Merge branch 'dygraph' of https://github.com/PaddlePaddle/PaddleOCR into fix_doc

ppocr/modeling/backbones/__init__.py

@@ -28,8 +28,10 @@ def build_backbone(config, model_type):
         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"
+            'MobileNetV1Enhance', 'MobileNetV3', 'ResNet', 'ResNetFPN', 'MTB',
+            "ResNet31", "ResNet_ASTER"
         ]
     elif model_type == "e2e":
         from .e2e_resnet_vd_pg import ResNet

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

ppocr/modeling/heads/__init__.py

@@ -29,13 +29,14 @@ def build_head(config):
     from .rec_srn_head import SRNHead
     from .rec_nrtr_head import Transformer
     from .rec_sar_head import SARHead
+    from .rec_aster_head import AsterHead
 
     # cls head
     from .cls_head import ClsHead
     support_dict = [
         'DBHead', 'PSEHead', 'EASTHead', 'SASTHead', 'CTCHead', 'ClsHead',
         'AttentionHead', 'SRNHead', 'PGHead', 'Transformer',
-        'TableAttentionHead', 'SARHead'
+        'TableAttentionHead', 'SARHead', 'AsterHead'
     ]
 
     #table head

ppocr/modeling/heads/rec_aster_head.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 sys
+
+import paddle
+from paddle import nn
+from paddle.nn import functional as F
+
+
+class AsterHead(nn.Layer):
+    def __init__(self,
+                 in_channels,
+                 out_channels,
+                 sDim,
+                 attDim,
+                 max_len_labels,
+                 time_step=25,
+                 beam_width=5,
+                 **kwargs):
+        super(AsterHead, self).__init__()
+        self.num_classes = out_channels
+        self.in_planes = in_channels
+        self.sDim = sDim
+        self.attDim = attDim
+        self.max_len_labels = max_len_labels
+        self.decoder = AttentionRecognitionHead(in_channels, out_channels, sDim,
+                                                attDim, max_len_labels)
+        self.time_step = time_step
+        self.embeder = Embedding(self.time_step, in_channels)
+        self.beam_width = beam_width
+        self.eos = self.num_classes - 1
+
+    def forward(self, x, targets=None, embed=None):
+        return_dict = {}
+        embedding_vectors = self.embeder(x)
+
+        if self.training:
+            rec_targets, rec_lengths, _ = targets
+            rec_pred = self.decoder([x, rec_targets, rec_lengths],
+                                    embedding_vectors)
+            return_dict['rec_pred'] = rec_pred
+            return_dict['embedding_vectors'] = embedding_vectors
+        else:
+            rec_pred, rec_pred_scores = self.decoder.beam_search(
+                x, self.beam_width, self.eos, embedding_vectors)
+            return_dict['rec_pred'] = rec_pred
+            return_dict['rec_pred_scores'] = rec_pred_scores
+            return_dict['embedding_vectors'] = embedding_vectors
+
+        return return_dict
+
+
+class Embedding(nn.Layer):
+    def __init__(self, in_timestep, in_planes, mid_dim=4096, embed_dim=300):
+        super(Embedding, self).__init__()
+        self.in_timestep = in_timestep
+        self.in_planes = in_planes
+        self.embed_dim = embed_dim
+        self.mid_dim = mid_dim
+        self.eEmbed = nn.Linear(
+            in_timestep * in_planes,
+            self.embed_dim)  # Embed encoder output to a word-embedding like
+
+    def forward(self, x):
+        x = paddle.reshape(x, [paddle.shape(x)[0], -1])
+        x = self.eEmbed(x)
+        return x
+
+
+class AttentionRecognitionHead(nn.Layer):
+    """
+  input: [b x 16 x 64 x in_planes]
+  output: probability sequence: [b x T x num_classes]
+  """
+
+    def __init__(self, in_channels, out_channels, sDim, attDim, max_len_labels):
+        super(AttentionRecognitionHead, self).__init__()
+        self.num_classes = out_channels  # this is the output classes. So it includes the <EOS>.
+        self.in_planes = in_channels
+        self.sDim = sDim
+        self.attDim = attDim
+        self.max_len_labels = max_len_labels
+
+        self.decoder = DecoderUnit(
+            sDim=sDim, xDim=in_channels, yDim=self.num_classes, attDim=attDim)
+
+    def forward(self, x, embed):
+        x, targets, lengths = x
+        batch_size = paddle.shape(x)[0]
+        # Decoder
+        state = self.decoder.get_initial_state(embed)
+        outputs = []
+        for i in range(max(lengths)):
+            if i == 0:
+                y_prev = paddle.full(
+                    shape=[batch_size], fill_value=self.num_classes)
+            else:
+                y_prev = targets[:, i - 1]
+            output, state = self.decoder(x, state, y_prev)
+            outputs.append(output)
+        outputs = paddle.concat([_.unsqueeze(1) for _ in outputs], 1)
+        return outputs
+
+    # inference stage.
+    def sample(self, x):
+        x, _, _ = x
+        batch_size = x.size(0)
+        # Decoder
+        state = paddle.zeros([1, batch_size, self.sDim])
+
+        predicted_ids, predicted_scores = [], []
+        for i in range(self.max_len_labels):
+            if i == 0:
+                y_prev = paddle.full(
+                    shape=[batch_size], fill_value=self.num_classes)
+            else:
+                y_prev = predicted
+
+            output, state = self.decoder(x, state, y_prev)
+            output = F.softmax(output, axis=1)
+            score, predicted = output.max(1)
+            predicted_ids.append(predicted.unsqueeze(1))
+            predicted_scores.append(score.unsqueeze(1))
