RealBlur data for training about mprnet

[adityac8]

Hi @asdf1996

In Table 4 of MPRNet, we report two results for RealBlur datasets.

For the top part, we simply take the pretrained model of GoPro dataset and test it on RealBlur datasets.
For the bottom part, we train two separate models, each for RealBlurR and RealBlurJ. For this, we follow the same training strategy as the authors of RealBlur dataset. We use GoPro+BSD+RealBlurR for training our RealBlurR model and GoPro+BSD+RealBlurJ for training our RealBlurJ model. We train the models for 100 epochs.

Thanks

[asdf1996]

Thanks for reply.
So is it better to train the model from scratch， instead of using pre-trained model of GoPro and fine-tuning on RealBlur dataset?

[BZboys]

I had the same problem. The loss becomes inf in the task of deblur, and reducing the learning rate is useless.

[swz30]

Thanks for reply.
So is it better to train the model from scratch， instead of using pre-trained model of GoPro and fine-tuning on RealBlur dataset?

We did not try finetuning.

[asdf1996]

We did not try finetuning.

Could you please also provide the model weights of trained on BSD and RealBlur dataset? THANKS!

[swz30]

Hi @asdf1996

We use a slightly lighter network for the GoPro+BSD+RealBlur dataset. Please change n_feat=80

MPRNet/Deblurring/MPRNet.py

Line 239 in 435f483

def __init__(self, in_c=3, out_c=3, n_feat=96, scale_unetfeats=48, scale_orsnetfeats=32, num_cab=8, kernel_size=3, reduction=4, bias=False):

You can download the weights from https://drive.google.com/drive/folders/1NY0nVT7A1w6cWTL8oig9e-ufaMT974mE?usp=sharing

[asdf1996]

Thank you!