## Contrib Operator Schemas *This file is automatically generated from the [def files](/onnxruntime/core/graph/contrib_ops/contrib_defs.cc) via [this script](/tools/python/gen_doc.py). Do not modify directly and instead edit operator definitions.* * ai.onnx.training * ai.onnx.training.Adagrad * ai.onnx.training.Gradient * ai.onnx.training.GraphCall * ai.onnx.training.Momentum * com.microsoft * com.microsoft.AttnLSTM * com.microsoft.CDist * com.microsoft.ConvTransposeWithDynamicPads * com.microsoft.CropAndResize * com.microsoft.DequantizeLinear * com.microsoft.ExpandDims * com.microsoft.FusedConv * com.microsoft.FusedGemm * com.microsoft.GatherND * com.microsoft.Irfft * com.microsoft.MatMulInteger16 * com.microsoft.MaxpoolWithMask * com.microsoft.MulInteger * com.microsoft.MurmurHash3 * com.microsoft.Pad * com.microsoft.QLinearAdd * com.microsoft.QLinearAveragePool * com.microsoft.QLinearMul * com.microsoft.QLinearReduceMean * com.microsoft.QuantizeLinear * com.microsoft.Range * com.microsoft.ReduceSumInteger * com.microsoft.Rfft * com.microsoft.SampleOp * com.microsoft.Tokenizer * com.microsoft.Unique * com.microsoft.WordConvEmbedding * experimental com.microsoft.Attention * experimental com.microsoft.BiasGelu * experimental com.microsoft.EmbedLayerNormalization * experimental com.microsoft.FastGelu * experimental com.microsoft.Gelu * experimental com.microsoft.SkipLayerNormalization * com.microsoft.nchwc * com.microsoft.nchwc.AveragePool * com.microsoft.nchwc.Conv * com.microsoft.nchwc.GlobalAveragePool * com.microsoft.nchwc.GlobalMaxPool * com.microsoft.nchwc.MaxPool * com.microsoft.nchwc.ReorderInput * com.microsoft.nchwc.ReorderOutput * com.microsoft.nchwc.Upsample ## ai.onnx.training ### **ai.onnx.training.Adagrad** Compute one iteration of ADAGRAD, a stochastic gradient based optimization algorithm. This operator can conduct the optimization of multiple tensor variables. Let's define the behavior of this operator. As you can imagine, ADAGRAD requires some parameters: - The initial learning-rate "R". - The update count "T". That is, the number of training iterations conducted. - A L2-norm regularization coefficient "norm_coefficient". - A learning-rate decay factor "decay_factor". - A small constant "epsilon" to avoid dividing-by-zero. At each ADAGRAD iteration, the optimized tensors are moved along a direction computed based on their estimated gradient and accumulated squared gradient. Assume that only a single tensor "X" is updated by this operator. We need the value of "X", its gradient "G", and its accumulated squared gradient "H". Therefore, variables in this operator's input list are sequentially "R", "T", "X", "G", and "H". Other parameters are given as attributes because they are usually constants. Also, the corresponding output tensors are the new value of "X" (called "X_new"), and then the new accumulated squared gradient (called "H_new"). Those outputs are computed from the given inputs following the pseudo code below. Let "+", "-", "*", and "/" are all element-wise arithmetic operations with numpy-style broadcasting support. The pseudo code to compute those outputs is: // Compute a scalar learning-rate factor. At the first update of X, T is generally // 0 (0-based update index) or 1 (1-based update index). r = R / (1 + T * decay_factor); // Add gradient of 0.5 * norm_coefficient * ||X||_2^2, where ||X||_2 is the 2-norm. G_regularized = norm_coefficient * X + G; // Compute new accumulated squared gradient. H_new = H + G_regularized * G_regularized; // Compute the adaptive part of per-coordinate learning rate. Note that Sqrt(...) // computes element-wise square-root. H_adaptive = Sqrt(H_new) + epsilon // Compute the new value of "X". X_new = X - r * G_regularized / H_adaptive; If one assign this operators to optimize multiple inputs, for example, "X_1" and "X_2", the same pseudo code may be extended to handle all tensors jointly. More specifically, we can view "X" as a concatenation of "X_1" and "X_2" (of course, their gradient and accumulate gradient should be concatenated too) and then just reuse the entire pseudo code. Note that ADAGRAD was first proposed in http://jmlr.org/papers/volume12/duchi11a/duchi11a.pdf. In that reference paper, this operator is a special case of the Figure 1's composite mirror descent update. #### Version This version of the operator has been available since version 1 of the 'ai.onnx.training' operator set. #### Attributes
decay_factor : float
The decay factor of learning rate after one update.The effective learning rate is computed by r = R / (1 + T * decay_factor). Default to 0 so that increasing update counts doesn't reduce the learning rate.
epsilon : float
Small scalar to avoid dividing by zero.
norm_coefficient : float
Regularization coefficient in 0.5 * norm_coefficient * ||X||_2^2. Default to 0, which means no regularization.
#### Inputs (3 - ∞)
R : T1
The initial learning rate.
T : T2
The update count of "X". It should be a scalar.
inputs (variadic, heterogeneous) : T3
The current values of optimized tensors, followed by their respective gradients, followed by their respective accumulated squared gradients.For example, if two tensor "X_1" and "X_2" are optimized, The input list would be ["X_1", "X_2", gradient of "X_1", gradient of "X_2", accumulated squared gradient of "X_1", accumulated squared gradient of "X_2"].
#### Outputs (1 - ∞)
outputs (variadic, heterogeneous) : T3
Updated values of optimized tensors, followed by their updated values of accumulated squared gradients. For example, if two tensor "X_1" and "X_2" are optimized, the output list would be [new value of "X_1," new value of "X_2" new accumulated squared gradient of "X_1", new accumulated squared gradient of "X_2"].
#### Type Constraints
T1 : tensor(float), tensor(double)
Constrain input types to float scalars.
T2 : tensor(int64)
Constrain input types to 64-bit integer scalars.
T3 : tensor(float), tensor(double)
Constrain input and output types to float tensors.
### **ai.onnx.training.Gradient** Gradient operator computes the partial derivatives of a specific tensor w.r.t. some other tensors. This operator is widely used in gradient-based training algorithms. To illustrate its use, let's consider a computation graph, ``` X -----. | v W --> Conv --> H --> Gemm --> Y ^ | Z ``` , where W and Z are trainable tensors. Note that operators' attributes are omitted for the sake of simplicity. Let dY/dW (dY/dZ) be the gradient of Y with respect to W (Z). The user can compute gradient by inserting Gradient operator to form another graph shown below. ``` W --> Conv --> H --> Gemm --> Y | ^ ^ | | | | X Z | | | | | .----------' | | | (W/Z/X is the 1st/2nd/3rd input of Gradient as shown in | | | "xs" followed by "zs") | v v '---> Gradient(xs=["W", "Z"], zs=["X"], y="Y") | | | '-----------------------------------> dY/dW (1st output of Gradient) | '---------------------------------------> dY/dZ (2nd output of Gradient) ``` By definition, the tensor "y" is a function of independent variables in "xs" and "zs". Since we only compute the gradient of "y" w.r.t. the differentiable variables in "xs", this Gradient only outputs dY/dW and dY/dZ. Note that "H" cannot appear in "xs" and "zs". The reason is that "H" can be determined by tensors "W" and "X" and therefore "H" is not an independent variable. All outputs are optional. If needed, for example, user can assign an empty string to the 1st output name of that Gradient to skip the generation of dY/dW. Note that the concept of optional outputs can also be found in ONNX's RNN, GRU, and LSTM. Gradient operator can compute derivative against intermediate tensors. For example, the gradient of Y with respect to H can be done via ``` W --> Conv --> H --> Gemm --> Y ^ | ^ | | | X | Z .-------' | | .----------' | | (H/Z is the 1st/2nd input of Gradient as shown in "xs") v v Gradient(xs=["H", "Z"], y="Y") | | | '-----------------------------------> dY/dH (1st output of Gradient) | '---------------------------------------> dY/dZ (2nd output of Gradient) ``` It is possible to represent high-order differentiation using Gradient operators. For example, given the following linear model: ``` W --> Gemm --> Y --> Loss --> O ^ ^ | | X L ``` To compute the 2nd order derivative of O with respect to W (denoted by d^2O/dW^2), one can do ``` W --> Gemm --> Y --> Loss --> O | ^ ^ | | | | X .------------L | | | | | | | v +------+-+> Gradient(xs=["X", "W"], zs=["L"], y="O") ---> dO/dX (1st output of Gradient) | | | | | | | '---> dO/dW (2nd output of Gradient) | v v '---> Gradient(xs=["X", "W"], zs=["L"], y="dO/dW") ---> d(dO/dW)dX (1st output of | Gradient) | | '---> d^2O/dW^2 (2nd output of Gradient) ``` The tensors named in attributes "xs", "zs", and "y" define the differentiated computation graph, and the inputs to Gradient node define the values at which the gradient is computed. We can feed different tensors to the identified graph. For example, one can compute the gradient of Y with respect to H at a specific value of H, H_1, by providing that value as an input to the Gradient node. ``` W --> Conv --> H --> Gemm --> Y ^ ^ | | X Z Z_1 (2nd input of Gradient) | v H_1 --> Gradient(xs=["H", "Z"], y="Y") ---> dY/dH when H = H_1 and Y = Y_1. | '------------------------------> dY/dZ (2nd output of Gradient) ``` When the inputs of Gradient are the tensors named in "xs" and "zs", the computation can be optimized. More specifically, intermediate variables in forward pass can be reused if the gradient is computed via reverse-mode auto-differentiation. #### Version This version of the operator has been available since version 1 of the 'ai.onnx.training' operator set. #### Attributes
xs : list of strings (required)
Input tensor names of the differentiated sub-graph. It contains only the necessary differentiated inputs of a (sub-)graph. Variables (usually called intermediate variables) that can be generated from inputs cannot be included in this attribute.
