fix conflict

ppocr/losses/combined_loss.py

+# Copyright (c) 2020 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.
+
+import paddle
+import paddle.nn as nn
+
+from .distillation_loss import DistillationCTCLoss
+from .distillation_loss import DistillationDMLLoss
+from .distillation_loss import DistillationDistanceLoss, DistillationDBLoss, DistillationDilaDBLoss
+
+
+class CombinedLoss(nn.Layer):
+    """
+    CombinedLoss:
+        a combionation of loss function
+    """
+
+    def __init__(self, loss_config_list=None):
+        super().__init__()
+        self.loss_func = []
+        self.loss_weight = []
+        assert isinstance(loss_config_list, list), (
+            'operator config should be a list')
+        for config in loss_config_list:
+            assert isinstance(config,
+                              dict) and len(config) == 1, "yaml format error"
+            name = list(config)[0]
+            param = config[name]
+            assert "weight" in param, "weight must be in param, but param just contains {}".format(
+                param.keys())
+            self.loss_weight.append(param.pop("weight"))
+            self.loss_func.append(eval(name)(**param))
+
+    def forward(self, input, batch, **kargs):
+        loss_dict = {}
+        loss_all = 0.
+        for idx, loss_func in enumerate(self.loss_func):
+            loss = loss_func(input, batch, **kargs)
+            if isinstance(loss, paddle.Tensor):
+                loss = {"loss_{}_{}".format(str(loss), idx): loss}
+            weight = self.loss_weight[idx]
+            for key in loss.keys():
+                if key == "loss":
+                    loss_all += loss[key] * weight
+                else:
+                    loss_dict["{}_{}".format(key, idx)] = loss[key]
+        loss_dict["loss"] = loss_all
+        return loss_dict

