Examples

Here are some simple examples of what you can do with augpy tensors.

Create a new tensor and do basic math on it:

>>> import augpy
>>> t = augpy.CudaTensor((3, 3)).fill(7)
>>> print(t.numpy())
[[7 7 7]
 [7 7 7]
 [7 7 7]]
>>> t += 11
>>> print(t.numpy())
[[18 18 18]
 [18 18 18]
 [18 18 18]]

Decode a JPEG image on the GPU:

>>> with open('cat.jpg', 'rb') as f:
...     data = f.read()
...
>>> decoder = augpy.Decoder()
>>> img = decoder.decode(data)
>>> print(img)
<CudaTensor shape=(480, 640, 3), device=0, dtype=uint8>

Convert a numpy array to augpy tensor:

>>> import numpy as np
>>> array = np.array([124, 116, 104], dtype=np.uint8)
>>> background = augpy.array_to_tensor(array)
>>> print(background)
<CudaTensor shape=(3), device=0, dtype=uint8>

Apply affine warp to an image:

>>> warped = augpy.CudaTensor((3, 224, 224))
>>> m, s = augpy.make_transform(img.shape[:2], warped.shape[1:], angle=10)
>>> print(m)
[[  2.1103022   -0.37210324 125.3217    ]
 [  0.37210324   2.1103022  -38.029423  ]]
>>> augpy.warp_affine(img, warped, m, background, s)
>>> print(warped.numpy())
[[[124 124 124 ... 138 138 138]
  [124 124 124 ... 138 138 138]
  [124 124 124 ... 138 138 137]
  ...
  [110 106 102 ... 124 124 124]
  [101  95  90 ... 124 124 124]
  [ 86  82  77 ... 124 124 124]]

 [[116 116 116 ... 101 101 101]
  [116 116 116 ... 101 101 101]
  [116 116 116 ... 100 101 100]
  ...
  [ 94  91  90 ... 116 116 116]
  [ 88  86  83 ... 116 116 116]
  [ 82  78  75 ... 116 116 116]]

 [[104 104 104 ...  68  68  68]
  [104 104 104 ...  68  68  68]
  [104 104 104 ...  67  68  67]
  ...
  [ 85  84  83 ... 104 104 104]
  [ 83  81  80 ... 104 104 104]
  [ 81  80  77 ... 104 104 104]]]

Add noise to red channel:

>>> gen = augpy.RandomNumberGenerator()
>>> noise = augpy.CudaTensor((224, 224), dtype=augpy.int8)
>>> gen.gaussian(noise, 5, 80)
>>> print(noise.numpy())
[[ -15   20   -3 ...    4   40   22]
 [  50   73  -30 ...   -8   45   58]
 [ -65  -32  -32 ...  127   35  -76]
 ...
 [  79  -23   94 ...   43  -17  127]
 [   0  -24   -4 ... -128   27   25]
 [  -1  -21  127 ...   52 -128  -34]]
>>> augpy.fma(0.5, noise, warped[0], warped[0])
>>> print(warped.numpy())
[[[116 134 122 ... 140 158 149]
  [149 160 109 ... 134 160 167]
  [ 92 108 108 ... 202 156  99]
  ...
  [150  94 149 ... 146 116 188]
  [101  83  88 ...  60 138 136]
  [ 86  72 140 ... 150  60 107]]

 [[116 116 116 ... 101 101 101]
  [116 116 116 ... 101 101 101]
  [116 116 116 ... 100 101 100]
  ...
  [ 94  91  90 ... 116 116 116]
  [ 88  86  83 ... 116 116 116]
  [ 82  78  75 ... 116 116 116]]

 [[104 104 104 ...  68  68  68]
  [104 104 104 ...  68  68  68]
  [104 104 104 ...  67  68  67]
  ...
  [ 85  84  83 ... 104 104 104]
  [ 83  81  80 ... 104 104 104]
  [ 81  80  77 ... 104 104 104]]]

Note

No clipping is required, since integer math is saturating. Over or underflows cannot occur.

Finally, export tensors to other frameworks like Pytorch:

>>> from torch.utils.dlpack import from_dlpack
>>> capsule = augpy.export_dltensor(warped)
>>> torch_tensor = from_dlpack(capsule)
>>> print(torch_tensor)
tensor([[[116, 134, 122,  ..., 140, 158, 149],
         [149, 160, 109,  ..., 134, 160, 167],
         [ 92, 108, 108,  ..., 202, 156,  99],
         ...,
         [150,  94, 149,  ..., 146, 116, 188],
         [101,  83,  88,  ...,  60, 138, 136],
         [ 86,  72, 140,  ..., 150,  60, 107]],

        [[116, 116, 116,  ..., 101, 101, 101],
         [116, 116, 116,  ..., 101, 101, 101],
         [116, 116, 116,  ..., 100, 101, 100],
         ...,
         [ 94,  91,  90,  ..., 116, 116, 116],
         [ 88,  86,  83,  ..., 116, 116, 116],
         [ 82,  78,  75,  ..., 116, 116, 116]],

        [[104, 104, 104,  ...,  68,  68,  68],
         [104, 104, 104,  ...,  68,  68,  68],
         [104, 104, 104,  ...,  67,  68,  67],
         ...,
         [ 85,  84,  83,  ..., 104, 104, 104],
         [ 83,  81,  80,  ..., 104, 104, 104],
         [ 81,  80,  77,  ..., 104, 104, 104]]], device='cuda:0',
       dtype=torch.uint8)