Open rose-jinyang opened 1 year ago
You can update the channel of 64 in the network. As far as I know, the larger the patch size, the better the final effect will be.(you also can use TCL)
Thanks for your reply.
Is it channel? Otherwise width?
Do u mean this parameter?
What do u think about the total training epochs?
Thanks for your reply. Is it channel? Otherwise width? Do u mean this parameter?
What do u think about the total training epochs? Details of our training will be published in the future, you can follow nafnet, mprnet training strategies
Hi @Tombs98 How many epochs did u train your model for?
Hi @Tombs98 How many epochs did u train your model for?
Details of our training will be published in the future, you can follow NAFNet, MPRNet training strategies
你好 用去雨预训练直接测试 出来的图像右下角都是有黑色条状是吗 然后计算PSNR SSIM需要对得到图像再处理?
Hello How are you? Thanks for contributing to this project. I am going to train a model on my custom dataset which contains about 5500 high-quality samples without corresponding low-quality samples. So I introduced to generate corresponding low-quality image for high-quality image in DataLoaderTrain class (dataset_RGB.py). For this, I used several blur augmentations of albumentations library. I trained a model for 2000 epochs. the input patch size into the model is still 256 as your setting. Do you think if there is any problem in my training strategy? For example, Is the total training epochs(2000 epochs) enough? May I use other patch size rather than 256? How can I train M3SNet-64 rather than M3SNet-32?