Hi @zhang0jhon

Thanks for the great work! The pretrained model you put in the docker image worked pretty well! However I want to fine-tune the parameters for my specific application. Is it possible to fine-tune the pretrained model? Since the model is different between the docker image and here, do you have the traning code for the docker version and/or do you have a fine-tuning code already?

Thanks in advance!

@zhang0jhon When running test.py with the model, i get error:

 Traceback (most recent call last):
  File "test.py", line 121, in <module>
    test(args)
  File "test.py", line 91, in test
    model = TextRecognition(args.pb_path, cfg.seq_len+1)
  File "test.py", line 23, in __init__
    self.init_model()
  File "test.py", line 37, in init_model
    self.label_ph = self.sess.graph.get_tensor_by_name('label:0')
  File "/home/home/anaconda3/envs/p3/lib/python3.6/site-packages/tensorflow/python/framework/ops.py", line 3972, in get_tensor_by_name
    return self.as_graph_element(name, allow_tensor=True, allow_operation=False)
  File "/home/home/anaconda3/envs/p3/lib/python3.6/site-packages/tensorflow/python/framework/ops.py", line 3796, in as_graph_element
    return self._as_graph_element_locked(obj, allow_tensor, allow_operation)
  File "/home/home/anaconda3/envs/p3/lib/python3.6/site-packages/tensorflow/python/framework/ops.py", line 3838, in _as_graph_element_locked
    "graph." % (repr(name), repr(op_name)))
KeyError: "The name 'label:0' refers to a Tensor which does not exist. The operation, 'label', does not exist in the graph."

你好： 请问我再运行完python flaskapp.py，上传图片之后，进行预测的时候，会显示下面的错误，请问这是什么原因导致的？ tensorflow.python.framework.errors_impl.InvalidArgumentError: Default MaxPoolingOp only supports NHWC on device type CPU [[node pool0/MaxPool (defined at /tensor_flow/OCRSpace/ocr/ocr/text_detection.py:29) ]]

Original stack trace for 'pool0/MaxPool':

请问应该怎么修改呢？多谢。。。

seq_len是单个文本行最多可识别的字符数，是这个意思吧。 我现在想训一个短文本行的模型，最长seq_len设置为16，请问还需要修改哪些地方？ 直接改为16报错，报ValueRrror，维度不匹配, 具体错误为cannot feed value of shape(16,33) for Tensor 'label:0', which has shape(?,17) 麻烦您指点一下，多谢

假设已经定位到文字部分（暂不考虑定位方法），若采用AttentionOCR去识别，识别结果是针对图片中文字整体识别还是针对图片中的文字一个一个进行识别，因为之前采用crnn-ctc的模型是对图片中的文字一起识别，但是我看到您的images文件夹中图片有标识每一个汉字的识别概率，不知道我表达清楚没有^~^

Hello, I have trained this network for about 80k steps but even though it started to detect a lot of the text parts it never predicted the EOS token. How much did you train it? Do I need some special post-processing?

Thank you for the awesome network 👍

I am unable to pull from the docker hub repository that you have mentioned. When I executed

docker pull zhang0jhon/demo:ocr

I got an error saying:

Error response from daemon: pull access denied for zhang0jhon/demo:ocr, repository does not exist or may require docker login

Is the repository still existing or have you given pull access to all?

W tensorflow/core/common_runtime/bfc_allocator.cc:211] Allocator (GPU_0_bfc) ran out of memory trying to allocate 358.89MiB. The caller indicates that this is not a failure, but may mean that there could be performance gains if more memory were available.

普通显卡 tensorflow/core/common_runtime/gpu/gpu_device.cc:1432] Found device 0 with properties: name: GeForce GT 1030 major: 6 minor: 1 memoryClockRate(GHz): 1.5185 pciBusID: 0000:01:00.0 totalMemory: 1.95GiB freeMemory: 1.63GiB

I have to questions. First what e2e accuracy mean? its accuracy for detection and recognition? And det-hmean is only for detection? And second on wich datasets was this model trained?

Thanks for your wonderful work! I observe that the text recongnition network outputs the text and probabilities. I wonder how the probilities calculate and what's the range of them? I can't find the funtion associated. Are the probilities just produced by the network? And they range in [0,1]? Thanks!

   First and foremost, thanks to you opensource the awesome STR algorithm,I have some questions about training and testing.

I cope the checkpoint from the docker image, And I find detecting and recognition separted two model, one is ICDAR_0.7.pb, another is text_recognition_5435.pb. this mode is not apropriate this repo code. Hence I train my model through dataset.py genderating icdar_datasets.npy , with a lot of reading code, I find that the test.py can not work.it appear the error: 2021-06-16 15:01:59.018589: I tensorflow/stream_executor/platform/default/dso_loader.cc:42] Successfully opened dynamic library libcudnn.so.7 2021-06-16 15:01:59.233909: E tensorflow/stream_executor/cuda/cuda_dnn.cc:329] Could not create cudnn handle: CUDNN_STATUS_INTERNAL_ERROR 2021-06-16 15:01:59.261653: E tensorflow/stream_executor/cuda/cuda_dnn.cc:329] Could not create cudnn handle: CUDNN_STATUS_INTERNAL_ERROR Traceback (most recent call last): File "/home/alan/anaconda3/envs/AttentionOCR/lib/python3.6/site-packages/tensorflow/python/client/session.py", line 1356, in _do_call return fn(*args) File "/home/alan/anaconda3/envs/AttentionOCR/lib/python3.6/site-packages/tensorflow/python/client/session.py", line 1341, in _run_fn options, feed_dict, fetch_list, target_list, run_metadata) File "/home/alan/anaconda3/envs/AttentionOCR/lib/python3.6/site-packages/tensorflow/python/client/session.py", line 1429, in _call_tf_sessionrun run_metadata) tensorflow.python.framework.errors_impl.UnknownError: 2 root error(s) found. (0) Unknown: Failed to get convolution algorithm. This is probably because cuDNN failed to initialize, so try looking to see if a warning log message was printed above. [[{{node InceptionV4/InceptionV4/Conv2d_1a_3x3/Conv2D}}]] (1) Unknown: Failed to get convolution algorithm. This is probably because cuDNN failed to initialize, so try looking to see if a warning log message was printed above. [[{{node InceptionV4/InceptionV4/Conv2d_1a_3x3/Conv2D}}]] [[sequence_preds/_17]]

Traceback (most recent call last): File "train.py", line 81, in train() File "train.py", line 64, in train starting_epoch=cfg.starting_epoch File "/usr/local/lib/python3.6/dist-packages/tensorpack/train/config.py", line 119, in init assert_type(model, ModelDescBase, 'model') File "/usr/local/lib/python3.6/dist-packages/tensorpack/train/config.py", line 107, in assert_type name, tp.name, v.class.name) AssertionError: model has to be type 'ModelDescBase', but an object of type 'AttentionOCR' found.

ZhangOjhon your pre-trained model is gr8 !! for recognising characters generically. gr8 work it would be nice if you share the config file which includes the values used for the hyper parameters.
I went through your code I just found out this formula discrepancy with the one in Bahadanus attention paper. 1.) the hidden state derived from the lstm decoder layer should only be given to the softmax. But in the code the attention_feature, prev_wemb, hidden_layer is concatenated and then given to the softmax. It would be helpful if you provide the explanation of why the formula is altered so and so. Thanks in advance.

Hi, I would love to use this amazing project for my project. Does it has MIT license? If yes, can you add the license to the repo? It will encourage people to use your amazing ocr model in their projects as well.