catvton

detectron2/layers/csrc/nms_rotated/nms_rotated.h

+// Copyright (c) Facebook, Inc. and its affiliates.
+#pragma once
+#include <torch/types.h>
+
+namespace detectron2 {
+
+at::Tensor nms_rotated_cpu(
+    const at::Tensor& dets,
+    const at::Tensor& scores,
+    const double iou_threshold);
+
+#if defined(WITH_CUDA) || defined(WITH_HIP)
+at::Tensor nms_rotated_cuda(
+    const at::Tensor& dets,
+    const at::Tensor& scores,
+    const double iou_threshold);
+#endif
+
+// Interface for Python
+// inline is needed to prevent multiple function definitions when this header is
+// included by different cpps
+inline at::Tensor nms_rotated(
+    const at::Tensor& dets,
+    const at::Tensor& scores,
+    const double iou_threshold) {
+  assert(dets.device().is_cuda() == scores.device().is_cuda());
+  if (dets.device().is_cuda()) {
+#if defined(WITH_CUDA) || defined(WITH_HIP)
+    return nms_rotated_cuda(
+        dets.contiguous(), scores.contiguous(), iou_threshold);
+#else
+    AT_ERROR("Detectron2 is not compiled with GPU support!");
+#endif
+  }
+
+  return nms_rotated_cpu(dets.contiguous(), scores.contiguous(), iou_threshold);
+}
+
+} // namespace detectron2

detectron2/layers/csrc/nms_rotated/nms_rotated_cpu.cpp

+// Copyright (c) Facebook, Inc. and its affiliates.
+#include "../box_iou_rotated/box_iou_rotated_utils.h"
+#include "nms_rotated.h"
+
+namespace detectron2 {
+
+template <typename scalar_t>
+at::Tensor nms_rotated_cpu_kernel(
+    const at::Tensor& dets,
+    const at::Tensor& scores,
+    const double iou_threshold) {
+  // nms_rotated_cpu_kernel is modified from torchvision's nms_cpu_kernel,
+  // however, the code in this function is much shorter because
+  // we delegate the IoU computation for rotated boxes to
+  // the single_box_iou_rotated function in box_iou_rotated_utils.h
+  AT_ASSERTM(dets.device().is_cpu(), "dets must be a CPU tensor");
+  AT_ASSERTM(scores.device().is_cpu(), "scores must be a CPU tensor");
+  AT_ASSERTM(
+      dets.scalar_type() == scores.scalar_type(),
+      "dets should have the same type as scores");
+
+  if (dets.numel() == 0) {
+    return at::empty({0}, dets.options().dtype(at::kLong));
+  }
+
+  auto order_t = std::get<1>(scores.sort(0, /* descending=*/true));
+
+  auto ndets = dets.size(0);
+  at::Tensor suppressed_t = at::zeros({ndets}, dets.options().dtype(at::kByte));
+  at::Tensor keep_t = at::zeros({ndets}, dets.options().dtype(at::kLong));
+
+  auto suppressed = suppressed_t.data_ptr<uint8_t>();
+  auto keep = keep_t.data_ptr<int64_t>();
+  auto order = order_t.data_ptr<int64_t>();
+
+  int64_t num_to_keep = 0;
+
+  for (int64_t _i = 0; _i < ndets; _i++) {
+    auto i = order[_i];
+    if (suppressed[i] == 1) {
+      continue;
+    }
+
+    keep[num_to_keep++] = i;
+
+    for (int64_t _j = _i + 1; _j < ndets; _j++) {
+      auto j = order[_j];
+      if (suppressed[j] == 1) {
+        continue;
+      }
+
+      auto ovr = single_box_iou_rotated<scalar_t>(
+          dets[i].data_ptr<scalar_t>(), dets[j].data_ptr<scalar_t>());
+      if (ovr >= iou_threshold) {
+        suppressed[j] = 1;
+      }
+    }
+  }
+  return keep_t.narrow(/*dim=*/0, /*start=*/0, /*length=*/num_to_keep);
+}
+
+at::Tensor nms_rotated_cpu(
+    // input must be contiguous
+    const at::Tensor& dets,
+    const at::Tensor& scores,
+    const double iou_threshold) {
+  auto result = at::empty({0}, dets.options());
+
+  AT_DISPATCH_FLOATING_TYPES(dets.scalar_type(), "nms_rotated", [&] {
+    result = nms_rotated_cpu_kernel<scalar_t>(dets, scores, iou_threshold);
+  });
+  return result;
+}
+
+} // namespace detectron2

detectron2/layers/csrc/nms_rotated/nms_rotated_cuda.cu

+// Copyright (c) Facebook, Inc. and its affiliates.
+#include <ATen/ATen.h>
+#include <ATen/cuda/CUDAContext.h>
+#include <c10/cuda/CUDAGuard.h>
+#include <ATen/cuda/CUDAApplyUtils.cuh>
+#ifdef WITH_CUDA
+#include "../box_iou_rotated/box_iou_rotated_utils.h"
+#endif
+// TODO avoid this when pytorch supports "same directory" hipification
+#ifdef WITH_HIP
+#include "box_iou_rotated/box_iou_rotated_utils.h"
+#endif
+
+using namespace detectron2;
+
+namespace {
+int const threadsPerBlock = sizeof(unsigned long long) * 8;
+}
+
+template <typename T>
+__global__ void nms_rotated_cuda_kernel(
+    const int n_boxes,
+    const double iou_threshold,
+    const T* dev_boxes,
+    unsigned long long* dev_mask) {
+  // nms_rotated_cuda_kernel is modified from torchvision's nms_cuda_kernel
+
+  const int row_start = blockIdx.y;
+  const int col_start = blockIdx.x;
+
+  // if (row_start > col_start) return;
+
+  const int row_size =
+      min(n_boxes - row_start * threadsPerBlock, threadsPerBlock);
+  const int col_size =
+      min(n_boxes - col_start * threadsPerBlock, threadsPerBlock);
+
+  // Compared to nms_cuda_kernel, where each box is represented with 4 values
+  // (x1, y1, x2, y2), each rotated box is represented with 5 values
+  // (x_center, y_center, width, height, angle_degrees) here.
+  __shared__ T block_boxes[threadsPerBlock * 5];
+  if (threadIdx.x < col_size) {
+    block_boxes[threadIdx.x * 5 + 0] =
+        dev_boxes[(threadsPerBlock * col_start + threadIdx.x) * 5 + 0];
+    block_boxes[threadIdx.x * 5 + 1] =
+        dev_boxes[(threadsPerBlock * col_start + threadIdx.x) * 5 + 1];
+    block_boxes[threadIdx.x * 5 + 2] =
+        dev_boxes[(threadsPerBlock * col_start + threadIdx.x) * 5 + 2];
+    block_boxes[threadIdx.x * 5 + 3] =
+        dev_boxes[(threadsPerBlock * col_start + threadIdx.x) * 5 + 3];
+    block_boxes[threadIdx.x * 5 + 4] =
+        dev_boxes[(threadsPerBlock * col_start + threadIdx.x) * 5 + 4];
+  }
+  __syncthreads();
+
+  if (threadIdx.x < row_size) {
+    const int cur_box_idx = threadsPerBlock * row_start + threadIdx.x;
+    const T* cur_box = dev_boxes + cur_box_idx * 5;
+    int i = 0;
+    unsigned long long t = 0;
+    int start = 0;
+    if (row_start == col_start) {
+      start = threadIdx.x + 1;
+    }
+    for (i = start; i < col_size; i++) {
+      // Instead of devIoU used by original horizontal nms, here
+      // we use the single_box_iou_rotated function from box_iou_rotated_utils.h
+      if (single_box_iou_rotated<T>(cur_box, block_boxes + i * 5) >
+          iou_threshold) {
+        t |= 1ULL << i;
+      }
+    }
+    const int col_blocks = at::cuda::ATenCeilDiv(n_boxes, threadsPerBlock);
+    dev_mask[cur_box_idx * col_blocks + col_start] = t;
+  }
+}
+
+namespace detectron2 {
+
+at::Tensor nms_rotated_cuda(
+    // input must be contiguous
+    const at::Tensor& dets,
+    const at::Tensor& scores,
+    double iou_threshold) {
+  // using scalar_t = float;
+  AT_ASSERTM(dets.is_cuda(), "dets must be a CUDA tensor");
+  AT_ASSERTM(scores.is_cuda(), "scores must be a CUDA tensor");
+  at::cuda::CUDAGuard device_guard(dets.device());
+
+  auto order_t = std::get<1>(scores.sort(0, /* descending=*/true));
+  auto dets_sorted = dets.index_select(0, order_t);
+
+  auto dets_num = dets.size(0);
+
+  const int col_blocks =
+      at::cuda::ATenCeilDiv(static_cast<int>(dets_num), threadsPerBlock);
+
+  at::Tensor mask =
+      at::empty({dets_num * col_blocks}, dets.options().dtype(at::kLong));
+
+  dim3 blocks(col_blocks, col_blocks);
+  dim3 threads(threadsPerBlock);
+  cudaStream_t stream = at::cuda::getCurrentCUDAStream();
+
+  AT_DISPATCH_FLOATING_TYPES(
+      dets_sorted.scalar_type(), "nms_rotated_kernel_cuda", [&] {
+        nms_rotated_cuda_kernel<scalar_t><<<blocks, threads, 0, stream>>>(
+            dets_num,
+            iou_threshold,
+            dets_sorted.data_ptr<scalar_t>(),
+            (unsigned long long*)mask.data_ptr<int64_t>());
+      });
+
+  at::Tensor mask_cpu = mask.to(at::kCPU);
+  unsigned long long* mask_host =
+      (unsigned long long*)mask_cpu.data_ptr<int64_t>();
+
+  std::vector<unsigned long long> remv(col_blocks);
+  memset(&remv[0], 0, sizeof(unsigned long long) * col_blocks);
+
+  at::Tensor keep =
+      at::empty({dets_num}, dets.options().dtype(at::kLong).device(at::kCPU));
+  int64_t* keep_out = keep.data_ptr<int64_t>();
+
+  int num_to_keep = 0;
+  for (int i = 0; i < dets_num; i++) {
+    int nblock = i / threadsPerBlock;
+    int inblock = i % threadsPerBlock;
+
+    if (!(remv[nblock] & (1ULL << inblock))) {
+      keep_out[num_to_keep++] = i;
+      unsigned long long* p = mask_host + i * col_blocks;
+      for (int j = nblock; j < col_blocks; j++) {
+        remv[j] |= p[j];
+      }
+    }
+  }
+
+  AT_CUDA_CHECK(cudaGetLastError());
+  return order_t.index(
+      {keep.narrow(/*dim=*/0, /*start=*/0, /*length=*/num_to_keep)
+           .to(order_t.device(), keep.scalar_type())});
+}
+
+} // namespace detectron2