+        predicted_ids = paddle.concat([predicted_ids, 1])
+        predicted_scores = paddle.concat([predicted_scores, 1])
+        # return predicted_ids.squeeze(), predicted_scores.squeeze()
+        return predicted_ids, predicted_scores
+
+    def beam_search(self, x, beam_width, eos, embed):
+        def _inflate(tensor, times, dim):
+            repeat_dims = [1] * tensor.dim()
+            repeat_dims[dim] = times
+            output = paddle.tile(tensor, repeat_dims)
+            return output
+
+        # https://github.com/IBM/pytorch-seq2seq/blob/fede87655ddce6c94b38886089e05321dc9802af/seq2seq/models/TopKDecoder.py
+        batch_size, l, d = x.shape
+        x = paddle.tile(
+            paddle.transpose(
+                x.unsqueeze(1), perm=[1, 0, 2, 3]), [beam_width, 1, 1, 1])
+        inflated_encoder_feats = paddle.reshape(
+            paddle.transpose(
+                x, perm=[1, 0, 2, 3]), [-1, l, d])
+
+        # Initialize the decoder
+        state = self.decoder.get_initial_state(embed, tile_times=beam_width)
+
+        pos_index = paddle.reshape(
+            paddle.arange(batch_size) * beam_width, shape=[-1, 1])
+
+        # Initialize the scores
+        sequence_scores = paddle.full(
+            shape=[batch_size * beam_width, 1], fill_value=-float('Inf'))
+        index = [i * beam_width for i in range(0, batch_size)]
+        sequence_scores[index] = 0.0
+
+        # Initialize the input vector
+        y_prev = paddle.full(
+            shape=[batch_size * beam_width], fill_value=self.num_classes)
+
+        # Store decisions for backtracking
+        stored_scores = list()
+        stored_predecessors = list()
+        stored_emitted_symbols = list()
+
+        for i in range(self.max_len_labels):
+            output, state = self.decoder(inflated_encoder_feats, state, y_prev)
+            state = paddle.unsqueeze(state, axis=0)
+            log_softmax_output = paddle.nn.functional.log_softmax(
+                output, axis=1)
+
+            sequence_scores = _inflate(sequence_scores, self.num_classes, 1)
+            sequence_scores += log_softmax_output
+            scores, candidates = paddle.topk(
+                paddle.reshape(sequence_scores, [batch_size, -1]),
+                beam_width,
+                axis=1)
+
+            # Reshape input = (bk, 1) and sequence_scores = (bk, 1)
+            y_prev = paddle.reshape(
+                candidates % self.num_classes, shape=[batch_size * beam_width])
+            sequence_scores = paddle.reshape(
+                scores, shape=[batch_size * beam_width, 1])
+
+            # Update fields for next timestep
+            pos_index = paddle.expand_as(pos_index, candidates)
+            predecessors = paddle.cast(
+                candidates / self.num_classes + pos_index, dtype='int64')
+            predecessors = paddle.reshape(
+                predecessors, shape=[batch_size * beam_width, 1])
+            state = paddle.index_select(
+                state, index=predecessors.squeeze(), axis=1)
+
+            # Update sequence socres and erase scores for <eos> symbol so that they aren't expanded
+            stored_scores.append(sequence_scores.clone())
+            y_prev = paddle.reshape(y_prev, shape=[-1, 1])
+            eos_prev = paddle.full_like(y_prev, fill_value=eos)
+            mask = eos_prev == y_prev
+            mask = paddle.nonzero(mask)
+            if mask.dim() > 0:
+                sequence_scores = sequence_scores.numpy()
+                mask = mask.numpy()
+                sequence_scores[mask] = -float('inf')
+                sequence_scores = paddle.to_tensor(sequence_scores)
+
+            # Cache results for backtracking
+            stored_predecessors.append(predecessors)
+            y_prev = paddle.squeeze(y_prev)
+            stored_emitted_symbols.append(y_prev)
+
+        # Do backtracking to return the optimal values
+        #====== backtrak ======#
+        # Initialize return variables given different types
+        p = list()
+        l = [[self.max_len_labels] * beam_width for _ in range(batch_size)
+             ]  # Placeholder for lengths of top-k sequences
+
+        # the last step output of the beams are not sorted
+        # thus they are sorted here
+        sorted_score, sorted_idx = paddle.topk(
+            paddle.reshape(
+                stored_scores[-1], shape=[batch_size, beam_width]),
+            beam_width)
+
+        # initialize the sequence scores with the sorted last step beam scores
+        s = sorted_score.clone()
+
+        batch_eos_found = [0] * batch_size  # the number of EOS found
+        # in the backward loop below for each batch
+        t = self.max_len_labels - 1
+        # initialize the back pointer with the sorted order of the last step beams.
+        # add pos_index for indexing variable with b*k as the first dimension.
+        t_predecessors = paddle.reshape(
+            sorted_idx + pos_index.expand_as(sorted_idx),
+            shape=[batch_size * beam_width])
+        while t >= 0:
+            # Re-order the variables with the back pointer
+            current_symbol = paddle.index_select(
+                stored_emitted_symbols[t], index=t_predecessors, axis=0)
+            t_predecessors = paddle.index_select(
+                stored_predecessors[t].squeeze(), index=t_predecessors, axis=0)
+            eos_indices = stored_emitted_symbols[t] == eos
+            eos_indices = paddle.nonzero(eos_indices)
+
+            if eos_indices.dim() > 0:
