xiumingzhang / genre-shapehd Goto Github PK

Hi, I'm reproducing the MarrNet training on the chair class, the reconstruction results are good but it somehow gives me kind of perspective surfaces as follow, the surfaces get transparent when rotating the object in different views. Do you know how to solve this? Is it because of the voxel_isosurf_th? voxel_isosurf_th =0.25 is used.

Hi!

Thanks for sharing the code and dataset.

Could you please provide some information about how to generate full spherical map as ground truth for training? Or I notice that there are rotated voxels in dataset dir for each spherical map and how to generate such aligned voxel?

I met some problem when I tried to align the shapenet object and depth when I rendered my own dataset with some specific poses.

Thanks!

Hi,

I am fine-tuning shapeHD with a pre-trained 3D GAN, but it seems all the test model will lead to this result.

Any idea about this?

Thank you!

Hi all!

I was trying to generate my own dataset, however the spherical maps I generate for my own dataset (from depth and from the voxel representation) do not match.

I'm using the camera back projection module:

Which is used here:

My question are the following:

• what is fl? Is it the focal length of your camera? In which units?
• If our camera is not at a distance of 2.2 from the object, is it sufficient to change it here, or we should change somewhere else?

Thanks in advance for your time.

Best,
Pedro

Hello,
I am trying to run GenRe-ShapeHD in a conda environment with CUDA 9.2 on Ubuntu 18.04.2 LTS and am getting stuck when I run sudo ./build_toolbox.sh
The error I am getting is "cuda available but nvcc not found. Please add nvcc to $PATH"
I tried to fix this by adding a file to the GenRe-ShapeHD directory that adds /usr/local/cuda-9.2/bin to the path.

After doing this and running nvcc -V, it looks like it was able to find nvcc becasue it returns
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2018 NVIDIA Corporation
Built on Tue_Jun_12_23:07:04_CDT_2018
Cuda compilation tools, release 9.2, V9.2.148

But it still is giving me the same error that it cannot find nvcc when I run sudo ./build_toolbox.sh
Do you know what could be causing this issue?

Thank you

Hi, Xiuming

Thank you very much for sharing your excellent work.
When I run scripts/test_genre.sh with your trained models, I couldn't get the results as you show.
Here are the results that I get for the 4 testing examples:

ShapeHD testing goes well.
Could you please help me? Are the models uploaded wrong or anything else?

Thanks a lot!

Dan

Thank you for your impressive work
How do I download GenRe data? None of the links provided in readme can be downloaded. Is the download link closed?

Hi, there.
Thank you for sharing. I tried to use multi-GPU to run this, but failed. I found there is a function named data_parallel_decorator in model/netinterface.py and tried to write self.net=data_parallel_decorator(self.net) for multi-GPU, but there is a bug about the input structure. Is there any demo or guidance about how to use multi-GPU?
Thank you !

Best regards,
Xuting

We are training the marrnet1 with our own dataset but found out that we are missing the depth_minmax, could you please tell how to generate the depth_minmax? Thanks in advance!

Thanks for your reply.
I was faced with a new problem as follows. My version is :gcc version 5.4.0 20160609 (Ubuntu 5.4.0-6ubuntu1~16.04.10) with CUDA 9.0

Traceback (most recent call last):
  File "build.py", line 43, in <module>
    ffi.build()
  File "/home/fxy/anaconda3/envs/shaperecon/lib/python3.6/site-packages/torch/utils/ffi/__init__.py", line 189, in build
    _build_extension(ffi, cffi_wrapper_name, target_dir, verbose)
  File "/home/fxy/anaconda3/envs/shaperecon/lib/python3.6/site-packages/torch/utils/ffi/__init__.py", line 111, in _build_extension
    outfile = ffi.compile(tmpdir=tmpdir, verbose=verbose, target=libname)
  File "/home/fxy/anaconda3/envs/shaperecon/lib/python3.6/site-packages/cffi/api.py", line 723, in compile
    compiler_verbose=verbose, debug=debug, **kwds)
  File "/home/fxy/anaconda3/envs/shaperecon/lib/python3.6/site-packages/cffi/recompiler.py", line 1527, in recompile
    compiler_verbose, debug)
  File "/home/fxy/anaconda3/envs/shaperecon/lib/python3.6/site-packages/cffi/ffiplatform.py", line 22, in compile
    outputfilename = _build(tmpdir, ext, compiler_verbose, debug)
  File "/home/fxy/anaconda3/envs/shaperecon/lib/python3.6/site-packages/cffi/ffiplatform.py", line 58, in _build
    raise VerificationError('%s: %s' % (e.__class__.__name__, e))
cffi.VerificationError: LinkError: command 'gcc' failed with exit status 1

I have printed the values of normals' ground truth, and got the results that they are in the range of 0-100. May I ask about how to get the value? Is it scale from RGB or something else? Thank you!

Hi,

I am training MarrNet2 using default setting on shapenet chair dataset. It seems it always stops at Eval 3/1000 as shown below:

Any idea what's going on?

Thanks!

Hi, Xiuming
We trained 300 epoches for MarrNet2 on Chairs and it has converged.
We picked the best.pt and wrote code of IOU to evaluate the MarrNet2 without finetuning, but only get 0.077 IoU on the whole validation set. This is too low. Did you get the similar results for 2.5D to 3D without finetuning? and then get much better IoU after finetuning?

I'm not sure what goes wrong, I attached the IoU code as below. Did you also set the threshold as 0.5 to binarize voxel values after sigmoid?

Cheers.

Hi Xiuming!

Thanks for sharing the code.
I would like to train GenRe on my own dataset.

I already have:

1. the RGB image
2. the silhouette image
3. the normals
4. the aligned 3D model (voxel representation)

I need, at least, the spherical maps as well, right?

How can I create my own dataset to train GenRe? Do you have any example on how you created yours on chairs, aeroplanes and cars based on shapenet?

Thanks in advance,
Best,
Pedro

Hi there,

Thanks for releasing the codes. I was trying to run it on Windows, but met some issues. So I wonder what changes need to be made for running on Windows, such as compiling cam_bp using CUDA myself and if there are some things cannot be done in Windows.

Thank you in advance.

Hi,

We are training the spherical inpainting network and found out that the network predicts the same spherical_full and spherical_partial for all objects, below is the result from 22 epoch:

spherical_full:

spherical_partial is totally white:

The arguments we are using:

--load_offline False
--joint_train False

Any ideas why this happens? Thanks in advance!

I want to train the wgangp model using my own voxel data. There are two questions bothered me.

1. The released data you use is normalized voxel data. The element of my own voxel data is 0 or 1. Need I normalize my voxel data? And how to normalize it?
2. In the paper, the wgangp model was trained for 80 epoches. But the epoch was set to 1000 in the training script(https://github.com/xiumingzhang/GenRe-ShapeHD/blob/master/scripts/train_wgangp.sh). Which one is correct?
   Thank you!

Hi,

I am wondering when training the marrnet2 with ShapeNet synthetic images, did you train the Marrnet2 only with ground truth normal, depth or did you train the Marrnet2 with ground truth normal, depth for some epochs and then keep training Marrnet2 with predictions from Marrnet1.

And did you use canon_sup when training Marrnet2? What are the batch_size and epoch_batches (how many data used in training per epoch?) For example, when training chair class, did you use 4 x 2500 training points or 6778 x 20 all training points (ignore invalid point)? Because if I use all the training point per epoch, the Eval loss fluctuate a lot and the reconstruction results are not good.

And are these setups the same for GenRe?