y : string (required)
The targeted tensor. It can be viewed as the output of the differentiated function. The attribute "xs" and attribute "zs" are the minimal independent variable set that determines the value of "y".
zs : list of strings
Input tensor names of the differentiated sub-graph. It contains only the necessary non-differentiated inputs of a (sub-)graph. Variables (usually called intermediate variables) that can be generated from inputs cannot be included in this attribute.
#### Inputs (1 - ∞)
Inputs (variadic, heterogeneous) : T1
The values fed into graph identified by the attributes. The i-th input is the value of the i-th tensor specified in the concatenated list of the attribute "xs" and the attribute "zs". For example, if xs=["A", "B"] and zs=["C"], the first input is used as the value of symbol "A" and the 3rd input is substituted for all the occurrences of "C".
#### Outputs (1 - ∞)
Outputs (variadic, heterogeneous) : T2
The gradient of the tensor specified by the attribute "y" with respect to each of tensors specified in the attribute "xs". The i-th output is the gradient of "y" with respect to the i-th tensor specified in the attribute "xs".
#### Type Constraints
T1 : tensor(uint8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(int8), tensor(int16), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double), tensor(string), tensor(bool), tensor(complex64), tensor(complex128)
Allow outputs to be any kind of tensor.
T2 : tensor(float16), tensor(float), tensor(double)
Allow inputs to be any kind of floating-point tensor.
### **ai.onnx.training.GraphCall** The GraphCall operator invokes a graph inside TrainingInfoProto's algorithm field. The GraphCall inputs and outputs are bound to those of invoked graph by position. If a graph input has an initializer, that input is considered optional. All graph outputs are optional. Below Python syntax is used for describing dictionary and list. Assume that ModelProto's graph field has - name: "MyInferenceGraph" - input: ["X", "W", "Z"] - initializer: [W] - output: ["Y"] as visualized below for inference. ``` X -----. | v W --> Conv --> H --> Gemm --> Y ^ | Z ``` Assume that the training algorithm contains - inputs: ["X_1", "Z_1", "C"] - initializer: [T] - outputs: ["W_new"] with a dictionary - update_binding: {"W": "W_new", "T": "T_new"} Inside the training algorithm graph, one can invoke the inference graph via adding a GraphCall node with - inputs: ["X_1", "W", Z_1"] - outputs: ["Y_1"] - an attribute graph_name="MyInferenceGraph", The initializers, "W" and "T" in this case, in update_binding are considered globally-visible and mutable variables, which can be used as inputs of operators in the training graph. An example training algorithm graph may look like ``` .-------- W (a global and mutable variable from | | the inference graph) | | | .-----'-----------. | | | | | v | | .-- X_1 --> GraphCall(graph_name="MyInferenceGraph") | | | | | | | | | | | | | Z_1 -----' | | | | | V | | | | Y_1 ---> Loss ---> O | | | | ^ | | | | | | | `--. | C | | | | | | | | | .----------------' | | | | | | | v v v | `--> Gradient(xs=["W"], zs=["X_1", "Z_1", "C"], y="O") | | | v | dO_dW (gradient of W) 1 (a scalar one) | | | | V v | Div <--- T ------------> Add ---> T_new | | (T is the number of training iterations. | | T is also globally visible and mutable.) | v `-----> Sub ----> W_new ``` where Loss is a dummy node which computes the minimized objective function. The variable "W" is an optional input in the called graph. If the user omits it, the input list of GraphCall becomes ["X_1", "", "Z_1"]. In this case, from the view of computation graph, the Conv operator invoked by GraphCall's may be still connected the global "W" variable and therefore the structure of the computation graph is unchanged. #### Version This version of the operator has been available since version 1 of the 'ai.onnx.training' operator set. #### Attributes
graph_name : string (required)
The invoked graph's name. The only allowed value is the name of the inference graph, which is stored in "ModelProto.graph.name" in the ONNX model format.
#### Inputs (1 - ∞)
Inputs (variadic, heterogeneous) : T
Inputs fed to the invoked graph. The i-th input here goes to the i-th input of the invoked graph. To omit an optional input in this field, the user can drop it or use an empty string.
#### Outputs (1 - ∞)
Outputs (variadic, heterogeneous) : T
The outputs generated by the called graph. Its i-th value is bound to the i-th output of the called graph. Similar to the inputs, all outputs are optional.
#### Type Constraints
T : tensor(uint8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(int8), tensor(int16), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double), tensor(string), tensor(bool), tensor(complex64), tensor(complex128)
Allow inputs and outputs to be any kind of tensor.
### **ai.onnx.training.Momentum** Compute one iteration of stochastic gradient update with momentum. This operator can conduct the optimization of multiple tensor variables. Let's define the behavior of this operator. As you can imagine, SG with momentum requires several parameters: - The learning-rate "R". - The update count "T". That is, the number of conducted training iterations. It should be zero in the first training iteration. - A L2-norm regularization coefficient "norm_coefficient". - A decay coefficient of previous accumulated gradient (i.e., momentum) "alpha". - The scaling coefficient of current gradient "beta". - An attribute to choose either standard momentum or Nesterov's momentum "mode" should be used. For the sake of simplicity, assume that there is only one tensor (called "X") to be optimized. Other necessary inputs are "X"'s gradient (called "G") and "X"'s momentum (called "V"). This Momentum operator maps all these inputs to the new value of "X" (called "X_new") and its new momentum (called "V_new"). This operator supports two different momentum algorithms. Set the attribute "mode" to "nesterov" if Nesterov's momentum is desired. Otherwise, set the attribute "model" to "standard" to use standard momentum. Computation details are described subsequently. Let "+", "-", "*", and "/" are all element-wise operations with numpy-style broadcasting. Pseudo code for SG with standard momentum: // Add gradient of 0.5 * norm_coefficient * ||X||^2, where ||X|| is the sum of squared // values of all elements in X. G_regularized = norm_coefficient * X + G // In the first training iteration, beta should always be 1. beta_adjusted = T > 0 ? beta : 1 // Compute the current momentum based on previous momentum and the current gradient. V_new = alpha * V + beta_adjusted * G_regularized // Update X. X_new = X - R * V_new Pseudo code for SG with Nesterov's momentum: // Add gradient of 0.5 * norm_coefficient * ||X||^2, where ||X|| is the sum of squared // values of all elements in X. G_regularized = norm_coefficient * X + G; // In the first training iteration, beta should always be 1. beta_adjusted = T > 0 ? beta : 1 // Compute the current momentum based on previous momentum and the current gradient. V_new = alpha * V + beta_adjusted * G_regularized; // Compute final update direction and then update X. X_new = X - R * (G_regularized + alpha * V_new) If one assign this operators to optimize multiple inputs, for example, "X_1" and "X_2". The same pseudo code would be extended to handle all tensors jointly. More specifically, we can view "X" as a concatenation of "X_1" and "X_2" (of course, their gradient and accumulate gradient should be concatenated too) and then our pseudo code becomes applicable. #### Version This version of the operator has been available since version 1 of the 'ai.onnx.training' operator set. #### Attributes
alpha : float (required)
The decay factor of momentum. It should be a scalar.
beta : float (required)
The coefficient of gradient in computing new momentum. It should be a scalar.
mode : string (required)
Its value should be either "nesterov" or "standard". The value "nesterov" leads to the use of Nesterov's momentum while "standard" invokes stochastic gradient method using standard momentum
norm_coefficient : float (required)
Coefficient of 0.5 * norm_coefficient * ||X||^2.
