关于char_dict.json和ord_map.json的疑问

[peterzhou1520]

你好，我在运行的是master分支的那个项目，按照readme里说的生成了tf文件，用项目自带的.json文件可以成功训练和测试，但是我想训练这个网络只识别数字，所以用chinese_version_debug分支的tools下的establish_char_dict.py生成了char_dict.json和index_2_ord_map.json，把后者改成ord_map.json之后都放回了master分支的char_dict目录下面

生成的char_dict.json如下： {"48": "0", "49": "1", "50": "2", "51": "3", "52": "4", "53": "5", "54": "6", "55": "7", "56": "8", "57": "9"}

生成的ord_map.json如下： {"0": "48", "1": "49", "2": "50", "3": "51", "4": "52", "5": "53", "6": "54", "7": "55", "8": "56", "9": "57"}

运行train_shadownet.py之后出现了一下错误： Traceback (most recent call last): File "C:/Users/Administrator/Desktop/CRNN_Tensorflow-master/tools/train_shadownet.py", line 179, in train_shadownet(args.dataset_dir, args.weights_path) File "C:/Users/Administrator/Desktop/CRNN_Tensorflow-master/tools/train_shadownet.py", line 129, in train_shadownet preds = decoder.sparse_tensor_to_str(preds[0]) File "C:\Users\Administrator\Desktop\CRNN_Tensorflow-master\local_utils\data_utils.py", line 141, in sparse_tensor_to_str values = np.array([self.ord_map[str(tmp)] for tmp in values]) File "C:\Users\Administrator\Desktop\CRNN_Tensorflow-master\local_utils\data_utils.py", line 141, in values = np.array([self.ord_map[str(tmp)] for tmp in values]) KeyError: '34'

最后的keyError不是固定的，还出现过'17','10','24'等数字，请问这个问题要怎么解决呢。

[MaybeShewill-CV]

@peterzhou1520 这个只要维护一个数字字符到标签的映射关系，你只要按照data文件夹中的char_dict.txt文件组织你自己的数据，然后调用tools里面的脚本这个映射关系就可以自动生成。

[peterzhou1520]

你好，我前面就是按照char_dict.txt文件组织的，从0-9就是 0 1 2 . . 9 这种格式，生成的json文件是 {"48": "0", "49": "1", "50": "2", "51": "3", "52": "4", "53": "5", "54": "6", "55": "7", "56": "8", "57": "9"} 和 {"0": "48", "1": "49", "2": "50", "3": "51", "4": "52", "5": "53", "6": "54", "7": "55", "8": "56", "9": "57"} 但是运行失败

我试了试改了train_shadownet.py里的num_classes，按照0-9再加上一类总共11类 t = crnn_model.ShadowNet(phase='Train', hidden_nums=256, layers_nums=2, seq_length=25, num_classes=11) 现在可以训练了，是不是问题就出在这里呢。

[peterzhou1520]

不好意思啊，今天我在0-9的基础上加了一个字母Y，本意是用Y来代替人民币符号做数字金额的识别，生成的json文件是这样： char_dict.json:{"48": "0", "49": "1", "50": "2", "51": "3", "52": "4", "53": "5", "54": "6", "55": "7", "56": "8", "57": "9"，"89": "10"} ord_map.json:{"0": "48", "1": "49", "2": "50", "3": "51", "4": "52", "5": "53", "6": "54", "7": "55", "8": "56", "9": "57","10": "89"} num_classes我改成了12，然后报错了，报错信息：

self._traceback = self._graph._extract_stack() # pylint: disable=protected-access

InvalidArgumentError (see above for traceback): Saw a non-null label (index >= num_classes - 1) following a null label, batch: 1 num_classes: 12 labels: [[Node: CTCLoss = CTCLoss[ctc_merge_repeated=true, ignore_longer_outputs_than_inputs=false, preprocess_collapse_repeated=fal

请问将num_classes改成12有问题吗，该怎么改

[frozenscrypt]

I need to recognise only letters from a to z and numbers from 0 to 9 and nothing else. For this purpose I need to have a mapping for characters that includes only those that I need. How do I generate a char_dict.json and order_map.json? I searched for a char_dict.txt file but couldn't find it. How do I train the network to recognize only the characters that I want?