Merge pull request #3721 from Topdu/dygraph

add rec_nrtr

configs/rec/rec_mtb_nrtr.yml

+Global:
+  use_gpu: True
+  epoch_num: 21
+  log_smooth_window: 20
+  print_batch_step: 10
+  save_model_dir: ./output/rec/nrtr/
+  save_epoch_step: 1
+  # evaluation is run every 2000 iterations
+  eval_batch_step: [0, 2000]
+  cal_metric_during_train: True
+  pretrained_model:
+  checkpoints: 
+  save_inference_dir:
+  use_visualdl: False
+  infer_img: doc/imgs_words_en/word_10.png
+  # for data or label process
+  character_dict_path: 
+  character_type: EN_symbol
+  max_text_length: 25
+  infer_mode: False
+  use_space_char: True
+  save_res_path: ./output/rec/predicts_nrtr.txt
+
+Optimizer:
+  name: Adam
+  beta1: 0.9
+  beta2: 0.99
+  clip_norm: 5.0
+  lr:
+    name: Cosine
+    learning_rate: 0.0005
+    warmup_epoch: 2
+  regularizer:
+    name: 'L2'
+    factor: 0.
+
+Architecture:
+  model_type: rec
+  algorithm: NRTR
+  in_channels: 1
+  Transform:
+  Backbone:
+    name: MTB
+    cnn_num: 2
+  Head:
+    name: Transformer
+    d_model: 512
+    num_encoder_layers: 6
+    beam_size: 10 # When Beam size is greater than 0, it means to use beam search when evaluation.
+    
+
+Loss:
+  name: NRTRLoss
+  smoothing: True
+
+PostProcess:
+  name: NRTRLabelDecode
+
+Metric:
+  name: RecMetric
+  main_indicator: acc
+
+Train:
+  dataset:
+    name: LMDBDataSet
+    data_dir: ./train_data/data_lmdb_release/training/
+    transforms:
+      - NRTRDecodeImage: # load image
+          img_mode: BGR
+          channel_first: False
+      - NRTRLabelEncode: # Class handling label
+      - NRTRRecResizeImg:
+          image_shape: [100, 32]
+          resize_type: PIL # PIL or OpenCV
+      - KeepKeys:
+          keep_keys: ['image', 'label', 'length'] # dataloader will return list in this order
+  loader:
+    shuffle: True
+    batch_size_per_card: 512
+    drop_last: True
+    num_workers: 8
+
+Eval:
+  dataset:
+    name: LMDBDataSet
+    data_dir: ./train_data/data_lmdb_release/evaluation/
+    transforms:
+      - NRTRDecodeImage: # load image
+          img_mode: BGR
+          channel_first: False
+      - NRTRLabelEncode: # Class handling label
+      - NRTRRecResizeImg:
+          image_shape: [100, 32]
+          resize_type: PIL # PIL or OpenCV
+      - KeepKeys:
+          keep_keys: ['image', 'label', 'length'] # dataloader will return list in this order
+  loader:
+    shuffle: False
+    drop_last: False
+    batch_size_per_card: 256
+    num_workers: 1
+    use_shared_memory: False

doc/doc_ch/algorithm_overview.md

@@ -44,6 +44,7 @@ PaddleOCR基于动态图开源的文本识别算法列表：
 - [x]  STAR-Net([paper](http://www.bmva.org/bmvc/2016/papers/paper043/index.html))[11]
 - [x]  RARE([paper](https://arxiv.org/abs/1603.03915v1))[12]
 - [x]  SRN([paper](https://arxiv.org/abs/2003.12294))[5]
+- [x]  NRTR([paper](https://arxiv.org/abs/1806.00926v2))
 
 参考[DTRB][3](https://arxiv.org/abs/1904.01906)文字识别训练和评估流程，使用MJSynth和SynthText两个文字识别数据集训练，在IIIT, SVT, IC03, IC13, IC15, SVTP, CUTE数据集上进行评估，算法效果如下：
 