detectron2/layers/csrc/vision.cpp

+// Copyright (c) Facebook, Inc. and its affiliates.
+
+#include <torch/extension.h>
+#include "ROIAlignRotated/ROIAlignRotated.h"
+#include "box_iou_rotated/box_iou_rotated.h"
+#include "cocoeval/cocoeval.h"
+#include "deformable/deform_conv.h"
+#include "nms_rotated/nms_rotated.h"
+
+namespace detectron2 {
+
+#if defined(WITH_CUDA) || defined(WITH_HIP)
+extern int get_cudart_version();
+#endif
+
+std::string get_cuda_version() {
+#if defined(WITH_CUDA) || defined(WITH_HIP)
+  std::ostringstream oss;
+
+#if defined(WITH_CUDA)
+  oss << "CUDA ";
+#else
+  oss << "HIP ";
+#endif
+
+  // copied from
+  // https://github.com/pytorch/pytorch/blob/master/aten/src/ATen/cuda/detail/CUDAHooks.cpp#L231
+  auto printCudaStyleVersion = [&](int v) {
+    oss << (v / 1000) << "." << (v / 10 % 100);
+    if (v % 10 != 0) {
+      oss << "." << (v % 10);
+    }
+  };
+  printCudaStyleVersion(get_cudart_version());
+  return oss.str();
+#else // neither CUDA nor HIP
+  return std::string("not available");
+#endif
+}
+
+bool has_cuda() {
+#if defined(WITH_CUDA)
+  return true;
+#else
+  return false;
+#endif
+}
+
+// similar to
+// https://github.com/pytorch/pytorch/blob/master/aten/src/ATen/Version.cpp
+std::string get_compiler_version() {
+  std::ostringstream ss;
+#if defined(__GNUC__)
+#ifndef __clang__
+
+#if ((__GNUC__ <= 4) && (__GNUC_MINOR__ <= 8))
+#error "GCC >= 4.9 is required!"
+#endif
+
+  { ss << "GCC " << __GNUC__ << "." << __GNUC_MINOR__; }
+#endif
+#endif
+
+#if defined(__clang_major__)
+  {
+    ss << "clang " << __clang_major__ << "." << __clang_minor__ << "."
+       << __clang_patchlevel__;
+  }
+#endif
+
+#if defined(_MSC_VER)
+  { ss << "MSVC " << _MSC_FULL_VER; }
+#endif
+  return ss.str();
+}
+
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
+  m.def("get_compiler_version", &get_compiler_version, "get_compiler_version");
+  m.def("get_cuda_version", &get_cuda_version, "get_cuda_version");
+  m.def("has_cuda", &has_cuda, "has_cuda");
+
+  m.def("deform_conv_forward", &deform_conv_forward, "deform_conv_forward");
+  m.def(
+      "deform_conv_backward_input",
+      &deform_conv_backward_input,
+      "deform_conv_backward_input");
+  m.def(
+      "deform_conv_backward_filter",
+      &deform_conv_backward_filter,
+      "deform_conv_backward_filter");
+  m.def(
+      "modulated_deform_conv_forward",
+      &modulated_deform_conv_forward,
+      "modulated_deform_conv_forward");
+  m.def(
+      "modulated_deform_conv_backward",
+      &modulated_deform_conv_backward,
+      "modulated_deform_conv_backward");
+
+  m.def("COCOevalAccumulate", &COCOeval::Accumulate, "COCOeval::Accumulate");
+  m.def(
+      "COCOevalEvaluateImages",
+      &COCOeval::EvaluateImages,
+      "COCOeval::EvaluateImages");
+  pybind11::class_<COCOeval::InstanceAnnotation>(m, "InstanceAnnotation")
+      .def(pybind11::init<uint64_t, double, double, bool, bool>());
+  pybind11::class_<COCOeval::ImageEvaluation>(m, "ImageEvaluation")
+      .def(pybind11::init<>());
+}
+
+TORCH_LIBRARY(detectron2, m) {
+  m.def("nms_rotated", &nms_rotated);
+  m.def("box_iou_rotated", &box_iou_rotated);
+  m.def("roi_align_rotated_forward", &ROIAlignRotated_forward);
+  m.def("roi_align_rotated_backward", &ROIAlignRotated_backward);
+}
+} // namespace detectron2