+                for i in range(eos_indices.shape[0] - 1, -1, -1):
+                    # Indices of the EOS symbol for both variables
+                    # with b*k as the first dimension, and b, k for
+                    # the first two dimensions
+                    idx = eos_indices[i]
+                    b_idx = int(idx[0] / beam_width)
+                    # The indices of the replacing position
+                    # according to the replacement strategy noted above
+                    res_k_idx = beam_width - (batch_eos_found[b_idx] %
+                                              beam_width) - 1
+                    batch_eos_found[b_idx] += 1
+                    res_idx = b_idx * beam_width + res_k_idx
+
+                    # Replace the old information in return variables
+                    # with the new ended sequence information
+                    t_predecessors[res_idx] = stored_predecessors[t][idx[0]]
+                    current_symbol[res_idx] = stored_emitted_symbols[t][idx[0]]
+                    s[b_idx, res_k_idx] = stored_scores[t][idx[0], 0]
+                    l[b_idx][res_k_idx] = t + 1
+
+            # record the back tracked results
+            p.append(current_symbol)
+            t -= 1
+
+        # Sort and re-order again as the added ended sequences may change
+        # the order (very unlikely)
+        s, re_sorted_idx = s.topk(beam_width)
+        for b_idx in range(batch_size):
+            l[b_idx] = [
+                l[b_idx][k_idx.item()] for k_idx in re_sorted_idx[b_idx, :]
+            ]
+
+        re_sorted_idx = paddle.reshape(
+            re_sorted_idx + pos_index.expand_as(re_sorted_idx),
+            [batch_size * beam_width])
+
+        # Reverse the sequences and re-order at the same time
+        # It is reversed because the backtracking happens in reverse time order
+        p = [
+            paddle.reshape(
+                paddle.index_select(step, re_sorted_idx, 0),
+                shape=[batch_size, beam_width, -1]) for step in reversed(p)
+        ]
+        p = paddle.concat(p, -1)[:, 0, :]
+        return p, paddle.ones_like(p)
+
+
+class AttentionUnit(nn.Layer):
+    def __init__(self, sDim, xDim, attDim):
+        super(AttentionUnit, self).__init__()
+
+        self.sDim = sDim
+        self.xDim = xDim
+        self.attDim = attDim
+
+        self.sEmbed = nn.Linear(sDim, attDim)
+        self.xEmbed = nn.Linear(xDim, attDim)
+        self.wEmbed = nn.Linear(attDim, 1)
+
+    def forward(self, x, sPrev):
+        batch_size, T, _ = x.shape  # [b x T x xDim]
+        x = paddle.reshape(x, [-1, self.xDim])  # [(b x T) x xDim]
+        xProj = self.xEmbed(x)  # [(b x T) x attDim]
+        xProj = paddle.reshape(xProj, [batch_size, T, -1])  # [b x T x attDim]
+
+        sPrev = sPrev.squeeze(0)
+        sProj = self.sEmbed(sPrev)  # [b x attDim]
+        sProj = paddle.unsqueeze(sProj, 1)  # [b x 1 x attDim]
+        sProj = paddle.expand(sProj,
+                              [batch_size, T, self.attDim])  # [b x T x attDim]
+
+        sumTanh = paddle.tanh(sProj + xProj)
+        sumTanh = paddle.reshape(sumTanh, [-1, self.attDim])
+
+        vProj = self.wEmbed(sumTanh)  # [(b x T) x 1]
+        vProj = paddle.reshape(vProj, [batch_size, T])
+        alpha = F.softmax(
+            vProj, axis=1)  # attention weights for each sample in the minibatch
+        return alpha
+
+
+class DecoderUnit(nn.Layer):
+    def __init__(self, sDim, xDim, yDim, attDim):
+        super(DecoderUnit, self).__init__()
+        self.sDim = sDim
+        self.xDim = xDim
+        self.yDim = yDim
+        self.attDim = attDim
+        self.emdDim = attDim
+
+        self.attention_unit = AttentionUnit(sDim, xDim, attDim)
+        self.tgt_embedding = nn.Embedding(
+            yDim + 1, self.emdDim, weight_attr=nn.initializer.Normal(
+                std=0.01))  # the last is used for <BOS>
+        self.gru = nn.GRUCell(input_size=xDim + self.emdDim, hidden_size=sDim)
+        self.fc = nn.Linear(
+            sDim,
+            yDim,
+            weight_attr=nn.initializer.Normal(std=0.01),
+            bias_attr=nn.initializer.Constant(value=0))
+        self.embed_fc = nn.Linear(300, self.sDim)
+
+    def get_initial_state(self, embed, tile_times=1):
+        assert embed.shape[1] == 300
+        state = self.embed_fc(embed)  # N * sDim
+        if tile_times != 1:
+            state = state.unsqueeze(1)
+            trans_state = paddle.transpose(state, perm=[1, 0, 2])
+            state = paddle.tile(trans_state, repeat_times=[tile_times, 1, 1])
+            trans_state = paddle.transpose(state, perm=[1, 0, 2])
+            state = paddle.reshape(trans_state, shape=[-1, self.sDim])
+        state = state.unsqueeze(0)  # 1 * N * sDim
+        return state
+
+    def forward(self, x, sPrev, yPrev):
+        # x: feature sequence from the image decoder.
+        batch_size, T, _ = x.shape
+        alpha = self.attention_unit(x, sPrev)
+        context = paddle.squeeze(paddle.matmul(alpha.unsqueeze(1), x), axis=1)
+        yPrev = paddle.cast(yPrev, dtype="int64")
+        yProj = self.tgt_embedding(yPrev)
+
+        concat_context = paddle.concat([yProj, context], 1)
+        concat_context = paddle.squeeze(concat_context, 1)
+        sPrev = paddle.squeeze(sPrev, 0)
+        output, state = self.gru(concat_context, sPrev)
+        output = paddle.squeeze(output, axis=1)
+        output = self.fc(output)
+        return output, state
\ No newline at end of file

