Conv - 1 vs 11

Files changed (1) hide show
  1. Conv1 → Conv11 +16 -9
Conv1 → Conv11 RENAMED
@@ -1 +1 @@
1
1
  The convolution operator consumes an input tensor and a filter, and
2
2
  computes the output.
3
3
  **Attributes**
4
4
  * **auto_pad**:
5
5
  auto_pad must be either NOTSET, SAME_UPPER, SAME_LOWER or VALID.
6
6
  Where default value is NOTSET, which means explicit padding is used.
7
- SAME_UPPER or SAME_LOWER mean pad the input so that the output
7
+ SAME_UPPER or SAME_LOWER mean pad the input so that output_shape[i]
8
+ = ceil(input_shape[i] / strides[i]) for each axis i. The padding
9
+ is split between the two sides equally or almost equally (depending
8
- spatial size match the input.In case of odd number add the extra
10
+ on whether it is even or odd). In case the padding is an odd number,
9
- padding at the end for SAME_UPPER and at the beginning for
11
+ the extra padding is added at the end for SAME_UPPER and at the
10
- SAME_LOWER. VALID mean no padding.
12
+ beginning for SAME_LOWER.
11
13
  * **dilations**:
12
- dilation value along each spatial axis of the filter.
14
+ dilation value along each spatial axis of the filter. If not
15
+ present, the dilation defaults is 1 along each spatial axis.
13
16
  * **group**:
14
17
  number of groups input channels and output channels are divided
15
18
  into.
16
19
  * **kernel_shape**:
17
20
  The shape of the convolution kernel. If not present, should be
18
21
  inferred from input W.
19
22
  * **pads**:
20
23
  Padding for the beginning and ending along each spatial axis, it can
21
24
  take any value greater than or equal to 0. The value represent the
22
25
  number of pixels added to the beginning and end part of the
23
26
  corresponding axis. pads format should be as follow [x1_begin,
24
27
  x2_begin...x1_end, x2_end,...], where xi_begin the number of pixels
25
28
  added at the beginning of axis i and xi_end, the number of pixels
26
29
  added at the end of axis i. This attribute cannot be used
27
30
  simultaneously with auto_pad attribute. If not present, the padding
28
31
  defaults to 0 along start and end of each spatial axis.
29
32
  * **strides**:
33
+ Stride along each spatial axis. If not present, the stride defaults
30
- Stride along each spatial axis.
34
+ is 1 along each spatial axis.
31
35
  **Inputs**
32
36
  Between 2 and 3 inputs.
33
37
  * **X** (heterogeneous) - **T**:
34
38
  Input data tensor from previous layer; has size (N x C x H x W),
35
39
  where N is the batch size, C is the number of channels, and H and W
36
40
  are the height and width. Note that this is for the 2D image.
37
41
  Otherwise the size is (N x C x D1 x D2 ... x Dn). Optionally, if
38
42
  dimension denotation is in effect, the operation expects input data
39
43
  tensor to arrive with the dimension denotation of [DATA_BATCH,
40
44
  DATA_CHANNEL, DATA_FEATURE, DATA_FEATURE ...].
41
45
  * **W** (heterogeneous) - **T**:
42
46
  The weight tensor that will be used in the convolutions; has size (M
43
47
  x C/group x kH x kW), where C is the number of channels, and kH and
44
48
  kW are the height and width of the kernel, and M is the number of
45
49
  feature maps. For more than 2 dimensions, the kernel shape will be
46
50
  (M x C/group x k1 x k2 x ... x kn), where (k1 x k2 x ... kn) is the
47
51
  dimension of the kernel. Optionally, if dimension denotation is in
48
52
  effect, the operation expects the weight tensor to arrive with the
49
53
  dimension denotation of [FILTER_OUT_CHANNEL, FILTER_IN_CHANNEL,
50
- FILTER_SPATIAL, FILTER_SPATIAL ...]. X.shape[1] == (W.shape[1] *
54
+ FILTER_SPATIAL, FILTER_SPATIAL ...]. Assuming zero based indices for
55
+ the shape array, X.shape[1] == (W.shape[1] * group) == C and
51
- group) == C (assuming zero based indices for the shape array). Or in
56
+ W.shape[0] mod G == 0. Or in other words FILTER_IN_CHANNEL
52
- other words FILTER_IN_CHANNEL should be equal to DATA_CHANNEL.
57
+ multiplied by the number of groups should be equal to DATA_CHANNEL
58
+ and the number of feature maps M should be a multiple of the number
59
+ of groups G.
53
60
  * **B** (optional, heterogeneous) - **T**:
54
61
  Optional 1D bias to be added to the convolution, has size of M.
55
62
  **Outputs**
56
63
  * **Y** (heterogeneous) - **T**:
57
64
  Output data tensor that contains the result of the convolution. The
58
65
  output dimensions are functions of the kernel size, stride size, and
59
66
  pad lengths.
60
67
  **Type Constraints**
61
68
  * **T** in (
62
69
  tensor(double),
63
70
  tensor(float),
64
71
  tensor(float16)
65
72
  ):
66
73
  Constrain input and output types to float tensors.