detectron2/layers/deform_conv.py

+# Copyright (c) Facebook, Inc. and its affiliates.
+import math
+from functools import lru_cache
+import torch
+from torch import nn
+from torch.autograd import Function
+from torch.autograd.function import once_differentiable
+from torch.nn.modules.utils import _pair
+from torchvision.ops import deform_conv2d
+
+from detectron2.utils.develop import create_dummy_class, create_dummy_func
+
+from .wrappers import _NewEmptyTensorOp
+
+
+class _DeformConv(Function):
+    @staticmethod
+    def forward(
+        ctx,
+        input,
+        offset,
+        weight,
+        stride=1,
+        padding=0,
+        dilation=1,
+        groups=1,
+        deformable_groups=1,
+        im2col_step=64,
+    ):
+        if input is not None and input.dim() != 4:
+            raise ValueError(
+                "Expected 4D tensor as input, got {}D tensor instead.".format(input.dim())
+            )
+        ctx.stride = _pair(stride)
+        ctx.padding = _pair(padding)
+        ctx.dilation = _pair(dilation)
+        ctx.groups = groups
+        ctx.deformable_groups = deformable_groups
+        ctx.im2col_step = im2col_step
+
+        ctx.save_for_backward(input, offset, weight)
+
+        output = input.new_empty(
+            _DeformConv._output_size(input, weight, ctx.padding, ctx.dilation, ctx.stride)
+        )
+
+        ctx.bufs_ = [input.new_empty(0), input.new_empty(0)]  # columns, ones
+
+        if not input.is_cuda:
+            # TODO: let torchvision support full features of our deformconv.
+            if deformable_groups != 1:
+                raise NotImplementedError(
+                    "Deformable Conv with deformable_groups != 1 is not supported on CPUs!"
+                )
+            return deform_conv2d(
+                input, offset, weight, stride=stride, padding=padding, dilation=dilation
+            )
+        else:
+            cur_im2col_step = _DeformConv._cal_im2col_step(input.shape[0], ctx.im2col_step)
+            assert (input.shape[0] % cur_im2col_step) == 0, "im2col step must divide batchsize"
+
+            _C.deform_conv_forward(
+                input,
+                weight,
+                offset,
+                output,
+                ctx.bufs_[0],
+                ctx.bufs_[1],
+                weight.size(3),
+                weight.size(2),
+                ctx.stride[1],
+                ctx.stride[0],
+                ctx.padding[1],
+                ctx.padding[0],
+                ctx.dilation[1],
+                ctx.dilation[0],
+                ctx.groups,
+                ctx.deformable_groups,
+                cur_im2col_step,
+            )
+        return output
+
+    @staticmethod
+    @once_differentiable
+    def backward(ctx, grad_output):
+        input, offset, weight = ctx.saved_tensors
+
+        grad_input = grad_offset = grad_weight = None
+
+        if not grad_output.is_cuda:
+            raise NotImplementedError("Deformable Conv is not supported on CPUs!")
+        else:
+            cur_im2col_step = _DeformConv._cal_im2col_step(input.shape[0], ctx.im2col_step)
+            assert (input.shape[0] % cur_im2col_step) == 0, "im2col step must divide batchsize"
+
+            if ctx.needs_input_grad[0] or ctx.needs_input_grad[1]:
+                grad_input = torch.zeros_like(input)
+                grad_offset = torch.zeros_like(offset)
+                _C.deform_conv_backward_input(
+                    input,
+                    offset,
+                    grad_output,
+                    grad_input,
+                    grad_offset,
+                    weight,
+                    ctx.bufs_[0],
+                    weight.size(3),
+                    weight.size(2),
+                    ctx.stride[1],
+                    ctx.stride[0],
+                    ctx.padding[1],
+                    ctx.padding[0],
+                    ctx.dilation[1],
+                    ctx.dilation[0],
+                    ctx.groups,
+                    ctx.deformable_groups,
+                    cur_im2col_step,
+                )
+
+            if ctx.needs_input_grad[2]:
+                grad_weight = torch.zeros_like(weight)
+                _C.deform_conv_backward_filter(
+                    input,
+                    offset,
+                    grad_output,
+                    grad_weight,
+                    ctx.bufs_[0],
+                    ctx.bufs_[1],
+                    weight.size(3),
+                    weight.size(2),
+                    ctx.stride[1],
+                    ctx.stride[0],
+                    ctx.padding[1],
+                    ctx.padding[0],
+                    ctx.dilation[1],
+                    ctx.dilation[0],
+                    ctx.groups,
+                    ctx.deformable_groups,
+                    1,
+                    cur_im2col_step,
+                )
+
+        return grad_input, grad_offset, grad_weight, None, None, None, None, None, None
+
+    @staticmethod
+    def _output_size(input, weight, padding, dilation, stride):
+        channels = weight.size(0)
+        output_size = (input.size(0), channels)
+        for d in range(input.dim() - 2):
+            in_size = input.size(d + 2)
+            pad = padding[d]
+            kernel = dilation[d] * (weight.size(d + 2) - 1) + 1
+            stride_ = stride[d]
+            output_size += ((in_size + (2 * pad) - kernel) // stride_ + 1,)
+        if not all(map(lambda s: s > 0, output_size)):
+            raise ValueError(
+                "convolution input is too small (output would be {})".format(
+                    "x".join(map(str, output_size))
+                )
+            )
+        return output_size
+
+    @staticmethod
+    @lru_cache(maxsize=128)
+    def _cal_im2col_step(input_size, default_size):
+        """
+        Calculate proper im2col step size, which should be divisible by input_size and not larger
+        than prefer_size. Meanwhile the step size should be as large as possible to be more
+        efficient. So we choose the largest one among all divisors of input_size which are smaller
+        than prefer_size.
+        :param input_size: input batch size .
+        :param default_size: default preferred im2col step size.
+        :return: the largest proper step size.
+        """
+        if input_size <= default_size:
+            return input_size
+        best_step = 1
+        for step in range(2, min(int(math.sqrt(input_size)) + 1, default_size)):
+            if input_size % step == 0:
+                if input_size // step <= default_size:
+                    return input_size // step
+                best_step = step
+
+        return best_step
+
+
+class _ModulatedDeformConv(Function):
+    @staticmethod
+    def forward(
+        ctx,
+        input,
+        offset,
+        mask,
+        weight,
+        bias=None,
+        stride=1,
+        padding=0,
+        dilation=1,
+        groups=1,
+        deformable_groups=1,
+    ):
+        ctx.stride = stride
+        ctx.padding = padding
+        ctx.dilation = dilation
+        ctx.groups = groups
+        ctx.deformable_groups = deformable_groups
+        ctx.with_bias = bias is not None
+        if not ctx.with_bias:
+            bias = input.new_empty(1)  # fake tensor
+        if not input.is_cuda:
+            raise NotImplementedError("Deformable Conv is not supported on CPUs!")
+        if (
+            weight.requires_grad
+            or mask.requires_grad
+            or offset.requires_grad
+            or input.requires_grad
+        ):
+            ctx.save_for_backward(input, offset, mask, weight, bias)
+        output = input.new_empty(_ModulatedDeformConv._infer_shape(ctx, input, weight))
+        ctx._bufs = [input.new_empty(0), input.new_empty(0)]
+        _C.modulated_deform_conv_forward(
+            input,
+            weight,
+            bias,
+            ctx._bufs[0],
+            offset,
+            mask,
+            output,
+            ctx._bufs[1],
+            weight.shape[2],
+            weight.shape[3],
+            ctx.stride,
+            ctx.stride,
+            ctx.padding,
+            ctx.padding,
+            ctx.dilation,
+            ctx.dilation,
+            ctx.groups,
+            ctx.deformable_groups,
+            ctx.with_bias,
+        )
+        return output
+
+    @staticmethod
+    @once_differentiable
+    def backward(ctx, grad_output):
+        if not grad_output.is_cuda:
+            raise NotImplementedError("Deformable Conv is not supported on CPUs!")
+        input, offset, mask, weight, bias = ctx.saved_tensors
+        grad_input = torch.zeros_like(input)
+        grad_offset = torch.zeros_like(offset)
+        grad_mask = torch.zeros_like(mask)
+        grad_weight = torch.zeros_like(weight)
+        grad_bias = torch.zeros_like(bias)
+        _C.modulated_deform_conv_backward(
+            input,
+            weight,
+            bias,
+            ctx._bufs[0],
+            offset,
+            mask,
+            ctx._bufs[1],
+            grad_input,
+            grad_weight,
+            grad_bias,
+            grad_offset,
+            grad_mask,
+            grad_output,
+            weight.shape[2],
+            weight.shape[3],
+            ctx.stride,
+            ctx.stride,
+            ctx.padding,
+            ctx.padding,
+            ctx.dilation,
+            ctx.dilation,
+            ctx.groups,
+            ctx.deformable_groups,
+            ctx.with_bias,
+        )
+        if not ctx.with_bias:
+            grad_bias = None
+
+        return (
+            grad_input,
+            grad_offset,
+            grad_mask,
+            grad_weight,
+            grad_bias,
+            None,
+            None,
+            None,
+            None,
+            None,
+        )
+
+    @staticmethod
+    def _infer_shape(ctx, input, weight):
+        n = input.size(0)
+        channels_out = weight.size(0)
+        height, width = input.shape[2:4]
+        kernel_h, kernel_w = weight.shape[2:4]
+        height_out = (
+            height + 2 * ctx.padding - (ctx.dilation * (kernel_h - 1) + 1)
+        ) // ctx.stride + 1
+        width_out = (
+            width + 2 * ctx.padding - (ctx.dilation * (kernel_w - 1) + 1)
+        ) // ctx.stride + 1
+        return n, channels_out, height_out, width_out
+
+
+deform_conv = _DeformConv.apply
+modulated_deform_conv = _ModulatedDeformConv.apply
+
+
+class DeformConv(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        out_channels,
+        kernel_size,
+        stride=1,
+        padding=0,
+        dilation=1,
+        groups=1,
+        deformable_groups=1,
+        bias=False,
+        norm=None,
+        activation=None,
+    ):
+        """
+        Deformable convolution from :paper:`deformconv`.
+
+        Arguments are similar to :class:`Conv2D`. Extra arguments:
+
+        Args:
+            deformable_groups (int): number of groups used in deformable convolution.
+            norm (nn.Module, optional): a normalization layer
+            activation (callable(Tensor) -> Tensor): a callable activation function
+        """
+        super(DeformConv, self).__init__()
+
+        assert not bias
+        assert in_channels % groups == 0, "in_channels {} cannot be divisible by groups {}".format(
+            in_channels, groups
+        )
+        assert (
+            out_channels % groups == 0
+        ), "out_channels {} cannot be divisible by groups {}".format(out_channels, groups)
+
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.kernel_size = _pair(kernel_size)
+        self.stride = _pair(stride)
+        self.padding = _pair(padding)
+        self.dilation = _pair(dilation)
+        self.groups = groups
+        self.deformable_groups = deformable_groups
+        self.norm = norm
+        self.activation = activation
+
+        self.weight = nn.Parameter(
+            torch.Tensor(out_channels, in_channels // self.groups, *self.kernel_size)
+        )
+        self.bias = None
+
+        nn.init.kaiming_uniform_(self.weight, nonlinearity="relu")
+
+    def forward(self, x, offset):
+        if x.numel() == 0:
+            # When input is empty, we want to return a empty tensor with "correct" shape,
+            # So that the following operations will not panic
+            # if they check for the shape of the tensor.
+            # This computes the height and width of the output tensor
+            output_shape = [
+                (i + 2 * p - (di * (k - 1) + 1)) // s + 1
+                for i, p, di, k, s in zip(
+                    x.shape[-2:], self.padding, self.dilation, self.kernel_size, self.stride
+                )
+            ]
+            output_shape = [x.shape[0], self.weight.shape[0]] + output_shape
+            return _NewEmptyTensorOp.apply(x, output_shape)
+
+        x = deform_conv(
+            x,
+            offset,
+            self.weight,
+            self.stride,
+            self.padding,
+            self.dilation,
+            self.groups,
+            self.deformable_groups,
+        )
+        if self.norm is not None:
+            x = self.norm(x)
+        if self.activation is not None:
+            x = self.activation(x)
+        return x
+
+    def extra_repr(self):
+        tmpstr = "in_channels=" + str(self.in_channels)
+        tmpstr += ", out_channels=" + str(self.out_channels)
+        tmpstr += ", kernel_size=" + str(self.kernel_size)
+        tmpstr += ", stride=" + str(self.stride)
+        tmpstr += ", padding=" + str(self.padding)
+        tmpstr += ", dilation=" + str(self.dilation)
+        tmpstr += ", groups=" + str(self.groups)
+        tmpstr += ", deformable_groups=" + str(self.deformable_groups)
+        tmpstr += ", bias=False"
+        return tmpstr
+
+
+class ModulatedDeformConv(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        out_channels,
+        kernel_size,
+        stride=1,
+        padding=0,
+        dilation=1,
+        groups=1,
+        deformable_groups=1,
+        bias=True,
+        norm=None,
+        activation=None,
+    ):
+        """
+        Modulated deformable convolution from :paper:`deformconv2`.
+
+        Arguments are similar to :class:`Conv2D`. Extra arguments:
+
+        Args:
+            deformable_groups (int): number of groups used in deformable convolution.
+            norm (nn.Module, optional): a normalization layer
+            activation (callable(Tensor) -> Tensor): a callable activation function
+        """
+        super(ModulatedDeformConv, self).__init__()
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.kernel_size = _pair(kernel_size)
+        self.stride = stride
+        self.padding = padding
+        self.dilation = dilation
+        self.groups = groups
+        self.deformable_groups = deformable_groups
+        self.with_bias = bias
+        self.norm = norm
+        self.activation = activation
+
+        self.weight = nn.Parameter(
+            torch.Tensor(out_channels, in_channels // groups, *self.kernel_size)
+        )
+        if bias:
+            self.bias = nn.Parameter(torch.Tensor(out_channels))
+        else:
+            self.bias = None
+
+        nn.init.kaiming_uniform_(self.weight, nonlinearity="relu")
+        if self.bias is not None:
+            nn.init.constant_(self.bias, 0)
+
+    def forward(self, x, offset, mask):
+        if x.numel() == 0:
+            output_shape = [
+                (i + 2 * p - (di * (k - 1) + 1)) // s + 1
+                for i, p, di, k, s in zip(
+                    x.shape[-2:], self.padding, self.dilation, self.kernel_size, self.stride
+                )
+            ]
+            output_shape = [x.shape[0], self.weight.shape[0]] + output_shape
+            return _NewEmptyTensorOp.apply(x, output_shape)
+
+        x = modulated_deform_conv(
+            x,
+            offset,
+            mask,
+            self.weight,
+            self.bias,
+            self.stride,
+            self.padding,
+            self.dilation,
+            self.groups,
+            self.deformable_groups,
+        )
+        if self.norm is not None:
+            x = self.norm(x)
+        if self.activation is not None:
+            x = self.activation(x)
+        return x
+
+    def extra_repr(self):
+        tmpstr = "in_channels=" + str(self.in_channels)
+        tmpstr += ", out_channels=" + str(self.out_channels)
+        tmpstr += ", kernel_size=" + str(self.kernel_size)
+        tmpstr += ", stride=" + str(self.stride)
+        tmpstr += ", padding=" + str(self.padding)
+        tmpstr += ", dilation=" + str(self.dilation)
+        tmpstr += ", groups=" + str(self.groups)
+        tmpstr += ", deformable_groups=" + str(self.deformable_groups)
+        tmpstr += ", bias=" + str(self.with_bias)
+        return tmpstr
+
+
+try:
+    from detectron2 import _C
+except ImportError:
+    # TODO: register ops natively so there is no need to import _C.
+    _msg = "detectron2 is not compiled successfully, please build following the instructions!"
+    _args = ("detectron2._C", _msg)
+    DeformConv = create_dummy_class("DeformConv", *_args)
+    ModulatedDeformConv = create_dummy_class("ModulatedDeformConv", *_args)
+    deform_conv = create_dummy_func("deform_conv", *_args)
+    modulated_deform_conv = create_dummy_func("modulated_deform_conv", *_args)

