search

sshan-zhao / GASDA

Geometry-Aware Symmetric Domain Adaptation for Monocular Depth Estimation, CVPR 2019
133 stars 23 forks source link

some problem about ablation experiment #13

Open leefly072 opened 3 years ago

leefly072 commented 3 years ago

Geometry-Aware Symmetric Domain Adaptation for Monocular Depth Estimation is a very interesting paper and gave me a lot of inspiration. When run the experiment the ablation study, there is some problem boring me. Could you tell me how to training Ft on the Gs2t? Thank you very much for you help.

sshan-zhao commented 3 years ago

Could you provide more details?

leefly072 commented 3 years ago

It is a great honor to receive your reply. When I try to train the Fs on {Xt,Xs2t}, I execute the command line : "python train.py --model fs --gpu_ids 0 --batchSize 8 --loadSize 256 1024 --g_src_premodel ./cyclegan/G_Src.pth "; But I have no idea when do the ablation experiment, such as try to train the Ft on Xs2t; If you can give me some hints will be very much appreciated. Do I use the command line or modify the code of network.py? I have no idea about how to train the Fs and Ft model with different data sets? Especially looking forward to your reply, thank you very much.

sshan-zhao commented 3 years ago

The command for training Ft is provided. Does it not work? python train.py --model ft --gpu_ids 0 --batchSize 8 --loadSize 256 1024 --g_tgt_premodel ./cyclegan/G_Tgt.pth For other datasets, you might need to first train an image-to-image translation model using CycleGAN. Then, prepare the datasets according to the provided file structure, or use your ownself file structure but you need to re-write the code for loading the data. After that, you can train the model using the provided commands (or with some modifications). Hope it can help.

leefly072 commented 3 years ago

it's very useful for you response. I still have some questions, hope you can give some advise. I use the paper dataset to training the model , and when I faced to the section 4.4, SYN(Fs trained on Xs); or SYN@REAL(Ft trained on Gs2t(Xs)). I have no idea how to run the experiment. in other words, if I want to training the Fs only on source domain (synthetic), what can I do , could I do some modification of the giving commend to run the Fs on Xs? Thank you very much for your valuable time to reply my question.

sshan-zhao commented 3 years ago

You may need to modify the code, such as the forward and backward function in ft_model.py / fs_model.py

leefly072 commented 3 years ago

Thank you very much for your valuable time to reply my question. Let me try to modify the related code~ Thank you very much.Thank you for the code.