Merge pull request #257 from open-mmlab/pytorch-1.0

Support Pytorch 1.0

INSTALL.md

@@ -4,13 +4,13 @@
 
 - Linux (tested on Ubuntu 16.04 and CentOS 7.2)
 - Python 3.4+
-- PyTorch 0.4.1
+- PyTorch 1.0
 - Cython
-- [mmcv](https://github.com/open-mmlab/mmcv)
+- [mmcv](https://github.com/open-mmlab/mmcv) >= 0.2.2
 
 ### Install mmdetection
 
-a. Install PyTorch 0.4.1 and torchvision following the [official instructions](https://pytorch.org/).
+a. Install PyTorch 1.0 and torchvision following the [official instructions](https://pytorch.org/).
 
 b. Clone the mmdetection repository.
 

MODEL_ZOO.md

@@ -10,11 +10,17 @@
 ### Software environment
 
 - Python 3.6 / 3.7
-- PyTorch 0.4.1
+- PyTorch 1.0
 - CUDA 9.0.176
 - CUDNN 7.0.4
 - NCCL 2.1.15
 
+Note: The train time was measured with PyTorch 0.4.1. We will update it later, which should be about 0.02s ~ 0.05s faster.
+
+## Mirror sites
+
+We use AWS as the main site to host our model zoo, and maintain a mirror on aliyun.
+You can replace `https://s3.ap-northeast-2.amazonaws.com` with `https://open-mmlab.oss-cn-beijing.aliyuncs.com` in model urls.
 
 ## Common settings
 

README.md

@@ -3,6 +3,9 @@
 
 ## Introduction
 
+The master branch works with **PyTorch 1.0**. If you would like to use PyTorch 0.4.1,
+please checkout to the [pytorch-0.4.1](https://github.com/open-mmlab/mmdetection/tree/pytorch-0.4.1) branch.
+
 mmdetection is an open source object detection toolbox based on PyTorch. It is
 a part of the open-mmlab project developed by [Multimedia Laboratory, CUHK](http://mmlab.ie.cuhk.edu.hk/).
 
@@ -36,6 +39,9 @@ This project is released under the [Apache 2.0 license](LICENSE).
 
 ## Updates
 
+v0.6rc0(06/02/2019)
+- Migrate to PyTorch 1.0.
+
 v0.5.7 (06/02/2019)
 - Add support for Deformable ConvNet v2. (Many thanks to the authors and [@chengdazhi](https://github.com/chengdazhi))
 - This is the last release based on PyTorch 0.4.1.

mmdet/core/loss/losses.py

@@ -34,13 +34,21 @@ def sigmoid_focal_loss(pred,
                        weight,
                        gamma=2.0,
                        alpha=0.25,
-                       reduction='elementwise_mean'):
+                       reduction='mean'):
     pred_sigmoid = pred.sigmoid()
     pt = (1 - pred_sigmoid) * target + pred_sigmoid * (1 - target)
     weight = (alpha * target + (1 - alpha) * (1 - target)) * weight
     weight = weight * pt.pow(gamma)
-    return F.binary_cross_entropy_with_logits(
-        pred, target, weight, reduction=reduction)
+    loss = F.binary_cross_entropy_with_logits(
+        pred, target, reduction='none') * weight
+    reduction_enum = F._Reduction.get_enum(reduction)
+    # none: 0, mean:1, sum: 2
+    if reduction_enum == 0:
+        return loss
+    elif reduction_enum == 1:
+        return loss.mean()
+    elif reduction_enum == 2:
+        return loss.sum()
 
 
 def weighted_sigmoid_focal_loss(pred,
@@ -62,22 +70,22 @@ def mask_cross_entropy(pred, target, label):
     inds = torch.arange(0, num_rois, dtype=torch.long, device=pred.device)
     pred_slice = pred[inds, label].squeeze(1)
     return F.binary_cross_entropy_with_logits(
-        pred_slice, target, reduction='elementwise_mean')[None]
+        pred_slice, target, reduction='mean')[None]
 
 
-def smooth_l1_loss(pred, target, beta=1.0, reduction='elementwise_mean'):
+def smooth_l1_loss(pred, target, beta=1.0, reduction='mean'):
     assert beta > 0
     assert pred.size() == target.size() and target.numel() > 0
     diff = torch.abs(pred - target)
     loss = torch.where(diff < beta, 0.5 * diff * diff / beta,
                        diff - 0.5 * beta)
-    reduction = F._Reduction.get_enum(reduction)
-    # none: 0, elementwise_mean:1, sum: 2
-    if reduction == 0:
+    reduction_enum = F._Reduction.get_enum(reduction)
+    # none: 0, mean:1, sum: 2
+    if reduction_enum == 0:
         return loss
-    elif reduction == 1:
+    elif reduction_enum == 1:
         return loss.sum() / pred.numel()
-    elif reduction == 2:
+    elif reduction_enum == 2:
         return loss.sum()
 