detectron2/layers/losses.py

+import math
+import torch
+
+
+def diou_loss(
+    boxes1: torch.Tensor,
+    boxes2: torch.Tensor,
+    reduction: str = "none",
+    eps: float = 1e-7,
+) -> torch.Tensor:
+    """
+    Distance Intersection over Union Loss (Zhaohui Zheng et. al)
+    https://arxiv.org/abs/1911.08287
+    Args:
+        boxes1, boxes2 (Tensor): box locations in XYXY format, shape (N, 4) or (4,).
+        reduction: 'none' | 'mean' | 'sum'
+                 'none': No reduction will be applied to the output.
+                 'mean': The output will be averaged.
+                 'sum': The output will be summed.
+        eps (float): small number to prevent division by zero
+    """
+
+    x1, y1, x2, y2 = boxes1.unbind(dim=-1)
+    x1g, y1g, x2g, y2g = boxes2.unbind(dim=-1)
+
+    # TODO: use torch._assert_async() when pytorch 1.8 support is dropped
+    assert (x2 >= x1).all(), "bad box: x1 larger than x2"
+    assert (y2 >= y1).all(), "bad box: y1 larger than y2"
+
+    # Intersection keypoints
+    xkis1 = torch.max(x1, x1g)
+    ykis1 = torch.max(y1, y1g)
+    xkis2 = torch.min(x2, x2g)
+    ykis2 = torch.min(y2, y2g)
+
+    intsct = torch.zeros_like(x1)
+    mask = (ykis2 > ykis1) & (xkis2 > xkis1)
+    intsct[mask] = (xkis2[mask] - xkis1[mask]) * (ykis2[mask] - ykis1[mask])
+    union = (x2 - x1) * (y2 - y1) + (x2g - x1g) * (y2g - y1g) - intsct + eps
+    iou = intsct / union
+
+    # smallest enclosing box
+    xc1 = torch.min(x1, x1g)
+    yc1 = torch.min(y1, y1g)
+    xc2 = torch.max(x2, x2g)
+    yc2 = torch.max(y2, y2g)
+    diag_len = ((xc2 - xc1) ** 2) + ((yc2 - yc1) ** 2) + eps
+
+    # centers of boxes
+    x_p = (x2 + x1) / 2
+    y_p = (y2 + y1) / 2
+    x_g = (x1g + x2g) / 2
+    y_g = (y1g + y2g) / 2
+    distance = ((x_p - x_g) ** 2) + ((y_p - y_g) ** 2)
+
+    # Eqn. (7)
+    loss = 1 - iou + (distance / diag_len)
+    if reduction == "mean":
+        loss = loss.mean() if loss.numel() > 0 else 0.0 * loss.sum()
+    elif reduction == "sum":
+        loss = loss.sum()
+
+    return loss
+
+
+def ciou_loss(
+    boxes1: torch.Tensor,
+    boxes2: torch.Tensor,
+    reduction: str = "none",
+    eps: float = 1e-7,
+) -> torch.Tensor:
+    """
+    Complete Intersection over Union Loss (Zhaohui Zheng et. al)
+    https://arxiv.org/abs/1911.08287
+    Args:
+        boxes1, boxes2 (Tensor): box locations in XYXY format, shape (N, 4) or (4,).
+        reduction: 'none' | 'mean' | 'sum'
+                 'none': No reduction will be applied to the output.
+                 'mean': The output will be averaged.
+                 'sum': The output will be summed.
+        eps (float): small number to prevent division by zero
+    """
+
+    x1, y1, x2, y2 = boxes1.unbind(dim=-1)
+    x1g, y1g, x2g, y2g = boxes2.unbind(dim=-1)
+
+    # TODO: use torch._assert_async() when pytorch 1.8 support is dropped
+    assert (x2 >= x1).all(), "bad box: x1 larger than x2"
+    assert (y2 >= y1).all(), "bad box: y1 larger than y2"
+
+    # Intersection keypoints
+    xkis1 = torch.max(x1, x1g)
+    ykis1 = torch.max(y1, y1g)
+    xkis2 = torch.min(x2, x2g)
+    ykis2 = torch.min(y2, y2g)
+
+    intsct = torch.zeros_like(x1)
+    mask = (ykis2 > ykis1) & (xkis2 > xkis1)
+    intsct[mask] = (xkis2[mask] - xkis1[mask]) * (ykis2[mask] - ykis1[mask])
+    union = (x2 - x1) * (y2 - y1) + (x2g - x1g) * (y2g - y1g) - intsct + eps
+    iou = intsct / union
+
+    # smallest enclosing box
+    xc1 = torch.min(x1, x1g)
+    yc1 = torch.min(y1, y1g)
+    xc2 = torch.max(x2, x2g)
+    yc2 = torch.max(y2, y2g)
+    diag_len = ((xc2 - xc1) ** 2) + ((yc2 - yc1) ** 2) + eps
+
+    # centers of boxes
+    x_p = (x2 + x1) / 2
+    y_p = (y2 + y1) / 2
+    x_g = (x1g + x2g) / 2
+    y_g = (y1g + y2g) / 2
+    distance = ((x_p - x_g) ** 2) + ((y_p - y_g) ** 2)
+
+    # width and height of boxes
+    w_pred = x2 - x1
+    h_pred = y2 - y1
+    w_gt = x2g - x1g
+    h_gt = y2g - y1g
+    v = (4 / (math.pi**2)) * torch.pow((torch.atan(w_gt / h_gt) - torch.atan(w_pred / h_pred)), 2)
+    with torch.no_grad():
+        alpha = v / (1 - iou + v + eps)
+
+    # Eqn. (10)
+    loss = 1 - iou + (distance / diag_len) + alpha * v
+    if reduction == "mean":
+        loss = loss.mean() if loss.numel() > 0 else 0.0 * loss.sum()
+    elif reduction == "sum":
+        loss = loss.sum()
+
+    return loss

