Merge branch 'PaddlePaddle:dygraph' into dygraph

doc/doc_en/training_en.md

@@ -94,7 +94,7 @@ The current open source models, data sets and magnitudes are as follows:
     - Chinese data set, LSVT street view data set crops the image according to the truth value, and performs position calibration, a total of 30w images. In addition, based on the LSVT corpus, 500w of synthesized data.
     - Small language data set, using different corpora and fonts, respectively generated 100w synthetic data set, and using ICDAR-MLT as the verification set.
 
-Among them, the public data sets are all open source, users can search and download by themselves, or refer to [Chinese data set](./datasets_en.md), synthetic data is not open source, users can use open source synthesis tools to synthesize by themselves. Synthesis tools include [text_renderer](https://github.com/Sanster/text_renderer), [SynthText](https://github.com/ankush-me/SynthText), [TextRecognitionDataGenerator](https://github.com/Belval/TextRecognitionDataGenerator) etc.
+Among them, the public data sets are all open source, users can search and download by themselves, or refer to [Chinese data set](dataset/datasets_en.md), synthetic data is not open source, users can use open source synthesis tools to synthesize by themselves. Synthesis tools include [text_renderer](https://github.com/Sanster/text_renderer), [SynthText](https://github.com/ankush-me/SynthText), [TextRecognitionDataGenerator](https://github.com/Belval/TextRecognitionDataGenerator) etc.
 
 <a name="22-vertical-scene"></a>
 

doc/doc_en/update_en.md

@@ -19,7 +19,7 @@
 - 2020.7.15, Add several related datasets, data annotation and synthesis tools.
 - 2020.7.9 Add a new model to support recognize the  character "space".
 - 2020.7.9 Add the data augument and learning rate decay strategies during training.
-- 2020.6.8 Add [datasets](./datasets_en.md) and keep updating
+- 2020.6.8 Add [datasets](dataset/datasets_en.md) and keep updating
 - 2020.6.5 Support exporting `attention` model to `inference_model`
 - 2020.6.5 Support separate prediction and recognition, output result score
 - 2020.5.30 Provide Lightweight Chinese OCR online experience

ppocr/data/__init__.py

@@ -72,6 +72,7 @@ def build_dataloader(config, mode, device, logger, seed=None):
         use_shared_memory = loader_config['use_shared_memory']
     else:
         use_shared_memory = True
+
     if mode == "Train":
         # Distribute data to multiple cards
         batch_sampler = DistributedBatchSampler(

ppocr/data/collate_fn.py

@@ -56,3 +56,17 @@ class ListCollator(object):
         for idx in to_tensor_idxs:
             data_dict[idx] = paddle.to_tensor(data_dict[idx])
         return list(data_dict.values())
+
+
+class SSLRotateCollate(object):
+    """
+    bach: [
+        [(4*3xH*W), (4,)]
+        [(4*3xH*W), (4,)]
+        ...
+    ]
+    """
+
+    def __call__(self, batch):
+        output = [np.concatenate(d, axis=0) for d in zip(*batch)]
+        return output

ppocr/data/imaug/__init__.py

@@ -24,6 +24,7 @@ from .make_pse_gt import MakePseGt
 
 from .rec_img_aug import RecAug, RecConAug, RecResizeImg, ClsResizeImg, \
     SRNRecResizeImg, NRTRRecResizeImg, SARRecResizeImg, PRENResizeImg, SVTRRecResizeImg
+from .ssl_img_aug import SSLRotateResize
 from .randaugment import RandAugment
 from .copy_paste import CopyPaste
 from .ColorJitter import ColorJitter

ppocr/data/imaug/label_ops.py

@@ -113,14 +113,14 @@ class BaseRecLabelEncode(object):
             dict_character = list(self.character_str)
             self.lower = True
         else:
-            self.character_str = ""
+            self.character_str = []
             with open(character_dict_path, "rb") as fin:
                 lines = fin.readlines()
                 for line in lines:
                     line = line.decode('utf-8').strip("\n").strip("\r\n")
-                    self.character_str += line
+                    self.character_str.append(line)
             if use_space_char:
-                self.character_str += " "
+                self.character_str.append(" ")
             dict_character = list(self.character_str)
         dict_character = self.add_special_char(dict_character)
         self.dict = {}

