Merge pull request #1139 from WenmuZhou/dygraph_rc

新增功能

ppocr/modeling/backbones/det_resnet_vd.py

@@ -19,6 +19,7 @@ from __future__ import print_function
 import paddle
 from paddle import ParamAttr
 import paddle.nn as nn
+import paddle.nn.functional as F
 
 __all__ = ["ResNet"]
 
@@ -37,9 +38,9 @@ class ConvBNLayer(nn.Layer):
         super(ConvBNLayer, self).__init__()
 
         self.is_vd_mode = is_vd_mode
-        self._pool2d_avg = nn.AvgPool2d(
+        self._pool2d_avg = nn.AvgPool2D(
             kernel_size=2, stride=2, padding=0, ceil_mode=True)
-        self._conv = nn.Conv2d(
+        self._conv = nn.Conv2D(
             in_channels=in_channels,
             out_channels=out_channels,
             kernel_size=kernel_size,
@@ -118,7 +119,8 @@ class BottleneckBlock(nn.Layer):
             short = inputs
         else:
             short = self.short(inputs)
-        y = paddle.elementwise_add(x=short, y=conv2, act='relu')
+        y = paddle.add(x=short, y=conv2)
+        y = F.relu(y)
         return y
 
 
@@ -165,7 +167,8 @@ class BasicBlock(nn.Layer):
             short = inputs
         else:
             short = self.short(inputs)
-        y = paddle.elementwise_add(x=short, y=conv1, act='relu')
+        y = paddle.add(x=short, y=conv1)
+        y = F.relu(y)
         return y
 
 
@@ -214,7 +217,7 @@ class ResNet(nn.Layer):
             stride=1,
             act='relu',
             name="conv1_3")
-        self.pool2d_max = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
+        self.pool2d_max = nn.MaxPool2D(kernel_size=3, stride=2, padding=1)
 
         self.stages = []
         self.out_channels = []

ppocr/modeling/backbones/rec_resnet_vd.py

@@ -19,6 +19,7 @@ from __future__ import print_function
 import paddle
 from paddle import ParamAttr
 import paddle.nn as nn
+import paddle.nn.functional as F
 
 __all__ = ["ResNet"]
 
@@ -37,9 +38,9 @@ class ConvBNLayer(nn.Layer):
         super(ConvBNLayer, self).__init__()
 
         self.is_vd_mode = is_vd_mode
-        self._pool2d_avg = nn.AvgPool2d(
+        self._pool2d_avg = nn.AvgPool2D(
             kernel_size=stride, stride=stride, padding=0, ceil_mode=True)
-        self._conv = nn.Conv2d(
+        self._conv = nn.Conv2D(
             in_channels=in_channels,
             out_channels=out_channels,
             kernel_size=kernel_size,
@@ -119,7 +120,8 @@ class BottleneckBlock(nn.Layer):
             short = inputs
         else:
             short = self.short(inputs)
-        y = paddle.elementwise_add(x=short, y=conv2, act='relu')
+        y = paddle.add(x=short, y=conv2)
+        y = F.relu(y)
         return y
 
 
@@ -166,7 +168,8 @@ class BasicBlock(nn.Layer):
             short = inputs
         else:
             short = self.short(inputs)
-        y = paddle.elementwise_add(x=short, y=conv1, act='relu')
+        y = paddle.add(x=short, y=conv1)
+        y = F.relu(y)
         return y
 
 
@@ -215,7 +218,7 @@ class ResNet(nn.Layer):
             stride=1,
             act='relu',
             name="conv1_3")
-        self.pool2d_max = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
+        self.pool2d_max = nn.MaxPool2D(kernel_size=3, stride=2, padding=1)
 
         self.block_list = []
         if layers >= 50:
@@ -270,7 +273,7 @@ class ResNet(nn.Layer):
                     shortcut = True
                     self.block_list.append(basic_block)
                 self.out_channels = num_filters[block]
-        self.out_pool = nn.MaxPool2d(kernel_size=2, stride=2, padding=0)
+        self.out_pool = nn.MaxPool2D(kernel_size=2, stride=2, padding=0)
 
     def forward(self, inputs):
         y = self.conv1_1(inputs)

ppocr/postprocess/db_postprocess.py

@@ -18,6 +18,7 @@ from __future__ import print_function
 
 import numpy as np
 import cv2
+import paddle
 from shapely.geometry import Polygon
 import pyclipper
 
