Cpt.Captain
/
connecting_the_dots
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

added validation set and pre-trained model; tested on pytorch1.8

Browse Source
master
Yiyi Liao 4 years ago
parent bdcc50a88f
commit 4617525f61
5 changed files with 68 additions and 60 deletions
Show all changes Ignore whitespace when comparing lines Ignore changes in amount of whitespace Ignore changes in whitespace at EOL
Show Stats Download Patch File Download Diff File
  1. 13
      readme.md
  2. 2
      renderer/Makefile
  3. 56
      torchext/ext/ext_cpu.cpp
  4. 54
      torchext/ext/ext_cuda.cpp
  5. 3
      torchext/setup.py

13
readme.md
Unescape View File

@ -32,6 +32,7 @@ The network training/evaluation code is based on `Pytorch`.
PyTorch>=1.1 PyTorch>=1.1
Cuda>=10.0 Cuda>=10.0
``` ```
Updated on 07.06.2021: The code is now compatible with the latest Pytorch version (1.8).
The other python packages can be installed with `anaconda`: The other python packages can be installed with `anaconda`:
``` ```
@ -65,11 +66,12 @@ python setup.py build_ext --inplace
## Running ## Running
### Creating synthetic data ### Creating Synthetic Data
To create synthetic data and save it locally, download [ShapeNet V2](https://www.shapenet.org/) and correct `SHAPENET_ROOT` in `config.json`. Then the data can be generated and saved to `DATA_ROOT` in `config.json` by running To create synthetic data and save it locally, download [ShapeNet V2](https://www.shapenet.org/) and correct `SHAPENET_ROOT` in `config.json`. Then the data can be generated and saved to `DATA_ROOT` in `config.json` by running
``` ```
./create_syn_data.sh ./create_syn_data.sh
``` ```
If you are only interested in evaluating our pre-trained model, [here (3.7G)](https://s3.eu-central-1.amazonaws.com/avg-projects/connecting_the_dots/val_data.zip) is a validation set that contains a small amount of images.
### Training Network ### Training Network
@ -90,6 +92,15 @@ To evaluate a specific checkpoint, e.g. the 50th epoch, one can run
python train_val.py --cmd retest --epoch 50 python train_val.py --cmd retest --epoch 50
``` ```
### Evaluating a Pre-trained Model
We provide a model pre-trained using the photometric loss. Once you have prepared the synthetic dataset and changed `DATA_ROOT` in `config.json`, the pre-trained model can be evaluated on the validation set by running:
```
mkdir -p output
mkdir -p output/exp_syn
wget -O output/exp_syn/net_0099.params https://s3.eu-central-1.amazonaws.com/avg-projects/connecting_the_dots/net_0099.params
python train_val.py --cmd retest --epoch 99
```
You can also download our validation set from [here (3.7G)](https://s3.eu-central-1.amazonaws.com/avg-projects/connecting_the_dots/val_data.zip).
## Acknowledgement ## Acknowledgement
This work was supported by the Intel Network on Intelligent Systems. This work was supported by the Intel Network on Intelligent Systems.