Thanks

I've been stopped by this issue for several days.
while running test_genre.sh,I got the following error:
Traceback (most recent call last):
File "test.py", line 95, in
model.test_on_batch(i, batch)
File "/home/zhanghao/models/genre_full_model.py", line 182, in test_on_batch
pred = self.forward_with_trimesh(batch)
File "/home/zhanghao/models/genre_full_model.py", line 207, in forward_with_trimesh
proj = self.net.depth_and_inpaint.proj_depth(pred_abs_depth)
File "/media/zhanghao/娱乐/anaconda3/envs/shaperecon/lib/python3.6/site-packages/torch/nn/modules/module.py", line 477, in call
result = self.forward(*input, **kwargs)
File "/home/zhanghao/toolbox/cam_bp/cam_bp/modules/camera_backprojection_module.py", line 22, in forward
df = CameraBackProjection.apply(depth_t, fl, cam_dist, self.res)
File "/home/zhanghao/toolbox/cam_bp/cam_bp/functions/cam_back_projection.py", line 25, in forward
cam_bp_lib.back_projection_forward(depth_t, cam_dist, fl, tdf, cnt)
File "/media/zhanghao/娱乐/anaconda3/envs/shaperecon/lib/python3.6/site-packages/torch/utils/ffi/init.py", line 202, in safe_call
result = torch._C._safe_call(*args, **kwargs)
torch.FatalError: aborting at /data/vision/billf/scratch/ztzhang/shape_oneshot/ShapeRecon/toolbox/cam_bp/cam_bp/src/back_projection.c:14

Does anyone have solution for that? thanks.

Hi, we tried to use the function mentioned in #24 to render the spherical maps and here is our code:

b = 64
sgrid = u.make_sgrid(b, 0, 0, 0)
mesh = trimesh.load('xxx.obj')
im_depth, im_texture = util.util_sph.render_model(mesh, sgrid)
im_depth = im_depth.reshape(2 * b, 2 * b)
im_depth = np.where(im_depth > 1, 1, im_depth)
np.save('xxx.npy', im_depth)

We rendered the spherical maps of shapenet obj using this code and it worked, but when we tried to render the spherical maps of our dataset, it did not work and give us the following result, half of the map is white:

Original RGB image:

Any ideas why this happens? Is it because the obj size is too large or the projection center is not at the center of the object? We normalized the obj in different scale but did not help. Thanks in advance!

Hi,there,
Thank you for releasing this repo. I wanna use CD metric for evaluation. It is said in the paper that we need to sample the point from the isosurfaces of the voxel pair(pred and gt) , and then calculate the CD of the sampled points.
I don't know the detail of this process, I would appreciate it if you provide some code for this.
Thank you.

Best regards,
Xuting

Hi,
I've downloaded the dataset for training the GenRE. In this dataset, the ‘**_gt_rotvox_samescale_128.npz’ contains a 'voxel' file (128128128 array). Each point in this array is 0 or 1 or a decimal from 0 to 1. I understand that the 0 and 1 can represent the voxel. But what does the decimal mean?

Hi there,

Thank you for releasing the codes. I wonder if you can provide or where I can find the parameters you used for the rendering, e.g. the camera position, whether orthographic or perspective, the range of depth(normalization using min/max or a specified range) and etc.

Thanks,
Ryan

python train.py
--net shapehd
--marrnet2 "$marrnet2"
--gan "$gan"
--dataset shapenet
--classes "$class"
--canon_sup
--w_gan_loss 1e-3
--batch_size 4
--epoch_batches 200
--eval_batches 10
--optim adam
--lr 1e-3
--epoch 1000
--vis_batches_vali 10
--gpu "$gpu"
--save_net 1
--workers 4
--logdir "$outdir"
--suffix '{classes}_w_ganloss{w_gan_loss}'
--tensorboard
$*

How to set $gan parameter?

Hi there,

Thanks for sharing the code.

As I don't have a powerful GPU and the dataset is huge, I'm wondering how many epochs roughly does MarrNet2 requires to train? (using all default settings, Adam)

Cheers!

it is difficult for training the wgangp model, could you provide your trained wgangp model? thank you!

How can I obtain 'obj_spherical' and 'depth_spherical' with another dataset? Am I supposed to use

Hi, I'm training the wgangp on our dataset but it generates the same result as following for all valid_voxel in every epoch, we also ran the code on shapenet and it generated the same result as our. is this correct? any idea why this happens? thanks! We only have one view per object, so I am assuming that each view is the canonical view.

this is our loss at epoch 19

Thank you for the code .
Is there a tensorflow version of the network part code?

Will the code about the reprojection consistency about the MarrNet model be opened？

Hi @ztzhang and @xiumingzhang

There is a small discrepancy (it's a minor issue) on the file name of the trained models (shapehd.pt and marrnet1_with_minmax.pt) and the expected ones on scripts/test_shapehd.sh

Best,
Pedro

hi，I have downloaded this，

This repo comes with a few Pix3D images and ShapeNet renderings, located in ```
downloads/data/test, for testing purposes.
For training, we make available our RGB and 2.5D sketch renderings, paired with their corresponding 3D shapes, for ShapeNet cars, chairs, and airplanes, with each object captured in 20 random views. Note that this .tar is 143 GB.
wget http://genre.csail.mit.edu/downloads/shapenet_cars_chairs_planes_20views.tar -P downloads/data/
mkdir downloads/data/shapenet/
tar -xvf downloads/data/shapenet_cars_chairs_planes_20views.tar -C downloads/data/shapenet/

Is 02691156_1a04e3eab45ca15dd86060f189eb133_view000_spherical.npz the 2.5D sketch ？And how can I read it by python?Use numpy?
Looking forward to your reply!

Hello, I am attempting to train marrnet on another dataset, but the depth I obtain from my dataset is very large and varies a lot in comparison to the shapenet dataset. Are the sizes of the meshes in the shapenet dataset constrained to a certain range?

Excuse me .I have this problem in the step ./build_toolbox.sh. My versions are CUDA8.0 pytorch0.4.1 gcc4.8.5/5.5( I have tried both).

hi,
I am runing the test_genre with your latest code and released model, following the instructions in README
The result is not even close, not even with the few test cases in the repo
I think I may did something wrong somewhere, but I have no clue where is it.
In my results, the predicted depth seems correct, but the sphere map and final reconstuction are really bad.
my results are shown as follows:
(seems the output pred_proj_depth do not have the same upside with pred_voxel, but pred_proj_depth all looks correct, however the pred_proj_sph_full and pred_voxel are not even close)

When I ran scripts/train_marrnet1.sh 0,1 03001627+02691156+02958343
The params shows bellow,
Namespace(adam_beta1=0.5, adam_beta2=0.9, batch_size=4, classes='03001627+02691156+02958343', dataset='shapenet', epoch=1000, epoch_batches=2500, eval_at_start=False, eval_batches=5, expr_id=0, gpu='0,1', log_batch=False, log_time=True, logdir='./output/marrnet1', lr=0.001, manual_seed=None, net='marrnet1', optim='adam', pred_depth_minmax=True, resume=0, save_net=10, save_net_opt=False, sgd_dampening=0, sgd_momentum=0.9, suffix='{classes}', tensorboard=True, vis_batches_train=10, vis_batches_vali=10, vis_every_train=1, vis_every_vali=1, vis_param_f=None, vis_workers=4, wdecay=0.0, workers=4)
But after Epoch 99/1000, it couldn't run anymore. Is there something wrong with my configs?

Could you generate texture after reconstruction? Thanks

rgb_pattern='./downloads/data/test/shapehd/*_rgb.*'
mask_pattern='./downloads/data/test/shapehd/*_mask.*'

I have downloaded the dataset http://genre.csail.mit.edu/downloads/shapenet_cars_chairs_planes_20views.tar, but I can't find where the test folder is. The dataset unzipped as follows,

Is this normal?

hello,
when I use the pretrained genre model to test, there is a problem called "Model loaded without optimizer states", it seems that the pretrained model did not save the optimizers?

