xt-chen / sarpn Goto Github PK

Hi @Xt-Chen ,
Thanks for your excellent work, but when I train your model on my own datasets, the losses are less than zero, could you give me some advice?

Processing the 501th image!
1: tensor([[[[0.1345, 0.1318, 0.1291, ..., 0.3593, 0.3570, 0.3546],
[0.1415, 0.1398, 0.1381, ..., 0.3608, 0.3596, 0.3584],
[0.1452, 0.1434, 0.1415, ..., 0.3629, 0.3608, 0.3616],
...,
[0.8342, 0.5683, 0.4067, ..., 0.2935, 0.2923, 0.2885],
[0.4220, 0.4454, 0.4706, ..., 0.3205, 0.3082, 0.2965],
[0.2269, 0.3603, 0.5547, ..., 0.3263, 0.3047, 0.2843]]]],
device='cuda:0', grad_fn=)
2: tensor(nan, device='cuda:0', grad_fn=)
3: tensor(16384., device='cuda:0')
4

Thank you for your great work. But when trying to train the model to reproduce your result, I met a problem.

line 201, in colormap
color_map = colors[indices].transpose(2, 3).transpose(1, 2)
RuntimeError: Dimension out of range (expected to be in range of [-3, 2], but got 3)

I do not know how to fix it. Looking forward to hear from you! Thank you so much!

In line 48 of train.py, your code shows loadckpt = os.path.join(args.logdir, all_saved_ckpts[-1])
Maybe loadckpt = os.path.join(args.checkpoint_dir, all_saved_ckpts[-1]) is right. I am not sure.

Thank for your great work!
I am a little confused about the dataloader code! Why divide the depth groundtruth by 255 and multiply it by 10 in training and divide it by 1000 in evaluate? Will this affect the evaluation results? If i need adjust it?

Dear @Xt-Chen, i got this problem how i can solve it, by the way i decreased Batch_size until 1 but does not solve
give me an Error on this line:
loading model ./checkpoints/SARPN_checkpoints_20.pth.tar
Processing the 0th image!
Traceback (most recent call last):
File "evaluate.py", line 95, in
test()
File "evaluate.py", line 52, in test
pred_depth = model(image)

RuntimeError: CUDA out of memory. Tried to allocate 22.00 MiB (GPU 0; 8.00 GiB total capacity; 1.27 GiB already allocated; 0 bytes free; 1.30 GiB reserved in total by PyTorch)

Hi Xt-Chen,

Thanks for your nice work. I find it interesting, so I cloned the code and re-trained on my PC without any changes except the dataset path. However, I tried several times, the results varies a lot. Take the ABS metric as an example, the results were 0.2%-0.4% worse than the reported results in your paper. I checked the code and could not find any operations to fix the rand seeds. I thought this might be the main potential issue that I could not reproduce the result.
As above, Would you please help me to cross the problem and I wonder if you tried many times and chosen the best one? or you just trained one time?

Thanks again.

Cheers,
cvgogogo

Hello, Xt-Chen.
I really thank you for releasing the SARPN codes for monocular depth estimation.

Unfortunately, the pretrained model that is uploaded on the MS OneDrive may be broken.
when trying to decompress the tar file, the file can't be decompressed.

Please check it one more time.
Thanks.

I did not find any files about get_models. But net.py need this file, where can I find it?

I want to know what is the definition of adff_num_features, and how is 1280 calculated?

你好 我在复现这篇论文的时候损失函数总是降不下去，使用的pytorch是1.1.0版本的

Is there any possible to estimate distance?

I can't Download Nyu v2 DataSet maybe the file is damaged, I will appreciate it if you upload it in the other way instead of google drive. Or you can send it to [email protected]. Thanks a lot!