FOTS_TF
FOTS_TF copied to clipboard
Pay20Y
models
I have trained a model on synth text dataset, the detection is okay, but recognition is very bad.
https://drive.google.com/open?id=1SZaPveIjdhpkQgv6UL75VRi2c_kZIIhr
@DecentMakeover Can the trained model recognize simple Numbers and English characters?
I think some training tricks are needed. Anyway these are not stated in the paper. I can't find any bugs in my code, but there's still no ideal result. :( I change some details in recognition, you can update the code and try again.
@Pay20Y
Thank you very much. I even changed the code to use icdar2015 to retrain the model, and the training did not last long, prompting a warning, but it did not stop and the training continued.
·····
RuntimeWarning:invalid value encountered in float_scalars
······
When I first started, I found that the training took a long time, but the GPU memory was not used very much, I suspected that I used CPU training. I recommend using this command.
CUDA_VISIBLE_DEVICES=0 python2 multigpu_train.py --gpu_list=0 --training_data_path=/path/to/DataSet
Otherwise it's going to be slow. How much memory and graphics memory is best?
@DecentMakeover I was running the model when the machine hung up again. Is 8G not enough memory?
look for this tf.app.flags.DEFINE_integer('batch_size_per_gpu', 8, '') in multigpu_train.py and change accordingly
@DecentMakeover Thank you very much. It's solved.I have set batch_size_per_gpu to 8, now I can train well, but the memory usage is very high, maybe my machine memory is too small.
@Pay20Y The accuracy of word recognition is very bad.Why is that?
Which testset do you use? If not icdar15 please set use_vocab False in eval.py
@Pay20Y Could you help me ?! How to train it on the Arabic language? CHAR.VECTOR contain letters which in Arabic is not like English, Arabic letters are concatenated with each other !!!! Looking forward to your help
@Pay20Y Could you help me ?! How to train it on the Arabic language? CHAR.VECTOR contain letters which in Arabic is not like English, Arabic letters are concatenated with each other !!!! Looking forward to your help Sorry, I'm not familiar with Arabic. You can refer to here. It's a multi-language recognizer.
@Pay20Y
Could you help me in evaluate the model it shows me this error I am not experienced in MAkeFile
@Pay20Y
Could you help me in evaluate the model it shows me this error I am not experienced in MAkeFile
Sorry, I have no idea about it. I copied the NMS code from EAST. Maybe you can find the solution from the issue.
@Pay20Y I know this is an old thread, however I came across your repository very recently. I'm trying to replicate results on SynthText data (English for now). I used the E2E-MLT , that model performed decent on detection and very bad on recognition. I was hoping this repo does a better job on SynthText.
Looks like there have already been some changes that make training better, before I run experiments wanted to get your thoughts to see what parameters and tricks are optimal.
Appreciate any help !
