0

I am working on word detection and using MMOCR’s ‘DBNetpp’ detection model, please refer [2202.10304] Real-Time Scene Text Detection with Differentiable Binarization and Adaptive Scale Fusion.

I want to use this pre-trained model and fine tune on my dataset. Therefore, I need to update the configuration file for ‘DBNetpp’. here is the part of config file. I would like to know which parameters (specifically the learning rates) can be changed to have descent first results (decrease in loss).

# optimizer

optimizer = dict(type="SGD", lr=0.007, momentum=0.9, weight_decay=0.0001)

optimizer_config = dict(grad_clip=None)

# learning policy

lr_config = dict(policy="poly", power=0.9, min_lr=1e-7, by_epoch=False)

Any help would be appreciated. Thank you in advance.

aarya
  • 83
  • 1
  • 8
  • The base learning rate is the parameter `lr=0.007` in the `optimizer` dictionary. Other parameters in that dictionary can also be adjusted to affect training accordingly. The `lr_config` dictionary also has parameters that affect how the learning rate changes throughout training (see section 4.2 of the paper). – Allie Jun 21 '22 at 13:38
  • Yes, thank you for your input. But I also need to add the warmup period, how do I do it ? – aarya Jun 21 '22 at 13:54

0 Answers0