 

mmdet/ops/dcn/src/deform_pool_cuda.cpp

+// modify from
+// https://github.com/chengdazhi/Deformable-Convolution-V2-PyTorch/blob/mmdetection/mmdet/ops/dcn/src/modulated_dcn_cuda.c
+
+// based on
 // author: Charles Shang
 // https://github.com/torch/cunn/blob/master/lib/THCUNN/generic/SpatialConvolutionMM.cu
 
-// modify from https://github.com/chengdazhi/Deformable-Convolution-V2-PyTorch/blob /mmdetection/mmdet/ops/dcn/src/modulated_dcn_cuda.c
-
-#include <torch/torch.h>
+#include <torch/extension.h>
 
 #include <cmath>
 #include <vector>
 
-void DeformablePSROIPoolForward(const at::Tensor data,
-                                const at::Tensor bbox,
-                                const at::Tensor trans,
-                                at::Tensor out,
-                                at::Tensor top_count,
-                                const int batch,
-                                const int channels,
-                                const int height,
-                                const int width,
-                                const int num_bbox,
-                                const int channels_trans,
-                                const int no_trans,
-                                const float spatial_scale,
-                                const int output_dim,
-                                const int group_size,
-                                const int pooled_size,
-                                const int part_size,
-                                const int sample_per_part,
-                                const float trans_std);
+void DeformablePSROIPoolForward(
+    const at::Tensor data, const at::Tensor bbox, const at::Tensor trans,
+    at::Tensor out, at::Tensor top_count, const int batch, const int channels,
+    const int height, const int width, const int num_bbox,
+    const int channels_trans, const int no_trans, const float spatial_scale,
+    const int output_dim, const int group_size, const int pooled_size,
+    const int part_size, const int sample_per_part, const float trans_std);
 
-void DeformablePSROIPoolBackwardAcc(const at::Tensor out_grad,
-                                    const at::Tensor data,
-                                    const at::Tensor bbox,
-                                    const at::Tensor trans,
-                                    const at::Tensor top_count,
-                                    at::Tensor in_grad,
-                                    at::Tensor trans_grad,
-                                    const int batch,
-                                    const int channels,
-                                    const int height,
-                                    const int width,
-                                    const int num_bbox,
-                                    const int channels_trans,
-                                    const int no_trans,
-                                    const float spatial_scale,
-                                    const int output_dim,
-                                    const int group_size,
-                                    const int pooled_size,
-                                    const int part_size,
-                                    const int sample_per_part,
-                                    const float trans_std);
+void DeformablePSROIPoolBackwardAcc(
+    const at::Tensor out_grad, const at::Tensor data, const at::Tensor bbox,
+    const at::Tensor trans, const at::Tensor top_count, at::Tensor in_grad,
+    at::Tensor trans_grad, const int batch, const int channels,
+    const int height, const int width, const int num_bbox,
+    const int channels_trans, const int no_trans, const float spatial_scale,
+    const int output_dim, const int group_size, const int pooled_size,
+    const int part_size, const int sample_per_part, const float trans_std);
 
-void deform_psroi_pooling_cuda_forward(at::Tensor input, at::Tensor bbox,
-                                       at::Tensor trans,
-                                       at::Tensor out, at::Tensor top_count,
-                                       const int no_trans,
-                                       const float spatial_scale,
-                                       const int output_dim,
-                                       const int group_size,
-                                       const int pooled_size,
-                                       const int part_size,
-                                       const int sample_per_part,
-                                       const float trans_std)
-{
-    AT_CHECK(input.is_contiguous(), "input tensor has to be contiguous");
+void deform_psroi_pooling_cuda_forward(
+    at::Tensor input, at::Tensor bbox, at::Tensor trans, at::Tensor out,
+    at::Tensor top_count, const int no_trans, const float spatial_scale,
+    const int output_dim, const int group_size, const int pooled_size,
+    const int part_size, const int sample_per_part, const float trans_std) {
+  AT_CHECK(input.is_contiguous(), "input tensor has to be contiguous");
 
