Restructuring C++ project: (#3146)

Summary:
* Reduce unnecessary header inclusions in models and io.

* Move autocast to separate folder and hide autograd implementation in an anonymous namespace.

* Moving files in subfolders.

Reviewed By: fmassa

Differential Revision: D25461523

fbshipit-source-id: 756eeb6848aacaa474de4825ed4c1045d17e2cea

torchvision/csrc/models/googlenet.cpp

 #include "googlenet.h"
 
-#include "modelsimpl.h"
-
 namespace vision {
 namespace models {
 

torchvision/csrc/models/googlenet.h

 #pragma once
 
-#include <torch/torch.h>
+#include <torch/nn.h>
 #include "../macros.h"
 
 namespace vision {

torchvision/csrc/models/inception.h

 #pragma once
 
-#include <torch/torch.h>
+#include <torch/nn.h>
 #include "../macros.h"
 
 namespace vision {

torchvision/csrc/models/mnasnet.h

 #pragma once
 
-#include <torch/torch.h>
+#include <torch/nn.h>
 #include "../macros.h"
 
 namespace vision {

torchvision/csrc/models/mobilenet.h

 #pragma once
 
-#include <torch/torch.h>
+#include <torch/nn.h>
 #include "../macros.h"
 
 namespace vision {

torchvision/csrc/models/modelsimpl.h

 #pragma once
 
-#include <torch/torch.h>
-
-#ifndef TORCH_CHECK
-#define TORCH_CHECK AT_CHECK
-#endif
+#include <torch/nn.h>
 
 namespace vision {
 namespace models {

torchvision/csrc/models/resnet.cpp

 #include "resnet.h"
 
-#include "modelsimpl.h"
-
 namespace vision {
 namespace models {
 namespace _resnetimpl {

torchvision/csrc/models/resnet.h

 #pragma once
 
-#include <torch/torch.h>
+#include <torch/nn.h>
 #include "../macros.h"
 
 namespace vision {

torchvision/csrc/models/shufflenetv2.h

 #pragma once
 
-#include <torch/torch.h>
+#include <torch/nn.h>
 #include "../macros.h"
 
 namespace vision {

torchvision/csrc/models/squeezenet.h

 #pragma once
 
-#include <torch/torch.h>
+#include <torch/nn.h>
 #include "../macros.h"
 
 namespace vision {

torchvision/csrc/models/vgg.h

 #pragma once
 
-#include <torch/torch.h>
+#include <torch/nn.h>
 #include "../macros.h"
 
 namespace vision {

torchvision/csrc/ops/autocast/deform_conv2d_kernel.cpp

+#include "../deform_conv2d.h"
+
+#include <ATen/autocast_mode.h>
+#include <torch/types.h>
+
+namespace vision {
+namespace ops {
+
+namespace {
+
+at::Tensor deform_conv2d_autocast(
+    const at::Tensor& input,
+    const at::Tensor& weight,
+    const at::Tensor& offset,
+    const at::Tensor& mask,
+    const at::Tensor& bias,
+    int64_t stride_h,
+    int64_t stride_w,
+    int64_t pad_h,
+    int64_t pad_w,
+    int64_t dilation_h,
+    int64_t dilation_w,
+    int64_t groups,
+    int64_t offset_groups,
+    bool use_mask) {
+  c10::impl::ExcludeDispatchKeyGuard no_autocast(c10::DispatchKey::Autocast);
+  return deform_conv2d(
+             at::autocast::cached_cast(at::kFloat, input),
+             at::autocast::cached_cast(at::kFloat, weight),
+             at::autocast::cached_cast(at::kFloat, offset),
+             at::autocast::cached_cast(at::kFloat, mask),
+             at::autocast::cached_cast(at::kFloat, bias),
+             stride_h,
+             stride_w,
+             pad_h,
+             pad_w,
+             dilation_h,
+             dilation_w,
+             groups,
+             offset_groups,
+             use_mask)
+      .to(input.scalar_type());
+}
+
+} // namespace
+
+TORCH_LIBRARY_IMPL(torchvision, Autocast, m) {
+  m.impl("deform_conv2d", deform_conv2d_autocast);
+}
+
+} // namespace ops
+} // namespace vision

torchvision/csrc/ops/autocast/nms_kernel.cpp

+#include "../nms.h"
+
+#include <ATen/autocast_mode.h>
+#include <torch/types.h>
+
+namespace vision {
+namespace ops {
+
+namespace {
+
+at::Tensor nms_autocast(
+    const at::Tensor& dets,
+    const at::Tensor& scores,
+    double iou_threshold) {
+  c10::impl::ExcludeDispatchKeyGuard no_autocast(c10::DispatchKey::Autocast);
+  return nms(
+      at::autocast::cached_cast(at::kFloat, dets),
+      at::autocast::cached_cast(at::kFloat, scores),
+      iou_threshold);
+}
+
+} // namespace
+
+TORCH_LIBRARY_IMPL(torchvision, Autocast, m) {
+  m.impl("nms", nms_autocast);
+}
+
+} // namespace ops
+} // namespace vision

