search

Karel911 / TRACER

TRACER: Extreme Attention Guided Salient Object Tracing Network (AAAI 2022) implementation in PyTorch
Apache License 2.0
195 stars 41 forks source link

Are training hyperparameters the same for all types of pretrained models with different backbones? #12

Closed OPHoperHPO closed 2 years ago

OPHoperHPO commented 2 years ago

Hi,

Thanks for your incredible research. I tried train network from scratch on dataset with objects and tried multiple pretrained backbones with removed edge generation. Model with backbone 0 or 1 came to very high accuracy during training with your model-training params, such channels, RFB_aggregated_channel, frequency_radius, denoise and other values in the config.py.

I tried move to heavy 7 backbone, but accuracy couldn't went to desireble values with same model parameters.

So, i got question, are model parameters such channels, RFB_aggregated_channel, frequency_radius, denoise and gamma used the same as in config.py for training your model with 7 backbone?