detectron2/layers/mask_ops.py

+# Copyright (c) Facebook, Inc. and its affiliates.
+import numpy as np
+from typing import Tuple
+import torch
+from PIL import Image
+from torch.nn import functional as F
+
+__all__ = ["paste_masks_in_image"]
+
+
+BYTES_PER_FLOAT = 4
+# TODO: This memory limit may be too much or too little. It would be better to
+# determine it based on available resources.
+GPU_MEM_LIMIT = 1024**3  # 1 GB memory limit
+
+
+def _do_paste_mask(masks, boxes, img_h: int, img_w: int, skip_empty: bool = True):
+    """
+    Args:
+        masks: N, 1, H, W
+        boxes: N, 4
+        img_h, img_w (int):
+        skip_empty (bool): only paste masks within the region that
+            tightly bound all boxes, and returns the results this region only.
+            An important optimization for CPU.
+
+    Returns:
+        if skip_empty == False, a mask of shape (N, img_h, img_w)
+        if skip_empty == True, a mask of shape (N, h', w'), and the slice
+            object for the corresponding region.
+    """
+    # On GPU, paste all masks together (up to chunk size)
+    # by using the entire image to sample the masks
+    # Compared to pasting them one by one,
+    # this has more operations but is faster on COCO-scale dataset.
+    device = masks.device
+
+    if skip_empty and not torch.jit.is_scripting():
+        x0_int, y0_int = torch.clamp(boxes.min(dim=0).values.floor()[:2] - 1, min=0).to(
+            dtype=torch.int32
+        )
+        x1_int = torch.clamp(boxes[:, 2].max().ceil() + 1, max=img_w).to(dtype=torch.int32)
+        y1_int = torch.clamp(boxes[:, 3].max().ceil() + 1, max=img_h).to(dtype=torch.int32)
+    else:
+        x0_int, y0_int = 0, 0
+        x1_int, y1_int = img_w, img_h
+    x0, y0, x1, y1 = torch.split(boxes, 1, dim=1)  # each is Nx1
+
+    N = masks.shape[0]
+
+    img_y = torch.arange(y0_int, y1_int, device=device, dtype=torch.float32) + 0.5
+    img_x = torch.arange(x0_int, x1_int, device=device, dtype=torch.float32) + 0.5
+    img_y = (img_y - y0) / (y1 - y0) * 2 - 1
+    img_x = (img_x - x0) / (x1 - x0) * 2 - 1
+    # img_x, img_y have shapes (N, w), (N, h)
+
+    gx = img_x[:, None, :].expand(N, img_y.size(1), img_x.size(1))
+    gy = img_y[:, :, None].expand(N, img_y.size(1), img_x.size(1))
+    grid = torch.stack([gx, gy], dim=3)
+
+    if not torch.jit.is_scripting():
+        if not masks.dtype.is_floating_point:
+            masks = masks.float()
+    img_masks = F.grid_sample(masks, grid.to(masks.dtype), align_corners=False)
+
+    if skip_empty and not torch.jit.is_scripting():
+        return img_masks[:, 0], (slice(y0_int, y1_int), slice(x0_int, x1_int))
+    else:
+        return img_masks[:, 0], ()
+
+
+# Annotate boxes as Tensor (but not Boxes) in order to use scripting
+@torch.jit.script_if_tracing
+def paste_masks_in_image(
+    masks: torch.Tensor, boxes: torch.Tensor, image_shape: Tuple[int, int], threshold: float = 0.5
+):
+    """
+    Paste a set of masks that are of a fixed resolution (e.g., 28 x 28) into an image.
+    The location, height, and width for pasting each mask is determined by their
+    corresponding bounding boxes in boxes.
+
+    Note:
+        This is a complicated but more accurate implementation. In actual deployment, it is
+        often enough to use a faster but less accurate implementation.
+        See :func:`paste_mask_in_image_old` in this file for an alternative implementation.
+
+    Args:
+        masks (tensor): Tensor of shape (Bimg, Hmask, Wmask), where Bimg is the number of
+            detected object instances in the image and Hmask, Wmask are the mask width and mask
+            height of the predicted mask (e.g., Hmask = Wmask = 28). Values are in [0, 1].
+        boxes (Boxes or Tensor): A Boxes of length Bimg or Tensor of shape (Bimg, 4).
+            boxes[i] and masks[i] correspond to the same object instance.
+        image_shape (tuple): height, width
+        threshold (float): A threshold in [0, 1] for converting the (soft) masks to
+            binary masks.
+
+    Returns:
+        img_masks (Tensor): A tensor of shape (Bimg, Himage, Wimage), where Bimg is the
+        number of detected object instances and Himage, Wimage are the image width
+        and height. img_masks[i] is a binary mask for object instance i.
+    """
+
+    assert masks.shape[-1] == masks.shape[-2], "Only square mask predictions are supported"
+    N = len(masks)
+    if N == 0:
+        return masks.new_empty((0,) + image_shape, dtype=torch.uint8)
+    if not isinstance(boxes, torch.Tensor):
+        boxes = boxes.tensor
+    device = boxes.device
+    assert len(boxes) == N, boxes.shape
+
+    img_h, img_w = image_shape
+
+    # The actual implementation split the input into chunks,
+    # and paste them chunk by chunk.
+    if device.type == "cpu" or torch.jit.is_scripting():
+        # CPU is most efficient when they are pasted one by one with skip_empty=True
+        # so that it performs minimal number of operations.
+        num_chunks = N
+    else:
+        # GPU benefits from parallelism for larger chunks, but may have memory issue
+        # int(img_h) because shape may be tensors in tracing
+        num_chunks = int(np.ceil(N * int(img_h) * int(img_w) * BYTES_PER_FLOAT / GPU_MEM_LIMIT))
+        assert (
+            num_chunks <= N
+        ), "Default GPU_MEM_LIMIT in mask_ops.py is too small; try increasing it"
+    chunks = torch.chunk(torch.arange(N, device=device), num_chunks)
+
+    img_masks = torch.zeros(
+        N, img_h, img_w, device=device, dtype=torch.bool if threshold >= 0 else torch.uint8
+    )
+    for inds in chunks:
+        masks_chunk, spatial_inds = _do_paste_mask(
+            masks[inds, None, :, :], boxes[inds], img_h, img_w, skip_empty=device.type == "cpu"
+        )
+
+        if threshold >= 0:
+            masks_chunk = (masks_chunk >= threshold).to(dtype=torch.bool)
+        else:
+            # for visualization and debugging
+            masks_chunk = (masks_chunk * 255).to(dtype=torch.uint8)
+
+        if torch.jit.is_scripting():  # Scripting does not use the optimized codepath
+            img_masks[inds] = masks_chunk
+        else:
+            img_masks[(inds,) + spatial_inds] = masks_chunk
+    return img_masks
+
+
+# The below are the original paste function (from Detectron1) which has
+# larger quantization error.
+# It is faster on CPU, while the aligned one is faster on GPU thanks to grid_sample.
+
+
+def paste_mask_in_image_old(mask, box, img_h, img_w, threshold):
+    """
+    Paste a single mask in an image.
+    This is a per-box implementation of :func:`paste_masks_in_image`.
+    This function has larger quantization error due to incorrect pixel
+    modeling and is not used any more.
+
+    Args:
+        mask (Tensor): A tensor of shape (Hmask, Wmask) storing the mask of a single
+            object instance. Values are in [0, 1].
+        box (Tensor): A tensor of shape (4, ) storing the x0, y0, x1, y1 box corners
+            of the object instance.
+        img_h, img_w (int): Image height and width.
+        threshold (float): Mask binarization threshold in [0, 1].
+
+    Returns:
+        im_mask (Tensor):
+            The resized and binarized object mask pasted into the original
+            image plane (a tensor of shape (img_h, img_w)).
+    """
+    # Conversion from continuous box coordinates to discrete pixel coordinates
+    # via truncation (cast to int32). This determines which pixels to paste the
+    # mask onto.
+    box = box.to(dtype=torch.int32)  # Continuous to discrete coordinate conversion
+    # An example (1D) box with continuous coordinates (x0=0.7, x1=4.3) will map to
+    # a discrete coordinates (x0=0, x1=4). Note that box is mapped to 5 = x1 - x0 + 1
+    # pixels (not x1 - x0 pixels).
+    samples_w = box[2] - box[0] + 1  # Number of pixel samples, *not* geometric width
+    samples_h = box[3] - box[1] + 1  # Number of pixel samples, *not* geometric height
+
+    # Resample the mask from it's original grid to the new samples_w x samples_h grid
+    mask = Image.fromarray(mask.cpu().numpy())
+    mask = mask.resize((samples_w, samples_h), resample=Image.BILINEAR)
+    mask = np.asarray(mask)
+
+    if threshold >= 0:
+        mask = np.array(mask > threshold, dtype=np.uint8)
+        mask = torch.from_numpy(mask)
+    else:
+        # for visualization and debugging, we also
+        # allow it to return an unmodified mask
+        mask = torch.from_numpy(mask * 255).to(torch.uint8)
+
+    im_mask = torch.zeros((img_h, img_w), dtype=torch.uint8)
+    x_0 = max(box[0], 0)
+    x_1 = min(box[2] + 1, img_w)
+    y_0 = max(box[1], 0)
+    y_1 = min(box[3] + 1, img_h)
+
+    im_mask[y_0:y_1, x_0:x_1] = mask[
+        (y_0 - box[1]) : (y_1 - box[1]), (x_0 - box[0]) : (x_1 - box[0])
+    ]
+    return im_mask
+
+
+# Our pixel modeling requires extrapolation for any continuous
+# coordinate < 0.5 or > length - 0.5. When sampling pixels on the masks,
+# we would like this extrapolation to be an interpolation between boundary values and zero,
+# instead of using absolute zero or boundary values.
+# Therefore `paste_mask_in_image_old` is often used with zero padding around the masks like this:
+# masks, scale = pad_masks(masks[:, 0, :, :], 1)
+# boxes = scale_boxes(boxes.tensor, scale)
+
+
+def pad_masks(masks, padding):
+    """
+    Args:
+        masks (tensor): A tensor of shape (B, M, M) representing B masks.
+        padding (int): Number of cells to pad on all sides.
+
+    Returns:
+        The padded masks and the scale factor of the padding size / original size.
+    """
+    B = masks.shape[0]
+    M = masks.shape[-1]
+    pad2 = 2 * padding
+    scale = float(M + pad2) / M
+    padded_masks = masks.new_zeros((B, M + pad2, M + pad2))
+    padded_masks[:, padding:-padding, padding:-padding] = masks
+    return padded_masks, scale
+
+
+def scale_boxes(boxes, scale):
+    """
+    Args:
+        boxes (tensor): A tensor of shape (B, 4) representing B boxes with 4
+            coords representing the corners x0, y0, x1, y1,
+        scale (float): The box scaling factor.
+
+    Returns:
+        Scaled boxes.
+    """
+    w_half = (boxes[:, 2] - boxes[:, 0]) * 0.5
+    h_half = (boxes[:, 3] - boxes[:, 1]) * 0.5
+    x_c = (boxes[:, 2] + boxes[:, 0]) * 0.5
+    y_c = (boxes[:, 3] + boxes[:, 1]) * 0.5
+
+    w_half *= scale
+    h_half *= scale
+
+    scaled_boxes = torch.zeros_like(boxes)
+    scaled_boxes[:, 0] = x_c - w_half
+    scaled_boxes[:, 2] = x_c + w_half
+    scaled_boxes[:, 1] = y_c - h_half
+    scaled_boxes[:, 3] = y_c + h_half
+    return scaled_boxes
+
+
+@torch.jit.script_if_tracing
+def _paste_masks_tensor_shape(
+    masks: torch.Tensor,
+    boxes: torch.Tensor,
+    image_shape: Tuple[torch.Tensor, torch.Tensor],
+    threshold: float = 0.5,
+):
+    """
+    A wrapper of paste_masks_in_image where image_shape is Tensor.
+    During tracing, shapes might be tensors instead of ints. The Tensor->int
+    conversion should be scripted rather than traced.
+    """
+    return paste_masks_in_image(masks, boxes, (int(image_shape[0]), int(image_shape[1])), threshold)