Testing Pipeline
==> Parsing arguments
Namespace(adam_beta1=0.5, adam_beta2=0.9, batch_size=1, classes='chair', dataset=None, epoch=0, epoch_batches=None, eval_at_start=False, eval_batches=None, expr_id=0, full_logdir=None, gpu='2', inpaint_path=None, input_mask='./downloads/data/test/genre/_silhouette.', input_rgb='./downloads/data/test/genre/_rgb.', joint_train=False, load_offline=False, log_batch=False, log_time=False, logdir=None, lr=0.0001, manual_seed=None, net='genre_full_model', net1_path=None, net_file='./downloads/models/full_model.pt', optim='adam', output_dir='./output/test', overwrite=True, padding_margin=16, pred_depth_minmax=True, resume=0, save_net=1, save_net_opt=False, sgd_dampening=0, sgd_momentum=0.9, suffix='{net}', surface_weight=1.0, tensorboard=False, vis_batches_train=10, vis_batches_vali=10, vis_every_train=1, vis_every_vali=1, vis_param_f=None, vis_workers=4, wdecay=0.0, workers=0)
==> Setting device
[Verbose] All designated GPU(s) free to use.
==> Setting up output directory
==> Setting up loggers
==> Setting up models
[Warning] Model loaded without optimizer states.
Traceback (most recent call last):
File "test.py", line 63, in
model = Model(opt, logger)
File "/mnt/disk/zhiyu/GenRe-ShapeHD/models/genre_full_model.py", line 152, in init
self.load_state_dict(opt.net_file, load_optimizer='auto')
File "/mnt/disk/zhiyu/GenRe-ShapeHD/models/netinterface.py", line 424, in load_state_dict
self._nets[i].load_state_dict(state_dicts['nets'][i])
File "/home/zhiyu/anaconda3/envs/shaperecon/lib/python3.6/site-packages/torch/nn/modules/module.py", line 719, in load_state_dict
self.class.name, "\n\t".join(error_msgs)))
RuntimeError: Error(s) in loading state_dict for Net:
Unexpected key(s) in state_dict: "depth_and_inpaint.net1.encoder.0.0.weight", "depth_and_inpaint.net1.encoder.0.1.weight", "depth_and_inpaint.net1.encoder.0.1.bias", "depth_and_inpaint.net1.encoder.0.1.running_mean", "depth_and_inpaint.net1.encoder.0.1.running_var", "depth_and_inpaint.net1.encoder.0.1.num_batches_tracked", "depth_and_inpaint.net1.encoder.1.0.conv1.weight", "depth_and_inpaint.net1.encoder.1.0.bn1.weight", "depth_and_inpaint.net1.encoder.1.0.bn1.bias", "depth_and_inpaint.net1.encoder.1.0.bn1.running_mean", "depth_and_inpaint.net1.encoder.1.0.bn1.running_var", "depth_and_inpaint.net1.encoder.1.0.bn1.num_batches_tracked", "depth_and_inpaint.net1.encoder.1.0.conv2.weight", "depth_and_inpaint.net1.encoder.1.0.bn2.weight", "depth_and_inpaint.net1.encoder.1.0.bn2.bias", "depth_and_inpaint.net1.encoder.1.0.bn2.running_mean", "depth_and_inpaint.net1.encoder.1.0.bn2.running_var", "depth_and_inpaint.net1.encoder.1.0.bn2.num_batches_tracked", "depth_and_inpaint.net1.encoder.1.1.conv1.weight", "depth_and_inpaint.net1.encoder.1.1.bn1.weight", "depth_and_inpaint.net1.encoder.1.1.bn1.bias", "depth_and_inpaint.net1.encoder.1.1.bn1.running_mean", "depth_and_inpaint.net1.encoder.1.1.bn1.running_var", "depth_and_inpaint.net1.encoder.1.1.bn1.num_batches_tracked", "depth_and_inpaint.net1.encoder.1.1.conv2.weight", "depth_and_inpaint.net1.encoder.1.1.bn2.weight", "depth_and_inpaint.net1.encoder.1.1.bn2.bias", "depth_and_inpaint.net1.encoder.1.1.bn2.running_mean", "depth_and_inpaint.net1.encoder.1.1.bn2.running_var", "depth_and_inpaint.net1.encoder.1.1.bn2.num_batches_tracked", "depth_and_inpaint.net1.encoder.2.0.conv1.weight", "depth_and_inpaint.net1.encoder.2.0.bn1.weight", "depth_and_inpaint.net1.encoder.2.0.bn1.bias", "depth_and_inpaint.net1.encoder.2.0.bn1.running_mean", "depth_and_inpaint.net1.encoder.2.0.bn1.running_var", "depth_and_inpaint.net1.encoder.2.0.bn1.num_batches_tracked", "depth_and_inpaint.net1.encoder.2.0.conv2.weight", "depth_and_inpaint.net1.encoder.2.0.bn2.weight", "depth_and_inpaint.net1.encoder.2.0.bn2.bias", "depth_and_inpaint.net1.encoder.2.0.bn2.running_mean", "depth_and_inpaint.net1.encoder.2.0.bn2.running_var", "depth_and_inpaint.net1.encoder.2.0.bn2.num_batches_tracked", "depth_and_inpaint.net1.encoder.2.0.downsample.0.weight", "depth_and_inpaint.net1.encoder.2.0.downsample.1.weight", "depth_and_inpaint.net1.encoder.2.0.downsample.1.bias", "depth_and_inpaint.net1.encoder.2.0.downsample.1.running_mean", "depth_and_inpaint.net1.encoder.2.0.downsample.1.running_var", "depth_and_inpaint.net1.encoder.2.0.downsample.1.num_batches_tracked", "depth_and_inpaint.net1.encoder.2.1.conv1.weight", "depth_and_inpaint.net1.encoder.2.1.bn1.weight", "depth_and_inpaint.net1.encoder.2.1.bn1.bias", "depth_and_inpaint.net1.encoder.2.1.bn1.running_mean", "depth_and_inpaint.net1.encoder.2.1.bn1.running_var", "depth_and_inpaint.net1.encoder.2.1.bn1.num_batches_tracked", "depth_and_inpaint.net1.encoder.2.1.conv2.weight", "depth_and_inpaint.net1.encoder.2.1.bn2.weight", "depth_and_inpaint.net1.encoder.2.1.bn2.bias", "depth_and_inpaint.net1.encoder.2.1.bn2.running_mean", "depth_and_inpaint.net1.encoder.2.1.bn2.running_var", "depth_and_inpaint.net1.encoder.2.1.bn2.num_batches_tracked", "depth_and_inpaint.net1.encoder.3.0.conv1.weight", "depth_and_inpaint.net1.encoder.3.0.bn1.weight", "depth_and_inpaint.net1.encoder.3.0.bn1.bias", "depth_and_inpaint.net1.encoder.3.0.bn1.running_mean", "depth_and_inpaint.net1.encoder.3.0.bn1.running_var", "depth_and_inpaint.net1.encoder.3.0.bn1.num_batches_tracked", "depth_and_inpaint.net1.encoder.3.0.conv2.weight", "depth_and_inpaint.net1.encoder.3.0.bn2.weight", "depth_and_inpaint.net1.encoder.3.0.bn2.bias", "depth_and_inpaint.net1.encoder.3.0.bn2.running_mean", "depth_and_inpaint.net1.encoder.3.0.bn2.running_var", "depth_and_inpaint.net1.encoder.3.0.bn2.num_batches_tracked", "depth_and_inpaint.net1.encoder.3.0.downsample.0.weight", "depth_and_inpaint.net1.encoder.3.0.downsample.1.weight", "depth_and_inpaint.net1.encoder.3.0.downsample.1.bias", "depth_and_inpaint.net1.encoder.3.0.downsample.1.running_mean", "depth_and_inpaint.net1.encoder.3.0.downsample.1.running_var", "depth_and_inpaint.net1.encoder.3.0.downsample.1.num_batches_tracked", "depth_and_inpaint.net1.encoder.3.1.conv1.weight", "depth_and_inpaint.net1.encoder.3.1.bn1.weight", "depth_and_inpaint.net1.encoder.3.1.bn1.bias", "depth_and_inpaint.net1.encoder.3.1.bn1.running_mean", "depth_and_inpaint.net1.encoder.3.1.bn1.running_var", "depth_and_inpaint.net1.encoder.3.1.bn1.num_batches_tracked", "depth_and_inpaint.net1.encoder.3.1.conv2.weight", "depth_and_inpaint.net1.encoder.3.1.bn2.weight", "depth_and_inpaint.net1.encoder.3.1.bn2.bias", "depth_and_inpaint.net1.encoder.3.1.bn2.running_mean", "depth_and_inpaint.net1.encoder.3.1.bn2.running