Media-Smart / vedastr

A scene text recognition toolbox based on PyTorch
Apache License 2.0
534 stars 100 forks source link

How to modify the net architecture? #32

Closed Wzj02200059 closed 4 years ago

Wzj02200059 commented 4 years ago

Hi~ Thanks for this awesome repo。 I have a question,i have noticed that the backbone network used several pooling layer, I think its not conducive to very small size、resolution img, so i just want to reduce the pooling layer and add dilated conv, pls give me some advice that how to modify it。 thanks~

ChaseMonsterAway commented 4 years ago

Hi, could you tell me which config file you want to modify?

Wzj02200059 commented 4 years ago

tps_resnet_bilstm_attn.py and small_satrn.py thx

ChaseMonsterAway commented 4 years ago

The params of ConvModule are here. So, if you want to add a dilation params, you can modify a ConvModule as follows:


 ('conv', dict(type='ConvModule', in_channels=1, out_channels=32, kernel_size=3,
               stride=1, padding=1, norm_cfg=norm_cfg, dilation=1))
Wzj02200059 commented 4 years ago

OK,i got it