fix conflicts

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/losses/rec_sar_loss.py

+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import paddle
+from paddle import nn
+
+
+class SARLoss(nn.Layer):
+    def __init__(self, **kwargs):
+        super(SARLoss, self).__init__()
+        self.loss_func = paddle.nn.loss.CrossEntropyLoss(reduction="mean", ignore_index=96)
+
+    def forward(self, predicts, batch):
+        predict = predicts[:, :-1, :] # ignore last index of outputs to be in same seq_len with targets
+        label = batch[1].astype("int64")[:, 1:] # ignore first index of target in loss calculation
+        batch_size, num_steps, num_classes = predict.shape[0], predict.shape[
+            1], predict.shape[2]
+        assert len(label.shape) == len(list(predict.shape)) - 1, \
+            "The target's shape and inputs's shape is [N, d] and [N, num_steps]"
+
+        inputs = paddle.reshape(predict, [-1, num_classes])
+        targets = paddle.reshape(label, [-1])
+        loss = self.loss_func(inputs, targets)
+        return {'loss': loss}

ppocr/losses/table_att_loss.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.
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import paddle
+from paddle import nn
+from paddle.nn import functional as F
+from paddle import fluid
+
+class TableAttentionLoss(nn.Layer):
+    def __init__(self, structure_weight, loc_weight, use_giou=False, giou_weight=1.0, **kwargs):
+        super(TableAttentionLoss, self).__init__()
+        self.loss_func = nn.CrossEntropyLoss(weight=None, reduction='none')
+        self.structure_weight = structure_weight
+        self.loc_weight = loc_weight
+        self.use_giou = use_giou
+        self.giou_weight = giou_weight
+        
+    def giou_loss(self, preds, bbox, eps=1e-7, reduction='mean'):
+        '''
+        :param preds:[[x1,y1,x2,y2], [x1,y1,x2,y2],,,]
+        :param bbox:[[x1,y1,x2,y2], [x1,y1,x2,y2],,,]
+        :return: loss
+        '''
+        ix1 = fluid.layers.elementwise_max(preds[:, 0], bbox[:, 0])
+        iy1 = fluid.layers.elementwise_max(preds[:, 1], bbox[:, 1])
+        ix2 = fluid.layers.elementwise_min(preds[:, 2], bbox[:, 2])
+        iy2 = fluid.layers.elementwise_min(preds[:, 3], bbox[:, 3])
+
+        iw = fluid.layers.clip(ix2 - ix1 + 1e-3, 0., 1e10)
+        ih = fluid.layers.clip(iy2 - iy1 + 1e-3, 0., 1e10)
+
+        # overlap
+        inters = iw * ih
+
+        # union
+        uni = (preds[:, 2] - preds[:, 0] + 1e-3) * (preds[:, 3] - preds[:, 1] + 1e-3
+            ) + (bbox[:, 2] - bbox[:, 0] + 1e-3) * (
+            bbox[:, 3] - bbox[:, 1] + 1e-3) - inters + eps
+
+        # ious
+        ious = inters / uni
+
+        ex1 = fluid.layers.elementwise_min(preds[:, 0], bbox[:, 0])
+        ey1 = fluid.layers.elementwise_min(preds[:, 1], bbox[:, 1])
+        ex2 = fluid.layers.elementwise_max(preds[:, 2], bbox[:, 2])
+        ey2 = fluid.layers.elementwise_max(preds[:, 3], bbox[:, 3])
+        ew = fluid.layers.clip(ex2 - ex1 + 1e-3, 0., 1e10)
+        eh = fluid.layers.clip(ey2 - ey1 + 1e-3, 0., 1e10)
+
+        # enclose erea
+        enclose = ew * eh + eps
+        giou = ious - (enclose - uni) / enclose
+
+        loss = 1 - giou
+
+        if reduction == 'mean':
+            loss = paddle.mean(loss)
+        elif reduction == 'sum':
+            loss = paddle.sum(loss)
+        else:
+            raise NotImplementedError
+        return loss
+
+    def forward(self, predicts, batch):
+        structure_probs = predicts['structure_probs']
+        structure_targets = batch[1].astype("int64")
+        structure_targets = structure_targets[:, 1:]
+        if len(batch) == 6:
+            structure_mask = batch[5].astype("int64")
+            structure_mask = structure_mask[:, 1:]
+            structure_mask = paddle.reshape(structure_mask, [-1])
+        structure_probs = paddle.reshape(structure_probs, [-1, structure_probs.shape[-1]])
+        structure_targets = paddle.reshape(structure_targets, [-1])
+        structure_loss = self.loss_func(structure_probs, structure_targets)
+        
+        if len(batch) == 6:
+             structure_loss = structure_loss * structure_mask
+            
+#         structure_loss = paddle.sum(structure_loss) * self.structure_weight
+        structure_loss = paddle.mean(structure_loss) * self.structure_weight
+        
+        loc_preds = predicts['loc_preds']
+        loc_targets = batch[2].astype("float32")
+        loc_targets_mask = batch[4].astype("float32")
+        loc_targets = loc_targets[:, 1:, :]
+        loc_targets_mask = loc_targets_mask[:, 1:, :]
+        loc_loss = F.mse_loss(loc_preds * loc_targets_mask, loc_targets) * self.loc_weight
+        if self.use_giou:
+            loc_loss_giou = self.giou_loss(loc_preds * loc_targets_mask, loc_targets) * self.giou_weight
+            total_loss = structure_loss + loc_loss + loc_loss_giou
+            return {'loss':total_loss, "structure_loss":structure_loss, "loc_loss":loc_loss, "loc_loss_giou":loc_loss_giou}
+        else:
+            total_loss = structure_loss + loc_loss            
+            return {'loss':total_loss, "structure_loss":structure_loss, "loc_loss":loc_loss}
\ No newline at end of file

ppocr/metrics/__init__.py

@@ -19,23 +19,26 @@ from __future__ import unicode_literals
 
 import copy
 
-__all__ = ['build_metric']
+__all__ = ["build_metric"]
 