_var", "depth_and_inpaint.net1.encoder.3.1.bn2.num_batches_tracked", "depth_and_inpaint.net1.encoder.4.0.conv1.weight", "depth_and_inpaint.net1.encoder.4.0.bn1.weight", "depth_and_inpaint.net1.encoder.4.0.bn1.bias", "depth_and_inpaint.net1.encoder.4.0.bn1.running_mean", "depth_and_inpaint.net1.encoder.4.0.bn1.running_var", "depth_and_inpaint.net1.encoder.4.0.bn1.num_batches_tracked", "depth_and_inpaint.net1.encoder.4.0.conv2.weight", "depth_and_inpaint.net1.encoder.4.0.bn2.weight", "depth_and_inpaint.net1.encoder.4.0.bn2.bias", "depth_and_inpaint.net1.encoder.4.0.bn2.running_mean", "depth_and_inpaint.net1.encoder.4.0.bn2.running_var", "depth_and_inpaint.net1.encoder.4.0.bn2.num_batches_tracked", "depth_and_inpaint.net1.encoder.4.0.downsample.0.weight", "depth_and_inpaint.net1.encoder.4.0.downsample.1.weight", "depth_and_inpaint.net1.encoder.4.0.downsample.1.bias", "depth_and_inpaint.net1.encoder.4.0.downsample.1.running_mean", "depth_and_inpaint.net1.encoder.4.0.downsample.1.running_var", "depth_and_inpaint.net1.encoder.4.0.downsample.1.num_batches_tracked", "depth_and_inpaint.net1.encoder.4.1.conv1.weight", "depth_and_inpaint.net1.encoder.4.1.bn1.weight", "depth_and_inpaint.net1.encoder.4.1.bn1.bias", "depth_and_inpaint.net1.encoder.4.1.bn1.running_mean", "depth_and_inpaint.net1.encoder.4.1.bn1.running_var", "depth_and_inpaint.net1.encoder.4.1.bn1.num_batches_tracked", "depth_and_inpaint.net1.encoder.4.1.conv2.weight", "depth_and_inpaint.net1.encoder.4.1.bn2.weight", "depth_and_inpaint.net1.encoder.4.1.bn2.bias", "depth_and_inpaint.net1.encoder.4.1.bn2.running_mean", "depth_and_inpaint.net1.encoder.4.1.bn2.running_var", "depth_and_inpaint.net1.encoder.4.1.bn2.num_batches_tracked", "depth_and_inpaint.net1.decoder_normal.0.0.deconv1.weight", "depth_and_inpaint.net1.decoder_normal.0.0.bn1.weight", "depth_and_inpaint.net1.decoder_normal.0.0.bn1.bias", "depth_and_inpaint.net1.decoder_normal.0.0.bn1.running_mean", "depth_and_inpaint.net1.decoder_normal.0.0.bn1.running_var", "depth_and_inpaint.net1.decoder_normal.0.0.bn1.num_batches_tracked", "depth_and_inpaint.net1.decoder_normal.0.0.deconv2.weight", "depth_and_inpaint.net1.decoder_normal.0.0.bn2.weight", "depth_and_inpaint.net1.decoder_normal.0.0.bn2.bias", "depth_and_inpaint.net1.decoder_normal.0.0.bn2.running_mean", "depth_and_inpaint.net1.decoder_normal.0.0.bn2.running_var", "depth_and_inpaint.net1.decoder_normal.0.0.bn2.num_batches_tracked", "depth_and_inpaint.net1.decoder_normal.0.0.upsample.0.weight", "depth_and_inpaint.net1.decoder_normal.0.0.upsample.1.weight", "depth_and_inpaint.net1.decoder_normal.0.0.upsample.1.bias", "depth_and_inpaint.net1.decoder_normal.0.0.upsample.1.running_mean", "depth_and_inpaint.net1.decoder_normal.0.0.upsample.1.running_var", "depth_and_inpaint.net1.decoder_normal.0.0.upsample.1.num_batches_tracked", "depth_and_inpaint.net1.decoder_normal.0.1.deconv1.weight", "depth_and_inpaint.net1.decoder_normal.0.1.bn1.weight", "depth_and_inpaint.net1.decoder_normal.0.1.bn1.bias", "depth_and_inpaint.net1.decoder_normal.0.1.bn1.running_mean", "depth_and_inpaint.net1.decoder_normal.0.1.bn1.running_var", "depth_and_inpaint.net1.decoder_normal.0.1.bn1.num_batches_tracked", "depth_and_inpaint.net1.decoder_normal.0.1.deconv2.weight", "depth_and_inpaint.net1.decoder_normal.0.1.bn2.weight", "depth_and_inpaint.net1.decoder_normal.0.1.bn2.bias", "depth_and_inpaint.net1.decoder_normal.0.1.bn2.running_mean", "depth_and_inpaint.net1.decoder_normal.0.1.bn2.running_var", "depth_and_inpaint.net1.decoder_normal.0.1.bn2.num_batches_tracked", "depth_and_inpaint.net1.decoder_normal.1.0.deconv1.weight", "depth_and_inpaint.net1.decoder_normal.1.0.bn1.weight", "depth_and_inpaint.net1.decoder_normal.1.0.bn1.bias", "depth_and_inpaint.net1.decoder_normal.1.0.bn1.running_mean", "depth_and_inpaint.net1.decoder_normal.1.0.bn1.running_var", "depth_and_inpaint.net1.decoder_normal.1.0.bn1.num_batches_tracked", "depth_and_inpaint.net1.decoder_normal.1.0.deconv2.weight", "depth_and_inpaint.net1.decoder_normal.1.0.bn2.weight", "depth_and_inpaint.net1.decoder_normal.1.0.bn2.bias", "depth_and_inpaint.net1.decoder_normal.1.0.bn2.running_mean", "depth_and_inpaint.net1.decoder_normal.1.0.bn2.running_var", "depth_and_inpaint.net1.decoder_normal.1.0.bn2.num_batches_tracked", "depth_and_inpaint.net1.decoder_normal.1.0.upsample.0.weight", "depth_and_inpaint.net1.decoder_normal.1.0.upsample.1.weight", "depth_and_inpaint.net1.decoder_normal.1.0.upsample.1.bias", "depth_and_inpaint.net1.decoder_normal.1.0.upsample.1.running_mean", "depth_and_inpaint.net1.decoder_normal.1.0.upsample.1.running_var", "depth_and_inpaint.net1.decoder_normal.1.0.upsample.1.num_batches_tracked", "depth_and_inpaint.net1.decoder_normal.1.1.deconv1.weight", "depth_and_inpaint.net1.decoder_normal.1.1.bn1.weight", "depth_and_inpaint.net1.decoder_normal.1.1.bn1.bias", "depth_and_inpaint.net1.decoder_normal.1.1.bn1.running_mean", "depth_and_inpaint.net1.decoder_normal.1.1.bn1.running_var", "depth_and_inpaint.net1.decoder_normal.1.1.bn1.num_batches_tracked", "depth_and_inpaint.net1.decoder_normal.1.1.deconv2.weight", "depth_and_inpaint.net1.decoder_normal.1.1.bn2.weight", "depth_and_inpaint.net1.decoder_normal.1.1.bn2.bias", "depth_and_inpaint.net1.decoder_normal.1.1.bn2.running_mean", "depth_and_inpaint.net1.decoder_normal.1.1.bn2.running_var", "depth_and_inpaint.net1.decoder_normal.1.1.bn2.num_batches_tracked", "depth_and_inpaint.net1.decoder_normal.2.0.deconv1.weight", "depth_and_inpaint.net1.decoder_normal.2.0.bn1.weight", "depth_and_inpaint.net1.decoder_normal.2.0.bn1.bias", "depth_and_inpaint.net1.decoder_normal.2.0.bn1.running_mean", "depth_and_inpaint.net1.decoder_normal.2.0.bn1.running_var", "depth_and_inpaint.net1.decoder_normal.2.0.bn1.num_batches_tracked", "depth_and_inpaint.net1.decoder_normal.2.0.deconv2.weight", "depth_and_inpaint.net1.decoder_normal.2.0.bn2.weight", "depth_and_inpaint.net1.decoder_normal.2.0.bn2.bias", "depth_and_inpaint.net1.decoder_normal.2.0.bn2.running_mean", "depth_and_inpaint.net1.decoder_normal.2.0.bn2.running_var", "depth_and_inpaint.net1.decoder_normal.2.0.bn2.num_batches_tracked", "depth_and_inpaint.net1.decoder_normal.2.0.upsample.0.weight", "depth_and_inpaint.net1.decoder_normal.2.0.upsample.1.weight", "depth_and_inpaint.net1.decoder_normal.2.0.upsample.1.bias", "depth_and_inpaint.net1.decoder_normal.2.0.upsample.1.running_mean", "depth_and_inpaint.net1.decoder_normal.2.0.upsample.1.running_var", "depth_and_inpaint.net1.decoder_normal.2.0.upsample.1.num_batches_tracked", "depth_and_inpaint.net1.decoder_normal.2.1.deconv1.weight", "depth_and_inpaint.net1.decoder_normal.2.1.bn1.weight", "depth_and_inpaint.net1.decoder_normal.2.1.bn1.bias", "depth_and_inpaint.net1.decoder_normal.2.1.bn1.running_mean", "depth_and_inpaint.net1.decoder_normal.2.1.bn1.running_var", "depth_and_inpaint.net1.decoder_normal.2.1.bn1.num_batches_tracked", "depth_and_inpaint.net1.decoder_normal.2.1.deconv2.weight", "depth_and_inpaint.net1.decoder_normal.2.1.bn2.weight", "depth_and_inpaint.net1.decoder_normal.2.1.bn2.bias", "depth_and_inpaint.net1.decoder_normal.2.1.bn2.running_mean", "depth_and_inpaint.net1.decoder_normal.2.1.bn2.running_var", "depth_and_inpaint.net1.decoder_normal.2.1.bn2.num_batches_tracked", "depth_and_inpaint.net1.decoder_normal.3.0.deconv1.weight", "depth_and_inpaint.net1.decoder_normal.3.0.bn1.weight", "depth_and_inpaint.net1.decoder_normal.3.0.bn1.bias", "depth_and_inpaint.net1.decoder_normal.3.0.bn1.running_mean", "depth_and_inpaint.net1.decoder_normal.3.0.bn1.running_var", "depth_and_inpaint.net1.decoder_normal.3.0.bn1.num_batches_tracked", "depth_and_inpaint.net1.decoder_normal.3.0.deconv2.weight", "depth_and_inpaint.net1.decoder_normal.3.0.bn2.weight", "depth_and_inpaint.net1.decoder_normal.3.0.bn2.bias", "depth_and_inpaint.net1.decoder_normal.3.0.bn2.running_mean", "depth_and_inpaint.net1.decoder_normal.3.0.bn2.running_var", "depth_and_inpaint.net1.decoder_normal.3.0.bn2.num_batches_tracked", "depth_and_inpaint.net1.decoder_normal.3.0.upsample.0.weight", "depth_and_inpaint.net1.decoder_normal.3.0.upsample.1.weight", "depth_and_inpaint.net1.decoder_normal.3.0.upsample.1.bias", "depth_and_inpaint.net1.decoder_normal.3.0.upsample.1.running_mean", "depth_and_inpaint.net1.decoder_normal.3.0.upsample.1.running_var", "depth_and_inpaint.net1.decoder_normal.3.0.upsample.1.num_batches_tracked", "depth_and_inpaint.net1.decoder_normal.3.1.deconv1.weight", "depth_and_inpaint.net1.decoder_normal.3.1.bn1.weight", "depth_and_inpaint.net1.decoder_normal.3.1.bn1.bias", "depth_and_inpaint.net1.decoder_normal.3.1.bn1.running_mean", "depth_and_inpaint.net1.decoder_normal.3.1.bn1.running_var", "depth_and_inpaint.net1.decoder_normal.3.1.bn1.num_batches_tracked", "depth_and_inpaint.net1.decoder_normal.3.1.deconv2.weight", "depth_and_inpaint.net1.decoder_normal.3.1.bn2.weight", "depth_and_inpaint.net1.decoder_normal.3.1.bn2.bias", "depth_and_inpaint.net1.decoder_normal.3.1.bn2.running_mean", "depth_and_inpaint.net1.decoder_normal.3.1.bn2.running_var", "depth_and_inpaint.net1.decoder_normal.3.1.bn2.num_batches_tracked", "depth_and_inpaint.net1.decoder_normal.4.0.weight", "depth_and_inpaint.net1.decoder_normal.4.0.bias", "depth_and_inpaint.net1.decoder_normal.4.1.weight", "depth_and_inpaint.net1.decoder_normal.4.1.bias", "depth_and_inpaint.net1.decoder_normal.4.1.running_mean", "depth_and_inpaint.net1.decoder_normal.4.1.running_var", "depth_and_inpaint.net1.decoder_normal.4.1.num_batches_tracked", "depth_and_inpaint.net1.decoder_normal.4.3.weight", "depth_and_inpaint.net1.decoder_depth.0.0.deconv1.weight", "depth_and_inpaint.net1.decoder_depth.0.0.bn1.weight", "depth_and_inpaint.net1.decoder_depth.0.0.bn1.bias", "depth_and_inpaint.net1.decoder_depth.0.0.bn1.running_mean", "depth_and_inpaint.net1.decoder_depth.0.0.bn1.running_var", "depth_and_inpaint.net1.decoder_depth.0.0.bn1.num_batches_tracked", "depth_and_inpaint.net1.decoder_depth.0.0.deconv2.weight", "depth_and_inpaint.net1.decoder_depth.0.0.bn2.weight", "depth_and_inpaint.net1.decoder_depth.0.0.bn2.bias", "depth_and_inpaint.net1.decoder_depth.0.0.bn2.running_mean", "depth_and_inpaint.net1.decoder_depth.0.0.bn2.running_var", "depth_and_inpaint.net1.decoder_depth.0.0.bn2.num_batches_tracked", "depth_and_inpaint.net1.decoder_depth.0.0.upsample.0.weight", "depth_and_inpaint.net1.decoder_depth.0.0.upsample.1.weight", "depth_and_inpaint.net1.decoder_depth.0.0.upsample.1.bias", "depth_and_inpaint.net1.decoder_depth.0.0.upsample.1.running_mean", "depth_and_inpaint.net1.decoder_depth.0.0.upsample.1.running_var", "depth_and_inpaint.net1.decoder_depth.0.0.upsample.1.num_batches_tracked", "depth_and_inpaint.net1.decoder_depth.0.1.deconv1.weight", "depth_and_inpaint.net1.decoder_depth.0.1.bn1.weight", "depth_and_inpaint.net1.decoder_depth.0.1.bn1.bias", "depth_and_inpaint.net1.decoder_depth.0.1.bn1.running_mean", "depth_and_inpaint.net1.decoder_depth.0.1.bn1.running_var", "depth_and_inpaint.net1.decoder_depth.0.1.bn1.num_batches_tracked", "depth_and_inpaint.net1.decoder_depth.0.1.deconv2.weight", "depth_and_inpaint.net1.decoder_depth.0.1.bn2.weight", "depth_and_inpaint.net1.decoder_depth.0.1.bn2.bias", "depth_and_inpaint.net1.decoder_depth.0.1.bn2.running_mean", "depth_and_inpaint.net1.decoder_depth.0.1.bn2.running_var", "depth_and_inpaint.net1.decoder_depth.0.1.bn2.num_batches_tracked", "depth_and_inpaint.net1.decoder_depth.1.0.deconv1.weight", "depth_and_inpaint.net1.decoder_depth.1.0.bn1.weight", "depth_and_inpaint.net1.decoder_depth.1.0.bn1.bias", "depth_and_inpaint.net1.decoder_depth.1.0.bn1.running_mean", "depth_and_inpaint.net1.decoder_depth.1.0.bn1.running_var", "depth_and_inpaint.net1.decoder_depth.1.0.bn1.num_batches_tracked", "depth_and_inpaint.net1.decoder_depth.1.0.deconv2.weight", "depth_and_inpaint.net1.decoder_depth.1.0.bn2.weight", "depth_and_inpaint.net1.decoder_depth.1.0.bn2.bias", "depth_and_inpaint.net1.decoder_depth.1.0.bn2.running_mean", "depth_and_inpaint.net1.decoder_depth.1.0.bn2.running_var", "depth_and_inpaint.net1.decoder_depth.1.0.bn2.num_batches_tracked", "depth_and_inpaint.net1.decoder_depth.1.0.upsample.0.weight", "depth_and_inpaint.net1.decoder_depth.1.0.upsample.1.weight", "depth_and_inpaint.net1.decoder_depth.1.0.upsample.1.bias", "depth_and_inpaint.net1.decoder_depth.1.0.upsample.1.running_mean", "depth_and_inpaint.net1.decoder_depth.1.0.upsample.1.running_var", "depth_and_inpaint.net1.decoder_depth.1.0.upsample.1.num_batches_tracked", "depth_and_inpaint.net1.decoder_depth.1.1.deconv1.weight", "depth_and_inpaint.net1.decoder_depth.1.1.bn1.weight", "depth_and_inpaint.net1.decoder_depth.1.1.bn1.bias", "depth_and_inpaint.net1.decoder_depth.1.1.bn1.running_mean", "depth_and_inpaint.net1.decoder_depth.1.1.bn1.running_var", "depth_and_inpaint.net1.decoder_depth.1.1.bn1.num_batches_tracked", "depth_and_inpaint.net1.decoder_depth.1.1.deconv2.weight", "depth_and_inpaint.net1.decoder_depth.1.1.bn2.weight", "depth_and_inpaint.net1.decoder_depth.1.1.bn2.bias", "depth_and_inpaint.net1.decoder_depth.1.1.bn2.running_mean", "depth_and_inpaint.net1.decoder_depth.1.1.bn2.running_var", "depth_and_inpaint.net1.decoder_depth.1.1.bn2.num_batches_tracked", "depth_and_inpaint.net1.decoder_depth.2.0.deconv1.weight", "depth_and_inpaint.net1.decoder_depth.2.0.bn1.weight", "depth_and_inpaint.net1.decoder_depth.2.0.bn1.bias", "depth_and_inpaint.net1.decoder_depth.2.0.bn1.running_mean", "depth_and_inpaint.net1.decoder_depth.2.0.bn1.running_var", "depth_and_inpaint.net1.decoder_depth.2.0.bn1.num_batches_tracked", "depth_and_inpaint.net1.decoder_depth.2.0.deconv2.weight", "depth_and_inpaint.net1.decoder_depth.2.0.bn2.weight", "depth_and_inpaint.net1.decoder_depth.2.0.bn2.bias", "depth_and_inpaint.net1.decoder_depth.2.0.bn2.running_mean", "depth_and_inpaint.net1.decoder_depth.2.0.bn2.running_var", "depth_and_inpaint.net1.decoder_depth.2.0.bn2.num_batches_tracked", "depth_and_inpaint.net1.decoder_depth.2.0.upsample.0.weight", "depth_and_inpaint.net1.decoder_depth.2.0.upsample.1.weight", "depth_and_inpaint.net1.decoder_depth.2.0.upsample.1.bias", "depth_and_inpaint.net1.decoder_depth.2.0.upsample.1.running_mean", "depth_and_inpaint.net1.decoder_depth.2.0.upsample.1.running_var", "depth_and_inpaint.net1.decoder_depth.2.0.upsample.1.num_batches_tracked", "depth_and_inpaint.net1.decoder_depth.2.1.deconv1.weight", "depth_and_inpaint.net1.decoder_depth.2.1.bn1.weight", "depth_and_inpaint.net1.decoder_depth.2.1.bn1.bias", "depth_and_inpaint.net1.decoder_depth.2.1.bn1.running_mean", "depth_and_inpaint.net1.decoder_depth.2.1.bn1.running_var", "depth_and_inpaint.net1.decoder_depth.2.1.bn1.num_batches_tracked", "depth_and_inpaint.net1.decoder_depth.2.1.deconv2.weight", "depth_and_inpaint.net1.decoder_depth.2.1.bn2.weight", "depth_and_inpaint.net1.decoder_depth.2.1.bn2.bias", "depth_and_inpaint.net1.decoder_depth.2.1.bn2.running_mean", "depth_and_inpaint.net1.decoder_depth.2.1.bn2.running_var", "depth_and_inpaint.net1.decoder_depth.2.1.bn2.num_batches_tracked", "depth_and_inpaint.net1.decoder_depth.3.0.deconv1.weight", "depth_and_inpaint.net1.decoder_depth.3.0.bn1.weight", "depth_and_inpaint.net1.decoder_depth.3.0.bn1.bias", "depth_and_inpaint.net1.decoder_depth.3.0.bn1.running_mean", "depth_and_inpaint.net1.decoder_depth.3.0.bn1.running_var", "depth_and_inpaint.net1.decoder_depth.3.0.bn1.num_batches_tracked", "depth_and_inpaint.net1.decoder_depth.3.0.deconv2.weight", "depth_and_inpaint.net1.decoder_depth.3.0.bn2.weight", "depth_and_inpaint.net1.decoder_depth.3.0.bn2.bias", "depth_and_inpaint.net1.decoder_depth.3.0.bn2.running_mean", "depth_and_inpaint.net1.decoder_depth.3.0.bn2.running_var", "depth_and_inpaint.net1.decoder_depth.3.0.bn2.num_batches_tracked", "depth_and_inpaint.net1.decoder_depth.3.0.upsample.0.weight", "depth_and_inpaint.net1.decoder_depth.3.0.upsample.1.weight", "depth_and_inpaint.net1.decoder_depth.3.0.upsample.1.bias", "depth_and_inpaint.net1.decoder_depth.3.0.upsample.1.running_mean", "depth_and_inpaint.net1.decoder_depth.3.0.upsample.1.running_var", "depth_and_inpaint.net1.decoder_depth.3.0.upsample.1.num_batches_tracked", "depth_and_inpaint.net1.decoder_depth.3.1.deconv1.weight", "depth_and_inpaint.net1.decoder_depth.3.1.bn1.weight", "depth_and_inpaint.net1.decoder_depth.3.1.bn1.bias", "depth_and_inpaint.net1.decoder_depth.3.1.bn1.running_mean", "depth_and_inpaint.net1.decoder_depth.3.1.bn1.running_var", "depth_and_inpaint.net1.decoder_depth.3.1.bn1.num_batches_tracked", "depth_and_inpaint.net1.decoder_depth.3.1.deconv2.weight", "depth_and_inpaint.net1.decoder_depth.3.1.bn2.weight", "depth_and_inpaint.net1.decoder_depth.3.1.bn2.bias", "depth_and_inpaint.net1.decoder_depth.3.1.bn2.running_mean", "depth_and_inpaint.net1.decoder_depth.3.1.bn2.running_var", "depth_and_inpaint.net1.decoder_depth.3.1.bn2.num_batches_tracked", "depth_and_inpaint.net1.decoder_depth.4.0.weight", "depth_and_inpaint.net1.decoder_depth.4.0.bias", "depth_and_inpaint.net1.decoder_depth.4.1.weight", "depth_and_inpaint.net1.decoder_depth.4.1.bias", "depth_and_inpaint.net1.decoder_depth.4.1.running_mean", "depth_and_inpaint.net1.decoder_depth.4.1.running_var", "depth_and_inpaint.net1.decoder_depth.4.1.num_batches_tracked", "depth_and_inpaint.net1.decoder_depth.4.3.weight", "depth_and_inpaint.net1.decoder_silhou.0.0.deconv1.weight", "depth_and_inpaint.net1.decoder_silhou.0.0.bn1.weight", "depth_and_inpaint.net1.decoder_silhou.0.0.bn1.bias", "depth_and_inpaint.net1.decoder_silhou.0.0.bn1.running_mean", "depth_and_inpaint.net1.decoder_silhou.0.0.bn1.running_var", "depth_and_inpaint.net1.decoder_silhou.0.0.bn1.num_batches_tracked", "depth_and_inpaint.net1.decoder_silhou.0.0.deconv2.weight", "depth_and_inpaint.net1.decoder_silhou.0.0.bn2.weight", "depth_and_inpaint.net1.decoder_silhou.0.0.bn2.bias", "depth_and_inpaint.net1.decoder_silhou.0.0.bn2.running_mean", "depth_and_inpaint.net1.decoder_silhou.0.0.bn2.running_var", "depth_and_inpaint.net1.decoder_silhou.0.0.bn2.num_batches_tracked", "depth_and_inpaint.net1.decoder_silhou.0.0.upsample.0.weight", "depth_and_inpaint.net1.decoder_silhou.0.0.upsample.1.weight", "depth_and_inpaint.net1.decoder_silhou.0.0.upsample.1.bias", "depth_and_inpaint.net1.decoder_silhou.0.0.upsample.1.running_mean", "depth_and_inpaint.net1.decoder_silhou.0.0.upsample.1.running_var", "depth_and_inpaint.net1.decoder_silhou.0.0.upsample.1.num_batches_tracked", "depth_and_inpaint.net1.decoder_silhou.0.1.deconv1.weight", "depth_and_inpaint.net1.decoder_silhou.0.1.bn1.weight", "depth_and_inpaint.net1.decoder_silhou.0.1.bn1.bias", "depth_and_inpaint.net1.decoder_silhou.0.1.bn1.running_mean", "depth_and_inpaint.net1.decoder_silhou.0.1.bn1.running_var", "depth_and_inpaint.net1.decoder_silhou.0.1.bn1.num_batches_tracked", "depth_and_inpaint.net1.decoder_silhou.0.1.deconv2.weight", "depth_and_inpaint.net1.decoder_silhou.0.1.bn2.weight", "depth_and_inpaint.net1.decoder_silhou.0.1.bn2.bias", "depth_and_inpaint.net1.decoder_silhou.0.1.bn2.running_mean", "depth_and_inpaint.net1.decoder_silhou.0.1.bn2.running_var", "depth_and_inpaint.net1.decoder_silhou.0.1.bn2.num_batches_tracked", "depth_and_inpaint.net1.decoder_silhou.1.0.deconv1.weight", "depth_and_inpaint.net1.decoder_silhou.1.0.bn1.weight", "depth_and_inpaint.net1.decoder_silhou.1.0.bn1.bias", "depth_and_inpaint.net1.decoder_silhou.1.0.bn1.running_mean", "depth_and_inpaint.net1.decoder_silhou.1.0.bn1.running_var", "depth_and_inpaint.net1.decoder_silhou.1.0.bn1.num_batches_tracked", "depth_and_inpaint.net1.decoder_silhou.1.0.deconv2.weight", "depth_and_inpaint.net1.decoder_silhou.1.0.bn2.weight", "depth_and_