-    const int batch = input.size(0);
-    const int channels = input.size(1);
-    const int height = input.size(2);
-    const int width = input.size(3);
-    const int channels_trans = no_trans ? 2 : trans.size(1);
+  const int batch = input.size(0);
+  const int channels = input.size(1);
+  const int height = input.size(2);
+  const int width = input.size(3);
+  const int channels_trans = no_trans ? 2 : trans.size(1);
 
-    const int num_bbox = bbox.size(0);
-    if (num_bbox != out.size(0))
-        AT_ERROR("Output shape and bbox number wont match: (%d vs %d).",
-                 out.size(0), num_bbox);
+  const int num_bbox = bbox.size(0);
+  if (num_bbox != out.size(0))
+    AT_ERROR("Output shape and bbox number wont match: (%d vs %d).",
+             out.size(0), num_bbox);
 
-    DeformablePSROIPoolForward(input, bbox, trans, out, top_count,
-                               batch, channels, height, width,
-                               num_bbox,
-                               channels_trans,
-                               no_trans,
-                               spatial_scale,
-                               output_dim,
-                               group_size,
-                               pooled_size,
-                               part_size,
-                               sample_per_part,
-                               trans_std);
+  DeformablePSROIPoolForward(
+      input, bbox, trans, out, top_count, batch, channels, height, width,
+      num_bbox, channels_trans, no_trans, spatial_scale, output_dim, group_size,
+      pooled_size, part_size, sample_per_part, trans_std);
 }
 
-void deform_psroi_pooling_cuda_backward(at::Tensor out_grad,
-                                        at::Tensor input, at::Tensor bbox,
-                                        at::Tensor trans, at::Tensor top_count,
-                                        at::Tensor input_grad, at::Tensor trans_grad,
-                                        const int no_trans,
-                                        const float spatial_scale,
-                                        const int output_dim,
-                                        const int group_size,
-                                        const int pooled_size,
-                                        const int part_size,
-                                        const int sample_per_part,
-                                        const float trans_std)
-{
-    AT_CHECK(out_grad.is_contiguous(), "out_grad tensor has to be contiguous");
-    AT_CHECK(input.is_contiguous(), "input tensor has to be contiguous");
+void deform_psroi_pooling_cuda_backward(
+    at::Tensor out_grad, at::Tensor input, at::Tensor bbox, at::Tensor trans,
+    at::Tensor top_count, at::Tensor input_grad, at::Tensor trans_grad,
+    const int no_trans, const float spatial_scale, const int output_dim,
+    const int group_size, const int pooled_size, const int part_size,
+    const int sample_per_part, const float trans_std) {
+  AT_CHECK(out_grad.is_contiguous(), "out_grad tensor has to be contiguous");
+  AT_CHECK(input.is_contiguous(), "input tensor has to be contiguous");
 
-    const int batch = input.size(0);
-    const int channels = input.size(1);
-    const int height = input.size(2);
-    const int width = input.size(3);
-    const int channels_trans = no_trans ? 2 : trans.size(1);
+  const int batch = input.size(0);
+  const int channels = input.size(1);
+  const int height = input.size(2);
+  const int width = input.size(3);
+  const int channels_trans = no_trans ? 2 : trans.size(1);
 
-    const int num_bbox = bbox.size(0);
-    if (num_bbox != out_grad.size(0))
-        AT_ERROR("Output shape and bbox number wont match: (%d vs %d).",
-                 out_grad.size(0), num_bbox);
+  const int num_bbox = bbox.size(0);
+  if (num_bbox != out_grad.size(0))
+    AT_ERROR("Output shape and bbox number wont match: (%d vs %d).",
+             out_grad.size(0), num_bbox);
 
-    DeformablePSROIPoolBackwardAcc(out_grad,
-                                   input,
-                                   bbox,
-                                   trans,
-                                   top_count,
-                                   input_grad,
-                                   trans_grad,
-                                   batch, channels, height, width, num_bbox,
-                                   channels_trans,
-                                   no_trans,
-                                   spatial_scale,
-                                   output_dim,
-                                   group_size,
-                                   pooled_size,
-                                   part_size,
-                                   sample_per_part,
-                                   trans_std);
+  DeformablePSROIPoolBackwardAcc(
+      out_grad, input, bbox, trans, top_count, input_grad, trans_grad, batch,
+      channels, height, width, num_bbox, channels_trans, no_trans,
+      spatial_scale, output_dim, group_size, pooled_size, part_size,
+      sample_per_part, trans_std);
 }
 