detectron2/layers/nms.py

+# -*- coding: utf-8 -*-
+# Copyright (c) Facebook, Inc. and its affiliates.
+
+import torch
+from torchvision.ops import boxes as box_ops
+from torchvision.ops import nms  # noqa . for compatibility
+
+from detectron2.layers.wrappers import disable_torch_compiler
+
+
+def batched_nms(
+    boxes: torch.Tensor, scores: torch.Tensor, idxs: torch.Tensor, iou_threshold: float
+):
+    """
+    Same as torchvision.ops.boxes.batched_nms, but with float().
+    """
+    assert boxes.shape[-1] == 4
+    # Note: Torchvision already has a strategy (https://github.com/pytorch/vision/issues/1311)
+    # to decide whether to use coordinate trick or for loop to implement batched_nms. So we
+    # just call it directly.
+    # Fp16 does not have enough range for batched NMS, so adding float().
+    return box_ops.batched_nms(boxes.float(), scores, idxs, iou_threshold)
+
+
+# Note: this function (nms_rotated) might be moved into
+# torchvision/ops/boxes.py in the future
+@disable_torch_compiler
+def nms_rotated(boxes: torch.Tensor, scores: torch.Tensor, iou_threshold: float):
+    """
+    Performs non-maximum suppression (NMS) on the rotated boxes according
+    to their intersection-over-union (IoU).
+
+    Rotated NMS iteratively removes lower scoring rotated boxes which have an
+    IoU greater than iou_threshold with another (higher scoring) rotated box.
+
+    Note that RotatedBox (5, 3, 4, 2, -90) covers exactly the same region as
+    RotatedBox (5, 3, 4, 2, 90) does, and their IoU will be 1. However, they
+    can be representing completely different objects in certain tasks, e.g., OCR.
+
+    As for the question of whether rotated-NMS should treat them as faraway boxes
+    even though their IOU is 1, it depends on the application and/or ground truth annotation.
+
+    As an extreme example, consider a single character v and the square box around it.
+
+    If the angle is 0 degree, the object (text) would be read as 'v';
+
+    If the angle is 90 degrees, the object (text) would become '>';
+
+    If the angle is 180 degrees, the object (text) would become '^';
+
+    If the angle is 270/-90 degrees, the object (text) would become '<'
+
+    All of these cases have IoU of 1 to each other, and rotated NMS that only
+    uses IoU as criterion would only keep one of them with the highest score -
+    which, practically, still makes sense in most cases because typically
+    only one of theses orientations is the correct one. Also, it does not matter
+    as much if the box is only used to classify the object (instead of transcribing
+    them with a sequential OCR recognition model) later.
+
+    On the other hand, when we use IoU to filter proposals that are close to the
+    ground truth during training, we should definitely take the angle into account if
+    we know the ground truth is labeled with the strictly correct orientation (as in,
+    upside-down words are annotated with -180 degrees even though they can be covered
+    with a 0/90/-90 degree box, etc.)
+
+    The way the original dataset is annotated also matters. For example, if the dataset
+    is a 4-point polygon dataset that does not enforce ordering of vertices/orientation,
+    we can estimate a minimum rotated bounding box to this polygon, but there's no way
+    we can tell the correct angle with 100% confidence (as shown above, there could be 4 different
+    rotated boxes, with angles differed by 90 degrees to each other, covering the exactly
+    same region). In that case we have to just use IoU to determine the box
+    proximity (as many detection benchmarks (even for text) do) unless there're other
+    assumptions we can make (like width is always larger than height, or the object is not
+    rotated by more than 90 degrees CCW/CW, etc.)
+
+    In summary, not considering angles in rotated NMS seems to be a good option for now,
+    but we should be aware of its implications.
+
+    Args:
+        boxes (Tensor[N, 5]): Rotated boxes to perform NMS on. They are expected to be in
+           (x_center, y_center, width, height, angle_degrees) format.
+        scores (Tensor[N]): Scores for each one of the rotated boxes
+        iou_threshold (float): Discards all overlapping rotated boxes with IoU < iou_threshold
+
+    Returns:
+        keep (Tensor): int64 tensor with the indices of the elements that have been kept
+        by Rotated NMS, sorted in decreasing order of scores
+    """
+    return torch.ops.detectron2.nms_rotated(boxes, scores, iou_threshold)
+
+
+# Note: this function (batched_nms_rotated) might be moved into
+# torchvision/ops/boxes.py in the future
+
+
+@torch.jit.script_if_tracing
+def batched_nms_rotated(
+    boxes: torch.Tensor, scores: torch.Tensor, idxs: torch.Tensor, iou_threshold: float
+):
+    """
+    Performs non-maximum suppression in a batched fashion.
+
+    Each index value correspond to a category, and NMS
+    will not be applied between elements of different categories.
+
+    Args:
+        boxes (Tensor[N, 5]):
+           boxes where NMS will be performed. They
+           are expected to be in (x_ctr, y_ctr, width, height, angle_degrees) format
+        scores (Tensor[N]):
+           scores for each one of the boxes
+        idxs (Tensor[N]):
+           indices of the categories for each one of the boxes.
+        iou_threshold (float):
+           discards all overlapping boxes
+           with IoU < iou_threshold
+
+    Returns:
+        Tensor:
+            int64 tensor with the indices of the elements that have been kept
+            by NMS, sorted in decreasing order of scores
+    """
+    assert boxes.shape[-1] == 5
+
+    if boxes.numel() == 0:
+        return torch.empty((0,), dtype=torch.int64, device=boxes.device)
+    boxes = boxes.float()  # fp16 does not have enough range for batched NMS
+    # Strategy: in order to perform NMS independently per class,
+    # we add an offset to all the boxes. The offset is dependent
+    # only on the class idx, and is large enough so that boxes
+    # from different classes do not overlap
+
+    # Note that batched_nms in torchvision/ops/boxes.py only uses max_coordinate,
+    # which won't handle negative coordinates correctly.
+    # Here by using min_coordinate we can make sure the negative coordinates are
+    # correctly handled.
+    max_coordinate = (
+        torch.max(boxes[:, 0], boxes[:, 1]) + torch.max(boxes[:, 2], boxes[:, 3]) / 2
+    ).max()
+    min_coordinate = (
+        torch.min(boxes[:, 0], boxes[:, 1]) - torch.max(boxes[:, 2], boxes[:, 3]) / 2
+    ).min()
+    offsets = idxs.to(boxes) * (max_coordinate - min_coordinate + 1)
+    boxes_for_nms = boxes.clone()  # avoid modifying the original values in boxes
+    boxes_for_nms[:, :2] += offsets[:, None]
+    keep = nms_rotated(boxes_for_nms, scores, iou_threshold)
+    return keep

detectron2/layers/roi_align.py

+# Copyright (c) Facebook, Inc. and its affiliates.
+from torch import nn
+from torchvision.ops import roi_align
+
+
+# NOTE: torchvision's RoIAlign has a different default aligned=False
+class ROIAlign(nn.Module):
+    def __init__(self, output_size, spatial_scale, sampling_ratio, aligned=True):
+        """
+        Args:
+            output_size (tuple): h, w
+            spatial_scale (float): scale the input boxes by this number
+            sampling_ratio (int): number of inputs samples to take for each output
+                sample. 0 to take samples densely.
+            aligned (bool): if False, use the legacy implementation in
+                Detectron. If True, align the results more perfectly.
+
+        Note:
+            The meaning of aligned=True:
+
+            Given a continuous coordinate c, its two neighboring pixel indices (in our
+            pixel model) are computed by floor(c - 0.5) and ceil(c - 0.5). For example,
+            c=1.3 has pixel neighbors with discrete indices [0] and [1] (which are sampled
+            from the underlying signal at continuous coordinates 0.5 and 1.5). But the original
+            roi_align (aligned=False) does not subtract the 0.5 when computing neighboring
+            pixel indices and therefore it uses pixels with a slightly incorrect alignment
+            (relative to our pixel model) when performing bilinear interpolation.
+
+            With `aligned=True`,
+            we first appropriately scale the ROI and then shift it by -0.5
+            prior to calling roi_align. This produces the correct neighbors; see
+            detectron2/tests/test_roi_align.py for verification.
+
+            The difference does not make a difference to the model's performance if
+            ROIAlign is used together with conv layers.
+        """
+        super().__init__()
+        self.output_size = output_size
+        self.spatial_scale = spatial_scale
+        self.sampling_ratio = sampling_ratio
+        self.aligned = aligned
+
+        from torchvision import __version__
+
+        version = tuple(int(x) for x in __version__.split(".")[:2])
+        # https://github.com/pytorch/vision/pull/2438
+        assert version >= (0, 7), "Require torchvision >= 0.7"
+
+    def forward(self, input, rois):
+        """
+        Args:
+            input: NCHW images
+            rois: Bx5 boxes. First column is the index into N. The other 4 columns are xyxy.
+        """
+        assert rois.dim() == 2 and rois.size(1) == 5
+        if input.is_quantized:
+            input = input.dequantize()
+        return roi_align(
+            input,
+            rois.to(dtype=input.dtype),
+            self.output_size,
+            self.spatial_scale,
+            self.sampling_ratio,
+            self.aligned,
+        )
+
+    def __repr__(self):
+        tmpstr = self.__class__.__name__ + "("
+        tmpstr += "output_size=" + str(self.output_size)
+        tmpstr += ", spatial_scale=" + str(self.spatial_scale)
+        tmpstr += ", sampling_ratio=" + str(self.sampling_ratio)
+        tmpstr += ", aligned=" + str(self.aligned)
+        tmpstr += ")"
+        return tmpstr

