shuochsu
training speed is slowly.
Hello! I'm from China and I'm researching the problem of multipath of ToF using your method now. And when I use your net structure and the synthetic data for training model, I find the training speed is very slow, more slower than you said in your article(my machine: ubuntu16.04 64bit, GPU GTX1080 8GB, i7-7700 CPU, 16GB RAM, I use CUDA and cuDNN), maybe exceed 1,000 hours, So I want to know is the Net parameter: batchSize is 1? Or other parameters are changing? Please, Thank you a lot.
Hi, thanks for reporting the issue. Can you provide a copy of the command line logs?
the command line is: DATA_ROOT=./datasets/mpi_correction_corrnormamp2depth_albedoaug_zpass niter=50 save_epoch_freq=200 mat=1 input_nc=5 output_nc=1
model=pix2pix align_data=1 name=my_resnet_lite_my_imageGAN_128 which_model_netG=my_resnet_lite which_model_netD=my_imageGAN_128
use_GAN=1 lambda_A=10 norm=none
noise=1 lr=0.00005 resize_or_crop=resize_and_crop_ratio_rand mask_nan=1 tv_strength=1e-4 tv_weight_scheme=1 tv_weight_scale=1
manualSeed=1 inv_lambda=1 display_port=1234 th train.lua;
and the default parameter of batchsize in the file of options.lua is "1", and here are some logs
,
,
, Maybe I should change the batchsize?
Hi, could you tell me your model use which train datasets, and which parameters you set when you train your model? I find the results when I test the model which is trained by myself, are worse than yours. Thank you very much.
hi，I have found the cause of the issue of training speed slowly, but also thank you very much
hello，how to resolve problem that the training speed is slow? Is your result is good as the author's?