2
renderer/Makefile
Unescape View File

@ -4,7 +4,7 @@ C_FLAGS = -O3 -msse -msse2 -msse3 -msse4.2 -fPIC -Wall
CXX = g++ -c CXX = g++ -c
CXX_FLAGS = -O3 -std=c++11 -msse -msse2 -msse3 -msse4.2 -fPIC -Wall CXX_FLAGS = -O3 -std=c++11 -msse -msse2 -msse3 -msse4.2 -fPIC -Wall
CUDA = nvcc -c CUDA = nvcc -c
CUDA_FLAGS = -x cu -Xcompiler -fPIC -arch=sm_30 -std=c++11 --expt-extended-lambda CUDA_FLAGS = -x cu -Xcompiler -fPIC -std=c++11 --expt-extended-lambda
PYRENDER_DEPENDENCIES = setup.py \ PYRENDER_DEPENDENCIES = setup.py \

56
torchext/ext/ext_cpu.cpp
Unescape View File

@ -12,17 +12,17 @@ void iterate_cpu(FunctorT functor, int N) {
} }
at::Tensor nn_cpu(at::Tensor in0, at::Tensor in1) { at::Tensor nn_cpu(at::Tensor in0, at::Tensor in1) {
CHECK_INPUT_CPU(in0) CHECK_INPUT_CPU(in0);
CHECK_INPUT_CPU(in1) CHECK_INPUT_CPU(in1);
auto nelem0 = in0.size(0); auto nelem0 = in0.size(0);
auto nelem1 = in1.size(0); auto nelem1 = in1.size(0);
auto dim = in0.size(1); auto dim = in0.size(1);
AT_ASSERTM(dim == in1.size(1), "in0 and in1 have to be the same shape") AT_ASSERTM(dim == in1.size(1), "in0 and in1 have to be the same shape");
AT_ASSERTM(dim == 3, "dim hast to be 3") AT_ASSERTM(dim == 3, "dim hast to be 3");
AT_ASSERTM(in0.dim() == 2, "in0 has to be N0 x 3") AT_ASSERTM(in0.dim() == 2, "in0 has to be N0 x 3");
AT_ASSERTM(in1.dim() == 2, "in1 has to be N1 x 3") AT_ASSERTM(in1.dim() == 2, "in1 has to be N1 x 3");
auto out = at::empty({nelem0}, torch::CPU(at::kLong)); auto out = at::empty({nelem0}, torch::CPU(at::kLong));
@ -37,11 +37,11 @@ at::Tensor nn_cpu(at::Tensor in0, at::Tensor in1) {
at::Tensor crosscheck_cpu(at::Tensor in0, at::Tensor in1) { at::Tensor crosscheck_cpu(at::Tensor in0, at::Tensor in1) {
CHECK_INPUT_CPU(in0) CHECK_INPUT_CPU(in0);
CHECK_INPUT_CPU(in1) CHECK_INPUT_CPU(in1);
AT_ASSERTM(in0.dim() == 1, "") AT_ASSERTM(in0.dim() == 1, "");
AT_ASSERTM(in1.dim() == 1, "") AT_ASSERTM(in1.dim() == 1, "");
auto nelem0 = in0.size(0); auto nelem0 = in0.size(0);
auto nelem1 = in1.size(0); auto nelem1 = in1.size(0);
@ -57,21 +57,21 @@ at::Tensor crosscheck_cpu(at::Tensor in0, at::Tensor in1) {
at::Tensor proj_nn_cpu(at::Tensor xyz0, at::Tensor xyz1, at::Tensor K, int patch_size) { at::Tensor proj_nn_cpu(at::Tensor xyz0, at::Tensor xyz1, at::Tensor K, int patch_size) {
CHECK_INPUT_CPU(xyz0) CHECK_INPUT_CPU(xyz0);
CHECK_INPUT_CPU(xyz1) CHECK_INPUT_CPU(xyz1);
CHECK_INPUT_CPU(K) CHECK_INPUT_CPU(K);
auto batch_size = xyz0.size(0); auto batch_size = xyz0.size(0);
auto height = xyz0.size(1); auto height = xyz0.size(1);
auto width = xyz0.size(2); auto width = xyz0.size(2);
AT_ASSERTM(xyz0.size(0) == xyz1.size(0), "") AT_ASSERTM(xyz0.size(0) == xyz1.size(0), "");
AT_ASSERTM(xyz0.size(1) == xyz1.size(1), "") AT_ASSERTM(xyz0.size(1) == xyz1.size(1), "");