torchvision/csrc/ops/autocast/ps_roi_align_kernel.cpp

+#include "../ps_roi_align.h"
+
+#include <ATen/autocast_mode.h>
+#include <torch/types.h>
+
+namespace vision {
+namespace ops {
+
+namespace {
+
+std::tuple<at::Tensor, at::Tensor> ps_roi_align_autocast(
+    const at::Tensor& input,
+    const at::Tensor& rois,
+    double spatial_scale,
+    int64_t pooled_height,
+    int64_t pooled_width,
+    int64_t sampling_ratio) {
+  c10::impl::ExcludeDispatchKeyGuard no_autocast(c10::DispatchKey::Autocast);
+  auto result = ps_roi_align(
+      at::autocast::cached_cast(at::kFloat, input),
+      at::autocast::cached_cast(at::kFloat, rois),
+      spatial_scale,
+      pooled_height,
+      pooled_width,
+      sampling_ratio);
+
+  return std::make_tuple(
+      std::get<0>(result).to(input.scalar_type()),
+      std::get<1>(result).to(input.scalar_type()));
+}
+
+} // namespace
+
+TORCH_LIBRARY_IMPL(torchvision, Autocast, m) {
+  m.impl("ps_roi_align", ps_roi_align_autocast);
+}
+
+} // namespace ops
+} // namespace vision

torchvision/csrc/ops/autocast/ps_roi_pool_kernel.cpp

+#include "../ps_roi_pool.h"
+
+#include <ATen/autocast_mode.h>
+#include <torch/types.h>
+
+namespace vision {
+namespace ops {
+
+namespace {
+
+std::tuple<at::Tensor, at::Tensor> ps_roi_pool_autocast(
+    const at::Tensor& input,
+    const at::Tensor& rois,
+    double spatial_scale,
+    int64_t pooled_height,
+    int64_t pooled_width) {
+  c10::impl::ExcludeDispatchKeyGuard no_autocast(c10::DispatchKey::Autocast);
+  auto result = ps_roi_pool(
+      at::autocast::cached_cast(at::kFloat, input),
+      at::autocast::cached_cast(at::kFloat, rois),
+      spatial_scale,
+      pooled_height,
+      pooled_width);
+
+  return std::make_tuple(
+      std::get<0>(result).to(input.scalar_type()),
+      std::get<1>(result).to(input.scalar_type()));
+}
+
+} // namespace
+
+TORCH_LIBRARY_IMPL(torchvision, Autocast, m) {
+  m.impl("ps_roi_pool", ps_roi_pool_autocast);
+}
+
+} // namespace ops
+} // namespace vision

torchvision/csrc/ops/autocast/roi_align_kernel.cpp

+#include "../roi_align.h"
+
+#include <ATen/autocast_mode.h>
+#include <torch/types.h>
+
+namespace vision {
+namespace ops {
+
+namespace {
+
+at::Tensor roi_align_autocast(
+    const at::Tensor& input,
+    const at::Tensor& rois,
+    double spatial_scale,
+    int64_t pooled_height,
+    int64_t pooled_width,
+    int64_t sampling_ratio,
+    bool aligned) {
+  c10::impl::ExcludeDispatchKeyGuard no_autocast(c10::DispatchKey::Autocast);
+  return roi_align(
+             at::autocast::cached_cast(at::kFloat, input),
+             at::autocast::cached_cast(at::kFloat, rois),
+             spatial_scale,
+             pooled_height,
+             pooled_width,
+             sampling_ratio,
+             aligned)
+      .to(input.scalar_type());
+}
+
+} // namespace
+
+TORCH_LIBRARY_IMPL(torchvision, Autocast, m) {
+  m.impl("roi_align", roi_align_autocast);
+}
+
+} // namespace ops
+} // namespace vision

torchvision/csrc/ops/autocast/roi_pool_kernel.cpp

+#include "../roi_pool.h"
+
+#include <ATen/autocast_mode.h>
+#include <torch/types.h>
+
+namespace vision {
+namespace ops {
+
+namespace {
+
+std::tuple<at::Tensor, at::Tensor> roi_pool_autocast(
+    const at::Tensor& input,
+    const at::Tensor& rois,
+    double spatial_scale,
+    int64_t pooled_height,
+    int64_t pooled_width) {
+  c10::impl::ExcludeDispatchKeyGuard no_autocast(c10::DispatchKey::Autocast);
+  auto result = roi_pool(
+      at::autocast::cached_cast(at::kFloat, input),
+      at::autocast::cached_cast(at::kFloat, rois),
+      spatial_scale,
+      pooled_height,
+      pooled_width);
+
+  return std::make_tuple(
+      std::get<0>(result).to(input.scalar_type()),
+      std::get<1>(result).to(input.scalar_type()));
+}
+
+} // namespace
+
+TORCH_LIBRARY_IMPL(torchvision, Autocast, m) {
+  m.impl("roi_pool", roi_pool_autocast);
+}
+
+} // namespace ops
+} // namespace vision