ppocr/modeling/heads/rec_ctc_head.py

@@ -38,6 +38,7 @@ class CTCHead(nn.Layer):
                  out_channels,
                  fc_decay=0.0004,
                  mid_channels=None,
+                 return_feats=False,
                  **kwargs):
         super(CTCHead, self).__init__()
         if mid_channels is None:
@@ -66,14 +67,22 @@ class CTCHead(nn.Layer):
                 bias_attr=bias_attr2)
         self.out_channels = out_channels
         self.mid_channels = mid_channels
+        self.return_feats = return_feats
 
     def forward(self, x, targets=None):
         if self.mid_channels is None:
             predicts = self.fc(x)
         else:
-            predicts = self.fc1(x)
-            predicts = self.fc2(predicts)
-            
+            x = self.fc1(x)
+            predicts = self.fc2(x)
+
+        if self.return_feats:
+            result = (x, predicts)
+        else:
+            result = predicts
+
         if not self.training:
             predicts = F.softmax(predicts, axis=2)
-        return predicts
+            result = predicts
+
+        return result

ppocr/modeling/transforms/__init__.py

@@ -17,8 +17,9 @@ __all__ = ['build_transform']
 
 def build_transform(config):
     from .tps import TPS
+    from .stn import STN_ON
 
-    support_dict = ['TPS']
+    support_dict = ['TPS', 'STN_ON']
 
     module_name = config.pop('name')
     assert module_name in support_dict, Exception(

ppocr/modeling/transforms/stn.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 math
+import paddle
+from paddle import nn, ParamAttr
+from paddle.nn import functional as F
+import numpy as np
+
+from .tps_spatial_transformer import TPSSpatialTransformer
+
+
+def conv3x3_block(in_channels, out_channels, stride=1):
+    n = 3 * 3 * out_channels
+    w = math.sqrt(2. / n)
+    conv_layer = nn.Conv2D(
+        in_channels,
+        out_channels,
+        kernel_size=3,
+        stride=stride,
+        padding=1,
+        weight_attr=nn.initializer.Normal(
+            mean=0.0, std=w),
+        bias_attr=nn.initializer.Constant(0))
+    block = nn.Sequential(conv_layer, nn.BatchNorm2D(out_channels), nn.ReLU())
+    return block
+
+
+class STN(nn.Layer):
+    def __init__(self, in_channels, num_ctrlpoints, activation='none'):
+        super(STN, self).__init__()
+        self.in_channels = in_channels
+        self.num_ctrlpoints = num_ctrlpoints
+        self.activation = activation
+        self.stn_convnet = nn.Sequential(
+            conv3x3_block(in_channels, 32),  #32x64
+            nn.MaxPool2D(
+                kernel_size=2, stride=2),
+            conv3x3_block(32, 64),  #16x32
+            nn.MaxPool2D(
+                kernel_size=2, stride=2),
+            conv3x3_block(64, 128),  # 8*16
+            nn.MaxPool2D(
+                kernel_size=2, stride=2),
+            conv3x3_block(128, 256),  # 4*8
+            nn.MaxPool2D(
+                kernel_size=2, stride=2),
+            conv3x3_block(256, 256),  # 2*4,
+            nn.MaxPool2D(
+                kernel_size=2, stride=2),
+            conv3x3_block(256, 256))  # 1*2
+        self.stn_fc1 = nn.Sequential(
+            nn.Linear(
+                2 * 256,
+                512,
+                weight_attr=nn.initializer.Normal(0, 0.001),
+                bias_attr=nn.initializer.Constant(0)),
+            nn.BatchNorm1D(512),
+            nn.ReLU())
+        fc2_bias = self.init_stn()
+        self.stn_fc2 = nn.Linear(
+            512,
+            num_ctrlpoints * 2,
+            weight_attr=nn.initializer.Constant(0.0),
+            bias_attr=nn.initializer.Assign(fc2_bias))
+
+    def init_stn(self):
+        margin = 0.01
+        sampling_num_per_side = int(self.num_ctrlpoints / 2)
+        ctrl_pts_x = np.linspace(margin, 1. - margin, sampling_num_per_side)
+        ctrl_pts_y_top = np.ones(sampling_num_per_side) * margin
+        ctrl_pts_y_bottom = np.ones(sampling_num_per_side) * (1 - margin)
+        ctrl_pts_top = np.stack([ctrl_pts_x, ctrl_pts_y_top], axis=1)
+        ctrl_pts_bottom = np.stack([ctrl_pts_x, ctrl_pts_y_bottom], axis=1)
+        ctrl_points = np.concatenate(
+            [ctrl_pts_top, ctrl_pts_bottom], axis=0).astype(np.float32)
+        if self.activation == 'none':
+            pass
+        elif self.activation == 'sigmoid':
+            ctrl_points = -np.log(1. / ctrl_points - 1.)
+        ctrl_points = paddle.to_tensor(ctrl_points)
+        fc2_bias = paddle.reshape(
+            ctrl_points, shape=[ctrl_points.shape[0] * ctrl_points.shape[1]])
+        return fc2_bias
+
+    def forward(self, x):
+        x = self.stn_convnet(x)
+        batch_size, _, h, w = x.shape
+        x = paddle.reshape(x, shape=(batch_size, -1))
+        img_feat = self.stn_fc1(x)
+        x = self.stn_fc2(0.1 * img_feat)
+        if self.activation == 'sigmoid':
+            x = F.sigmoid(x)
+        x = paddle.reshape(x, shape=[-1, self.num_ctrlpoints, 2])
+        return img_feat, x
+
+
+class STN_ON(nn.Layer):
+    def __init__(self, in_channels, tps_inputsize, tps_outputsize,
+                 num_control_points, tps_margins, stn_activation):
+        super(STN_ON, self).__init__()
+        self.tps = TPSSpatialTransformer(
+            output_image_size=tuple(tps_outputsize),
+            num_control_points=num_control_points,
+            margins=tuple(tps_margins))
+        self.stn_head = STN(in_channels=in_channels,
+                            num_ctrlpoints=num_control_points,
+                            activation=stn_activation)
+        self.tps_inputsize = tps_inputsize
+        self.out_channels = in_channels
+
+    def forward(self, image):
+        stn_input = paddle.nn.functional.interpolate(
+            image, self.tps_inputsize, mode="bilinear", align_corners=True)
+        stn_img_feat, ctrl_points = self.stn_head(stn_input)
+        x, _ = self.tps(image, ctrl_points)
+        return x

ppocr/modeling/transforms/tps.py

@@ -231,7 +231,8 @@ class GridGenerator(nn.Layer):
         """ Return inv_delta_C which is needed to calculate T """
         F = self.F
         hat_eye = paddle.eye(F, dtype='float64')  # F x F
-        hat_C = paddle.norm(C.reshape([1, F, 2]) - C.reshape([F, 1, 2]), axis=2) + hat_eye
+        hat_C = paddle.norm(
+            C.reshape([1, F, 2]) - C.reshape([F, 1, 2]), axis=2) + hat_eye
         hat_C = (hat_C**2) * paddle.log(hat_C)
         delta_C = paddle.concat(  # F+3 x F+3
             [

