add version branch

README.md

@@ -4,6 +4,10 @@
 
 ***
 最近跟新：
+- 2021.05.19 更新基于DBnet的多语种文本检测。
+- 2021.05.01 更新CRNN 训练，解决了多gpu训练问题，更换成lmdb训练，需要将图片先转成lmdb（在script文件夹中有多进程将图片转成lmdb的代码），做了一些训练优化，模型结构更改（训练时使用名字中带lmdb的yaml文件），实际训练效果如下表。
+- 2021.03.26 更新CRNN 训练效果，代码整理后上传
+- 2021.03.06 更新CRNN backbone resnet 和 mobilev3 以及配置文件
 - 2020.12.22 更新CRNN+CTCLoss+CenterLoss训练
 - 2020.09.18 更新文本检测说明文档
 - 2020.09.12 更新DB,pse,pan,sast,crnn训练测试代码和预训练模型
@@ -27,6 +31,17 @@
 - [ ] 训练通用化ocr模型
 - [ ] 结合chinese_lite进行部署
 - [ ] 手机端部署
+***
+### crnn模型效果(实验中)
+使用 MJSynth(MJ) 和 SynthText(ST) 训练，以batchsize=512训练，在以下数据集上测试：
+
+| 模型      |迭代次数|  CUTE80 |   IC03_867 |IC13_1015|IC13_857|IC15_1811|IC15_2077|IIIT5k_3000|SVT|SVTP|mean|
+|-|-|-|-|-|-|-|-|-|-|-|-|
+| resnet34+lstm+ctc |120000|  82.98|  91.92|90.93|91.59|73.10|67.98|90.16|85.16|78.29|83.56|
+| mobilev3_large+lstm+ctc | 210000| 73.61| 92.50|90.34|91.59|74.82|68.89|87.56|83.46|77.20|82.21|
+| mobilev3_small+lstm+ctc | 210000| 66.31| 90.77|88.76|91.13|73.66|69.52|88.80|84.54|72.24|80.64|
+
+
 ***
 ### 检测模型效果(实验中)
 
@@ -93,6 +108,18 @@
 <img src="./doc/show/ocr1.jpg" width=600 height=600 />     
 <img src="./doc/show/ocr2.jpg" width=600 height=600 />
 
+***
+
+### Dbnet多语种文本检测效果
+
+#### 生成数据集：
+<img src="./doc/show/2.jpg" width=600 height=600 />
+
+#### 公开数据集：
+<img src="./doc/show/1.jpg" width=600 height=600 />     
+<img src="./doc/show/3.jpg" width=500 height=600 />
+
+
 ***
 
 ### 有问题及交流加微信

config/det_DB_mobilev3.yaml

@@ -9,12 +9,12 @@ base:
   crop_shape: [640,640]
   shrink_ratio: 0.4
   n_epoch: 1200
-  start_val: 400
+  start_val: 700
   show_step: 20
   checkpoints: ./checkpoint
   save_epoch: 100
-  restore: True
-  restore_file : ./checkpoint/ag_DB_bb_mobilenet_v3_small_he_DB_Head_bs_16_ep_1200_mobile_slim_all/DB_best.pth.tar
+  restore: False
+  restore_file : ./checkpoint/DB_best.pth.tar
 
 backbone:
   function: ptocr.model.backbone.det_mobilev3,mobilenet_v3_small
@@ -82,6 +82,6 @@ postprocess:
   min_size: 3
 
 infer:
-  model_path: './checkpoint/ag_DB_bb_mobilenet_v3_small_he_DB_Head_bs_16_ep_1200/DB_best.pth.tar'
+  model_path: './checkpoint/DB_best.pth.tar'
   path: '/src/notebooks/detect_text/icdar2015/ch4_test_images'
   save_path: './result'

config/det_DB_resnet50_3_3.yaml

@@ -1,27 +1,27 @@
 base:
   gpu_id: '0'
   algorithm: DB
-  pretrained: True
+  pretrained: False
   in_channels: [256, 512, 1024, 2048]
   inner_channels: 256
   k: 50
   adaptive: True
   crop_shape: [640,640]
   shrink_ratio: 0.4
-  n_epoch: 1201
-  start_val: 400
+  n_epoch: 600
+  start_val: 6000
   show_step: 20
   checkpoints: ./checkpoint
   save_epoch: 100
   restore: False
   restore_file : ./DB.pth.tar
 
 backbone:
-  function: ptocr.model.backbone.det_resnet_3*3,resnet50
+  function: ptocr.model.backbone.det_resnet_3_3,resnet50
 
 head:
-  function: ptocr.model.head.det_DBHead,DB_Head
-#  function: ptocr.model.head.det_FPEM_FFM_Head,FPEM_FFM_Head
+#   function: ptocr.model.head.det_DBHead,DB_Head
+ function: ptocr.model.head.det_FPEM_FFM_Head,FPEM_FFM_Head
 #  function: ptocr.model.head.det_FPNHead,FPN_Head
 
 segout:
@@ -59,7 +59,7 @@ optimizer_decay:
 
 trainload:
   function: ptocr.dataloader.DetLoad.DBProcess,DBProcessTrain
-  train_file: /src/notebooks/detect_text/icdar2015/train_list.txt
+  train_file: /src/notebooks/MyworkData/huayandang/train_list.txt
   num_workers: 10
   batch_size: 8
 
