t2 = time.time()
for j in range(100000):
z1 = bgr.transpose(2, 0, 1)[::-1, ...]
print(time.time() - t2)
t1=time.time()
for i in range(100000):
z=bgr2[:, :, ::-1].swapaxes(1,2).swapaxes(0,1)
print(time.time()-t1)
print(z.shape)
print(z1.shape)
'''0.06399011611938477
0.09599995613098145
(3, 1067, 650)
(3, 1067, 650)'''
可以看出相差不大,但是总体来说numpy的速度快
交换维度,扩充维度,组合batch方式对比 for j in range(10):
z1 = bgr.transpose(2, 0, 1)[::-1, ...]
img1 = np.array([z1, z1])
print(img1.shape,time.time() - t2)
t1=time.time()
for i in range(10):
z=bgr2[:, :, ::-1].swapaxes(1,2).swapaxes(0,1)
img2=np.concatenate([np.expand_dims(z, axis=0),np.expand_dims(z, axis=0)])
print(img2.shape,time.time()-t1)
t1=time.time()
for i in range(10):
z=bgr2[:, :, ::-1].swapaxes(1,2).swapaxes(0,1)
img2=np.concatenate([z.reshape(1, 3, 1067, 650),z.reshape(1, 3, 1067, 650)])
print(img2.shape,time.time()-t1)
'''(2, 3, 1067, 650) 0.02399611473083496
(2, 3, 1067, 650) 0.05620718002319336
(2, 3, 1067, 650) 0.05584287643432617'''
transpose之后使用array是最优的
扩充维度对比
for i in range(100000):
z11=z.reshape(1, 3, 1067, 650)
print("reshape time",time.time() - t1)
t1 = time.time()
for i in range(100000):
z12=np.expand_dims(z1,axis=0)
print("np.expand_dims time",time.time() - t1)
'''reshape time 0.0431513786315918
np.expand_dims time 0.4829425811767578'''
reshpe的速度是最快的
结论:#多batch最优方案 z1 = img.transpose(2, 0, 1)[::-1, ...] img1 = np.array([z1, z1]) #扩充维度最优方案 z1 = img.transpose(2, 0, 1)[::-1, ...] img2=z1.reshape(1,3,h,w)



