Resize - 11 vs 13

Next section compares an older to a newer version of the same operator after both definition are converted into markdown text. Green means an addition to the newer version, red means a deletion. Anything else is unchanged.

Files changed (1) hide show

1. Resize11 → Resize13 +9 -10

Resize11 → Resize13 RENAMED Viewed

	@@ -1 +1 @@
1 1	Resize the input tensor. In general, it calculates every value in the output tensor as a weighted average of neighborhood (a.k.a. sampling locations) in the input tensor.
2 2	Each dimension value of the output tensor is:
3 3	output_dimension = floor(input_dimension * (roi_end - roi_start) * scale) if input "sizes" is not specified.
4 4	**Attributes**
5 5	* **coordinate_transformation_mode**:
6 6	This attribute describes how to transform the coordinate in the
7 7	resized tensor to the coordinate in the original tensor. <br/> The
8 8	coordinate of each dimension is transformed individually. Let's
9 9	describe a case using axis x as an example. Denote x_resized as the
10 10	coordinate of axis x in the resized tensor, x_original as the
11 11	coordinate of axis x in the original tensor, length_original as the
12 12	length of the original tensor in axis x, length_resized as the
13 13	length of the resized tensor in axis x, roi_x = (start_x, end_x) of
14 14	the axis x in input "roi", scale = length_resized / length_original,
15 15	<br/> if coordinate_transformation_mode is "half_pixel", <br/>
16 16	x_original = (x_resized + 0.5) / scale - 0.5, <br/> if
17 17	coordinate_transformation_mode is "pytorch_half_pixel", <br/>
18 18	x_original = length_resized > 1 ? (x_resized + 0.5) / scale - 0.5 :
19 19	0, <br/> if coordinate_transformation_mode is "align_corners",
20 20	<br/> x_original = x_resized * (length_original - 1) /
21 21	(length_resized - 1), <br/> if coordinate_transformation_mode is
22 22	"asymmetric", <br/> x_original = x_resized / scale, <br/> if
23	+ coordinate_transformation_mode is "tf_half_pixel_for_nn", <br/>
24	+ x_original = (x_resized + 0.5) / scale, <br/> if
23 25	coordinate_transformation_mode is "tf_crop_and_resize", <br/>
24 26	x_original = length_resized > 1 ? start_x * (length_original - 1) +
25 27	x_resized * (end_x - start_x) * (length_original - 1) /
26 28	(length_resized - 1) : 0.5 * (start_x + end_x) * (length_original -
27 29	1).
28 30	* **cubic_coeff_a**:
29 31	The coefficient 'a' used in cubic interpolation. Two common choice
30 32	are -0.5 (in some cases of TensorFlow) and -0.75 (in PyTorch). Check
31 33	out Equation (4) in https://ieeexplore.ieee.org/document/1163711 for
32 34	the details. This attribute is valid only if "mode" is "cubic".
33 35	* **exclude_outside**:
34 36	If set to 1, the weight of sampling locations outside the tensor
35 37	will be set to 0 and the weight will be renormalized so that their
36 38	sum is 1.0. The default value is 0.
37 39	* **extrapolation_value**:
38 40	When coordinate_transformation_mode is "tf_crop_and_resize" and
39 41	x_original is outside the range [0, length_original - 1], this value
40 42	is used as the corresponding output value. Default is 0.0f.
41 43	* **mode**:
42 44	Three interpolation modes: nearest (default), linear and cubic. The
43 45	"linear" mode includes linear interpolation for 1D tensor and
44 46	N-linear interpolation for N-D tensor (for example, bilinear
45 47	interpolation for 2D tensor). The "cubic" mode includes cubic
46 48	interpolation for 1D tensor and N-cubic interpolation for N-D tensor
47 49	(for example, bicubic interpolation for 2D tensor).
48 50	* **nearest_mode**:
49 51	Four modes: round_prefer_floor (default, as known as round half
50 52	down), round_prefer_ceil (as known as round half up), floor, ceil.
51 53	Only used by nearest interpolation. It indicates how to get
52 54	"nearest" pixel in input tensor from x_original, so this attribute
53 55	is valid only if "mode" is "nearest".
54 56	**Inputs**
55	- Between 1 and 4 inputs.
57	+ Between 3 and 4 inputs.
56 58	* **X** (heterogeneous) - **T1**:
57 59	N-D tensor
58	- * **roi** (optional, heterogeneous) - **T2**:
60	+ * **roi** (heterogeneous) - **T2**:
59 61	1-D tensor given as [start1, ..., startN, end1, ..., endN], where N
60 62	is the rank of X. The RoIs' coordinates are normalized in the
61 63	coordinate system of the input image. It only takes effect when
62 64	coordinate_transformation_mode is "tf_crop_and_resize"
63	- * **scales** (optional, heterogeneous) - **tensor(float)**:
65	+ * **scales** (heterogeneous) - **tensor(float)**:
64 66	The scale array along each dimension. It takes value greater than 0.
65 67	If it's less than 1, it's sampling down, otherwise, it's upsampling.
66 68	The number of elements of 'scales' should be the same as the rank of
69	+ input 'X'. If 'size' is needed, the user must set 'scales' to an
70	+ empty tensor.
67	- input 'X'. One of 'scales' and 'sizes' MUST be specified and it is
68	- an error if both are specified. If 'sizes' is needed, the user can
69	- use an empty string as the name of 'scales' in this operator's input
70	- list.
71 71	* **sizes** (optional, heterogeneous) - **tensor(int64)**:
72 72	The size of the output tensor. The number of elements of 'sizes'
73	- should be the same as the rank of input 'X'. Only one of 'scales'
73	+ should be the same as the rank of input 'X'. May only be set if
74	- and 'sizes' can be specified.
74	+ 'scales' is set to an empty tensor.
75 75	**Outputs**
76 76	* **Y** (heterogeneous) - **T1**:
77 77	N-D tensor after resizing
78 78	**Type Constraints**
79 79	* **T1** in (
80	- tensor(bfloat16),
81 80	tensor(bool),
82 81	tensor(complex128),
83 82	tensor(complex64),
84 83	tensor(double),
85 84	tensor(float),
86 85	tensor(float16),
87 86	tensor(int16),
88 87	tensor(int32),
89 88	tensor(int64),
90 89	tensor(int8),
91 90	tensor(string),
92 91	tensor(uint16),
93 92	tensor(uint32),
94 93	tensor(uint64),
95 94	tensor(uint8)
96 95	):
97 96	Constrain input 'X' and output 'Y' to all tensor types.
98 97	* **T2** in (
99 98	tensor(double),
100 99	tensor(float),
101 100	tensor(float16)
102 101	):
103 102	Constrain roi type to float or double.