xinghedyc / mxnet-cnn-lstm-ctc-ocr Goto Github PK

Hi Yuchen, I can't find your email address on the Internet, so please forgive me for asking here.
I recently read your paper "Fused Text Segmentation Networks for Multi-oriented Scene Text Detection". And I am very interested in your method FTSN. You propose "Mask-NMS " to obtain final detection results in the paper, can you share this part code with me, thank you very much.

Hi,there
is it possible to change warpctc with mx.contrib.sym.ctc_loss ?
thx!

@xinghedyc I have trained the model with my data, but how to predict one image!
the project does not include predict code?

@xinghedyc Hi,could you tell me why you use SGD instead of Adadelta or Adam in your trainner? I'm not very similar to how to choose a proper optimizer......
Thanks a lot. 😓

('seq_len : ', 100)
('seq_len : ', 100)
[02:42:42] e:\work\testroot\mxnet\src\operator./cudnn_algoreg-inl.h:106: Runnin
g performance tests to find the best convolution algorithm, this can take a whil
e... (setting env variable MXNET_CUDNN_AUTOTUNE_DEFAULT to 0 to disable)
('seq_len : ', 20L)
Traceback (most recent call last):
File "predict.py", line 128, in
model.forward(data_batch, is_train=False)
File "D:\Anaconda2\lib\site-packages\mxnet-0.11.1-py2.7.egg\mxnet\module\bucke
ting_module.py", line 420, in forward
data_batch.provide_label)
File "D:\Anaconda2\lib\site-packages\mxnet-0.11.1-py2.7.egg\mxnet\module\bucke
ting_module.py", line 347, in switch_bucket
symbol, data_names, label_names = self._sym_gen(bucket_key)
File "predict.py", line 87, in sym_gen
num_label=25, dropout=0.5), ('data', 'l0_init_c', 'l1_init_c', 'l0_init_h',
'l1_init_h'), (
File "D:\DeepLearning\MXNET\example\mxnet-cnn-lstm-ctc-ocr\text_lstm.py", line
156, in bi_lstm_unroll
hidden =mx.sym.Flatten(data=column_features[k])
File "D:\Anaconda2\lib\site-packages\mxnet-0.11.1-py2.7.egg\mxnet\symbol\symbo
l.py", line 511, in getitem
raise TypeError('Symbol only support integer index to fetch i-th output')
TypeError: Symbol only support integer index to fetch i-th output

how to handle this ?
thx!

Hi,
I do the same thing that replaces the cnn part to resnet in Torch7, but I don't see the performance improvement. Do you test your model on the public datasets?