+from .det_metric import DetMetric
+from .rec_metric import RecMetric
+from .cls_metric import ClsMetric
+from .e2e_metric import E2EMetric
+from .distillation_metric import DistillationMetric
+from .table_metric import TableMetric
+from .kie_metric import KIEMetric
 
-def build_metric(config):
-    from .det_metric import DetMetric
-    from .rec_metric import RecMetric
-    from .cls_metric import ClsMetric
-    from .e2e_metric import E2EMetric
-    from .kie_metric import KIEMetric
 
+def build_metric(config):
     support_dict = [
-        'DetMetric', 'RecMetric', 'ClsMetric', 'E2EMetric', 'KIEMetric'
+        "DetMetric", "RecMetric", "ClsMetric", "E2EMetric",
+        "DistillationMetric", "TableMetric", 'KIEMetric'
     ]
 
     config = copy.deepcopy(config)
-    module_name = config.pop('name')
+    module_name = config.pop("name")
     assert module_name in support_dict, Exception(
-        'metric only support {}'.format(support_dict))
+        "metric only support {}".format(support_dict))
     module_class = eval(module_name)(**config)
     return module_class

ppocr/metrics/det_metric.py

@@ -55,6 +55,7 @@ class DetMetric(object):
             result = self.evaluator.evaluate_image(gt_info_list, det_info_list)
             self.results.append(result)
 
+
     def get_metric(self):
         """
         return metrics {

ppocr/metrics/distillation_metric.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 importlib
+import copy
+
+from .rec_metric import RecMetric
+from .det_metric import DetMetric
+from .e2e_metric import E2EMetric
+from .cls_metric import ClsMetric
+
+
+class DistillationMetric(object):
+    def __init__(self,
+                 key=None,
+                 base_metric_name=None,
+                 main_indicator=None,
+                 **kwargs):
+        self.main_indicator = main_indicator
+        self.key = key
+        self.main_indicator = main_indicator
+        self.base_metric_name = base_metric_name
+        self.kwargs = kwargs
+        self.metrics = None
+
+    def _init_metrcis(self, preds):
+        self.metrics = dict()
+        mod = importlib.import_module(__name__)
+        for key in preds:
+            self.metrics[key] = getattr(mod, self.base_metric_name)(
+                main_indicator=self.main_indicator, **self.kwargs)
+            self.metrics[key].reset()
+
+    def __call__(self, preds, batch, **kwargs):
+        assert isinstance(preds, dict)
+        if self.metrics is None:
+            self._init_metrcis(preds)
+        output = dict()
+        for key in preds:
+            self.metrics[key].__call__(preds[key], batch, **kwargs)
+
+    def get_metric(self):
+        """
+        return metrics {
+                 'acc': 0,
+                 'norm_edit_dis': 0,
+            }
+        """
+        output = dict()
+        for key in self.metrics:
+            metric = self.metrics[key].get_metric()
+            # main indicator
+            if key == self.key:
+                output.update(metric)
+            else:
+                for sub_key in metric:
+                    output["{}_{}".format(key, sub_key)] = metric[sub_key]
+        return output
+
+    def reset(self):
+        for key in self.metrics:
+            self.metrics[key].reset()

ppocr/metrics/e2e_metric.py

@@ -18,16 +18,18 @@ from __future__ import print_function
 
 __all__ = ['E2EMetric']
 
-from ppocr.utils.e2e_metric.Deteval import get_socre, combine_results
+from ppocr.utils.e2e_metric.Deteval import get_socre_A, get_socre_B, combine_results
 from ppocr.utils.e2e_utils.extract_textpoint_slow import get_dict
 
 
 class E2EMetric(object):
     def __init__(self,
+                 mode,
                  gt_mat_dir,
                  character_dict_path,
                  main_indicator='f_score_e2e',
                  **kwargs):
+        self.mode = mode
         self.gt_mat_dir = gt_mat_dir
         self.label_list = get_dict(character_dict_path)
         self.max_index = len(self.label_list)
@@ -35,13 +37,45 @@ class E2EMetric(object):
         self.reset()
 
     def __call__(self, preds, batch, **kwargs):
-        img_id = batch[5][0]
-        e2e_info_list = [{
-            'points': det_polyon,
-            'text': pred_str
-        } for det_polyon, pred_str in zip(preds['points'], preds['strs'])]
-        result = get_socre(self.gt_mat_dir, img_id, e2e_info_list)
-        self.results.append(result)
+        if self.mode == 'A':
+            gt_polyons_batch = batch[2]
+            temp_gt_strs_batch = batch[3][0]
+            ignore_tags_batch = batch[4]
+            gt_strs_batch = []
+
+            for temp_list in temp_gt_strs_batch:
+                t = ""
+                for index in temp_list:
+                    if index < self.max_index:
+                        t += self.label_list[index]
+                gt_strs_batch.append(t)
+
+            for pred, gt_polyons, gt_strs, ignore_tags in zip(
+                [preds], gt_polyons_batch, [gt_strs_batch], ignore_tags_batch):
+                # prepare gt
+                gt_info_list = [{
+                    'points': gt_polyon,
+                    'text': gt_str,
+                    'ignore': ignore_tag
+                } for gt_polyon, gt_str, ignore_tag in
+                                zip(gt_polyons, gt_strs, ignore_tags)]
+                # prepare det
+                e2e_info_list = [{
+                    'points': det_polyon,
+                    'texts': pred_str
+                } for det_polyon, pred_str in
+                                 zip(pred['points'], pred['texts'])]
+
+                result = get_socre_A(gt_info_list, e2e_info_list)
+                self.results.append(result)
+        else:
+            img_id = batch[5][0]
+            e2e_info_list = [{
+                'points': det_polyon,
+                'texts': pred_str
+            } for det_polyon, pred_str in zip(preds['points'], preds['texts'])]
+            result = get_socre_B(self.gt_mat_dir, img_id, e2e_info_list)
+            self.results.append(result)
 
