first commit

ppocr/losses/rec_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
+
+
+class AttentionLoss(nn.Layer):
+    def __init__(self, **kwargs):
+        super(AttentionLoss, self).__init__()
+        self.loss_func = nn.CrossEntropyLoss(weight=None, reduction='none')
+
+    def forward(self, predicts, batch):
+        targets = batch[1].astype("int64")
+        label_lengths = batch[2].astype('int64')
+        batch_size, num_steps, num_classes = predicts.shape[0], predicts.shape[
+            1], predicts.shape[2]
+        assert len(targets.shape) == len(list(predicts.shape)) - 1, \
+            "The target's shape and inputs's shape is [N, d] and [N, num_steps]"
+
+        inputs = paddle.reshape(predicts, [-1, predicts.shape[-1]])
+        targets = paddle.reshape(targets, [-1])
+
+        return {'loss': paddle.sum(self.loss_func(inputs, targets))}

ppocr/losses/rec_ctc_loss.py

+# copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import paddle
+from paddle import nn
+
+
+class CTCLoss(nn.Layer):
+    def __init__(self, use_focal_loss=False, **kwargs):
+        super(CTCLoss, self).__init__()
+        self.loss_func = nn.CTCLoss(blank=0, reduction='none')
+        self.use_focal_loss = use_focal_loss
+
+    def forward(self, predicts, batch):
+        if isinstance(predicts, (list, tuple)):
+            predicts = predicts[-1]
+        predicts = predicts.transpose((1, 0, 2))
+        N, B, _ = predicts.shape
+        preds_lengths = paddle.to_tensor(
+            [N] * B, dtype='int64', place=paddle.CPUPlace())
+        labels = batch[1].astype("int32")
+        label_lengths = batch[2].astype('int64')
+        loss = self.loss_func(predicts, labels, preds_lengths, label_lengths)
+        if self.use_focal_loss:
+            weight = paddle.exp(-loss)
+            weight = paddle.subtract(paddle.to_tensor([1.0]), weight)
+            weight = paddle.square(weight)
+            loss = paddle.multiply(loss, weight)
+        loss = loss.mean()
+        return {'loss': loss}

ppocr/losses/rec_enhanced_ctc_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 .ace_loss import ACELoss
+from .center_loss import CenterLoss
+from .rec_ctc_loss import CTCLoss
+
+
+class EnhancedCTCLoss(nn.Layer):
+    def __init__(self,
+                 use_focal_loss=False,
+                 use_ace_loss=False,
+                 ace_loss_weight=0.1,
+                 use_center_loss=False,
+                 center_loss_weight=0.05,
+                 num_classes=6625,
+                 feat_dim=96,
+                 init_center=False,
+                 center_file_path=None,
+                 **kwargs):
+        super(EnhancedCTCLoss, self).__init__()
+        self.ctc_loss_func = CTCLoss(use_focal_loss=use_focal_loss)
+
+        self.use_ace_loss = False
+        if use_ace_loss:
+            self.use_ace_loss = use_ace_loss
+            self.ace_loss_func = ACELoss()
+            self.ace_loss_weight = ace_loss_weight
+
+        self.use_center_loss = False
+        if use_center_loss:
+            self.use_center_loss = use_center_loss
+            self.center_loss_func = CenterLoss(
+                num_classes=num_classes,
+                feat_dim=feat_dim,
+                init_center=init_center,
+                center_file_path=center_file_path)
+            self.center_loss_weight = center_loss_weight
+
+    def __call__(self, predicts, batch):
+        loss = self.ctc_loss_func(predicts, batch)["loss"]
+
+        if self.use_center_loss:
+            center_loss = self.center_loss_func(
+                predicts, batch)["loss_center"] * self.center_loss_weight
+            loss = loss + center_loss
+
+        if self.use_ace_loss:
+            ace_loss = self.ace_loss_func(
+                predicts, batch)["loss_ace"] * self.ace_loss_weight
+            loss = loss + ace_loss
+
+        return {'enhanced_ctc_loss': loss}

ppocr/losses/rec_multi_loss.py

+# copyright (c) 2022 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 .rec_ctc_loss import CTCLoss
+from .rec_sar_loss import SARLoss
+
+
+class MultiLoss(nn.Layer):
+    def __init__(self, **kwargs):
+        super().__init__()
+        self.loss_funcs = {}
+        self.loss_list = kwargs.pop('loss_config_list')
+        self.weight_1 = kwargs.get('weight_1', 1.0)
+        self.weight_2 = kwargs.get('weight_2', 1.0)
+        self.gtc_loss = kwargs.get('gtc_loss', 'sar')
+        for loss_info in self.loss_list:
+            for name, param in loss_info.items():
+                if param is not None:
+                    kwargs.update(param)
+                loss = eval(name)(**kwargs)
+                self.loss_funcs[name] = loss
+
+    def forward(self, predicts, batch):
+        self.total_loss = {}
+        total_loss = 0.0
+        # batch [image, label_ctc, label_sar, length, valid_ratio]
+        for name, loss_func in self.loss_funcs.items():
+            if name == 'CTCLoss':
+                loss = loss_func(predicts['ctc'],
+                                 batch[:2] + batch[3:])['loss'] * self.weight_1
+            elif name == 'SARLoss':
+                loss = loss_func(predicts['sar'],
+                                 batch[:1] + batch[2:])['loss'] * self.weight_2
+            else:
+                raise NotImplementedError(
+                    '{} is not supported in MultiLoss yet'.format(name))
+            self.total_loss[name] = loss
+            total_loss += loss
+        self.total_loss['loss'] = total_loss
+        return self.total_loss

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=tgt.dtype))
+            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_pren_loss.py

+# copyright (c) 2022 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
+
+from paddle import nn
+
+
+class PRENLoss(nn.Layer):
+    def __init__(self, **kwargs):
+        super(PRENLoss, self).__init__()
+        # note: 0 is padding idx
+        self.loss_func = nn.CrossEntropyLoss(reduction='mean', ignore_index=0)
+
+    def forward(self, predicts, batch):
+        loss = self.loss_func(predicts, batch[1].astype('int64'))
+        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__()
+        ignore_index = kwargs.get('ignore_index', 92)  # 6626
+        self.loss_func = paddle.nn.loss.CrossEntropyLoss(
+            reduction="mean", ignore_index=ignore_index)
+
+    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/rec_srn_loss.py

+# copyright (c) 2020 PaddlePaddle Authors. All Rights Reserve.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import paddle
+from paddle import nn
+
+
+class SRNLoss(nn.Layer):
+    def __init__(self, **kwargs):
+        super(SRNLoss, self).__init__()
+        self.loss_func = paddle.nn.loss.CrossEntropyLoss(reduction="sum")
+
+    def forward(self, predicts, batch):
+        predict = predicts['predict']
+        word_predict = predicts['word_out']
+        gsrm_predict = predicts['gsrm_out']
+        label = batch[1]
+
+        casted_label = paddle.cast(x=label, dtype='int64')
+        casted_label = paddle.reshape(x=casted_label, shape=[-1, 1])
+
+        cost_word = self.loss_func(word_predict, label=casted_label)
+        cost_gsrm = self.loss_func(gsrm_predict, label=casted_label)
+        cost_vsfd = self.loss_func(predict, label=casted_label)
+
+        cost_word = paddle.reshape(x=paddle.sum(cost_word), shape=[1])
+        cost_gsrm = paddle.reshape(x=paddle.sum(cost_gsrm), shape=[1])
+        cost_vsfd = paddle.reshape(x=paddle.sum(cost_vsfd), shape=[1])
+
+        sum_cost = cost_word * 3.0 + cost_vsfd + cost_gsrm * 0.15
+
+        return {'loss': sum_cost, 'word_loss': cost_word, 'img_loss': cost_vsfd}

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
+
+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 = paddle.maximum(preds[:, 0], bbox[:, 0])
+        iy1 = paddle.maximum(preds[:, 1], bbox[:, 1])
+        ix2 = paddle.minimum(preds[:, 2], bbox[:, 2])
+        iy2 = paddle.minimum(preds[:, 3], bbox[:, 3])
+
+        iw = paddle.clip(ix2 - ix1 + 1e-3, 0., 1e10)
+        ih = paddle.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 = paddle.minimum(preds[:, 0], bbox[:, 0])
+        ey1 = paddle.minimum(preds[:, 1], bbox[:, 1])
+        ex2 = paddle.maximum(preds[:, 2], bbox[:, 2])
+        ey2 = paddle.maximum(preds[:, 3], bbox[:, 3])
+        ew = paddle.clip(ex2 - ex1 + 1e-3, 0., 1e10)
+        eh = paddle.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/losses/vqa_token_layoutlm_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
+
+from paddle import nn
+
+
+class VQASerTokenLayoutLMLoss(nn.Layer):
+    def __init__(self, num_classes):
+        super().__init__()
+        self.loss_class = nn.CrossEntropyLoss()
+        self.num_classes = num_classes
+        self.ignore_index = self.loss_class.ignore_index
+
+    def forward(self, predicts, batch):
+        labels = batch[1]
+        attention_mask = batch[4]
+        if attention_mask is not None:
+            active_loss = attention_mask.reshape([-1, ]) == 1
+            active_outputs = predicts.reshape(
+                [-1, self.num_classes])[active_loss]
+            active_labels = labels.reshape([-1, ])[active_loss]
+            loss = self.loss_class(active_outputs, active_labels)
+        else:
+            loss = self.loss_class(
+                predicts.reshape([-1, self.num_classes]),
+                labels.reshape([-1, ]))
+        return {'loss': loss}

ppocr/metrics/__init__.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
+from __future__ import unicode_literals
+
+import copy
+
+__all__ = ["build_metric"]
+
+from .det_metric import DetMetric, DetFCEMetric
+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
+from .vqa_token_ser_metric import VQASerTokenMetric
+from .vqa_token_re_metric import VQAReTokenMetric
+
+
+def build_metric(config):
+    support_dict = [
+        "DetMetric", "DetFCEMetric", "RecMetric", "ClsMetric", "E2EMetric",
+        "DistillationMetric", "TableMetric", 'KIEMetric', 'VQASerTokenMetric',
+        'VQAReTokenMetric'
+    ]
+
+    config = copy.deepcopy(config)
+    module_name = config.pop("name")
+    assert module_name in support_dict, Exception(
+        "metric only support {}".format(support_dict))
+    module_class = eval(module_name)(**config)
+    return module_class