experiments/STT-tensorflow
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

Update ops-related pbtxt files.

Browse Source 
Change: 154614872
This commit is contained in:
A. Unique TensorFlower 2017-04-28 21:48:56 -08:00 committed by TensorFlower Gardener
parent 1d679a0476
commit 5652c3418b
1 changed files with 40 additions and 40 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

80
tensorflow/core/ops/ops.pbtxt
View File

@ -3212,7 +3212,7 @@ op {
}
output_arg {
name: "output"
description: "4-D with shape `[batch, height, width, depth]`, where:\n\n height = height_pad - crop_top - crop_bottom\n width = width_pad - crop_left - crop_right\n\nThe attr `block_size` must be greater than one. It indicates the block size.\n\nSome examples:\n\n(1) For the following input of shape `[4, 1, 1, 1]` and block_size of 2:\n\n```prettyprint\n[[[[1]]], [[[2]]], [[[3]]], [[[4]]]]\n```\n\nThe output tensor has shape `[1, 2, 2, 1]` and value:\n\n```prettyprint\nx = [[[[1], [2]], [[3], [4]]]]\n```\n\n(2) For the following input of shape `[4, 1, 1, 3]` and block_size of 2:\n\n```prettyprint\n[[[1, 2, 3]], [[4, 5, 6]], [[7, 8, 9]], [[10, 11, 12]]]\n```\n\nThe output tensor has shape `[1, 2, 2, 3]` and value:\n\n```prettyprint\nx = [[[[1, 2, 3], [4, 5, 6]],\n [[7, 8, 9], [10, 11, 12]]]]\n```\n\n(3) For the following input of shape `[4, 2, 2, 1]` and block_size of 2:\n\n```prettyprint\nx = [[[[1], [3]], [[9], [11]]],\n [[[2], [4]], [[10], [12]]],\n [[[5], [7]], [[13], [15]]],\n [[[6], [8]], [[14], [16]]]]\n```\n\nThe output tensor has shape `[1, 4, 4, 1]` and value:\n\n```prettyprint\nx = [[[1], [2], [3], [4]],\n [[5], [6], [7], [8]],\n [[9], [10], [11], [12]],\n [[13], [14], [15], [16]]]\n```\n\n(4) For the following input of shape `[8, 1, 2, 1]` and block_size of 2:\n\n```prettyprint\nx = [[[[1], [3]]], [[[9], [11]]], [[[2], [4]]], [[[10], [12]]],\n [[[5], [7]]], [[[13], [15]]], [[[6], [8]]], [[[14], [16]]]]\n```\n\nThe output tensor has shape `[2, 2, 4, 1]` and value:\n\n```prettyprint\nx = [[[[1], [3]], [[5], [7]]],\n [[[2], [4]], [[10], [12]]],\n [[[5], [7]], [[13], [15]]],\n [[[6], [8]], [[14], [16]]]]\n```"
description: "4-D with shape `[batch, height, width, depth]`, where:\n\n height = height_pad - crop_top - crop_bottom\n width = width_pad - crop_left - crop_right\n\nThe attr `block_size` must be greater than one. It indicates the block size.\n\nSome examples:\n\n(1) For the following input of shape `[4, 1, 1, 1]` and block_size of 2:\n\n```\n[[[[1]]], [[[2]]], [[[3]]], [[[4]]]]\n```\n\nThe output tensor has shape `[1, 2, 2, 1]` and value:\n\n```\nx = [[[[1], [2]], [[3], [4]]]]\n```\n\n(2) For the following input of shape `[4, 1, 1, 3]` and block_size of 2:\n\n```\n[[[1, 2, 3]], [[4, 5, 6]], [[7, 8, 9]], [[10, 11, 12]]]\n```\n\nThe output tensor has shape `[1, 2, 2, 3]` and value:\n\n```\nx = [[[[1, 2, 3], [4, 5, 6]],\n [[7, 8, 9], [10, 11, 12]]]]\n```\n\n(3) For the following input of shape `[4, 2, 2, 1]` and block_size of 2:\n\n```\nx = [[[[1], [3]], [[9], [11]]],\n [[[2], [4]], [[10], [12]]],\n [[[5], [7]], [[13], [15]]],\n [[[6], [8]], [[14], [16]]]]\n```\n\nThe output tensor has shape `[1, 4, 4, 1]` and value:\n\n```\nx = [[[1], [2], [3], [4]],\n [[5], [6], [7], [8]],\n [[9], [10], [11], [12]],\n [[13], [14], [15], [16]]]\n```\n\n(4) For the following input of shape `[8, 1, 2, 1]` and block_size of 2:\n\n```\nx = [[[[1], [3]]], [[[9], [11]]], [[[2], [4]]], [[[10], [12]]],\n [[[5], [7]]], [[[13], [15]]], [[[6], [8]]], [[[14], [16]]]]\n```\n\nThe output tensor has shape `[2, 2, 4, 1]` and value:\n\n```\nx = [[[[1], [3]], [[5], [7]]],\n [[[2], [4]], [[10], [12]]],\n [[[5], [7]], [[13], [15]]],\n [[[6], [8]], [[14], [16]]]]\n```"
type_attr: "T"
}
attr {
@ -3255,7 +3255,7 @@ op {
}
input_arg {
name: "crops"
description: "2-D with shape `[M, 2]`, all values must be >= 0.\n `crops[i] = [crop_start, crop_end]` specifies the amount to crop from input\n dimension `i + 1`, which corresponds to spatial dimension `i`. It is\n required that\n `crop_start[i] + crop_end[i] <= block_shape[i] * input_shape[i + 1]`.\n\nThis operation is equivalent to the following steps:\n\n1. Reshape `input` to `reshaped` of shape:\n [block_shape[0], ..., block_shape[M-1],\n batch / prod(block_shape),\n input_shape[1], ..., input_shape[N-1]]\n\n2. Permute dimensions of `reshaped` to produce `permuted` of shape\n [batch / prod(block_shape),\n\n input_shape[1], block_shape[0],\n ...,\n input_shape[M], block_shape[M-1],\n\n input_shape[M+1], ..., input_shape[N-1]]\n\n3. Reshape `permuted` to produce `reshaped_permuted` of shape\n [batch / prod(block_shape),\n\n input_shape[1] * block_shape[0],\n ...,\n input_shape[M] * block_shape[M-1],\n\n input_shape[M+1],\n ...,\n input_shape[N-1]]\n\n4. Crop the start and end of dimensions `[1, ..., M]` of\n `reshaped_permuted` according to `crops` to produce the output of shape:\n [batch / prod(block_shape),\n\n input_shape[1] * block_shape[0] - crops[0,0] - crops[0,1],\n ...,\n input_shape[M] * block_shape[M-1] - crops[M-1,0] - crops[M-1,1],\n\n input_shape[M+1], ..., input_shape[N-1]]\n\nSome examples:\n\n(1) For the following input of shape `[4, 1, 1, 1]`, `block_shape = [2, 2]`, and\n `crops = [[0, 0], [0, 0]]`:\n\n```prettyprint\n[[[[1]]], [[[2]]], [[[3]]], [[[4]]]]\n```\n\nThe output tensor has shape `[1, 2, 2, 1]` and value:\n\n```prettyprint\nx = [[[[1], [2]], [[3], [4]]]]\n```\n\n(2) For the following input of shape `[4, 1, 1, 3]`, `block_shape = [2, 2]`, and\n `crops = [[0, 0], [0, 0]]`:\n\n```prettyprint\n[[[1, 2, 3]], [[4, 5, 6]], [[7, 8, 9]], [[10, 11, 12]]]\n```\n\nThe output tensor has shape `[1, 2, 2, 3]` and value:\n\n```prettyprint\nx = [[[[1, 2, 3], [4, 5, 6]],\n [[7, 8, 9], [10, 11, 12]]]]\n```\n\n(3) For the following input of shape `[4, 2, 2, 1]`, `block_shape = [2, 2]`, and\n `crops = [[0, 0], [0, 0]]`:\n\n```prettyprint\nx = [[[[1], [3]], [[9], [11]]],\n [[[2], [4]], [[10], [12]]],\n [[[5], [7]], [[13], [15]]],\n [[[6], [8]], [[14], [16]]]]\n```\n\nThe output tensor has shape `[1, 4, 4, 1]` and value:\n\n```prettyprint\nx = [[[1], [2], [3], [4]],\n [[5], [6], [7], [8]],\n [[9], [10], [11], [12]],\n [[13], [14], [15], [16]]]\n```\n\n(4) For the following input of shape `[8, 1, 3, 1]`, `block_shape = [2, 2]`, and\n `crops = [[0, 0], [2, 0]]`:\n\n```prettyprint\nx = [[[[0], [1], [3]]], [[[0], [9], [11]]],\n [[[0], [2], [4]]], [[[0], [10], [12]]],\n [[[0], [5], [7]]], [[[0], [13], [15]]],\n [[[0], [6], [8]]], [[[0], [14], [16]]]]\n```\n\nThe output tensor has shape `[2, 2, 4, 1]` and value:\n\n```prettyprint\nx = [[[[1], [2], [3], [4]],\n [[5], [6], [7], [8]]],\n [[[9], [10], [11], [12]],\n [[13], [14], [15], [16]]]]\n```"
description: "2-D with shape `[M, 2]`, all values must be >= 0.\n `crops[i] = [crop_start, crop_end]` specifies the amount to crop from input\n dimension `i + 1`, which corresponds to spatial dimension `i`. It is\n required that\n `crop_start[i] + crop_end[i] <= block_shape[i] * input_shape[i + 1]`.\n\nThis operation is equivalent to the following steps:\n\n1. Reshape `input` to `reshaped` of shape:\n [block_shape[0], ..., block_shape[M-1],\n batch / prod(block_shape),\n input_shape[1], ..., input_shape[N-1]]\n\n2. Permute dimensions of `reshaped` to produce `permuted` of shape\n [batch / prod(block_shape),\n\n input_shape[1], block_shape[0],\n ...,\n input_shape[M], block_shape[M-1],\n\n input_shape[M+1], ..., input_shape[N-1]]\n\n3. Reshape `permuted` to produce `reshaped_permuted` of shape\n [batch / prod(block_shape),\n\n input_shape[1] * block_shape[0],\n ...,\n input_shape[M] * block_shape[M-1],\n\n input_shape[M+1],\n ...,\n input_shape[N-1]]\n\n4. Crop the start and end of dimensions `[1, ..., M]` of\n `reshaped_permuted` according to `crops` to produce the output of shape:\n [batch / prod(block_shape),\n\n input_shape[1] * block_shape[0] - crops[0,0] - crops[0,1],\n ...,\n input_shape[M] * block_shape[M-1] - crops[M-1,0] - crops[M-1,1],\n\n input_shape[M+1], ..., input_shape[N-1]]\n\nSome examples:\n\n(1) For the following input of shape `[4, 1, 1, 1]`, `block_shape = [2, 2]`, and\n `crops = [[0, 0], [0, 0]]`:\n\n```\n[[[[1]]], [[[2]]], [[[3]]], [[[4]]]]\n```\n\nThe output tensor has shape `[1, 2, 2, 1]` and value:\n\n```\nx = [[[[1], [2]], [[3], [4]]]]\n```\n\n(2) For the following input of shape `[4, 1, 1, 3]`, `block_shape = [2, 2]`, and\n `crops = [[0, 0], [0, 0]]`:\n\n```\n[[[1, 2, 3]], [[4, 5, 6]], [[7, 8, 9]], [[10, 11, 12]]]\n```\n\nThe output tensor has shape `[1, 2, 2, 3]` and value:\n\n```\nx = [[[[1, 2, 3], [4, 5, 6]],\n [[7, 8, 9], [10, 11, 12]]]]\n```\n\n(3) For the following input of shape `[4, 2, 2, 1]`, `block_shape = [2, 2]`, and\n `crops = [[0, 0], [0, 0]]`:\n\n```\nx = [[[[1], [3]], [[9], [11]]],\n [[[2], [4]], [[10], [12]]],\n [[[5], [7]], [[13], [15]]],\n [[[6], [8]], [[14], [16]]]]\n```\n\nThe output tensor has shape `[1, 4, 4, 1]` and value:\n\n```\nx = [[[1], [2], [3], [4]],\n [[5], [6], [7], [8]],\n [[9], [10], [11], [12]],\n [[13], [14], [15], [16]]]\n```\n\n(4) For the following input of shape `[8, 1, 3, 1]`, `block_shape = [2, 2]`, and\n `crops = [[0, 0], [2, 0]]`:\n\n```\nx = [[[[0], [1], [3]]], [[[0], [9], [11]]],\n [[[0], [2], [4]]], [[[0], [10], [12]]],\n [[[0], [5], [7]]], [[[0], [13], [15]]],\n [[[0], [6], [8]]], [[[0], [14], [16]]]]\n```\n\nThe output tensor has shape `[2, 2, 4, 1]` and value:\n\n```\nx = [[[[1], [2], [3], [4]],\n [[5], [6], [7], [8]]],\n [[[9], [10], [11], [12]],\n [[13], [14], [15], [16]]]]\n```"
type_attr: "Tcrops"
}
output_arg {
@ -4103,7 +4103,7 @@ op {
minimum: 2
}
summary: "Computes offsets of concat inputs within its output."
description: "For example:\n\n```prettyprint\n# \'x\' is [2, 2, 7]\n# \'y\' is [2, 3, 7]\n# \'z\' is [2, 5, 7]\nconcat_offset(2, [x, y, z]) => [0, 0, 0], [0, 2, 0], [0, 5, 0]\n```"
description: "For example:\n\n```\n# \'x\' is [2, 2, 7]\n# \'y\' is [2, 3, 7]\n# \'z\' is [2, 5, 7]\nconcat_offset(2, [x, y, z]) => [0, 0, 0], [0, 2, 0], [0, 5, 0]\n```"
}
op {
name: "ConcatV2"
@ -5810,7 +5810,7 @@ op {
minimum: 2
}
summary: "DepthToSpace for tensors of type T."