#### Inputs (3 - ∞)
R : T1
The learning rate.
T : T2
Update count of "X". It should be a scalar.
inputs (variadic, heterogeneous) : T3
It sequentially contains the current values of optimized tensors, then their gradient tensors, and finally their momentum tensors. For example, if two tensors "X_1" and "X_2" are optimized, The expected input list would be ["X_1", "X_2", gradient of "X_1", gradient of "X_2", momentum of "X_1", momentum of "X_2"].
#### Outputs (1 - ∞)
outputs (variadic, heterogeneous) : T3
It sequentially contains the new values of optimized tensors and then the new values of their momentum tensors. For example, if two tensors "X_1" and "X_2" are optimized, the output list would be [new value of "X_1," new value of "X_2" new momentum of "X_1", new momentum of "X_2"].
#### Type Constraints
T1 : tensor(float), tensor(double)
Constrain input types to float scalars.
T2 : tensor(int64)
Constrain input types to 64-bit integer scalars.
T3 : tensor(float), tensor(double)
Constrain input types to float tensors.
## com.microsoft ### **com.microsoft.AttnLSTM** Computes an one-layer RNN where its RNN Cell is an AttentionWrapper wrapped a LSTM Cell. The RNN layer contains following basic component: LSTM Cell, Bahdanau Attention Mechanism, AttentionWrapp. Activation functions: Relu(x) - max(0, x) Tanh(x) - (1 - e^{-2x})/(1 + e^{-2x}) Sigmoid(x) - 1/(1 + e^{-x}) (NOTE: Below are optional) Affine(x) - alpha*x + beta LeakyRelu(x) - x if x >= 0 else alpha * x ThresholdedRelu(x) - x if x >= alpha else 0 ScaledTanh(x) - alpha*Tanh(beta*x) HardSigmoid(x) - min(max(alpha*x + beta, 0), 1) Elu(x) - x if x >= 0 else alpha*(e^x - 1) Softsign(x) - x/(1 + |x|) Softplus(x) - log(1 + e^x) Softmax(x) - exp(x) / sum(exp(x)) Bahdanau Attention Mechanism: `M` - Memory tensor. `VALUES` - masked Memory by its real sequence length. `MW` - Memory layer weight. `KEYS` - Processed memory tensor by the memory layer. KEYS = M * MW `Query` - Query tensor, normally at specific time step in sequence. `QW` - Query layer weight in the attention mechanism `PQ` - processed query, = `Query` * `QW` `V' - attention vector `ALIGN` - calculated alignment based on Query and KEYS ALIGN = softmax(reduce_sum(`V` * Tanh(`KEYS` + `PQ`))) `CONTEXT` - context based on `ALIGN` and `VALUES` CONTEXT = `ALIGN` * `VALUES` LSTM Cell: `X` - input tensor concat with attention state in the attention wrapper `i` - input gate `o` - output gate `f` - forget gate `c` - cell gate `t` - time step (t-1 means previous time step) `W[iofc]` - W parameter weight matrix for input, output, forget, and cell gates `R[iofc]` - R recurrence weight matrix for input, output, forget, and cell gates `Wb[iofc]` - W bias vectors for input, output, forget, and cell gates `Rb[iofc]` - R bias vectors for input, output, forget, and cell gates `P[iof]` - P peephole weight vector for input, output, and forget gates `WB[iofc]` - W parameter weight matrix for backward input, output, forget, and cell gates `RB[iofc]` - R recurrence weight matrix for backward input, output, forget, and cell gates `WBb[iofc]` - W bias vectors for backward input, output, forget, and cell gates `RBb[iofc]` - R bias vectors for backward input, output, forget, and cell gates `PB[iof]` - P peephole weight vector for backward input, output, and forget gates `H` - Hidden state `num_directions` - 2 if direction == bidirectional else 1 Equations (Default: f=Sigmoid, g=Tanh, h=Tanh): - it = f(Xt*(Wi^T) + Ht-1*(Ri^T) + Pi (.) Ct-1 + Wbi + Rbi) - ft = f(Xt*(Wf^T) + Ht-1*(Rf^T) + Pf (.) Ct-1 + Wbf + Rbf) - ct = g(Xt*(Wc^T) + Ht-1*(Rc^T) + Wbc + Rbc) - Ct = ft (.) Ct-1 + it (.) ct - ot = f(Xt*(Wo^T) + Ht-1*(Ro^T) + Po (.) Ct + Wbo + Rbo) - Ht = ot (.) h(Ct) AttentionWrapp Notations: `lstm()' - wrapped inner cell. Ht, Ct = lstm(concat(Xt, ATTNt-1), Ct-1) `am()` - attention mechanism the wrapper used. CONTEXTt, ALIGNt = am(Ht, ALIGNt-1) `AW` - attention layer weights, optional. `ATTN` - attention state, initial is zero. If `AW` provided, it is the output of the attention layer, ATTNt = concat(Ht, CONTEXTt) * AW otherwise, ATTNt = CONTEXTt RNN layer output: `Y` - if needed is the sequence of Ht from lstm cell. `Y_h` - is the last valid H from lstm cell. `Y_c` - is the last valid C from lstm cell. #### Version This version of the operator has been available since version 1 of the 'com.microsoft' operator set. #### Attributes
activation_alpha : list of floats
Optional scaling values used by some activation functions. The values are consumed in the order of activation functions, for example (f, g, h) in LSTM. Default values are the same as of corresponding ONNX operators.For example with LeakyRelu, the default alpha is 0.01.
activation_beta : list of floats
Optional scaling values used by some activation functions. The values are consumed in the order of activation functions, for example (f, g, h) in LSTM. Default values are the same as of corresponding ONNX operators.
activations : list of strings
A list of 3 (or 6 if bidirectional) activation functions for input, output, forget, cell, and hidden. The activation functions must be one of the activation functions specified above. Optional: See the equations for default if not specified.
clip : float
Cell clip threshold. Clipping bounds the elements of a tensor in the range of [-threshold, +threshold] and is applied to the input of activations. No clip if not specified.
direction : string
Specify if the RNN is forward, reverse, or bidirectional. Must be one of forward (default), reverse, or bidirectional.
hidden_size : int
Number of neurons in the hidden layer.
input_forget : int
Couple the input and forget gates if 1, default 0.
#### Inputs (3 - 14)
X : T
The input sequences packed (and potentially padded) into one 3-D tensor with the shape of `[seq_length, batch_size, input_size]`
W : T
The weight tensor for the gates. Concatenation of `W[iofc]` and `WB[iofc]` (if bidirectional) along dimension 0. The tensor has shape `[num_directions, 4*hidden_size, input_size]`.
R : T
The recurrence weight tensor. Concatenation of `R[iofc]` and `RB[iofc]` (if bidirectional) along dimension 0. This tensor has shape `[num_directions, 4*hidden_size, hidden_size]`.
B (optional) : T
The bias tensor for input gate. Concatenation of `[Wb[iofc], Rb[iofc]]`, and `[WBb[iofc], RBb[iofc]]` (if bidirectional) along dimension 0. This tensor has shape `[num_directions, 8*hidden_size]`. Optional: If not specified - assumed to be 0.
sequence_lens (optional) : T1
Optional tensor specifying lengths of the sequences in a batch. If not specified - assumed all sequences in the batch to have length `seq_length`. It has shape `[batch_size]`
initial_h (optional) : T
Optional initial value of the hidden. If not specified - assumed to be 0. It has shape `[num_directions, batch_size, hidden_size]`.
initial_c (optional) : T
Optional initial value of the cell. If not specified - assumed to be 0. It has shape `[num_directions, batch_size, hidden_size]`.
P (optional) : T
The weight tensor for peepholes. Concatenation of `P[iof]` and `PB[iof]` (if bidirectional) along dimension 0. It has shape `[num_directions, 3*hidde_size]`. Optional: If not specified - assumed to be 0.