@@ -130,7 +131,9 @@ class DBPostProcess(object):
         return cv2.mean(bitmap[ymin:ymax + 1, xmin:xmax + 1], mask)[0]
 
     def __call__(self, pred, shape_list):
-        pred = pred.numpy()[:, 0, :, :]
+        if isinstance(pred, paddle.Tensor):
+            pred = pred.numpy()
+        pred = pred[:, 0, :, :]
         segmentation = pred > self.thresh
 
         boxes_batch = []

ppocr/postprocess/db_postprocess_torch.py

 import cv2
+import paddle
 import numpy as np
 import pyclipper
 from shapely.geometry import Polygon
@@ -23,7 +24,9 @@ class DBPostProcess():
         pred:
             binary: text region segmentation map, with shape (N, 1,H, W)
         '''
-        pred = pred.numpy()[:, 0, :, :]
+        if isinstance(pred, paddle.Tensor):
+            pred = pred.numpy()
+        pred = pred[:, 0, :, :]
         segmentation = self.binarize(pred)
         batch_out = []
         for batch_index in range(pred.shape[0]):

ppocr/postprocess/rec_postprocess.py

@@ -100,9 +100,10 @@ class CTCLabelDecode(BaseRecLabelDecode):
                                              character_type, use_space_char)
 
     def __call__(self, preds, label=None, *args, **kwargs):
+        if isinstance(preds, paddle.Tensor):
+            preds = preds.numpy()
         # out = self.decode_preds(preds)
 
-        preds = F.softmax(preds, axis=2).numpy()
         preds_idx = preds.argmax(axis=2)
         preds_prob = preds.max(axis=2)
         text = self.decode(preds_idx, preds_prob)
@@ -116,19 +117,18 @@ class CTCLabelDecode(BaseRecLabelDecode):
         return dict_character
 
     def decode_preds(self, preds):
-        probs = F.softmax(preds, axis=2).numpy()
-        probs_ind = np.argmax(probs, axis=2)
+        probs_ind = np.argmax(preds, axis=2)
 
         B, N, _ = preds.shape
         l = np.ones(B).astype(np.int64) * N
-        length = paddle.to_variable(l)
+        length = paddle.to_tensor(l)
         out = paddle.fluid.layers.ctc_greedy_decoder(preds, 0, length)
         batch_res = [
             x[:idx[0]] for x, idx in zip(out[0].numpy(), out[1].numpy())
         ]
 
         result_list = []
-        for sample_idx, ind, prob in zip(batch_res, probs_ind, probs):
+        for sample_idx, ind, prob in zip(batch_res, probs_ind, preds):
             char_list = [self.character[idx] for idx in sample_idx]
             valid_ind = np.where(ind != 0)[0]
             if len(valid_ind) == 0:

ppocr/utils/save_load.py

@@ -68,11 +68,11 @@ def load_dygraph_pretrain(model, logger, path=None, load_static_weights=False):
                 param_state_dict[key] = pre_state_dict[weight_name]
             else:
                 param_state_dict[key] = model_dict[key]
-        model.set_dict(param_state_dict)
+        model.set_state_dict(param_state_dict)
         return
 
-    param_state_dict, optim_state_dict = paddle.load(path)
-    model.set_dict(param_state_dict)
+    param_state_dict = paddle.load(path + '.pdparams')
+    model.set_state_dict(param_state_dict)
     return
 
 
@@ -91,7 +91,7 @@ def init_model(config, model, logger, optimizer=None, lr_scheduler=None):
             "Given dir {}.pdopt not exist.".format(checkpoints)
         para_dict = paddle.load(checkpoints + '.pdparams')
         opti_dict = paddle.load(checkpoints + '.pdopt')
-        model.set_dict(para_dict)
+        model.set_state_dict(para_dict)
         if optimizer is not None:
             optimizer.set_state_dict(opti_dict)
 

tools/export_model.py

@@ -12,6 +12,13 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+import os
+import sys
+
+__dir__ = os.path.dirname(os.path.abspath(__file__))
+sys.path.append(__dir__)
+sys.path.append(os.path.abspath(os.path.join(__dir__, '..')))
+
 import argparse
 