@@ -58,6 +59,7 @@ PaddleOCR基于动态图开源的文本识别算法列表：
 |RARE|MobileNetV3|82.5%|rec_mv3_tps_bilstm_att |[下载链接](https://paddleocr.bj.bcebos.com/dygraph_v2.0/en/rec_mv3_tps_bilstm_att_v2.0_train.tar)|
 |RARE|Resnet34_vd|83.6%|rec_r34_vd_tps_bilstm_att |[下载链接](https://paddleocr.bj.bcebos.com/dygraph_v2.0/en/rec_r34_vd_tps_bilstm_att_v2.0_train.tar)|
 |SRN|Resnet50_vd_fpn| 88.52% | rec_r50fpn_vd_none_srn | [下载链接](https://paddleocr.bj.bcebos.com/dygraph_v2.0/en/rec_r50_vd_srn_train.tar) |
+|NRTR|NRTR_MTB| 84.3% | rec_mtb_nrtr | [下载链接](https://paddleocr.bj.bcebos.com/dygraph_v2.0/en/rec_mtb_nrtr_train.tar) |
 
 
 PaddleOCR文本识别算法的训练和使用请参考文档教程中[模型训练/评估中的文本识别部分](./recognition.md)。

doc/doc_ch/recognition.md

@@ -215,6 +215,7 @@ PaddleOCR支持训练和评估交替进行, 可以在 `configs/rec/rec_icdar15_t
 | rec_mv3_tps_bilstm_att.yml |  CRNN |   Mobilenet_v3 |  TPS   |  BiLSTM |  att  |
 | rec_r34_vd_tps_bilstm_att.yml |  CRNN |   Resnet34_vd |  TPS   |  BiLSTM |  att  |
 | rec_r50fpn_vd_none_srn.yml    | SRN | Resnet50_fpn_vd    | None    | rnn | srn |
+| rec_mtb_nrtr.yml    | NRTR | nrtr_mtb    | None    | transformer encoder | transformer decoder |
 
 训练中文数据，推荐使用[rec_chinese_lite_train_v2.0.yml](../../configs/rec/ch_ppocr_v2.0/rec_chinese_lite_train_v2.0.yml)，如您希望尝试其他算法在中文数据集上的效果，请参考下列说明修改配置文件：
 

doc/doc_en/algorithm_overview_en.md

@@ -46,6 +46,7 @@ PaddleOCR open-source text recognition algorithms list:
 - [x]  STAR-Net([paper](http://www.bmva.org/bmvc/2016/papers/paper043/index.html))[11]
 - [x]  RARE([paper](https://arxiv.org/abs/1603.03915v1))[12]
 - [x]  SRN([paper](https://arxiv.org/abs/2003.12294))[5]
+- [x]  NRTR([paper](https://arxiv.org/abs/1806.00926v2))
 
 Refer to [DTRB](https://arxiv.org/abs/1904.01906), the training and evaluation result of these above text recognition (using MJSynth and SynthText for training, evaluate on IIIT, SVT, IC03, IC13, IC15, SVTP, CUTE) is as follow:
 
@@ -60,5 +61,6 @@ Refer to [DTRB](https://arxiv.org/abs/1904.01906), the training and evaluation r
 |RARE|MobileNetV3|82.5%|rec_mv3_tps_bilstm_att |[Download link](https://paddleocr.bj.bcebos.com/dygraph_v2.0/en/rec_mv3_tps_bilstm_att_v2.0_train.tar)|
 |RARE|Resnet34_vd|83.6%|rec_r34_vd_tps_bilstm_att |[Download link](https://paddleocr.bj.bcebos.com/dygraph_v2.0/en/rec_r34_vd_tps_bilstm_att_v2.0_train.tar)|
 |SRN|Resnet50_vd_fpn| 88.52% | rec_r50fpn_vd_none_srn |[Download link](https://paddleocr.bj.bcebos.com/dygraph_v2.0/en/rec_r50_vd_srn_train.tar)|
+|NRTR|NRTR_MTB| 84.3% | rec_mtb_nrtr | [Download link](https://paddleocr.bj.bcebos.com/dygraph_v2.0/en/rec_mtb_nrtr_train.tar) |
 
 Please refer to the document for training guide and use of PaddleOCR text recognition algorithms [Text recognition model training/evaluation/prediction](./recognition_en.md)

doc/doc_en/recognition_en.md

@@ -207,7 +207,7 @@ If the evaluation set is large, the test will be time-consuming. It is recommend
 | rec_mv3_tps_bilstm_att.yml |  CRNN |   Mobilenet_v3 |  TPS   |  BiLSTM |  att  |
 | rec_r34_vd_tps_bilstm_att.yml |  CRNN |   Resnet34_vd |  TPS   |  BiLSTM |  att  |
 | rec_r50fpn_vd_none_srn.yml    | SRN | Resnet50_fpn_vd    | None    | rnn | srn |
-
+| rec_mtb_nrtr.yml    | NRTR | nrtr_mtb    | None    | transformer encoder | transformer decoder |
 
