search

LoSealL / VideoSuperResolution

A collection of state-of-the-art video or single-image super-resolution architectures, reimplemented in tensorflow.
MIT License
1.62k stars 296 forks source link

RDN train problems~ #74

Closed lhwwhl closed 5 years ago

lhwwhl commented 5 years ago

Hi, Thank you very much for sharing. Do you train RDN model to getting psnr 32.47 x4 in set5 from paper? How much time do you spending? How many epoch you have been trained?

LoSealL commented 5 years ago

No, currently not. As I follow the training step in the official code, it takes about 3 days for the 1st step (x2, init 128x128) on a Maxwell Titan X.

lhwwhl commented 5 years ago

I’am a newcomer to deep learning. Thank for your response. You mean that training step should train the 1st step x2, and then train the 2nd step x3 with x2 result as pre-train? X4 is same as follow?

LoSealL commented 5 years ago

In the author's code README:

    # BI, scale 2, 3, 4
    # BIX2F64D18C6G64P48, input=48x48, output=96x96
    th main.lua -scale 2 -netType RDN -nFeat 64 -nFeaSDB 64 -nDenseBlock 16 -nDenseConv 8 -growthRate 64 -patchSize 96 -dataset div2k -datatype t7  -DownKernel BI -splitBatch 4 -trainOnly true

    # BIX3F64D18C6G64P32, input=32x32, output=96x96, fine-tune on RDN_BIX2.t7
    th main.lua -scale 3 -netType resnet_cu -nFeat 64 -nFeaSDB 64 -nDenseBlock 16 -nDenseConv 8 -growthRate 64 -patchSize 96 -dataset div2k -datatype t7  -DownKernel BI -splitBatch 4 -trainOnly true  -preTrained ../experiment/model/RDN_BIX2.t7

    # BIX4F64D18C6G64P32, input=32x32, output=128x128, fine-tune on RDN_BIX2.t7
    th main.lua -scale 4 -nGPU 1 -netType resnet_cu -nFeat 64 -nFeaSDB 64 -nDenseBlock 16 -nDenseConv 8 -growthRate 64 -patchSize 128 -dataset div2k -datatype t7  -DownKernel BI -splitBatch 4 -trainOnly true -nEpochs 1000 -preTrained ../experiment/model/RDN_BIX2.t7 

    # BD, scale 3
    # BDX3F64D18C6G64P32, input=32x32, output=96x96, fine-tune on RDN_BIX3.t7
    th main.lua -scale 3 -nGPU 1 -netType resnet_cu -nFeat 64 -nFeaSDB 64 -nDenseBlock 16 -nDenseConv 8 -growthRate 64 -patchSize 96 -dataset div2k -datatype t7  -DownKernel BD -splitBatch 4 -trainOnly true -nEpochs 200 -preTrained ../experiment/model/RDN_BIX3.t7

    # DN, scale 3
    # DNX3F64D18C6G64P32, input=32x32, output=96x96, fine-tune on RDN_BIX3.t7
    th main.lua -scale 3 -nGPU 1 -netType resnet_cu -nFeat 64 -nFeaSDB 64 -nDenseBlock 16 -nDenseConv 8 -growthRate 64 -patchSize 96 -dataset div2k -datatype t7  -DownKernel DN -splitBatch 4 -trainOnly true  -nEpochs 200 -preTrained ../experiment/model/RDN_BIX3.t7

In the 1st step, train a x2 model using patch_size=96; In the 2nd step, fune-tuning x4 model, by replacing the last upsample layer from x2 to x4, using patch_size=128.

LoSealL commented 5 years ago

Another option is directly training a x4 model from scratch. I think it could also reach the minima.

lhwwhl commented 5 years ago

I understand, thank you very much for your answer.