remove .type().tensor() calls in favor of the new approach to tensor initialization

Author: varunagrawal

torchvision/csrc/cpu/ROIAlign_cpu.cpp

@@ -231,8 +231,8 @@ at::Tensor ROIAlign_forward_cpu(const at::Tensor& input,
   auto height = input.size(2);
   auto width = input.size(3);
 
-  at::Tensor output = input.type().tensor({num_rois, channels, pooled_height, pooled_width});
-
+  at::Tensor output = at::zeros({num_rois, channels, pooled_height, pooled_width}, input.type());
+  
   auto output_size = num_rois * pooled_height * pooled_width * channels;
 
   if (output.numel() == 0)

torchvision/csrc/cpu/ROIPool_cpu.cpp

@@ -16,8 +16,8 @@ std::tuple<at::Tensor, at::Tensor> ROIPool_forward_cpu(const at::Tensor &input,
     int input_height = input.size(2);
     int input_width = input.size(3);
 
-    at::Tensor output = input.type().tensor({num_rois, channels, pooled_height, pooled_width});
-    at::Tensor argmax = input.type().toScalarType(at::kInt).tensor({num_rois, channels, pooled_height, pooled_width}).zero_();
+    at::Tensor output = at::zeros({num_rois, channels, pooled_height, pooled_width}, input.type());
+    at::Tensor argmax = at::zeros({num_rois, channels, pooled_height, pooled_width}, input.type().toScalarType(at::kInt));
 
     // define accessors for indexing
     auto input_a = input.accessor<float, 4>();

torchvision/csrc/cpu/nms_cpu.cpp

@@ -1,6 +1,5 @@
 #include "cpu/vision.h"
 
-
 template <typename scalar_t>
 at::Tensor nms_cpu_kernel(const at::Tensor& dets,
                           const at::Tensor& scores,
@@ -10,7 +9,7 @@ at::Tensor nms_cpu_kernel(const at::Tensor& dets,
   AT_ASSERTM(dets.type() == scores.type(), "dets should have the same type as scores");
 
   if (dets.numel() == 0)
-    return torch::CPU(at::kLong).tensor();
+    return at::empty({0}, at::device(at::kCPU).dtype(at::kLong));
 
   auto x1_t = dets.select(1, 0).contiguous();
   auto y1_t = dets.select(1, 1).contiguous();
@@ -22,7 +21,7 @@ at::Tensor nms_cpu_kernel(const at::Tensor& dets,
   auto order_t = std::get<1>(scores.sort(0, /* descending=*/true));
 
   auto ndets = dets.size(0);
-  at::Tensor suppressed_t = at::zeros(torch::CPU(at::kByte), {ndets});
+  at::Tensor suppressed_t = at::zeros({ndets}, at::device(at::kCPU).dtype(at::kByte));
 
   auto suppressed = suppressed_t.data<uint8_t>();
   auto order = order_t.data<int64_t>();
@@ -66,7 +65,7 @@ at::Tensor nms_cpu(const at::Tensor& dets,
                const at::Tensor& scores,
                const float threshold) {
 
-  auto result = dets.type().tensor();
+  auto result = at::empty({0}, dets.type());
 
   AT_DISPATCH_FLOATING_TYPES(dets.type(), "nms", [&] {
     result = nms_cpu_kernel<scalar_t>(dets, scores, threshold);

torchvision/csrc/cuda/ROIAlign_cuda.cu

@@ -267,7 +267,7 @@ at::Tensor ROIAlign_forward_cuda(const at::Tensor& input,
   auto height = input.size(2);
   auto width = input.size(3);
 
-  at::Tensor output = input.type().tensor({num_rois, channels, pooled_height, pooled_width});
+  at::Tensor output = at::zeros({num_rois, channels, pooled_height, pooled_width}, input.options());
 
   auto output_size = num_rois * pooled_height * pooled_width * channels;
   cudaStream_t stream = at::cuda::getCurrentCUDAStream();
@@ -313,7 +313,7 @@ at::Tensor ROIAlign_backward_cuda(const at::Tensor& grad,
   AT_ASSERTM(rois.type().is_cuda(), "rois must be a CUDA tensor");
 
   auto num_rois = rois.size(0);
-  at::Tensor grad_input = grad.type().tensor({batch_size, channels, height, width}).zero_();
+  at::Tensor grad_input = at::zeros({batch_size, channels, height, width}, grad.options());
 
   cudaStream_t stream = at::cuda::getCurrentCUDAStream();
 

torchvision/csrc/cuda/ROIPool_cuda.cu

@@ -116,8 +116,8 @@ std::tuple<at::Tensor, at::Tensor> ROIPool_forward_cuda(const at::Tensor& input,
   auto height = input.size(2);
   auto width = input.size(3);
 
-  at::Tensor output = input.type().tensor({num_rois, channels, pooled_height, pooled_width});
-  at::Tensor argmax = input.type().toScalarType(at::kInt).tensor({num_rois, channels, pooled_height, pooled_width}).zero_();
+  at::Tensor output = at::zeros({num_rois, channels, pooled_height, pooled_width}, input.type());
+  at::Tensor argmax = at::zeros({num_rois, channels, pooled_height, pooled_width}, input.type().toScalarType(at::kInt));
 
   auto output_size = num_rois * pooled_height * pooled_width * channels;
   cudaStream_t stream = at::cuda::getCurrentCUDAStream();