 For training Chinese data, it is recommended to use
 [rec_chinese_lite_train_v2.0.yml](../../configs/rec/ch_ppocr_v2.0/rec_chinese_lite_train_v2.0.yml). If you want to try the result of other algorithms on the Chinese data set, please refer to the following instructions to modify the configuration file:

ppocr/data/imaug/__init__.py

@@ -21,7 +21,7 @@ from .make_border_map import MakeBorderMap
 from .make_shrink_map import MakeShrinkMap
 from .random_crop_data import EastRandomCropData, PSERandomCrop
 
-from .rec_img_aug import RecAug, RecResizeImg, ClsResizeImg, SRNRecResizeImg
+from .rec_img_aug import RecAug, RecResizeImg, ClsResizeImg, SRNRecResizeImg, NRTRRecResizeImg
 from .randaugment import RandAugment
 from .copy_paste import CopyPaste
 from .operators import *

ppocr/data/imaug/label_ops.py

@@ -161,6 +161,34 @@ class BaseRecLabelEncode(object):
         return text_list
 
 
+class NRTRLabelEncode(BaseRecLabelEncode):
+    """ Convert between text-label and text-index """
+
+    def __init__(self,
+                 max_text_length,
+                 character_dict_path=None,
+                 character_type='EN_symbol',
+                 use_space_char=False,
+                 **kwargs):
+
+        super(NRTRLabelEncode,
+              self).__init__(max_text_length, character_dict_path,
+                             character_type, use_space_char)
+    def __call__(self, data):
+        text = data['label']
+        text = self.encode(text)
+        if text is None:
+            return None
+        data['length'] = np.array(len(text))
+        text.insert(0, 2)
+        text.append(3)
+        text = text + [0] * (self.max_text_len - len(text))
+        data['label'] = np.array(text)
+        return data
+    def add_special_char(self, dict_character):
+        dict_character = ['blank','<unk>','<s>','</s>'] + dict_character
+        return dict_character
+
 class CTCLabelEncode(BaseRecLabelEncode):
     """ Convert between text-label and text-index """
 

ppocr/data/imaug/operators.py

@@ -57,6 +57,38 @@ class DecodeImage(object):
         return data
 
 
+class NRTRDecodeImage(object):
+    """ decode image """
+
+    def __init__(self, img_mode='RGB', channel_first=False, **kwargs):
+        self.img_mode = img_mode
+        self.channel_first = channel_first
+
+    def __call__(self, data):
+        img = data['image']
+        if six.PY2:
+            assert type(img) is str and len(
+                img) > 0, "invalid input 'img' in DecodeImage"
+        else:
+            assert type(img) is bytes and len(
+                img) > 0, "invalid input 'img' in DecodeImage"
+        img = np.frombuffer(img, dtype='uint8')
+
+        img = cv2.imdecode(img, 1)
+
+        if img is None:
+            return None
+        if self.img_mode == 'GRAY':
+            img = cv2.cvtColor(img, cv2.COLOR_GRAY2BGR)
+        elif self.img_mode == 'RGB':
+            assert img.shape[2] == 3, 'invalid shape of image[%s]' % (img.shape)
+            img = img[:, :, ::-1]
+        img = cv2.cvtColor(img,cv2.COLOR_BGR2GRAY)
+        if self.channel_first:
+            img = img.transpose((2, 0, 1))
+        data['image'] = img
+        return data
+
 class NormalizeImage(object):
     """ normalize image such as substract mean, divide std
     """

ppocr/data/imaug/rec_img_aug.py

@@ -16,7 +16,7 @@ import math
 import cv2
 import numpy as np
 import random
-
+from PIL import Image
 from .text_image_aug import tia_perspective, tia_stretch, tia_distort
 
