")
+        self.score_conv = ConvBNLayer(
+            in_channels=num_outputs[1],
+            out_channels=num_outputs[2],
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            if_act=False,
+            act=None,
+            name="f_score")
+        self.geo_conv = ConvBNLayer(
+            in_channels=num_outputs[1],
+            out_channels=num_outputs[3],
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            if_act=False,
+            act=None,
+            name="f_geo")
+
+    def forward(self, x):
+        f_det = self.det_conv1(x)
+        f_det = self.det_conv2(f_det)
+        f_score = self.score_conv(f_det)
+        f_score = F.sigmoid(f_score)
+        f_geo = self.geo_conv(f_det)
+        f_geo = (F.sigmoid(f_geo) - 0.5) * 2 * 800
+
+        pred = {'f_score': f_score, 'f_geo': f_geo}
+        return pred

ppocr/modeling/heads/det_sast_head.py

+# copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import math
+import paddle
+from paddle import nn
+import paddle.nn.functional as F
+from paddle import ParamAttr
+
+
+class ConvBNLayer(nn.Layer):
+    def __init__(self,
+                 in_channels,
+                 out_channels,
+                 kernel_size,
+                 stride,
+                 groups=1,
+                 if_act=True,
+                 act=None,
+                 name=None):
+        super(ConvBNLayer, self).__init__()
+        self.if_act = if_act
+        self.act = act
+        self.conv = nn.Conv2D(
+            in_channels=in_channels,
+            out_channels=out_channels,
+            kernel_size=kernel_size,
+            stride=stride,
+            padding=(kernel_size - 1) // 2,
+            groups=groups,
+            weight_attr=ParamAttr(name=name + '_weights'),
+            bias_attr=False)
+
+        self.bn = nn.BatchNorm(
+            num_channels=out_channels,
+            act=act,
+            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")
+
+    def forward(self, x):
+        x = self.conv(x)
+        x = self.bn(x)
+        return x
+
+
+class SAST_Header1(nn.Layer):
+    def __init__(self, in_channels, **kwargs):
+        super(SAST_Header1, self).__init__()
+        out_channels = [64, 64, 128]
+        self.score_conv = nn.Sequential(
+            ConvBNLayer(in_channels, out_channels[0], 1, 1, act='relu', name='f_score1'),
+            ConvBNLayer(out_channels[0], out_channels[1], 3, 1, act='relu', name='f_score2'),
+            ConvBNLayer(out_channels[1], out_channels[2], 1, 1, act='relu', name='f_score3'),
+            ConvBNLayer(out_channels[2], 1, 3, 1, act=None, name='f_score4')
+        )
+        self.border_conv = nn.Sequential(
+            ConvBNLayer(in_channels, out_channels[0], 1, 1, act='relu', name='f_border1'),
+            ConvBNLayer(out_channels[0], out_channels[1], 3, 1, act='relu', name='f_border2'),
+            ConvBNLayer(out_channels[1], out_channels[2], 1, 1, act='relu', name='f_border3'),
+            ConvBNLayer(out_channels[2], 4, 3, 1, act=None, name='f_border4')            
+        )
+
+    def forward(self, x):
+        f_score = self.score_conv(x)
+        f_score = F.sigmoid(f_score)
+        f_border = self.border_conv(x)
+        return f_score, f_border
+
+
+class SAST_Header2(nn.Layer):
+    def __init__(self, in_channels, **kwargs):
+        super(SAST_Header2, self).__init__()
+        out_channels = [64, 64, 128]
+        self.tvo_conv = nn.Sequential(
+            ConvBNLayer(in_channels, out_channels[0], 1, 1, act='relu', name='f_tvo1'),
+            ConvBNLayer(out_channels[0], out_channels[1], 3, 1, act='relu', name='f_tvo2'),
+            ConvBNLayer(out_channels[1], out_channels[2], 1, 1, act='relu', name='f_tvo3'),
+            ConvBNLayer(out_channels[2], 8, 3, 1, act=None, name='f_tvo4')
+        )
+        self.tco_conv = nn.Sequential(
+            ConvBNLayer(in_channels, out_channels[0], 1, 1, act='relu', name='f_tco1'),
+            ConvBNLayer(out_channels[0], out_channels[1], 3, 1, act='relu', name='f_tco2'),
+            ConvBNLayer(out_channels[1], out_channels[2], 1, 1, act='relu', name='f_tco3'),
+            ConvBNLayer(out_channels[2], 2, 3, 1, act=None, name='f_tco4')            
+        )
+
+    def forward(self, x):
+        f_tvo = self.tvo_conv(x)
+        f_tco = self.tco_conv(x)
+        return f_tvo, f_tco
+
+
+class SASTHead(nn.Layer):
+    """
+    """
+    def __init__(self, in_channels, **kwargs):
+        super(SASTHead, self).__init__()
+
+        self.head1 = SAST_Header1(in_channels)
+        self.head2 = SAST_Header2(in_channels)
+
+    def forward(self, x):
+        f_score, f_border = self.head1(x)
+        f_tvo, f_tco = self.head2(x)
+
+        predicts = {}
+        predicts['f_score'] = f_score
+        predicts['f_border'] = f_border
+        predicts['f_tvo'] = f_tvo
+        predicts['f_tco'] = f_tco
+        return predicts
\ No newline at end of file

