CTPN + DenseNet + CTC based end-to-end Chinese OCR implemented using tensorflow and keras

请问楼主,我采用您的代码,自己做了几个数据集,但是每次都是乱码,请问这是怎么回事呢,应该怎么解决呀 Recognition Result:

蝇俛蝇俛蝇蝇蝇蝇蝇蝇蝇蝇蝇蝇蝇蝇蝇 蝇蝇蝇蝇蝇健蝇蝇蝇蝇蝇拽蝇蝇蝇蝇蝇蝇蝇健蝇蝇蝇健蝇蝇蝇蝇蝇蝇 蝇蝇蝇蝇蝇蝇蝇蝇蝇蝇蝇蝇蝇蝇蝇蝇蝇蝇蝇蝇蝇蝇蝇蝇蝇蝇蝇蝇蝇蝇蝇蝇蝇蝇蝇膜蝇蝇蝇蝇蝇蝇蝇

自己训练出的模型是包括ctc和loss参数的，直接在densenet/models下使用会报错： ValueError: Dimension 1 in both shapes must be equal, but are 5990 and 17765 for 'Assign_159' (op: 'Assign') with input shapes: [768,5990], [768,17765].

求大神指导怎么修改模型?

Hi All,

I have tried training SynthText data and i observe that after some epochs its getting in to below exception.

tensorflow/core/framework/op_kernel.cc:1273] OP_REQUIRES failed at ctc_loss_op.cc:168 : Invalid argument: All labels must be nonnegative integers, batch: 94 labels: -1 Traceback (most recent call last): File "train.py", line 181, in callbacks = [checkpoint, earlystop, changelr, tensorboard]) File "/home/ubuntu/srijith/dl/lib/python3.5/site-packages/keras/legacy/interfaces.py", line 91, in wrapper return func(*args, **kwargs) File "/home/ubuntu/srijith/dl/lib/python3.5/site-packages/keras/engine/training.py", line 1418, in fit_generator initial_epoch=initial_epoch) File "/home/ubuntu/srijith/dl/lib/python3.5/site-packages/keras/engine/training_generator.py", line 217, in fit_generator class_weight=class_weight) File "/home/ubuntu/srijith/dl/lib/python3.5/site-packages/keras/engine/training.py", line 1217, in train_on_batch outputs = self.train_function(ins) File "/home/ubuntu/srijith/dl/lib/python3.5/site-packages/keras/backend/tensorflow_backend.py", line 2715, in call return self._call(inputs) File "/home/ubuntu/srijith/dl/lib/python3.5/site-packages/keras/backend/tensorflow_backend.py", line 2675, in _call fetched = self._callable_fn(*array_vals) File "/home/ubuntu/srijith/dl/lib/python3.5/site-packages/tensorflow/python/client/session.py", line 1439, in call run_metadata_ptr) File "/home/ubuntu/srijith/dl/lib/python3.5/site-packages/tensorflow/python/framework/errors_impl.py", line 528, in exit c_api.TF_GetCode(self.status.status)) tensorflow.python.framework.errors_impl.InvalidArgumentError: All labels must be nonnegative integers, batch: 94 labels: -1 [[{{node ctc/CTCLoss}} = CTCLoss[_class=["loc:@training/Adam/gradients/ctc/CTCLoss_grad/mul"], ctc_merge_repeated=true, ignore_longer_outputs_than_inputs=false, preprocess_collapse_repeated=false, _device="/job:localhost/replica:0/task:0/device:CPU:0"](ctc/Log, ctc/ToInt64, ctc/ToInt32_2, ctc/ToInt32_1)]]

Thank You, Srijith

请问训练代码里的 basemodel 是什么作用？ 我看代码在 loadweights 之后，没有后续的操作，请问这个 basemodel 是干嘛的呢？ 另外，我想在作者提供的 weights 基础之上，添加训练集，继续训练，应该怎么训练呢？

modelPath = './models/pretrain_model/keras.h5' if os.path.exists(modelPath): print("Loading model weights...") basemodel.load_weights(modelPath) print('done!')

Not enough time for target transition sequence (required: 38, available: 35)58You can turn this error into a warning by using the flag ignore_longer_outputs_than_inputs

不知道是不是因为数据集的原因，我的数据集是不定字长的，然后resize成了280*32，然后运行train.py的时候出错了。

++ python setup.py build_ext --inplace running build_ext skipping 'bbox.c' Cython extension (up-to-date) building 'utils.bbox' extension creating build creating build/temp.linux-x86_64-3.6 {'gcc': ['-Wno-cpp', '-Wno-unused-function']} gcc -pthread -B /home/test/anaconda3/compiler_compat -Wl,--sysroot=/ -Wsign-compare -DNDEBUG -g -fwrapv -O3 -Wall -Wstrict-prototypes -fPIC -I/home/test/anaconda3/lib/python3.6/site-packages/numpy/core/include -I/home/test/anaconda3/include/python3.6m -c bbox.c -o build/temp.linux-x86_64-3.6/bbox.o -Wno-cpp -Wno-unused-function creating /home/test/orc_gc/chinese_ocr-master/ctpn/lib/utils/utils gcc -pthread -shared -B /home/test/anaconda3/compiler_compat -L/home/test/anaconda3/lib -Wl,-rpath=/home/test/anaconda3/lib -Wl,--no-as-needed -Wl,--sysroot=/ build/temp.linux-x86_64-3.6/bbox.o -o /home/test/orc_gc/chinese_ocr-master/ctpn/lib/utils/utils/bbox.cpython-36m-x86_64-linux-gnu.so /home/test/anaconda3/compiler_compat/ld: cannot find -lpthread /home/test/anaconda3/compiler_compat/ld: cannot find -lc collect2: error: ld returned 1 exit status error: command 'gcc' failed with exit status 1