ppocr/data/imaug/ssl_img_aug.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 random
+from PIL import Image
+
+from .rec_img_aug import resize_norm_img
+
+
+class SSLRotateResize(object):
+    def __init__(self,
+                 image_shape,
+                 padding=False,
+                 select_all=True,
+                 mode="train",
+                 **kwargs):
+        self.image_shape = image_shape
+        self.padding = padding
+        self.select_all = select_all
+        self.mode = mode
+
+    def __call__(self, data):
+        img = data["image"]
+
+        data["image_r90"] = cv2.rotate(img, cv2.ROTATE_90_CLOCKWISE)
+        data["image_r180"] = cv2.rotate(data["image_r90"],
+                                        cv2.ROTATE_90_CLOCKWISE)
+        data["image_r270"] = cv2.rotate(data["image_r180"],
+                                        cv2.ROTATE_90_CLOCKWISE)
+
+        images = []
+        for key in ["image", "image_r90", "image_r180", "image_r270"]:
+            images.append(
+                resize_norm_img(
+                    data.pop(key),
+                    image_shape=self.image_shape,
+                    padding=self.padding)[0])
+        data["image"] = np.stack(images, axis=0)
+        data["label"] = np.array(list(range(4)))
+        if not self.select_all:
+            data["image"] = data["image"][0::2]  # just choose 0 and 180
+            data["label"] = data["label"][0:2]  # label needs to be continuous
+        if self.mode == "test":
+            data["image"] = data["image"][0]
+            data["label"] = data["label"][0]
+        return data

ppocr/modeling/heads/det_db_head.py

@@ -31,13 +31,14 @@ def get_bias_attr(k):
 
 
 class Head(nn.Layer):
-    def __init__(self, in_channels, name_list):
+    def __init__(self, in_channels, name_list, kernel_list=[3, 2, 2], **kwargs):
         super(Head, self).__init__()
+
         self.conv1 = nn.Conv2D(
             in_channels=in_channels,
             out_channels=in_channels // 4,
-            kernel_size=3,
-            padding=1,
+            kernel_size=kernel_list[0],
+            padding=int(kernel_list[0] // 2),
             weight_attr=ParamAttr(),
             bias_attr=False)
         self.conv_bn1 = nn.BatchNorm(
@@ -50,7 +51,7 @@ class Head(nn.Layer):
         self.conv2 = nn.Conv2DTranspose(
             in_channels=in_channels // 4,
             out_channels=in_channels // 4,
-            kernel_size=2,
+            kernel_size=kernel_list[1],
             stride=2,
             weight_attr=ParamAttr(
                 initializer=paddle.nn.initializer.KaimingUniform()),
@@ -65,7 +66,7 @@ class Head(nn.Layer):
         self.conv3 = nn.Conv2DTranspose(
             in_channels=in_channels // 4,
             out_channels=1,
-            kernel_size=2,
+            kernel_size=kernel_list[2],
             stride=2,
             weight_attr=ParamAttr(
                 initializer=paddle.nn.initializer.KaimingUniform()),
@@ -100,8 +101,8 @@ class DBHead(nn.Layer):
             'conv2d_57', 'batch_norm_49', 'conv2d_transpose_2', 'batch_norm_50',
             'conv2d_transpose_3', 'thresh'
         ]
-        self.binarize = Head(in_channels, binarize_name_list)
-        self.thresh = Head(in_channels, thresh_name_list)
+        self.binarize = Head(in_channels, binarize_name_list, **kwargs)
+        self.thresh = Head(in_channels, thresh_name_list, **kwargs)
 
     def step_function(self, x, y):
         return paddle.reciprocal(1 + paddle.exp(-self.k * (x - y)))

ppocr/modeling/necks/__init__.py

@@ -16,7 +16,7 @@ __all__ = ['build_neck']
 
 
 def build_neck(config):
-    from .db_fpn import DBFPN
+    from .db_fpn import DBFPN, RSEFPN, LKPAN
     from .east_fpn import EASTFPN
     from .sast_fpn import SASTFPN
     from .rnn import SequenceEncoder
@@ -26,8 +26,8 @@ def build_neck(config):
     from .fce_fpn import FCEFPN
     from .pren_fpn import PRENFPN
     support_dict = [
-        'FPN', 'FCEFPN', 'DBFPN', 'EASTFPN', 'SASTFPN', 'SequenceEncoder',
-        'PGFPN', 'TableFPN', 'PRENFPN'
+        'FPN', 'FCEFPN', 'LKPAN', 'DBFPN', 'RSEFPN', 'EASTFPN', 'SASTFPN',
+        'SequenceEncoder', 'PGFPN', 'TableFPN', 'PRENFPN'
     ]
 