ppocr/losses/distillation_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.
+
+import paddle
+import paddle.nn as nn
+import numpy as np
+import cv2
+
+from .rec_ctc_loss import CTCLoss
+from .basic_loss import DMLLoss
+from .basic_loss import DistanceLoss
+from .det_db_loss import DBLoss
+from .det_basic_loss import BalanceLoss, MaskL1Loss, DiceLoss
+
+
+def _sum_loss(loss_dict):
+    if "loss" in loss_dict.keys():
+        return loss_dict
+    else:
+        loss_dict["loss"] = 0.
+        for k, value in loss_dict.items():
+            if k == "loss":
+                continue
+            else:
+                loss_dict["loss"] += value
+        return loss_dict
+
+
+class DistillationDMLLoss(DMLLoss):
+    """
+    """
+
+    def __init__(self,
+                 model_name_pairs=[],
+                 act=None,
+                 key=None,
+                 maps_name=None,
+                 name="dml"):
+        super().__init__(act=act)
+        assert isinstance(model_name_pairs, list)
+        self.key = key
+        self.model_name_pairs = self._check_model_name_pairs(model_name_pairs)
+        self.name = name
+        self.maps_name = self._check_maps_name(maps_name)
+    
+    def _check_model_name_pairs(self, model_name_pairs):
+        if not isinstance(model_name_pairs, list):
+            return []
+        elif isinstance(model_name_pairs[0], list) and isinstance(model_name_pairs[0][0], str):
+            return model_name_pairs
+        else:
+            return [model_name_pairs]
+
+    def _check_maps_name(self, maps_name):
+        if maps_name is None:
+            return None
+        elif type(maps_name) == str:
+            return [maps_name]
+        elif type(maps_name) == list:
+            return [maps_name]
+        else:
+            return None
+
+    def _slice_out(self, outs):
+        new_outs = {}
+        for k in self.maps_name:
+            if k == "thrink_maps":
+                new_outs[k] = outs[:, 0, :, :]
+            elif k == "threshold_maps":
+                new_outs[k] = outs[:, 1, :, :]
+            elif k == "binary_maps":
+                new_outs[k] = outs[:, 2, :, :]
+            else:
+                continue
+        return new_outs
+
+    def forward(self, predicts, batch):
+        loss_dict = dict()
+        for idx, pair in enumerate(self.model_name_pairs):
+            out1 = predicts[pair[0]]
+            out2 = predicts[pair[1]]
+            if self.key is not None:
+                out1 = out1[self.key]
+                out2 = out2[self.key]
+
+            if self.maps_name is None:
+                loss = super().forward(out1, out2)
+                if isinstance(loss, dict):
+                    for key in loss:
+                        loss_dict["{}_{}_{}_{}".format(key, pair[0], pair[1],
+                                                       idx)] = loss[key]
+                else:
+                    loss_dict["{}_{}".format(self.name, idx)] = loss
+            else:
+                outs1 = self._slice_out(out1)
+                outs2 = self._slice_out(out2)
+                for _c, k in enumerate(outs1.keys()):
+                    loss = super().forward(outs1[k], outs2[k])
+                    if isinstance(loss, dict):
+                        for key in loss:
+                            loss_dict["{}_{}_{}_{}_{}".format(key, pair[
+                                0], pair[1], map_name, idx)] = loss[key]
+                    else:
+                        loss_dict["{}_{}_{}".format(self.name, self.maps_name[_c],
+                                                    idx)] = loss
+        
+        loss_dict = _sum_loss(loss_dict)
+
+        return loss_dict
+
+
+class DistillationCTCLoss(CTCLoss):
+    def __init__(self, model_name_list=[], key=None, name="loss_ctc"):
+        super().__init__()
+        self.model_name_list = model_name_list
+        self.key = key
+        self.name = name
+
+    def forward(self, predicts, batch):
+        loss_dict = dict()
+        for idx, model_name in enumerate(self.model_name_list):
+            out = predicts[model_name]
+            if self.key is not None:
+                out = out[self.key]
+            loss = super().forward(out, batch)
+            if isinstance(loss, dict):
+                for key in loss:
+                    loss_dict["{}_{}_{}".format(self.name, model_name,
+                                                idx)] = loss[key]
+            else:
+                loss_dict["{}_{}".format(self.name, model_name)] = loss
+        return loss_dict
+
+
+class DistillationDBLoss(DBLoss):
+    def __init__(self,
+                 model_name_list=[],
+                 balance_loss=True,
+                 main_loss_type='DiceLoss',
+                 alpha=5,
+                 beta=10,
+                 ohem_ratio=3,
+                 eps=1e-6,
+                 name="db",
+                 **kwargs):
+        super().__init__()
+        self.model_name_list = model_name_list
+        self.name = name
+        self.key = None
+
+    def forward(self, predicts, batch):
+        loss_dict = {}
+        for idx, model_name in enumerate(self.model_name_list):
+            out = predicts[model_name]
+            if self.key is not None:
+                out = out[self.key]
+            loss = super().forward(out, batch)
+
+            if isinstance(loss, dict):
+                for key in loss.keys():
+                    if key == "loss":
+                        continue
+                    name = "{}_{}_{}".format(self.name, model_name, key)
+                    loss_dict[name] = loss[key]
+            else:
+                loss_dict["{}_{}".format(self.name, model_name)] = loss
+
+        loss_dict = _sum_loss(loss_dict)
+        return loss_dict
+
+
+class DistillationDilaDBLoss(DBLoss):
+    def __init__(self,
+                 model_name_pairs=[],
+                 key=None,
+                 balance_loss=True,
+                 main_loss_type='DiceLoss',
+                 alpha=5,
+                 beta=10,
+                 ohem_ratio=3,
+                 eps=1e-6,
+                 name="dila_dbloss"):
+        super().__init__()
+        self.model_name_pairs = model_name_pairs
+        self.name = name
+        self.key = key
+
+    def forward(self, predicts, batch):
+        loss_dict = dict()
+        for idx, pair in enumerate(self.model_name_pairs):
+            stu_outs = predicts[pair[0]]
+            tch_outs = predicts[pair[1]]
+            if self.key is not None:
+                stu_preds = stu_outs[self.key]
+                tch_preds = tch_outs[self.key]
+
+            stu_shrink_maps = stu_preds[:, 0, :, :]
+            stu_binary_maps = stu_preds[:, 2, :, :]
+
+            # dilation to teacher prediction
+            dilation_w = np.array([[1, 1], [1, 1]])
+            th_shrink_maps = tch_preds[:, 0, :, :]
+            th_shrink_maps = th_shrink_maps.numpy() > 0.3  # thresh = 0.3 
+            dilate_maps = np.zeros_like(th_shrink_maps).astype(np.float32)
+            for i in range(th_shrink_maps.shape[0]):
+                dilate_maps[i] = cv2.dilate(
+                    th_shrink_maps[i, :, :].astype(np.uint8), dilation_w)
+            th_shrink_maps = paddle.to_tensor(dilate_maps)
+
+            label_threshold_map, label_threshold_mask, label_shrink_map, label_shrink_mask = batch[
+                1:]
+
+            # calculate the shrink map loss
+            bce_loss = self.alpha * self.bce_loss(
+                stu_shrink_maps, th_shrink_maps, label_shrink_mask)
+            loss_binary_maps = self.dice_loss(stu_binary_maps, th_shrink_maps,
+                                              label_shrink_mask)
+
+            # k = f"{self.name}_{pair[0]}_{pair[1]}"
+            k = "{}_{}_{}".format(self.name, pair[0], pair[1])
+            loss_dict[k] = bce_loss + loss_binary_maps
+
+        loss_dict = _sum_loss(loss_dict)
+        return loss_dict
+
+
+class DistillationDistanceLoss(DistanceLoss):
+    """
+    """
+
+    def __init__(self,
+                 mode="l2",
+                 model_name_pairs=[],
+                 key=None,
+                 name="loss_distance",
+                 **kargs):
+        super().__init__(mode=mode, **kargs)
+        assert isinstance(model_name_pairs, list)
+        self.key = key
+        self.model_name_pairs = model_name_pairs
+        self.name = name + "_l2"
+
+    def forward(self, predicts, batch):
+        loss_dict = dict()
+        for idx, pair in enumerate(self.model_name_pairs):
+            out1 = predicts[pair[0]]
+            out2 = predicts[pair[1]]
+            if self.key is not None:
+                out1 = out1[self.key]
+                out2 = out2[self.key]
+            loss = super().forward(out1, out2)
+            if isinstance(loss, dict):
+                for key in loss:
+                    loss_dict["{}_{}_{}".format(self.name, key, idx)] = loss[
+                        key]
+            else:
+                loss_dict["{}_{}_{}_{}".format(self.name, pair[0], pair[1],
+                                               idx)] = loss
+        return loss_dict