ppocr/modeling/necks/__init__.py

@@ -16,8 +16,10 @@ __all__ = ['build_neck']
 
 def build_neck(config):
     from .db_fpn import DBFPN
+    from .east_fpn import EASTFPN
+    from .sast_fpn import SASTFPN
     from .rnn import SequenceEncoder
-    support_dict = ['DBFPN', 'SequenceEncoder']
+    support_dict = ['DBFPN', 'EASTFPN', 'SASTFPN', 'SequenceEncoder']
 
     module_name = config.pop('name')
     assert module_name in support_dict, Exception('neck only support {}'.format(

ppocr/modeling/necks/east_fpn.py

+# copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import paddle
+from paddle import nn
+import paddle.nn.functional as F
+from paddle import ParamAttr
+
+
+class ConvBNLayer(nn.Layer):
+    def __init__(self,
+                 in_channels,
+                 out_channels,
+                 kernel_size,
+                 stride,
+                 padding,
+                 groups=1,
+                 if_act=True,
+                 act=None,
+                 name=None):
+        super(ConvBNLayer, self).__init__()
+        self.if_act = if_act
+        self.act = act
+        self.conv = nn.Conv2D(
+            in_channels=in_channels,
+            out_channels=out_channels,
+            kernel_size=kernel_size,
+            stride=stride,
+            padding=padding,
+            groups=groups,
+            weight_attr=ParamAttr(name=name + '_weights'),
+            bias_attr=False)
+
+        self.bn = nn.BatchNorm(
+            num_channels=out_channels,
+            act=act,
+            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")
+
+    def forward(self, x):
+        x = self.conv(x)
+        x = self.bn(x)
+        return x
+
+
+class DeConvBNLayer(nn.Layer):
+    def __init__(self,
+                 in_channels,
+                 out_channels,
+                 kernel_size,
+                 stride,
+                 padding,
+                 groups=1,
+                 if_act=True,
+                 act=None,
+                 name=None):
+        super(DeConvBNLayer, self).__init__()
+        self.if_act = if_act
+        self.act = act
+        self.deconv = nn.Conv2DTranspose(
+            in_channels=in_channels,
+            out_channels=out_channels,
+            kernel_size=kernel_size,
+            stride=stride,
+            padding=padding,
+            groups=groups,
+            weight_attr=ParamAttr(name=name + '_weights'),
+            bias_attr=False)
+        self.bn = nn.BatchNorm(
+            num_channels=out_channels,
+            act=act,
+            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")
+
+    def forward(self, x):
+        x = self.deconv(x)
+        x = self.bn(x)
+        return x
+
+
+class EASTFPN(nn.Layer):
+    def __init__(self, in_channels, model_name, **kwargs):
+        super(EASTFPN, self).__init__()
+        self.model_name = model_name
+        if self.model_name == "large":
+            self.out_channels = 128
+        else:
+            self.out_channels = 64
+        self.in_channels = in_channels[::-1]
+        self.h1_conv = ConvBNLayer(
+            in_channels=self.out_channels+self.in_channels[1],
+            out_channels=self.out_channels,
+            kernel_size=3,
+            stride=1,
+            padding=1,
+            if_act=True,
+            act='relu',
+            name="unet_h_1")
+        self.h2_conv = ConvBNLayer(
+            in_channels=self.out_channels+self.in_channels[2],
+            out_channels=self.out_channels,
+            kernel_size=3,
+            stride=1,
+            padding=1,
+            if_act=True,
+            act='relu',
+            name="unet_h_2")
+        self.h3_conv = ConvBNLayer(
+            in_channels=self.out_channels+self.in_channels[3],
+            out_channels=self.out_channels,
+            kernel_size=3,
+            stride=1,
+            padding=1,
+            if_act=True,
+            act='relu',
+            name="unet_h_3")
+        self.g0_deconv = DeConvBNLayer(
+            in_channels=self.in_channels[0],
+            out_channels=self.out_channels,
+            kernel_size=4,
+            stride=2,
+            padding=1,
+            if_act=True,
+            act='relu',
+            name="unet_g_0")
+        self.g1_deconv = DeConvBNLayer(
+            in_channels=self.out_channels,
+            out_channels=self.out_channels,
+            kernel_size=4,
+            stride=2,
+            padding=1,
+            if_act=True,
+            act='relu',
+            name="unet_g_1")
+        self.g2_deconv = DeConvBNLayer(
+            in_channels=self.out_channels,
+            out_channels=self.out_channels,
+            kernel_size=4,
+            stride=2,
+            padding=1,
+            if_act=True,
+            act='relu',
+            name="unet_g_2")
+        self.g3_conv = ConvBNLayer(
+            in_channels=self.out_channels,
+            out_channels=self.out_channels,
+            kernel_size=3,
+            stride=1,
+            padding=1,
+            if_act=True,
+            act='relu',
+            name="unet_g_3")
+
+    def forward(self, x):
+        f = x[::-1]
+
+        h = f[0]
+        g = self.g0_deconv(h)
+        h = paddle.concat([g, f[1]], axis=1)
+        h = self.h1_conv(h)
+        g = self.g1_deconv(h)
+        h = paddle.concat([g, f[2]], axis=1)
+        h = self.h2_conv(h)
+        g = self.g2_deconv(h)
+        h = paddle.concat([g, f[3]], axis=1)
+        h = self.h3_conv(h)
+        g = self.g3_conv(h)
+
+        return g
\ No newline at end of file