QW (optional) : T
The weight tensor of the query layer in the attention mechanism. Should be of shape `[num_directions, am_query_depth(hidden_size of lstm), am_attn_size]`
MW (optional) : T
The weight tensor of the memory layer in the attention mechanism. Should be of shape `[num_directions, memory_depth, am_attn_size]`
V (optional) : T
The attention_v tensor in the attention mechanism. Should be of shape `[num_directions, am_attn_size]`
M (optional) : T
The sequence of the memory (input) for attention mechanism. Should be of `[batch_size, max_memory_step, memory_depth]`
memory_seq_lens (optional) : T1
The sequence length of the input memory for the attention mechanism. Should be of `[batch_size]`
AW (optional) : T
The weights of attention layer in the attention wrapper. If exists, should be of shape `[num_directions, memory_depth+hidden_size, aw_attn_size]. Please note that attention mechanism context depth is also memory_depth in the attention mechanism.`
#### Outputs (0 - 3)
Y (optional) : T
A tensor that concats all the intermediate output values of the hidden. It has shape `[seq_length, num_directions, batch_size, hidden_size]`
Y_h (optional) : T
The last output value of the hidden. It has shape `[num_directions, batch_size, hidden_size]`.
Y_c (optional) : T
The last output value of the cell. It has shape `[num_directions, batch_size, hidden_size]`.
#### Type Constraints
T : tensor(float), tensor(double)
Constrain input and output types to float tensors.
T1 : tensor(int32)
Constrain seq_lens to integral tensors.
### **com.microsoft.CDist** #### Version This version of the operator has been available since version 1 of the 'com.microsoft' operator set. #### Attributes
metric : string
The distance metric to use. If a string, the distance function can be "braycurtis", "canberra", "chebyshev", "cityblock", "correlation", "cosine", "dice", "euclidean", "hamming", "jaccard", "jensenshannon", "kulsinski", "mahalanobis", "matching", "minkowski", "rogerstanimoto", "russellrao", "seuclidean", "sokalmichener", "sokalsneath", "sqeuclidean", "wminkowski", "yule".
#### Inputs
A : T
2D matrix with shape (M,N)
B : T
2D matrix with shape (K,N)
#### Outputs
C : T
A 2D Matrix that represents the distance between each pair of the two collections of inputs.
#### Type Constraints
T : tensor(float), tensor(double)
Constrains input to only numeric types.
### **com.microsoft.ConvTransposeWithDynamicPads** #### Version This version of the operator has been available since version 1 of the 'com.microsoft' operator set. #### Attributes
auto_pad : string
dilations : list of ints
group : int
kernel_shape : list of ints
output_padding : list of ints
strides : list of ints
#### Inputs (2 - 4)
X : T
W : T
Pads (optional) : tensor(int64)
B (optional) : T
#### Outputs
Y : T
#### Type Constraints
T : tensor(float16), tensor(float), tensor(double)
Constrain input and output types to float tensors
### **com.microsoft.CropAndResize** Extracts crops from the input image tensor and resizes them using bilinear sampling or nearest neighbor sampling (possibly with aspect ratio change) to a common output size specified by crop_height and crop_width. Returns a tensor with crops from the input image at positions defined at the bounding box locations in boxes. The cropped boxes are all resized (with bilinear or nearest neighbor interpolation) to a fixed size = [crop_height, crop_width]. The result is a 4-D tensor [num_boxes, crop_height, crop_width, depth]. The resizing is corner aligned. #### Version This version of the operator has been available since version 1 of the 'com.microsoft' operator set. #### Attributes
extrapolation_value : float
Value used for extrapolation, when applicable. Default is 0.0f.
mode : string
The pooling method. Two modes are supported: 'bilinear' and 'nearest'. Default is 'bilinear'.
#### Inputs
X : T1
Input data tensor from the previous operator; 4-D feature map of shape (N, C, H, W), where N is the batch size, C is the number of channels, and H and W are the height and the width of the data.
rois : T1
RoIs (Regions of Interest) to pool over; rois is 2-D input of shape (num_rois, 4) given as [[y1, x1, y2, x2], ...]. The RoIs' coordinates are normalized in the coordinate system of the input image. Each coordinate set has a 1:1 correspondence with the 'batch_indices' input.
batch_indices : T2
1-D tensor of shape (num_rois,) with each element denoting the index of the corresponding image in the batch.
crop_size : T2
1-D tensor of 2 elements: [crop_height, crop_width]. All cropped image patches are resized to this size. Both crop_height and crop_width need to be positive.
#### Outputs
Y : T1
RoI pooled output, 4-D tensor of shape (num_rois, C, crop_height, crop_width). The r-th batch element Y[r-1] is a pooled feature map corresponding to the r-th RoI X[r-1].
#### Type Constraints
T1 : tensor(float16), tensor(float), tensor(double)
Constrain types to float tensors.
T2 : tensor(int32)
Constrain types to int tensors.
### **com.microsoft.DequantizeLinear** The linear dequantization operator. It consumes a quantized data, a scale, a zero point and computes the full precision data. The dequantization formula is y = (x - x_zero_point) * x_scale. Scale and zero point must have same shape. They must be either scalar (per tensor) or 1-D tensor (per 'axis'). #### Version This version of the operator has been available since version 1 of the 'com.microsoft' operator set. #### Attributes
axis : int
The axis along which same quantization parameters are applied. It's optional.If it's not specified, it means per-tensor quantization and input 'x_scale' and 'x_zero_point' must be scalars.If it's specified, it means per 'axis' quantization and input 'x_scale' and 'x_zero_point' must be 1-D tensors.
#### Inputs
x : T1
N-D quantized Input tensor to be de-quantized.
x_scale : T2
Scale for input 'x'. It could be a scalar or a 1-D tensor, which means a per-tensor or per-axis quantization.If it's a 1-D tensor, its number of elements should be equal to the dimension value of 'axis' dimension of input 'x'.
x_zero_point : T1
Zero point for input 'x'. It could be a scalar or a 1-D tensor, which means a per-tensor or per-axis quantization.If it's a 1-D tensor, its number of elements should be equal to the dimension value of 'axis' dimension of input 'x'.
#### Outputs
y : T2
N-D full precision output tensor. It has same shape as input 'x'.
#### Type Constraints
T1 : tensor(int8), tensor(uint8)
Constrain 'x' and 'x_zero_point' to 8-bit integer tensors.
T2 : tensor(float16), tensor(float)
Constrain 'y', 'x_scale' to float tensors.
### **com.microsoft.ExpandDims** ExpandDims echo operator. #### Version This version of the operator has been available since version 1 of the 'com.microsoft' operator set. #### Inputs
X : T
input
axis : tensor(int32)
Specified axis to insert a dimension
#### Outputs
Y : T
output
#### Type Constraints
T : tensor(uint8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(int8), tensor(int16), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double), tensor(string), tensor(bool), tensor(complex64), tensor(complex128)
Constrain to any tensor type. If the dtype attribute is not provided this must be a valid output type.
### **com.microsoft.FusedConv** The fused convolution operator schema is the same as Conv besides it includes an attribute activation. #### Version This version of the operator has been available since version 1 of the 'com.microsoft' operator set. #### Attributes
activation : string
activation_params : list of floats
auto_pad : string
dilations : list of ints
group : int
kernel_shape : list of ints
pads : list of ints
strides : list of ints
#### Inputs (2 - 3)
X : T
W : T
B (optional) : T
#### Outputs
Y : T
#### Type Constraints
T : tensor(float16), tensor(float), tensor(double)
Constrain input and output types to float tensors
### **com.microsoft.FusedGemm** The FusedGemm operator schema is the same as Gemm besides it includes attributes activation and leaky_relu_alpha. #### Version This version of the operator has been available since version 1 of the 'com.microsoft' operator set. #### Attributes
activation : string
alpha : float
Scalar multiplier for the product of input tensors A * B.
beta : float
Scalar multiplier for input tensor C.
leaky_relu_alpha : float
transA : int
Whether A should be transposed
transB : int
Whether B should be transposed
#### Inputs
A : T
Input tensor A. The shape of A should be (M, K) if transA is 0, or (K, M) if transA is non-zero.
B : T
Input tensor B. The shape of B should be (K, N) if transB is 0, or (N, K) if transB is non-zero.
C : T
Input tensor C. The shape of C should be unidirectional broadcastable to (M, N).
