ConvTranspose - 1 vs 11

Next section compares an older to a newer version of the same operator after both definition are converted into markdown text.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. ConvTranspose1 → ConvTranspose11 +9 -22

ConvTranspose1 → ConvTranspose11 RENAMED Viewed

	@@ -1 +1 @@
1 1	The convolution transpose operator consumes an input tensor and a filter,
2 2	and computes the output.
3 3	If the pads parameter is provided the shape of the output is calculated via the following equation:
4 4	output_shape[i] = stride[i] * (input_size[i] - 1) + output_padding[i] + ((kernel_shape[i] - 1) * dilations[i] + 1) - pads[start_i] - pads[end_i]
5 5	output_shape can also be explicitly specified in which case pads values are auto generated using these equations:
6 6	total_padding[i] = stride[i] * (input_size[i] - 1) + output_padding[i] + ((kernel_shape[i] - 1) * dilations[i] + 1) - output_shape[i]
7	- If (auto_pads == SAME_UPPER): pads[start_i] = total_padding[i]/2; pads[end_i] = total_padding[i] - (total_padding[i]/2)
7	+ If (auto_pads != SAME_UPPER): pads[start_i] = total_padding[i]/2; pads[end_i] = total_padding[i] - (total_padding[i]/2)
8 8	Else: pads[start_i] = total_padding[i] - (total_padding[i]/2); pads[end_i] = (total_padding[i]/2).
9 9	**Attributes**
10 10	* **auto_pad**:
11 11	auto_pad must be either NOTSET, SAME_UPPER, SAME_LOWER or VALID.
12 12	Where default value is NOTSET, which means explicit padding is used.
13	- SAME_UPPER or SAME_LOWER mean pad the input so that output_shape[i]
13	+ SAME_UPPER or SAME_LOWER mean pad the input so that the output
14	+ spatial size match the input.In case of odd number add the extra
15	+ padding at the end for SAME_UPPER and at the beginning for
16	+ SAME_LOWER. VALID mean no padding.
14	- = input_shape[i] * strides[i] for each axis i. The padding is
15	- split between the two sides equally or almost equally (depending on
16	- whether it is even or odd). In case the padding is an odd number,
17	- the extra padding is added at the end for SAME_UPPER and at the
18	- beginning for SAME_LOWER.
19 17	* **dilations**:
20	- dilation value along each spatial axis of the filter. If not
18	+ dilation value along each spatial axis of the filter.
21	- present, the dilation defaults to 1 along each spatial axis.
22 19	* **group**:
23 20	number of groups input channels and output channels are divided
24 21	into.
25 22	* **kernel_shape**:
26 23	The shape of the convolution kernel. If not present, should be
27 24	inferred from input W.
28 25	* **output_padding**:
26	+ The zero-padding added to one side of the output. This is also
27	+ called adjs/adjustment in some frameworks.
29	- Additional elements added to the side with higher coordinate indices
30	- in the output. Each padding value in "output_padding" must be less
31	- than the corresponding stride/dilation dimension. By default, this
32	- attribute is a zero vector. Note that this attribute doesn't
33	- directly affect the computed output values. It only controls the
34	- selection of the computed values, so changing this attribute only
35	- adds or removes output elements. If "output_shape" is explicitly
36	- provided, "output_padding" does not contribute additional size to
37	- "output_shape" but participates in the computation of the needed
38	- padding amount. This is also called adjs or adjustment in some
39	- frameworks.
40 28	* **output_shape**:
41 29	The shape of the output can be explicitly set which will cause pads
42 30	values to be auto generated. If output_shape is specified pads
43 31	values are ignored. See doc for details for equations to generate
44 32	pads
45 33	* **pads**:
46 34	Padding for the beginning and ending along each spatial axis, it can
47 35	take any value greater than or equal to 0. The value represent the
48 36	number of pixels added to the beginning and end part of the
49 37	corresponding axis. pads format should be as follow [x1_begin,
50 38	x2_begin...x1_end, x2_end,...], where xi_begin the number of pixels
51 39	added at the beginning of axis i and xi_end, the number of pixels
52 40	added at the end of axis i. This attribute cannot be used
53 41	simultaneously with auto_pad attribute. If not present, the padding
54 42	defaults to 0 along start and end of each spatial axis.
55 43	* **strides**:
56	- Stride along each spatial axis. If not present, the stride defaults
57	- to 1 along each spatial axis.
44	+ Stride along each spatial axis.
58 45	**Inputs**
59 46	Between 2 and 3 inputs.
60 47	* **X** (heterogeneous) - **T**:
61 48	Input data tensor from previous layer; has size (N x C x H x W),
62 49	where N is the batch size, C is the number of channels, and H and W
63 50	are the height and width. Note that this is for the 2D image.
64 51	Otherwise the size is (N x C x D1 x D2 ... x Dn)
65 52	* **W** (heterogeneous) - **T**:
66 53	The weight tensor that will be used in the convolutions; has size (C
67 54	x M/group x kH x kW), where C is the number of channels, and kH and
68 55	kW are the height and width of the kernel, and M is the number of
69 56	feature maps. For more than 2 dimensions, the weight shape will be
70 57	(C x M/group x k1 x k2 x ... x kn), where (k1 x k2 x ... x kn) is
71 58	the dimension of the kernel. The number of channels in the output
72 59	should be equal to W.shape[1] * group (assuming zero based indices
73 60	of the shape array)
74 61	* **B** (optional, heterogeneous) - **T**:
75 62	Optional 1D bias to be added to the convolution, has size of M.
76 63	**Outputs**
77 64	* **Y** (heterogeneous) - **T**:
78 65	Output data tensor that contains the result of the convolution. The
79 66	output dimensions are functions of the kernel size, stride size, pad
80 67	lengths and group count. The number of channels in the output should
81 68	be equal to W.shape[1] * group (assuming zero based indices of the
82 69	shape array)
83 70	**Type Constraints**
84 71	* **T** in (
85 72	tensor(double),
86 73	tensor(float),
87 74	tensor(float16)
88 75	):
89 76	Constrain input and output types to float tensors.