AT_ASSERTM(xyz0.size(2) == xyz1.size(2), "") AT_ASSERTM(xyz0.size(2) == xyz1.size(2), "");
AT_ASSERTM(xyz0.size(3) == xyz1.size(3), "") AT_ASSERTM(xyz0.size(3) == xyz1.size(3), "");
AT_ASSERTM(xyz0.size(3) == 3, "") AT_ASSERTM(xyz0.size(3) == 3, "");
AT_ASSERTM(xyz0.dim() == 4, "") AT_ASSERTM(xyz0.dim() == 4, "");
AT_ASSERTM(xyz1.dim() == 4, "") AT_ASSERTM(xyz1.dim() == 4, "");
auto out = at::empty({batch_size, height, width}, torch::CPU(at::kLong)); auto out = at::empty({batch_size, height, width}, torch::CPU(at::kLong));
@ -86,8 +86,8 @@ at::Tensor proj_nn_cpu(at::Tensor xyz0, at::Tensor xyz1, at::Tensor K, int patch
at::Tensor xcorrvol_cpu(at::Tensor in0, at::Tensor in1, int n_disps, int block_size) { at::Tensor xcorrvol_cpu(at::Tensor in0, at::Tensor in1, int n_disps, int block_size) {
CHECK_INPUT_CPU(in0) CHECK_INPUT_CPU(in0);
CHECK_INPUT_CPU(in1) CHECK_INPUT_CPU(in1);
auto channels = in0.size(0); auto channels = in0.size(0);
auto height = in0.size(1); auto height = in0.size(1);
@ -108,8 +108,8 @@ at::Tensor xcorrvol_cpu(at::Tensor in0, at::Tensor in1, int n_disps, int block_s
at::Tensor photometric_loss_forward(at::Tensor es, at::Tensor ta, int block_size, int type, float eps) { at::Tensor photometric_loss_forward(at::Tensor es, at::Tensor ta, int block_size, int type, float eps) {
CHECK_INPUT_CPU(es) CHECK_INPUT_CPU(es);
CHECK_INPUT_CPU(ta) CHECK_INPUT_CPU(ta);
auto batch_size = es.size(0); auto batch_size = es.size(0);
auto channels = es.size(1); auto channels = es.size(1);
@ -145,16 +145,16 @@ at::Tensor photometric_loss_forward(at::Tensor es, at::Tensor ta, int block_size
} }
at::Tensor photometric_loss_backward(at::Tensor es, at::Tensor ta, at::Tensor grad_out, int block_size, int type, float eps) { at::Tensor photometric_loss_backward(at::Tensor es, at::Tensor ta, at::Tensor grad_out, int block_size, int type, float eps) {
CHECK_INPUT_CPU(es) CHECK_INPUT_CPU(es);
CHECK_INPUT_CPU(ta) CHECK_INPUT_CPU(ta);
CHECK_INPUT_CPU(grad_out) CHECK_INPUT_CPU(grad_out);
auto batch_size = es.size(0); auto batch_size = es.size(0);
auto channels = es.size(1); auto channels = es.size(1);
auto height = es.size(2); auto height = es.size(2);
auto width = es.size(3); auto width = es.size(3);
CHECK_INPUT_CPU(ta) CHECK_INPUT_CPU(ta);
auto grad_in = at::zeros({batch_size, channels, height, width}, grad_out.options()); auto grad_in = at::zeros({batch_size, channels, height, width}, grad_out.options());
AT_DISPATCH_FLOATING_TYPES(es.scalar_type(), "photometric_loss_backward_cpu", ([&] { AT_DISPATCH_FLOATING_TYPES(es.scalar_type(), "photometric_loss_backward_cpu", ([&] {

54
torchext/ext/ext_cuda.cpp
Unescape View File

@ -7,16 +7,16 @@
void nn_kernel(at::Tensor in0, at::Tensor in1, at::Tensor out); void nn_kernel(at::Tensor in0, at::Tensor in1, at::Tensor out);