@@ -75,10 +75,10 @@ testload:
 postprocess:
   function: ptocr.postprocess.DBpostprocess,DBPostProcess
   is_poly: False
-  thresh: 0.5
-  box_thresh: 0.6
+  thresh: 0.2
+  box_thresh: 0.3
   max_candidates: 1000
-  unclip_ratio: 2
+  unclip_ratio: 1.5
   min_size: 3
 
 infer:

config/det_DB_resnet50_mul.yaml

@@ -0,0 +1,89 @@
+base:
+  gpu_id: '1'  # 设置训练的gpu id，多卡训练设置为 '0,1,2'
+  algorithm: DB  # 算法名称
+  pretrained: True # 是否加载预训练
+  in_channels: [256, 512, 1024, 2048] # 
+  inner_channels: 256 #
+  k: 50
+  n_class: 3
+  adaptive: True
+  crop_shape: [640,640]  #训练时crop图片的大小
+  shrink_ratio: 0.4 # kernel向内收缩比率
+  n_epoch: 1200 # 训练的epoch
+  start_val: 400 #开始验证的epoch，如果不想验证直接设置数值大于n_epoch
+  show_step: 20 #设置迭代多少次输出一次loss
+  checkpoints: ./checkpoint #保存模型地址
+  save_epoch: 100 #设置每多少个epoch保存一次模型
+  restore: False #是否恢复训练
+  restore_file : ./DB.pth.tar #恢复训练所需加载模型的地址
+
+backbone:
+  function: ptocr.model.backbone.det_resnet,resnet50
+
+head:
+  function: ptocr.model.head.det_DBHead,DB_Head
+#  function: ptocr.model.head.det_FPEM_FFM_Head,FPEM_FFM_Head
+#  function: ptocr.model.head.det_FPNHead,FPN_Head
+
+segout:
+  function: ptocr.model.segout.det_DB_segout,SegDetectorMul
+
+architectures:
+  model_function: ptocr.model.architectures.det_model,DetModel
+  loss_function: ptocr.model.architectures.det_model,DetLoss
+
+loss:
+  function: ptocr.model.loss.db_loss,DBLossMul
+  l1_scale: 10
+  bce_scale: 1
+  class_scale: 1
+
+#optimizer:
+#  function: ptocr.optimizer,AdamDecay
+#  base_lr: 0.002
+#  beta1: 0.9
+#  beta2: 0.999
+
+optimizer:  
+  function: ptocr.optimizer,SGDDecay
+  base_lr: 0.002
+  momentum: 0.99
+  weight_decay: 0.0005
+
+optimizer_decay:
+  function: ptocr.optimizer,adjust_learning_rate_poly
+  factor: 0.9
+
+#optimizer_decay:
+#  function: ptocr.optimizer,adjust_learning_rate
+#  schedule: [1,2]
+#  gama: 0.1
+
+trainload:
+  function: ptocr.dataloader.DetLoad.DBProcess,DBProcessTrainMul
+  train_file: /src/notebooks/fangxuwei_96/TextGenerator-master/output/train/train_list.txt
+  num_workers: 10
+  batch_size: 8
+
+testload:
+  function: ptocr.dataloader.DetLoad.DBProcess,DBProcessTest
+  test_file: /src/notebooks/detect_text/icdar2015/test_list.txt
+  test_gt_path: /src/notebooks/detect_text/icdar2015/ch4_test_gts/
+  test_size: 736
+  stride: 32
+  num_workers: 5
+  batch_size: 4
+
+postprocess:
+  function: ptocr.postprocess.DBpostprocess,DBPostProcessMul
+  is_poly: False #测试时，检测弯曲文本设置成 True，否则就是输出矩形框
+  thresh: 0.5
+  box_thresh: 0.6
+  max_candidates: 1000
+  unclip_ratio: 2
+  min_size: 3
+
+infer:
+  model_path: './checkpoint/ag_DB_bb_resnet50_he_DB_Head_bs_8_ep_601_train_mul/DB_best.pth.tar'
+  path: '/src/notebooks/fangxuwei_96/TextGenerator-master/output/img/'
+  save_path: './result'