英文:
Any way to optimize this NumPy indexed reassignation?
问题
我试图实现的基本上是HDR：从每个曝光级别中获取最强烈的像素，并替换主图像中相应的像素（如果它已经是最强烈的则保持不变）。
我主要是想改进for循环，因为它在处理大型数组时是主要的瓶颈。
hdr_data = np.array([main_scan.data, lower_scan.data, higher_scan.data])maximum_intensities = hdr_data.max(axis=-1)frame_indices = np.argmax(maximum_intensities, axis=0)for i, index in enumerate(frame_indices):main_scan.data[i] = hdr_data[index, i, :]
main_scan，lower_scan 和 higher_scan 的形状是 X*Y，而 hdr_data 的形状是 3*X*Y。
英文:
What I'm trying to achieve is basically HDR: grab the most intense pixel out of each exposure level and replace the corresponding pixel in the main image (or keep as-is if it's already the most intense).
I'm mainly looking to improve the for loop, as it's the main bottleneck when processing large arrays.
hdr_data = np.array([main_scan.data, lower_scan.data, higher_scan.data])maximum_intensities = hdr_data.max(axis=-1)frame_indices = np.argmax(maximum_intensities, axis=0)for i, index in enumerate(frame_indices):main_scan.data[i] = hdr_data[index, i, :]
main_scan, lower_scan and higher_scan are in X*Y shapes, and hdr_data is in 3*X*Y shape
答案1
得分: 1
hdr_data # (3,x,y) 形状maximum_intensities # (3,x) 形状frame_indices # (x,) 形状，值为 0,1,2iterative:for i, index in enumerate(frame_indices):main_scan.data[i] = hdr_data[index, i, :]try:max_data = hdr_data[np.arange(x)[:,None], frame_indices] # (x,y)`np.take_along_axis` 也可以做到这一点；参见其文档testIn [340]: arr = np.random.randint(0,10,(3,10,4))In [341]: idx=np.argmax(arr.max(axis=-1), axis=0)In [342]: idxOut[342]: array([0, 0, 0, 0, 1, 0, 0, 2, 0, 0], dtype=int64)In [343]: arr[np.arange(10),idx,:]---------------------------------------------------------------------------IndexError Traceback (most recent call last)Cell In[343], line 1----> 1 arr[np.arange(10),idx,:]IndexError: index 3 is out of bounds for axis 0 with size 3oops，颠倒索引顺序：In [344]: arr[idx,np.arange(10),:]Out[344]:array([[4, 2, 4, 9],[1, 9, 1, 4],[7, 0, 1, 4],[1, 6, 8, 9],[5, 9, 5, 1],[8, 5, 0, 7],[7, 1, 1, 5],[7, 9, 8, 2],[2, 6, 4, 9],[7, 4, 1, 9]])
英文:
hdr_data # (3,x,y) shapemaximum_intensities # (3,x) shapeframe_indices # (x,) shape, values 0,1,2
iterative:
for i, index in enumerate(frame_indices):main_scan.data[i] = hdr_data[index, i, :]
try:
max_data = hdr_data[np.arange(x)[:,None], frame_indices] # (x,y)
np.take_along_axis can also do this; see its docs
test
In [340]: arr = np.random.randint(0,10,(3,10,4))In [341]: idx=np.argmax(arr.max(axis=-1), axis=0)In [342]: idxOut[342]: array([0, 0, 0, 0, 1, 0, 0, 2, 0, 0], dtype=int64)In [343]: arr[np.arange(10),idx,:]---------------------------------------------------------------------------IndexError Traceback (most recent call last)Cell In[343], line 1----> 1 arr[np.arange(10),idx,:]IndexError: index 3 is out of bounds for axis 0 with size 3
oops, reverse the index order:
In [344]: arr[idx,np.arange(10),:]Out[344]:array([[4, 2, 4, 9],[1, 9, 1, 4],[7, 0, 1, 4],[1, 6, 8, 9],[5, 9, 5, 1],[8, 5, 0, 7],[7, 1, 1, 5],[7, 9, 8, 2],[2, 6, 4, 9],[7, 4, 1, 9]])
通过集体智慧和协作来改善编程学习和解决问题的方式。致力于成为全球开发者共同参与的知识库，让每个人都能够通过互相帮助和分享经验来进步。
评论