ppocr/modeling/transforms/tps_spatial_transformer.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 math
+import paddle
+from paddle import nn, ParamAttr
+from paddle.nn import functional as F
+import numpy as np
+import itertools
+
+
+def grid_sample(input, grid, canvas=None):
+    input.stop_gradient = False
+    output = F.grid_sample(input, grid)
+    if canvas is None:
+        return output
+    else:
+        input_mask = paddle.ones(shape=input.shape)
+        output_mask = F.grid_sample(input_mask, grid)
+        padded_output = output * output_mask + canvas * (1 - output_mask)
+        return padded_output
+
+
+# phi(x1, x2) = r^2 * log(r), where r = ||x1 - x2||_2
+def compute_partial_repr(input_points, control_points):
+    N = input_points.shape[0]
+    M = control_points.shape[0]
+    pairwise_diff = paddle.reshape(
+        input_points, shape=[N, 1, 2]) - paddle.reshape(
+            control_points, shape=[1, M, 2])
+    # original implementation, very slow
+    # pairwise_dist = torch.sum(pairwise_diff ** 2, dim = 2) # square of distance
+    pairwise_diff_square = pairwise_diff * pairwise_diff
+    pairwise_dist = pairwise_diff_square[:, :, 0] + pairwise_diff_square[:, :,
+                                                                         1]
+    repr_matrix = 0.5 * pairwise_dist * paddle.log(pairwise_dist)
+    # fix numerical error for 0 * log(0), substitute all nan with 0
+    mask = repr_matrix != repr_matrix
+    repr_matrix[mask] = 0
+    return repr_matrix
+
+
+# output_ctrl_pts are specified, according to our task.
+def build_output_control_points(num_control_points, margins):
+    margin_x, margin_y = margins
+    num_ctrl_pts_per_side = num_control_points // 2
+    ctrl_pts_x = np.linspace(margin_x, 1.0 - margin_x, num_ctrl_pts_per_side)
+    ctrl_pts_y_top = np.ones(num_ctrl_pts_per_side) * margin_y
+    ctrl_pts_y_bottom = np.ones(num_ctrl_pts_per_side) * (1.0 - margin_y)
+    ctrl_pts_top = np.stack([ctrl_pts_x, ctrl_pts_y_top], axis=1)
+    ctrl_pts_bottom = np.stack([ctrl_pts_x, ctrl_pts_y_bottom], axis=1)
+    output_ctrl_pts_arr = np.concatenate(
+        [ctrl_pts_top, ctrl_pts_bottom], axis=0)
+    output_ctrl_pts = paddle.to_tensor(output_ctrl_pts_arr)
+    return output_ctrl_pts
+
+
+class TPSSpatialTransformer(nn.Layer):
+    def __init__(self,
+                 output_image_size=None,
+                 num_control_points=None,
+                 margins=None):
+        super(TPSSpatialTransformer, self).__init__()
+        self.output_image_size = output_image_size
+        self.num_control_points = num_control_points
+        self.margins = margins
+
+        self.target_height, self.target_width = output_image_size
+        target_control_points = build_output_control_points(num_control_points,
+                                                            margins)
+        N = num_control_points
+
+        # create padded kernel matrix
+        forward_kernel = paddle.zeros(shape=[N + 3, N + 3])
+        target_control_partial_repr = compute_partial_repr(
+            target_control_points, target_control_points)
+        target_control_partial_repr = paddle.cast(target_control_partial_repr,
+                                                  forward_kernel.dtype)
+        forward_kernel[:N, :N] = target_control_partial_repr
+        forward_kernel[:N, -3] = 1
+        forward_kernel[-3, :N] = 1
+        target_control_points = paddle.cast(target_control_points,
+                                            forward_kernel.dtype)
+        forward_kernel[:N, -2:] = target_control_points
+        forward_kernel[-2:, :N] = paddle.transpose(
+            target_control_points, perm=[1, 0])
+        # compute inverse matrix
+        inverse_kernel = paddle.inverse(forward_kernel)
+
+        # create target cordinate matrix
+        HW = self.target_height * self.target_width
+        target_coordinate = list(
+            itertools.product(
+                range(self.target_height), range(self.target_width)))
+        target_coordinate = paddle.to_tensor(target_coordinate)  # HW x 2
+        Y, X = paddle.split(
+            target_coordinate, target_coordinate.shape[1], axis=1)
+        Y = Y / (self.target_height - 1)
+        X = X / (self.target_width - 1)
+        target_coordinate = paddle.concat(
+            [X, Y], axis=1)  # convert from (y, x) to (x, y)
+        target_coordinate_partial_repr = compute_partial_repr(
+            target_coordinate, target_control_points)
+        target_coordinate_repr = paddle.concat(
+            [
+                target_coordinate_partial_repr, paddle.ones(shape=[HW, 1]),
+                target_coordinate
+            ],
+            axis=1)
+
+        # register precomputed matrices
+        self.inverse_kernel = inverse_kernel
+        self.padding_matrix = paddle.zeros(shape=[3, 2])
+        self.target_coordinate_repr = target_coordinate_repr
+        self.target_control_points = target_control_points
+
+    def forward(self, input, source_control_points):
+        assert source_control_points.ndimension() == 3
+        assert source_control_points.shape[1] == self.num_control_points
+        assert source_control_points.shape[2] == 2
+        batch_size = paddle.shape(source_control_points)[0]
+
+        self.padding_matrix = paddle.expand(
+            self.padding_matrix, shape=[batch_size, 3, 2])
+        Y = paddle.concat([source_control_points, self.padding_matrix], 1)
+        mapping_matrix = paddle.matmul(self.inverse_kernel, Y)
+        source_coordinate = paddle.matmul(self.target_coordinate_repr,
+                                          mapping_matrix)
+
+        grid = paddle.reshape(
+            source_coordinate,
+            shape=[-1, self.target_height, self.target_width, 2])
+        grid = paddle.clip(grid, 0,
+                           1)  # the source_control_points may be out of [0, 1].
+        # the input to grid_sample is normalized [-1, 1], but what we get is [0, 1]
+        grid = 2.0 * grid - 1.0
+        output_maps = grid_sample(input, grid, canvas=None)
+        return output_maps, source_coordinate