 import paddle
@@ -20,14 +27,11 @@ from paddle.jit import to_static
 from ppocr.modeling.architectures import build_model
 from ppocr.postprocess import build_post_process
 from ppocr.utils.save_load import init_model
+from ppocr.utils.logging import get_logger
 from tools.program import load_config
-from tools.program import merge_config
 
 
 def parse_args():
-    def str2bool(v):
-        return v.lower() in ("true", "t", "1")
-
     parser = argparse.ArgumentParser()
     parser.add_argument("-c", "--config", help="configuration file to use")
     parser.add_argument(
@@ -43,7 +47,7 @@ class Model(paddle.nn.Layer):
     # Please modify the 'shape' according to actual needs
     @to_static(input_spec=[
         paddle.static.InputSpec(
-            shape=[None, 3, 32, None], dtype='float32')
+            shape=[None, 3, 640, 640], dtype='float32')
     ])
     def forward(self, inputs):
         x = self.pre_model(inputs)
@@ -53,14 +57,13 @@ class Model(paddle.nn.Layer):
 def main():
     FLAGS = parse_args()
     config = load_config(FLAGS.config)
-    merge_config(FLAGS.opt)
-
+    logger = get_logger()
     # build post process
     post_process_class = build_post_process(config['PostProcess'],
                                             config['Global'])
 
     # build model
-    #for rec algorithm
+    # for rec algorithm
     if hasattr(post_process_class, 'character'):
         char_num = len(getattr(post_process_class, 'character'))
         config['Architecture']["Head"]['out_channels'] = char_num
@@ -69,7 +72,10 @@ def main():
     model.eval()
 
     model = Model(model)
-    paddle.jit.save(model, FLAGS.output_path)
+    save_path = '{}/{}'.format(FLAGS.output_path,
+                               config['Architecture']['model_type'])
+    paddle.jit.save(model, save_path)
+    logger.info('inference model is saved to {}'.format(save_path))
 
 
 if __name__ == "__main__":

tools/infer/predict_det.py

@@ -22,7 +22,6 @@ import cv2
 import numpy as np
 import time
 import sys
-
 import paddle
 
 import tools.infer.utility as utility
@@ -39,7 +38,7 @@ class TextDetector(object):
         postprocess_params = {}
         if self.det_algorithm == "DB":
             pre_process_list = [{
-                'ResizeForTest': {
+                'DetResizeForTest': {
                     'limit_side_len': args.det_limit_side_len,
                     'limit_type': args.det_limit_type
                 }
@@ -53,7 +52,7 @@ class TextDetector(object):
             }, {
                 'ToCHWImage': None
             }, {
-                'keepKeys': {
+                'KeepKeys': {
                     'keep_keys': ['image', 'shape']
                 }
             }]
@@ -68,8 +67,9 @@ class TextDetector(object):
 
         self.preprocess_op = create_operators(pre_process_list)
         self.postprocess_op = build_post_process(postprocess_params)
-        self.predictor = paddle.jit.load(args.det_model_dir)
-        self.predictor.eval()
+        self.predictor, self.input_tensor, self.output_tensors = utility.create_predictor(
+            args, 'det', logger)  # paddle.jit.load(args.det_model_dir)
+        # self.predictor.eval()
 
     def order_points_clockwise(self, pts):
         """
@@ -133,11 +133,23 @@ class TextDetector(object):
             return None, 0
         img = np.expand_dims(img, axis=0)
         shape_list = np.expand_dims(shape_list, axis=0)
+        img = img.copy()
         starttime = time.time()
 