 
@@ -43,6 +43,25 @@ class ClsResizeImg(object):
         return data
 
 
+class NRTRRecResizeImg(object):
+    def __init__(self, image_shape, resize_type, **kwargs):
+        self.image_shape = image_shape
+        self.resize_type = resize_type
+
+    def __call__(self, data):
+        img = data['image']
+        if self.resize_type == 'PIL':
+            image_pil = Image.fromarray(np.uint8(img))
+            img = image_pil.resize(self.image_shape, Image.ANTIALIAS)
+            img = np.array(img)
+        if self.resize_type == 'OpenCV':
+            img = cv2.resize(img, self.image_shape)
+        norm_img = np.expand_dims(img, -1)
+        norm_img = norm_img.transpose((2, 0, 1))
+        data['image'] = norm_img.astype(np.float32) / 128. - 1.
+        return data
+
+
 class RecResizeImg(object):
     def __init__(self,
                  image_shape,

ppocr/losses/__init__.py

@@ -25,7 +25,7 @@ from .det_sast_loss import SASTLoss
 from .rec_ctc_loss import CTCLoss
 from .rec_att_loss import AttentionLoss
 from .rec_srn_loss import SRNLoss
-
+from .rec_nrtr_loss import NRTRLoss
 # cls loss
 from .cls_loss import ClsLoss
 
@@ -44,8 +44,9 @@ from .table_att_loss import TableAttentionLoss
 def build_loss(config):
     support_dict = [
         'DBLoss', 'EASTLoss', 'SASTLoss', 'CTCLoss', 'ClsLoss', 'AttentionLoss',
-        'SRNLoss', 'PGLoss', 'CombinedLoss', 'TableAttentionLoss'
+        'SRNLoss', 'PGLoss', 'CombinedLoss', 'NRTRLoss', 'TableAttentionLoss'
     ]
+
     config = copy.deepcopy(config)
     module_name = config.pop('name')
     assert module_name in support_dict, Exception('loss only support {}'.format(

ppocr/losses/rec_nrtr_loss.py

+import paddle
+from paddle import nn
+import paddle.nn.functional as F
+
+
+class NRTRLoss(nn.Layer):
+    def __init__(self, smoothing=True, **kwargs):
+        super(NRTRLoss, self).__init__()
+        self.loss_func = nn.CrossEntropyLoss(reduction='mean', ignore_index=0)
+        self.smoothing = smoothing
+
+    def forward(self, pred, batch):
+        pred = pred.reshape([-1, pred.shape[2]])
+        max_len = batch[2].max()
+        tgt = batch[1][:, 1:2 + max_len]
+        tgt = tgt.reshape([-1])
+        if self.smoothing:
+            eps = 0.1
+            n_class = pred.shape[1]
+            one_hot = F.one_hot(tgt, pred.shape[1])
+            one_hot = one_hot * (1 - eps) + (1 - one_hot) * eps / (n_class - 1)
+            log_prb = F.log_softmax(pred, axis=1)
+            non_pad_mask = paddle.not_equal(
+                tgt, paddle.zeros(
+                    tgt.shape, dtype='int64'))
+            loss = -(one_hot * log_prb).sum(axis=1)
+            loss = loss.masked_select(non_pad_mask).mean()
+        else:
+            loss = self.loss_func(pred, tgt)
+        return {'loss': loss}

ppocr/metrics/rec_metric.py

@@ -57,3 +57,4 @@ class RecMetric(object):
         self.correct_num = 0
         self.all_num = 0
         self.norm_edit_dis = 0
+        

ppocr/modeling/architectures/base_model.py

@@ -14,7 +14,6 @@
 from __future__ import absolute_import
 from __future__ import division
 from __future__ import print_function
-
 from paddle import nn
 from ppocr.modeling.transforms import build_transform
 from ppocr.modeling.backbones import build_backbone

ppocr/modeling/backbones/__init__.py

@@ -26,8 +26,9 @@ def build_backbone(config, model_type):
         from .rec_resnet_vd import ResNet
         from .rec_resnet_fpn import ResNetFPN
         from .rec_mv1_enhance import MobileNetV1Enhance
+        from .rec_nrtr_mtb import MTB
         support_dict = [
-            "MobileNetV1Enhance", "MobileNetV3", "ResNet", "ResNetFPN"
+            'MobileNetV1Enhance', 'MobileNetV3', 'ResNet', 'ResNetFPN', 'MTB'
         ]
     elif model_type == "e2e":
         from .e2e_resnet_vd_pg import ResNet