ppocr/optimizer/optimizer.py

@@ -127,3 +127,34 @@ class RMSProp(object):
             grad_clip=self.grad_clip,
             parameters=parameters)
         return opt
+
+
+class Adadelta(object):
+    def __init__(self,
+                 learning_rate=0.001,
+                 epsilon=1e-08,
+                 rho=0.95,
+                 parameter_list=None,
+                 weight_decay=None,
+                 grad_clip=None,
+                 name=None,
+                 **kwargs):
+        self.learning_rate = learning_rate
+        self.epsilon = epsilon
+        self.rho = rho
+        self.parameter_list = parameter_list
+        self.learning_rate = learning_rate
+        self.weight_decay = weight_decay
+        self.grad_clip = grad_clip
+        self.name = name
+
+    def __call__(self, parameters):
+        opt = optim.Adadelta(
+            learning_rate=self.learning_rate,
+            epsilon=self.epsilon,
+            rho=self.rho,
+            weight_decay=self.weight_decay,
+            grad_clip=self.grad_clip,
+            name=self.name,
+            parameters=parameters)
+        return opt

ppocr/postprocess/__init__.py

@@ -18,17 +18,21 @@ from __future__ import print_function
 from __future__ import unicode_literals
 
 import copy
+import platform
 
 __all__ = ['build_post_process']
 
 from .db_postprocess import DBPostProcess, DistillationDBPostProcess
 from .east_postprocess import EASTPostProcess
 from .sast_postprocess import SASTPostProcess
-from .rec_postprocess import CTCLabelDecode, AttnLabelDecode, SRNLabelDecode, DistillationCTCLabelDecode, NRTRLabelDecode, \
-    TableLabelDecode, SARLabelDecode
+from .rec_postprocess import CTCLabelDecode, AttnLabelDecode, SRNLabelDecode, DistillationCTCLabelDecode, \
+    TableLabelDecode, NRTRLabelDecode, SARLabelDecode , SEEDLabelDecode
 from .cls_postprocess import ClsPostProcess
 from .pg_postprocess import PGPostProcess
-from .pse_postprocess import PSEPostProcess
+
+if platform.system() != "Windows":
+    # pse is not support in Windows
+    from .pse_postprocess import PSEPostProcess
 
 
 def build_post_process(config, global_config=None):
@@ -36,7 +40,8 @@ def build_post_process(config, global_config=None):
         'DBPostProcess', 'PSEPostProcess', 'EASTPostProcess', 'SASTPostProcess',
         'CTCLabelDecode', 'AttnLabelDecode', 'ClsPostProcess', 'SRNLabelDecode',
         'PGPostProcess', 'DistillationCTCLabelDecode', 'TableLabelDecode',
-        'DistillationDBPostProcess', 'NRTRLabelDecode', 'SARLabelDecode'
+        'DistillationDBPostProcess', 'NRTRLabelDecode', 'SARLabelDecode',
+        'SEEDLabelDecode'
     ]
 
     config = copy.deepcopy(config)

ppocr/postprocess/rec_postprocess.py

@@ -111,6 +111,8 @@ class CTCLabelDecode(BaseRecLabelDecode):
                                              character_type, use_space_char)
 
     def __call__(self, preds, label=None, *args, **kwargs):
+        if isinstance(preds, tuple):
+            preds = preds[-1]
         if isinstance(preds, paddle.Tensor):
             preds = preds.numpy()
         preds_idx = preds.argmax(axis=2)
@@ -308,6 +310,87 @@ class AttnLabelDecode(BaseRecLabelDecode):
         return idx
 
 
+class SEEDLabelDecode(BaseRecLabelDecode):
+    """ Convert between text-label and text-index """
+
+    def __init__(self,
+                 character_dict_path=None,
+                 character_type='ch',
+                 use_space_char=False,
+                 **kwargs):
+        super(SEEDLabelDecode, self).__init__(character_dict_path,
+                                              character_type, use_space_char)
+
+    def add_special_char(self, dict_character):
+        self.beg_str = "sos"
+        self.end_str = "eos"
+        dict_character = dict_character + [self.end_str]
+        return dict_character
+
+    def get_ignored_tokens(self):
+        end_idx = self.get_beg_end_flag_idx("eos")
+        return [end_idx]
+
+    def get_beg_end_flag_idx(self, beg_or_end):
+        if beg_or_end == "sos":
+            idx = np.array(self.dict[self.beg_str])
+        elif beg_or_end == "eos":
+            idx = np.array(self.dict[self.end_str])
+        else:
+            assert False, "unsupport type %s in get_beg_end_flag_idx" % beg_or_end
+        return idx
+
+    def decode(self, text_index, text_prob=None, is_remove_duplicate=False):
+        """ convert text-index into text-label. """
+        result_list = []
+        [end_idx] = self.get_ignored_tokens()
+        batch_size = len(text_index)
+        for batch_idx in range(batch_size):
+            char_list = []
+            conf_list = []
+            for idx in range(len(text_index[batch_idx])):
+                if int(text_index[batch_idx][idx]) == int(end_idx):
+                    break
+                if is_remove_duplicate:
+                    # only for predict
+                    if idx > 0 and text_index[batch_idx][idx - 1] == text_index[
+                            batch_idx][idx]:
+                        continue
+                char_list.append(self.character[int(text_index[batch_idx][
+                    idx])])
+                if text_prob is not None:
+                    conf_list.append(text_prob[batch_idx][idx])
+                else:
+                    conf_list.append(1)
+            text = ''.join(char_list)
+            result_list.append((text, np.mean(conf_list)))
+        return result_list
+
+    def __call__(self, preds, label=None, *args, **kwargs):
+        """
+        text = self.decode(text)
+        if label is None:
+            return text
+        else:
+            label = self.decode(label, is_remove_duplicate=False)
+            return text, label
+        """
+        preds_idx = preds["rec_pred"]
+        if isinstance(preds_idx, paddle.Tensor):
+            preds_idx = preds_idx.numpy()
+        if "rec_pred_scores" in preds:
+            preds_idx = preds["rec_pred"]
+            preds_prob = preds["rec_pred_scores"]
+        else:
+            preds_idx = preds["rec_pred"].argmax(axis=2)
+            preds_prob = preds["rec_pred"].max(axis=2)
+        text = self.decode(preds_idx, preds_prob, is_remove_duplicate=False)
+        if label is None:
+            return text
+        label = self.decode(label, is_remove_duplicate=False)
+        return text, label
+
+
 class SRNLabelDecode(BaseRecLabelDecode):
     """ Convert between text-label and text-index """
 