     def get_metric(self):
         metircs = combine_results(self.results)

ppocr/metrics/eval_det_iou.py

@@ -169,21 +169,10 @@ class DetectionIoUEvaluator(object):
         numGlobalCareDet += numDetCare
 
         perSampleMetrics = {
-            'precision': precision,
-            'recall': recall,
-            'hmean': hmean,
-            'pairs': pairs,
-            'iouMat': [] if len(detPols) > 100 else iouMat.tolist(),
-            'gtPolPoints': gtPolPoints,
-            'detPolPoints': detPolPoints,
             'gtCare': numGtCare,
             'detCare': numDetCare,
-            'gtDontCare': gtDontCarePolsNum,
-            'detDontCare': detDontCarePolsNum,
             'detMatched': detMatched,
-            'evaluationLog': evaluationLog
         }
-
         return perSampleMetrics
 
     def combine_results(self, results):

ppocr/metrics/rec_metric.py

@@ -13,13 +13,20 @@
 # limitations under the License.
 
 import Levenshtein
+import string
 
 
 class RecMetric(object):
-    def __init__(self, main_indicator='acc', **kwargs):
+    def __init__(self, main_indicator='acc', is_filter=False, **kwargs):
         self.main_indicator = main_indicator
+        self.is_filter = is_filter
         self.reset()
 
+    def _normalize_text(self, text):
+        text = ''.join(
+            filter(lambda x: x in (string.digits + string.ascii_letters), text))
+        return text.lower()
+
     def __call__(self, pred_label, *args, **kwargs):
         preds, labels = pred_label
         correct_num = 0
@@ -28,6 +35,9 @@ class RecMetric(object):
         for (pred, pred_conf), (target, _) in zip(preds, labels):
             pred = pred.replace(" ", "")
             target = target.replace(" ", "")
+            if self.is_filter:
+                pred = self._normalize_text(pred)
+                target = self._normalize_text(target)
             norm_edit_dis += Levenshtein.distance(pred, target) / max(
                 len(pred), len(target), 1)
             if pred == target:

ppocr/metrics/table_metric.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 numpy as np
+class TableMetric(object):
+    def __init__(self, main_indicator='acc', **kwargs):
+        self.main_indicator = main_indicator
+        self.reset()
+
+    def __call__(self, pred, batch, *args, **kwargs):
+        structure_probs = pred['structure_probs'].numpy()
+        structure_labels = batch[1]
+        correct_num = 0
+        all_num = 0
+        structure_probs = np.argmax(structure_probs, axis=2)
+        structure_labels = structure_labels[:, 1:]
+        batch_size = structure_probs.shape[0]
+        for bno in range(batch_size):
+            all_num += 1
+            if (structure_probs[bno] == structure_labels[bno]).all():
+                correct_num += 1
+        self.correct_num += correct_num
+        self.all_num += all_num
+        return {
+            'acc': correct_num * 1.0 / all_num,
+        }
+
+    def get_metric(self):
+        """
+        return metrics {
+                 'acc': 0,
+            }
+        """
+        acc = 1.0 * self.correct_num / self.all_num
+        self.reset()
+        return {'acc': acc}
+
+    def reset(self):
+        self.correct_num = 0
+        self.all_num = 0

ppocr/modeling/architectures/__init__.py

@@ -13,12 +13,20 @@
 # limitations under the License.
 
 import copy
+import importlib
+
+from .base_model import BaseModel
+from .distillation_model import DistillationModel
 
 __all__ = ['build_model']
 
+
 def build_model(config):
-    from .base_model import BaseModel
-    
     config = copy.deepcopy(config)
-    module_class = BaseModel(config)
-    return module_class
+    if not "name" in config:
+        arch = BaseModel(config)
\ No newline at end of file
+    else:
+        name = config.pop("name")
+        mod = importlib.import_module(__name__)
+        arch = getattr(mod, name)(config)
+    return arch

ppocr/modeling/architectures/base_model.py

-# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -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
@@ -32,7 +31,6 @@ class BaseModel(nn.Layer):
             config (dict): the super parameters for module.
         """
         super(BaseModel, self).__init__()
-
         in_channels = config.get('in_channels', 3)
         model_type = config['model_type']
         # build transfrom,
@@ -68,14 +66,23 @@ class BaseModel(nn.Layer):
         config["Head"]['in_channels'] = in_channels
         self.head = build_head(config["Head"])
 