at::Tensor nn_cuda(at::Tensor in0, at::Tensor in1) { at::Tensor nn_cuda(at::Tensor in0, at::Tensor in1) {
CHECK_INPUT_CUDA(in0) CHECK_INPUT_CUDA(in0);
CHECK_INPUT_CUDA(in1) CHECK_INPUT_CUDA(in1);
auto nelem0 = in0.size(0); auto nelem0 = in0.size(0);
auto dim = in0.size(1); auto dim = in0.size(1);
AT_ASSERTM(dim == in1.size(1), "in0 and in1 have to be the same shape") AT_ASSERTM(dim == in1.size(1), "in0 and in1 have to be the same shape");
AT_ASSERTM(dim == 3, "dim hast to be 3") AT_ASSERTM(dim == 3, "dim hast to be 3");
AT_ASSERTM(in0.dim() == 2, "in0 has to be N0 x 3") AT_ASSERTM(in0.dim() == 2, "in0 has to be N0 x 3");
AT_ASSERTM(in1.dim() == 2, "in1 has to be N1 x 3") AT_ASSERTM(in1.dim() == 2, "in1 has to be N1 x 3");
auto out = at::empty({nelem0}, torch::CUDA(at::kLong)); auto out = at::empty({nelem0}, torch::CUDA(at::kLong));
@ -29,11 +29,11 @@ at::Tensor nn_cuda(at::Tensor in0, at::Tensor in1) {
void crosscheck_kernel(at::Tensor in0, at::Tensor in1, at::Tensor out); void crosscheck_kernel(at::Tensor in0, at::Tensor in1, at::Tensor out);
at::Tensor crosscheck_cuda(at::Tensor in0, at::Tensor in1) { at::Tensor crosscheck_cuda(at::Tensor in0, at::Tensor in1) {
CHECK_INPUT_CUDA(in0) CHECK_INPUT_CUDA(in0);
CHECK_INPUT_CUDA(in1) CHECK_INPUT_CUDA(in1);
AT_ASSERTM(in0.dim() == 1, "") AT_ASSERTM(in0.dim() == 1, "");
AT_ASSERTM(in1.dim() == 1, "") AT_ASSERTM(in1.dim() == 1, "");
auto nelem0 = in0.size(0); auto nelem0 = in0.size(0);
auto out = at::empty({nelem0}, torch::CUDA(at::kByte)); auto out = at::empty({nelem0}, torch::CUDA(at::kByte));
@ -45,21 +45,21 @@ at::Tensor crosscheck_cuda(at::Tensor in0, at::Tensor in1) {
void proj_nn_kernel(at::Tensor xyz0, at::Tensor xyz1, at::Tensor K, int patch_size, at::Tensor out); void proj_nn_kernel(at::Tensor xyz0, at::Tensor xyz1, at::Tensor K, int patch_size, at::Tensor out);
at::Tensor proj_nn_cuda(at::Tensor xyz0, at::Tensor xyz1, at::Tensor K, int patch_size) { at::Tensor proj_nn_cuda(at::Tensor xyz0, at::Tensor xyz1, at::Tensor K, int patch_size) {
CHECK_INPUT_CUDA(xyz0) CHECK_INPUT_CUDA(xyz0);
CHECK_INPUT_CUDA(xyz1) CHECK_INPUT_CUDA(xyz1);
CHECK_INPUT_CUDA(K) CHECK_INPUT_CUDA(K);
auto batch_size = xyz0.size(0); auto batch_size = xyz0.size(0);
auto height = xyz0.size(1); auto height = xyz0.size(1);
auto width = xyz0.size(2); auto width = xyz0.size(2);
AT_ASSERTM(xyz0.size(0) == xyz1.size(0), "") AT_ASSERTM(xyz0.size(0) == xyz1.size(0), "");
AT_ASSERTM(xyz0.size(1) == xyz1.size(1), "") AT_ASSERTM(xyz0.size(1) == xyz1.size(1), "");
AT_ASSERTM(xyz0.size(2) == xyz1.size(2), "") AT_ASSERTM(xyz0.size(2) == xyz1.size(2), "");
AT_ASSERTM(xyz0.size(3) == xyz1.size(3), "") AT_ASSERTM(xyz0.size(3) == xyz1.size(3), "");
AT_ASSERTM(xyz0.size(3) == 3, "") AT_ASSERTM(xyz0.size(3) == 3, "");
AT_ASSERTM(xyz0.dim() == 4, "") AT_ASSERTM(xyz0.dim() == 4, "");