#### Outputs
Y : T
Output tensor of shape (M, N).
#### Type Constraints
T : tensor(float16), tensor(float), tensor(double), tensor(uint32), tensor(uint64), tensor(int32), tensor(int64)
Constrain input and output types to float/int tensors.
### **com.microsoft.GatherND** Given `data` tensor of rank r >= 1, and `indices` tensor of rank q >= 1, gather slices of `data` into an output tensor of rank q - 1 + r - indices[-1]. Example 1: data = [[0,1],[2,3]] indices = [[0,0],[1,1]] output = [0,3] Example 2: data = [[0,1],[2,3]] indices = [[1],[0]] output = [[2,3],[0,1]] Example 3: data = [[[0,1],[2,3]],[[4,5],[6,7]]] indices = [[0,1],[1,0]] output = [[2,3],[4,5]] Example 4: data = [[[0,1],[2,3]],[[4,5],[6,7]]] indices = [[[0,1]],[[1,0]]] output = [[[2,3]],[[4,5]]] #### Version This version of the operator has been available since version 1 of the 'com.microsoft' operator set. #### Inputs
data : T
Tensor of rank r >= 1.
indices : Tind
Tensor of rank q >= 1.
#### Outputs
output : T
Tensor of rank q-1+r-indices[-1].
#### Type Constraints
T : tensor(uint8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(int8), tensor(int16), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double), tensor(string), tensor(bool), tensor(complex64), tensor(complex128)
Constrain input and output types to any tensor type.
Tind : tensor(int32), tensor(int64)
Constrain indice type to int32 or int64
### **com.microsoft.Irfft** #### Version This version of the operator has been available since version 1 of the 'com.microsoft' operator set. #### Attributes
normalized : int
onesided : int
signal_ndim : int (required)
#### Inputs
X : T
input tensor
#### Outputs
Y : T
output tensor
#### Type Constraints
T : tensor(float), tensor(double), tensor(float16)
Constrain input and output types to float or half tensors.
### **com.microsoft.MatMulInteger16** Matrix product that behaves like numpy.matmul: https://docs.scipy.org/doc/numpy-1.13.0/reference/generated/numpy.matmul.html. The production MUST never overflow. The accumulation may overflow if and only if in 32 bits. #### Version This version of the operator has been available since version 1 of the 'com.microsoft' operator set. #### Inputs
A : T1
N-dimensional matrix A
B : T2
N-dimensional matrix B
#### Outputs
Y : T3
Matrix multiply results from A * B
#### Type Constraints
T1 : tensor(int16), tensor(uint16)
Constrain input A data types as 16-bit integer tensor
T2 : tensor(int16), tensor(uint16)
Constrain input B data types as 16-bit integer tensor
T3 : tensor(int32), tensor(uint32)
Constrain output Y data types as 32-bit integer tensor.T3 must be tensor(uint32) when both T1 and T2 are tensor(uint16),or must be tensor(int32) when either T1 or T2 is tensor(int16).
### **com.microsoft.MaxpoolWithMask** For internal use. #### Version This version of the operator has been available since version 1 of the 'com.microsoft' operator set. #### Attributes
auto_pad : string
kernel_shape : list of ints
pads : list of ints
storage_order : int
strides : list of ints
#### Inputs
X : T
M : tensor(int32)
mask
#### Outputs
Y : T
#### Type Constraints
T : tensor(float)
Constrain input0 and output types to float tensors
### **com.microsoft.MulInteger** Performs element-wise binary quantized multiplication (with Numpy-style broadcasting support). "This operator supports **multidirectional (i.e., Numpy-style) broadcasting**" The output of this op is the int32 accumulated result of the mul operation ``` C (int32) = (A - A_zero_point) * (B - B_zero_point) ``` #### Version This version of the operator has been available since version 1 of the 'com.microsoft' operator set. #### Inputs (3 - 4)
A : T
First operand.
A_zero_point (optional) : T
Input A zero point. Default value is 0 if it's not specified. It's a scalar, which means a per-tensor/layer quantization.
B : T
Second operand.
B_zero_point (optional) : T
Input B zero point. Default value is 0 if it's not specified. It's a scalar, which means a per-tensor/layer quantization.
#### Outputs
C : T1
Constrain output to 32 bit tensor
#### Type Constraints
T : tensor(uint8), tensor(int8)
Constrain input types to 8 bit signed and unsigned tensors.
T1 : tensor(int32)
Constrain output types to 32 bit tensors.
### **com.microsoft.MurmurHash3** The underlying implementation is MurmurHash3_x86_32 generating low latency 32bits hash suitable for implementing lookup tables, Bloom filters, count min sketch or feature hashing. #### Version This version of the operator has been available since version 1 of the 'com.microsoft' operator set. #### Attributes
positive : int
If value is 1, output type is uint32_t, else int32_t. Default value is 1.
seed : int
Seed for the hashing algorithm, unsigned 32-bit integer, default to 0.
#### Inputs
X : T1
An input tensor to hash.
#### Outputs
Y : T2
32-bit hash value.
#### Type Constraints
T1 : tensor(uint32), tensor(int32), tensor(string)
Constrain input type to unsigned or signed 32-bit integer tensor, or string tensor. It should be utf-8 encoded if using unicode.
T2 : tensor(uint32), tensor(int32)
Constrain output type to unsigned and signed 32-bit integer tensor.
### **com.microsoft.Pad** Given `data` tensor, pads, mode, and value. Example: Insert 0 pads to the beginning of the second dimension. data = [ [1.0, 1.2], [2.3, 3.4], [4.5, 5.7], ] pads = [0, 2, 0, 0] output = [ [ [0.0, 0.0, 1.0, 1.2], [0.0, 0.0, 2.3, 3.4], [0.0, 0.0, 4.5, 5.7], ], ] #### Version This version of the operator has been available since version 1 of the 'com.microsoft' operator set. #### Attributes
mode : string
Three modes: `constant`(default) - pads with a given constant value, `reflect` - pads with the reflection of the vector mirrored on the first and last values of the vector along each axis, `edge` - pads with the edge values of array
#### Inputs (2 - 3)
data : T
Input tensor.
pads : tensor(int64)
Tensor of integers indicating the number of padding elements to add or remove (if negative) at the beginning and end of each axis. For 2D input tensor, it is the number of pixels. `pads` should be a 1D tensor of shape [2 * input_rank] or a 2D tensor of shape [1, 2 * input_rank]. `pads` format (1D example) should be as follow [x1_begin, x2_begin,...,x1_end, x2_end,...], where xi_begin is the number of pixels added at the beginning of axis `i` and xi_end, the number of pixels added at the end of axis `i`.
value (optional) : T
(Optional) A scalar or rank 1 tensor containing a single value to be filled if the mode chosen is `constant` (by default it is 0.0).
#### Outputs
output : T
Tensor after padding.
#### Type Constraints
T : tensor(float16), tensor(float), tensor(double)
Constrain input and output types to float tensors.
### **com.microsoft.QLinearAdd** Performs element-wise binary addition on 8 bit data types (with Numpy-style broadcasting support). C = (A_scale * (A - A_zero_point) + B_scale * (B - B_zero_point))/C_scale + C_zero_point #### Version This version of the operator has been available since version 1 of the 'com.microsoft' operator set. #### Inputs (7 - 8)
A : T
First operand.
A_scale : tensor(float)
Input A's scale. It's a scalar, which means a per-tensor/layer quantization.
A_zero_point (optional) : T
Input A zero point. Default value is 0 if it's not specified. It's a scalar, which means a per-tensor/layer quantization.
B : T
Second operand.
B_scale : tensor(float)
Input B's scale. It's a scalar, which means a per-tensor/layer quantization.
B_zero_point (optional) : T
Input B zero point. Default value is 0 if it's not specified. It's a scalar, which means a per-tensor/layer quantization.
C_scale : tensor(float)
Output scale. It's a scalar, which means a per-tensor/layer quantization.
C_zero_point (optional) : T
Output zero point. Default value is 0 if it's not specified. It's a scalar, which means a per-tensor/layer quantization.
#### Outputs
C : T
Result, has same element type as two inputs
#### Type Constraints
T : tensor(uint8), tensor(int8)
Constrain input and output types to 8 bit signed and unsigned tensors.
