Build

README.md

@@ -34,6 +34,17 @@ Currently it only works using the default GPU (index 0)
     ./test.sh
     ```
 
+## Adjust
++ Build and test
+    ```
+    export CUDA_HOME=/usr/local/cuda-11.3
+    python setup.py build
+    ```
+    Move the compiled dynamic link library to the *roi_align* file directory, Then start testing
+    ```
+    ./test.sh
+    ``` 
+
 + Use RoIAlign or crop_and_resize 
 
     Since PyTorch 1.2.0 [Legacy autograd function with non-static forward method is deprecated.](https://github.com/pytorch/pytorch/blob/fdfc676eb6c4d9f50496e564976fbe6d124e23a5/torch/csrc/autograd/python_function.cpp#L636-L638)

tests/crop_and_resize_example.py

@@ -1,3 +1,5 @@
+import sys, os
+sys.path.append(os.getcwd())
 import torch
 from torch import nn
 from torchvision import transforms, utils

tests/test.py

@@ -1,3 +1,5 @@
+import sys, os
+sys.path.append(os.getcwd())
 import numpy as np
 import torch
 import sys

tests/test2.py

@@ -1,3 +1,5 @@
+import sys, os
+sys.path.append(os.getcwd())
 import numpy as np
 import torch
 from torch.autograd import Variable
@@ -25,4 +27,6 @@ def to_varabile(arr, requires_grad=False, is_cuda=True):
 box_index = to_varabile(box_index_data, requires_grad=False, is_cuda=is_cuda)
 
 # set transform_fpcoor to False is the crop_and_resize
-print(RoIAlign.apply(image_torch, boxes, box_index, 3, 3, transform_fpcoor=True))
+roi_align =RoIAlign(3, 3, transform_fpcoor=True)
+print(roi_align(image_torch, boxes, box_index))
+# print(RoIAlign.apply(image_torch, boxes, box_index, 3, 3, transform_fpcoor=True))

tests/test3.py

@@ -0,0 +1,27 @@
+import torch
+import sys,os
+sys.path.append(os.getcwd())
+from roi_align import RoIAlign      # RoIAlign module
+from roi_align import CropAndResize # crop_and_resize module
+
+# input feature maps (suppose that we have batch_size==2)
+image = torch.arange(0., 49).view(1, 1, 7, 7).repeat(2, 1, 1, 1)
+image[0] += 10
+print('image: ', image)
+
+
+# for example, we have two bboxes with coords xyxy (first with batch_id=0, second with batch_id=1).
+boxes = torch.Tensor([[1, 0, 5, 4],
+                     [0.5, 3.5, 4, 7]])
+
+box_index = torch.tensor([0, 1], dtype=torch.int) # index of bbox in batch
+
+# RoIAlign layer with crop sizes:
+crop_height = 4
+crop_width = 4
+roi_align = RoIAlign(crop_height, crop_width)
+
+# make crops:
+crops = roi_align(image, boxes, box_index)
+
+print('crops:', crops)