ppocr/modeling/backbones/rec_nrtr_mtb.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 paddle import nn
+
+
+class MTB(nn.Layer):
+    def __init__(self, cnn_num, in_channels):
+        super(MTB, self).__init__()
+        self.block = nn.Sequential()
+        self.out_channels = in_channels
+        self.cnn_num = cnn_num
+        if self.cnn_num == 2:
+            for i in range(self.cnn_num):
+                self.block.add_sublayer(
+                    'conv_{}'.format(i),
+                    nn.Conv2D(
+                        in_channels=in_channels
+                        if i == 0 else 32 * (2**(i - 1)),
+                        out_channels=32 * (2**i),
+                        kernel_size=3,
+                        stride=2,
+                        padding=1))
+                self.block.add_sublayer('relu_{}'.format(i), nn.ReLU())
+                self.block.add_sublayer('bn_{}'.format(i),
+                                        nn.BatchNorm2D(32 * (2**i)))
+
+    def forward(self, images):
+        x = self.block(images)
+        if self.cnn_num == 2:
+            # (b, w, h, c)
+            x = x.transpose([0, 3, 2, 1])
+            x_shape = x.shape
+            x = x.reshape([x_shape[0], x_shape[1], x_shape[2] * x_shape[3]])
+        return x

ppocr/modeling/heads/__init__.py

@@ -26,12 +26,14 @@ def build_head(config):
     from .rec_ctc_head import CTCHead
     from .rec_att_head import AttentionHead
     from .rec_srn_head import SRNHead
+    from .rec_nrtr_head import Transformer
 
     # cls head
     from .cls_head import ClsHead
     support_dict = [
         'DBHead', 'EASTHead', 'SASTHead', 'CTCHead', 'ClsHead', 'AttentionHead',
-        'SRNHead', 'PGHead', 'TableAttentionHead']
+        'SRNHead', 'PGHead', 'Transformer', 'TableAttentionHead'
+    ]
 