+        self.return_all_feats = config.get("return_all_feats", False)
+
     def forward(self, x, data=None):
+        y = dict()
         if self.use_transform:
             x = self.transform(x)
         x = self.backbone(x)
+        y["backbone_out"] = x
         if self.use_neck:
             x = self.neck(x)
-        if data is None:
-            x = self.head(x)
+        y["neck_out"] = x
+        x = self.head(x, targets=data)
+        if isinstance(x, dict):
+            y.update(x)
+        else:
+            y["head_out"] = x
+        if self.return_all_feats:
+            return y
         else:
-            x = self.head(x, data)
-        return x
+            return x

ppocr/modeling/architectures/distillation_model.py

+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+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
+from ppocr.modeling.necks import build_neck
+from ppocr.modeling.heads import build_head
+from .base_model import BaseModel
+from ppocr.utils.save_load import init_model, load_pretrained_params
+
+__all__ = ['DistillationModel']
+
+
+class DistillationModel(nn.Layer):
+    def __init__(self, config):
+        """
+        the module for OCR distillation.
+        args:
+            config (dict): the super parameters for module.
+        """
+        super().__init__()
+        self.model_list = []
+        self.model_name_list = []
+        for key in config["Models"]:
+            model_config = config["Models"][key]
+            freeze_params = False
+            pretrained = None
+            if "freeze_params" in model_config:
+                freeze_params = model_config.pop("freeze_params")
+            if "pretrained" in model_config:
+                pretrained = model_config.pop("pretrained")
+            model = BaseModel(model_config)
+            if pretrained is not None:
+                load_pretrained_params(model, pretrained)
+            if freeze_params:
+                for param in model.parameters():
+                    param.trainable = False
+            self.model_list.append(self.add_sublayer(key, model))
+            self.model_name_list.append(key)
+
+    def forward(self, x):
+        result_dict = dict()
+        for idx, model_name in enumerate(self.model_name_list):
+            result_dict[model_name] = self.model_list[idx](x)
+        return result_dict

ppocr/modeling/backbones/__init__.py

@@ -12,33 +12,43 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-__all__ = ['build_backbone']
+__all__ = ["build_backbone"]
 
 
 def build_backbone(config, model_type):
-    if model_type == 'det':
+    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':
+        support_dict = ["MobileNetV3", "ResNet", "ResNet_SAST"]
+    elif model_type == "rec" or model_type == "cls":
         from .rec_mobilenet_v3 import MobileNetV3
         from .rec_resnet_vd import ResNet
         from .rec_resnet_fpn import ResNetFPN
-        support_dict = ['MobileNetV3', 'ResNet', 'ResNetFPN']
-    elif model_type == 'e2e':
+        from .rec_mv1_enhance import MobileNetV1Enhance
+        from .rec_nrtr_mtb import MTB
+        from .rec_resnet_31 import ResNet31
+        from .rec_resnet_aster import ResNet_ASTER
+        support_dict = [
+            'MobileNetV1Enhance', 'MobileNetV3', 'ResNet', 'ResNetFPN', 'MTB',
+            "ResNet31", "ResNet_ASTER"
+        ]
+    elif model_type == "e2e":
         from .e2e_resnet_vd_pg import ResNet
         support_dict = ['ResNet']
     elif model_type == 'kie':
         from .kie_unet_sdmgr import Kie_backbone
         support_dict = ['Kie_backbone']
-
+    elif model_type == "table":
+        from .table_resnet_vd import ResNet
+        from .table_mobilenet_v3 import MobileNetV3
+        support_dict = ["ResNet", "MobileNetV3"]
     else:
         raise NotImplementedError
 
-    module_name = config.pop('name')
+    module_name = config.pop("name")
     assert module_name in support_dict, Exception(
-        'when model typs is {}, backbone only support {}'.format(model_type,
+        "when model typs is {}, backbone only support {}".format(model_type,
                                                                  support_dict))
     module_class = eval(module_name)(**config)
     return module_class

ppocr/modeling/backbones/det_mobilenet_v3.py

@@ -102,8 +102,7 @@ class MobileNetV3(nn.Layer):
             padding=1,
             groups=1,
             if_act=True,
-            act='hardswish',
-            name='conv1')
+            act='hardswish')
 
         self.stages = []
         self.out_channels = []
@@ -125,8 +124,7 @@ class MobileNetV3(nn.Layer):
                     kernel_size=k,
                     stride=s,
                     use_se=se,
-                    act=nl,
-                    name="conv" + str(i + 2)))
+                    act=nl))
             inplanes = make_divisible(scale * c)
             i += 1
         block_list.append(
@@ -138,8 +136,7 @@ class MobileNetV3(nn.Layer):
                 padding=0,
                 groups=1,
                 if_act=True,
-                act='hardswish',
-                name='conv_last'))
+                act='hardswish'))
         self.stages.append(nn.Sequential(*block_list))
         self.out_channels.append(make_divisible(scale * cls_ch_squeeze))
         for i, stage in enumerate(self.stages):
@@ -163,8 +160,7 @@ class ConvBNLayer(nn.Layer):
                  padding,
                  groups=1,
                  if_act=True,
-                 act=None,
-                 name=None):
+                 act=None):
         super(ConvBNLayer, self).__init__()
         self.if_act = if_act
         self.act = act
@@ -175,16 +171,9 @@ class ConvBNLayer(nn.Layer):
             stride=stride,
             padding=padding,
             groups=groups,
-            weight_attr=ParamAttr(name=name + '_weights'),
             bias_attr=False)
 
