Low accuracy Performance
Opened this issue · 4 comments
Hi, I trained a model with "ReCTS" dataset for 1000 Epoch (almost one week). I get really low accuracy. Your model works quite well. How did you train your model? Which dataset and parameters you used, could you please give information?
I followed all the steps as you described at tutorial.
@zhang0jhon @Tian14267
I trained my recognition model with "ReCTS" , "LSVT" and "ArT" datasets in ICDAR-2019. Because of weak supervised attention mechanism, you can feel free to use arbitrary-shaped images for training even with synthesized text images. Furthermore, text images with sufficient transformations help to learn robust 2D attention maps.
@zhang0jhon How was your loss graph? I got loss=1 at 1000 epochs. Also how many epochs did you train? Don't you think after 1000 epochs (with only ReCTS dataset), I should have at least average performance?
Thanks in advance.
Multiple datasets with plenty of arbitrary-shaped images is the key point for robust text recognition model.
@etatbak Hello, I fail to train the model, can you send me the icdar_datasets.npy to my email: zhou19920226@126.com ? Thank you very much.