Pytorch 1.6 and lib knn build with cuda 10.2

[drapado]

I changed the Pytorch-1.0 branch to make it work with Pytorch 1.6.0 and CUDA 10.2 (on python 3.6.11).

To install you'll need to:

1. Download the code with these changes
2. Open a terminal and go to the densefusion directory
3. Run the following commands:

pip install opencv-python scipy pyyaml
cd lib/knn
python setup.py install
cd dist
unzip knn_pytorch-0.1-py3.6-linux-x86_64.egg 
cp knn_pytorch/knn_pytorch.py ../knn_pytorch.py
cp knn_pytorch/knn_pytorch.cpython-36m-x86_64-linux-gnu.so ../knn_pytorch.cpython-36m-x86_64-linux-gnu.so
cd ../../..

The changes are:

1. Line 38 on lib/knn/src/knn.h since THCState_getCurrentStream(state) is deprecated (see here and here)
2. Adapted class KNearestNeighbor from /lib/knn/init.py and /lib/loss.py accordingly to solve the issue Legacy autograd function with non-static forward method is deprecated following the documentation

[wjalex2000]

Hello, I have a question about this. When I applied the changes and ran the commands, I still get an error saying that

File "./tools/train.py", line 140, in main
loss, dis, new_points, new_target = criterion(pred_r, pred_t, pred_c, target, model_points, idx, points, opt.w, opt.refine_start)
File "/home/motion/anaconda3/envs/wj6d_env/lib/python3.6/site-packages/torch/nn/modules/module.py", line 722, in _call_impl
result = self.forward(*input, **kwargs)
File "/home/motion/anaconda3/envs/wj6d_env/lib/python3.6/site-packages/torch/nn/modules/module.py", line 223, in _forward_unimplemented
raise NotImplementedError
NotImplementedError

Is it possible to get some help with this?
Thank you!

[drapado]

Mmm, are you sure you are using pytorch 1.6? The error seems to point to a non existing __init__ on class KNearestNeighbor(Function):. But from pytorch 1.3 you don't need it anymore because you use an static forward I think.

Also, are you using LineMOD or YCB?

[wjalex2000]

Hi! That's weird... I just double-checked my pytorch version and I'm using 1.6.0. I'm using LineMOD right now. Again, thank you so much,

[drapado]

TBH I haven't checked with LineMOD, I tested the code with YCB so maybe there's a problem there. I'll try to test in LineMOD if I have some extra time

[wjalex2000]

Hmm... Ok, I'll try with YCB too. Let me know if you get time to try it! Thank you!

[emigmo]

great

[a-pizzi]

Hello, i followed step by step the instructions provided, but when the .egg file is generated it does not contain the file named "knn_pytorch.cpython-36m-x86_64-linux-gnu.so", but instead it's created a file named "knn_pytorch.so" that weights 3.5MB; i tried to rename it so to verify if it was just an error related to the name, but if i use it as .so file it returns me an error: unrecognized statement inside the .so file. I'm using pytorch 1.0.0, torchvision 0.2.2 and python 2.7.17 on an ubuntu environment.
The procedure i follow is:

• clone the repository
• install all the dependencies
• download the datasets
• follow the above instructions (provided by @drapado)

What am i doing wrong?

Thank you!

[gntoni]

loss_refiner.py should be modified similarly to loss.py

index 6496c9a..2e1b21c 100755
--- a/lib/loss_refiner.py
+++ b/lib/loss_refiner.py
@@ -10,7 +10,7 @@ from lib.knn.__init__ import KNearestNeighbor
 
 
 def loss_calculation(pred_r, pred_t, target, model_points, idx, points, num_point_mesh, sym_list):
-    knn = KNearestNeighbor(1)
+    #knn = KNearestNeighbor(1)
     pred_r = pred_r.view(1, 1, -1)
     pred_t = pred_t.view(1, 1, -1)
     bs, num_p, _ = pred_r.size()
@@ -41,7 +41,7 @@ def loss_calculation(pred_r, pred_t, target, model_points, idx, points, num_poin
     if idx[0].item() in sym_list:
         target = target[0].transpose(1, 0).contiguous().view(3, -1)
         pred = pred.permute(2, 0, 1).contiguous().view(3, -1)
-        inds = knn(target.unsqueeze(0), pred.unsqueeze(0))
+        inds = KNearestNeighbor.apply(target.unsqueeze(0), pred.unsqueeze(0))
         target = torch.index_select(target, 1, inds.view(-1) - 1)
         target = target.view(3, bs * num_p, num_point_mesh).permute(1, 2, 0).contiguous()
         pred = pred.view(3, bs * num_p, num_point_mesh).permute(1, 2, 0).contiguous()
@@ -60,7 +60,7 @@ def loss_calculation(pred_r, pred_t, target, model_points, idx, points, num_poin
     new_target = torch.bmm((new_target - ori_t), ori_base).contiguous()
 
     # print('------------> ', dis.item(), idx[0].item())
-    del knn
+    #del knn
     return dis, new_points.detach(), new_target.detach()