-        self.bn = nn.BatchNorm(
-            num_channels=out_channels,
-            act=None,
-            param_attr=ParamAttr(name=name + "_bn_scale"),
-            bias_attr=ParamAttr(name=name + "_bn_offset"),
-            moving_mean_name=name + "_bn_mean",
-            moving_variance_name=name + "_bn_variance")
+        self.bn = nn.BatchNorm(num_channels=out_channels, act=None)
 
     def forward(self, x):
         x = self.conv(x)
@@ -209,8 +198,7 @@ class ResidualUnit(nn.Layer):
                  kernel_size,
                  stride,
                  use_se,
-                 act=None,
-                 name=''):
+                 act=None):
         super(ResidualUnit, self).__init__()
         self.if_shortcut = stride == 1 and in_channels == out_channels
         self.if_se = use_se
@@ -222,8 +210,7 @@ class ResidualUnit(nn.Layer):
             stride=1,
             padding=0,
             if_act=True,
-            act=act,
-            name=name + "_expand")
+            act=act)
         self.bottleneck_conv = ConvBNLayer(
             in_channels=mid_channels,
             out_channels=mid_channels,
@@ -232,10 +219,9 @@ class ResidualUnit(nn.Layer):
             padding=int((kernel_size - 1) // 2),
             groups=mid_channels,
             if_act=True,
-            act=act,
-            name=name + "_depthwise")
+            act=act)
         if self.if_se:
-            self.mid_se = SEModule(mid_channels, name=name + "_se")
+            self.mid_se = SEModule(mid_channels)
         self.linear_conv = ConvBNLayer(
             in_channels=mid_channels,
             out_channels=out_channels,
@@ -243,8 +229,7 @@ class ResidualUnit(nn.Layer):
             stride=1,
             padding=0,
             if_act=False,
-            act=None,
-            name=name + "_linear")
+            act=None)
 
     def forward(self, inputs):
         x = self.expand_conv(inputs)
@@ -258,7 +243,7 @@ class ResidualUnit(nn.Layer):
 
 
 class SEModule(nn.Layer):
-    def __init__(self, in_channels, reduction=4, name=""):
+    def __init__(self, in_channels, reduction=4):
         super(SEModule, self).__init__()
         self.avg_pool = nn.AdaptiveAvgPool2D(1)
         self.conv1 = nn.Conv2D(
@@ -266,17 +251,13 @@ class SEModule(nn.Layer):
             out_channels=in_channels // reduction,
             kernel_size=1,
             stride=1,
-            padding=0,
-            weight_attr=ParamAttr(name=name + "_1_weights"),
-            bias_attr=ParamAttr(name=name + "_1_offset"))
+            padding=0)
         self.conv2 = nn.Conv2D(
             in_channels=in_channels // reduction,
             out_channels=in_channels,
             kernel_size=1,
             stride=1,
-            padding=0,
-            weight_attr=ParamAttr(name + "_2_weights"),
-            bias_attr=ParamAttr(name=name + "_2_offset"))
+            padding=0)
 
     def forward(self, inputs):
         outputs = self.avg_pool(inputs)

ppocr/modeling/backbones/rec_mobilenet_v3.py

@@ -96,8 +96,7 @@ class MobileNetV3(nn.Layer):
             padding=1,
             groups=1,
             if_act=True,
-            act='hardswish',
-            name='conv1')
+            act='hardswish')
         i = 0
         block_list = []
         inplanes = make_divisible(inplanes * scale)
@@ -110,8 +109,7 @@ class MobileNetV3(nn.Layer):
                     kernel_size=k,
                     stride=s,
                     use_se=se,
-                    act=nl,
-                    name='conv' + str(i + 2)))
+                    act=nl))
             inplanes = make_divisible(scale * c)
             i += 1
         self.blocks = nn.Sequential(*block_list)
@@ -124,8 +122,7 @@ class MobileNetV3(nn.Layer):
             padding=0,
             groups=1,
             if_act=True,
-            act='hardswish',
-            name='conv_last')
+            act='hardswish')
 
         self.pool = nn.MaxPool2D(kernel_size=2, stride=2, padding=0)
         self.out_channels = make_divisible(scale * cls_ch_squeeze)