detectron2/layers/roi_align_rotated.py

+# Copyright (c) Facebook, Inc. and its affiliates.
+import torch
+from torch import nn
+from torch.autograd import Function
+from torch.autograd.function import once_differentiable
+from torch.nn.modules.utils import _pair
+
+from detectron2.layers.wrappers import disable_torch_compiler
+
+
+class _ROIAlignRotated(Function):
+    @staticmethod
+    @disable_torch_compiler
+    def forward(ctx, input, roi, output_size, spatial_scale, sampling_ratio):
+        ctx.save_for_backward(roi)
+        ctx.output_size = _pair(output_size)
+        ctx.spatial_scale = spatial_scale
+        ctx.sampling_ratio = sampling_ratio
+        ctx.input_shape = input.size()
+        output = torch.ops.detectron2.roi_align_rotated_forward(
+            input, roi, spatial_scale, output_size[0], output_size[1], sampling_ratio
+        )
+        return output
+
+    @staticmethod
+    @once_differentiable
+    def backward(ctx, grad_output):
+        (rois,) = ctx.saved_tensors
+        output_size = ctx.output_size
+        spatial_scale = ctx.spatial_scale
+        sampling_ratio = ctx.sampling_ratio
+        bs, ch, h, w = ctx.input_shape
+        grad_input = torch.ops.detectron2.roi_align_rotated_backward(
+            grad_output,
+            rois,
+            spatial_scale,
+            output_size[0],
+            output_size[1],
+            bs,
+            ch,
+            h,
+            w,
+            sampling_ratio,
+        )
+        return grad_input, None, None, None, None, None
+
+
+roi_align_rotated = _ROIAlignRotated.apply
+
+
+class ROIAlignRotated(nn.Module):
+    def __init__(self, output_size, spatial_scale, sampling_ratio):
+        """
+        Args:
+            output_size (tuple): h, w
+            spatial_scale (float): scale the input boxes by this number
+            sampling_ratio (int): number of inputs samples to take for each output
+                sample. 0 to take samples densely.
+
+        Note:
+            ROIAlignRotated supports continuous coordinate by default:
+            Given a continuous coordinate c, its two neighboring pixel indices (in our
+            pixel model) are computed by floor(c - 0.5) and ceil(c - 0.5). For example,
+            c=1.3 has pixel neighbors with discrete indices [0] and [1] (which are sampled
+            from the underlying signal at continuous coordinates 0.5 and 1.5).
+        """
+        super(ROIAlignRotated, self).__init__()
+        self.output_size = output_size
+        self.spatial_scale = spatial_scale
+        self.sampling_ratio = sampling_ratio
+
+    def forward(self, input, rois):
+        """
+        Args:
+            input: NCHW images
+            rois: Bx6 boxes. First column is the index into N.
+                The other 5 columns are (x_ctr, y_ctr, width, height, angle_degrees).
+        """
+        assert rois.dim() == 2 and rois.size(1) == 6
+        orig_dtype = input.dtype
+        if orig_dtype == torch.float16:
+            input = input.float()
+            rois = rois.float()
+        output_size = _pair(self.output_size)
+
+        # Scripting for Autograd is currently unsupported.
+        # This is a quick fix without having to rewrite code on the C++ side
+        if torch.jit.is_scripting() or torch.jit.is_tracing():
+            return torch.ops.detectron2.roi_align_rotated_forward(
+                input, rois, self.spatial_scale, output_size[0], output_size[1], self.sampling_ratio
+            ).to(dtype=orig_dtype)
+
+        return roi_align_rotated(
+            input, rois, self.output_size, self.spatial_scale, self.sampling_ratio
+        ).to(dtype=orig_dtype)
+
+    def __repr__(self):
+        tmpstr = self.__class__.__name__ + "("
+        tmpstr += "output_size=" + str(self.output_size)
+        tmpstr += ", spatial_scale=" + str(self.spatial_scale)
+        tmpstr += ", sampling_ratio=" + str(self.sampling_ratio)
+        tmpstr += ")"
+        return tmpstr

detectron2/layers/rotated_boxes.py

+# Copyright (c) Facebook, Inc. and its affiliates.
+from __future__ import absolute_import, division, print_function, unicode_literals
+import torch
+
+
+def pairwise_iou_rotated(boxes1, boxes2):
+    """
+    Return intersection-over-union (Jaccard index) of boxes.
+
+    Both sets of boxes are expected to be in
+    (x_center, y_center, width, height, angle) format.
+
+    Arguments:
+        boxes1 (Tensor[N, 5])
+        boxes2 (Tensor[M, 5])
+
+    Returns:
+        iou (Tensor[N, M]): the NxM matrix containing the pairwise
+            IoU values for every element in boxes1 and boxes2
+    """
+    return torch.ops.detectron2.box_iou_rotated(boxes1, boxes2)

detectron2/layers/shape_spec.py

+# -*- coding: utf-8 -*-
+# Copyright (c) Facebook, Inc. and its affiliates.
+from dataclasses import dataclass
+from typing import Optional
+
+
+@dataclass
+class ShapeSpec:
+    """
+    A simple structure that contains basic shape specification about a tensor.
+    It is often used as the auxiliary inputs/outputs of models,
+    to complement the lack of shape inference ability among pytorch modules.
+    """
+
+    channels: Optional[int] = None
+    height: Optional[int] = None
+    width: Optional[int] = None
+    stride: Optional[int] = None

detectron2/layers/wrappers.py

+# Copyright (c) Facebook, Inc. and its affiliates.
+"""
+Wrappers around on some nn functions, mainly to support empty tensors.
+
+Ideally, add support directly in PyTorch to empty tensors in those functions.
+
+These can be removed once https://github.com/pytorch/pytorch/issues/12013
+is implemented
+"""
+
+import functools
+import warnings
+from typing import List, Optional
+import torch
+from torch.nn import functional as F
+
+from detectron2.utils.env import TORCH_VERSION
+
+
+def shapes_to_tensor(x: List[int], device: Optional[torch.device] = None) -> torch.Tensor:
+    """
+    Turn a list of integer scalars or integer Tensor scalars into a vector,
+    in a way that's both traceable and scriptable.
+
+    In tracing, `x` should be a list of scalar Tensor, so the output can trace to the inputs.
+    In scripting or eager, `x` should be a list of int.
+    """
+    if torch.jit.is_scripting():
+        return torch.as_tensor(x, device=device)
+    if torch.jit.is_tracing():
+        assert all(
+            [isinstance(t, torch.Tensor) for t in x]
+        ), "Shape should be tensor during tracing!"
+        # as_tensor should not be used in tracing because it records a constant
+        ret = torch.stack(x)
+        if ret.device != device:  # avoid recording a hard-coded device if not necessary
+            ret = ret.to(device=device)
+        return ret
+    return torch.as_tensor(x, device=device)
+
+
+def check_if_dynamo_compiling():
+    if TORCH_VERSION >= (2, 1):
+        from torch._dynamo import is_compiling
+
+        return is_compiling()
+    else:
+        return False
+
+
+def disable_torch_compiler(func):
+    if TORCH_VERSION >= (2, 1):
+        # Use the torch.compiler.disable decorator if supported
+        @torch.compiler.disable
+        @functools.wraps(func)
+        def wrapper(*args, **kwargs):
+            return func(*args, **kwargs)
+
+        return wrapper
+    else:
+        # Return the function unchanged if torch.compiler.disable is not supported
+        return func
+
+
+def cat(tensors: List[torch.Tensor], dim: int = 0):
+    """
+    Efficient version of torch.cat that avoids a copy if there is only a single element in a list
+    """
+    assert isinstance(tensors, (list, tuple))
+    if len(tensors) == 1:
+        return tensors[0]
+    return torch.cat(tensors, dim)
+
+
+def empty_input_loss_func_wrapper(loss_func):
+    def wrapped_loss_func(input, target, *, reduction="mean", **kwargs):
+        """
+        Same as `loss_func`, but returns 0 (instead of nan) for empty inputs.
+        """
+        if target.numel() == 0 and reduction == "mean":
+            return input.sum() * 0.0  # connect the gradient
+        return loss_func(input, target, reduction=reduction, **kwargs)
+
+    return wrapped_loss_func
+
+
+cross_entropy = empty_input_loss_func_wrapper(F.cross_entropy)
+
+
+class _NewEmptyTensorOp(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, x, new_shape):
+        ctx.shape = x.shape
+        return x.new_empty(new_shape)
+
+    @staticmethod
+    def backward(ctx, grad):
+        shape = ctx.shape
+        return _NewEmptyTensorOp.apply(grad, shape), None
+
+
+class Conv2d(torch.nn.Conv2d):
+    """
+    A wrapper around :class:`torch.nn.Conv2d` to support empty inputs and more features.
+    """
+
+    def __init__(self, *args, **kwargs):
+        """
+        Extra keyword arguments supported in addition to those in `torch.nn.Conv2d`:
+
+        Args:
+            norm (nn.Module, optional): a normalization layer
+            activation (callable(Tensor) -> Tensor): a callable activation function
+
+        It assumes that norm layer is used before activation.
+        """
+        norm = kwargs.pop("norm", None)
+        activation = kwargs.pop("activation", None)
+        super().__init__(*args, **kwargs)
+
+        self.norm = norm
+        self.activation = activation
+
+    def forward(self, x):
+        # torchscript does not support SyncBatchNorm yet
+        # https://github.com/pytorch/pytorch/issues/40507
+        # and we skip these codes in torchscript since:
+        # 1. currently we only support torchscript in evaluation mode
+        # 2. features needed by exporting module to torchscript are added in PyTorch 1.6 or
+        # later version, `Conv2d` in these PyTorch versions has already supported empty inputs.
+        if not torch.jit.is_scripting():
+            # Dynamo doesn't support context managers yet
+            is_dynamo_compiling = check_if_dynamo_compiling()
+            if not is_dynamo_compiling:
+                with warnings.catch_warnings(record=True):
+                    if x.numel() == 0 and self.training:
+                        # https://github.com/pytorch/pytorch/issues/12013
+                        assert not isinstance(
+                            self.norm, torch.nn.SyncBatchNorm
+                        ), "SyncBatchNorm does not support empty inputs!"
+
+        x = F.conv2d(
+            x, self.weight, self.bias, self.stride, self.padding, self.dilation, self.groups
+        )
+        if self.norm is not None:
+            x = self.norm(x)
+        if self.activation is not None:
+            x = self.activation(x)
+        return x
+
+
+ConvTranspose2d = torch.nn.ConvTranspose2d
+BatchNorm2d = torch.nn.BatchNorm2d
+interpolate = F.interpolate
+Linear = torch.nn.Linear
+
+
+def nonzero_tuple(x):
+    """
+    A 'as_tuple=True' version of torch.nonzero to support torchscript.
+    because of https://github.com/pytorch/pytorch/issues/38718
+    """
+    if torch.jit.is_scripting():
+        if x.dim() == 0:
+            return x.unsqueeze(0).nonzero().unbind(1)
+        return x.nonzero().unbind(1)
+    else:
+        return x.nonzero(as_tuple=True)
+
+
+@torch.jit.script_if_tracing
+def move_device_like(src: torch.Tensor, dst: torch.Tensor) -> torch.Tensor:
+    """
+    Tracing friendly way to cast tensor to another tensor's device. Device will be treated
+    as constant during tracing, scripting the casting process as whole can workaround this issue.
+    """
+    return src.to(dst.device)

