PyTorch 1.0

README.md

@@ -128,10 +128,10 @@ python VGGplus.py
 
 ## Setup
 
-Tested with Ubuntu 16.04, Python 3.6 in [Miniconda](https://conda.io/miniconda.html) and PyTorch v0.4 (with merged Tensors/Variables).
+Tested with Ubuntu 16.04, Python 3.6 in [Miniconda](https://conda.io/miniconda.html) and PyTorch 1.0.
 
 ```
-conda install pytorch -c pytorch
+conda install pytorch-nightly -c pytorch # See https://pytorch.org/get-started/locally/
 conda install google-sparsehash -c bioconda   # OR apt-get install libsparsehash-dev
 conda install -c anaconda pillow
 git clone git@github.com:facebookresearch/SparseConvNet.git

build.sh

 #!/bin/bash
-rm -rf build/ dist/ sparseconvnet.egg-info sparseconvnet_SCN*.so
+rm -rf build/ dist/ sparseconvnet.egg-info
 python setup.py install

sparseconvnet/SCN/CPU/Convolution.cpp

@@ -62,7 +62,7 @@ double cpu_Convolution_updateOutput(
       // auto w = weight.select(0, i);
       // auto output_rows = at::mm(input_rows, w);
       // output_features.index_add_(0, rt.select(1, 1), output_rows);
-      auto input_rows = input_features.type().tensor({nRules, ip});
+      auto input_rows = at::empty({nRules, ip}, input_features.type());
       rule_index_select<T>(input_rows, input_features, nRules, &r[0]);
       auto w = weight.select(0, i);
       auto output_rows = at::mm(input_rows, w);
@@ -105,9 +105,9 @@ void cpu_Convolution_backward(
       // at::mm_out(dw, input_rows.t(), d_output_rows);
       // auto d_input_rows = at::mm(d_output_rows, w.t());
       // d_input_features.index_add_(0, rt.select(1, 0), d_input_rows);
-      auto input_rows = input_features.type().tensor({nRules, ip});
+      auto input_rows = at::empty({nRules, ip}, input_features.type());
       rule_index_select<T>(input_rows, input_features, nRules, &r[0]);
-      auto d_output_rows = d_output_features.type().tensor({nRules, op});
+      auto d_output_rows = at::empty({nRules, op}, d_output_features.type());
       rule_index_select<T>(d_output_rows, d_output_features, nRules, &r[1]);
       at::mm_out(dw, input_rows.t(), d_output_rows);
       auto d_input_rows = at::mm(d_output_rows, w.t());
@@ -144,7 +144,7 @@ double cpu_SubmanifoldConvolution_updateOutput(
       // auto w = weight.select(0, i);
       // auto output_rows = at::mm(input_rows, w);
       // output_features.index_add_(0, rt.select(1, 1), output_rows);
-      auto input_rows = input_features.type().tensor({nRules, ip});
+      auto input_rows = at::empty({nRules, ip}, input_features.type());
       rule_index_select<T>(input_rows, input_features, nRules, &r[0]);
       auto w = weight.select(0, i);
       auto output_rows = at::mm(input_rows, w);
@@ -186,9 +186,9 @@ void cpu_SubmanifoldConvolution_backward(
       // at::mm_out(dw, input_rows.t(), d_output_rows);
       // auto d_input_rows = at::mm(d_output_rows, w.t());
       // d_input_features.index_add_(0, rt.select(1, 0), d_input_rows);
-      auto input_rows = input_features.type().tensor({nRules, ip});
+      auto input_rows = at::empty({nRules, ip}, input_features.type());
       rule_index_select<T>(input_rows, input_features, nRules, &r[0]);
-      auto d_output_rows = d_output_features.type().tensor({nRules, op});
+      auto d_output_rows = at::empty({nRules, op}, d_output_features.type());
       rule_index_select<T>(d_output_rows, d_output_features, nRules, &r[1]);
       at::mm_out(dw, input_rows.t(), d_output_rows);
       auto d_input_rows = at::mm(d_output_rows, w.t());
@@ -224,7 +224,7 @@ double cpu_PermutohedralSubmanifoldConvolution_updateOutput(
       // auto w = weight.select(0, i);
       // auto output_rows = at::mm(input_rows, w);
       // output_features.index_add_(0, rt.select(1, 1), output_rows);
-      auto input_rows = input_features.type().tensor({nRules, ip});
+      auto input_rows = at::empty({nRules, ip}, input_features.type());
       rule_index_select<T>(input_rows, input_features, nRules, &r[0]);
       auto w = weight.select(0, i);
       auto output_rows = at::mm(input_rows, w);
@@ -265,9 +265,9 @@ void cpu_PermutohedralSubmanifoldConvolution_backward(
       // at::mm_out(dw, input_rows.t(), d_output_rows);
       // auto d_input_rows = at::mm(d_output_rows, w.t());
       // d_input_features.index_add_(0, rt.select(1, 0), d_input_rows);
-      auto input_rows = input_features.type().tensor({nRules, ip});
+      auto input_rows = at::empty({nRules, ip}, input_features.type());
       rule_index_select<T>(input_rows, input_features, nRules, &r[0]);
-      auto d_output_rows = d_output_features.type().tensor({nRules, op});
+      auto d_output_rows = at::empty({nRules, op}, d_output_features.type());
       rule_index_select<T>(d_output_rows, d_output_features, nRules, &r[1]);
       at::mm_out(dw, input_rows.t(), d_output_rows);
       auto d_input_rows = at::mm(d_output_rows, w.t());
@@ -307,7 +307,7 @@ double cpu_FullConvolution_updateOutput(
       // auto w = weight.select(0, i);
       // auto output_rows = at::mm(input_rows, w);