AT_ASSERTM(xyz1.dim() == 4, "") AT_ASSERTM(xyz1.dim() == 4, "");
auto out = at::empty({batch_size, height, width}, torch::CUDA(at::kLong)); auto out = at::empty({batch_size, height, width}, torch::CUDA(at::kLong));
@ -71,8 +71,8 @@ at::Tensor proj_nn_cuda(at::Tensor xyz0, at::Tensor xyz1, at::Tensor K, int patc
void xcorrvol_kernel(at::Tensor in0, at::Tensor in1, int n_disps, int block_size, at::Tensor out); void xcorrvol_kernel(at::Tensor in0, at::Tensor in1, int n_disps, int block_size, at::Tensor out);
at::Tensor xcorrvol_cuda(at::Tensor in0, at::Tensor in1, int n_disps, int block_size) { at::Tensor xcorrvol_cuda(at::Tensor in0, at::Tensor in1, int n_disps, int block_size) {
CHECK_INPUT_CUDA(in0) CHECK_INPUT_CUDA(in0);
CHECK_INPUT_CUDA(in1) CHECK_INPUT_CUDA(in1);
// auto channels = in0.size(0); // auto channels = in0.size(0);
auto height = in0.size(1); auto height = in0.size(1);
@ -90,8 +90,8 @@ at::Tensor xcorrvol_cuda(at::Tensor in0, at::Tensor in1, int n_disps, int block_
void photometric_loss_forward_kernel(at::Tensor es, at::Tensor ta, int block_size, int type, float eps, at::Tensor out); void photometric_loss_forward_kernel(at::Tensor es, at::Tensor ta, int block_size, int type, float eps, at::Tensor out);
at::Tensor photometric_loss_forward(at::Tensor es, at::Tensor ta, int block_size, int type, float eps) { at::Tensor photometric_loss_forward(at::Tensor es, at::Tensor ta, int block_size, int type, float eps) {
CHECK_INPUT_CUDA(es) CHECK_INPUT_CUDA(es);
CHECK_INPUT_CUDA(ta) CHECK_INPUT_CUDA(ta);
auto batch_size = es.size(0); auto batch_size = es.size(0);
auto height = es.size(2); auto height = es.size(2);
@ -107,9 +107,9 @@ at::Tensor photometric_loss_forward(at::Tensor es, at::Tensor ta, int block_size
void photometric_loss_backward_kernel(at::Tensor es, at::Tensor ta, at::Tensor grad_out, int block_size, int type, float eps, at::Tensor grad_in); void photometric_loss_backward_kernel(at::Tensor es, at::Tensor ta, at::Tensor grad_out, int block_size, int type, float eps, at::Tensor grad_in);
at::Tensor photometric_loss_backward(at::Tensor es, at::Tensor ta, at::Tensor grad_out, int block_size, int type, float eps) { at::Tensor photometric_loss_backward(at::Tensor es, at::Tensor ta, at::Tensor grad_out, int block_size, int type, float eps) {
CHECK_INPUT_CUDA(es) CHECK_INPUT_CUDA(es);
CHECK_INPUT_CUDA(ta) CHECK_INPUT_CUDA(ta);
CHECK_INPUT_CUDA(grad_out) CHECK_INPUT_CUDA(grad_out);
auto batch_size = es.size(0); auto batch_size = es.size(0);
auto channels = es.size(1); auto channels = es.size(1);

3
torchext/setup.py
Unescape View File

@ -8,9 +8,6 @@ include_dirs = [
] ]
nvcc_args = [ nvcc_args = [
'-arch=sm_30',
'-gencode=arch=compute_30,code=sm_30',
'-gencode=arch=compute_35,code=sm_35',
] ]
setup( setup(

Write Preview
Loading…
Cancel
Save