ppocr/modeling/backbones/rec_mv1_enhance.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 numpy as np
+import paddle
+from paddle import ParamAttr
+import paddle.nn as nn
+import paddle.nn.functional as F
+from paddle.nn import Conv2D, BatchNorm, Linear, Dropout
+from paddle.nn import AdaptiveAvgPool2D, MaxPool2D, AvgPool2D
+from paddle.nn.initializer import KaimingNormal
+import math
+import numpy as np
+import paddle
+from paddle import ParamAttr, reshape, transpose, concat, split
+import paddle.nn as nn
+import paddle.nn.functional as F
+from paddle.nn import Conv2D, BatchNorm, Linear, Dropout
+from paddle.nn import AdaptiveAvgPool2D, MaxPool2D, AvgPool2D
+from paddle.nn.initializer import KaimingNormal
+import math
+from paddle.nn.functional import hardswish, hardsigmoid
+from paddle.regularizer import L2Decay
+
+
+class ConvBNLayer(nn.Layer):
+    def __init__(self,
+                 num_channels,
+                 filter_size,
+                 num_filters,
+                 stride,
+                 padding,
+                 channels=None,
+                 num_groups=1,
+                 act='hard_swish'):
+        super(ConvBNLayer, self).__init__()
+
+        self._conv = Conv2D(
+            in_channels=num_channels,
+            out_channels=num_filters,
+            kernel_size=filter_size,
+            stride=stride,
+            padding=padding,
+            groups=num_groups,
+            weight_attr=ParamAttr(initializer=KaimingNormal()),
+            bias_attr=False)
+
+        self._batch_norm = BatchNorm(
+            num_filters,
+            act=act,
+            param_attr=ParamAttr(regularizer=L2Decay(0.0)),
+            bias_attr=ParamAttr(regularizer=L2Decay(0.0)))
+
+    def forward(self, inputs):
+        y = self._conv(inputs)
+        y = self._batch_norm(y)
+        return y
+
+
+class DepthwiseSeparable(nn.Layer):
+    def __init__(self,
+                 num_channels,
+                 num_filters1,
+                 num_filters2,
+                 num_groups,
+                 stride,
+                 scale,
+                 dw_size=3,
+                 padding=1,
+                 use_se=False):
+        super(DepthwiseSeparable, self).__init__()
+        self.use_se = use_se
+        self._depthwise_conv = ConvBNLayer(
+            num_channels=num_channels,
+            num_filters=int(num_filters1 * scale),
+            filter_size=dw_size,
+            stride=stride,
+            padding=padding,
+            num_groups=int(num_groups * scale))
+        if use_se:
+            self._se = SEModule(int(num_filters1 * scale))
+        self._pointwise_conv = ConvBNLayer(
+            num_channels=int(num_filters1 * scale),
+            filter_size=1,
+            num_filters=int(num_filters2 * scale),
+            stride=1,
+            padding=0)
+
+    def forward(self, inputs):
+        y = self._depthwise_conv(inputs)
+        if self.use_se:
+            y = self._se(y)
+        y = self._pointwise_conv(y)
+        return y
+
+
+class MobileNetV1Enhance(nn.Layer):
+    def __init__(self, in_channels=3, scale=0.5, **kwargs):
+        super().__init__()
+        self.scale = scale
+        self.block_list = []
+
+        self.conv1 = ConvBNLayer(
+            num_channels=3,
+            filter_size=3,
+            channels=3,
+            num_filters=int(32 * scale),
+            stride=2,
+            padding=1)
+
+        conv2_1 = DepthwiseSeparable(
+            num_channels=int(32 * scale),
+            num_filters1=32,
+            num_filters2=64,
+            num_groups=32,
+            stride=1,
+            scale=scale)
+        self.block_list.append(conv2_1)
+
+        conv2_2 = DepthwiseSeparable(
+            num_channels=int(64 * scale),
+            num_filters1=64,
+            num_filters2=128,
+            num_groups=64,
+            stride=1,
+            scale=scale)
+        self.block_list.append(conv2_2)
+
+        conv3_1 = DepthwiseSeparable(
+            num_channels=int(128 * scale),
+            num_filters1=128,
+            num_filters2=128,
+            num_groups=128,
+            stride=1,
+            scale=scale)
+        self.block_list.append(conv3_1)
+
+        conv3_2 = DepthwiseSeparable(
+            num_channels=int(128 * scale),
+            num_filters1=128,
+            num_filters2=256,
+            num_groups=128,
+            stride=(2, 1),
+            scale=scale)
+        self.block_list.append(conv3_2)
+
+        conv4_1 = DepthwiseSeparable(
+            num_channels=int(256 * scale),
+            num_filters1=256,
+            num_filters2=256,
+            num_groups=256,
+            stride=1,
+            scale=scale)
+        self.block_list.append(conv4_1)
+
+        conv4_2 = DepthwiseSeparable(
+            num_channels=int(256 * scale),
+            num_filters1=256,
+            num_filters2=512,
+            num_groups=256,
+            stride=(2, 1),
+            scale=scale)
+        self.block_list.append(conv4_2)
+
+        for _ in range(5):
+            conv5 = DepthwiseSeparable(
+                num_channels=int(512 * scale),
+                num_filters1=512,
+                num_filters2=512,
+                num_groups=512,
+                stride=1,
+                dw_size=5,
+                padding=2,
+                scale=scale,
+                use_se=False)
+            self.block_list.append(conv5)
+
+        conv5_6 = DepthwiseSeparable(
+            num_channels=int(512 * scale),
+            num_filters1=512,
+            num_filters2=1024,
+            num_groups=512,
+            stride=(2, 1),
+            dw_size=5,
+            padding=2,
+            scale=scale,
+            use_se=True)
+        self.block_list.append(conv5_6)
+
+        conv6 = DepthwiseSeparable(
+            num_channels=int(1024 * scale),
+            num_filters1=1024,
+            num_filters2=1024,
+            num_groups=1024,
+            stride=1,
+            dw_size=5,
+            padding=2,
+            use_se=True,
+            scale=scale)
+        self.block_list.append(conv6)
+
+        self.block_list = nn.Sequential(*self.block_list)
+
+        self.pool = nn.MaxPool2D(kernel_size=2, stride=2, padding=0)
+        self.out_channels = int(1024 * scale)
+
+    def forward(self, inputs):
+        y = self.conv1(inputs)
+        y = self.block_list(y)
+        y = self.pool(y)
+        return y
+
+
+class SEModule(nn.Layer):
+    def __init__(self, channel, reduction=4):
+        super(SEModule, self).__init__()
+        self.avg_pool = AdaptiveAvgPool2D(1)
+        self.conv1 = Conv2D(
+            in_channels=channel,
+            out_channels=channel // reduction,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            weight_attr=ParamAttr(),
+            bias_attr=ParamAttr())
+        self.conv2 = Conv2D(
+            in_channels=channel // reduction,
+            out_channels=channel,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            weight_attr=ParamAttr(),
+            bias_attr=ParamAttr())
+
+    def forward(self, inputs):
+        outputs = self.avg_pool(inputs)
+        outputs = self.conv1(outputs)
+        outputs = F.relu(outputs)
+        outputs = self.conv2(outputs)
+        outputs = hardsigmoid(outputs)
+        return paddle.multiply(x=inputs, y=outputs)