### **com.microsoft.QLinearAveragePool** QLinearAveragePool consumes an input tensor X and applies average pooling across the tensor according to kernel sizes, stride sizes, and pad lengths. average pooling consisting of computing the average on all values of a subset of the input tensor according to the kernel size and downsampling the data into the output tensor Y for further processing. The output spatial shape will be following: ``` output_spatial_shape[i] = floor((input_spatial_shape[i] + pad_shape[i] - kernel_spatial_shape[i]) / strides_spatial_shape[i] + 1) ``` or ``` output_spatial_shape[i] = ceil((input_spatial_shape[i] + pad_shape[i] - kernel_spatial_shape[i]) / strides_spatial_shape[i] + 1) ``` if ceil_mode is enabled ``` * pad_shape[i] is sum of pads along axis i ``` `auto_pad` is a DEPRECATED attribute. If you are using them currently, the output spatial shape will be following: ``` VALID: output_spatial_shape[i] = ceil((input_spatial_shape[i] - kernel_spatial_shape[i] + 1) / strides_spatial_shape[i]) SAME_UPPER or SAME_LOWER: output_spatial_shape[i] = ceil(input_spatial_shape[i] / strides_spatial_shape[i]) ``` And pad shape will be following if `SAME_UPPER` or `SAME_LOWER`: ``` pad_shape[i] = (output_spatial_shape[i] - 1) * strides_spatial_shape[i] + kernel_spatial_shape[i] - input_spatial_shape[i] ``` The output of each pooling window is divided by the number of elements (exclude pad when attribute count_include_pad is zero). Input and output scales and zero points are used to convert the output to a new quantization range. Output = Dequantize(Input) -> AveragePool on fp32 data -> Quantize(output) #### Version This version of the operator has been available since version 1 of the 'com.microsoft' operator set. #### Attributes
auto_pad : string
auto_pad must be either NOTSET, SAME_UPPER, SAME_LOWER or VALID. Where default value is NOTSET, which means explicit padding is used. SAME_UPPER or SAME_LOWER mean pad the input so that the output spatial size match the input.In case of odd number add the extra padding at the end for SAME_UPPER and at the beginning for SAME_LOWER. VALID mean no padding.
ceil_mode : int
Whether to use ceil or floor (default) to compute the output shape.
count_include_pad : int
Whether include pad pixels when calculating values for the edges. Default is 0, doesn't count include pad.
kernel_shape : list of ints (required)
The size of the kernel along each axis.
pads : list of ints
Padding for the beginning and ending along each spatial axis, it can take any value greater than or equal to 0. The value represent the number of pixels added to the beginning and end part of the corresponding axis. `pads` format should be as follow [x1_begin, x2_begin...x1_end, x2_end,...], where xi_begin the number of pixels added at the beginning of axis `i` and xi_end, the number of pixels added at the end of axis `i`. This attribute cannot be used simultaneously with auto_pad attribute. If not present, the padding defaults to 0 along start and end of each spatial axis.
strides : list of ints
Stride along each spatial axis. If not present, the stride defaults to 1 along each spatial axis.
#### Inputs (4 - 5)
X : T
Input data tensor from the previous operator; dimensions for image case are (N x C x H x W), where N is the batch size, C is the number of channels, and H and W are the height and the width of the data. For non image case, the dimensions are in the form of (N x C x D1 x D2 ... Dn), where N is the batch size. Optionally, if dimension denotation is in effect, the operation expects the input data tensor to arrive with the dimension denotation of [DATA_BATCH, DATA_CHANNEL, DATA_FEATURE, DATA_FEATURE ...].
x_scale : tensor(float)
Input scale. It's a scalar, which means a per-tensor/layer quantization.
x_zero_point (optional) : T
Input zero point. Default value is 0 if it's not specified. It's a scalar, which means a per-tensor/layer quantization.
y_scale : tensor(float)
Output scale. It's a scalar, which means a per-tensor/layer quantization.
y_zero_point (optional) : T
Output zero point. Default value is 0 if it's not specified. It's a scalar, which means a per-tensor/layer quantization.
#### Outputs
Y : T
Output data tensor from average or max pooling across the input tensor. Dimensions will vary based on various kernel, stride, and pad sizes. Floor value of the dimension is used
#### Type Constraints
T : tensor(uint8), tensor(int8)
Constrain input and output types to 8 bit tensors.
### **com.microsoft.QLinearMul** Performs element-wise binary multiplication on 8 bit data types (with Numpy-style broadcasting support). C = ((A - A_zero_point) * (B - B_zero_point)) * (A_scale * B_scale)/C_scale + C_zero_point #### Version This version of the operator has been available since version 1 of the 'com.microsoft' operator set. #### Inputs (7 - 8)
A : T
First operand.
A_scale : tensor(float)
Input A's scale. It's a scalar, which means a per-tensor/layer quantization.
A_zero_point (optional) : T
Input A zero point. Default value is 0 if it's not specified. It's a scalar, which means a per-tensor/layer quantization.
B : T
Second operand.
B_scale : tensor(float)
Input B's scale. It's a scalar, which means a per-tensor/layer quantization.
B_zero_point (optional) : T
Input B zero point. Default value is 0 if it's not specified. It's a scalar, which means a per-tensor/layer quantization.
C_scale : tensor(float)
Output scale. It's a scalar, which means a per-tensor/layer quantization.
C_zero_point (optional) : T
Output zero point. Default value is 0 if it's not specified. It's a scalar, which means a per-tensor/layer quantization.
#### Outputs
C : T
Result, has same element type as two inputs
#### Type Constraints
T : tensor(uint8), tensor(int8)
Constrain input and output types to 8 bit signed and unsigned tensors.
### **com.microsoft.QLinearReduceMean** Computes the mean of the low-precision input tensor's element along the provided axes. The resulting tensor has the same rank as the input if keepdims equal 1. If keepdims equal 0, then the resulting tensor have the reduced dimension pruned. The above behavior is similar to numpy, with the exception that numpy default keepdims to False instead of True. Input and Output scales and zero points are used to requantize the output in a new range. This helps to improve accuracy as after ReduceMean operation the range of the output is expected to decrease. ``` "Output = Dequantize(Input) -> ReduceMean on fp32 data -> Quantize(output)", ``` #### Version This version of the operator has been available since version 1 of the 'com.microsoft' operator set. #### Attributes
axes : list of ints (required)
A list of integers, along which to reduce. The default is to reduce over all the dimensions of the input tensor.
keepdims : int (required)
Keep the reduced dimension or not, default 1 mean keep reduced dimension.
#### Inputs (4 - 5)
data : T
An input tensor.
data_scale : tensor(float)
Input scale. It's a scalar, which means a per-tensor/layer quantization.
data_zero_point (optional) : T
Input zero point. Default value is 0 if it's not specified. It's a scalar, which means a per-tensor/layer quantization.
reduced_scale : tensor(float)
Output scale. It's a scalar, which means a per-tensor/layer quantization.
reduced_zero_point (optional) : T
Output zero point. Default value is 0 if it's not specified. It's a scalar, which means a per-tensor/layer quantization.
#### Outputs
reduced : T
Reduced output tensor.
#### Type Constraints
T : tensor(uint8), tensor(int8)
Constrain input types to 8 bit signed and unsigned tensors.
### **com.microsoft.QuantizeLinear** The linear quantization operator. It consumes a full precision data, a scale, a zero point to compute the low precision / quantized tensor. The quantization formula is y = saturate ((x / y_scale) + y_zero_point).For saturation, it saturates to [0, 255] if it's uint8, or [-128, 127] if it's int8. For (x / y_scale), it's rounding to nearest ties to even. Refer to https://en.wikipedia.org/wiki/Rounding for details. Scale and zero point must have same shape. They must be either scalar (per tensor) or 1-D tensor (per 'axis'). #### Version This version of the operator has been available since version 1 of the 'com.microsoft' operator set. #### Attributes
axis : int
The axis along which same quantization parameters are applied. It's optional.If it's not specified, it means per-tensor quantization and input 'x_scale' and 'x_zero_point' must be scalars.If it's specified, it means per 'axis' quantization and input 'x_scale' and 'x_zero_point' must be 1-D tensors.
#### Inputs
x : T1
N-D full precision Input tensor to be quantized.
y_scale : T1
Scale for doing quantization to get 'y'. It could be a scalar or a 1-D tensor,which means a per-tensor or per-axis quantization. If it's a 1-D tensor, its number of elements should be equal to the dimension value of 'axis' dimension of input 'x'.
y_zero_point : T2
Zero point for doing quantization to get 'y'. It could be a scalar or a 1-D tensor, which means a per-tensoror per-axis quantization. If it's a 1-D tensor, its number of elements should be equal to the dimension value of 'axis' dimension of input 'x'.