ppocr/modeling/necks/sast_fpn.py

+# copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import paddle
+from paddle import nn
+import paddle.nn.functional as F
+from paddle import ParamAttr
+
+
+class ConvBNLayer(nn.Layer):
+    def __init__(self,
+                 in_channels,
+                 out_channels,
+                 kernel_size,
+                 stride,
+                 groups=1,
+                 if_act=True,
+                 act=None,
+                 name=None):
+        super(ConvBNLayer, self).__init__()
+        self.if_act = if_act
+        self.act = act
+        self.conv = nn.Conv2D(
+            in_channels=in_channels,
+            out_channels=out_channels,
+            kernel_size=kernel_size,
+            stride=stride,
+            padding=(kernel_size - 1) // 2,
+            groups=groups,
+            weight_attr=ParamAttr(name=name + '_weights'),
+            bias_attr=False)
+  
+        self.bn = nn.BatchNorm(
+            num_channels=out_channels,
+            act=act,
+            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")
+
+    def forward(self, x):
+        x = self.conv(x)
+        x = self.bn(x)
+        return x
+
+
+class DeConvBNLayer(nn.Layer):
+    def __init__(self,
+                 in_channels,
+                 out_channels,
+                 kernel_size,
+                 stride,
+                 groups=1,
+                 if_act=True,
+                 act=None,
+                 name=None):
+        super(DeConvBNLayer, self).__init__()
+        self.if_act = if_act
+        self.act = act
+        self.deconv = nn.Conv2DTranspose(
+            in_channels=in_channels,
+            out_channels=out_channels,
+            kernel_size=kernel_size,
+            stride=stride,
+            padding=(kernel_size - 1) // 2,
+            groups=groups,
+            weight_attr=ParamAttr(name=name + '_weights'),
+            bias_attr=False)
+        self.bn = nn.BatchNorm(
+            num_channels=out_channels,
+            act=act,
+            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")
+
+    def forward(self, x):
+        x = self.deconv(x)
+        x = self.bn(x)
+        return x
+
+
+class FPN_Up_Fusion(nn.Layer):
+    def __init__(self, in_channels):
+        super(FPN_Up_Fusion, self).__init__()
+        in_channels = in_channels[::-1]
+        out_channels = [256, 256, 192, 192, 128]
+                
+        self.h0_conv = ConvBNLayer(in_channels[0], out_channels[0], 1, 1, act=None, name='fpn_up_h0')
+        self.h1_conv = ConvBNLayer(in_channels[1], out_channels[1], 1, 1, act=None, name='fpn_up_h1')
+        self.h2_conv = ConvBNLayer(in_channels[2], out_channels[2], 1, 1, act=None, name='fpn_up_h2')
+        self.h3_conv = ConvBNLayer(in_channels[3], out_channels[3], 1, 1, act=None, name='fpn_up_h3')
+        self.h4_conv = ConvBNLayer(in_channels[4], out_channels[4], 1, 1, act=None, name='fpn_up_h4')
+
+        self.g0_conv = DeConvBNLayer(out_channels[0], out_channels[1], 4, 2, act=None, name='fpn_up_g0')
+
+        self.g1_conv = nn.Sequential(
+            ConvBNLayer(out_channels[1], out_channels[1], 3, 1, act='relu', name='fpn_up_g1_1'),
+            DeConvBNLayer(out_channels[1], out_channels[2], 4, 2, act=None, name='fpn_up_g1_2')
+        )
+        self.g2_conv = nn.Sequential(