detectron2/model_zoo/__init__.py

+# Copyright (c) Facebook, Inc. and its affiliates.
+"""
+Model Zoo API for Detectron2: a collection of functions to create common model architectures
+listed in `MODEL_ZOO.md <https://github.com/facebookresearch/detectron2/blob/main/MODEL_ZOO.md>`_,
+and optionally load their pre-trained weights.
+"""
+
+from .model_zoo import get, get_config_file, get_checkpoint_url, get_config
+
+__all__ = ["get_checkpoint_url", "get", "get_config_file", "get_config"]

detectron2/model_zoo/configs/Base-RCNN-C4.yaml

+MODEL:
+  META_ARCHITECTURE: "GeneralizedRCNN"
+  RPN:
+    PRE_NMS_TOPK_TEST: 6000
+    POST_NMS_TOPK_TEST: 1000
+  ROI_HEADS:
+    NAME: "Res5ROIHeads"
+DATASETS:
+  TRAIN: ("coco_2017_train",)
+  TEST: ("coco_2017_val",)
+SOLVER:
+  IMS_PER_BATCH: 16
+  BASE_LR: 0.02
+  STEPS: (60000, 80000)
+  MAX_ITER: 90000
+INPUT:
+  MIN_SIZE_TRAIN: (640, 672, 704, 736, 768, 800)
+VERSION: 2

detectron2/model_zoo/configs/Base-RCNN-DilatedC5.yaml

+MODEL:
+  META_ARCHITECTURE: "GeneralizedRCNN"
+  RESNETS:
+    OUT_FEATURES: ["res5"]
+    RES5_DILATION: 2
+  RPN:
+    IN_FEATURES: ["res5"]
+    PRE_NMS_TOPK_TEST: 6000
+    POST_NMS_TOPK_TEST: 1000
+  ROI_HEADS:
+    NAME: "StandardROIHeads"
+    IN_FEATURES: ["res5"]
+  ROI_BOX_HEAD:
+    NAME: "FastRCNNConvFCHead"
+    NUM_FC: 2
+    POOLER_RESOLUTION: 7
+  ROI_MASK_HEAD:
+    NAME: "MaskRCNNConvUpsampleHead"
+    NUM_CONV: 4
+    POOLER_RESOLUTION: 14
+DATASETS:
+  TRAIN: ("coco_2017_train",)
+  TEST: ("coco_2017_val",)
+SOLVER:
+  IMS_PER_BATCH: 16
+  BASE_LR: 0.02
+  STEPS: (60000, 80000)
+  MAX_ITER: 90000
+INPUT:
+  MIN_SIZE_TRAIN: (640, 672, 704, 736, 768, 800)
+VERSION: 2

detectron2/model_zoo/configs/Base-RCNN-FPN.yaml

+MODEL:
+  META_ARCHITECTURE: "GeneralizedRCNN"
+  BACKBONE:
+    NAME: "build_resnet_fpn_backbone"
+  RESNETS:
+    OUT_FEATURES: ["res2", "res3", "res4", "res5"]
+  FPN:
+    IN_FEATURES: ["res2", "res3", "res4", "res5"]
+  ANCHOR_GENERATOR:
+    SIZES: [[32], [64], [128], [256], [512]]  # One size for each in feature map
+    ASPECT_RATIOS: [[0.5, 1.0, 2.0]]  # Three aspect ratios (same for all in feature maps)
+  RPN:
+    IN_FEATURES: ["p2", "p3", "p4", "p5", "p6"]
+    PRE_NMS_TOPK_TRAIN: 2000  # Per FPN level
+    PRE_NMS_TOPK_TEST: 1000  # Per FPN level
+    # Detectron1 uses 2000 proposals per-batch,
+    # (See "modeling/rpn/rpn_outputs.py" for details of this legacy issue)
+    # which is approximately 1000 proposals per-image since the default batch size for FPN is 2.
+    POST_NMS_TOPK_TRAIN: 1000
+    POST_NMS_TOPK_TEST: 1000
+  ROI_HEADS:
+    NAME: "StandardROIHeads"
+    IN_FEATURES: ["p2", "p3", "p4", "p5"]
+  ROI_BOX_HEAD:
+    NAME: "FastRCNNConvFCHead"
+    NUM_FC: 2
+    POOLER_RESOLUTION: 7
+  ROI_MASK_HEAD:
+    NAME: "MaskRCNNConvUpsampleHead"
+    NUM_CONV: 4
+    POOLER_RESOLUTION: 14
+DATASETS:
+  TRAIN: ("coco_2017_train",)
+  TEST: ("coco_2017_val",)
+SOLVER:
+  IMS_PER_BATCH: 16
+  BASE_LR: 0.02
+  STEPS: (60000, 80000)
+  MAX_ITER: 90000
+INPUT:
+  MIN_SIZE_TRAIN: (640, 672, 704, 736, 768, 800)
+VERSION: 2

detectron2/model_zoo/configs/Base-RetinaNet.yaml

+MODEL:
+  META_ARCHITECTURE: "RetinaNet"
+  BACKBONE:
+    NAME: "build_retinanet_resnet_fpn_backbone"
+  RESNETS:
+    OUT_FEATURES: ["res3", "res4", "res5"]
+  ANCHOR_GENERATOR:
+    SIZES: !!python/object/apply:eval ["[[x, x * 2**(1.0/3), x * 2**(2.0/3) ] for x in [32, 64, 128, 256, 512 ]]"]
+  FPN:
+    IN_FEATURES: ["res3", "res4", "res5"]
+  RETINANET:
+    IOU_THRESHOLDS: [0.4, 0.5]
+    IOU_LABELS: [0, -1, 1]
+    SMOOTH_L1_LOSS_BETA: 0.0
+DATASETS:
+  TRAIN: ("coco_2017_train",)
+  TEST: ("coco_2017_val",)
+SOLVER:
+  IMS_PER_BATCH: 16
+  BASE_LR: 0.01  # Note that RetinaNet uses a different default learning rate
+  STEPS: (60000, 80000)
+  MAX_ITER: 90000
+INPUT:
+  MIN_SIZE_TRAIN: (640, 672, 704, 736, 768, 800)
+VERSION: 2

detectron2/model_zoo/configs/COCO-Detection/fast_rcnn_R_50_FPN_1x.yaml

+_BASE_: "../Base-RCNN-FPN.yaml"
+MODEL:
+  WEIGHTS: "detectron2://ImageNetPretrained/MSRA/R-50.pkl"
+  MASK_ON: False
+  LOAD_PROPOSALS: True
+  RESNETS:
+    DEPTH: 50
+  PROPOSAL_GENERATOR:
+    NAME: "PrecomputedProposals"
+DATASETS:
+  TRAIN: ("coco_2017_train",)
+  PROPOSAL_FILES_TRAIN: ("detectron2://COCO-Detection/rpn_R_50_FPN_1x/137258492/coco_2017_train_box_proposals_21bc3a.pkl", )
+  TEST: ("coco_2017_val",)
+  PROPOSAL_FILES_TEST: ("detectron2://COCO-Detection/rpn_R_50_FPN_1x/137258492/coco_2017_val_box_proposals_ee0dad.pkl", )
+DATALOADER:
+  # proposals are part of the dataset_dicts, and take a lot of RAM
+  NUM_WORKERS: 2

detectron2/model_zoo/configs/COCO-Detection/faster_rcnn_R_101_C4_3x.yaml

+_BASE_: "../Base-RCNN-C4.yaml"
+MODEL:
+  WEIGHTS: "detectron2://ImageNetPretrained/MSRA/R-101.pkl"
+  MASK_ON: False
+  RESNETS:
+    DEPTH: 101
+SOLVER:
+  STEPS: (210000, 250000)
+  MAX_ITER: 270000