-PYBIND11_MODULE(TORCH_EXTENSION_NAME, m)
-{
-    m.def("deform_psroi_pooling_cuda_forward", &deform_psroi_pooling_cuda_forward,
-          "deform psroi pooling forward(CUDA)");
-    m.def("deform_psroi_pooling_cuda_backward", &deform_psroi_pooling_cuda_backward,
-          "deform psroi pooling backward(CUDA)");
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
+  m.def("deform_psroi_pooling_cuda_forward", &deform_psroi_pooling_cuda_forward,
+        "deform psroi pooling forward(CUDA)");
+  m.def("deform_psroi_pooling_cuda_backward",
+        &deform_psroi_pooling_cuda_backward,
+        "deform psroi pooling backward(CUDA)");
 }
\ No newline at end of file

mmdet/ops/nms/cpu_nms.pyx

@@ -5,6 +5,8 @@
 # Written by Ross Girshick
 # --------------------------------------------------------
 
+# cython: language_level=3, boundscheck=False
+
 import numpy as np
 cimport numpy as np
 

mmdet/ops/nms/cpu_soft_nms.pyx

@@ -6,6 +6,8 @@
 # Modified by Kai Chen
 # ----------------------------------------------------------
 
+# cython: language_level=3, boundscheck=False
+
 import numpy as np
 cimport numpy as np
 

mmdet/ops/nms/gpu_nms.pyx

@@ -5,6 +5,8 @@
 # Written by Ross Girshick
 # --------------------------------------------------------
 
+# cython: language_level=3, boundscheck=False
+
 import numpy as np
 cimport numpy as np
 

mmdet/ops/roi_align/functions/roi_align.py

-from torch.autograd import Function, Variable
+from torch.autograd import Function
 
 from .. import roi_align_cuda
 
@@ -49,11 +49,11 @@ class RoIAlignFunction(Function):
 
         grad_input = grad_rois = None
         if ctx.needs_input_grad[0]:
-            grad_input = Variable(
-                rois.new(batch_size, num_channels, data_height, data_width)
-                .zero_())
-            roi_align_cuda.backward(grad_output, rois, out_h, out_w,
-                                    spatial_scale, sample_num, grad_input)
+            grad_input = rois.new_zeros(batch_size, num_channels, data_height,
+                                        data_width)
+            roi_align_cuda.backward(grad_output.contiguous(), rois, out_h,
+                                    out_w, spatial_scale, sample_num,
+                                    grad_input)
 
         return grad_input, grad_rois, None, None, None
 

mmdet/ops/roi_align/src/roi_align_cuda.cpp

-#include <torch/torch.h>
+#include <torch/extension.h>
 
 #include <cmath>
 #include <vector>

mmdet/ops/roi_align/src/roi_align_kernel.cu

 #include <ATen/ATen.h>
 #include <THC/THCAtomics.cuh>
 
-using namespace at;  // temporal fix for pytorch<=0.4.1 (see #9848)
-
 #define CUDA_1D_KERNEL_LOOP(i, n)                            \
   for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < n; \
        i += blockDim.x * gridDim.x)
@@ -144,12 +142,7 @@ int ROIAlignForwardLaucher(const at::Tensor features, const at::Tensor rois,
                 sample_num, channels, height, width, pooled_height,
                 pooled_width, top_data);
       }));
-  cudaError_t err = cudaGetLastError();
-  if (cudaSuccess != err) {
-    fprintf(stderr, "cudaCheckError() failed : %s\n", cudaGetErrorString(err));
-    exit(-1);
-  }
-
+  THCudaCheck(cudaGetLastError());
   return 1;
 }
 
@@ -280,8 +273,7 @@ int ROIAlignBackwardLaucher(const at::Tensor top_grad, const at::Tensor rois,
                             at::Tensor bottom_grad) {
   const int output_size = num_rois * pooled_height * pooled_width * channels;
 