#### Outputs
y : T2
N-D quantized output tensor. It has same shape as input 'x'.
#### Type Constraints
T1 : tensor(float16), tensor(float)
Constrain 'x', 'y_scale' to float tensors.
T2 : tensor(int8), tensor(uint8)
Constrain 'y_zero_point' and 'y' to 8-bit integer tensors.
### **com.microsoft.Range** Creates a sequence of numbers that begins at `start` and extends by increments of `delta` up to but not including `limit`. #### Version This version of the operator has been available since version 1 of the 'com.microsoft' operator set. #### Inputs (2 - 3)
start : T
Tensor(scalar, or dims=[1]). First entry in the range.
limit : T
Tensor(scalar, or dims=[1]). Upper limit of sequence, exclusive.
delta (optional) : T
Tensor(scalar, or dims=[1]). Number that increments start. Defaults to 1.
#### Outputs
Y : T
1-D Tensor of the range.
#### Type Constraints
T : tensor(float), tensor(double), tensor(int16), tensor(int32), tensor(int64)
Constrain input and output types.
### **com.microsoft.ReduceSumInteger** Computes the sum of the low-precision input tensor's element along the provided axes. The resulting tensor has the same rank as the input if keepdims equal 1. If keepdims equal 0, then the resulting tensor have the reduced dimension pruned. The above behavior is similar to numpy, with the exception that numpy default keepdims to False instead of True. #### Version This version of the operator has been available since version 1 of the 'com.microsoft' operator set. #### Attributes
axes : list of ints (required)
A list of integers, along which to reduce. The default is to reduce over all the dimensions of the input tensor.
keepdims : int (required)
Keep the reduced dimension or not, default 1 mean keep reduced dimension.
#### Inputs
data : T1
An input tensor.
#### Outputs
reduced : T2
Reduced output tensor.
#### Type Constraints
T1 : tensor(int8), tensor(uint8)
Constrain input type to 8-bit integer tensor.
T2 : tensor(int32), tensor(uint32)
Constrain output data type to 32-bit integer tensor.T2 must be tensor(uint32) when T1 is tensor(uint8),or must be tensor(int32) when T1 is tensor(int8).
### **com.microsoft.Rfft** #### Version This version of the operator has been available since version 1 of the 'com.microsoft' operator set. #### Attributes
normalized : int
onesided : int
signal_ndim : int (required)
#### Inputs
X : T
input tensor
#### Outputs
Y : T
output tensor
#### Type Constraints
T : tensor(float), tensor(double), tensor(float16)
Constrain input and output types to float or half tensors.
### **com.microsoft.SampleOp** Sample echo operator. #### Version This version of the operator has been available since version 1 of the 'com.microsoft' operator set. #### Inputs
X : T
input
#### Outputs
Y : T
output
#### Type Constraints
T : tensor(uint32), tensor(uint64), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double)
Constrain to any tensor type. If the dtype attribute is not provided this must be a valid output type.
### **com.microsoft.Tokenizer** Tokenizer divides each string in X into a vector of strings along the last axis. Allowed input shapes are [C] and [N, C]. If the maximum number of tokens found per input string is D, the output shape would be [N, C, D] when input shape is [N, C]. Similarly, if input shape is [C] then the output should be [C, D]. Tokenizer has two different operation modes. The first mode is selected when "tokenexp" is not set and "separators" is set. If "tokenexp" is set and "separators" is not set, the second mode will be used. The first mode breaks each input string into tokens by matching and removing separators. "separators" is a list of strings which are regular expressions. "tokenexp" is a single regular expression. Let's assume "separators" is [" "] and consider an example. If input is ["Hello World", "I love computer science !"] whose shape is [2], then the output would be [["Hello", "World", padvalue, padvalue, padvalue], ["I", "love", "computer", "science", "!"]] whose shape is [2, 5] because you can find at most 5 tokens per input string. Note that the input at most can have two axes, so 3-D and higher dimension are not supported. If "separators" contains a single empty string, the Tokenizer will enter into character tokenezation mode. This means all strings will be broken part into individual characters. For each input string, the second mode searches matches of "tokenexp" and each match will be a token in Y. The matching of "tokenexp" is conducted greedily (i.e., a match should be as long as possible). This operator searches for the first match starting from the beginning of the considered string, and then launches another search starting from the first remained character after the first matched token. If no match found, this operator will remove the first character from the remained string and do another search. This procedure will be repeated until reaching the end of the considered string. Let's consider another example to illustrate the effect of setting "mark" to true. If input is ["Hello", "World"], then the corresponding output would be [0x02, "Hello", "World", 0x03]. This implies that if mark is true, [C]/[N, C] - input's output shape becomes [C, D+2]/[N, C, D+2]. If tokenizer removes the entire content of [C]-input, it will produce [[]]. I.e. the output shape should be [C][0] or [N][C][0] if input shape was [N][C]. If the tokenizer receives empty input of [0] then the output is [0] if empty input of [N, 0] then [N, 0]. #### Version This version of the operator has been available since version 1 of the 'com.microsoft' operator set. #### Attributes
mark : int (required)
Boolean whether to mark the beginning/end character with start of text character (0x02)/end of text character (0x03).
mincharnum : int (required)
Minimum number of characters allowed in the output. For example, if mincharnum is 2, tokens such as "A" and "B" would be ignored
pad_value : string (required)
The string used to pad output tensors when the tokens extracted doesn't match the maximum number of tokens found. If start/end markers are needed, padding will appear outside the markers.
separators : list of strings
an optional list of strings attribute that contains a list of separators - regular expressions to match separators Two consecutive segments in X connected by a separator would be divided into two tokens. For example, if the input is "Hello World!" and this attribute contains only one space character, the corresponding output would be ["Hello", "World!"]. To achieve character-level tokenization, one should set the 'separators' to [""], which contains an empty string.
tokenexp : string
An optional string. Token's regular expression in basic POSIX format (http://pubs.opengroup.org/onlinepubs/9699919799/basedefs/V1_chap09.html#tag_09_03). If set, tokenizer may produce tokens matching the specified pattern. Note that one and only of 'tokenexp' and 'separators' should be set.
#### Inputs
X : T
Strings to tokenize
#### Outputs
Y : T
Tokenized strings
#### Type Constraints
T : tensor(string)
Input/Output is a string tensor
### **com.microsoft.Unique** Finds all the unique values (deduped list) present in the given input tensor. This operator returns 3 outputs. The first output tensor 'uniques' contains all of the unique elements of the input, sorted in the same order that they occur in the input. The second output tensor 'idx' is the same size as the input and it contains the index of each value of the input in 'uniques'. The third output tensor 'counts' contains the count of each element of 'uniques' in the input. Example: input_x = [2, 1, 1, 3, 4, 3] output_uniques = [2, 1, 3, 4] output_idx = [0, 1, 1, 2, 3, 2] output_counts = [1, 2, 2, 1] #### Version This version of the operator has been available since version 1 of the 'com.microsoft' operator set. #### Inputs
x : T
A 1-D input tensor that is to be processed.
#### Outputs
y : T
A 1-D tensor of the same type as 'x' containing all the unique values in 'x' sorted in the same order that they occur in the input 'x'
idx : tensor(int64)
A 1-D INT64 tensor of the same size as 'x' containing the indices for each value in 'x' in the output 'uniques'
counts : tensor(int64)
A 1-D INT64 tensor containing the the count of each element of 'uniques' in the input 'x'
#### Type Constraints
T : tensor(uint8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(int8), tensor(int16), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double), tensor(string), tensor(bool), tensor(complex64), tensor(complex128)
Input can be of any tensor type.
### **com.microsoft.WordConvEmbedding** The WordConvEmbedding takes in a batch of sequence words and embed each word to a vector. #### Version This version of the operator has been available since version 1 of the 'com.microsoft' operator set. #### Attributes
char_embedding_size : int
Integer representing the embedding vector size for each char.If not provide, use the char embedding size of embedding vector.
conv_window_size : int
This operator applies convolution to word from left to right with window equal to conv_window_size and stride to 1.Take word 'example' for example, with conv_window_size equal to 2, conv is applied to [ex],[xa], [am], [mp]...If not provide, use the first dimension of conv kernal shape.
embedding_size : int
Integer representing the embedding vector size for each word.If not provide, use the fileter size of conv weight
#### Inputs
Sequence : T
Specify batchs of sequence words to embedding
W : T1
Specify weights of conv
B : T1
Specify bias of conv
C : T1
Specify embedding vector of char
#### Outputs
Y : T1
output
#### Type Constraints
T : tensor(int32)
Constrain to tensor(int32).