     module_name = config.pop('name')

ppocr/modeling/necks/db_fpn.py

@@ -20,6 +20,88 @@ import paddle
 from paddle import nn
 import paddle.nn.functional as F
 from paddle import ParamAttr
+import os
+import sys
+
+__dir__ = os.path.dirname(os.path.abspath(__file__))
+sys.path.append(__dir__)
+sys.path.insert(0, os.path.abspath(os.path.join(__dir__, '../../..')))
+
+from ppocr.modeling.backbones.det_mobilenet_v3 import SEModule
+
+
+class DSConv(nn.Layer):
+    def __init__(self,
+                 in_channels,
+                 out_channels,
+                 kernel_size,
+                 padding,
+                 stride=1,
+                 groups=None,
+                 if_act=True,
+                 act="relu",
+                 **kwargs):
+        super(DSConv, self).__init__()
+        if groups == None:
+            groups = in_channels
+        self.if_act = if_act
+        self.act = act
+        self.conv1 = nn.Conv2D(
+            in_channels=in_channels,
+            out_channels=in_channels,
+            kernel_size=kernel_size,
+            stride=stride,
+            padding=padding,
+            groups=groups,
+            bias_attr=False)
+
+        self.bn1 = nn.BatchNorm(num_channels=in_channels, act=None)
+
+        self.conv2 = nn.Conv2D(
+            in_channels=in_channels,
+            out_channels=int(in_channels * 4),
+            kernel_size=1,
+            stride=1,
+            bias_attr=False)
+
+        self.bn2 = nn.BatchNorm(num_channels=int(in_channels * 4), act=None)
+
+        self.conv3 = nn.Conv2D(
+            in_channels=int(in_channels * 4),
+            out_channels=out_channels,
+            kernel_size=1,
+            stride=1,
+            bias_attr=False)
+        self._c = [in_channels, out_channels]
+        if in_channels != out_channels:
+            self.conv_end = nn.Conv2D(
+                in_channels=in_channels,
+                out_channels=out_channels,
+                kernel_size=1,
+                stride=1,
+                bias_attr=False)
+
+    def forward(self, inputs):
+
+        x = self.conv1(inputs)
+        x = self.bn1(x)
+
+        x = self.conv2(x)
+        x = self.bn2(x)
+        if self.if_act:
+            if self.act == "relu":
+                x = F.relu(x)
+            elif self.act == "hardswish":
+                x = F.hardswish(x)
+            else:
+                print("The activation function({}) is selected incorrectly.".
+                      format(self.act))
+                exit()
+
+        x = self.conv3(x)
+        if self._c[0] != self._c[1]:
+            x = x + self.conv_end(inputs)
+        return x
 
 
 class DBFPN(nn.Layer):
@@ -106,3 +188,171 @@ class DBFPN(nn.Layer):
 