请问我是缺少了什么必要的库吗

windows下用cygwin运行setup.sh报出如下错误，请问该怎么解决呢？

Traceback (most recent call last): File "setup_cpu.py", line 57, in cmdclass={'build_ext': custom_build_ext}, File "D:\Program Files\Python\lib\distutils\core.py", line 148, in setup dist.run_commands() File "D:\Program Files\Python\lib\distutils\dist.py", line 955, in run_commands self.run_command(cmd) File "D:\Program Files\Python\lib\distutils\dist.py", line 974, in run_command cmd_obj.run() File "D:\Program Files\Python\lib\site-packages\Cython\Distutils\build_ext.py", line 164, in run _build_ext.build_ext.run(self) File "D:\Program Files\Python\lib\distutils\command\build_ext.py", line 339, in run self.build_extensions() File "setup_cpu.py", line 37, in build_extensions customize_compiler_for_nvcc(self.compiler) File "setup_cpu.py", line 23, in customize_compiler_for_nvcc default_compiler_so = self.compiler_so AttributeError: 'MSVCCompiler' object has no attribute 'compiler_so' mv: 无法获取'utils/*' 的文件状态(stat): No such file or directory

train.py 147-151行 modelPath = './models/pretrain_model/keras.h5' if os.path.exists(modelPath): print("Loading model weights...") basemodel.load_weights(modelPath) print('done!')

load_weight 加载不进去啊 ，虽然输出了done，训练的时候acc还是从0开始

ValueError: Dimension 1 in both shapes must be equal, but are 5988 and 5987 for 'Assign_159' (op: 'Assign') with input shapes: [768,5988], [768,5987] 算上最开始的空格，我的char_std_5990.txt 一共5987个字符，请问该怎么改啊？

我用46*120进行训练得到的h5文件 在model加载时 报错 ValueError: Dimension 0 in both shapes must be equal, but are 768 and 960. Shapes are [768,5990] and [960,5990]. for 'Assign_159' (op: 'Assign') with input shapes: [768,5990], [960,5990].

怎么解决啊 @YCG09

训练时报错 tensorflow.python.framework.errors_impl.InvalidArgumentError: Labels length is zero in batch 8 [[{{node ctc/CTCLoss}}]]

发现batchsize设置很小的时候（比如说2）可以正常运行，但是batchsize设置为32，64或者更大的时候就会报这个错，图片和label都能对上，有时可以正常跑完一个epoch但是第二个epoch报这个错...卡了几天了，我的label_length都是10没有是0的情况...求大神指点

I find that the line-nms delete some boxes. It seems a little confusing. The left-top and right-top point are used as left-top and right-bottom point , respectively. The x-coordinate of left-bottom point is used as confidence. It is a little difficult to understand it.

https://github.com/YCG09/chinese_ocr/blob/9af64905f7e1a226d902feac5c0dfd64399ce6b2/ctpn/lib/text_connector/detectors.py#L42

首先很感谢repo主无私的分享，该repo非常经典。在使用该repo过程中走过很多坑以下简要介绍：（1）模型推断阶段BN层未关闭，（2）模型推断时dropout未关闭（但测试了很多，发现未关闭也没有出现影响），（3）模型中图像宽度缩小8倍进入ctc loss计算这一点限制很大，使得你训练使用的图像宽度和图像中的文字序列长度有较好的比例，不然就会导致ctc loss无法计算出现ctc loss 为inf的情况，即模型结构使得图像宽度缩小8倍很坑，（4）模型采用的原始训练数据是很规整的同一32*280的图片，统一的10个字符长度，但如果我们拿自己的图像（或者开源的数据）是不定长图像宽度以及字符长度，所以为了保证我们同一个batch中图像的宽度是一致的，我们有2种策略选择：一是直接统一resize成同一高和宽（一定要确保宽度满足条件：宽//8>=字符长度+连续字符重复总个数）;二是在每个batch中以最大宽度为基准进行pad。这两种策略我测试了在一些图像宽度变化较大，仿射变换严重的数据集上都不太好，首先策略1直接resize会造成图像中文字发生变形，其次策略2在每个batch中以最大宽度为基准进行pad会造成原始图像（特别是一些图与具有最大宽度的图之间宽度相差较大）出现较多非文字空白区域，使得模型难以识别这样的图。 综上：如果想要在宽度一致性较差，宽高比例不好以及图像倾斜度较大的数据集上进行训练建议放弃该repo，选择其他的repo（比如crnn的）