-  // TODO: use AT_DISPATCH_FLOATING_TYPES_AND_HALF when atomicAdd is resolved
-  AT_DISPATCH_FLOATING_TYPES(
+  AT_DISPATCH_FLOATING_TYPES_AND_HALF(
       top_grad.type(), "ROIAlignLaucherBackward", ([&] {
         const scalar_t *top_diff = top_grad.data<scalar_t>();
         const scalar_t *rois_data = rois.data<scalar_t>();
@@ -297,11 +289,6 @@ int ROIAlignBackwardLaucher(const at::Tensor top_grad, const at::Tensor rois,
                 channels, height, width, pooled_height, pooled_width,
                 bottom_diff);
       }));
-  cudaError_t err = cudaGetLastError();
-  if (cudaSuccess != err) {
-    fprintf(stderr, "cudaCheckError() failed : %s\n", cudaGetErrorString(err));
-    exit(-1);
-  }
-
+  THCudaCheck(cudaGetLastError());
   return 1;
 }

mmdet/ops/roi_pool/functions/roi_pool.py

@@ -24,9 +24,8 @@ class RoIPoolFunction(Function):
         num_channels = features.size(1)
         num_rois = rois.size(0)
         out_size = (num_rois, num_channels, out_h, out_w)
-        output = features.new_zeros(*out_size)
-
-        argmax = features.new_zeros(*out_size, dtype=torch.int)
+        output = features.new_zeros(out_size)
+        argmax = features.new_zeros(out_size, dtype=torch.int)
         roi_pool_cuda.forward(features, rois, out_h, out_w, spatial_scale,
                               output, argmax)
         ctx.spatial_scale = spatial_scale
@@ -46,9 +45,9 @@ class RoIPoolFunction(Function):
 
         grad_input = grad_rois = None
         if ctx.needs_input_grad[0]:
-            grad_input = grad_output.new(feature_size).zero_()
-            roi_pool_cuda.backward(grad_output, rois, argmax, spatial_scale,
-                                   grad_input)
+            grad_input = grad_output.new_zeros(feature_size)
+            roi_pool_cuda.backward(grad_output.contiguous(), rois, argmax,
+                                   spatial_scale, grad_input)
 
         return grad_input, grad_rois, None, None
 

mmdet/ops/roi_pool/src/roi_pool_cuda.cpp

-#include <torch/torch.h>
+#include <torch/extension.h>
 
 #include <cmath>
 #include <vector>

mmdet/ops/roi_pool/src/roi_pool_kernel.cu

 #include <ATen/ATen.h>
 #include <THC/THCAtomics.cuh>
 
-using namespace at;  // temporal fix for pytorch<=0.4.1 (see #9848)
-
 #define CUDA_1D_KERNEL_LOOP(i, n)                            \
   for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < n; \
        i += blockDim.x * gridDim.x)
@@ -100,11 +98,7 @@ int ROIPoolForwardLaucher(const at::Tensor features, const at::Tensor rois,
                 channels, height, width, pooled_h, pooled_w, top_data,
                 argmax_data);
       }));
-  cudaError_t err = cudaGetLastError();
-  if (cudaSuccess != err) {
-    fprintf(stderr, "cudaCheckError() failed : %s\n", cudaGetErrorString(err));
-    exit(-1);
-  }
+  THCudaCheck(cudaGetLastError());
   return 1;
 }
 
@@ -139,8 +133,7 @@ int ROIPoolBackwardLaucher(const at::Tensor top_grad, const at::Tensor rois,
                            const int pooled_w, at::Tensor bottom_grad) {
   const int output_size = num_rois * pooled_h * pooled_w * channels;
 
-  // TODO: use AT_DISPATCH_FLOATING_TYPES_AND_HALF when atomicAdd is resolved
-  AT_DISPATCH_FLOATING_TYPES(
+  AT_DISPATCH_FLOATING_TYPES_AND_HALF(
       top_grad.type(), "ROIPoolLaucherBackward", ([&] {
         const scalar_t *top_diff = top_grad.data<scalar_t>();
         const scalar_t *rois_data = rois.data<scalar_t>();
@@ -158,11 +151,6 @@ int ROIPoolBackwardLaucher(const at::Tensor top_grad, const at::Tensor rois,
                 scalar_t(spatial_scale), channels, height, width, pooled_h,
                 pooled_w, bottom_diff);
       }));
-  cudaError_t err = cudaGetLastError();
-  if (cudaSuccess != err) {
-    fprintf(stderr, "cudaCheckError() failed : %s\n", cudaGetErrorString(err));
-    exit(-1);
-  }
-
+  THCudaCheck(cudaGetLastError());
   return 1;
 }

setup.py

@@ -11,8 +11,8 @@ def readme():
 
 
 MAJOR = 0
-MINOR = 5
-PATCH = 7
+MINOR = 6
+PATCH = 'rc0'
 SUFFIX = ''
 SHORT_VERSION = '{}.{}.{}{}'.format(MAJOR, MINOR, PATCH, SUFFIX)