requirements.txt

@@ -11,4 +11,5 @@ opencv-contrib-python==4.4.0.46
 cython
 lxml
 premailer
-openpyxl
\ No newline at end of file
+openpyxl
+fasttext==0.9.1
\ No newline at end of file

tools/eval.py

@@ -54,8 +54,7 @@ def main():
             config['Architecture']["Head"]['out_channels'] = char_num
 
     model = build_model(config['Architecture'])
-    use_srn = config['Architecture']['algorithm'] == "SRN"
-    use_sar = config['Architecture']['algorithm'] == "SAR"
+    extra_input = config['Architecture']['algorithm'] in ["SRN", "SAR"]
     if "model_type" in config['Architecture'].keys():
         model_type = config['Architecture']['model_type']
     else:
@@ -72,7 +71,7 @@ def main():
 
     # start eval
     metric = program.eval(model, valid_dataloader, post_process_class,
-                        eval_class, model_type, use_srn, use_sar)
+                          eval_class, model_type, extra_input)
     logger.info('metric eval ***************')
     for k, v in metric.items():
         logger.info('{}:{}'.format(k, v))

tools/export_model.py

@@ -49,6 +49,12 @@ def export_single_model(model, arch_config, save_path, logger):
                 ]
         ]
         model = to_static(model, input_spec=other_shape)
+    elif arch_config["algorithm"] == "SAR":
+        other_shape = [
+            paddle.static.InputSpec(
+                shape=[None, 3, 48, 160], dtype="float32"),
+        ]
+        model = to_static(model, input_spec=other_shape)
     else:
         infer_shape = [3, -1, -1]
         if arch_config["model_type"] == "rec":

tools/infer/predict_rec.py

@@ -68,6 +68,13 @@ class TextRecognizer(object):
                 "character_dict_path": args.rec_char_dict_path,
                 "use_space_char": args.use_space_char
             }
+        elif self.rec_algorithm == "SAR":
+            postprocess_params = {
+                'name': 'SARLabelDecode',
+                "character_type": args.rec_char_type,
+                "character_dict_path": args.rec_char_dict_path,
+                "use_space_char": args.use_space_char
+            }
         self.postprocess_op = build_post_process(postprocess_params)
         self.predictor, self.input_tensor, self.output_tensors, self.config = \
             utility.create_predictor(args, 'rec', logger)
@@ -194,6 +201,41 @@ class TextRecognizer(object):
         return (norm_img, encoder_word_pos, gsrm_word_pos, gsrm_slf_attn_bias1,
                 gsrm_slf_attn_bias2)
 
+    def resize_norm_img_sar(self, img, image_shape,
+                            width_downsample_ratio=0.25):
+        imgC, imgH, imgW_min, imgW_max = image_shape
+        h = img.shape[0]
+        w = img.shape[1]
+        valid_ratio = 1.0
+        # make sure new_width is an integral multiple of width_divisor.
+        width_divisor = int(1 / width_downsample_ratio)
+        # resize
+        ratio = w / float(h)
+        resize_w = math.ceil(imgH * ratio)
+        if resize_w % width_divisor != 0:
+            resize_w = round(resize_w / width_divisor) * width_divisor
+        if imgW_min is not None:
+            resize_w = max(imgW_min, resize_w)
+        if imgW_max is not None:
+            valid_ratio = min(1.0, 1.0 * resize_w / imgW_max)
+            resize_w = min(imgW_max, resize_w)
+        resized_image = cv2.resize(img, (resize_w, imgH))
+        resized_image = resized_image.astype('float32')
+        # norm 
+        if image_shape[0] == 1:
+            resized_image = resized_image / 255
+            resized_image = resized_image[np.newaxis, :]
+        else:
+            resized_image = resized_image.transpose((2, 0, 1)) / 255
+        resized_image -= 0.5
+        resized_image /= 0.5
+        resize_shape = resized_image.shape
+        padding_im = -1.0 * np.ones((imgC, imgH, imgW_max), dtype=np.float32)
+        padding_im[:, :, 0:resize_w] = resized_image
+        pad_shape = padding_im.shape
+
+        return padding_im, resize_shape, pad_shape, valid_ratio
+
     def __call__(self, img_list):
         img_num = len(img_list)
         # Calculate the aspect ratio of all text bars
@@ -216,11 +258,19 @@ class TextRecognizer(object):
                 wh_ratio = w * 1.0 / h
                 max_wh_ratio = max(max_wh_ratio, wh_ratio)
             for ino in range(beg_img_no, end_img_no):
-                if self.rec_algorithm != "SRN":
+                if self.rec_algorithm != "SRN" and self.rec_algorithm != "SAR":
                     norm_img = self.resize_norm_img(img_list[indices[ino]],
                                                     max_wh_ratio)
                     norm_img = norm_img[np.newaxis, :]
                     norm_img_batch.append(norm_img)
+                elif self.rec_algorithm == "SAR":
+                    norm_img, _, _, valid_ratio = self.resize_norm_img_sar(
+                        img_list[indices[ino]], self.rec_image_shape)
+                    norm_img = norm_img[np.newaxis, :]
+                    valid_ratio = np.expand_dims(valid_ratio, axis=0)
+                    valid_ratios = []
+                    valid_ratios.append(valid_ratio)
+                    norm_img_batch.append(norm_img)
                 else:
                     norm_img = self.process_image_srn(
                         img_list[indices[ino]], self.rec_image_shape, 8, 25)
@@ -266,6 +316,25 @@ class TextRecognizer(object):
                 if self.benchmark:
                     self.autolog.times.stamp()
                 preds = {"predict": outputs[2]}
+            elif self.rec_algorithm == "SAR":
+                valid_ratios = np.concatenate(valid_ratios)
+                inputs = [
+                    norm_img_batch,
+                    valid_ratios,
+                ]
+                input_names = self.predictor.get_input_names()
+                for i in range(len(input_names)):
+                    input_tensor = self.predictor.get_input_handle(input_names[
+                        i])
+                    input_tensor.copy_from_cpu(inputs[i])
+                self.predictor.run()
+                outputs = []
+                for output_tensor in self.output_tensors:
+                    output = output_tensor.copy_to_cpu()
+                    outputs.append(output)
+                if self.benchmark:
+                    self.autolog.times.stamp()
+                preds = outputs[0]
             else:
                 self.input_tensor.copy_from_cpu(norm_img_batch)
                 self.predictor.run()

