Can you share pretrained model on CVUSA dataset?

[happsky]

Can you share pretrained model on CVUSA dataset?

[kregmi]

Please find the pretrained models here .

[happsky]

@kregmi Models of X-fork and X-seq work for me, however I have encountered the following error when I use the pretrained pix2pix model,

checkpoints_dir ./checkpoints
/home/csdept/torch/install/bin/luajit: /home/csdept/torch/install/share/lua/5.1/torch/File.lua:272: read error: read 5499 blocks instead of 1484150018 at /home/csdept/torch/pkg/torch/lib/TH/THDiskFile.c:344
stack traceback:
[C]: in function 'readChar'
/home/csdept/torch/install/share/lua/5.1/torch/File.lua:272: in function 'readObject'
/home/csdept/torch/install/share/lua/5.1/torch/File.lua:368: in function 'readObject'
/home/csdept/torch/install/share/lua/5.1/torch/File.lua:353: in function 'readObject'
/home/csdept/torch/install/share/lua/5.1/torch/File.lua:369: in function 'readObject'
/home/csdept/torch/install/share/lua/5.1/torch/File.lua:369: in function 'readObject'
/home/csdept/torch/install/share/lua/5.1/torch/File.lua:353: in function 'readObject'
/home/csdept/torch/install/share/lua/5.1/torch/File.lua:369: in function 'readObject'
...e/csdept/torch/install/share/lua/5.1/nngraph/gmodule.lua:495: in function 'read'
/home/csdept/torch/install/share/lua/5.1/torch/File.lua:351: in function 'readObject'
/home/csdept/torch/install/share/lua/5.1/torch/File.lua:409: in function 'load'
...dept/projects/cross_view_image_synthesis_2/util/util.lua:282: in function 'load'
test_pix2pix.lua:82: in main chunk
[C]: in function 'dofile'
...dept/torch/install/lib/luarocks/rocks/trepl/scm-1/bin/th:150: in main chunk
[C]: at 0x00405d50

Commands I used,

DATA_ROOT=/home/csdept/projects/crossnet/CVUSA/cvusa name=X-Fork-cvusa which_direction=g2a batchSize=4 phase=test which_epoch=35 th test_fork.lua;

DATA_ROOT=/home/csdept/projects/crossnet/CVUSA/cvusa name=x_pix2pix_cvusa which_direction=g2a batchSize=4 phase=test which_epoch=35 th test_pix2pix.lua;

[kregmi]

Seems like the model was not uploaded correctly. I have uploaded the models again here.
Also, I can see in your commands that you tried to synthesize aerial images (which_direction=g2a) but the experiments on CVUSA datasets were performed on a2g direction only and the shared models work to synthesize streetview images.

[happsky]

checkpoints_dir ./checkpoints
/home/csdept/torch/install/bin/luajit: /home/csdept/torch/install/share/lua/5.1/torch/File.lua:375: unknown object
stack traceback:
[C]: in function 'error'
/home/csdept/torch/install/share/lua/5.1/torch/File.lua:375: in function 'readObject'
/home/csdept/torch/install/share/lua/5.1/torch/File.lua:368: in function 'readObject'
/home/csdept/torch/install/share/lua/5.1/torch/File.lua:369: in function 'readObject'
/home/csdept/torch/install/share/lua/5.1/torch/File.lua:353: in function 'readObject'
/home/csdept/torch/install/share/lua/5.1/torch/File.lua:369: in function 'readObject'
...e/csdept/torch/install/share/lua/5.1/nngraph/gmodule.lua:495: in function 'read'
/home/csdept/torch/install/share/lua/5.1/torch/File.lua:351: in function 'readObject'
/home/csdept/torch/install/share/lua/5.1/torch/File.lua:409: in function 'load'
...dept/projects/cross_view_image_synthesis_2/util/util.lua:282: in function 'load'
test_pix2pix.lua:82: in main chunk
[C]: in function 'dofile'
...dept/torch/install/lib/luarocks/rocks/trepl/scm-1/bin/th:150: in main chunk
[C]: at 0x00405d50

Still not work, any ideas?
For pix2pix, which image format should I input, 4 images or just 2 images?
Thank you for remaining me the problem, I have a different image order so this would not be a problem.

[kregmi]

Not sure what was the problem. I re-uploaded the model, then downloaded and tried in my machine. Its working for me now. Hope it will work for you as well. Let me know!

[happsky]

My bad, it works like a charm now. Thanks for your help.

[kregmi]

From the command line, you can do> th and hit enter to get into torch environment. There you can load required packages by doing: require 'nngraph' and require 'cudnn'
After this you can load the model as: model = torch.load('35_net_G.t7')

Also about the image format, you should input 4 images, same as x-fork and x-seq; the dataloader is written to expect 4 images concatenated.

[happsky]

It works now, can you share the pretrained X-Seq model for segmentation generation on this dataset?

[kregmi]

Unfortunately I could not locate the X-seq model for segmentation map generation on CVUSA dataset. I can train the network and share the model with you, this will take about couple of days. Alternatively, you can train it yourself.

[happsky]

Ok, thank you.
Since there is no segmentation map for the aerial image on this dataset, what did you do for data preparation
{streetview image, aerial image, segmentation map for streetview image, segmentation map for aerial image}?

[kregmi]

You can use any image at the position of segmentation map for aerial image, lets say use segmentation map of streetview image. This is a quick fix and a good option.
eg. {streetview image, aerial image, segmentation map for streetview image, segmentation map for streetview image}.
This will not affect our training because we are training a2g direction only.

[happsky]

Hi @kregmi, the pretrained models for X-Seq and Pix2pix have the same size on this dataset, i.e., 535.1M, and the results produced by both models are quite similar.
Can you check it? Thank you so much.

[happsky]

@kregmi can you help me?