Attention-ocr-Chinese-Version
Attention-ocr-Chinese-Version copied to clipboard
A-bone1
Attention OCR Based On Tensorflow
最近在做古籍识别，用到了paddleocr中的attention部分作为序列预测部分，但是碰到了一个问题： 在attention的decoder部分，如果将前一时刻decoder的输出作为当前时刻的输入，模型训练效果很差，收敛很慢，准确率上不去；但是如果将前一时刻的真实标签作为当前时刻的输入，模型收敛速度直接起飞，很快训练准确率就到1，但是预测准确率一直是0，似乎是这样做直接把真实标签作为了训练模型的输入，导致模型根本没有得到训练。 但就我个人对seq2seq模型的理解，在训练时将前一时刻的真实标签作为当前时刻的输入，应该是更容易将模型往理想的方向训练，更容易收敛，模型理应训练得更好，但是出现了预测准确了一直为0的情况。我真的很困惑，不知道大佬是否可以解决一下我的疑问。
How to predict a single image (random size)?? what file do you run I run demo_reference.py file, it requere 600x150 include 4 image are combined Please, thanks
How to pass the images to the model those having different width images(Long width and short width) without resizing?
训练到后面ｌｏｓｓ越来越大，从ｌｏｇ里面看text 都是这样的东西 AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA uuuuuuuuOOOOOOOOOOOOOOOOOOOOOOOOOOOOO -- 我制造ｔｆ-record 都是用的楼主的那个程序，请问有人知道什么原因吗
Instructions for updating: Please switch to tf.train.get_or_create_global_step INFO:tensorflow:Restoring parameters from /home/ucmed/opt/python/models-master/research/attention_ocr/python/logs/model.ckpt-0 INFO 2019-01-03 02:14:41.000888: tf_logging.py: 82 Restoring parameters from /home/ucmed/opt/python/models-master/research/attention_ocr/python/logs/model.ckpt-0 INFO:tensorflow:Starting Session. INFO 2019-01-03 02:14:55.000713: tf_logging.py: 82 Starting Session. INFO:tensorflow:Saving...
During handling of the above exception, another exception occurred: Traceback (most recent call last): File "train.py", line 211, in app.run() File "/home/shakey/.local/lib/python3.6/site-packages/tensorflow/python/platform/app.py", line 125, in run _sys.exit(main(argv)) File "train.py", line...
Can we feed the Document images dataset instead of a small word dataset to this Architecture? What is the max-sequence length that can be used? can you please suggest me...