tools/program.py

@@ -186,12 +186,13 @@ def train(config,
     model.train()
 
     use_srn = config['Architecture']['algorithm'] == "SRN"
-    use_nrtr = config['Architecture']['algorithm'] == "NRTR"
-    use_sar = config['Architecture']['algorithm'] == 'SAR'
+    extra_input = config['Architecture'][
+        'algorithm'] in ["SRN", "NRTR", "SAR", "SEED"]
     try:
         model_type = config['Architecture']['model_type']
     except:
         model_type = None
+    algorithm = config['Architecture']['algorithm']
 
     if 'start_epoch' in best_model_dict:
         start_epoch = best_model_dict['start_epoch']
@@ -215,7 +216,7 @@ def train(config,
             images = batch[0]
             if use_srn:
                 model_average = True
-            if use_srn or model_type == 'table' or use_nrtr or use_sar:
+            if model_type == 'table' or extra_input:
                 preds = model(images, data=batch[1:])
             else:
                 preds = model(images)
@@ -279,8 +280,7 @@ def train(config,
                     post_process_class,
                     eval_class,
                     model_type,
-                    use_srn=use_srn,
-                    use_sar=use_sar)
+                    extra_input=extra_input)
                 cur_metric_str = 'cur metric, {}'.format(', '.join(
                     ['{}: {}'.format(k, v) for k, v in cur_metric.items()]))
                 logger.info(cur_metric_str)
@@ -351,9 +351,8 @@ def eval(model,
          valid_dataloader,
          post_process_class,
          eval_class,
-         model_type,
-         use_srn=False,
-         use_sar=False):
+         model_type=None,
+         extra_input=False):
     model.eval()
     with paddle.no_grad():
         total_frame = 0.0
@@ -366,7 +365,7 @@ def eval(model,
                 break
             images = batch[0]
             start = time.time()
-            if use_srn or model_type == 'table' or use_sar:
+            if model_type == 'table' or extra_input:
                 preds = model(images, data=batch[1:])
             else:
                 preds = model(images)
@@ -395,6 +394,18 @@ def preprocess(is_train=False):
     config = load_config(FLAGS.config)
     merge_config(FLAGS.opt)
 
+    if is_train:
+        # save_config
+        save_model_dir = config['Global']['save_model_dir']
+        os.makedirs(save_model_dir, exist_ok=True)
+        with open(os.path.join(save_model_dir, 'config.yml'), 'w') as f:
+            yaml.dump(
+                dict(config), f, default_flow_style=False, sort_keys=False)
+        log_file = '{}/train.log'.format(save_model_dir)
+    else:
+        log_file = None
+    logger = get_logger(name='root', log_file=log_file)
+
     # check if set use_gpu=True in paddlepaddle cpu version
     use_gpu = config['Global']['use_gpu']
     check_gpu(use_gpu)
@@ -402,24 +413,19 @@ def preprocess(is_train=False):
     alg = config['Architecture']['algorithm']
     assert alg in [
         'EAST', 'DB', 'SAST', 'Rosetta', 'CRNN', 'STARNet', 'RARE', 'SRN',
-        'CLS', 'PGNet', 'Distillation', 'NRTR', 'TableAttn', 'SAR', 'PSE'
-    ]
+        'CLS', 'PGNet', 'Distillation', 'NRTR', 'TableAttn', 'SAR', 'PSE',
+        'SEED']
+    windows_not_support_list = ['PSE']
+    if platform.system() == "Windows" and alg in windows_not_support_list:
+        logger.warning('{} is not support in Windows now'.format(
+            windows_not_support_list))
+        sys.exit()
 
     device = 'gpu:{}'.format(dist.ParallelEnv().dev_id) if use_gpu else 'cpu'
     device = paddle.set_device(device)
 
     config['Global']['distributed'] = dist.get_world_size() != 1
-    if is_train:
-        # save_config
-        save_model_dir = config['Global']['save_model_dir']
-        os.makedirs(save_model_dir, exist_ok=True)
-        with open(os.path.join(save_model_dir, 'config.yml'), 'w') as f:
-            yaml.dump(
-                dict(config), f, default_flow_style=False, sort_keys=False)
-        log_file = '{}/train.log'.format(save_model_dir)
-    else:
-        log_file = None
-    logger = get_logger(name='root', log_file=log_file)
+
     if config['Global']['use_visualdl']:
         from visualdl import LogWriter
         save_model_dir = config['Global']['save_model_dir']