ppocr/losses/rec_ctc_loss.py

@@ -25,7 +25,7 @@ class CTCLoss(nn.Layer):
         super(CTCLoss, self).__init__()
         self.loss_func = nn.CTCLoss(blank=0, reduction='none')
 
-    def __call__(self, predicts, batch):
+    def forward(self, predicts, batch):
         predicts = predicts.transpose((1, 0, 2))
         N, B, _ = predicts.shape
         preds_lengths = paddle.to_tensor([N] * B, dtype='int64')

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

@@ -19,20 +19,23 @@ 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
 
 def build_metric(config):
-    from .det_metric import DetMetric
-    from .rec_metric import RecMetric
-    from .cls_metric import ClsMetric
-    from .e2e_metric import E2EMetric
-
-    support_dict = ['DetMetric', 'RecMetric', 'ClsMetric', 'E2EMetric']
+    support_dict = [
+        "DetMetric", "RecMetric", "ClsMetric", "E2EMetric", "DistillationMetric", "TableMetric"
+    ]
 
     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/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.
@@ -32,7 +32,6 @@ 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 +67,23 @@ 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

@@ -12,29 +12,36 @@
 # 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
+        support_dict = [
+            "MobileNetV1Enhance", "MobileNetV3", "ResNet", "ResNetFPN"
+        ]
+    elif model_type == "e2e":
         from .e2e_resnet_vd_pg import ResNet
-        support_dict = ['ResNet']
+        support_dict = ["ResNet"]
+    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/heads/__init__.py

@@ -31,8 +31,10 @@ def build_head(config):
     from .cls_head import ClsHead
     support_dict = [
         'DBHead', 'EASTHead', 'SASTHead', 'CTCHead', 'ClsHead', 'AttentionHead',
-        'SRNHead', 'PGHead']
+        'SRNHead', 'PGHead', 'TableAttentionHead']
 
+    #table head
+    from .table_att_head import TableAttentionHead
 
     module_name = config.pop('name')
     assert module_name in support_dict, Exception('head only support {}'.format(

ppocr/modeling/heads/cls_head.py

@@ -43,7 +43,7 @@ class ClsHead(nn.Layer):
                 initializer=nn.initializer.Uniform(-stdv, stdv)),
             bias_attr=ParamAttr(name="fc_0.b_0"), )
 
-    def forward(self, x):
+    def forward(self, x, targets=None):
         x = self.pool(x)
         x = paddle.reshape(x, shape=[x.shape[0], x.shape[1]])
         x = self.fc(x)