       // output_features.index_add_(0, rt.select(1, 1), output_rows);
-      auto input_rows = input_features.type().tensor({nRules, ip});
+      auto input_rows = at::empty({nRules, ip}, input_features.type());
       rule_index_select<T>(input_rows, input_features, nRules, &r[0]);
       auto w = weight.select(0, i);
       auto output_rows = at::mm(input_rows, w);
@@ -352,9 +352,9 @@ void cpu_FullConvolution_backward(
       // at::mm_out(dw, input_rows.t(), d_output_rows);
       // auto d_input_rows = at::mm(d_output_rows, w.t());
       // d_input_features.index_add_(0, rt.select(1, 0), d_input_rows);
-      auto input_rows = input_features.type().tensor({nRules, ip});
+      auto input_rows = at::empty({nRules, ip}, input_features.type());
       rule_index_select<T>(input_rows, input_features, nRules, &r[0]);
-      auto d_output_rows = d_output_features.type().tensor({nRules, op});
+      auto d_output_rows = at::empty({nRules, op}, d_output_features.type());
       rule_index_select<T>(d_output_rows, d_output_features, nRules, &r[1]);
       at::mm_out(dw, input_rows.t(), d_output_rows);
       auto d_input_rows = at::mm(d_output_rows, w.t());
@@ -393,7 +393,7 @@ double cpu_RandomizedStrideConvolution_updateOutput(
       // auto w = weight.select(0, i);
       // auto output_rows = at::mm(input_rows, w);
       // output_features.index_add_(0, rt.select(1, 1), output_rows);
-      auto input_rows = input_features.type().tensor({nRules, ip});
+      auto input_rows = at::empty({nRules, ip}, input_features.type());
       rule_index_select<T>(input_rows, input_features, nRules, &r[0]);
       auto w = weight.select(0, i);
       auto output_rows = at::mm(input_rows, w);
@@ -436,9 +436,9 @@ void cpu_RandomizedStrideConvolution_backward(
       // at::mm_out(dw, input_rows.t(), d_output_rows);
       // auto d_input_rows = at::mm(d_output_rows, w.t());
       // d_input_features.index_add_(0, rt.select(1, 0), d_input_rows);
-      auto input_rows = input_features.type().tensor({nRules, ip});
+      auto input_rows = at::empty({nRules, ip}, input_features.type());
       rule_index_select<T>(input_rows, input_features, nRules, &r[0]);
-      auto d_output_rows = d_output_features.type().tensor({nRules, op});
+      auto d_output_rows = at::empty({nRules, op}, d_output_features.type());
       rule_index_select<T>(d_output_rows, d_output_features, nRules, &r[1]);
       at::mm_out(dw, input_rows.t(), d_output_rows);
       auto d_input_rows = at::mm(d_output_rows, w.t());

sparseconvnet/SCN/CPU/Deconvolution.cpp

@@ -34,7 +34,7 @@ double cpu_Deconvolution_updateOutput(
       // auto w = weight.select(0, i);
       // auto output_rows = at::mm(input_rows, w);
       // output_features.index_add_(0, rt.select(1, 0), output_rows);
-      auto input_rows = input_features.type().tensor({nRules, ip});
+      auto input_rows = at::empty({nRules, ip}, input_features.type());
       rule_index_select<T>(input_rows, input_features, nRules, &r[1]);
       auto w = weight.select(0, i);
       auto output_rows = at::mm(input_rows, w);
@@ -77,9 +77,9 @@ void cpu_Deconvolution_backward(
       // at::mm_out(dw, input_rows.t(), d_output_rows);
       // auto d_input_rows = at::mm(d_output_rows, w.t());
       // d_input_features.index_add_(0, rt.select(1, 1), d_input_rows);
-      auto input_rows = input_features.type().tensor({nRules, ip});
+      auto input_rows = at::empty({nRules, ip}, d_output_features.type());
       rule_index_select<T>(input_rows, input_features, nRules, &r[1]);
-      auto d_output_rows = d_output_features.type().tensor({nRules, op});
+      auto d_output_rows = at::empty({nRules, op}, d_output_features.type());
       rule_index_select<T>(d_output_rows, d_output_features, nRules, &r[0]);
       at::mm_out(dw, input_rows.t(), d_output_rows);
       auto d_input_rows = at::mm(d_output_rows, w.t());

sparseconvnet/SCN/pybind.cpp

@@ -4,7 +4,7 @@
 // This source code is licensed under the license found in the
 // LICENSE file in the root directory of this source tree.
 
-#include <torch/torch.h>
+#include <torch/extension.h>
 
 #include "sparseconvnet.h"
 

sparseconvnet/SCN/sparseconvnet_cpu.cpp

@@ -9,7 +9,7 @@
 #include <omp.h>
 #endif
 
-#include <torch/torch.h>
+#include <torch/extension.h>
 
 #include "Metadata/Metadata.cpp"
 template class Metadata<1>;

sparseconvnet/SCN/sparseconvnet_cuda.cpp

@@ -9,10 +9,9 @@
 #include <omp.h>
 #endif
 
-//#include <ATen/ATen.h>
 #include <cuda.h>
 #include <cuda_runtime.h>
-#include <torch/torch.h>
+#include <torch/extension.h>
 
 #include "Metadata/Metadata.cpp"
 template class Metadata<1>;

sparseconvnet/sparseConvNetTensor.py

@@ -29,8 +29,8 @@ class SparseConvNetTensor(object):
             return self
         return self.features.type()
 
-    def cuda(self, async=False):
-        self.features = self.features.cuda(async=async)
+    def cuda(self):
+        self.features = self.features.cuda()
         return self
 
     def cpu(self):