Training time of the RestoreFormer++ model

[jiangjin1246]

Hi, thanks for your remarkable work. I've several questions as follows:

1. How many epochs was the RestoreFormer++ model trained?

2. How long does it take to train the RestoreFormer++ model？

Thanks again and look forward to your feedback.

[wzhouxiff]

The training details are outlined below:

• Step 1: Achieve ROHQD. We fine-tune the model provided in taming-transformers using FFHQ with an input size of 512x512. It is important to note that the pre-trained model is trained with an input size of 256x256. This step consists of 180 epochs and takes approximately 8 days.

• Step 2: Obtain RestoreFormer++. Initially, we freeze the decoder and train the encoder exclusively with degraded-clear pairs. After about 100 epochs, we fine-tune both the encoder and decoder together for approximately 50 epochs. This process takes around 10 days.

[jiangjin1246]

Thank you for your detailed answer! It really helps! By the way, what kind of GPUs are utilized, and how many GPUs are needed to attain such training efficiency?

[wzhouxiff]

4 V100s with batchsize = 4 for each gpu.

[luser350]

Hi, @wzhouxiff Thanks for sharing your great work I want to ask a question:

1. For training on a custom dataset, did I need to train ROHQD or can I directly proceed with the RF++ config file?
2. If ROHQD is required then from where I can download this

"/group/30042/zhouxiawang/checkpoints/RestoreFormer/release/logs/2022-11-11T18-36-57_ROHQDROHQD_gpus7_lmdb_h4_seed30/checkpoints/last.ckpt.62"