description: "Rearranges data from depth into blocks of spatial data.\nThis is the reverse transformation of SpaceToDepth. More specifically,\nthis op outputs a copy of the input tensor where values from the `depth`\ndimension are moved in spatial blocks to the `height` and `width` dimensions.\nThe attr `block_size` indicates the input block size and how the data is moved.\n\n * Chunks of data of size `block_size * block_size` from depth are rearranged\n into non-overlapping blocks of size `block_size x block_size`\n * The width the output tensor is `input_depth * block_size`, whereas the\n height is `input_height * block_size`.\n * The depth of the input tensor must be divisible by\n `block_size * block_size`.\n\nThat is, assuming the input is in the shape:\n`[batch, height, width, depth]`,\nthe shape of the output will be:\n`[batch, height*block_size, width*block_size, depth/(block_size*block_size)]`\n\nThis operation requires that the input tensor be of rank 4, and that\n`block_size` be >=1 and that `block_size * block_size` be a divisor of the\ninput depth.\n\nThis operation is useful for resizing the activations between convolutions\n(but keeping all data), e.g. instead of pooling. It is also useful for training\npurely convolutional models.\n\nFor example, given this input of shape `[1, 1, 1, 4]`, and a block size of 2:\n\n```prettyprint\nx = [[[[1, 2, 3, 4]]]]\n\n```\n\nThis operation will output a tensor of shape `[1, 2, 2, 1]`:\n\n```prettyprint\n [[[[1], [2]],\n [[3], [4]]]]\n```\n\nHere, the input has a batch of 1 and each batch element has shape `[1, 1, 4]`,\nthe corresponding output will have 2x2 elements and will have a depth of\n1 channel (1 = `4 / (block_size * block_size)`).\nThe output element shape is `[2, 2, 1]`.\n\nFor an input tensor with larger depth, here of shape `[1, 1, 1, 12]`, e.g.\n\n```prettyprint\nx = [[[[1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12]]]]\n```\n\nThis operation, for block size of 2, will return the following tensor of shape\n`[1, 2, 2, 3]`\n\n```prettyprint\n [[[[1, 2, 3], [4, 5, 6]],\n [[7, 8, 9], [10, 11, 12]]]]\n\n```\n\nSimilarly, for the following input of shape `[1 2 2 4]`, and a block size of 2:\n\n```prettyprint\nx = [[[[1, 2, 3, 4],\n [5, 6, 7, 8]],\n [[9, 10, 11, 12],\n [13, 14, 15, 16]]]]\n```\n\nthe operator will return the following tensor of shape `[1 4 4 1]`:\n\n```prettyprint\nx = [[ [1], [2], [5], [6]],\n [ [3], [4], [7], [8]],\n [ [9], [10], [13], [14]],\n [ [11], [12], [15], [16]]]\n\n```"
description: "Rearranges data from depth into blocks of spatial data.\nThis is the reverse transformation of SpaceToDepth. More specifically,\nthis op outputs a copy of the input tensor where values from the `depth`\ndimension are moved in spatial blocks to the `height` and `width` dimensions.\nThe attr `block_size` indicates the input block size and how the data is moved.\n\n * Chunks of data of size `block_size * block_size` from depth are rearranged\n into non-overlapping blocks of size `block_size x block_size`\n * The width the output tensor is `input_depth * block_size`, whereas the\n height is `input_height * block_size`.\n * The depth of the input tensor must be divisible by\n `block_size * block_size`.\n\nThat is, assuming the input is in the shape:\n`[batch, height, width, depth]`,\nthe shape of the output will be:\n`[batch, height*block_size, width*block_size, depth/(block_size*block_size)]`\n\nThis operation requires that the input tensor be of rank 4, and that\n`block_size` be >=1 and that `block_size * block_size` be a divisor of the\ninput depth.\n\nThis operation is useful for resizing the activations between convolutions\n(but keeping all data), e.g. instead of pooling. It is also useful for training\npurely convolutional models.\n\nFor example, given this input of shape `[1, 1, 1, 4]`, and a block size of 2:\n\n```\nx = [[[[1, 2, 3, 4]]]]\n\n```\n\nThis operation will output a tensor of shape `[1, 2, 2, 1]`:\n\n```\n [[[[1], [2]],\n [[3], [4]]]]\n```\n\nHere, the input has a batch of 1 and each batch element has shape `[1, 1, 4]`,\nthe corresponding output will have 2x2 elements and will have a depth of\n1 channel (1 = `4 / (block_size * block_size)`).\nThe output element shape is `[2, 2, 1]`.\n\nFor an input tensor with larger depth, here of shape `[1, 1, 1, 12]`, e.g.\n\n```\nx = [[[[1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12]]]]\n```\n\nThis operation, for block size of 2, will return the following tensor of shape\n`[1, 2, 2, 3]`\n\n```\n [[[[1, 2, 3], [4, 5, 6]],\n [[7, 8, 9], [10, 11, 12]]]]\n\n```\n\nSimilarly, for the following input of shape `[1 2 2 4]`, and a block size of 2:\n\n```\nx = [[[[1, 2, 3, 4],\n [5, 6, 7, 8]],\n [[9, 10, 11, 12],\n [13, 14, 15, 16]]]]\n```\n\nthe operator will return the following tensor of shape `[1 4 4 1]`:\n\n```\nx = [[ [1], [2], [5], [6]],\n [ [3], [4], [7], [8]],\n [ [9], [10], [13], [14]],\n [ [11], [12], [15], [16]]]\n\n```"
}
op {
name: "DepthwiseConv2dNative"
@ -6123,7 +6123,7 @@ op {
}
}
summary: "Returns a diagonal tensor with a given diagonal values."
description: "Given a `diagonal`, this operation returns a tensor with the `diagonal` and\neverything else padded with zeros. The diagonal is computed as follows:\n\nAssume `diagonal` has dimensions [D1,..., Dk], then the output is a tensor of\nrank 2k with dimensions [D1,..., Dk, D1,..., Dk] where:\n\n`output[i1,..., ik, i1,..., ik] = diagonal[i1, ..., ik]` and 0 everywhere else.\n\nFor example:\n\n```prettyprint\n# \'diagonal\' is [1, 2, 3, 4]\ntf.diag(diagonal) ==> [[1, 0, 0, 0]\n [0, 2, 0, 0]\n [0, 0, 3, 0]\n [0, 0, 0, 4]]\n```"
description: "Given a `diagonal`, this operation returns a tensor with the `diagonal` and\neverything else padded with zeros. The diagonal is computed as follows:\n\nAssume `diagonal` has dimensions [D1,..., Dk], then the output is a tensor of\nrank 2k with dimensions [D1,..., Dk, D1,..., Dk] where:\n\n`output[i1,..., ik, i1,..., ik] = diagonal[i1, ..., ik]` and 0 everywhere else.\n\nFor example:\n\n```\n# \'diagonal\' is [1, 2, 3, 4]\ntf.diag(diagonal) ==> [[1, 0, 0, 0]\n [0, 2, 0, 0]\n [0, 0, 3, 0]\n [0, 0, 0, 4]]\n```"
}
op {
name: "DiagPart"
@ -6152,7 +6152,7 @@ op {
}
}
summary: "Returns the diagonal part of the tensor."
description: "This operation returns a tensor with the `diagonal` part\nof the `input`. The `diagonal` part is computed as follows:\n\nAssume `input` has dimensions `[D1,..., Dk, D1,..., Dk]`, then the output is a\ntensor of rank `k` with dimensions `[D1,..., Dk]` where:\n\n`diagonal[i1,..., ik] = input[i1, ..., ik, i1,..., ik]`.\n\nFor example:\n\n```prettyprint\n# \'input\' is [[1, 0, 0, 0]\n [0, 2, 0, 0]\n [0, 0, 3, 0]\n [0, 0, 0, 4]]\n\ntf.diag_part(input) ==> [1, 2, 3, 4]\n```"
description: "This operation returns a tensor with the `diagonal` part\nof the `input`. The `diagonal` part is computed as follows:\n\nAssume `input` has dimensions `[D1,..., Dk, D1,..., Dk]`, then the output is a\ntensor of rank `k` with dimensions `[D1,..., Dk]` where:\n\n`diagonal[i1,..., ik] = input[i1, ..., ik, i1,..., ik]`.\n\nFor example:\n\n```\n# \'input\' is [[1, 0, 0, 0]\n [0, 2, 0, 0]\n [0, 0, 3, 0]\n [0, 0, 0, 4]]\n\ntf.diag_part(input) ==> [1, 2, 3, 4]\n```"
}
op {
name: "Digamma"
@ -6469,7 +6469,7 @@ op {
type: "type"
}
summary: "Partitions `data` into `num_partitions` tensors using indices from `partitions`."
description: "For each index tuple `js` of size `partitions.ndim`, the slice `data[js, ...]`\nbecomes part of `outputs[partitions[js]]`. The slices with `partitions[js] = i`\nare placed in `outputs[i]` in lexicographic order of `js`, and the first\ndimension of `outputs[i]` is the number of entries in `partitions` equal to `i`.\nIn detail,\n\n```python\n outputs[i].shape = [sum(partitions == i)] + data.shape[partitions.ndim:]\n\n outputs[i] = pack([data[js, ...] for js if partitions[js] == i])\n```\n\n`data.shape` must start with `partitions.shape`.\n\nFor example:\n\n```python\n # Scalar partitions.\n partitions = 1\n num_partitions = 2\n data = [10, 20]\n outputs[0] = [] # Empty with shape [0, 2]\n outputs[1] = [[10, 20]]\n\n # Vector partitions.\n partitions = [0, 0, 1, 1, 0]\n num_partitions = 2\n data = [10, 20, 30, 40, 50]\n outputs[0] = [10, 20, 50]\n outputs[1] = [30, 40]\n```\n\n<div style=\"width:70%; margin:auto; margin-bottom:10px; margin-top:20px;\">\n<img style=\"width:100%\" src=\"../../images/DynamicPartition.png\" alt>\n</div>"
description: "For each index tuple `js` of size `partitions.ndim`, the slice `data[js, ...]`\nbecomes part of `outputs[partitions[js]]`. The slices with `partitions[js] = i`\nare placed in `outputs[i]` in lexicographic order of `js`, and the first\ndimension of `outputs[i]` is the number of entries in `partitions` equal to `i`.\nIn detail,\n\n```python\n outputs[i].shape = [sum(partitions == i)] + data.shape[partitions.ndim:]\n\n outputs[i] = pack([data[js, ...] for js if partitions[js] == i])\n```\n\n`data.shape` must start with `partitions.shape`.\n\nFor example:\n\n```python\n # Scalar partitions.\n partitions = 1\n num_partitions = 2\n data = [10, 20]\n outputs[0] = [] # Empty with shape [0, 2]\n outputs[1] = [[10, 20]]\n\n # Vector partitions.\n partitions = [0, 0, 1, 1, 0]\n num_partitions = 2\n data = [10, 20, 30, 40, 50]\n outputs[0] = [10, 20, 50]\n outputs[1] = [30, 40]\n```\n\n<div style=\"width:70%; margin:auto; margin-bottom:10px; margin-top:20px;\">\n<img style=\"width:100%\" src=\"https://www.tensorflow.org/images/DynamicPartition.png\" alt>\n</div>"
}
op {
name: "DynamicStitch"
@ -6498,7 +6498,7 @@ op {
type: "type"
}
summary: "Interleave the values from the `data` tensors into a single tensor."
description: "Builds a merged tensor such that\n\n```python\n merged[indices[m][i, ..., j], ...] = data[m][i, ..., j, ...]\n```\n\nFor example, if each `indices[m]` is scalar or vector, we have\n\n```python\n # Scalar indices:\n merged[indices[m], ...] = data[m][...]\n\n # Vector indices:\n merged[indices[m][i], ...] = data[m][i, ...]\n```\n\nEach `data[i].shape` must start with the corresponding `indices[i].shape`,\nand the rest of `data[i].shape` must be constant w.r.t. `i`. That is, we\nmust have `data[i].shape = indices[i].shape + constant`. In terms of this\n`constant`, the output shape is\n\n merged.shape = [max(indices)] + constant\n\nValues are merged in order, so if an index appears in both `indices[m][i]` and\n`indices[n][j]` for `(m,i) < (n,j)` the slice `data[n][j]` will appear in the\nmerged result.\n\nFor example:\n\n```python\n indices[0] = 6\n indices[1] = [4, 1]\n indices[2] = [[5, 2], [0, 3]]\n data[0] = [61, 62]\n data[1] = [[41, 42], [11, 12]]\n data[2] = [[[51, 52], [21, 22]], [[1, 2], [31, 32]]]\n merged = [[1, 2], [11, 12], [21, 22], [31, 32], [41, 42],\n [51, 52], [61, 62]]\n```\n\n<div style=\"width:70%; margin:auto; margin-bottom:10px; margin-top:20px;\">\n<img style=\"width:100%\" src=\"../../images/DynamicStitch.png\" alt>\n</div>"
description: "Builds a merged tensor such that\n\n```python\n merged[indices[m][i, ..., j], ...] = data[m][i, ..., j, ...]\n```\n\nFor example, if each `indices[m]` is scalar or vector, we have\n\n```python\n # Scalar indices:\n merged[indices[m], ...] = data[m][...]\n\n # Vector indices:\n merged[indices[m][i], ...] = data[m][i, ...]\n```\n\nEach `data[i].shape` must start with the corresponding `indices[i].shape`,\nand the rest of `data[i].shape` must be constant w.r.t. `i`. That is, we\nmust have `data[i].shape = indices[i].shape + constant`. In terms of this\n`constant`, the output shape is\n\n merged.shape = [max(indices)] + constant\n\nValues are merged in order, so if an index appears in both `indices[m][i]` and\n`indices[n][j]` for `(m,i) < (n,j)` the slice `data[n][j]` will appear in the\nmerged result.\n\nFor example:\n\n```python\n indices[0] = 6\n indices[1] = [4, 1]\n indices[2] = [[5, 2], [0, 3]]\n data[0] = [61, 62]\n data[1] = [[41, 42], [11, 12]]\n data[2] = [[[51, 52], [21, 22]], [[1, 2], [31, 32]]]\n merged = [[1, 2], [11, 12], [21, 22], [31, 32], [41, 42],\n [51, 52], [61, 62]]\n```\n\n<div style=\"width:70%; margin:auto; margin-bottom:10px; margin-top:20px;\">\n<img style=\"width:100%\" src=\"https://www.tensorflow.org/images/DynamicStitch.png\" alt>\n</div>"
}
op {
name: "EditDistance"
@ -6990,7 +6990,7 @@ op {
}
}
summary: "Inserts a dimension of 1 into a tensor\'s shape."
description: "Given a tensor `input`, this operation inserts a dimension of 1 at the\ndimension index `dim` of `input`\'s shape. The dimension index `dim` starts at\nzero; if you specify a negative number for `dim` it is counted backward from\nthe end.\n\nThis operation is useful if you want to add a batch dimension to a single\nelement. For example, if you have a single image of shape `[height, width,\nchannels]`, you can make it a batch of 1 image with `expand_dims(image, 0)`,\nwhich will make the shape `[1, height, width, channels]`.\n\nOther examples:\n\n```prettyprint\n# \'t\' is a tensor of shape [2]\nshape(expand_dims(t, 0)) ==> [1, 2]\nshape(expand_dims(t, 1)) ==> [2, 1]\nshape(expand_dims(t, -1)) ==> [2, 1]\n\n# \'t2\' is a tensor of shape [2, 3, 5]\nshape(expand_dims(t2, 0)) ==> [1, 2, 3, 5]\nshape(expand_dims(t2, 2)) ==> [2, 3, 1, 5]\nshape(expand_dims(t2, 3)) ==> [2, 3, 5, 1]\n```\n\nThis operation requires that:\n\n`-1-input.dims() <= dim <= input.dims()`\n\nThis operation is related to `squeeze()`, which removes dimensions of\nsize 1."
description: "Given a tensor `input`, this operation inserts a dimension of 1 at the\ndimension index `dim` of `input`\'s shape. The dimension index `dim` starts at\nzero; if you specify a negative number for `dim` it is counted backward from\nthe end.\n\nThis operation is useful if you want to add a batch dimension to a single\nelement. For example, if you have a single image of shape `[height, width,\nchannels]`, you can make it a batch of 1 image with `expand_dims(image, 0)`,\nwhich will make the shape `[1, height, width, channels]`.\n\nOther examples:\n\n```\n# \'t\' is a tensor of shape [2]\nshape(expand_dims(t, 0)) ==> [1, 2]\nshape(expand_dims(t, 1)) ==> [2, 1]\nshape(expand_dims(t, -1)) ==> [2, 1]\n\n# \'t2\' is a tensor of shape [2, 3, 5]\nshape(expand_dims(t2, 0)) ==> [1, 2, 3, 5]\nshape(expand_dims(t2, 2)) ==> [2, 3, 1, 5]\nshape(expand_dims(t2, 3)) ==> [2, 3, 5, 1]\n```\n\nThis operation requires that:\n\n`-1-input.dims() <= dim <= input.dims()`\n\nThis operation is related to `squeeze()`, which removes dimensions of\nsize 1."
}
op {
name: "Expm1"
@ -7541,7 +7541,7 @@ op {
type: "type"
}
summary: "Creates a tensor filled with a scalar value."
description: "This operation creates a tensor of shape `dims` and fills it with `value`.\n\nFor example:\n\n```prettyprint\n# Output tensor has shape [2, 3].\nfill([2, 3], 9) ==> [[9, 9, 9]\n [9, 9, 9]]\n```"
description: "This operation creates a tensor of shape `dims` and fills it with `value`.\n\nFor example:\n\n```\n# Output tensor has shape [2, 3].\nfill([2, 3], 9) ==> [[9, 9, 9]\n [9, 9, 9]]\n```"
}
op {
name: "FixedLengthRecordReader"
@ -8479,7 +8479,7 @@ op {
}
}
summary: "Gather slices from `params` according to `indices`."
description: "`indices` must be an integer tensor of any dimension (usually 0-D or 1-D).\nProduces an output tensor with shape `indices.shape + params.shape[1:]` where:\n\n```python\n # Scalar indices\n output[:, ..., :] = params[indices, :, ... :]\n\n # Vector indices\n output[i, :, ..., :] = params[indices[i], :, ... :]\n\n # Higher rank indices\n output[i, ..., j, :, ... :] = params[indices[i, ..., j], :, ..., :]\n```\n\nIf `indices` is a permutation and `len(indices) == params.shape[0]` then\nthis operation will permute `params` accordingly.\n\n`validate_indices`: DEPRECATED. If this operation is assigned to CPU, values in\n`indices` are always validated to be within range. If assigned to GPU,\nout-of-bound indices result in safe but unspecified behavior, which may include\nraising an error.\n\n<div style=\"width:70%; margin:auto; margin-bottom:10px; margin-top:20px;\">\n<img style=\"width:100%\" src=\"../../../images/Gather.png\" alt>\n</div>"
description: "`indices` must be an integer tensor of any dimension (usually 0-D or 1-D).\nProduces an output tensor with shape `indices.shape + params.shape[1:]` where:\n\n```python\n # Scalar indices\n output[:, ..., :] = params[indices, :, ... :]\n\n # Vector indices\n output[i, :, ..., :] = params[indices[i], :, ... :]\n\n # Higher rank indices\n output[i, ..., j, :, ... :] = params[indices[i, ..., j], :, ..., :]\n```\n\nIf `indices` is a permutation and `len(indices) == params.shape[0]` then\nthis operation will permute `params` accordingly.\n\n`validate_indices`: DEPRECATED. If this operation is assigned to CPU, values in\n`indices` are always validated to be within range. If assigned to GPU,\nout-of-bound indices result in safe but unspecified behavior, which may include\nraising an error.\n\n<div style=\"width:70%; margin:auto; margin-bottom:10px; margin-top:20px;\">\n<img style=\"width:100%\" src=\"https://www.tensorflow.org/images/Gather.png\" alt>\n</div>"
}
op {
name: "GatherNd"
@ -9315,7 +9315,7 @@ op {
}
}
summary: "Computes the inverse permutation of a tensor."
description: "This operation computes the inverse of an index permutation. It takes a 1-D\ninteger tensor `x`, which represents the indices of a zero-based array, and\nswaps each value with its index position. In other words, for an output tensor\n`y` and an input tensor `x`, this operation computes the following:\n\n`y[x[i]] = i for i in [0, 1, ..., len(x) - 1]`\n\nThe values must include 0. There can be no duplicate values or negative values.\n\nFor example:\n\n```prettyprint\n# tensor `x` is [3, 4, 0, 2, 1]\ninvert_permutation(x) ==> [2, 4, 3, 0, 1]\n```"
description: "This operation computes the inverse of an index permutation. It takes a 1-D\ninteger tensor `x`, which represents the indices of a zero-based array, and\nswaps each value with its index position. In other words, for an output tensor\n`y` and an input tensor `x`, this operation computes the following:\n\n`y[x[i]] = i for i in [0, 1, ..., len(x) - 1]`\n\nThe values must include 0. There can be no duplicate values or negative values.\n\nFor example:\n\n```\n# tensor `x` is [3, 4, 0, 2, 1]\ninvert_permutation(x) ==> [2, 4, 3, 0, 1]\n```"
}
op {
name: "IsFinite"
@ -9811,7 +9811,7 @@ op {
}
}
summary: "Computes the difference between two lists of numbers or strings."
description: "Given a list `x` and a list `y`, this operation returns a list `out` that\nrepresents all values that are in `x` but not in `y`. The returned list `out`\nis sorted in the same order that the numbers appear in `x` (duplicates are\npreserved). This operation also returns a list `idx` that represents the\nposition of each `out` element in `x`. In other words:\n\n`out[i] = x[idx[i]] for i in [0, 1, ..., len(out) - 1]`\n\nFor example, given this input:\n\n```prettyprint\nx = [1, 2, 3, 4, 5, 6]\ny = [1, 3, 5]\n```\n\nThis operation would return:\n\n```prettyprint\nout ==> [2, 4, 6]\nidx ==> [1, 3, 5]\n```"
description: "Given a list `x` and a list `y`, this operation returns a list `out` that\nrepresents all values that are in `x` but not in `y`. The returned list `out`\nis sorted in the same order that the numbers appear in `x` (duplicates are\npreserved). This operation also returns a list `idx` that represents the\nposition of each `out` element in `x`. In other words:\n\n`out[i] = x[idx[i]] for i in [0, 1, ..., len(out) - 1]`\n\nFor example, given this input:\n\n```\nx = [1, 2, 3, 4, 5, 6]\ny = [1, 3, 5]\n```\n\nThis operation would return:\n\n```\nout ==> [2, 4, 6]\nidx ==> [1, 3, 5]\n```"
}
op {
name: "Log"
@ -10245,7 +10245,7 @@ op {
type: "type"
}
summary: "Copy a tensor setting everything outside a central band in each innermost matrix"
description: "to zero.\n\nThe `band` part is computed as follows:\nAssume `input` has `k` dimensions `[I, J, K, ..., M, N]`, then the output is a\ntensor with the same shape where\n\n`band[i, j, k, ..., m, n] = in_band(m, n) * input[i, j, k, ..., m, n]`.\n\nThe indicator function\n\n`in_band(m, n) = (num_lower < 0 || (m-n) <= num_lower)) &&\n (num_upper < 0 || (n-m) <= num_upper)`.\n\nFor example:\n\n```prettyprint\n# if \'input\' is [[ 0, 1, 2, 3]\n [-1, 0, 1, 2]\n [-2, -1, 0, 1]\n [-3, -2, -1, 0]],\n\ntf.matrix_band_part(input, 1, -1) ==> [[ 0, 1, 2, 3]\n [-1, 0, 1, 2]\n [ 0, -1, 0, 1]\n [ 0, 0, -1, 0]],\n\ntf.matrix_band_part(input, 2, 1) ==> [[ 0, 1, 0, 0]\n [-1, 0, 1, 0]\n [-2, -1, 0, 1]\n [ 0, -2, -1, 0]]\n```\n\nUseful special cases:\n\n```prettyprint\n tf.matrix_band_part(input, 0, -1) ==> Upper triangular part.\n tf.matrix_band_part(input, -1, 0) ==> Lower triangular part.\n tf.matrix_band_part(input, 0, 0) ==> Diagonal.\n```"
description: "to zero.\n\nThe `band` part is computed as follows:\nAssume `input` has `k` dimensions `[I, J, K, ..., M, N]`, then the output is a\ntensor with the same shape where\n\n`band[i, j, k, ..., m, n] = in_band(m, n) * input[i, j, k, ..., m, n]`.\n\nThe indicator function\n\n`in_band(m, n) = (num_lower < 0 || (m-n) <= num_lower)) &&\n (num_upper < 0 || (n-m) <= num_upper)`.\n\nFor example:\n\n```\n# if \'input\' is [[ 0, 1, 2, 3]\n [-1, 0, 1, 2]\n [-2, -1, 0, 1]\n [-3, -2, -1, 0]],\n\ntf.matrix_band_part(input, 1, -1) ==> [[ 0, 1, 2, 3]\n [-1, 0, 1, 2]\n [ 0, -1, 0, 1]\n [ 0, 0, -1, 0]],\n\ntf.matrix_band_part(input, 2, 1) ==> [[ 0, 1, 0, 0]\n [-1, 0, 1, 0]\n [-2, -1, 0, 1]\n [ 0, -2, -1, 0]]\n```\n\nUseful special cases:\n\n```\n tf.matrix_band_part(input, 0, -1) ==> Upper triangular part.\n tf.matrix_band_part(input, -1, 0) ==> Lower triangular part.\n tf.matrix_band_part(input, 0, 0) ==> Diagonal.\n```"
}
op {
name: "MatrixDeterminant"
@ -10289,7 +10289,7 @@ op {
type: "type"
}
summary: "Returns a batched diagonal tensor with a given batched diagonal values."
description: "Given a `diagonal`, this operation returns a tensor with the `diagonal` and\neverything else padded with zeros. The diagonal is computed as follows:\n\nAssume `diagonal` has `k` dimensions `[I, J, K, ..., N]`, then the output is a\ntensor of rank `k+1` with dimensions [I, J, K, ..., N, N]` where:\n\n`output[i, j, k, ..., m, n] = 1{m=n} * diagonal[i, j, k, ..., n]`.\n\nFor example:\n\n```prettyprint\n# \'diagonal\' is [[1, 2, 3, 4], [5, 6, 7, 8]]\n\nand diagonal.shape = (2, 4)\n\ntf.matrix_diag(diagonal) ==> [[[1, 0, 0, 0]\n [0, 2, 0, 0]\n [0, 0, 3, 0]\n [0, 0, 0, 4]],\n [[5, 0, 0, 0]\n [0, 6, 0, 0]\n [0, 0, 7, 0]\n [0, 0, 0, 8]]]\n\nwhich has shape (2, 4, 4)\n```"
description: "Given a `diagonal`, this operation returns a tensor with the `diagonal` and\neverything else padded with zeros. The diagonal is computed as follows:\n\nAssume `diagonal` has `k` dimensions `[I, J, K, ..., N]`, then the output is a\ntensor of rank `k+1` with dimensions [I, J, K, ..., N, N]` where:\n\n`output[i, j, k, ..., m, n] = 1{m=n} * diagonal[i, j, k, ..., n]`.\n\nFor example:\n\n```\n# \'diagonal\' is [[1, 2, 3, 4], [5, 6, 7, 8]]\n\nand diagonal.shape = (2, 4)\n\ntf.matrix_diag(diagonal) ==> [[[1, 0, 0, 0]\n [0, 2, 0, 0]\n [0, 0, 3, 0]\n [0, 0, 0, 4]],\n [[5, 0, 0, 0]\n [0, 6, 0, 0]\n [0, 0, 7, 0]\n [0, 0, 0, 8]]]\n\nwhich has shape (2, 4, 4)\n```"
}
op {
name: "MatrixDiagPart"
@ -10308,7 +10308,7 @@ op {
type: "type"
}
summary: "Returns the batched diagonal part of a batched tensor."
description: "This operation returns a tensor with the `diagonal` part\nof the batched `input`. The `diagonal` part is computed as follows:\n\nAssume `input` has `k` dimensions `[I, J, K, ..., M, N]`, then the output is a\ntensor of rank `k - 1` with dimensions `[I, J, K, ..., min(M, N)]` where:\n\n`diagonal[i, j, k, ..., n] = input[i, j, k, ..., n, n]`.\n\nThe input must be at least a matrix.\n\nFor example:\n\n```prettyprint\n# \'input\' is [[[1, 0, 0, 0]\n [0, 2, 0, 0]\n [0, 0, 3, 0]\n [0, 0, 0, 4]],\n [[5, 0, 0, 0]\n [0, 6, 0, 0]\n [0, 0, 7, 0]\n [0, 0, 0, 8]]]\n\nand input.shape = (2, 4, 4)\n\ntf.matrix_diag_part(input) ==> [[1, 2, 3, 4], [5, 6, 7, 8]]\n\nwhich has shape (2, 4)\n```"
description: "This operation returns a tensor with the `diagonal` part\nof the batched `input`. The `diagonal` part is computed as follows:\n\nAssume `input` has `k` dimensions `[I, J, K, ..., M, N]`, then the output is a\ntensor of rank `k - 1` with dimensions `[I, J, K, ..., min(M, N)]` where:\n\n`diagonal[i, j, k, ..., n] = input[i, j, k, ..., n, n]`.\n\nThe input must be at least a matrix.\n\nFor example:\n\n```\n# \'input\' is [[[1, 0, 0, 0]\n [0, 2, 0, 0]\n [0, 0, 3, 0]\n [0, 0, 0, 4]],\n [[5, 0, 0, 0]\n [0, 6, 0, 0]\n [0, 0, 7, 0]\n [0, 0, 0, 8]]]\n\nand input.shape = (2, 4, 4)\n\ntf.matrix_diag_part(input) ==> [[1, 2, 3, 4], [5, 6, 7, 8]]\n\nwhich has shape (2, 4)\n```"
}
op {
name: "MatrixInverse"
@ -11601,7 +11601,7 @@ op {
}
}
summary: "Pads a tensor with mirrored values."
description: "This operation pads a `input` with mirrored values according to the `paddings`\nyou specify. `paddings` is an integer tensor with shape `[n, 2]`, where n is\nthe rank of `input`. For each dimension D of `input`, `paddings[D, 0]` indicates\nhow many values to add before the contents of `input` in that dimension, and\n`paddings[D, 1]` indicates how many values to add after the contents of `input`\nin that dimension. Both `paddings[D, 0]` and `paddings[D, 1]` must be no greater\nthan `input.dim_size(D)` (or `input.dim_size(D) - 1`) if `copy_border` is true\n(if false, respectively).\n\nThe padded size of each dimension D of the output is:\n\n`paddings(D, 0) + input.dim_size(D) + paddings(D, 1)`\n\nFor example:\n\n```prettyprint\n# \'t\' is [[1, 2, 3], [4, 5, 6]].\n# \'paddings\' is [[1, 1]], [2, 2]].\n# \'mode\' is SYMMETRIC.\n# rank of \'t\' is 2.\npad(t, paddings) ==> [[2, 1, 1, 2, 3, 3, 2]\n [2, 1, 1, 2, 3, 3, 2]\n [5, 4, 4, 5, 6, 6, 5]\n [5, 4, 4, 5, 6, 6, 5]]\n```"
description: "This operation pads a `input` with mirrored values according to the `paddings`\nyou specify. `paddings` is an integer tensor with shape `[n, 2]`, where n is\nthe rank of `input`. For each dimension D of `input`, `paddings[D, 0]` indicates\nhow many values to add before the contents of `input` in that dimension, and\n`paddings[D, 1]` indicates how many values to add after the contents of `input`\nin that dimension. Both `paddings[D, 0]` and `paddings[D, 1]` must be no greater\nthan `input.dim_size(D)` (or `input.dim_size(D) - 1`) if `copy_border` is true\n(if false, respectively).\n\nThe padded size of each dimension D of the output is:\n\n`paddings(D, 0) + input.dim_size(D) + paddings(D, 1)`\n\nFor example:\n\n```\n# \'t\' is [[1, 2, 3], [4, 5, 6]].\n# \'paddings\' is [[1, 1]], [2, 2]].\n# \'mode\' is SYMMETRIC.\n# rank of \'t\' is 2.\npad(t, paddings) ==> [[2, 1, 1, 2, 3, 3, 2]\n [2, 1, 1, 2, 3, 3, 2]\n [5, 4, 4, 5, 6, 6, 5]\n [5, 4, 4, 5, 6, 6, 5]]\n```"
}
op {
name: "MirrorPadGrad"
@ -11649,7 +11649,7 @@ op {
}
}
summary: "Gradient op for `MirrorPad` op. This op folds a mirror-padded tensor."
description: "This operation folds the padded areas of `input` by `MirrorPad` according to the\n`paddings` you specify. `paddings` must be the same as `paddings` argument\ngiven to the corresponding `MirrorPad` op.\n\nThe folded size of each dimension D of the output is:\n\n`input.dim_size(D) - paddings(D, 0) - paddings(D, 1)`\n\nFor example:\n\n```prettyprint\n# \'t\' is [[1, 2, 3], [4, 5, 6], [7, 8, 9]].\n# \'paddings\' is [[0, 1]], [0, 1]].\n# \'mode\' is SYMMETRIC.\n# rank of \'t\' is 2.\npad(t, paddings) ==> [[ 1, 5]\n [11, 28]]\n```"
description: "This operation folds the padded areas of `input` by `MirrorPad` according to the\n`paddings` you specify. `paddings` must be the same as `paddings` argument\ngiven to the corresponding `MirrorPad` op.\n\nThe folded size of each dimension D of the output is:\n\n`input.dim_size(D) - paddings(D, 0) - paddings(D, 1)`\n\nFor example:\n\n```\n# \'t\' is [[1, 2, 3], [4, 5, 6], [7, 8, 9]].\n# \'paddings\' is [[0, 1]], [0, 1]].\n# \'mode\' is SYMMETRIC.\n# rank of \'t\' is 2.\npad(t, paddings) ==> [[ 1, 5]\n [11, 28]]\n```"
}
op {
name: "Mod"
@ -12229,7 +12229,7 @@ op {
description: "Dimension along which to pack. Negative values wrap around, so the\nvalid range is `[-(R+1), R+1)`."
}
summary: "Packs a list of `N` rank-`R` tensors into one rank-`(R+1)` tensor."
description: "Packs the `N` tensors in `values` into a tensor with rank one higher than each\ntensor in `values`, by packing them along the `axis` dimension.\nGiven a list of tensors of shape `(A, B, C)`;\n\nif `axis == 0` then the `output` tensor will have the shape `(N, A, B, C)`.\nif `axis == 1` then the `output` tensor will have the shape `(A, N, B, C)`.\nEtc.\n\nFor example:\n\n```prettyprint\n# \'x\' is [1, 4]\n# \'y\' is [2, 5]\n# \'z\' is [3, 6]\npack([x, y, z]) => [[1, 4], [2, 5], [3, 6]] # Pack along first dim.\npack([x, y, z], axis=1) => [[1, 2, 3], [4, 5, 6]]\n```\n\nThis is the opposite of `unpack`."
description: "Packs the `N` tensors in `values` into a tensor with rank one higher than each\ntensor in `values`, by packing them along the `axis` dimension.\nGiven a list of tensors of shape `(A, B, C)`;\n\nif `axis == 0` then the `output` tensor will have the shape `(N, A, B, C)`.\nif `axis == 1` then the `output` tensor will have the shape `(A, N, B, C)`.\nEtc.\n\nFor example:\n\n```\n# \'x\' is [1, 4]\n# \'y\' is [2, 5]\n# \'z\' is [3, 6]\npack([x, y, z]) => [[1, 4], [2, 5], [3, 6]] # Pack along first dim.\npack([x, y, z], axis=1) => [[1, 2, 3], [4, 5, 6]]\n```\n\nThis is the opposite of `unpack`."
}
op {
name: "Pad"
@ -12263,7 +12263,7 @@ op {
}
}
summary: "Pads a tensor with zeros."
description: "This operation pads a `input` with zeros according to the `paddings` you\nspecify. `paddings` is an integer tensor with shape `[Dn, 2]`, where n is the\nrank of `input`. For each dimension D of `input`, `paddings[D, 0]` indicates\nhow many zeros to add before the contents of `input` in that dimension, and\n`paddings[D, 1]` indicates how many zeros to add after the contents of `input`\nin that dimension.\n\nThe padded size of each dimension D of the output is:\n\n`paddings(D, 0) + input.dim_size(D) + paddings(D, 1)`\n\nFor example:\n\n```prettyprint\n# \'t\' is [[1, 1], [2, 2]]\n# \'paddings\' is [[1, 1], [2, 2]]\n# rank of \'t\' is 2\npad(t, paddings) ==> [[0, 0, 0, 0, 0, 0]\n [0, 0, 1, 1, 0, 0]\n [0, 0, 2, 2, 0, 0]\n [0, 0, 0, 0, 0, 0]]\n```"
description: "This operation pads a `input` with zeros according to the `paddings` you\nspecify. `paddings` is an integer tensor with shape `[Dn, 2]`, where n is the\nrank of `input`. For each dimension D of `input`, `paddings[D, 0]` indicates\nhow many zeros to add before the contents of `input` in that dimension, and\n`paddings[D, 1]` indicates how many zeros to add after the contents of `input`\nin that dimension.\n\nThe padded size of each dimension D of the output is:\n\n`paddings(D, 0) + input.dim_size(D) + paddings(D, 1)`\n\nFor example:\n\n```\n# \'t\' is [[1, 1], [2, 2]]\n# \'paddings\' is [[1, 1], [2, 2]]\n# rank of \'t\' is 2\npad(t, paddings) ==> [[0, 0, 0, 0, 0, 0]\n [0, 0, 1, 1, 0, 0]\n [0, 0, 2, 2, 0, 0]\n [0, 0, 0, 0, 0, 0]]\n```"
}
op {
name: "PaddingFIFOQueue"
@ -12399,7 +12399,7 @@ op {
description: "the final shape of the result; should be equal to the shapes of any input\nbut with the number of input values in the first dimension."
}
summary: "Concatenates a list of `N` tensors along the first dimension."
description: "The input tensors are all required to have size 1 in the first dimension.\n\nFor example:\n\n```prettyprint\n# \'x\' is [[1, 4]]\n# \'y\' is [[2, 5]]\n# \'z\' is [[3, 6]]\nparallel_concat([x, y, z]) => [[1, 4], [2, 5], [3, 6]] # Pack along first dim.\n```\n\nThe difference between concat and parallel_concat is that concat requires all\nof the inputs be computed before the operation will begin but doesn\'t require\nthat the input shapes be known during graph construction. Parallel concat\nwill copy pieces of the input into the output as they become available, in\nsome situations this can provide a performance benefit."
description: "The input tensors are all required to have size 1 in the first dimension.\n\nFor example:\n\n```\n# \'x\' is [[1, 4]]\n# \'y\' is [[2, 5]]\n# \'z\' is [[3, 6]]\nparallel_concat([x, y, z]) => [[1, 4], [2, 5], [3, 6]] # Pack along first dim.\n```\n\nThe difference between concat and parallel_concat is that concat requires all\nof the inputs be computed before the operation will begin but doesn\'t require\nthat the input shapes be known during graph construction. Parallel concat\nwill copy pieces of the input into the output as they become available, in\nsome situations this can provide a performance benefit."
}
op {
name: "ParameterizedTruncatedNormal"
@ -15623,7 +15623,7 @@ op {
type: "type"
}
summary: "Returns the rank of a tensor."
description: "This operation returns an integer representing the rank of `input`.\n\nFor example:\n\n```prettyprint\n# \'t\' is [[[1, 1, 1], [2, 2, 2]], [[3, 3, 3], [4, 4, 4]]]\n# shape of tensor \'t\' is [2, 2, 3]\nrank(t) ==> 3\n```\n\n**Note**: The rank of a tensor is not the same as the rank of a matrix. The rank\nof a tensor is the number of indices required to uniquely select each element\nof the tensor. Rank is also known as \"order\", \"degree\", or \"ndims.\""
description: "This operation returns an integer representing the rank of `input`.\n\nFor example:\n\n```\n# \'t\' is [[[1, 1, 1], [2, 2, 2]], [[3, 3, 3], [4, 4, 4]]]\n# shape of tensor \'t\' is [2, 2, 3]\nrank(t) ==> 3\n```\n\n**Note**: The rank of a tensor is not the same as the rank of a matrix. The rank\nof a tensor is the number of indices required to uniquely select each element\nof the tensor. Rank is also known as \"order\", \"degree\", or \"ndims.\""
}
op {
name: "ReadFile"
@ -16594,7 +16594,7 @@ op {
}
}
summary: "Reshapes a tensor."
description: "Given `tensor`, this operation returns a tensor that has the same values\nas `tensor` with shape `shape`.\n\nIf one component of `shape` is the special value -1, the size of that dimension\nis computed so that the total size remains constant. In particular, a `shape`\nof `[-1]` flattens into 1-D. At most one component of `shape` can be -1.\n\nIf `shape` is 1-D or higher, then the operation returns a tensor with shape\n`shape` filled with the values of `tensor`. In this case, the number of elements\nimplied by `shape` must be the same as the number of elements in `tensor`.\n\nFor example:\n\n```prettyprint\n# tensor \'t\' is [1, 2, 3, 4, 5, 6, 7, 8, 9]\n# tensor \'t\' has shape [9]\nreshape(t, [3, 3]) ==> [[1, 2, 3],\n [4, 5, 6],\n [7, 8, 9]]\n\n# tensor \'t\' is [[[1, 1], [2, 2]],\n# [[3, 3], [4, 4]]]\n# tensor \'t\' has shape [2, 2, 2]\nreshape(t, [2, 4]) ==> [[1, 1, 2, 2],\n [3, 3, 4, 4]]\n\n# tensor \'t\' is [[[1, 1, 1],\n# [2, 2, 2]],\n# [[3, 3, 3],\n# [4, 4, 4]],\n# [[5, 5, 5],\n# [6, 6, 6]]]\n# tensor \'t\' has shape [3, 2, 3]\n# pass \'[-1]\' to flatten \'t\'\nreshape(t, [-1]) ==> [1, 1, 1, 2, 2, 2, 3, 3, 3, 4, 4, 4, 5, 5, 5, 6, 6, 6]\n\n# -1 can also be used to infer the shape\n\n# -1 is inferred to be 9:\nreshape(t, [2, -1]) ==> [[1, 1, 1, 2, 2, 2, 3, 3, 3],\n [4, 4, 4, 5, 5, 5, 6, 6, 6]]\n# -1 is inferred to be 2:\nreshape(t, [-1, 9]) ==> [[1, 1, 1, 2, 2, 2, 3, 3, 3],\n [4, 4, 4, 5, 5, 5, 6, 6, 6]]\n# -1 is inferred to be 3:\nreshape(t, [ 2, -1, 3]) ==> [[[1, 1, 1],\n [2, 2, 2],\n [3, 3, 3]],\n [[4, 4, 4],\n [5, 5, 5],\n [6, 6, 6]]]\n\n# tensor \'t\' is [7]\n# shape `[]` reshapes to a scalar\nreshape(t, []) ==> 7\n```"
description: "Given `tensor`, this operation returns a tensor that has the same values\nas `tensor` with shape `shape`.\n\nIf one component of `shape` is the special value -1, the size of that dimension\nis computed so that the total size remains constant. In particular, a `shape`\nof `[-1]` flattens into 1-D. At most one component of `shape` can be -1.\n\nIf `shape` is 1-D or higher, then the operation returns a tensor with shape\n`shape` filled with the values of `tensor`. In this case, the number of elements\nimplied by `shape` must be the same as the number of elements in `tensor`.\n\nFor example:\n\n```\n# tensor \'t\' is [1, 2, 3, 4, 5, 6, 7, 8, 9]\n# tensor \'t\' has shape [9]\nreshape(t, [3, 3]) ==> [[1, 2, 3],\n [4, 5, 6],\n [7, 8, 9]]\n\n# tensor \'t\' is [[[1, 1], [2, 2]],\n# [[3, 3], [4, 4]]]\n# tensor \'t\' has shape [2, 2, 2]\nreshape(t, [2, 4]) ==> [[1, 1, 2, 2],\n [3, 3, 4, 4]]\n\n# tensor \'t\' is [[[1, 1, 1],\n# [2, 2, 2]],\n# [[3, 3, 3],\n# [4, 4, 4]],\n# [[5, 5, 5],\n# [6, 6, 6]]]\n# tensor \'t\' has shape [3, 2, 3]\n# pass \'[-1]\' to flatten \'t\'\nreshape(t, [-1]) ==> [1, 1, 1, 2, 2, 2, 3, 3, 3, 4, 4, 4, 5, 5, 5, 6, 6, 6]\n\n# -1 can also be used to infer the shape\n\n# -1 is inferred to be 9:\nreshape(t, [2, -1]) ==> [[1, 1, 1, 2, 2, 2, 3, 3, 3],\n [4, 4, 4, 5, 5, 5, 6, 6, 6]]\n# -1 is inferred to be 2:\nreshape(t, [-1, 9]) ==> [[1, 1, 1, 2, 2, 2, 3, 3, 3],\n [4, 4, 4, 5, 5, 5, 6, 6, 6]]\n# -1 is inferred to be 3:\nreshape(t, [ 2, -1, 3]) ==> [[[1, 1, 1],\n [2, 2, 2],\n [3, 3, 3]],\n [[4, 4, 4],\n [5, 5, 5],\n [6, 6, 6]]]\n\n# tensor \'t\' is [7]\n# shape `[]` reshapes to a scalar\nreshape(t, []) ==> 7\n```"
}
op {
name: "ResizeArea"
@ -18516,7 +18516,7 @@ op {
}
}
summary: "Reverses specific dimensions of a tensor."
description: "Given a `tensor`, and a `bool` tensor `dims` representing the dimensions\nof `tensor`, this operation reverses each dimension i of `tensor` where\n`dims[i]` is `True`.\n\n`tensor` can have up to 8 dimensions. The number of dimensions\nof `tensor` must equal the number of elements in `dims`. In other words:\n\n`rank(tensor) = size(dims)`\n\nFor example:\n\n```prettyprint\n# tensor \'t\' is [[[[ 0, 1, 2, 3],\n# [ 4, 5, 6, 7],\n# [ 8, 9, 10, 11]],\n# [[12, 13, 14, 15],\n# [16, 17, 18, 19],\n# [20, 21, 22, 23]]]]\n# tensor \'t\' shape is [1, 2, 3, 4]\n\n# \'dims\' is [False, False, False, True]\nreverse(t, dims) ==> [[[[ 3, 2, 1, 0],\n [ 7, 6, 5, 4],\n [ 11, 10, 9, 8]],\n [[15, 14, 13, 12],\n [19, 18, 17, 16],\n [23, 22, 21, 20]]]]\n\n# \'dims\' is [False, True, False, False]\nreverse(t, dims) ==> [[[[12, 13, 14, 15],\n [16, 17, 18, 19],\n [20, 21, 22, 23]\n [[ 0, 1, 2, 3],\n [ 4, 5, 6, 7],\n [ 8, 9, 10, 11]]]]\n\n# \'dims\' is [False, False, True, False]\nreverse(t, dims) ==> [[[[8, 9, 10, 11],\n [4, 5, 6, 7],\n [0, 1, 2, 3]]\n [[20, 21, 22, 23],\n [16, 17, 18, 19],\n [12, 13, 14, 15]]]]\n```"
description: "Given a `tensor`, and a `bool` tensor `dims` representing the dimensions\nof `tensor`, this operation reverses each dimension i of `tensor` where\n`dims[i]` is `True`.\n\n`tensor` can have up to 8 dimensions. The number of dimensions\nof `tensor` must equal the number of elements in `dims`. In other words:\n\n`rank(tensor) = size(dims)`\n\nFor example:\n\n```\n# tensor \'t\' is [[[[ 0, 1, 2, 3],\n# [ 4, 5, 6, 7],\n# [ 8, 9, 10, 11]],\n# [[12, 13, 14, 15],\n# [16, 17, 18, 19],\n# [20, 21, 22, 23]]]]\n# tensor \'t\' shape is [1, 2, 3, 4]\n\n# \'dims\' is [False, False, False, True]\nreverse(t, dims) ==> [[[[ 3, 2, 1, 0],\n [ 7, 6, 5, 4],\n [ 11, 10, 9, 8]],\n [[15, 14, 13, 12],\n [19, 18, 17, 16],\n [23, 22, 21, 20]]]]\n\n# \'dims\' is [False, True, False, False]\nreverse(t, dims) ==> [[[[12, 13, 14, 15],\n [16, 17, 18, 19],\n [20, 21, 22, 23]\n [[ 0, 1, 2, 3],\n [ 4, 5, 6, 7],\n [ 8, 9, 10, 11]]]]\n\n# \'dims\' is [False, False, True, False]\nreverse(t, dims) ==> [[[[8, 9, 10, 11],\n [4, 5, 6, 7],\n [0, 1, 2, 3]]\n [[20, 21, 22, 23],\n [16, 17, 18, 19],\n [12, 13, 14, 15]]]]\n```"
}
op {
name: "ReverseSequence"
@ -18566,7 +18566,7 @@ op {
}
}
summary: "Reverses variable length slices."
description: "This op first slices `input` along the dimension `batch_dim`, and for each\nslice `i`, reverses the first `seq_lengths[i]` elements along\nthe dimension `seq_dim`.\n\nThe elements of `seq_lengths` must obey `seq_lengths[i] <= input.dims[seq_dim]`,\nand `seq_lengths` must be a vector of length `input.dims[batch_dim]`.\n\nThe output slice `i` along dimension `batch_dim` is then given by input\nslice `i`, with the first `seq_lengths[i]` slices along dimension\n`seq_dim` reversed.\n\nFor example:\n\n```prettyprint\n# Given this:\nbatch_dim = 0\nseq_dim = 1\ninput.dims = (4, 8, ...)\nseq_lengths = [7, 2, 3, 5]\n\n# then slices of input are reversed on seq_dim, but only up to seq_lengths:\noutput[0, 0:7, :, ...] = input[0, 7:0:-1, :, ...]\noutput[1, 0:2, :, ...] = input[1, 2:0:-1, :, ...]\noutput[2, 0:3, :, ...] = input[2, 3:0:-1, :, ...]\noutput[3, 0:5, :, ...] = input[3, 5:0:-1, :, ...]\n\n# while entries past seq_lens are copied through:\noutput[0, 7:, :, ...] = input[0, 7:, :, ...]\noutput[1, 2:, :, ...] = input[1, 2:, :, ...]\noutput[2, 3:, :, ...] = input[2, 3:, :, ...]\noutput[3, 2:, :, ...] = input[3, 2:, :, ...]\n```\n\nIn contrast, if:\n\n```prettyprint\n# Given this:\nbatch_dim = 2\nseq_dim = 0\ninput.dims = (8, ?, 4, ...)\nseq_lengths = [7, 2, 3, 5]\n\n# then slices of input are reversed on seq_dim, but only up to seq_lengths:\noutput[0:7, :, 0, :, ...] = input[7:0:-1, :, 0, :, ...]\noutput[0:2, :, 1, :, ...] = input[2:0:-1, :, 1, :, ...]\noutput[0:3, :, 2, :, ...] = input[3:0:-1, :, 2, :, ...]\noutput[0:5, :, 3, :, ...] = input[5:0:-1, :, 3, :, ...]\n\n# while entries past seq_lens are copied through:\noutput[7:, :, 0, :, ...] = input[7:, :, 0, :, ...]\noutput[2:, :, 1, :, ...] = input[2:, :, 1, :, ...]\noutput[3:, :, 2, :, ...] = input[3:, :, 2, :, ...]\noutput[2:, :, 3, :, ...] = input[2:, :, 3, :, ...]\n```"
description: "This op first slices `input` along the dimension `batch_dim`, and for each\nslice `i`, reverses the first `seq_lengths[i]` elements along\nthe dimension `seq_dim`.\n\nThe elements of `seq_lengths` must obey `seq_lengths[i] <= input.dims[seq_dim]`,\nand `seq_lengths` must be a vector of length `input.dims[batch_dim]`.\n\nThe output slice `i` along dimension `batch_dim` is then given by input\nslice `i`, with the first `seq_lengths[i]` slices along dimension\n`seq_dim` reversed.\n\nFor example:\n\n```\n# Given this:\nbatch_dim = 0\nseq_dim = 1\ninput.dims = (4, 8, ...)\nseq_lengths = [7, 2, 3, 5]\n\n# then slices of input are reversed on seq_dim, but only up to seq_lengths:\noutput[0, 0:7, :, ...] = input[0, 7:0:-1, :, ...]\noutput[1, 0:2, :, ...] = input[1, 2:0:-1, :, ...]\noutput[2, 0:3, :, ...] = input[2, 3:0:-1, :, ...]\noutput[3, 0:5, :, ...] = input[3, 5:0:-1, :, ...]\n\n# while entries past seq_lens are copied through:\noutput[0, 7:, :, ...] = input[0, 7:, :, ...]\noutput[1, 2:, :, ...] = input[1, 2:, :, ...]\noutput[2, 3:, :, ...] = input[2, 3:, :, ...]\noutput[3, 2:, :, ...] = input[3, 2:, :, ...]\n```\n\nIn contrast, if:\n\n```\n# Given this:\nbatch_dim = 2\nseq_dim = 0\ninput.dims = (8, ?, 4, ...)\nseq_lengths = [7, 2, 3, 5]\n\n# then slices of input are reversed on seq_dim, but only up to seq_lengths:\noutput[0:7, :, 0, :, ...] = input[7:0:-1, :, 0, :, ...]\noutput[0:2, :, 1, :, ...] = input[2:0:-1, :, 1, :, ...]\noutput[0:3, :, 2, :, ...] = input[3:0:-1, :, 2, :, ...]\noutput[0:5, :, 3, :, ...] = input[5:0:-1, :, 3, :, ...]\n\n# while entries past seq_lens are copied through:\noutput[7:, :, 0, :, ...] = input[7:, :, 0, :, ...]\noutput[2:, :, 1, :, ...] = input[2:, :, 1, :, ...]\noutput[3:, :, 2, :, ...] = input[3:, :, 2, :, ...]\noutput[2:, :, 3, :, ...] = input[2:, :, 3, :, ...]\n```"
}
op {
name: "ReverseV2"
@ -18617,7 +18617,7 @@ op {
}
}
summary: "Reverses specific dimensions of a tensor."
description: "NOTE `tf.reverse` has now changed behavior in preparation for 1.0.\n`tf.reverse_v2` is currently an alias that will be deprecated before TF 1.0.\n\nGiven a `tensor`, and a `int32` tensor `axis` representing the set of\ndimensions of `tensor` to reverse. This operation reverses each dimension\n`i` for which there exists `j` s.t. `axis[j] == i`.\n\n`tensor` can have up to 8 dimensions. The number of dimensions specified\nin `axis` may be 0 or more entries. If an index is specified more than\nonce, a InvalidArgument error is raised.\n\nFor example:\n\n```prettyprint\n# tensor \'t\' is [[[[ 0, 1, 2, 3],\n# [ 4, 5, 6, 7],\n# [ 8, 9, 10, 11]],\n# [[12, 13, 14, 15],\n# [16, 17, 18, 19],\n# [20, 21, 22, 23]]]]\n# tensor \'t\' shape is [1, 2, 3, 4]\n\n# \'dims\' is [3] or \'dims\' is -1\nreverse(t, dims) ==> [[[[ 3, 2, 1, 0],\n [ 7, 6, 5, 4],\n [ 11, 10, 9, 8]],\n [[15, 14, 13, 12],\n [19, 18, 17, 16],\n [23, 22, 21, 20]]]]\n\n# \'dims\' is \'[1]\' (or \'dims\' is \'[-3]\')\nreverse(t, dims) ==> [[[[12, 13, 14, 15],\n [16, 17, 18, 19],\n [20, 21, 22, 23]\n [[ 0, 1, 2, 3],\n [ 4, 5, 6, 7],\n [ 8, 9, 10, 11]]]]\n\n# \'dims\' is \'[2]\' (or \'dims\' is \'[-2]\')\nreverse(t, dims) ==> [[[[8, 9, 10, 11],\n [4, 5, 6, 7],\n [0, 1, 2, 3]]\n [[20, 21, 22, 23],\n [16, 17, 18, 19],\n [12, 13, 14, 15]]]]\n```"
description: "NOTE `tf.reverse` has now changed behavior in preparation for 1.0.\n`tf.reverse_v2` is currently an alias that will be deprecated before TF 1.0.\n\nGiven a `tensor`, and a `int32` tensor `axis` representing the set of\ndimensions of `tensor` to reverse. This operation reverses each dimension\n`i` for which there exists `j` s.t. `axis[j] == i`.\n\n`tensor` can have up to 8 dimensions. The number of dimensions specified\nin `axis` may be 0 or more entries. If an index is specified more than\nonce, a InvalidArgument error is raised.\n\nFor example:\n\n```\n# tensor \'t\' is [[[[ 0, 1, 2, 3],\n# [ 4, 5, 6, 7],\n# [ 8, 9, 10, 11]],\n# [[12, 13, 14, 15],\n# [16, 17, 18, 19],\n# [20, 21, 22, 23]]]]\n# tensor \'t\' shape is [1, 2, 3, 4]\n\n# \'dims\' is [3] or \'dims\' is -1\nreverse(t, dims) ==> [[[[ 3, 2, 1, 0],\n [ 7, 6, 5, 4],\n [ 11, 10, 9, 8]],\n [[15, 14, 13, 12],\n [19, 18, 17, 16],\n [23, 22, 21, 20]]]]\n\n# \'dims\' is \'[1]\' (or \'dims\' is \'[-3]\')\nreverse(t, dims) ==> [[[[12, 13, 14, 15],\n [16, 17, 18, 19],\n [20, 21, 22, 23]\n [[ 0, 1, 2, 3],\n [ 4, 5, 6, 7],\n [ 8, 9, 10, 11]]]]\n\n# \'dims\' is \'[2]\' (or \'dims\' is \'[-2]\')\nreverse(t, dims) ==> [[[[8, 9, 10, 11],\n [4, 5, 6, 7],\n [0, 1, 2, 3]]\n [[20, 21, 22, 23],\n [16, 17, 18, 19],\n [12, 13, 14, 15]]]]\n```"
}
op {
name: "Rint"
@ -19022,7 +19022,7 @@ op {
description: "If True, the addition will be protected by a lock;\notherwise the behavior is undefined, but may exhibit less contention."
}
summary: "Adds sparse updates to a variable reference."
description: "This operation computes\n\n # Scalar indices\n ref[indices, ...] += updates[...]\n\n # Vector indices (for each i)\n ref[indices[i], ...] += updates[i, ...]\n\n # High rank indices (for each i, ..., j)\n ref[indices[i, ..., j], ...] += updates[i, ..., j, ...]\n\nThis operation outputs `ref` after the update is done.\nThis makes it easier to chain operations that need to use the reset value.\n\nDuplicate entries are handled correctly: if multiple `indices` reference\nthe same location, their contributions add.\n\nRequires `updates.shape = indices.shape + ref.shape[1:]`.\n\n<div style=\"width:70%; margin:auto; margin-bottom:10px; margin-top:20px;\">\n<img style=\"width:100%\" src=\"../../images/ScatterAdd.png\" alt>\n</div>"
description: "This operation computes\n\n # Scalar indices\n ref[indices, ...] += updates[...]\n\n # Vector indices (for each i)\n ref[indices[i], ...] += updates[i, ...]\n\n # High rank indices (for each i, ..., j)\n ref[indices[i, ..., j], ...] += updates[i, ..., j, ...]\n\nThis operation outputs `ref` after the update is done.\nThis makes it easier to chain operations that need to use the reset value.\n\nDuplicate entries are handled correctly: if multiple `indices` reference\nthe same location, their contributions add.\n\nRequires `updates.shape = indices.shape + ref.shape[1:]`.\n\n<div style=\"width:70%; margin:auto; margin-bottom:10px; margin-top:20px;\">\n<img style=\"width:100%\" src=\"https://www.tensorflow.org/images/ScatterAdd.png\" alt>\n</div>"
}
op {
name: "ScatterDiv"
@ -19195,7 +19195,7 @@ op {
}
}
summary: "Scatter `updates` into a new (initially zero) tensor according to `indices`."
description: "Creates a new tensor by applying sparse `updates` to individual\nvalues or slices within a zero tensor of the given `shape` according to\nindices. This operator is the inverse of the [tf.gather_nd](#gather_nd)\noperator which extracts values or slices from a given tensor.\n\n**WARNING**: The order in which updates are applied is nondeterministic, so the\noutput will be nondeterministic if `indices` contains duplicates.\n\n`indices` is an integer tensor containing indices into a new tensor of shape\n`shape`. The last dimension of `indices` can be at most the rank of `shape`:\n\n indices.shape[-1] <= shape.rank\n\nThe last dimension of `indices` corresponds to indices into elements\n(if `indices.shape[-1] = shape.rank`) or slices\n(if `indices.shape[-1] < shape.rank`) along dimension `indices.shape[-1]` of\n`shape`. `updates` is a tensor with shape\n\n indices.shape[:-1] + shape[indices.shape[-1]:]\n\nThe simplest form of scatter is to insert individual elements in a tensor by\nindex. For example, say we want to insert 4 scattered elements in a rank-1\ntensor with 8 elements.\n\n<div style=\"width:70%; margin:auto; margin-bottom:10px; margin-top:20px;\">\n<img style=\"width:100%\" src=\"../../images/ScatterNd1.png\" alt>\n</div>\n\nIn Python, this scatter operation would look like this:\n\n```python\n indices = tf.constant([[4], [3], [1], [7]])\n updates = tf.constant([9, 10, 11, 12])\n shape = tf.constant([8])\n scatter = tf.scatter_nd(indices, updates, shape)\n with tf.Session() as sess:\n print(sess.run(scatter))\n```\n\nThe resulting tensor would look like this:\n\n [0, 11, 0, 10, 9, 0, 0, 12]\n\nWe can also, insert entire slices of a higher rank tensor all at once. For\nexample, if we wanted to insert two slices in the first dimension of a\nrank-3 tensor with two matrices of new values.\n\n<div style=\"width:70%; margin:auto; margin-bottom:10px; margin-top:20px;\">\n<img style=\"width:100%\" src=\"../../images/ScatterNd2.png\" alt>\n</div>\n\nIn Python, this scatter operation would look like this:\n\n```python\n indices = tf.constant([[0], [2]])\n updates = tf.constant([[[5, 5, 5, 5], [6, 6, 6, 6],\n [7, 7, 7, 7], [8, 8, 8, 8]],\n [[5, 5, 5, 5], [6, 6, 6, 6],\n [7, 7, 7, 7], [8, 8, 8, 8]]])\n shape = tf.constant([4, 4, 4])\n scatter = tf.scatter_nd(indices, updates, shape)\n with tf.Session() as sess:\n print(sess.run(scatter))\n```\n\nThe resulting tensor would look like this:\n\n [[[5, 5, 5, 5], [6, 6, 6, 6], [7, 7, 7, 7], [8, 8, 8, 8]],\n [[0, 0, 0, 0], [0, 0, 0, 0], [0, 0, 0, 0], [0, 0, 0, 0]],\n [[5, 5, 5, 5], [6, 6, 6, 6], [7, 7, 7, 7], [8, 8, 8, 8]],\n [[0, 0, 0, 0], [0, 0, 0, 0], [0, 0, 0, 0], [0, 0, 0, 0]]]"
description: "Creates a new tensor by applying sparse `updates` to individual\nvalues or slices within a zero tensor of the given `shape` according to\nindices. This operator is the inverse of the [tf.gather_nd](#gather_nd)\noperator which extracts values or slices from a given tensor.\n\n**WARNING**: The order in which updates are applied is nondeterministic, so the\noutput will be nondeterministic if `indices` contains duplicates.\n\n`indices` is an integer tensor containing indices into a new tensor of shape\n`shape`. The last dimension of `indices` can be at most the rank of `shape`:\n\n indices.shape[-1] <= shape.rank\n\nThe last dimension of `indices` corresponds to indices into elements\n(if `indices.shape[-1] = shape.rank`) or slices\n(if `indices.shape[-1] < shape.rank`) along dimension `indices.shape[-1]` of\n`shape`. `updates` is a tensor with shape\n\n indices.shape[:-1] + shape[indices.shape[-1]:]\n\nThe simplest form of scatter is to insert individual elements in a tensor by\nindex. For example, say we want to insert 4 scattered elements in a rank-1\ntensor with 8 elements.\n\n<div style=\"width:70%; margin:auto; margin-bottom:10px; margin-top:20px;\">\n<img style=\"width:100%\" src=\"https://www.tensorflow.org/images/ScatterNd1.png\" alt>\n</div>\n\nIn Python, this scatter operation would look like this:\n\n```python\n indices = tf.constant([[4], [3], [1], [7]])\n updates = tf.constant([9, 10, 11, 12])\n shape = tf.constant([8])\n scatter = tf.scatter_nd(indices, updates, shape)\n with tf.Session() as sess:\n print(sess.run(scatter))\n```\n\nThe resulting tensor would look like this:\n\n [0, 11, 0, 10, 9, 0, 0, 12]\n\nWe can also, insert entire slices of a higher rank tensor all at once. For\nexample, if we wanted to insert two slices in the first dimension of a\nrank-3 tensor with two matrices of new values.\n\n<div style=\"width:70%; margin:auto; margin-bottom:10px; margin-top:20px;\">\n<img style=\"width:100%\" src=\"https://www.tensorflow.org/images/ScatterNd2.png\" alt>\n</div>\n\nIn Python, this scatter operation would look like this:\n\n```python\n indices = tf.constant([[0], [2]])\n updates = tf.constant([[[5, 5, 5, 5], [6, 6, 6, 6],\n [7, 7, 7, 7], [8, 8, 8, 8]],\n [[5, 5, 5, 5], [6, 6, 6, 6],\n [7, 7, 7, 7], [8, 8, 8, 8]]])\n shape = tf.constant([4, 4, 4])\n scatter = tf.scatter_nd(indices, updates, shape)\n with tf.Session() as sess:\n print(sess.run(scatter))\n```\n\nThe resulting tensor would look like this:\n\n [[[5, 5, 5, 5], [6, 6, 6, 6], [7, 7, 7, 7], [8, 8, 8, 8]],\n [[0, 0, 0, 0], [0, 0, 0, 0], [0, 0, 0, 0], [0, 0, 0, 0]],\n [[5, 5, 5, 5], [6, 6, 6, 6], [7, 7, 7, 7], [8, 8, 8, 8]],\n [[0, 0, 0, 0], [0, 0, 0, 0], [0, 0, 0, 0], [0, 0, 0, 0]]]"
}
op {
name: "ScatterNdAdd"
@ -19445,7 +19445,7 @@ op {
description: "If True, the subtraction will be protected by a lock;\notherwise the behavior is undefined, but may exhibit less contention."
}
summary: "Subtracts sparse updates to a variable reference."
description: "```python\n # Scalar indices\n ref[indices, ...] -= updates[...]\n\n # Vector indices (for each i)\n ref[indices[i], ...] -= updates[i, ...]\n\n # High rank indices (for each i, ..., j)\n ref[indices[i, ..., j], ...] -= updates[i, ..., j, ...]\n```\n\nThis operation outputs `ref` after the update is done.\nThis makes it easier to chain operations that need to use the reset value.\n\nDuplicate entries are handled correctly: if multiple `indices` reference\nthe same location, their (negated) contributions add.\n\nRequires `updates.shape = indices.shape + ref.shape[1:]`.\n\n<div style=\"width:70%; margin:auto; margin-bottom:10px; margin-top:20px;\">\n<img style=\"width:100%\" src=\"../../images/ScatterSub.png\" alt>\n</div>"
description: "```python\n # Scalar indices\n ref[indices, ...] -= updates[...]\n\n # Vector indices (for each i)\n ref[indices[i], ...] -= updates[i, ...]\n\n # High rank indices (for each i, ..., j)\n ref[indices[i, ..., j], ...] -= updates[i, ..., j, ...]\n```\n\nThis operation outputs `ref` after the update is done.\nThis makes it easier to chain operations that need to use the reset value.\n\nDuplicate entries are handled correctly: if multiple `indices` reference\nthe same location, their (negated) contributions add.\n\nRequires `updates.shape = indices.shape + ref.shape[1:]`.\n\n<div style=\"width:70%; margin:auto; margin-bottom:10px; margin-top:20px;\">\n<img style=\"width:100%\" src=\"https://www.tensorflow.org/images/ScatterSub.png\" alt>\n</div>"
}
op {
name: "ScatterUpdate"
@ -19494,7 +19494,7 @@ op {
description: "If True, the assignment will be protected by a lock;\notherwise the behavior is undefined, but may exhibit less contention."
}
summary: "Applies sparse updates to a variable reference."
description: "This operation computes\n\n```python\n # Scalar indices\n ref[indices, ...] = updates[...]\n\n # Vector indices (for each i)\n ref[indices[i], ...] = updates[i, ...]\n\n # High rank indices (for each i, ..., j)\n ref[indices[i, ..., j], ...] = updates[i, ..., j, ...]\n```\n\nThis operation outputs `ref` after the update is done.\nThis makes it easier to chain operations that need to use the reset value.\n\nIf values in `ref` is to be updated more than once, because there are\nduplicate entries in `indices`, the order at which the updates happen\nfor each value is undefined.\n\nRequires `updates.shape = indices.shape + ref.shape[1:]`.\n\n<div style=\"width:70%; margin:auto; margin-bottom:10px; margin-top:20px;\">\n<img style=\"width:100%\" src=\"../../images/ScatterUpdate.png\" alt>\n</div>"
description: "This operation computes\n\n```python\n # Scalar indices\n ref[indices, ...] = updates[...]\n\n # Vector indices (for each i)\n ref[indices[i], ...] = updates[i, ...]\n\n # High rank indices (for each i, ..., j)\n ref[indices[i, ..., j], ...] = updates[i, ..., j, ...]\n```\n\nThis operation outputs `ref` after the update is done.\nThis makes it easier to chain operations that need to use the reset value.\n\nIf values in `ref` is to be updated more than once, because there are\nduplicate entries in `indices`, the order at which the updates happen\nfor each value is undefined.\n\nRequires `updates.shape = indices.shape + ref.shape[1:]`.\n\n<div style=\"width:70%; margin:auto; margin-bottom:10px; margin-top:20px;\">\n<img style=\"width:100%\" src=\"https://www.tensorflow.org/images/ScatterUpdate.png\" alt>\n</div>"
}
op {
name: "SdcaFprint"
@ -20146,7 +20146,7 @@ op {
}
}
summary: "Returns the shape of a tensor."
description: "This operation returns a 1-D integer tensor representing the shape of `input`.\n\nFor example:\n\n```prettyprint\n# \'t\' is [[[1, 1, 1], [2, 2, 2]], [[3, 3, 3], [4, 4, 4]]]\nshape(t) ==> [2, 2, 3]\n```"
description: "This operation returns a 1-D integer tensor representing the shape of `input`.\n\nFor example:\n\n```\n# \'t\' is [[[1, 1, 1], [2, 2, 2]], [[3, 3, 3], [4, 4, 4]]]\nshape(t) ==> [2, 2, 3]\n```"
}
op {
name: "ShapeN"
@ -20360,7 +20360,7 @@ op {
}
}
summary: "Returns the size of a tensor."
description: "This operation returns an integer representing the number of elements in\n`input`.\n\nFor example:\n\n```prettyprint\n# \'t\' is [[[1, 1,, 1], [2, 2, 2]], [[3, 3, 3], [4, 4, 4]]]]\nsize(t) ==> 12\n```"
description: "This operation returns an integer representing the number of elements in\n`input`.\n\nFor example:\n\n```\n# \'t\' is [[[1, 1,, 1], [2, 2, 2]], [[3, 3, 3], [4, 4, 4]]]]\nsize(t) ==> 12\n```"
}
op {
name: "Skipgram"
@ -20678,7 +20678,7 @@ op {
}
input_arg {
name: "paddings"
description: "2-D tensor of non-negative integers with shape `[2, 2]`. It specifies\n the padding of the input with zeros across the spatial dimensions as follows:\n\n paddings = [[pad_top, pad_bottom], [pad_left, pad_right]]\n\n The effective spatial dimensions of the zero-padded input tensor will be:\n\n height_pad = pad_top + height + pad_bottom\n width_pad = pad_left + width + pad_right\n\nThe attr `block_size` must be greater than one. It indicates the block size.\n\n * Non-overlapping blocks of size `block_size x block size` in the height and\n width dimensions are rearranged into the batch dimension at each location.\n * The batch of the output tensor is `batch * block_size * block_size`.\n * Both height_pad and width_pad must be divisible by block_size.\n\nThe shape of the output will be:\n\n [batch*block_size*block_size, height_pad/block_size, width_pad/block_size,\n depth]\n\nSome examples:\n\n(1) For the following input of shape `[1, 2, 2, 1]` and block_size of 2:\n\n```prettyprint\nx = [[[[1], [2]], [[3], [4]]]]\n```\n\nThe output tensor has shape `[4, 1, 1, 1]` and value:\n\n```prettyprint\n[[[[1]]], [[[2]]], [[[3]]], [[[4]]]]\n```\n\n(2) For the following input of shape `[1, 2, 2, 3]` and block_size of 2:\n\n```prettyprint\nx = [[[[1, 2, 3], [4, 5, 6]],\n [[7, 8, 9], [10, 11, 12]]]]\n```\n\nThe output tensor has shape `[4, 1, 1, 3]` and value:\n\n```prettyprint\n[[[1, 2, 3]], [[4, 5, 6]], [[7, 8, 9]], [[10, 11, 12]]]\n```\n\n(3) For the following input of shape `[1, 4, 4, 1]` and block_size of 2:\n\n```prettyprint\nx = [[[[1], [2], [3], [4]],\n [[5], [6], [7], [8]],\n [[9], [10], [11], [12]],\n [[13], [14], [15], [16]]]]\n```\n\nThe output tensor has shape `[4, 2, 2, 1]` and value:\n\n```prettyprint\nx = [[[[1], [3]], [[9], [11]]],\n [[[2], [4]], [[10], [12]]],\n [[[5], [7]], [[13], [15]]],\n [[[6], [8]], [[14], [16]]]]\n```\n\n(4) For the following input of shape `[2, 2, 4, 1]` and block_size of 2:\n\n```prettyprint\nx = [[[[1], [2], [3], [4]],\n [[5], [6], [7], [8]]],\n [[[9], [10], [11], [12]],\n [[13], [14], [15], [16]]]]\n```\n\nThe output tensor has shape `[8, 1, 2, 1]` and value:\n\n```prettyprint\nx = [[[[1], [3]]], [[[9], [11]]], [[[2], [4]]], [[[10], [12]]],\n [[[5], [7]]], [[[13], [15]]], [[[6], [8]]], [[[14], [16]]]]\n```\n\nAmong others, this operation is useful for reducing atrous convolution into\nregular convolution."
description: "2-D tensor of non-negative integers with shape `[2, 2]`. It specifies\n the padding of the input with zeros across the spatial dimensions as follows:\n\n paddings = [[pad_top, pad_bottom], [pad_left, pad_right]]\n\n The effective spatial dimensions of the zero-padded input tensor will be:\n\n height_pad = pad_top + height + pad_bottom\n width_pad = pad_left + width + pad_right\n\nThe attr `block_size` must be greater than one. It indicates the block size.\n\n * Non-overlapping blocks of size `block_size x block size` in the height and\n width dimensions are rearranged into the batch dimension at each location.\n * The batch of the output tensor is `batch * block_size * block_size`.\n * Both height_pad and width_pad must be divisible by block_size.\n\nThe shape of the output will be:\n\n [batch*block_size*block_size, height_pad/block_size, width_pad/block_size,\n depth]\n\nSome examples:\n\n(1) For the following input of shape `[1, 2, 2, 1]` and block_size of 2:\n\n```\nx = [[[[1], [2]], [[3], [4]]]]\n```\n\nThe output tensor has shape `[4, 1, 1, 1]` and value:\n\n```\n[[[[1]]], [[[2]]], [[[3]]], [[[4]]]]\n```\n\n(2) For the following input of shape `[1, 2, 2, 3]` and block_size of 2:\n\n```\nx = [[[[1, 2, 3], [4, 5, 6]],\n [[7, 8, 9], [10, 11, 12]]]]\n```\n\nThe output tensor has shape `[4, 1, 1, 3]` and value:\n\n```\n[[[1, 2, 3]], [[4, 5, 6]], [[7, 8, 9]], [[10, 11, 12]]]\n```\n\n(3) For the following input of shape `[1, 4, 4, 1]` and block_size of 2:\n\n```\nx = [[[[1], [2], [3], [4]],\n [[5], [6], [7], [8]],\n [[9], [10], [11], [12]],\n [[13], [14], [15], [16]]]]\n```\n\nThe output tensor has shape `[4, 2, 2, 1]` and value:\n\n```\nx = [[[[1], [3]], [[9], [11]]],\n [[[2], [4]], [[10], [12]]],\n [[[5], [7]], [[13], [15]]],\n [[[6], [8]], [[14], [16]]]]\n```\n\n(4) For the following input of shape `[2, 2, 4, 1]` and block_size of 2:\n\n```\nx = [[[[1], [2], [3], [4]],\n [[5], [6], [7], [8]]],\n [[[9], [10], [11], [12]],\n [[13], [14], [15], [16]]]]\n```\n\nThe output tensor has shape `[8, 1, 2, 1]` and value:\n\n```\nx = [[[[1], [3]]], [[[9], [11]]], [[[2], [4]]], [[[10], [12]]],\n [[[5], [7]]], [[[13], [15]]], [[[6], [8]]], [[[14], [16]]]]\n```\n\nAmong others, this operation is useful for reducing atrous convolution into\nregular convolution."
type_attr: "Tpaddings"
}
output_arg {
@ -20725,7 +20725,7 @@ op {
}
input_arg {
name: "paddings"
description: "2-D with shape `[M, 2]`, all values must be >= 0.\n `paddings[i] = [pad_start, pad_end]` specifies the padding for input dimension\n `i + 1`, which corresponds to spatial dimension `i`. It is required that\n `block_shape[i]` divides `input_shape[i + 1] + pad_start + pad_end`.\n\nThis operation is equivalent to the following steps:\n\n1. Zero-pad the start and end of dimensions `[1, ..., M]` of the\n input according to `paddings` to produce `padded` of shape `padded_shape`.\n\n2. Reshape `padded` to `reshaped_padded` of shape:\n\n [batch] +\n [padded_shape[1] / block_shape[0],\n block_shape[0],\n ...,\n padded_shape[M] / block_shape[M-1],\n block_shape[M-1]] +\n remaining_shape\n\n3. Permute dimensions of `reshaped_padded` to produce\n `permuted_reshaped_padded` of shape:\n\n block_shape +\n [batch] +\n [padded_shape[1] / block_shape[0],\n ...,\n padded_shape[M] / block_shape[M-1]] +\n remaining_shape\n\n4. Reshape `permuted_reshaped_padded` to flatten `block_shape` into the batch\n dimension, producing an output tensor of shape:\n\n [batch * prod(block_shape)] +\n [padded_shape[1] / block_shape[0],\n ...,\n padded_shape[M] / block_shape[M-1]] +\n remaining_shape\n\nSome examples:\n\n(1) For the following input of shape `[1, 2, 2, 1]`, `block_shape = [2, 2]`, and\n `paddings = [[0, 0], [0, 0]]`:\n\n```prettyprint\nx = [[[[1], [2]], [[3], [4]]]]\n```\n\nThe output tensor has shape `[4, 1, 1, 1]` and value:\n\n```prettyprint\n[[[[1]]], [[[2]]], [[[3]]], [[[4]]]]\n```\n\n(2) For the following input of shape `[1, 2, 2, 3]`, `block_shape = [2, 2]`, and\n `paddings = [[0, 0], [0, 0]]`:\n\n```prettyprint\nx = [[[[1, 2, 3], [4, 5, 6]],\n [[7, 8, 9], [10, 11, 12]]]]\n```\n\nThe output tensor has shape `[4, 1, 1, 3]` and value:\n\n```prettyprint\n[[[1, 2, 3]], [[4, 5, 6]], [[7, 8, 9]], [[10, 11, 12]]]\n```\n\n(3) For the following input of shape `[1, 4, 4, 1]`, `block_shape = [2, 2]`, and\n `paddings = [[0, 0], [0, 0]]`:\n\n```prettyprint\nx = [[[[1], [2], [3], [4]],\n [[5], [6], [7], [8]],\n [[9], [10], [11], [12]],\n [[13], [14], [15], [16]]]]\n```\n\nThe output tensor has shape `[4, 2, 2, 1]` and value:\n\n```prettyprint\nx = [[[[1], [3]], [[9], [11]]],\n [[[2], [4]], [[10], [12]]],\n [[[5], [7]], [[13], [15]]],\n [[[6], [8]], [[14], [16]]]]\n```\n\n(4) For the following input of shape `[2, 2, 4, 1]`, block_shape = `[2, 2]`, and\n paddings = `[[0, 0], [2, 0]]`:\n\n```prettyprint\nx = [[[[1], [2], [3], [4]],\n [[5], [6], [7], [8]]],\n [[[9], [10], [11], [12]],\n [[13], [14], [15], [16]]]]\n```\n\nThe output tensor has shape `[8, 1, 3, 1]` and value:\n\n```prettyprint\nx = [[[[0], [1], [3]]], [[[0], [9], [11]]],\n [[[0], [2], [4]]], [[[0], [10], [12]]],\n [[[0], [5], [7]]], [[[0], [13], [15]]],\n [[[0], [6], [8]]], [[[0], [14], [16]]]]\n```\n\nAmong others, this operation is useful for reducing atrous convolution into\nregular convolution."
description: "2-D with shape `[M, 2]`, all values must be >= 0.\n `paddings[i] = [pad_start, pad_end]` specifies the padding for input dimension\n `i + 1`, which corresponds to spatial dimension `i`. It is required that\n `block_shape[i]` divides `input_shape[i + 1] + pad_start + pad_end`.\n\nThis operation is equivalent to the following steps:\n\n1. Zero-pad the start and end of dimensions `[1, ..., M]` of the\n input according to `paddings` to produce `padded` of shape `padded_shape`.\n\n2. Reshape `padded` to `reshaped_padded` of shape:\n\n [batch] +\n [padded_shape[1] / block_shape[0],\n block_shape[0],\n ...,\n padded_shape[M] / block_shape[M-1],\n block_shape[M-1]] +\n remaining_shape\n\n3. Permute dimensions of `reshaped_padded` to produce\n `permuted_reshaped_padded` of shape:\n\n block_shape +\n [batch] +\n [padded_shape[1] / block_shape[0],\n ...,\n padded_shape[M] / block_shape[M-1]] +\n remaining_shape\n\n4. Reshape `permuted_reshaped_padded` to flatten `block_shape` into the batch\n dimension, producing an output tensor of shape:\n\n [batch * prod(block_shape)] +\n [padded_shape[1] / block_shape[0],\n ...,\n padded_shape[M] / block_shape[M-1]] +\n remaining_shape\n\nSome examples:\n\n(1) For the following input of shape `[1, 2, 2, 1]`, `block_shape = [2, 2]`, and\n `paddings = [[0, 0], [0, 0]]`:\n\n```\nx = [[[[1], [2]], [[3], [4]]]]\n```\n\nThe output tensor has shape `[4, 1, 1, 1]` and value:\n\n```\n[[[[1]]], [[[2]]], [[[3]]], [[[4]]]]\n```\n\n(2) For the following input of shape `[1, 2, 2, 3]`, `block_shape = [2, 2]`, and\n `paddings = [[0, 0], [0, 0]]`:\n\n```\nx = [[[[1, 2, 3], [4, 5, 6]],\n [[7, 8, 9], [10, 11, 12]]]]\n```\n\nThe output tensor has shape `[4, 1, 1, 3]` and value:\n\n```\n[[[1, 2, 3]], [[4, 5, 6]], [[7, 8, 9]], [[10, 11, 12]]]\n```\n\n(3) For the following input of shape `[1, 4, 4, 1]`, `block_shape = [2, 2]`, and\n `paddings = [[0, 0], [0, 0]]`:\n\n```\nx = [[[[1], [2], [3], [4]],\n [[5], [6], [7], [8]],\n [[9], [10], [11], [12]],\n [[13], [14], [15], [16]]]]\n```\n\nThe output tensor has shape `[4, 2, 2, 1]` and value:\n\n```\nx = [[[[1], [3]], [[9], [11]]],\n [[[2], [4]], [[10], [12]]],\n [[[5], [7]], [[13], [15]]],\n [[[6], [8]], [[14], [16]]]]\n```\n\n(4) For the following input of shape `[2, 2, 4, 1]`, block_shape = `[2, 2]`, and\n paddings = `[[0, 0], [2, 0]]`:\n\n```\nx = [[[[1], [2], [3], [4]],\n [[5], [6], [7], [8]]],\n [[[9], [10], [11], [12]],\n [[13], [14], [15], [16]]]]\n```\n\nThe output tensor has shape `[8, 1, 3, 1]` and value:\n\n```\nx = [[[[0], [1], [3]]], [[[0], [9], [11]]],\n [[[0], [2], [4]]], [[[0], [10], [12]]],\n [[[0], [5], [7]]], [[[0], [13], [15]]],\n [[[0], [6], [8]]], [[[0], [14], [16]]]]\n```\n\nAmong others, this operation is useful for reducing atrous convolution into\nregular convolution."
type_attr: "Tpaddings"
}
output_arg {
@ -20787,7 +20787,7 @@ op {
minimum: 2
}
summary: "SpaceToDepth for tensors of type T."
description: "Rearranges blocks of spatial data, into depth. More specifically,\nthis op outputs a copy of the input tensor where values from the `height`\nand `width` dimensions are moved to the `depth` dimension.\nThe attr `block_size` indicates the input block size and how the data is moved.\n\n * Non-overlapping blocks of size `block_size x block size` are rearranged\n into depth at each location.\n * The depth of the output tensor is `input_depth * block_size * block_size`.\n * The input tensor\'s height and width must be divisible by block_size.\n\nThat is, assuming the input is in the shape:\n`[batch, height, width, depth]`,\nthe shape of the output will be:\n`[batch, height/block_size, width/block_size, depth*block_size*block_size]`\n\nThis operation requires that the input tensor be of rank 4, and that\n`block_size` be >=1 and a divisor of both the input `height` and `width`.\n\nThis operation is useful for resizing the activations between convolutions\n(but keeping all data), e.g. instead of pooling. It is also useful for training\npurely convolutional models.\n\nFor example, given this input of shape `[1, 2, 2, 1]`, and block_size of 2:\n\n```prettyprint\nx = [[[[1], [2]],\n [[3], [4]]]]\n```\n\nThis operation will output a tensor of shape `[1, 1, 1, 4]`:\n\n```prettyprint\n[[[[1, 2, 3, 4]]]]\n```\n\nHere, the input has a batch of 1 and each batch element has shape `[2, 2, 1]`,\nthe corresponding output will have a single element (i.e. width and height are\nboth 1) and will have a depth of 4 channels (1 * block_size * block_size).\nThe output element shape is `[1, 1, 4]`.\n\nFor an input tensor with larger depth, here of shape `[1, 2, 2, 3]`, e.g.\n\n```prettyprint\nx = [[[[1, 2, 3], [4, 5, 6]],\n [[7, 8, 9], [10, 11, 12]]]]\n```\n\nThis operation, for block_size of 2, will return the following tensor of shape\n`[1, 1, 1, 12]`\n\n```prettyprint\n[[[[1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12]]]]\n```\n\nSimilarly, for the following input of shape `[1 4 4 1]`, and a block size of 2:\n\n```prettyprint\nx = [[[[1], [2], [5], [6]],\n [[3], [4], [7], [8]],\n [[9], [10], [13], [14]],\n [[11], [12], [15], [16]]]]\n```\n\nthe operator will return the following tensor of shape `[1 2 2 4]`:\n\n```prettyprint\nx = [[[[1, 2, 3, 4],\n [5, 6, 7, 8]],\n [[9, 10, 11, 12],\n [13, 14, 15, 16]]]]\n```"
description: "Rearranges blocks of spatial data, into depth. More specifically,\nthis op outputs a copy of the input tensor where values from the `height`\nand `width` dimensions are moved to the `depth` dimension.\nThe attr `block_size` indicates the input block size and how the data is moved.\n\n * Non-overlapping blocks of size `block_size x block size` are rearranged\n into depth at each location.\n * The depth of the output tensor is `input_depth * block_size * block_size`.\n * The input tensor\'s height and width must be divisible by block_size.\n\nThat is, assuming the input is in the shape:\n`[batch, height, width, depth]`,\nthe shape of the output will be:\n`[batch, height/block_size, width/block_size, depth*block_size*block_size]`\n\nThis operation requires that the input tensor be of rank 4, and that\n`block_size` be >=1 and a divisor of both the input `height` and `width`.\n\nThis operation is useful for resizing the activations between convolutions\n(but keeping all data), e.g. instead of pooling. It is also useful for training\npurely convolutional models.\n\nFor example, given this input of shape `[1, 2, 2, 1]`, and block_size of 2:\n\n```\nx = [[[[1], [2]],\n [[3], [4]]]]\n```\n\nThis operation will output a tensor of shape `[1, 1, 1, 4]`:\n\n```\n[[[[1, 2, 3, 4]]]]\n```\n\nHere, the input has a batch of 1 and each batch element has shape `[2, 2, 1]`,\nthe corresponding output will have a single element (i.e. width and height are\nboth 1) and will have a depth of 4 channels (1 * block_size * block_size).\nThe output element shape is `[1, 1, 4]`.\n\nFor an input tensor with larger depth, here of shape `[1, 2, 2, 3]`, e.g.\n\n```\nx = [[[[1, 2, 3], [4, 5, 6]],\n [[7, 8, 9], [10, 11, 12]]]]\n```\n\nThis operation, for block_size of 2, will return the following tensor of shape\n`[1, 1, 1, 12]`\n\n```\n[[[[1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12]]]]\n```\n\nSimilarly, for the following input of shape `[1 4 4 1]`, and a block size of 2:\n\n```\nx = [[[[1], [2], [5], [6]],\n [[3], [4], [7], [8]],\n [[9], [10], [13], [14]],\n [[11], [12], [15], [16]]]]\n```\n\nthe operator will return the following tensor of shape `[1 2 2 4]`:\n\n```\nx = [[[[1, 2, 3, 4],\n [5, 6, 7, 8]],\n [[9, 10, 11, 12],\n [13, 14, 15, 16]]]]\n```"
}
op {
name: "SparseAccumulatorApplyGradient"
@ -23399,7 +23399,7 @@ op {
has_minimum: true
}
summary: "Removes dimensions of size 1 from the shape of a tensor."
description: "Given a tensor `input`, this operation returns a tensor of the same type with\nall dimensions of size 1 removed. If you don\'t want to remove all size 1\ndimensions, you can remove specific size 1 dimensions by specifying\n`squeeze_dims`.\n\nFor example:\n\n```prettyprint\n# \'t\' is a tensor of shape [1, 2, 1, 3, 1, 1]\nshape(squeeze(t)) ==> [2, 3]\n```\n\nOr, to remove specific size 1 dimensions:\n\n```prettyprint\n# \'t\' is a tensor of shape [1, 2, 1, 3, 1, 1]\nshape(squeeze(t, [2, 4])) ==> [1, 2, 3, 1]\n```"
description: "Given a tensor `input`, this operation returns a tensor of the same type with\nall dimensions of size 1 removed. If you don\'t want to remove all size 1\ndimensions, you can remove specific size 1 dimensions by specifying\n`squeeze_dims`.\n\nFor example:\n\n```\n# \'t\' is a tensor of shape [1, 2, 1, 3, 1, 1]\nshape(squeeze(t)) ==> [2, 3]\n```\n\nOr, to remove specific size 1 dimensions:\n\n```\n# \'t\' is a tensor of shape [1, 2, 1, 3, 1, 1]\nshape(squeeze(t, [2, 4])) ==> [1, 2, 3, 1]\n```"
}
op {
name: "Stack"
@ -23761,7 +23761,7 @@ op {
description: "a bitmask where bit `i` implies that the `i`th\nspecification should shrink the dimensionality. begin and end\nmust imply a slice of size 1 in the dimension. For example in\npython one might do `foo[:, 3, :]` which would result in\n`shrink_axis_mask` being 2."