T1 : tensor(float)
Constrain to tensor(float).
### experimental **com.microsoft.Attention** Multi-Head Self Attention that can be either unidirectional (like GPT2) or bidirectional (like BERT). The mask_index input is optional. Unidirectional and mask_index input are mutually exclusive. When unidirectional is 1, the mask_index shall not be provided. #### Version No versioning maintained for experimental ops. #### Attributes
num_heads : int (required)
Number of attention heads
unidirectional : int
Whether every token can only attend to previous tokens. Default value is 0.
#### Inputs (3 - 4)
input : T
3D input tensor with shape (batch_size, sequence_length, hidden_size), hidden_size = num_heads * head_size
weight : T
2D input tensor with shape (hidden_size, 3 * hidden_size)
bias : T
1D input tensor with shape (3 * hidden_size)
mask_index (optional) : M
Attention mask index with shape (batch_size).
#### Outputs
output : T
3D output tensor with shape (batch_size, sequence_length, hidden_size)
#### Type Constraints
T : tensor(float), tensor(float16)
Constrain input and output types to float tensors.
M : tensor(int32)
Constrain mask index to integer types
### experimental **com.microsoft.BiasGelu** Bias Gelu. It's an extension of Gelu. It takes the sum of input A and bias input B as the input of Gelu activation. #### Version No versioning maintained for experimental ops. #### Inputs
A : T
The normal input data.
B : T
The bias input data that is a 1D tensor.
#### Outputs
C : T
The output.
#### Type Constraints
T : tensor(float16), tensor(float), tensor(double)
Constrain input and output types to float tensors.
### experimental **com.microsoft.EmbedLayerNormalization** EmbedLayerNormalization is the fusion of embedding layer in BERT model, with optional mask processing. The embedding layer takes input_ids (word IDs) and segment_ids (sentence IDs) to look up word_embedding, position_embedding, and segment_emedding; the embeddings are added then applied layer normalization using gamma and beta tensors. The last input mask is optional. If mask is provided, mask index (that is position of first 0 in mask, or number of words) will be calculated. #### Version No versioning maintained for experimental ops. #### Inputs (7 - 8)
input_ids : T1
2D words IDs with shape (batch_size, sequence_length)
segment_ids : T1
2D segment IDs with shape (batch_size, sequence_length)
word_embedding : T
2D with shape (,hidden_size)
position_embedding : T
2D with shape (, hidden_size)
segment_embedding : T
2D with shape (, hidden_size)
gamma : T
1D gamma tensor for layer normalization with shape (hidden_size)
beta : T
1D beta tensor for layer normalization with shape (hidden_size)
mask (optional) : T1
2D attention mask with shape (batch_size, sequence_length)
#### Outputs
output : T
3D output tensor with shape (batch_size, sequence_length, hidden_size)
mask_index : T1
1D mask_index tensor with shape (batch_size)
#### Type Constraints
T1 : tensor(int32)
Constrain input and output integer tensors types
T : tensor(float), tensor(float16)
Constrain input and output float tensors types.
### experimental **com.microsoft.FastGelu** GELU (Gaussian Error Linear Unit) approximation: Y=0.5*X*(1+tanh(0.797885*X+0.035677*X*X*X))) with an optional input of bias that will be added to X before GELU. #### Version No versioning maintained for experimental ops. #### Inputs (1 - 2)
X : T
input tensor
bias (optional) : T
bias tensor
#### Outputs
Y : T
output tensor
#### Type Constraints
T : tensor(float), tensor(float16)
Constrain input and output types to float or half tensors.
### experimental **com.microsoft.Gelu** Gaussian Error Linear Unit. A high-performing neural network activation function.The GELU nonlinearity is the expected transformation of a stochastic regularizer which randomly applies the identity or zero map to a neuron's input. The GELU nonlinearity weights inputs by their magnitude, rather than gates inputs by their sign as in ReLUs. #### Version No versioning maintained for experimental ops. #### Inputs
X : T
The input data as Tensor.
#### Outputs
Y : T
The output.
#### Type Constraints
T : tensor(float16), tensor(float), tensor(double)
Constrain input and output types to float tensors.
### experimental **com.microsoft.SkipLayerNormalization** Skip and Layer Normalization Fusion #### Version No versioning maintained for experimental ops. #### Inputs (4 - 5)
input : T
3D input tensor with shape (batch_size, sequence_length, hidden_size)
skip : T
3D skip tensor with shape (batch_size, sequence_length, hidden_size)
gamma : T
1D input tensor with shape (hidden_size)
beta : T
1D skip tensor with shape (hidden_size
bias (optional) : T
1D bias tensor with shape (hidden_size
#### Outputs (1 - 3)
output : T
3D output tensor with shape (batch_size, sequence_length, hidden_size)
mean (optional) : U
Saved mean used during training to speed up gradient computation
inv_std_var (optional) : U
Saved inverse standard variance used during training to speed up gradient computation.
#### Type Constraints
T : tensor(float), tensor(float16)
Constrain input and output types to float or half tensors.
U : tensor(float)
Constrain mean and inv_std_var to float tensors.
## com.microsoft.nchwc ### **com.microsoft.nchwc.AveragePool** For internal use. #### Version This version of the operator has been available since version 1 of the 'com.microsoft.nchwc' operator set. #### Attributes
auto_pad : string
ceil_mode : int
count_include_pad : int
dilations : list of ints
kernel_shape : list of ints (required)
pads : list of ints
strides : list of ints
#### Inputs
X : T
#### Outputs
Y : T
#### Type Constraints
T : tensor(float)
Constrain input and output types to float tensors
### **com.microsoft.nchwc.Conv** For internal use. #### Version This version of the operator has been available since version 1 of the 'com.microsoft.nchwc' operator set. #### Attributes
activation : string
activation_params : list of floats
auto_pad : string
dilations : list of ints
group : int
kernel_shape : list of ints
pads : list of ints
strides : list of ints
#### Inputs (2 - 4)
X : T
W : T
B (optional) : T
Sum (optional) : T
#### Outputs
Y : T
#### Type Constraints
T : tensor(float)
Constrain input and output types to float tensors
### **com.microsoft.nchwc.GlobalAveragePool** For internal use. #### Version This version of the operator has been available since version 1 of the 'com.microsoft.nchwc' operator set. #### Inputs
X : T
#### Outputs
Y : T
#### Type Constraints
T : tensor(float)
Constrain input and output types to float tensors
### **com.microsoft.nchwc.GlobalMaxPool** For internal use. #### Version This version of the operator has been available since version 1 of the 'com.microsoft.nchwc' operator set. #### Inputs
X : T
#### Outputs
Y : T
#### Type Constraints
T : tensor(float)
Constrain input and output types to float tensors
### **com.microsoft.nchwc.MaxPool** For internal use. #### Version This version of the operator has been available since version 1 of the 'com.microsoft.nchwc' operator set. #### Attributes
auto_pad : string
ceil_mode : int
dilations : list of ints
kernel_shape : list of ints (required)
pads : list of ints
storage_order : int
strides : list of ints
#### Inputs
X : T
#### Outputs
Y : T
#### Type Constraints
T : tensor(float)
Constrain input and output types to float tensors
### **com.microsoft.nchwc.ReorderInput** For internal use. #### Version This version of the operator has been available since version 1 of the 'com.microsoft.nchwc' operator set. #### Inputs
X : T
#### Outputs
Y : T
#### Type Constraints
T : tensor(float)
Constrain input and output types to float tensors
### **com.microsoft.nchwc.ReorderOutput** For internal use. #### Version This version of the operator has been available since version 1 of the 'com.microsoft.nchwc' operator set. #### Attributes
channels : int
channels_last : int
#### Inputs
X : T
#### Outputs
Y : T
#### Type Constraints
T : tensor(float)
Constrain input and output types to float tensors
### **com.microsoft.nchwc.Upsample** For internal use. #### Version This version of the operator has been available since version 1 of the 'com.microsoft.nchwc' operator set. #### Attributes
scales : list of ints
#### Inputs
X : T
#### Outputs
Y : T
#### Type Constraints
T : tensor(float)
Constrain input and output types to float tensors