+            ConvBNLayer(out_channels[2], out_channels[2], 3, 1, act='relu', name='fpn_up_g2_1'),
+            DeConvBNLayer(out_channels[2], out_channels[3], 4, 2, act=None, name='fpn_up_g2_2')
+        )
+        self.g3_conv = nn.Sequential(
+            ConvBNLayer(out_channels[3], out_channels[3], 3, 1, act='relu', name='fpn_up_g3_1'),
+            DeConvBNLayer(out_channels[3], out_channels[4], 4, 2, act=None, name='fpn_up_g3_2')
+        )
+
+        self.g4_conv = nn.Sequential(
+            ConvBNLayer(out_channels[4], out_channels[4], 3, 1, act='relu', name='fpn_up_fusion_1'),
+            ConvBNLayer(out_channels[4], out_channels[4], 1, 1, act=None, name='fpn_up_fusion_2')
+        )
+
+    def _add_relu(self, x1, x2):
+        x = paddle.add(x=x1, y=x2)
+        x = F.relu(x)
+        return x
+
+    def forward(self, x):
+        f = x[2:][::-1]
+        h0 = self.h0_conv(f[0])
+        h1 = self.h1_conv(f[1])
+        h2 = self.h2_conv(f[2])
+        h3 = self.h3_conv(f[3])
+        h4 = self.h4_conv(f[4])
+
+        g0 = self.g0_conv(h0)
+        g1 = self._add_relu(g0, h1)
+        g1 = self.g1_conv(g1)
+        g2 = self.g2_conv(self._add_relu(g1, h2))
+        g3 = self.g3_conv(self._add_relu(g2, h3))
+        g4 = self.g4_conv(self._add_relu(g3, h4))
+
+        return g4
+
+
+class FPN_Down_Fusion(nn.Layer):
+    def __init__(self, in_channels):
+        super(FPN_Down_Fusion, self).__init__()
+        out_channels = [32, 64, 128]
+
+        self.h0_conv = ConvBNLayer(in_channels[0], out_channels[0], 3, 1, act=None, name='fpn_down_h0')
+        self.h1_conv = ConvBNLayer(in_channels[1], out_channels[1], 3, 1, act=None, name='fpn_down_h1')
+        self.h2_conv = ConvBNLayer(in_channels[2], out_channels[2], 3, 1, act=None, name='fpn_down_h2')
+
+        self.g0_conv = ConvBNLayer(out_channels[0], out_channels[1], 3, 2, act=None, name='fpn_down_g0')
+
+        self.g1_conv = nn.Sequential(
+            ConvBNLayer(out_channels[1], out_channels[1], 3, 1, act='relu', name='fpn_down_g1_1'),
+            ConvBNLayer(out_channels[1], out_channels[2], 3, 2, act=None, name='fpn_down_g1_2')            
+        )
+
+        self.g2_conv = nn.Sequential(
+            ConvBNLayer(out_channels[2], out_channels[2], 3, 1, act='relu', name='fpn_down_fusion_1'),
+            ConvBNLayer(out_channels[2], out_channels[2], 1, 1, act=None, name='fpn_down_fusion_2')            
+        )
+
+    def forward(self, x):
+        f = x[:3]
+        h0 = self.h0_conv(f[0])
+        h1 = self.h1_conv(f[1])
+        h2 = self.h2_conv(f[2])
+        g0 = self.g0_conv(h0)
+        g1 = paddle.add(x=g0, y=h1)
+        g1 = F.relu(g1)
+        g1 = self.g1_conv(g1)
+        g2 = paddle.add(x=g1, y=h2)
+        g2 = F.relu(g2)
+        g2 = self.g2_conv(g2)
+        return g2
+
+
+class Cross_Attention(nn.Layer):
+    def __init__(self, in_channels):
+        super(Cross_Attention, self).__init__()
+        self.theta_conv = ConvBNLayer(in_channels, in_channels, 1, 1, act='relu', name='f_theta')
+        self.phi_conv = ConvBNLayer(in_channels, in_channels, 1, 1, act='relu', name='f_phi')
+        self.g_conv = ConvBNLayer(in_channels, in_channels, 1, 1, act='relu', name='f_g')
+
+        self.fh_weight_conv = ConvBNLayer(in_channels, in_channels, 1, 1, act=None, name='fh_weight')
+        self.fh_sc_conv = ConvBNLayer(in_channels, in_channels, 1, 1, act=None, name='fh_sc')
+