     #table head
     from .table_att_head import TableAttentionHead

ppocr/modeling/heads/multiheadAttention.py

+# copyright (c) 2021 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 paddle
+from paddle import nn
+import paddle.nn.functional as F
+from paddle.nn import Linear
+from paddle.nn.initializer import XavierUniform as xavier_uniform_
+from paddle.nn.initializer import Constant as constant_
+from paddle.nn.initializer import XavierNormal as xavier_normal_
+
+zeros_ = constant_(value=0.)
+ones_ = constant_(value=1.)
+
+
+class MultiheadAttention(nn.Layer):
+    """Allows the model to jointly attend to information
+    from different representation subspaces.
+    See reference: Attention Is All You Need
+
+    .. math::
+        \text{MultiHead}(Q, K, V) = \text{Concat}(head_1,\dots,head_h)W^O
+        \text{where} head_i = \text{Attention}(QW_i^Q, KW_i^K, VW_i^V)
+
+    Args:
+        embed_dim: total dimension of the model
+        num_heads: parallel attention layers, or heads
+
+    """
+
+    def __init__(self,
+                 embed_dim,
+                 num_heads,
+                 dropout=0.,
+                 bias=True,
+                 add_bias_kv=False,
+                 add_zero_attn=False):
+        super(MultiheadAttention, self).__init__()
+        self.embed_dim = embed_dim
+        self.num_heads = num_heads
+        self.dropout = dropout
+        self.head_dim = embed_dim // num_heads
+        assert self.head_dim * num_heads == self.embed_dim, "embed_dim must be divisible by num_heads"
+        self.scaling = self.head_dim**-0.5
+        self.out_proj = Linear(embed_dim, embed_dim, bias_attr=bias)
+        self._reset_parameters()
+        self.conv1 = paddle.nn.Conv2D(
+            in_channels=embed_dim, out_channels=embed_dim, kernel_size=(1, 1))
+        self.conv2 = paddle.nn.Conv2D(
+            in_channels=embed_dim, out_channels=embed_dim, kernel_size=(1, 1))
+        self.conv3 = paddle.nn.Conv2D(
+            in_channels=embed_dim, out_channels=embed_dim, kernel_size=(1, 1))
+
+    def _reset_parameters(self):
+        xavier_uniform_(self.out_proj.weight)
+
+    def forward(self,
+                query,
+                key,
+                value,
+                key_padding_mask=None,
+                incremental_state=None,
+                need_weights=True,
+                static_kv=False,
+                attn_mask=None):
+        """
+        Inputs of forward function
+            query: [target length, batch size, embed dim]
+            key: [sequence length, batch size, embed dim]
+            value: [sequence length, batch size, embed dim]
+            key_padding_mask: if True, mask padding based on batch size
+            incremental_state: if provided, previous time steps are cashed
+            need_weights: output attn_output_weights
+            static_kv: key and value are static
+
+        Outputs of forward function
+            attn_output: [target length, batch size, embed dim]
+            attn_output_weights: [batch size, target length, sequence length]
+        """
+        tgt_len, bsz, embed_dim = query.shape
+        assert embed_dim == self.embed_dim
+        assert list(query.shape) == [tgt_len, bsz, embed_dim]
+        assert key.shape == value.shape
+
+        q = self._in_proj_q(query)
+        k = self._in_proj_k(key)
+        v = self._in_proj_v(value)
+        q *= self.scaling
+
+        q = q.reshape([tgt_len, bsz * self.num_heads, self.head_dim]).transpose(
+            [1, 0, 2])
+        k = k.reshape([-1, bsz * self.num_heads, self.head_dim]).transpose(
+            [1, 0, 2])
+        v = v.reshape([-1, bsz * self.num_heads, self.head_dim]).transpose(
+            [1, 0, 2])
+
+        src_len = k.shape[1]
+
+        if key_padding_mask is not None:
+            assert key_padding_mask.shape[0] == bsz
+            assert key_padding_mask.shape[1] == src_len
+
+        attn_output_weights = paddle.bmm(q, k.transpose([0, 2, 1]))
+        assert list(attn_output_weights.
+                    shape) == [bsz * self.num_heads, tgt_len, src_len]
+
+        if attn_mask is not None:
+            attn_mask = attn_mask.unsqueeze(0)
+            attn_output_weights += attn_mask
+        if key_padding_mask is not None:
+            attn_output_weights = attn_output_weights.reshape(
+                [bsz, self.num_heads, tgt_len, src_len])
+            key = key_padding_mask.unsqueeze(1).unsqueeze(2).astype('float32')
+            y = paddle.full(shape=key.shape, dtype='float32', fill_value='-inf')
+            y = paddle.where(key == 0., key, y)
+            attn_output_weights += y
+            attn_output_weights = attn_output_weights.reshape(
+                [bsz * self.num_heads, tgt_len, src_len])
+
+        attn_output_weights = F.softmax(
+            attn_output_weights.astype('float32'),
+            axis=-1,
+            dtype=paddle.float32 if attn_output_weights.dtype == paddle.float16
+            else attn_output_weights.dtype)
+        attn_output_weights = F.dropout(
+            attn_output_weights, p=self.dropout, training=self.training)
+
+        attn_output = paddle.bmm(attn_output_weights, v)
+        assert list(attn_output.
+                    shape) == [bsz * self.num_heads, tgt_len, self.head_dim]
+        attn_output = attn_output.transpose([1, 0, 2]).reshape(
+            [tgt_len, bsz, embed_dim])
+        attn_output = self.out_proj(attn_output)
+
+        if need_weights:
+            # average attention weights over heads
+            attn_output_weights = attn_output_weights.reshape(
+                [bsz, self.num_heads, tgt_len, src_len])
+            attn_output_weights = attn_output_weights.sum(
+                axis=1) / self.num_heads
+        else:
+            attn_output_weights = None
+        return attn_output, attn_output_weights
+
+    def _in_proj_q(self, query):
+        query = query.transpose([1, 2, 0])
+        query = paddle.unsqueeze(query, axis=2)
+        res = self.conv1(query)
+        res = paddle.squeeze(res, axis=2)
+        res = res.transpose([2, 0, 1])
+        return res
+
+    def _in_proj_k(self, key):
+        key = key.transpose([1, 2, 0])
+        key = paddle.unsqueeze(key, axis=2)
+        res = self.conv2(key)
+        res = paddle.squeeze(res, axis=2)
+        res = res.transpose([2, 0, 1])
+        return res
+
+    def _in_proj_v(self, value):
+        value = value.transpose([1, 2, 0])  #(1, 2, 0)
+        value = paddle.unsqueeze(value, axis=2)
+        res = self.conv3(value)
+        res = paddle.squeeze(res, axis=2)
+        res = res.transpose([2, 0, 1])
+        return res