inpaint.net1.decoder_silhou.1.0.bn2.bias", "depth_and_inpaint.net1.decoder_silhou.1.0.bn2.running_mean", "depth_and_inpaint.net1.decoder_silhou.1.0.bn2.running_var", "depth_and_inpaint.net1.decoder_silhou.1.0.bn2.num_batches_tracked", "depth_and_inpaint.net1.decoder_silhou.1.0.upsample.0.weight", "depth_and_inpaint.net1.decoder_silhou.1.0.upsample.1.weight", "depth_and_inpaint.net1.decoder_silhou.1.0.upsample.1.bias", "depth_and_inpaint.net1.decoder_silhou.1.0.upsample.1.running_mean", "depth_and_inpaint.net1.decoder_silhou.1.0.upsample.1.running_var", "depth_and_inpaint.net1.decoder_silhou.1.0.upsample.1.num_batches_tracked", "depth_and_inpaint.net1.decoder_silhou.1.1.deconv1.weight", "depth_and_inpaint.net1.decoder_silhou.1.1.bn1.weight", "depth_and_inpaint.net1.decoder_silhou.1.1.bn1.bias", "depth_and_inpaint.net1.decoder_silhou.1.1.bn1.running_mean", "depth_and_inpaint.net1.decoder_silhou.1.1.bn1.running_var", "depth_and_inpaint.net1.decoder_silhou.1.1.bn1.num_batches_tracked", "depth_and_inpaint.net1.decoder_silhou.1.1.deconv2.weight", "depth_and_inpaint.net1.decoder_silhou.1.1.bn2.weight", "depth_and_inpaint.net1.decoder_silhou.1.1.bn2.bias", "depth_and_inpaint.net1.decoder_silhou.1.1.bn2.running_mean", "depth_and_inpaint.net1.decoder_silhou.1.1.bn2.running_var", "depth_and_inpaint.net1.decoder_silhou.1.1.bn2.num_batches_tracked", "depth_and_inpaint.net1.decoder_silhou.2.0.deconv1.weight", "depth_and_inpaint.net1.decoder_silhou.2.0.bn1.weight", "depth_and_inpaint.net1.decoder_silhou.2.0.bn1.bias", "depth_and_inpaint.net1.decoder_silhou.2.0.bn1.running_mean", "depth_and_inpaint.net1.decoder_silhou.2.0.bn1.running_var", "depth_and_inpaint.net1.decoder_silhou.2.0.bn1.num_batches_tracked", "depth_and_inpaint.net1.decoder_silhou.2.0.deconv2.weight", "depth_and_inpaint.net1.decoder_silhou.2.0.bn2.weight", "depth_and_inpaint.net1.decoder_silhou.2.0.bn2.bias", "depth_and_inpaint.net1.decoder_silhou.2.0.bn2.running_mean", "depth_and_inpaint.net1.decoder_silhou.2.0.bn2.running_var", "depth_and_inpaint.net1.decoder_silhou.2.0.bn2.num_batches_tracked", "depth_and_inpaint.net1.decoder_silhou.2.0.upsample.0.weight", "depth_and_inpaint.net1.decoder_silhou.2.0.upsample.1.weight", "depth_and_inpaint.net1.decoder_silhou.2.0.upsample.1.bias", "depth_and_inpaint.net1.decoder_silhou.2.0.upsample.1.running_mean", "depth_and_inpaint.net1.decoder_silhou.2.0.upsample.1.running_var", "depth_and_inpaint.net1.decoder_silhou.2.0.upsample.1.num_batches_tracked", "depth_and_inpaint.net1.decoder_silhou.2.1.deconv1.weight", "depth_and_inpaint.net1.decoder_silhou.2.1.bn1.weight", "depth_and_inpaint.net1.decoder_silhou.2.1.bn1.bias", "depth_and_inpaint.net1.decoder_silhou.2.1.bn1.running_mean", "depth_and_inpaint.net1.decoder_silhou.2.1.bn1.running_var", "depth_and_inpaint.net1.decoder_silhou.2.1.bn1.num_batches_tracked", "depth_and_inpaint.net1.decoder_silhou.2.1.deconv2.weight", "depth_and_inpaint.net1.decoder_silhou.2.1.bn2.weight", "depth_and_inpaint.net1.decoder_silhou.2.1.bn2.bias", "depth_and_inpaint.net1.decoder_silhou.2.1.bn2.running_mean", "depth_and_inpaint.net1.decoder_silhou.2.1.bn2.running_var", "depth_and_inpaint.net1.decoder_silhou.2.1.bn2.num_batches_tracked", "depth_and_inpaint.net1.decoder_silhou.3.0.deconv1.weight", "depth_and_inpaint.net1.decoder_silhou.3.0.bn1.weight", "depth_and_inpaint.net1.decoder_silhou.3.0.bn1.bias", "depth_and_inpaint.net1.decoder_silhou.3.0.bn1.running_mean", "depth_and_inpaint.net1.decoder_silhou.3.0.bn1.running_var", "depth_and_inpaint.net1.decoder_silhou.3.0.bn1.num_batches_tracked", "depth_and_inpaint.net1.decoder_silhou.3.0.deconv2.weight", "depth_and_inpaint.net1.decoder_silhou.3.0.bn2.weight", "depth_and_inpaint.net1.decoder_silhou.3.0.bn2.bias", "depth_and_inpaint.net1.decoder_silhou.3.0.bn2.running_mean", "depth_and_inpaint.net1.decoder_silhou.3.0.bn2.running_var", "depth_and_inpaint.net1.decoder_silhou.3.0.bn2.num_batches_tracked", "depth_and_inpaint.net1.decoder_silhou.3.0.upsample.0.weight", "depth_and_inpaint.net1.decoder_silhou.3.0.upsample.1.weight", "depth_and_inpaint.net1.decoder_silhou.3.0.upsample.1.bias", "depth_and_inpaint.net1.decoder_silhou.3.0.upsample.1.running_mean", "depth_and_inpaint.net1.decoder_silhou.3.0.upsample.1.running_var", "depth_and_inpaint.net1.decoder_silhou.3.0.upsample.1.num_batches_tracked", "depth_and_inpaint.net1.decoder_silhou.3.1.deconv1.weight", "depth_and_inpaint.net1.decoder_silhou.3.1.bn1.weight", "depth_and_inpaint.net1.decoder_silhou.3.1.bn1.bias", "depth_and_inpaint.net1.decoder_silhou.3.1.bn1.running_mean", "depth_and_inpaint.net1.decoder_silhou.3.1.bn1.running_var", "depth_and_inpaint.net1.decoder_silhou.3.1.bn1.num_batches_tracked", "depth_and_inpaint.net1.decoder_silhou.3.1.deconv2.weight", "depth_and_inpaint.net1.decoder_silhou.3.1.bn2.weight", "depth_and_inpaint.net1.decoder_silhou.3.1.bn2.bias", "depth_and_inpaint.net1.decoder_silhou.3.1.bn2.running_mean", "depth_and_inpaint.net1.decoder_silhou.3.1.bn2.running_var", "depth_and_inpaint.net1.decoder_silhou.3.1.bn2.num_batches_tracked", "depth_and_inpaint.net1.decoder_silhou.4.0.weight", "depth_and_inpaint.net1.decoder_silhou.4.0.bias", "depth_and_inpaint.net1.decoder_silhou.4.1.weight", "depth_and_inpaint.net1.decoder_silhou.4.1.bias", "depth_and_inpaint.net1.decoder_silhou.4.1.running_mean", "depth_and_inpaint.net1.decoder_silhou.4.1.running_var", "depth_and_inpaint.net1.decoder_silhou.4.1.num_batches_tracked", "depth_and_inpaint.net1.decoder_silhou.4.3.weight", "depth_and_inpaint.net1.decoder_minmax.0.weight", "depth_and_inpaint.net1.decoder_minmax.0.bias", "depth_and_inpaint.net1.decoder_minmax.1.weight", "depth_and_inpaint.net1.decoder_minmax.1.bias", "depth_and_inpaint.net1.decoder_minmax.3.weight", "depth_and_inpaint.net1.decoder_minmax.3.bias", "depth_and_inpaint.net1.decoder_minmax.4.weight", "depth_and_inpaint.net1.decoder_minmax.4.bias", "depth_and_inpaint.net1.decoder_minmax.4.running_mean", "depth_and_inpaint.net1.decoder_minmax.4.running_var", "depth_and_inpaint.net1.decoder_minmax.4.num_batches_tracked", "depth_and_inpaint.net1.decoder_minmax.6.weight", "depth_and_inpaint.net1.decoder_minmax.6.bias", "depth_and_inpaint.net1.decoder_minmax.7.weight", "depth_and_inpaint.net1.decoder_minmax.7.bias", "depth_and_inpaint.net1.decoder_minmax.7.running_mean", "depth_and_inpaint.net1.decoder_minmax.7.running_var", "depth_and_inpaint.net1.decoder_minmax.7.num_batches_tracked", "depth_and_inpaint.net1.decoder_minmax.9.weight", "depth_and_inpaint.net1.decoder_minmax.9.bias".

