clovaai / deep-text-recognition-benchmark

Text recognition (optical character recognition) with deep learning methods, ICCV 2019
Apache License 2.0
3.75k stars 1.1k forks source link

Training time #89

Closed SealQ closed 5 years ago

SealQ commented 5 years ago

How long do you need to train your data? Why do I need a long time to train with my own data? It still takes me about a week to train the model once with three 1080ti GPUs.

ku21fan commented 5 years ago

Hello,

With our training dataset which we released, we took about 2.5~3 days with TPS-ResNet-BiLSTM-Attn. and took about 10 Hours with CRNN (None-VGG-BiLSTM-CTC). (with imgH 32, imgW 100, num_iter 300K, batch size 192, batch_max_length 25)

In our code, when we use multi-GPU, we multiply the batch size with the number of GPUs (by default setting). I recommend adjusting batch_size (or change this line.) https://github.com/clovaai/deep-text-recognition-benchmark/blob/4381a5fb46f328984dad1d2fe255728eab328f17/train.py#L284

Best