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
+import paddle
+
+
+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 = paddle.transpose(x, [0, 3, 2, 1])
+            x_shape = paddle.shape(x)
+            x = paddle.reshape(
+                x, [x_shape[0], x_shape[1], x_shape[2] * x_shape[3]])
+        return x

ppocr/modeling/backbones/rec_resnet_31.py

+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
+import numpy as np
+
+__all__ = ["ResNet31"]
+
+
+def conv3x3(in_channel, out_channel, stride=1):
+    return nn.Conv2D(
+        in_channel,
+        out_channel,
+        kernel_size=3,
+        stride=stride,
+        padding=1,
+        bias_attr=False
+    )
+
+
+class BasicBlock(nn.Layer):
+    expansion = 1
+    def __init__(self, in_channels, channels, stride=1, downsample=False):
+        super().__init__()
+        self.conv1 = conv3x3(in_channels, channels, stride)
+        self.bn1 = nn.BatchNorm2D(channels)
+        self.relu = nn.ReLU()
+        self.conv2 = conv3x3(channels, channels)
+        self.bn2 = nn.BatchNorm2D(channels)
+        self.downsample = downsample
+        if downsample:
+            self.downsample = nn.Sequential(
+                nn.Conv2D(in_channels, channels * self.expansion, 1, stride, bias_attr=False),
+                nn.BatchNorm2D(channels * self.expansion),
+            )
+        else:
+            self.downsample = nn.Sequential()
+        self.stride = stride
+
+    def forward(self, x):
+        residual = x
+
+        out = self.conv1(x)
+        out = self.bn1(out)
+        out = self.relu(out)
+
+        out = self.conv2(out)
+        out = self.bn2(out)
+
+        if self.downsample:
+            residual = self.downsample(x)
+
+        out += residual
+        out = self.relu(out)
+
+        return out        
+
+
+class ResNet31(nn.Layer):
+    '''
+    Args:
+        in_channels (int): Number of channels of input image tensor.
+        layers (list[int]): List of BasicBlock number for each stage.
+        channels (list[int]): List of out_channels of Conv2d layer.
+        out_indices (None | Sequence[int]): Indices of output stages.
+        last_stage_pool (bool): If True, add `MaxPool2d` layer to last stage.
+    '''
+    def __init__(self, 
+                in_channels=3, 
+                layers=[1, 2, 5, 3],
+                channels=[64, 128, 256, 256, 512, 512, 512],
+                out_indices=None,
+                last_stage_pool=False):
+        super(ResNet31, self).__init__()
+        assert isinstance(in_channels, int)
+        assert isinstance(last_stage_pool, bool)
+
+        self.out_indices = out_indices
+        self.last_stage_pool = last_stage_pool
+
+        # conv 1 (Conv Conv)
+        self.conv1_1 = nn.Conv2D(in_channels, channels[0], kernel_size=3, stride=1, padding=1)
+        self.bn1_1 = nn.BatchNorm2D(channels[0])
+        self.relu1_1 = nn.ReLU()
+
+        self.conv1_2 = nn.Conv2D(channels[0], channels[1], kernel_size=3, stride=1, padding=1)
+        self.bn1_2 = nn.BatchNorm2D(channels[1])
+        self.relu1_2 = nn.ReLU()
+
+        # conv 2 (Max-pooling, Residual block, Conv)
+        self.pool2 = nn.MaxPool2D(kernel_size=2, stride=2, padding=0, ceil_mode=True)
+        self.block2 = self._make_layer(channels[1], channels[2], layers[0])
+        self.conv2 = nn.Conv2D(channels[2], channels[2], kernel_size=3, stride=1, padding=1)
+        self.bn2 = nn.BatchNorm2D(channels[2])
+        self.relu2 = nn.ReLU()
+
+        # conv 3 (Max-pooling, Residual block, Conv)
+        self.pool3 = nn.MaxPool2D(kernel_size=2, stride=2, padding=0, ceil_mode=True)
+        self.block3 = self._make_layer(channels[2], channels[3], layers[1])
+        self.conv3 = nn.Conv2D(channels[3], channels[3], kernel_size=3, stride=1, padding=1)
+        self.bn3 = nn.BatchNorm2D(channels[3])
+        self.relu3 = nn.ReLU()
+
+        # conv 4 (Max-pooling, Residual block, Conv)
+        self.pool4 = nn.MaxPool2D(kernel_size=(2, 1), stride=(2, 1), padding=0, ceil_mode=True)
+        self.block4 = self._make_layer(channels[3], channels[4], layers[2])
+        self.conv4 = nn.Conv2D(channels[4], channels[4], kernel_size=3, stride=1, padding=1)
+        self.bn4 = nn.BatchNorm2D(channels[4])
+        self.relu4 = nn.ReLU()
+
+        # conv 5 ((Max-pooling), Residual block, Conv)
+        self.pool5 = None
+        if self.last_stage_pool:
+            self.pool5 = nn.MaxPool2D(kernel_size=2, stride=2, padding=0, ceil_mode=True)
+        self.block5 = self._make_layer(channels[4], channels[5], layers[3])
+        self.conv5 = nn.Conv2D(channels[5], channels[5], kernel_size=3, stride=1, padding=1)
+        self.bn5 = nn.BatchNorm2D(channels[5])
+        self.relu5 = nn.ReLU()
+
+        self.out_channels = channels[-1]
+    
+    def _make_layer(self, input_channels, output_channels, blocks):
+        layers = []
+        for _ in range(blocks):
+            downsample = None
+            if input_channels != output_channels:
+                downsample = nn.Sequential(
+                    nn.Conv2D(
+                        input_channels, 
+                        output_channels, 
+                        kernel_size=1, 
+                        stride=1, 
+                        bias_attr=False),
+                    nn.BatchNorm2D(output_channels),
+                )
+                
+            layers.append(BasicBlock(input_channels, output_channels, downsample=downsample))
+            input_channels = output_channels
+        return nn.Sequential(*layers)
+
+
+    def forward(self, x):
+        x = self.conv1_1(x)
+        x = self.bn1_1(x)
+        x = self.relu1_1(x)
+
+        x = self.conv1_2(x)
+        x = self.bn1_2(x)
+        x = self.relu1_2(x)
+
+        outs = []
+        for i in range(4):
+            layer_index = i + 2
+            pool_layer = getattr(self, f'pool{layer_index}')
+            block_layer = getattr(self, f'block{layer_index}')
+            conv_layer = getattr(self, f'conv{layer_index}')
+            bn_layer = getattr(self, f'bn{layer_index}')
+            relu_layer = getattr(self, f'relu{layer_index}')
+
+            if pool_layer is not None:
+                x = pool_layer(x)
+            x = block_layer(x)
+            x = conv_layer(x)
+            x = bn_layer(x)
+            x= relu_layer(x)
+
+            outs.append(x)
+        
+        if self.out_indices is not None:
+            return tuple([outs[i] for i in self.out_indices])
+        
+        return x