+        self.fv_weight_conv = ConvBNLayer(in_channels, in_channels, 1, 1, act=None, name='fv_weight')
+        self.fv_sc_conv = ConvBNLayer(in_channels, in_channels, 1, 1, act=None, name='fv_sc')
+
+        self.f_attn_conv = ConvBNLayer(in_channels * 2, in_channels, 1, 1, act='relu', name='f_attn')
+
+    def _cal_fweight(self, f, shape):
+        f_theta, f_phi, f_g = f
+        #flatten
+        f_theta = paddle.transpose(f_theta, [0, 2, 3, 1])
+        f_theta = paddle.reshape(f_theta, [shape[0] * shape[1], shape[2], 128])
+        f_phi = paddle.transpose(f_phi, [0, 2, 3, 1])
+        f_phi = paddle.reshape(f_phi, [shape[0] * shape[1], shape[2], 128])
+        f_g = paddle.transpose(f_g, [0, 2, 3, 1])
+        f_g = paddle.reshape(f_g, [shape[0] * shape[1], shape[2], 128])
+        #correlation
+        f_attn = paddle.matmul(f_theta, paddle.transpose(f_phi, [0, 2, 1]))
+        #scale
+        f_attn = f_attn / (128**0.5)
+        f_attn = F.softmax(f_attn)
+        #weighted sum
+        f_weight = paddle.matmul(f_attn, f_g)
+        f_weight = paddle.reshape(
+            f_weight, [shape[0], shape[1], shape[2], 128])
+        return f_weight
+
+    def forward(self, f_common):
+        f_shape = paddle.shape(f_common)
+        # print('f_shape: ', f_shape)
+
+        f_theta = self.theta_conv(f_common)
+        f_phi = self.phi_conv(f_common)
+        f_g = self.g_conv(f_common)
+
+        ######## horizon ########
+        fh_weight = self._cal_fweight([f_theta, f_phi, f_g], 
+                                        [f_shape[0], f_shape[2], f_shape[3]])
+        fh_weight = paddle.transpose(fh_weight, [0, 3, 1, 2])
+        fh_weight = self.fh_weight_conv(fh_weight)
+        #short cut
+        fh_sc = self.fh_sc_conv(f_common)
+        f_h = F.relu(fh_weight + fh_sc)
+
+        ######## vertical ########
+        fv_theta = paddle.transpose(f_theta, [0, 1, 3, 2])
+        fv_phi = paddle.transpose(f_phi, [0, 1, 3, 2])
+        fv_g = paddle.transpose(f_g, [0, 1, 3, 2])
+        fv_weight = self._cal_fweight([fv_theta, fv_phi, fv_g], 
+                                        [f_shape[0], f_shape[3], f_shape[2]])
+        fv_weight = paddle.transpose(fv_weight, [0, 3, 2, 1])
+        fv_weight = self.fv_weight_conv(fv_weight)
+        #short cut
+        fv_sc = self.fv_sc_conv(f_common)
+        f_v = F.relu(fv_weight + fv_sc)
+
+        ######## merge ########
+        f_attn = paddle.concat([f_h, f_v], axis=1)
+        f_attn = self.f_attn_conv(f_attn)
+        return f_attn
+
+
+class SASTFPN(nn.Layer):
+    def __init__(self, in_channels, with_cab=False, **kwargs):
+        super(SASTFPN, self).__init__()
+        self.in_channels = in_channels
+        self.with_cab = with_cab
+        self.FPN_Down_Fusion = FPN_Down_Fusion(self.in_channels)
+        self.FPN_Up_Fusion = FPN_Up_Fusion(self.in_channels)
+        self.out_channels = 128
+        self.cross_attention = Cross_Attention(self.out_channels)
+
+    def forward(self, x):
+        #down fpn
+        f_down = self.FPN_Down_Fusion(x)
+
+        #up fpn
+        f_up = self.FPN_Up_Fusion(x)
+
+        #fusion
+        f_common = paddle.add(x=f_down, y=f_up)
+        f_common = F.relu(f_common)
+
+        if self.with_cab:
+            # print('enhence f_common with CAB.')
+            f_common = self.cross_attention(f_common)
+
+        return f_common