         fuse = paddle.concat([p5, p4, p3, p2], axis=1)
         return fuse
+
+
+class RSELayer(nn.Layer):
+    def __init__(self, in_channels, out_channels, kernel_size, shortcut=True):
+        super(RSELayer, self).__init__()
+        weight_attr = paddle.nn.initializer.KaimingUniform()
+        self.out_channels = out_channels
+        self.in_conv = nn.Conv2D(
+            in_channels=in_channels,
+            out_channels=self.out_channels,
+            kernel_size=kernel_size,
+            padding=int(kernel_size // 2),
+            weight_attr=ParamAttr(initializer=weight_attr),
+            bias_attr=False)
+        self.se_block = SEModule(self.out_channels)
+        self.shortcut = shortcut
+
+    def forward(self, ins):
+        x = self.in_conv(ins)
+        if self.shortcut:
+            out = x + self.se_block(x)
+        else:
+            out = self.se_block(x)
+        return out
+
+
+class RSEFPN(nn.Layer):
+    def __init__(self, in_channels, out_channels, shortcut=True, **kwargs):
+        super(RSEFPN, self).__init__()
+        self.out_channels = out_channels
+        self.ins_conv = nn.LayerList()
+        self.inp_conv = nn.LayerList()
+
+        for i in range(len(in_channels)):
+            self.ins_conv.append(
+                RSELayer(
+                    in_channels[i],
+                    out_channels,
+                    kernel_size=1,
+                    shortcut=shortcut))
+            self.inp_conv.append(
+                RSELayer(
+                    out_channels,
+                    out_channels // 4,
+                    kernel_size=3,
+                    shortcut=shortcut))
+
+    def forward(self, x):
+        c2, c3, c4, c5 = x
+
+        in5 = self.ins_conv[3](c5)
+        in4 = self.ins_conv[2](c4)
+        in3 = self.ins_conv[1](c3)
+        in2 = self.ins_conv[0](c2)
+
+        out4 = in4 + F.upsample(
+            in5, scale_factor=2, mode="nearest", align_mode=1)  # 1/16
+        out3 = in3 + F.upsample(
+            out4, scale_factor=2, mode="nearest", align_mode=1)  # 1/8
+        out2 = in2 + F.upsample(
+            out3, scale_factor=2, mode="nearest", align_mode=1)  # 1/4
+
+        p5 = self.inp_conv[3](in5)
+        p4 = self.inp_conv[2](out4)
+        p3 = self.inp_conv[1](out3)
+        p2 = self.inp_conv[0](out2)
+
+        p5 = F.upsample(p5, scale_factor=8, mode="nearest", align_mode=1)
+        p4 = F.upsample(p4, scale_factor=4, mode="nearest", align_mode=1)
+        p3 = F.upsample(p3, scale_factor=2, mode="nearest", align_mode=1)
+
+        fuse = paddle.concat([p5, p4, p3, p2], axis=1)
+        return fuse
+
+
+class LKPAN(nn.Layer):
+    def __init__(self, in_channels, out_channels, mode='large', **kwargs):
+        super(LKPAN, self).__init__()
+        self.out_channels = out_channels
+        weight_attr = paddle.nn.initializer.KaimingUniform()
+
+        self.ins_conv = nn.LayerList()
+        self.inp_conv = nn.LayerList()
+        # pan head
+        self.pan_head_conv = nn.LayerList()
+        self.pan_lat_conv = nn.LayerList()
+
+        if mode.lower() == 'lite':
+            p_layer = DSConv
+        elif mode.lower() == 'large':
+            p_layer = nn.Conv2D
+        else:
+            raise ValueError(
+                "mode can only be one of ['lite', 'large'], but received {}".
+                format(mode))
+
+        for i in range(len(in_channels)):
+            self.ins_conv.append(
+                nn.Conv2D(
+                    in_channels=in_channels[i],
+                    out_channels=self.out_channels,
+                    kernel_size=1,
+                    weight_attr=ParamAttr(initializer=weight_attr),
+                    bias_attr=False))
+
+            self.inp_conv.append(
+                p_layer(
+                    in_channels=self.out_channels,
+                    out_channels=self.out_channels // 4,
+                    kernel_size=9,
+                    padding=4,
+                    weight_attr=ParamAttr(initializer=weight_attr),
+                    bias_attr=False))
+
+            if i > 0:
+                self.pan_head_conv.append(
+                    nn.Conv2D(
+                        in_channels=self.out_channels // 4,
+                        out_channels=self.out_channels // 4,
+                        kernel_size=3,
+                        padding=1,
+                        stride=2,
+                        weight_attr=ParamAttr(initializer=weight_attr),
+                        bias_attr=False))
+            self.pan_lat_conv.append(
+                p_layer(
+                    in_channels=self.out_channels // 4,
+                    out_channels=self.out_channels // 4,
+                    kernel_size=9,
+                    padding=4,
+                    weight_attr=ParamAttr(initializer=weight_attr),
+                    bias_attr=False))
+
+    def forward(self, x):
+        c2, c3, c4, c5 = x
+
+        in5 = self.ins_conv[3](c5)
+        in4 = self.ins_conv[2](c4)
+        in3 = self.ins_conv[1](c3)
+        in2 = self.ins_conv[0](c2)
+
+        out4 = in4 + F.upsample(
+            in5, scale_factor=2, mode="nearest", align_mode=1)  # 1/16
+        out3 = in3 + F.upsample(
+            out4, scale_factor=2, mode="nearest", align_mode=1)  # 1/8
+        out2 = in2 + F.upsample(
+            out3, scale_factor=2, mode="nearest", align_mode=1)  # 1/4
+
+        f5 = self.inp_conv[3](in5)
+        f4 = self.inp_conv[2](out4)
+        f3 = self.inp_conv[1](out3)
+        f2 = self.inp_conv[0](out2)
+
+        pan3 = f3 + self.pan_head_conv[0](f2)
+        pan4 = f4 + self.pan_head_conv[1](pan3)
+        pan5 = f5 + self.pan_head_conv[2](pan4)
+
+        p2 = self.pan_lat_conv[0](f2)
+        p3 = self.pan_lat_conv[1](pan3)
+        p4 = self.pan_lat_conv[2](pan4)
+        p5 = self.pan_lat_conv[3](pan5)
+
+        p5 = F.upsample(p5, scale_factor=8, mode="nearest", align_mode=1)
+        p4 = F.upsample(p4, scale_factor=4, mode="nearest", align_mode=1)
+        p3 = F.upsample(p3, scale_factor=2, mode="nearest", align_mode=1)
+
+        fuse = paddle.concat([p5, p4, p3, p2], axis=1)
+        return fuse