ppocr/modeling/backbones/rec_resnet_aster.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 paddle
+import paddle.nn as nn
+
+import sys
+import math
+
+
+def conv3x3(in_planes, out_planes, stride=1):
+    """3x3 convolution with padding"""
+    return nn.Conv2D(
+        in_planes,
+        out_planes,
+        kernel_size=3,
+        stride=stride,
+        padding=1,
+        bias_attr=False)
+
+
+def conv1x1(in_planes, out_planes, stride=1):
+    """1x1 convolution"""
+    return nn.Conv2D(
+        in_planes, out_planes, kernel_size=1, stride=stride, bias_attr=False)
+
+
+def get_sinusoid_encoding(n_position, feat_dim, wave_length=10000):
+    # [n_position]
+    positions = paddle.arange(0, n_position)
+    # [feat_dim]
+    dim_range = paddle.arange(0, feat_dim)
+    dim_range = paddle.pow(wave_length, 2 * (dim_range // 2) / feat_dim)
+    # [n_position, feat_dim]
+    angles = paddle.unsqueeze(
+        positions, axis=1) / paddle.unsqueeze(
+            dim_range, axis=0)
+    angles = paddle.cast(angles, "float32")
+    angles[:, 0::2] = paddle.sin(angles[:, 0::2])
+    angles[:, 1::2] = paddle.cos(angles[:, 1::2])
+    return angles
+
+
+class AsterBlock(nn.Layer):
+    def __init__(self, inplanes, planes, stride=1, downsample=None):
+        super(AsterBlock, self).__init__()
+        self.conv1 = conv1x1(inplanes, planes, stride)
+        self.bn1 = nn.BatchNorm2D(planes)
+        self.relu = nn.ReLU()
+        self.conv2 = conv3x3(planes, planes)
+        self.bn2 = nn.BatchNorm2D(planes)
+        self.downsample = downsample
+        self.stride = stride
+
+    def forward(self, x):
+        residual = x
+        out = self.conv1(x)
+        out = self.bn1(out)
+        out = self.relu(out)
+        out = self.conv2(out)
+        out = self.bn2(out)
+
+        if self.downsample is not None:
+            residual = self.downsample(x)
+        out += residual
+        out = self.relu(out)
+        return out
+
+
+class ResNet_ASTER(nn.Layer):
+    """For aster or crnn"""
+
+    def __init__(self, with_lstm=True, n_group=1, in_channels=3):
+        super(ResNet_ASTER, self).__init__()
+        self.with_lstm = with_lstm
+        self.n_group = n_group
+
+        self.layer0 = nn.Sequential(
+            nn.Conv2D(
+                in_channels,
+                32,
+                kernel_size=(3, 3),
+                stride=1,
+                padding=1,
+                bias_attr=False),
+            nn.BatchNorm2D(32),
+            nn.ReLU())
+
+        self.inplanes = 32
+        self.layer1 = self._make_layer(32, 3, [2, 2])  # [16, 50]
+        self.layer2 = self._make_layer(64, 4, [2, 2])  # [8, 25]
+        self.layer3 = self._make_layer(128, 6, [2, 1])  # [4, 25]
+        self.layer4 = self._make_layer(256, 6, [2, 1])  # [2, 25]
+        self.layer5 = self._make_layer(512, 3, [2, 1])  # [1, 25]
+
+        if with_lstm:
+            self.rnn = nn.LSTM(512, 256, direction="bidirect", num_layers=2)
+            self.out_channels = 2 * 256
+        else:
+            self.out_channels = 512
+
+    def _make_layer(self, planes, blocks, stride):
+        downsample = None
+        if stride != [1, 1] or self.inplanes != planes:
+            downsample = nn.Sequential(
+                conv1x1(self.inplanes, planes, stride), nn.BatchNorm2D(planes))
+
+        layers = []
+        layers.append(AsterBlock(self.inplanes, planes, stride, downsample))
+        self.inplanes = planes
+        for _ in range(1, blocks):
+            layers.append(AsterBlock(self.inplanes, planes))
+        return nn.Sequential(*layers)
+
+    def forward(self, x):
+        x0 = self.layer0(x)
+        x1 = self.layer1(x0)
+        x2 = self.layer2(x1)
+        x3 = self.layer3(x2)
+        x4 = self.layer4(x3)
+        x5 = self.layer5(x4)
+
+        cnn_feat = x5.squeeze(2)  # [N, c, w]
+        cnn_feat = paddle.transpose(cnn_feat, perm=[0, 2, 1])
+        if self.with_lstm:
+            rnn_feat, _ = self.rnn(cnn_feat)
+            return rnn_feat
+        else:
+            return cnn_feat