As mentioned in the ShapeHD paper, ``to boost the realism
of the rendered RGB images, we put three different types of backgrounds behind
the object during rendering. One third of the images are rendered in a clean white
background; one third are rendered in high-dynamic-range backgrounds with
illumination channels that produce realistic lighting. We render the remaining one
third images with backgrounds randomly sampled from the SUN database [61].''

I wonder how to render images with high-dynamic-range backgrounds with illumination channels.

Hi Xiuming,

I'd like to test the ShapeHD model on airplanes? Is the pre-trained model on airplanes available?

Thanks!

Hi, I noticed that the IoU number under "4.1 3D Reconstruction on ShapeNet" section in MarrNet paper is 0.57, what class is this IoU for? Or is it the average IoU over the chair, plane and car classes? The best IoU I can get with MarrNet on the chair class is around 0.45, is this IoU reasonable since Reprojection consistency loss is not needed in synthetic data?

By the way, the IoU in ShapeHD paper on chair class is 0.488 which is lower than MarrNet, shouldn't it be higher since ShapeHD generates better reconstructions? Or maybe it is supposed to be lower because of the naturalness loss? Or we cannot just simply compare MarrNet and ShapeHD?

Thanks

Hi, there.
Thank you for sharing.

I tried to build the cuda extension using "./build_toolbox.sh", but it failed....

(shaperecon) root@d31b5ed138c4:~/GenRe-ShapeHD# ./build_toolbox.sh           
Add -gencode to match all the GPU architectures you have.
Check 'https://en.wikipedia.org/wiki/CUDA#GPUs_supported' for list of architecture.
Check 'http://docs.nvidia.com/cuda/cuda-compiler-driver-nvcc/index.html' for GPU compilation based on architecture.
nvcc -c -o calc_prob_kernel.cu.o calc_prob_kernel.cu -x cu -Xcompiler -std=c++0x -fPIC -I /root/anaconda3/envs/shaperecon/lib/python3.6/site-packages/torch/lib/include -I /root/anaconda3/envs/shaperecon/lib/python3.6/site-packages/torch/lib/include/TH -I /root/anaconda3/envs/shaperecon/lib/python3.6/site-packages/torch/lib/include/THC -I /root/GenRe-ShapeHD/toolbox/calc_prob/calc_prob/src         -gencode arch=compute_30,code=sm_30         -gencode arch=compute_35,code=sm_35         -gencode arch=compute_52,code=sm_52         -gencode arch=compute_61,code=sm_61 
nvcc fatal   : Unknown option 'fPIC'
/root/GenRe-ShapeHD/toolbox/calc_prob
/root/anaconda3/envs/shaperecon/lib/python3.6/distutils/extension.py:131: UserWarning: Unknown Extension options: 'headers', 'package', 'relative_to', 'with_cuda'
  warnings.warn(msg)
Traceback (most recent call last):
  File "build.py", line 42, in <module>
    BuildExtension(ext)
  File "/root/anaconda3/envs/shaperecon/lib/python3.6/site-packages/setuptools/__init__.py", line 163, in __init__
    _Command.__init__(self, dist)
  File "/root/anaconda3/envs/shaperecon/lib/python3.6/distutils/cmd.py", line 57, in __init__
    raise TypeError("dist must be a Distribution instance")
TypeError: dist must be a Distribution instance
Add -gencode to match all the GPU architectures you have.
Check 'https://en.wikipedia.org/wiki/CUDA#GPUs_supported' for list of architecture.
......

Is there any way to solve this problem?
Thank you !

Hi,

Is it possible to have the pascal 3D+ dataset you used for testing? Especially the rgb and corresponding voxels? Thanks!

Hello, Xiuming. I ran codes following your steps, while when I ran ./build_toolbox.sh to build the environment, errors occurred as follows. Could you please help me solve this problem?

Traceback (most recent call last):
  File "<string>", line 1, in <module>
  File "/home/fxy/anaconda3/envs/shaperecon/lib/python3.6/site-packages/torch/__init__.py", line 80, in <module>
    from torch._C import *
ImportError: numpy.core.multiarray failed to import

Thanks a lot!

I get the ShapeHD result of the three in download folder, but could you tell me how to visualization the results as what you show in the picture ?

I download shapenet_cars_chairs_planes_20views.tar,and I want to generate other category 2.5D sketch from the RGB(like 02691156_1a04e3eab45ca15dd86060f189eb133_view000_spherical.npz), is there any download link,or how can I generate them?

Hi there,

When Training marrnet1, how can I pass multiple classes. ?
(i.e. --classes [ ] )
Cheers

Hi,

I am trying to retrain marrnet2+3D-GAN and fine-tune them. But it seems each epoch takes ~1500s, if train 200epch, it would take ~100h only for Marrnet2. It seems 3D-GAN would take much longer time. Did I miss something or is there any method to speed up the training? Thank you so much.

Hi, I defined some the Net class like the following, after training, I found out that the parameters of some networks I defined, for example self.fc and self.classifier, in Net class are not saved because when I resume training, the loss and accuracy didn't start around where it stopped, do you have any suggestions where to look at?

class Net(nn.Module):
    def __init__(self, npf, npc, ndims):
        super().__init__()
        self.Encoder = VoxelEncoder(ndims=ndims).cuda()
        self.classifier = {}
        self.fc = FC(in_dims=ndims, out_dims=100).cuda()
        for i in range(2):
            self.classifier[i] = ClassificationLayer(100, npc[i]).cuda()