-        preds = self.predictor(img)
+        if self.use_zero_copy_run:
+            self.input_tensor.copy_from_cpu(img)
+            self.predictor.zero_copy_run()
+        else:
+            im = paddle.fluid.core.PaddleTensor(img)
+            self.predictor.run([im])
+        outputs = []
+        for output_tensor in self.output_tensors:
+            output = output_tensor.copy_to_cpu()
+            outputs.append(output)
+        preds = outputs[0]
+
+        # preds = self.predictor(img)
         post_result = self.postprocess_op(preds, shape_list)
-
         dt_boxes = post_result[0]['points']
         dt_boxes = self.filter_tag_det_res(dt_boxes, ori_im.shape)
         elapse = time.time() - starttime
@@ -146,8 +158,6 @@ class TextDetector(object):
 
 if __name__ == "__main__":
     args = utility.parse_args()
-    place = paddle.CPUPlace()
-    paddle.disable_static(place)
 
     image_file_list = get_image_file_list(args.image_dir)
     logger = get_logger()

tools/infer_det.py

@@ -29,12 +29,11 @@ import cv2
 import json
 import paddle
 
-from ppocr.utils.logging import get_logger
 from ppocr.data import create_operators, transform
-from ppocr.modeling import build_model
+from ppocr.modeling.architectures import build_model
 from ppocr.postprocess import build_post_process
 from ppocr.utils.save_load import init_model
-from ppocr.utils.utility import print_dict, get_image_file_list
+from ppocr.utils.utility import get_image_file_list
 import tools.program as program
 
 
@@ -67,11 +66,11 @@ def main():
 
     # create data ops
     transforms = []
-    for op in config['EVAL']['dataset']['transforms']:
+    for op in config['Eval']['dataset']['transforms']:
         op_name = list(op)[0]
         if 'Label' in op_name:
             continue
-        elif op_name == 'keepKeys':
+        elif op_name == 'KeepKeys':
             op[op_name]['keep_keys'] = ['image', 'shape']
         transforms.append(op)
 
@@ -92,8 +91,7 @@ def main():
 
             images = np.expand_dims(batch[0], axis=0)
             shape_list = np.expand_dims(batch[1], axis=0)
-            images = paddle.to_variable(images)
-            print(images.shape)
+            images = paddle.to_tensor(images)
             preds = model(images)
             post_result = post_process_class(preds, shape_list)
             boxes = post_result[0]['points']
@@ -109,14 +107,7 @@ def main():
             draw_det_res(boxes, config, src_img, file)
     logger.info("success!")
 
-    # save inference model
-    # paddle.jit.save(model, 'output/model')
-
 
 if __name__ == '__main__':
-    place, config = program.preprocess()
-    paddle.disable_static(place)
-
-    logger = get_logger()
-    print_dict(config, logger)
+    config, device, logger, vdl_writer = program.preprocess()
     main()
\ No newline at end of file

tools/infer_rec.py

@@ -27,12 +27,11 @@ sys.path.append(os.path.abspath(os.path.join(__dir__, '..')))
 
 import paddle
 
-from ppocr.utils.logging import get_logger
 from ppocr.data import create_operators, transform
-from ppocr.modeling import build_model
+from ppocr.modeling.architectures import build_model
 from ppocr.postprocess import build_post_process
 from ppocr.utils.save_load import init_model
-from ppocr.utils.utility import print_dict, get_image_file_list
+from ppocr.utils.utility import get_image_file_list
 import tools.program as program
 
 
@@ -54,13 +53,13 @@ def main():
 
     # create data ops
     transforms = []
-    for op in config['EVAL']['dataset']['transforms']:
+    for op in config['Eval']['dataset']['transforms']:
         op_name = list(op)[0]
         if 'Label' in op_name:
             continue
         elif op_name in ['RecResizeImg']:
             op[op_name]['infer_mode'] = True
-        elif op_name == 'keepKeys':
+        elif op_name == 'KeepKeys':
             op[op_name]['keep_keys'] = ['image']
         transforms.append(op)
     global_config['infer_mode'] = True
@@ -75,22 +74,14 @@ def main():
         batch = transform(data, ops)
 
         images = np.expand_dims(batch[0], axis=0)
-        images = paddle.to_variable(images)
+        images = paddle.to_tensor(images)
         preds = model(images)
         post_result = post_process_class(preds)
         for rec_reuslt in post_result:
             logger.info('\t result: {}'.format(rec_reuslt))
     logger.info("success!")
 
-    # save inference model
-    # currently, paddle.jit.to_static not support rnn
-    # paddle.jit.save(model, 'output/rec/model')
-
 
 if __name__ == '__main__':
-    place, config = program.preprocess()
-    paddle.disable_static(place)
-
-    logger = get_logger()
-    print_dict(config, logger)
+    config, device, logger, vdl_writer = program.preprocess()
     main()