ppocr/postprocess/__init__.py

@@ -24,18 +24,16 @@ __all__ = ['build_post_process']
 from .db_postprocess import DBPostProcess, DistillationDBPostProcess
 from .east_postprocess import EASTPostProcess
 from .sast_postprocess import SASTPostProcess
-from .rec_postprocess import CTCLabelDecode, AttnLabelDecode, SRNLabelDecode, DistillationCTCLabelDecode, \
+from .rec_postprocess import CTCLabelDecode, AttnLabelDecode, SRNLabelDecode, DistillationCTCLabelDecode, NRTRLabelDecode, \
     TableLabelDecode
 from .cls_postprocess import ClsPostProcess
 from .pg_postprocess import PGPostProcess
 
-
 def build_post_process(config, global_config=None):
     support_dict = [
         'DBPostProcess', 'EASTPostProcess', 'SASTPostProcess', 'CTCLabelDecode',
         'AttnLabelDecode', 'ClsPostProcess', 'SRNLabelDecode', 'PGPostProcess',
-        'DistillationCTCLabelDecode', 'TableLabelDecode',
-        'DistillationDBPostProcess'
+        'DistillationCTCLabelDecode', 'NRTRLabelDecode', 'TableLabelDecode', 'DistillationDBPostProcess'
     ]
 
     config = copy.deepcopy(config)

ppocr/postprocess/rec_postprocess.py

@@ -156,6 +156,69 @@ class DistillationCTCLabelDecode(CTCLabelDecode):
         return output
 
 
+class NRTRLabelDecode(BaseRecLabelDecode):
+    """ Convert between text-label and text-index """
+
+    def __init__(self,
+                 character_dict_path=None,
+                 character_type='EN_symbol',
+                 use_space_char=True,
+                 **kwargs):
+        super(NRTRLabelDecode, self).__init__(character_dict_path,
+                                             character_type, use_space_char)
+
+    def __call__(self, preds, label=None, *args, **kwargs):
+        if preds.dtype == paddle.int64:
+            if isinstance(preds, paddle.Tensor):
+                preds = preds.numpy()
+            if preds[0][0]==2:
+                preds_idx = preds[:,1:]
+            else:
+                preds_idx = preds
+
+            text = self.decode(preds_idx)
+            if label is None:
+                return text
+            label = self.decode(label[:,1:])
+        else:
+            if isinstance(preds, paddle.Tensor):
+                preds = preds.numpy()
+            preds_idx = preds.argmax(axis=2)
+            preds_prob = preds.max(axis=2)
+            text = self.decode(preds_idx, preds_prob, is_remove_duplicate=False)
+            if label is None:
+                return text
+            label = self.decode(label[:,1:])
+        return text, label
+
+    def add_special_char(self, dict_character):
+        dict_character = ['blank','<unk>','<s>','</s>'] + dict_character
+        return dict_character
+    
+    def decode(self, text_index, text_prob=None, is_remove_duplicate=False):
+        """ convert text-index into text-label. """
+        result_list = []
+        batch_size = len(text_index)
+        for batch_idx in range(batch_size):
+            char_list = []
+            conf_list = []
+            for idx in range(len(text_index[batch_idx])):
+                if text_index[batch_idx][idx] == 3: # end
+                    break
+                try:
+                    char_list.append(self.character[int(text_index[batch_idx][idx])])
+                except:
+                    continue
+                if text_prob is not None:
+                    conf_list.append(text_prob[batch_idx][idx])
+                else:
+                    conf_list.append(1)
+            text = ''.join(char_list)
+            result_list.append((text.lower(), np.mean(conf_list)))
+        return result_list
+
+
+
 class AttnLabelDecode(BaseRecLabelDecode):
     """ Convert between text-label and text-index """
 
@@ -193,8 +256,7 @@ class AttnLabelDecode(BaseRecLabelDecode):
                     if idx > 0 and text_index[batch_idx][idx - 1] == text_index[
                             batch_idx][idx]:
                         continue
-                char_list.append(self.character[int(text_index[batch_idx][
-                    idx])])
+                char_list.append(self.character[int(text_index[batch_idx][idx])])
                 if text_prob is not None:
                     conf_list.append(text_prob[batch_idx][idx])
                 else: