"comment":"(string, default: tanh)The activation for candidate hidden state, `tanh` by default.",
"comment":"(string, default: tanh)The activation for candidate hidden state, `tanh` by default.",
"generated":0
"generated":0
}]
}]
},{
"type":"lstmp",
"comment":"\nLong-Short Term Memory with recurrent Projection layer (LSTMP) Operator.\n\nLSTMP has a separate projection layer after the LSTM layer, projecting the \noriginal hidden state to a lower-dimensional one, which is proposed to reduce \nthe number of total parameters and furthermore computational complexity for \nthe LSTM, espeacially for the case that the size of output units is relative \nlarge (https://research.google.com/pubs/archive/43905.pdf). \n\nThe formula is as follows:\n\n$$\ni_t = \\sigma(W_{ix}x_{t} + W_{ir}r_{t-1} + W_{ic}c_{t-1} + b_i) \\\\\n\nf_t = \\sigma(W_{fx}x_{t} + W_{fr}r_{t-1} + W_{fc}c_{t-1} + b_f) \\\\\n\n\\tilde{c_t} = act_g(W_{cx}x_t + W_{cr}r_{t-1} + b_c) \\\\\n\no_t = \\sigma(W_{ox}x_{t} + W_{or}r_{t-1} + W_{oc}c_t + b_o) \\\\\n\nc_t = f_t \\odot c_{t-1} + i_t \\odot \\tilde{c_t} \\\\\n\nh_t = o_t \\odot act_h(c_t) \\\\\n\nr_t = \\overline{act_h}(W_{rh}h_t)\n$$\n\nwhere the W terms denote weight matrices (e.g. $W_{xi}$ is the matrix\nof weights from the input gate to the input), $W_{ic}, W_{fc}, W_{oc}$\nare diagonal weight matrices for peephole connections. In our implementation,\nwe use vectors to reprenset these diagonal weight matrices. The b terms\ndenote bias vectors ($b_i$ is the input gate bias vector), $\\sigma$\nis the activation, such as logistic sigmoid function, and\n$i, f, o$ and $c$ are the input gate, forget gate, output gate,\nand cell activation vectors, respectively, all of which have the same size as\nthe cell output activation vector $h$. Here $h$ is usually called the hidden \nstate and $r$ denotes its recurrent projection. And $\\tilde{c_t}$ is also \ncalled the candidate hidden state, whose computation is based on the current \ninput and previous hidden state.\n\nThe $\\odot$ is the element-wise product of the vectors. $act_g$ and $act_h$\nare the cell input and cell output activation functions and `tanh` is usually\nused for them. $\\overline{act_h}$ is the activation function for the \nprojection output, usually using `identity` or same as $act_h$.\n\nNote that these $W_{xi}x_{t}, W_{xf}x_{t}, W_{xc}x_{t}, W_{xo}x_{t}$\noperations on the input $x_{t}$ are NOT included in this operator.\nUsers can choose to use fully-connected operator before LSTMP operator.\n\n",
"inputs":[
{
"name":"Input",
"comment":"(LoDTensor) the input for sequence data, which supports variable-time length input sequence. The underlying tensor in this LoDTensor is a matrix with shape (T X 4D), where T is the total time steps in this mini-batch, D is the hidden size.",
"duplicable":0,
"intermediate":0
},{
"name":"H0",
"comment":"(Tensor, optional) the initial hidden state is an optional input. This is a tensor with shape (N x D), where N is the batch size and D is the hidden size.",
"duplicable":0,
"intermediate":0
},{
"name":"C0",
"comment":"(Tensor, optional) the initial cell state is an optional input. This is a tensor with shape (N x D), where N is the batch size. `C0` should not be null if `H0` provided.",
"duplicable":0,
"intermediate":0
},{
"name":"Weight",
"comment":"(Tensor) the learnable hidden-hidden weights. - The shape is (P x 4D), where P is the projection layer size and D is the hidden size. - Weight = {W_cr, W_ir, W_fr, W_or}",
"duplicable":0,
"intermediate":0
},{
"name":"ProjWeight",
"comment":"(Tensor) the learnable weight of the projection layer. - The shape is (D x P), where P is the recurrent projection layer size and D is the hidden size. - ProjWeight = {W_rh}",
"duplicable":0,
"intermediate":0
},{
"name":"Bias",
"comment":"(Tensor) the learnable biases, which contains two parts: input-hidden biases and peephole connections weights if setting `use_peepholes` to `True`. 1. `use_peepholes = False` - The shape is (1 x 4D). - Bias = {b_c, b_i, b_f, b_o}.2. `use_peepholes = True` - The shape is (1 x 7D). - Bias = {b_c, b_i, b_f, b_o, W_ic, W_fc, W_oc}.",
"duplicable":0,
"intermediate":0
}],
"outputs":[
{
"name":"Projection",
"comment":"(LoDTensor) the projection of the hidden state of LSTMP operator. The shape is (T x P), and LoD is the same with the `Input`.",
"duplicable":0,
"intermediate":0
},{
"name":"Cell",
"comment":"(LoDTensor) the cell state of LSTMP operator. The shape is (T x D), and lod is the same with the `Input`.",
"duplicable":0,
"intermediate":0
},{
"name":"BatchGate",
"comment":"(LoDTensor) This LoDTensor contains input gate, forget gate and output gate after the activations. This LoDTensor has the same shape as the reorganized input, which is also be called batch input. The LoD size is 2. The first-level LoD is the batch offsets and the second contains the indices, which denotes the position of reorganized sequence in the raw input.",
"duplicable":0,
"intermediate":1
},{
"name":"BatchCellPreAct",
"comment":"(LoDTensor) the pre-activation cell state reorganized in batch. This LoDTensor is obtained in the forward and used in the backward.",
"duplicable":0,
"intermediate":1
},{
"name":"BatchHidden",
"comment":"(LoDTensor) the hidden state reorganized in batch. This LoDTensor is obtained in the forward and used in the backward.",
"duplicable":0,
"intermediate":1
},{
"name":"OrderedP0",
"comment":"(Tensor) the projection of the initial hidden state H0. This is a tensor with shape (N x P), where N is the batch size and P is the hidden size.",
"duplicable":0,
"intermediate":1
}],
"attrs":[
{
"name":"use_peepholes",
"type":"bool",
"comment":"(bool, defalut: True) whether to enable diagonal/peephole connections.",
"generated":0
},{
"name":"is_reverse",
"type":"bool",
"comment":"(bool, defalut: False) whether to compute reversed LSTMP.",
"generated":0
},{
"name":"gate_activation",
"type":"string",
"comment":"(string, default: sigmoid)The activation for input gate, forget gate and output gate, `sigmoid` by default.",
"generated":0
},{
"name":"cell_activation",
"type":"string",
"comment":"(string, default: tanh)The activation for cell output, `tanh` by defalut.",
"generated":0
},{
"name":"candidate_activation",
"type":"string",
"comment":"(string, default: tanh)The activation for candidate hidden state, `tanh` by default.",
"generated":0
},{
"name":"proj_activation",
"type":"string",
"comment":"(string, default: tanh)The activation for projection output, `tanh` by defalut.",
"generated":0
}]
},{
},{
"type":"warpctc",
"type":"warpctc",
"comment":"\nAn operator integrating the open-source\n[warp-ctc](https://github.com/baidu-research/warp-ctc) library, which is used in\n[Deep Speech 2: End-toEnd Speech Recognition in English and Mandarin](\nhttps://arxiv.org/pdf/1512.02595v1.pdf),\nto compute Connectionist Temporal Classification (CTC) loss.\nIt can be aliased as softmax with ctc, since a native softmax activation is\ninterated to the warp-ctc library, to to normlize values for each row of the\ninput tensor.\n\nMore detail of CTC loss can be found by refering to\n[Connectionist Temporal Classification: Labelling Unsegmented Sequence Data with\nRecurrent Neural Networks](\nhttp://machinelearning.wustl.edu/mlpapers/paper_files/icml2006_GravesFGS06.pdf).\n",
"comment":"\nAn operator integrating the open-source\n[warp-ctc](https://github.com/baidu-research/warp-ctc) library, which is used in\n[Deep Speech 2: End-toEnd Speech Recognition in English and Mandarin](\nhttps://arxiv.org/pdf/1512.02595v1.pdf),\nto compute Connectionist Temporal Classification (CTC) loss.\nIt can be aliased as softmax with ctc, since a native softmax activation is\ninterated to the warp-ctc library, to to normlize values for each row of the\ninput tensor.\n\nMore detail of CTC loss can be found by refering to\n[Connectionist Temporal Classification: Labelling Unsegmented Sequence Data with\nRecurrent Neural Networks](\nhttp://machinelearning.wustl.edu/mlpapers/paper_files/icml2006_GravesFGS06.pdf).\n",
...
@@ -1808,47 +1709,6 @@
...
@@ -1808,47 +1709,6 @@
"intermediate":0
"intermediate":0
}],
}],
"attrs":[]
"attrs":[]
},{
"type":"logical_and",
"comment":"logical_and Operator\n\nIt operates element-wise on X and Y, and returns the Out. X, Y and Out are N-dim boolean tensors.\nEach element of Out is calculated by $$Out = X \\&\\& Y$$\n",
"inputs":[
{
"name":"X",
"comment":"(LoDTensor) Left hand operand of logical_and operator",
"duplicable":0,
"intermediate":0
},{
"name":"Y",
"comment":"(LoDTensor) Right hand operand of logical_and operator",
"duplicable":0,
"intermediate":0
}],
"outputs":[
{
"name":"Out",
"comment":"(LoDTensor) n-dim bool tensor. Each element is $$Out = X \\&\\& Y$$",
"duplicable":0,
"intermediate":0
}],
"attrs":[]
},{
"type":"logical_not",
"comment":"logical_not Operator\n\nIt operates element-wise on X, and returns the Out. X and Out are N-dim boolean tensors.\nEach element of Out is calculated by $$Out = !X$$\n",
"inputs":[
{
"name":"X",
"comment":"(LoDTensor) Operand of logical_not operator",
"duplicable":0,
"intermediate":0
}],
"outputs":[
{
"name":"Out",
"comment":"(LoDTensor) n-dim bool tensor. Each element is $$Out = !X$$",
"duplicable":0,
"intermediate":0
}],
"attrs":[]
},{
},{
"type":"max_sequence_len",
"type":"max_sequence_len",
"comment":"Calculate the max sequence length through lod_rank_table.",
"comment":"Calculate the max sequence length through lod_rank_table.",
...
@@ -2421,30 +2281,6 @@
...
@@ -2421,30 +2281,6 @@
"comment":"(int, default -1). The start dimension index for broadcasting Y onto X.",
"comment":"(int, default -1). The start dimension index for broadcasting Y onto X.",
"generated":0
"generated":0
}]
}]
},{
"type":"rnn_memory_helper",
"comment":"",
"inputs":[
{
"name":"X",
"comment":"",
"duplicable":0,
"intermediate":0
}],
"outputs":[
{
"name":"Out",
"comment":"",
"duplicable":0,
"intermediate":0
}],
"attrs":[
{
"name":"dtype",
"type":"int",
"comment":"(int, default 5 (FP32)) Output data type",
"generated":0
}]
},{
},{
"type":"smooth_l1_loss",
"type":"smooth_l1_loss",
"comment":"\nSmooth L1 Loss Operator.\n\nThis operator computes the smooth l1 loss for X and Y.\nThe operator takes the first dimension of X and Y as batch size.\nFor each instance, it computes the smooth l1 loss element by element first\nand then sums all the losses. So the shape of Out is [batch_size, 1].\n\nThe equation is:\n$$\nOut_{\\sigma}(X, Y)_i = \\begin{cases}\n0.5 * (\\sigma * (X_i - Y_i)) ^ 2\n\\quad |X_i - Y_i| \\lt \\frac{1} {{\\sigma} ^ 2} \\\\\n\\frac{|X_i - Y_i| - 0.5}{{\\sigma}^2},\n\\quad otherwise\n\\end{cases}\n$$\n\nIn the above equation, $Out_{\\sigma}(X, Y)_i$, $X_i$ and $Y_i$ represent the ith\nelement of Out, X and Y.\n\n",
"comment":"\nSmooth L1 Loss Operator.\n\nThis operator computes the smooth l1 loss for X and Y.\nThe operator takes the first dimension of X and Y as batch size.\nFor each instance, it computes the smooth l1 loss element by element first\nand then sums all the losses. So the shape of Out is [batch_size, 1].\n\nThe equation is:\n$$\nOut_{\\sigma}(X, Y)_i = \\begin{cases}\n0.5 * (\\sigma * (X_i - Y_i)) ^ 2\n\\quad |X_i - Y_i| \\lt \\frac{1} {{\\sigma} ^ 2} \\\\\n\\frac{|X_i - Y_i| - 0.5}{{\\sigma}^2},\n\\quad otherwise\n\\end{cases}\n$$\n\nIn the above equation, $Out_{\\sigma}(X, Y)_i$, $X_i$ and $Y_i$ represent the ith\nelement of Out, X and Y.\n\n",
...
@@ -2489,6 +2325,58 @@
...
@@ -2489,6 +2325,58 @@
"comment":"Hyper parameter of smooth l1 loss op.A float scalar with default value 3.0.",
"comment":"Hyper parameter of smooth l1 loss op.A float scalar with default value 3.0.",
"generated":0
"generated":0
}]
}]
},{
"type":"reorder_lod_tensor_by_rank",
"comment":"ReorderLoDTensorByRankTable operator.\n\nInput(X) is a batch of sequences. Input(RankTable) stores new orders of the\ninput sequence batch. The reorder_lod_tensor_by_rank operator reorders the\nInput(X) according to the information provided by Input(RankTable).\n\nFor example:\n\nIf the indices stored in the Input(RankTable) are [3, 0, 2, 1], the\nInput(X) will be reordered that the fourth sequence in Input(X) will become the\nfirst one, and then followed by the original first, third, and the second one.\n\nThis is:\nX = [Seq0, Seq1, Seq2, Seq3]. The indices in RankTable are [3, 0, 2, 1].\nOut = [Seq3, Seq0, Seq2, Seq1] with a new LoD information.\n\nIf the LoD information of Input(X) is empty, this means Input(X) is not sequence\ndata. This is also identical to a batch of sequences where each sequence has a\nfixed length 1. In this case, the reorder_lod_tensor_by_rank operator reorders\neach slice of Input(X) along the first axis according to Input(RankTable).\n\nThis is:\nX = [Slice0, Slice1, Slice2, Slice3] and its LoD information is empty. The\nindices in RankTable are [3, 0, 2, 1].\nOut = [Slice3, Slice0, Slice2, Slice1] with no LoD information is appended.\n\nNOTE: This operator sorts Input(X) according to a given LoDRankTable which does\nnot need to be calculated according to Input(X). It can be calculated according\nto another different sequence, and then this operator sorts Input(X) according\nto the given LoDRankTable.\n\n",
"inputs":[
{
"name":"X",
"comment":"(LoDTensor), the input lod tensor to be reordered according to Input(RankTable).",
"duplicable":0,
"intermediate":0
},{
"name":"RankTable",
"comment":"(LoDRankTable), the rank table according to which Input(X) is reordered.",
"duplicable":0,
"intermediate":0
}],
"outputs":[
{
"name":"Out",
"comment":"(LoDTensor), the reordered lod tensor.",
"duplicable":0,
"intermediate":0
}],
"attrs":[]
},{
"type":"pad",
"comment":"\nPad Operator.\n\nPad input into output, as specified by paddings and pad_value. \nThe input should be a k-D tensor(k > 0 and k < 7). As an example:\n\nGiven:\n\nX = [[1, 2],\n [3, 4]],\n\npaddings = [0, 1, 1, 2],\n\nand\n\npad_value = 0,\n\nwe have:\n\nOut = [[0, 1, 2, 0, 0]\n [0, 3, 4, 0, 0]\n [0, 0, 0, 0, 0]]\n\n",
"inputs":[
{
"name":"X",
"comment":"The input of pad op. The input should be a k-D tensor(k > 0 and k < 7)",
"duplicable":0,
"intermediate":0
}],
"outputs":[
{
"name":"Out",
"comment":"The output of pad op. A tensor with the same shape as X.",
"duplicable":0,
"intermediate":0
}],
"attrs":[
{
"name":"paddings",
"type":"int array",
"comment":"(vector<int>) A list<int> to describe the padding rules for each dimension. For 2-D image tensor, paddings=[0, 1, 2, 3] means padding 0 row to top, 1 row to bottom, 2 columns to left and 3 columns to right. Size of paddings should be equal to 2 * dimension size of the input tensor.",
"generated":0
},{
"name":"pad_value",
"type":"float",
"comment":"(float, default 0.0) The value to fill the padded areas.",
"generated":0
}]
},{
},{
"type":"lstm_unit",
"type":"lstm_unit",
"comment":"\nLstm Unit Operator\n\nEquation:\n\n$$\ni, f, o, j = split(X) \\\\\nC = C_{prev} * sigm(f + forget\\_bias) + sigm(i) * tanh(j) \\\\\nH = C * sigm(o)\n$$\n\n",
"comment":"\nLstm Unit Operator\n\nEquation:\n\n$$\ni, f, o, j = split(X) \\\\\nC = C_{prev} * sigm(f + forget\\_bias) + sigm(i) * tanh(j) \\\\\nH = C * sigm(o)\n$$\n\n",
...
@@ -2652,98 +2540,46 @@
...
@@ -2652,98 +2540,46 @@
"generated":0
"generated":0
}]
}]
},{
},{
"type":"pad",
"type":"split_selected_rows",
"comment":"\nPad Operator.\n\nPad input into output, as specified by paddings and pad_value. \nThe input should be a k-D tensor(k > 0 and k < 7). As an example:\n\nGiven:\n\nX = [[1, 2],\n [3, 4]],\n\npaddings = [0, 1, 1, 2],\n\nand\n\npad_value = 0,\n\nwe have:\n\nOut = [[0, 1, 2, 0, 0]\n [0, 3, 4, 0, 0]\n [0, 0, 0, 0, 0]]\n\n",
"comment":"\nSplit a SelectedRows with a specified rows section.\nheight_sections is only needed when need to split the dims of the original tensor.\n\nExample:\n Input:\n X.rows = {7, 5}\n X.height = 12\n Attr:\n height_sections = {4, 8}\n Out:\n out0.rows = {}\n out0.height = 4\n\n out1.rows = {5, 7}\n out2.height = 8\n\n",
"inputs":[
"inputs":[
{
{
"name":"X",
"name":"X",
"comment":"The input of pad op. The input should be a k-D tensor(k > 0 and k < 7)",
"comment":"The input SelectedRows.",
"duplicable":0,
"duplicable":0,
"intermediate":0
"intermediate":0
}],
}],
"outputs":[
"outputs":[
{
{
"name":"Out",
"name":"Out",
"comment":"The output of pad op. A tensor with the same shape as X.",
"comment":"The outputs of input SelectedRows.",
"duplicable":0,
"duplicable":1,
"intermediate":0
"intermediate":0
}],
}],
"attrs":[
"attrs":[
{
{
"name":"paddings",
"name":"height_sections",
"type":"int array",
"type":"int array",
"comment":"(vector<int>) A list<int> to describe the padding rules for each dimension. For 2-D image tensor, paddings=[0, 1, 2, 3] means padding 0 row to top, 1 row to bottom, 2 columns to left and 3 columns to right. Size of paddings should be equal to 2 * dimension size of the input tensor.",
"comment":"Height for each output SelectedRows.",
"generated":0
},{
"name":"pad_value",
"type":"float",
"comment":"(float, default 0.0) The value to fill the padded areas.",
"generated":0
"generated":0
}]
}]
},{
},{
"type":"reorder_lod_tensor_by_rank",
"type":"adam",
"comment":"ReorderLoDTensorByRankTable operator.\n\nInput(X) is a batch of sequences. Input(RankTable) stores new orders of the\ninput sequence batch. The reorder_lod_tensor_by_rank operator reorders the\nInput(X) according to the information provided by Input(RankTable).\n\nFor example:\n\nIf the indices stored in the Input(RankTable) are [3, 0, 2, 1], the\nInput(X) will be reordered that the fourth sequence in Input(X) will become the\nfirst one, and then followed by the original first, third, and the second one.\n\nThis is:\nX = [Seq0, Seq1, Seq2, Seq3]. The indices in RankTable are [3, 0, 2, 1].\nOut = [Seq3, Seq0, Seq2, Seq1] with a new LoD information.\n\nIf the LoD information of Input(X) is empty, this means Input(X) is not sequence\ndata. This is also identical to a batch of sequences where each sequence has a\nfixed length 1. In this case, the reorder_lod_tensor_by_rank operator reorders\neach slice of Input(X) along the first axis according to Input(RankTable).\n\nThis is:\nX = [Slice0, Slice1, Slice2, Slice3] and its LoD information is empty. The\nindices in RankTable are [3, 0, 2, 1].\nOut = [Slice3, Slice0, Slice2, Slice1] with no LoD information is appended.\n\nNOTE: This operator sorts Input(X) according to a given LoDRankTable which does\nnot need to be calculated according to Input(X). It can be calculated according\nto another different sequence, and then this operator sorts Input(X) according\nto the given LoDRankTable.\n\n",
"comment":"\nAdam Optimizer.\n\nThis implements the Adam optimizer from Section 2 of the Adam\npaper : https://arxiv.org/abs/1412.6980.\nAdam is a first-order gradient-based optimization method based on\nadaptive estimates of lower-order moments.\n\nAdam updates:\n\n$$\nmoment\\_1\\_out = \\beta_1 * moment\\_1 + (1 - \\beta_1) * grad \\\\\nmoment\\_2_\\out = \\beta_2 * moment\\_2 + (1 - \\beta_2) * grad * grad \\\\\nlearning\\_rate = learning\\_rate *\n\\frac{\\sqrt{1 - \\beta_{2\\_pow}}}{1 - \\beta_{1\\_pow}} \\\\\nparam\\_out = param - learning\\_rate * \\frac{moment\\_1}{\\sqrt{moment\\_2} + \\epsilon}\n$$\n\n",
"inputs":[
"inputs":[
{
{
"name":"X",
"name":"Param",
"comment":"(LoDTensor), the input lod tensor to be reordered according to Input(RankTable).",
"comment":"(Tensor) Input parameter",
"duplicable":0,
"duplicable":0,
"intermediate":0
"intermediate":0
},{
},{
"name":"RankTable",
"name":"Grad",
"comment":"(LoDRankTable), the rank table according to which Input(X) is reordered.",
"comment":"(Tensor) Input gradient",
"duplicable":0,
"duplicable":0,
"intermediate":0
"intermediate":0
}],
},{
"outputs":[
"name":"LearningRate",
{
"comment":"(Tensor) Learning rate",
"name":"Out",
"comment":"(LoDTensor), the reordered lod tensor.",
"duplicable":0,
"intermediate":0
}],
"attrs":[]
},{
"type":"split_selected_rows",
"comment":"\nSplit a SelectedRows with a specified rows section.\nheight_sections is only needed when need to split the dims of the original tensor.\n\nExample:\n Input:\n X.rows = {7, 5}\n X.height = 12\n Attr:\n height_sections = {4, 8}\n Out:\n out0.rows = {}\n out0.height = 4\n\n out1.rows = {5, 7}\n out2.height = 8\n\n",
"inputs":[
{
"name":"X",
"comment":"The input SelectedRows.",
"duplicable":0,
"intermediate":0
}],
"outputs":[
{
"name":"Out",
"comment":"The outputs of input SelectedRows.",
"duplicable":1,
"intermediate":0
}],
"attrs":[
{
"name":"height_sections",
"type":"int array",
"comment":"Height for each output SelectedRows.",
"generated":0
}]
},{
"type":"adam",
"comment":"\nAdam Optimizer.\n\nThis implements the Adam optimizer from Section 2 of the Adam\npaper : https://arxiv.org/abs/1412.6980.\nAdam is a first-order gradient-based optimization method based on\nadaptive estimates of lower-order moments.\n\nAdam updates:\n\n$$\nmoment\\_1\\_out = \\beta_1 * moment\\_1 + (1 - \\beta_1) * grad \\\\\nmoment\\_2_\\out = \\beta_2 * moment\\_2 + (1 - \\beta_2) * grad * grad \\\\\nlearning\\_rate = learning\\_rate *\n\\frac{\\sqrt{1 - \\beta_{2\\_pow}}}{1 - \\beta_{1\\_pow}} \\\\\nparam\\_out = param - learning\\_rate * \\frac{moment\\_1}{\\sqrt{moment\\_2} + \\epsilon}\n$$\n\n",
"comment":"\nLong-Short Term Memory with recurrent Projection layer (LSTMP) Operator.\n\nLSTMP has a separate projection layer after the LSTM layer, projecting the \noriginal hidden state to a lower-dimensional one, which is proposed to reduce \nthe number of total parameters and furthermore computational complexity for \nthe LSTM, espeacially for the case that the size of output units is relative \nlarge (https://research.google.com/pubs/archive/43905.pdf). \n\nThe formula is as follows:\n\n$$\ni_t = \\sigma(W_{ix}x_{t} + W_{ir}r_{t-1} + W_{ic}c_{t-1} + b_i) \\\\\n\nf_t = \\sigma(W_{fx}x_{t} + W_{fr}r_{t-1} + W_{fc}c_{t-1} + b_f) \\\\\n\n\\tilde{c_t} = act_g(W_{cx}x_t + W_{cr}r_{t-1} + b_c) \\\\\n\no_t = \\sigma(W_{ox}x_{t} + W_{or}r_{t-1} + W_{oc}c_t + b_o) \\\\\n\nc_t = f_t \\odot c_{t-1} + i_t \\odot \\tilde{c_t} \\\\\n\nh_t = o_t \\odot act_h(c_t) \\\\\n\nr_t = \\overline{act_h}(W_{rh}h_t)\n$$\n\nwhere the W terms denote weight matrices (e.g. $W_{xi}$ is the matrix\nof weights from the input gate to the input), $W_{ic}, W_{fc}, W_{oc}$\nare diagonal weight matrices for peephole connections. In our implementation,\nwe use vectors to reprenset these diagonal weight matrices. The b terms\ndenote bias vectors ($b_i$ is the input gate bias vector), $\\sigma$\nis the activation, such as logistic sigmoid function, and\n$i, f, o$ and $c$ are the input gate, forget gate, output gate,\nand cell activation vectors, respectively, all of which have the same size as\nthe cell output activation vector $h$. Here $h$ is usually called the hidden \nstate and $r$ denotes its recurrent projection. And $\\tilde{c_t}$ is also \ncalled the candidate hidden state, whose computation is based on the current \ninput and previous hidden state.\n\nThe $\\odot$ is the element-wise product of the vectors. $act_g$ and $act_h$\nare the cell input and cell output activation functions and `tanh` is usually\nused for them. $\\overline{act_h}$ is the activation function for the \nprojection output, usually using `identity` or same as $act_h$.\n\nNote that these $W_{xi}x_{t}, W_{xf}x_{t}, W_{xc}x_{t}, W_{xo}x_{t}$\noperations on the input $x_{t}$ are NOT included in this operator.\nUsers can choose to use fully-connected operator before LSTMP operator.\n\n",
"inputs":[
"inputs":[
{
{
"name":"X",
"name":"Input",
"comment":"Input of Softplus operator",
"comment":"(LoDTensor) the input for sequence data, which supports variable-time length input sequence. The underlying tensor in this LoDTensor is a matrix with shape (T X 4D), where T is the total time steps in this mini-batch, D is the hidden size.",
"duplicable":0,
"duplicable":0,
"intermediate":0
"intermediate":0
}],
},{
"outputs":[
"name":"H0",
{
"comment":"(Tensor, optional) the initial hidden state is an optional input. This is a tensor with shape (N x D), where N is the batch size and D is the hidden size.",
"name":"Out",
"duplicable":0,
"comment":"Output of Softplus operator",
"intermediate":0
},{
"name":"C0",
"comment":"(Tensor, optional) the initial cell state is an optional input. This is a tensor with shape (N x D), where N is the batch size. `C0` should not be null if `H0` provided.",
"duplicable":0,
"intermediate":0
},{
"name":"Weight",
"comment":"(Tensor) the learnable hidden-hidden weights. - The shape is (P x 4D), where P is the projection layer size and D is the hidden size. - Weight = {W_cr, W_ir, W_fr, W_or}",
"duplicable":0,
"intermediate":0
},{
"name":"ProjWeight",
"comment":"(Tensor) the learnable weight of the projection layer. - The shape is (D x P), where P is the recurrent projection layer size and D is the hidden size. - ProjWeight = {W_rh}",
"duplicable":0,
"intermediate":0
},{
"name":"Bias",
"comment":"(Tensor) the learnable biases, which contains two parts: input-hidden biases and peephole connections weights if setting `use_peepholes` to `True`. 1. `use_peepholes = False` - The shape is (1 x 4D). - Bias = {b_c, b_i, b_f, b_o}.2. `use_peepholes = True` - The shape is (1 x 7D). - Bias = {b_c, b_i, b_f, b_o, W_ic, W_fc, W_oc}.",
"duplicable":0,
"duplicable":0,
"intermediate":0
"intermediate":0
}],
}],
"attrs":[]
},{
"type":"get_places",
"comment":"\nReturns a list of places based on flags. The list will be used for parallel\nexecution.\n",
"inputs":[],
"outputs":[
"outputs":[
{
{
"name":"Out",
"name":"Projection",
"comment":"vector of Place",
"comment":"(LoDTensor) the projection of the hidden state of LSTMP operator. The shape is (T x P), and LoD is the same with the `Input`.",
"duplicable":0,
"intermediate":0
},{
"name":"Cell",
"comment":"(LoDTensor) the cell state of LSTMP operator. The shape is (T x D), and lod is the same with the `Input`.",
"duplicable":0,
"duplicable":0,
"intermediate":0
"intermediate":0
},{
"name":"BatchGate",
"comment":"(LoDTensor) This LoDTensor contains input gate, forget gate and output gate after the activations. This LoDTensor has the same shape as the reorganized input, which is also be called batch input. The LoD size is 2. The first-level LoD is the batch offsets and the second contains the indices, which denotes the position of reorganized sequence in the raw input.",
"duplicable":0,
"intermediate":1
},{
"name":"BatchCellPreAct",
"comment":"(LoDTensor) the pre-activation cell state reorganized in batch. This LoDTensor is obtained in the forward and used in the backward.",
"duplicable":0,
"intermediate":1
},{
"name":"BatchHidden",
"comment":"(LoDTensor) the hidden state reorganized in batch. This LoDTensor is obtained in the forward and used in the backward.",
"duplicable":0,
"intermediate":1
},{
"name":"OrderedP0",
"comment":"(Tensor) the projection of the initial hidden state H0. This is a tensor with shape (N x P), where N is the batch size and P is the hidden size.",
"duplicable":0,
"intermediate":1
}],
}],
"attrs":[
"attrs":[
{
{
"name":"device_count",
"name":"use_peepholes",
"type":"int",
"type":"bool",
"comment":"device count",
"comment":"(bool, defalut: True) whether to enable diagonal/peephole connections.",
"generated":0
"generated":0
},{
},{
"name":"device_type",
"name":"is_reverse",
"type":"bool",
"comment":"(bool, defalut: False) whether to compute reversed LSTMP.",
"generated":0
},{
"name":"gate_activation",
"type":"string",
"type":"string",
"comment":"device type",
"comment":"(string, default: sigmoid)The activation for input gate, forget gate and output gate, `sigmoid` by default.",
"generated":0
},{
"name":"cell_activation",
"type":"string",
"comment":"(string, default: tanh)The activation for cell output, `tanh` by defalut.",
"generated":0
},{
"name":"candidate_activation",
"type":"string",
"comment":"(string, default: tanh)The activation for candidate hidden state, `tanh` by default.",
"generated":0
},{
"name":"proj_activation",
"type":"string",
"comment":"(string, default: tanh)The activation for projection output, `tanh` by defalut.",
"generated":0
"generated":0
}]
}]
},{
},{
"type":"read_from_array",
"type":"target_assign",
"comment":"\nReadFromArray Operator.\n\nRead a LoDTensor from a LoDTensor Array.\n\nAssume $T$ is LoDTensor, $i$ is the subscript of the array, and $A$ is the array. The\nequation is\n\n$$T = A[i]$$\n\n",
"comment":"\nThis operator is, for given the encoded boxes between prior boxes and\nground-truth boxes and ground-truth class labels, to assign classification\nand regression targets to each prior box as well as weights to each\nprior box. The weights is used to specify which prior box would not contribute\nto training loss.\n\nFor each instance, the output `PredBBoxLabel`, `PredBBoxWeight`,\n`PredScoreLabel` and `PredScoreWeight` are assigned based on `MatchIndices`.\nAssumed that the row offset for each instance in `EncodedGTBBox` is called lod,\nthis operato assigns classification/regression targets by performing the\nfollowing steps:\n\n1. Assigning all outpts based on `MatchIndices`:\n\nIf id = MatchIndices[i][j] > 0,\n\n PredBBoxLabel[i][j] = EncodedGTBBox[lod[i] + id][j]\n PredBBoxWeight[i][j] = 1.\n PredScoreLabel[i][j] = GTScoreLabel[lod[i] + id]\n PredScoreWeight[i][j] = 1.\n\nOtherwise, \n\n PredBBoxLabel[j][j] = [0., 0., 0., 0.]\n PredBBoxWeight[i][j] = 0.\n PredScoreLabel[i][j] = background_label\n PredScoreWeight[i][j] = 0.\n\n2. Assigning PredScoreWeight based on `NegIndices`:\n\nAssumed that the row offset for each instance in `NegIndices` is caleed neg_lod,\nfor i-th instance and all ids of NegIndices in this instance:\n\n PredScoreLabel[i][id] = background_label\n PredScoreWeight[i][id] = 1.0\n\n",
"inputs":[
"inputs":[
{
{
"name":"X",
"name":"EncodedGTBBox",
"comment":"(TensorArray) the array will be read from.",
"comment":"(LoDTensor), The encoded ground-truth bounding boxes with shape [Ng, Np, 4], where Ng is the total number of ground-truth boxes in this mini-batch, Np the number of predictions, 4 is the number of coordinate in [xmin, ymin, xmax, ymax] layout.",
"duplicable":0,
"duplicable":0,
"intermediate":0
"intermediate":0
},{
},{
"name":"I",
"name":"GTScoreLabel",
"comment":"(Tensor) the subscript index in tensor array. The number of element should be 1",
"comment":"(LoDTensor, default LoDTensor<int>), The input ground-truth labels with shape [Ng, 1], where the Ng is the same as it in the input of EncodedGTBBox.",
"duplicable":0,
"intermediate":0
},{
"name":"MatchIndices",
"comment":"(Tensor, default Tensor<int>), The input matched indices with shape [N, Np], where N is the batch size, Np is the same as it in the input of EncodedGTBBox. If MatchIndices[i][j] is -1, the j-th prior box is not matched to any ground-truh box in i-th instance.",
"duplicable":0,
"intermediate":0
},{
"name":"NegIndices",
"comment":"(LoDTensor, default LoDTensor<int>), The input negative example indices with shape [Neg, 1], where is the total number of negative example indices.",
"duplicable":0,
"duplicable":0,
"intermediate":0
"intermediate":0
}],
}],
"outputs":[
"outputs":[
{
{
"name":"Out",
"name":"PredBBoxLabel",
"comment":"(LoDTensor) the tensor will be read from.",
"comment":"(Tensor), The output encoded ground-truth labels with shape [N, Np, 4], N is the batch size and Np, 4 is the same as they in input of EncodedGTBBox. If MatchIndices[i][j] is -1, the PredBBoxLabel[i][j][:] is the encoded ground-truth box for background_label in i-th instance.",
"duplicable":0,
"duplicable":0,
"intermediate":0
"intermediate":0
}],
},{
"attrs":[]
"name":"PredBBoxWeight",
},{
"comment":"(Tensor), The weight for PredBBoxLabel with the shape of [N, Np, 1]",
"type":"shrink_rnn_memory",
"comment":"\nThis operator is used to shrink output batch of memory defined in dynamic RNN.\n\nDynamic RNN is able to handle variable-length sequences, in which, sequences in\na mini-batch are sorted by their lengths first. After that, the longest sequence\nbecomes the first one in the sorted batch, followed by the second longest, the\nthird longest, and so on. Dynamic RNN then slices a batch input timestep by\ntimestep from the sorted input. Once any sequence in the input batch reaches its\nend, memory defined in dynamicRNN has to shrink its outputs to adapt to the input\nbatch size for the next time step.\n",
"inputs":[
{
"name":"X",
"comment":"(LoDTensor) The RNN step memory to be shrinked.",
"duplicable":0,
"duplicable":0,
"intermediate":0
"intermediate":0
},{
},{
"name":"RankTable",
"name":"PredScoreLabel",
"comment":"(LoDRankTable) The lod_rank_table of dynamic RNN.",
"comment":"(Tensor, default Tensor<int>), The output score labels for each predictions with shape [N, Np, 1]. If MatchIndices[i][j] is -1, PredScoreLabel[i][j] = background_label.",
"duplicable":0,
"duplicable":0,
"intermediate":0
"intermediate":0
},{
},{
"name":"I",
"name":"PredScoreWeight",
"comment":"(LoDTensor) The step index. The RNN step memory 'X' will be shrinked to match the size of the input of the index'th step.",
"comment":"(Tensor), The weight for PredScoreLabel with the shape of [N, Np, 1]",
"duplicable":0,
"intermediate":0
}],
"attrs":[
{
"name":"background_label",
"type":"int",
"comment":"(int, default 0), Label index of background class.",
"generated":0
}]
},{
"type":"mean",
"comment":"\nMean Operator.\n\nOut is a scalar which is the mean of all elements in X. \n\n",
"inputs":[
{
"name":"X",
"comment":"The input of mean op",
"duplicable":0,
"duplicable":0,
"intermediate":0
"intermediate":0
}],
}],
"outputs":[
"outputs":[
{
{
"name":"Out",
"name":"Out",
"comment":"(LoDTensor) The shrinked RNN step memory.",
"comment":"The output of mean op",
"duplicable":0,
"duplicable":0,
"intermediate":0
"intermediate":0
}],
}],
...
@@ -3628,6 +3543,122 @@
...
@@ -3628,6 +3543,122 @@
"comment":"(int) Number of classes to be evaluated.",
"comment":"(int) Number of classes to be evaluated.",
"comment":"\nReturns a list of places based on flags. The list will be used for parallel\nexecution.\n",
"inputs":[],
"outputs":[
{
"name":"Out",
"comment":"vector of Place",
"duplicable":0,
"intermediate":0
}],
"attrs":[
{
"name":"device_count",
"type":"int",
"comment":"device count",
"generated":0
},{
"name":"device_type",
"type":"string",
"comment":"device type",
"generated":0
}]
},{
"type":"read_from_array",
"comment":"\nReadFromArray Operator.\n\nRead a LoDTensor from a LoDTensor Array.\n\nAssume $T$ is LoDTensor, $i$ is the subscript of the array, and $A$ is the array. The\nequation is\n\n$$T = A[i]$$\n\n",
"inputs":[
{
"name":"X",
"comment":"(TensorArray) the array will be read from.",
"duplicable":0,
"intermediate":0
},{
"name":"I",
"comment":"(Tensor) the subscript index in tensor array. The number of element should be 1",
"duplicable":0,
"intermediate":0
}],
"outputs":[
{
"name":"Out",
"comment":"(LoDTensor) the tensor will be read from.",
"duplicable":0,
"intermediate":0
}],
"attrs":[]
},{
"type":"rnn_memory_helper",
"comment":"",
"inputs":[
{
"name":"X",
"comment":"",
"duplicable":0,
"intermediate":0
}],
"outputs":[
{
"name":"Out",
"comment":"",
"duplicable":0,
"intermediate":0
}],
"attrs":[
{
"name":"dtype",
"type":"int",
"comment":"(int, default 5 (FP32)) Output data type",
"generated":0
}]
},{
"type":"shrink_rnn_memory",
"comment":"\nThis operator is used to shrink output batch of memory defined in dynamic RNN.\n\nDynamic RNN is able to handle variable-length sequences, in which, sequences in\na mini-batch are sorted by their lengths first. After that, the longest sequence\nbecomes the first one in the sorted batch, followed by the second longest, the\nthird longest, and so on. Dynamic RNN then slices a batch input timestep by\ntimestep from the sorted input. Once any sequence in the input batch reaches its\nend, memory defined in dynamicRNN has to shrink its outputs to adapt to the input\nbatch size for the next time step.\n",
"inputs":[
{
"name":"X",
"comment":"(LoDTensor) The RNN step memory to be shrinked.",
"duplicable":0,
"intermediate":0
},{
"name":"RankTable",
"comment":"(LoDRankTable) The lod_rank_table of dynamic RNN.",
"duplicable":0,
"intermediate":0
},{
"name":"I",
"comment":"(LoDTensor) The step index. The RNN step memory 'X' will be shrinked to match the size of the input of the index'th step.",
"duplicable":0,
"intermediate":0
}],
"outputs":[
{
"name":"Out",
"comment":"(LoDTensor) The shrinked RNN step memory.",
"duplicable":0,
"intermediate":0
}],
"attrs":[]
},{
},{
"type":"merge_lod_tensor",
"type":"merge_lod_tensor",
"comment":"\n Merge True and False branches of LoDTensor into a single Output,\n with a mask at certain lod level. X is used to obtain complete\n lod information. Please refer to SplitLoDTensorOp.",
"comment":"\n Merge True and False branches of LoDTensor into a single Output,\n with a mask at certain lod level. X is used to obtain complete\n lod information. Please refer to SplitLoDTensorOp.",
...
@@ -3925,24 +3956,6 @@
...
@@ -3925,24 +3956,6 @@
"comment":"(float, default 1.0)The scaling factor of the scale operator.",
"comment":"(float, default 1.0)The scaling factor of the scale operator.",
"generated":0
"generated":0
}]
}]
},{
"type":"mean",
"comment":"\nMean Operator.\n\nOut is a scalar which is the mean of all elements in X. \n\n",
"inputs":[
{
"name":"X",
"comment":"The input of mean op",
"duplicable":0,
"intermediate":0
}],
"outputs":[
{
"name":"Out",
"comment":"The output of mean op",
"duplicable":0,
"intermediate":0
}],
"attrs":[]
},{
},{
"type":"lookup_table",
"type":"lookup_table",
"comment":"\nLookup Table Operator.\n\nThis operator is used to perform lookups on the parameter W,\nthen concatenated into a dense tensor.\n\nThe input Ids can carry the LoD (Level of Details) information,\nor not. And the output only shares the LoD information with input Ids.\n\n",
"comment":"\nLookup Table Operator.\n\nThis operator is used to perform lookups on the parameter W,\nthen concatenated into a dense tensor.\n\nThe input Ids can carry the LoD (Level of Details) information,\nor not. And the output only shares the LoD information with input Ids.\n\n",
...
@@ -4000,6 +4013,47 @@
...
@@ -4000,6 +4013,47 @@
"intermediate":0
"intermediate":0
}],
}],
"attrs":[]
"attrs":[]
},{
"type":"logical_not",
"comment":"logical_not Operator\n\nIt operates element-wise on X, and returns the Out. X and Out are N-dim boolean tensors.\nEach element of Out is calculated by $$Out = !X$$\n",
"inputs":[
{
"name":"X",
"comment":"(LoDTensor) Operand of logical_not operator",
"duplicable":0,
"intermediate":0
}],
"outputs":[
{
"name":"Out",
"comment":"(LoDTensor) n-dim bool tensor. Each element is $$Out = !X$$",
"duplicable":0,
"intermediate":0
}],
"attrs":[]
},{
"type":"logical_and",
"comment":"logical_and Operator\n\nIt operates element-wise on X and Y, and returns the Out. X, Y and Out are N-dim boolean tensors.\nEach element of Out is calculated by $$Out = X \\&\\& Y$$\n",
"inputs":[
{
"name":"X",
"comment":"(LoDTensor) Left hand operand of logical_and operator",
"duplicable":0,
"intermediate":0
},{
"name":"Y",
"comment":"(LoDTensor) Right hand operand of logical_and operator",
"duplicable":0,
"intermediate":0
}],
"outputs":[
{
"name":"Out",
"comment":"(LoDTensor) n-dim bool tensor. Each element is $$Out = X \\&\\& Y$$",
"duplicable":0,
"intermediate":0
}],
"attrs":[]
},{
},{
"type":"logical_or",
"type":"logical_or",
"comment":"logical_or Operator\n\nIt operates element-wise on X and Y, and returns the Out. X, Y and Out are N-dim boolean tensors.\nEach element of Out is calculated by $$Out = X || Y$$\n",
"comment":"logical_or Operator\n\nIt operates element-wise on X and Y, and returns the Out. X, Y and Out are N-dim boolean tensors.\nEach element of Out is calculated by $$Out = X || Y$$\n",