ppocr/postprocess/cls_postprocess.py

@@ -17,17 +17,26 @@ import paddle
 class ClsPostProcess(object):
     """ Convert between text-label and text-index """
 
-    def __init__(self, label_list, **kwargs):
+    def __init__(self, label_list=None, key=None, **kwargs):
         super(ClsPostProcess, self).__init__()
         self.label_list = label_list
+        self.key = key
 
     def __call__(self, preds, label=None, *args, **kwargs):
+        if self.key is not None:
+            preds = preds[self.key]
+
+        label_list = self.label_list
+        if label_list is None:
+            label_list = {idx: idx for idx in range(preds.shape[-1])}
+
         if isinstance(preds, paddle.Tensor):
             preds = preds.numpy()
+
         pred_idxs = preds.argmax(axis=1)
-        decode_out = [(self.label_list[idx], preds[i, idx])
+        decode_out = [(label_list[idx], preds[i, idx])
                       for i, idx in enumerate(pred_idxs)]
         if label is None:
             return decode_out
-        label = [(self.label_list[idx], 1.0) for idx in label]
+        label = [(label_list[idx], 1.0) for idx in label]
         return decode_out, label

tools/export_model.py

@@ -31,7 +31,7 @@ from ppocr.utils.logging import get_logger
 from tools.program import load_config, merge_config, ArgsParser
 
 
-def export_single_model(model, arch_config, save_path, logger):
+def export_single_model(model, arch_config, save_path, logger, quanter=None):
     if arch_config["algorithm"] == "SRN":
         max_text_length = arch_config["Head"]["max_text_length"]
         other_shape = [
@@ -95,7 +95,10 @@ def export_single_model(model, arch_config, save_path, logger):
                     shape=[None] + infer_shape, dtype="float32")
             ])
 
+    if quanter is None:
         paddle.jit.save(model, save_path)
+    else:
+        quanter.save_quantized_model(model, save_path)
     logger.info("inference model is saved to {}".format(save_path))
     return
 
@@ -125,7 +128,6 @@ def main():
                         char_num = char_num - 2
                     out_channels_list['CTCLabelDecode'] = char_num
                     out_channels_list['SARLabelDecode'] = char_num + 2
-                    loss_list = config['Loss']['loss_config_list']
                     config['Architecture']['Models'][key]['Head'][
                         'out_channels_list'] = out_channels_list
                 else:

tools/infer/predict_det.py

@@ -284,7 +284,7 @@ if __name__ == "__main__":
             total_time += elapse
         count += 1
         save_pred = os.path.basename(image_file) + "\t" + str(
-            json.dumps(np.array(dt_boxes).astype(np.int32).tolist())) + "\n"
+            json.dumps([x.tolist() for x in dt_boxes])) + "\n"
         save_results.append(save_pred)
         logger.info(save_pred)
         logger.info("The predict time of {}: {}".format(image_file, elapse))

tools/infer/utility.py

@@ -301,8 +301,8 @@ def create_predictor(args, mode, logger):
         # enable memory optim
         config.enable_memory_optim()
         config.disable_glog_info()
-
         config.delete_pass("conv_transpose_eltwiseadd_bn_fuse_pass")
+        config.delete_pass("matmul_transpose_reshape_fuse_pass")
         if mode == 'table':
             config.delete_pass("fc_fuse_pass")  # not supported for table
         config.switch_use_feed_fetch_ops(False)