Implement on CPU

.gitignore

@@ -0,0 +1,56 @@
+# output dir
+output
+instant_test_output
+inference_test_output
+
+
+*.jpg
+*.png
+*.txt
+
+# compilation and distribution
+__pycache__
+_ext
+*.pyc
+*.so
+AdelaiDet.egg-info/
+build/
+dist/
+
+# pytorch/python/numpy formats
+*.pth
+*.pkl
+*.npy
+
+# ipython/jupyter notebooks
+*.ipynb
+**/.ipynb_checkpoints/
+
+# Editor temporaries
+*.swn
+*.swo
+*.swp
+*~
+
+# Pycharm editor settings
+.idea
+.vscode
+.python-version
+
+# project dirs
+/datasets/coco
+/datasets/lvis
+/datasets/pic
+/datasets/ytvos
+/models
+/demo_outputs
+/example_inputs
+/debug
+/weights
+/export
+eval.sh
+
+demo/performance.py
+demo/demo2.py
+train.sh
+benchmark.sh

adet/layers/csrc/DeformAttn/ms_deform_attn_cpu.cpp → adet/layers/csrc/DeformAttn/cpu/ms_deform_attn_cpu.cpp

@@ -11,7 +11,7 @@
 #include <vector>
 
 #include <ATen/ATen.h>
-#include <ATen/cuda/CUDAContext.h>
+// #include <ATen/cuda/CUDAContext.h>
 
 
 at::Tensor

adet/layers/csrc/DeformAttn/ms_deform_attn.h

@@ -10,7 +10,7 @@
 
 #pragma once
 
-#include "ms_deform_attn_cpu.h"
+#include "cpu/ms_deform_attn_cpu.h"
 
 #ifdef WITH_CUDA
 #include "ms_deform_attn_cuda.h"

adet/layers/ms_deform_attn.py

@@ -148,8 +148,18 @@ def forward(self, query, reference_points, input_flatten, input_spatial_shapes,
         else:
             raise ValueError(
                 'Last dim of reference_points must be 2 or 4, but get {} instead.'.format(reference_points.shape[-1]))
-        output = _MSDeformAttnFunction.apply(
-            value, input_spatial_shapes, input_level_start_index, sampling_locations, attention_weights, self.im2col_step)
+
+        try:
+            # CUDA implementation
+            output = _MSDeformAttnFunction.apply(
+                value, input_spatial_shapes, input_level_start_index, sampling_locations, attention_weights, self.im2col_step)
+        except Exception:
+            # PyTorch implementation
+            output = ms_deform_attn_core_pytorch(
+                value, 
+                input_spatial_shapes, 
+                sampling_locations, 
+                attention_weights)
         output = self.output_proj(output)
         # if decoder:
         #     return output, sampling_locations, attention_weights