The accuracy rises slowly

[DrewdropLife]

I used to train Chinese text recognition. After training for 10w iterations, the accuracy rate is still less than 0.5，while the loss value has dropped below 0.2.What might be wrong? Thanks. This is my configs: workers = 8 batch_max_length = 16 batch_size = 112 imgH = 32 imgW = 280 keep_ratio = True rgb = False sensitive = True data_filtering_off = False keep_ratio_with_pad = True total_data_usage_ratio = 1.0

num_fiducial = 20
input_channel = 1
output_channel = 512
hidden_size = 256
num_gpu = 1

[yy9568]

@changxinC Have you solved the problem now?

[DrewdropLife]

@changxinC Have you solved the problem now?

No

[opconty]

sorry for late reply.

As far as I know, you can replace the linear layer (output part) with encoder from Transformer w.r.t. Chinese Text recognition. Since I have no more GPUs available to evaluate the idea, so plz tell me the result if it works.

[DrewdropLife]

sorry for late reply.

As far as I know, you can replace the linear layer (output part) with encoder from Transformer w.r.t. Chinese Text recognition. Since I have no more GPUs available to evaluate the idea, so plz tell me the result if it works.

Thanks for reply, I planned to use the decoder from Transformer before, but I didn't do it because I was too busy. I will try it when I have time.