}
summary: "Return a strided slice from `input`."
description: "Note, most python users will want to use the Python `Tensor.__getitem__`\nor `Variable.__getitem__` rather than this op directly.\n\nThe goal of this op is to produce a new tensor with a subset of\nthe elements from the `n` dimensional `input` tensor. The subset is chosen using\na sequence of `m` sparse range specifications encoded into the arguments\nof this function. Note, in some cases\n`m` could be equal to `n`, but this need not be the case. Each\nrange specification entry can be one of the following:\n\n- An ellipsis (...). Ellipses are used to imply zero or more\n dimensions of full-dimension selection and are produced using\n `ellipsis_mask`. For example, `foo[...]` is the identity slice.\n\n- A new axis. This is used to insert a new shape=1 dimension and is\n produced using `new_axis_mask`. For example, `foo[:, ...]` where\n `foo` is shape `(3, 4)` produces a `(1, 3, 4)` tensor.\n\n\n- A range `begin:end:stride`. This is used to specify how much to choose from\n a given dimension. `stride` can be any integer but 0. `begin` is an integer\n which represents the index of the first value to select while `end` represents\n the index of the last value to select. The number of values selected in each\n dimension is `end - begin` if `stride > 0` and `begin - end` if `stride < 0`.\n `begin` and `end` can be negative where `-1` is the last element, `-2` is\n the second to last. `begin_mask` controls whether to replace the explicitly\n given `begin` with an implicit effective value of `0` if `stride > 0` and\n `-1` if `stride < 0`. `end_mask` is analogous but produces the number\n required to create the largest open interval. For example, given a shape\n `(3,)` tensor `foo[:]`, the effective `begin` and `end` are `0` and `3`. Do\n not assume this is equivalent to `foo[0:-1]` which has an effective `begin`\n and `end` of `0` and `2`. Another example is `foo[-2::-1]` which reverses the\n first dimension of a tensor while dropping the last two (in the original\n order elements). For example `foo = [1,2,3,4]; foo[-2::-1]` is `[4,3]`.\n\n- A single index. This is used to keep only elements that have a given\n index. For example (`foo[2, :]` on a shape `(5,6)` tensor produces a\n shape `(6,)` tensor. This is encoded in `begin` and `end` and\n `shrink_axis_mask`.\n\nEach conceptual range specification is encoded in the op\'s argument. This\nencoding is best understand by considering a non-trivial example. In\nparticular,\n`foo[1, 2:4, None, ..., :-3:-1, :]` will be encoded as\n\n```prettyprint\nbegin = [1, 2, x, x, 0, x] # x denotes don\'t care (usually 0)\nend = [2, 4, x, x, -3, x]\nstrides = [1, 1, x, x, -1, 1]\nbegin_mask = 1<<4 | 1 << 5 = 48\nend_mask = 1<<5 = 32\nellipsis_mask = 1<<3 = 8\nnew_axis_mask = 1<<2 4\nshrink_axis_mask = 1<<0\n```\n\nIn this case if `foo.shape` is (5, 5, 5, 5, 5, 5) the final shape of\nthe slice becomes (2, 1, 5, 5, 2, 5).\nLet us walk step by step through each argument specification.\n\n1. The first argument in the example slice is turned into `begin = 1` and\n`end = begin + 1 = 2`. To disambiguate from the original spec `2:4` we\nalso set the appropriate bit in `shrink_axis_mask`.\n\n2. `2:4` is contributes 2, 4, 1 to begin, end, and stride. All masks have\nzero bits contributed.\n\n3. None is a synonym for `tf.newaxis`. This means insert a dimension of size 1\ndimension in the final shape. Dummy values are contributed to begin,\nend and stride, while the new_axis_mask bit is set.\n\n4. `...` grab the full ranges from as many dimensions as needed to\nfully specify a slice for every dimension of the input shape.\n\n5. `:-3:-1` shows the use of negative indices. A negative index `i` associated\nwith a dimension that has shape `s` is converted to a positive index\n`s + i`. So `-1` becomes `s-1` (i.e. the last element). This conversion\nis done internally so begin, end and strides receive x, -3, and -1.\nThe appropriate begin_mask bit is set to indicate the start range is the\nfull range (ignoring the x).\n\n6. `:` indicates that the entire contents of the corresponding dimension\nis selected. This is equivalent to `::` or `0::1`. begin, end, and strides\nreceive 0, 0, and 1, respectively. The appropriate bits in `begin_mask` and\n`end_mask` are also set.\n\n*Requirements*:\n `0 != strides[i] for i in [0, m)`\n `ellipsis_mask must be a power of two (only one ellipsis)`"
description: "Note, most python users will want to use the Python `Tensor.__getitem__`\nor `Variable.__getitem__` rather than this op directly.\n\nThe goal of this op is to produce a new tensor with a subset of\nthe elements from the `n` dimensional `input` tensor. The subset is chosen using\na sequence of `m` sparse range specifications encoded into the arguments\nof this function. Note, in some cases\n`m` could be equal to `n`, but this need not be the case. Each\nrange specification entry can be one of the following:\n\n- An ellipsis (...). Ellipses are used to imply zero or more\n dimensions of full-dimension selection and are produced using\n `ellipsis_mask`. For example, `foo[...]` is the identity slice.\n\n- A new axis. This is used to insert a new shape=1 dimension and is\n produced using `new_axis_mask`. For example, `foo[:, ...]` where\n `foo` is shape `(3, 4)` produces a `(1, 3, 4)` tensor.\n\n\n- A range `begin:end:stride`. This is used to specify how much to choose from\n a given dimension. `stride` can be any integer but 0. `begin` is an integer\n which represents the index of the first value to select while `end` represents\n the index of the last value to select. The number of values selected in each\n dimension is `end - begin` if `stride > 0` and `begin - end` if `stride < 0`.\n `begin` and `end` can be negative where `-1` is the last element, `-2` is\n the second to last. `begin_mask` controls whether to replace the explicitly\n given `begin` with an implicit effective value of `0` if `stride > 0` and\n `-1` if `stride < 0`. `end_mask` is analogous but produces the number\n required to create the largest open interval. For example, given a shape\n `(3,)` tensor `foo[:]`, the effective `begin` and `end` are `0` and `3`. Do\n not assume this is equivalent to `foo[0:-1]` which has an effective `begin`\n and `end` of `0` and `2`. Another example is `foo[-2::-1]` which reverses the\n first dimension of a tensor while dropping the last two (in the original\n order elements). For example `foo = [1,2,3,4]; foo[-2::-1]` is `[4,3]`.\n\n- A single index. This is used to keep only elements that have a given\n index. For example (`foo[2, :]` on a shape `(5,6)` tensor produces a\n shape `(6,)` tensor. This is encoded in `begin` and `end` and\n `shrink_axis_mask`.\n\nEach conceptual range specification is encoded in the op\'s argument. This\nencoding is best understand by considering a non-trivial example. In\nparticular,\n`foo[1, 2:4, None, ..., :-3:-1, :]` will be encoded as\n\n```\nbegin = [1, 2, x, x, 0, x] # x denotes don\'t care (usually 0)\nend = [2, 4, x, x, -3, x]\nstrides = [1, 1, x, x, -1, 1]\nbegin_mask = 1<<4 | 1 << 5 = 48\nend_mask = 1<<5 = 32\nellipsis_mask = 1<<3 = 8\nnew_axis_mask = 1<<2 4\nshrink_axis_mask = 1<<0\n```\n\nIn this case if `foo.shape` is (5, 5, 5, 5, 5, 5) the final shape of\nthe slice becomes (2, 1, 5, 5, 2, 5).\nLet us walk step by step through each argument specification.\n\n1. The first argument in the example slice is turned into `begin = 1` and\n`end = begin + 1 = 2`. To disambiguate from the original spec `2:4` we\nalso set the appropriate bit in `shrink_axis_mask`.\n\n2. `2:4` is contributes 2, 4, 1 to begin, end, and stride. All masks have\nzero bits contributed.\n\n3. None is a synonym for `tf.newaxis`. This means insert a dimension of size 1\ndimension in the final shape. Dummy values are contributed to begin,\nend and stride, while the new_axis_mask bit is set.\n\n4. `...` grab the full ranges from as many dimensions as needed to\nfully specify a slice for every dimension of the input shape.\n\n5. `:-3:-1` shows the use of negative indices. A negative index `i` associated\nwith a dimension that has shape `s` is converted to a positive index\n`s + i`. So `-1` becomes `s-1` (i.e. the last element). This conversion\nis done internally so begin, end and strides receive x, -3, and -1.\nThe appropriate begin_mask bit is set to indicate the start range is the\nfull range (ignoring the x).\n\n6. `:` indicates that the entire contents of the corresponding dimension\nis selected. This is equivalent to `::` or `0::1`. begin, end, and strides\nreceive 0, 0, and 1, respectively. The appropriate bits in `begin_mask` and\n`end_mask` are also set.\n\n*Requirements*:\n `0 != strides[i] for i in [0, m)`\n `ellipsis_mask must be a power of two (only one ellipsis)`"
}
op {
name: "StridedSliceAssign"
@ -26120,7 +26120,7 @@ op {
}
}
summary: "Finds unique elements in a 1-D tensor."
description: "This operation returns a tensor `y` containing all of the unique elements of `x`\nsorted in the same order that they occur in `x`. This operation also returns a\ntensor `idx` the same size as `x` that contains the index of each value of `x`\nin the unique output `y`. In other words:\n\n`y[idx[i]] = x[i] for i in [0, 1,...,rank(x) - 1]`\n\nFor example:\n\n```prettyprint\n# tensor \'x\' is [1, 1, 2, 4, 4, 4, 7, 8, 8]\ny, idx = unique(x)\ny ==> [1, 2, 4, 7, 8]\nidx ==> [0, 0, 1, 2, 2, 2, 3, 4, 4]\n```"
description: "This operation returns a tensor `y` containing all of the unique elements of `x`\nsorted in the same order that they occur in `x`. This operation also returns a\ntensor `idx` the same size as `x` that contains the index of each value of `x`\nin the unique output `y`. In other words:\n\n`y[idx[i]] = x[i] for i in [0, 1,...,rank(x) - 1]`\n\nFor example:\n\n```\n# tensor \'x\' is [1, 1, 2, 4, 4, 4, 7, 8, 8]\ny, idx = unique(x)\ny ==> [1, 2, 4, 7, 8]\nidx ==> [0, 0, 1, 2, 2, 2, 3, 4, 4]\n```"
}
op {
name: "UniqueWithCounts"
@ -26162,7 +26162,7 @@ op {
}
}
summary: "Finds unique elements in a 1-D tensor."
description: "This operation returns a tensor `y` containing all of the unique elements of `x`\nsorted in the same order that they occur in `x`. This operation also returns a\ntensor `idx` the same size as `x` that contains the index of each value of `x`\nin the unique output `y`. Finally, it returns a third tensor `count` that\ncontains the count of each element of `y` in `x`. In other words:\n\n`y[idx[i]] = x[i] for i in [0, 1,...,rank(x) - 1]`\n\nFor example:\n\n```prettyprint\n# tensor \'x\' is [1, 1, 2, 4, 4, 4, 7, 8, 8]\ny, idx, count = unique_with_counts(x)\ny ==> [1, 2, 4, 7, 8]\nidx ==> [0, 0, 1, 2, 2, 2, 3, 4, 4]\ncount ==> [2, 1, 3, 1, 2]\n```"
description: "This operation returns a tensor `y` containing all of the unique elements of `x`\nsorted in the same order that they occur in `x`. This operation also returns a\ntensor `idx` the same size as `x` that contains the index of each value of `x`\nin the unique output `y`. Finally, it returns a third tensor `count` that\ncontains the count of each element of `y` in `x`. In other words:\n\n`y[idx[i]] = x[i] for i in [0, 1,...,rank(x) - 1]`\n\nFor example:\n\n```\n# tensor \'x\' is [1, 1, 2, 4, 4, 4, 7, 8, 8]\ny, idx, count = unique_with_counts(x)\ny ==> [1, 2, 4, 7, 8]\nidx ==> [0, 0, 1, 2, 2, 2, 3, 4, 4]\ncount ==> [2, 1, 3, 1, 2]\n```"
}
op {
name: "Unpack"
@ -26413,7 +26413,7 @@ op {
type: DT_INT64
}
summary: "Returns locations of true values in a boolean tensor."
description: "This operation returns the coordinates of true elements in `input`. The\ncoordinates are returned in a 2-D tensor where the first dimension (rows)\nrepresents the number of true elements, and the second dimension (columns)\nrepresents the coordinates of the true elements. Keep in mind, the shape of\nthe output tensor can vary depending on how many true values there are in\n`input`. Indices are output in row-major order.\n\nFor example:\n\n```prettyprint\n# \'input\' tensor is [[True, False]\n# [True, False]]\n# \'input\' has two true values, so output has two coordinates.\n# \'input\' has rank of 2, so coordinates have two indices.\nwhere(input) ==> [[0, 0],\n [1, 0]]\n\n# `input` tensor is [[[True, False]\n# [True, False]]\n# [[False, True]\n# [False, True]]\n# [[False, False]\n# [False, True]]]\n# \'input\' has 5 true values, so output has 5 coordinates.\n# \'input\' has rank of 3, so coordinates have three indices.\nwhere(input) ==> [[0, 0, 0],\n [0, 1, 0],\n [1, 0, 1],\n [1, 1, 1],\n [2, 1, 1]]\n```"
description: "This operation returns the coordinates of true elements in `input`. The\ncoordinates are returned in a 2-D tensor where the first dimension (rows)\nrepresents the number of true elements, and the second dimension (columns)\nrepresents the coordinates of the true elements. Keep in mind, the shape of\nthe output tensor can vary depending on how many true values there are in\n`input`. Indices are output in row-major order.\n\nFor example:\n\n```\n# \'input\' tensor is [[True, False]\n# [True, False]]\n# \'input\' has two true values, so output has two coordinates.\n# \'input\' has rank of 2, so coordinates have two indices.\nwhere(input) ==> [[0, 0],\n [1, 0]]\n\n# `input` tensor is [[[True, False]\n# [True, False]]\n# [[False, True]\n# [False, True]]\n# [[False, False]\n# [False, True]]]\n# \'input\' has 5 true values, so output has 5 coordinates.\n# \'input\' has rank of 3, so coordinates have three indices.\nwhere(input) ==> [[0, 0, 0],\n [0, 1, 0],\n [1, 0, 1],\n [1, 1, 1],\n [2, 1, 1]]\n```"
}
op {
name: "WholeFileReader"