codeslake/IFAN

training the code with pretrained weights

xianrui-luo opened this issue · 7 comments

xianrui-luo commented

Hi, I would like to ask about how to load the pretrained weights and start training without interfering other process?

codeslake commented

Hi, are you trying to resume training? or to start the training with pretrained weights with a new learning rate scheduling?

xianrui-luo commented

start the training with pretrained weights with a new learning rate scheduling

codeslake commented

I see. The current code does not provide the functionality at the moment. I will spend some of my time adding the functionality.
Do your pretrained weights include weights for the reblurring network also?

xianrui-luo commented

I think so. And i have encountered a warning while training. It said "Warning! No positional inputs found for a module, assuming batch size is 1." But I have set the batchsize as your training details. I wonder what happened?

image

codeslake commented

Hi, I've just pushed some commits. Pull the repo and try the following command:

  • Use a mode name different from the mode name of the pretrained model to prevent the process from overriding the folder of the pretrained model.
  • Give absolute path of the pretrained weights to -ra or --resume_absolute option.
  • The learning rate scheduleling will start from the beginning following the config file given with the --config option.
CUDA_VISIBLE_DEVICES=0,1,2,3 python -B -m torch.distributed.launch --nproc_per_node=4 --master_port=9000 run.py \
            --is_train \
            --mode IFAN_finetune \
            --config config_IFAN \
            --trainer trainer \
            --network IFAN \
            -b 2 \
            -th 8 \
            -ra /path/to/the/pretriained/weights \
            -ss \
            -dist
xianrui-luo commented

Thank you for your help! I will try it now.

makeyyp commented

I am a beginner and when replicating your code this error occurred.The dataset was placed as required。Can you help me
image