NeuralNetworks

NO fused activation function.

Fused ReLU activation function.

Fused ReLU1 activation function.

Fused ReLU6 activation function.

A 32 bit floating point scalar value.

A signed 32 bit integer scalar value.

A tensor of 32 bit floating point values.

A tensor of 32 bit integer values.

A tensor of 8 bit integers that represent real numbers.

Attached to this tensor are two numbers that can be used to convert the 8 bit integer to the real value and vice versa. These two numbers are:

• scale: a 32 bit floating point value greater than zero.
• zeroPoint: a 32 bit integer, in range [0, 255].

The formula is: real_value = (integer_value - zeroPoint) * scale.

An unsigned 32 bit integer scalar value.

Adds two tensors, element-wise.

Takes two input tensors of identical OperandCode and compatible dimensions. The output is the sum of both input tensors, optionally modified by an activation function.

Two dimensions are compatible when:

1. they are equal, or
2. one of them is 1

The size of the output is the maximum size along each dimension of the input operands. It starts with the trailing dimensions, and works its way forward.

Example:

input1.dimension = {4, 1, 2}
input2.dimension = {5, 4, 3, 1}
output.dimension = {5, 4, 3, 2}

Supported tensor OperandCode:

• ANEURALNETWORKS_TENSOR_FLOAT32
• ANEURALNETWORKS_TENSOR_QUANT8_ASYMM

Supported tensor rank: up to 4

Inputs:

• 0: A tensor.
• 1: A tensor of the same OperandCode, and compatible dimensions as input0.
• 2: An ANEURALNETWORKS_INT32 scalar, and has to be one of the FuseCode values. Specifies the activation to invoke on the result.

Outputs:

• 0: The sum, a tensor of the same OperandCode as input0.

Available since API level 27.

Performs a 2-D average pooling operation.

The output dimensions are functions of the filter dimensions, stride, and padding.

The values in the output tensor are computed as:

output[batch, row, col, channel] =
    sum_{i, j}(input[batch, row + i, col + j, channel]) / sum(1)

Supported tensor OperandCode:

• ANEURALNETWORKS_TENSOR_FLOAT32
• ANEURALNETWORKS_TENSOR_QUANT8_ASYMM

Supported tensor rank: 4, with "NHWC" (i.e., Num_samples, Height, Width, and Channels) data layout.

Both explicit padding and implicit padding are supported.

Inputs (explicit padding):

• 0: A 4-D tensor, of shape [batches, height, width, depth], specifying the input.
• 1: An ANEURALNETWORKS_INT32 scalar, specifying the padding on the left, in the ‘width’ dimension.
• 2: An ANEURALNETWORKS_INT32 scalar, specifying the padding on the right, in the ‘width’ dimension.
• 3: An ANEURALNETWORKS_INT32 scalar, specifying the padding on the top, in the ‘height’ dimension.
• 4: An ANEURALNETWORKS_INT32 scalar, specifying the padding on the bottom, in the ‘height’ dimension.
• 5: An ANEURALNETWORKS_INT32 scalar, specifying the stride when walking through input in the ‘width’ dimension.
• 6: An ANEURALNETWORKS_INT32 scalar, specifying the stride when walking through input in the ‘height’ dimension.
• 7: An ANEURALNETWORKS_INT32 scalar, specifying the filter width.
• 8: An ANEURALNETWORKS_INT32 scalar, specifying the filter height.
• 9: An ANEURALNETWORKS_INT32 scalar, and has to be one of the FuseCode values. Specifies the activation to invoke on the result.

Inputs (implicit padding):

• 0: A 4-D tensor, of shape [batches, height, width, depth], specifying the input.
• 1: An ANEURALNETWORKS_INT32 scalar, specifying the implicit padding scheme, has to be one of the PaddingCode values.
• 2: An ANEURALNETWORKS_INT32 scalar, specifying the stride when walking through input in the ‘width’ dimension.
• 3: An ANEURALNETWORKS_INT32 scalar, specifying the stride when walking through input in the ‘height’ dimension.
• 4: An ANEURALNETWORKS_INT32 scalar, specifying the filter width.
• 5: An ANEURALNETWORKS_INT32 scalar, specifying the filter height.
• 6: An ANEURALNETWORKS_INT32 scalar, and has to be one of the FuseCode values. Specifies the activation to invoke on the result.

Outputs:

• 0: The output 4-D tensor, of shape [batches, out_height, out_width, depth].

Available since API level 27.

BatchToSpace for N-dimensional tensors.

This operation reshapes the batch dimension (dimension 0) into M + 1 dimensions of shape block_shape + [batch], interleaves these blocks back into the grid defined by the spatial dimensions [1, ..., M], to obtain a result with the same rank as the input.

This is the reverse of SpaceToBatch.

Supported tensor OperandCode:

• ANEURALNETWORKS_TENSOR_FLOAT32
• ANEURALNETWORKS_TENSOR_QUANT8_ASYMM

Supported tensor rank: 4

Inputs:

• 0: An n-D tensor, specifying the tensor to be reshaped
• 1: A 1-D Tensor of ANEURALNETWORKS_TENSOR_INT32, the block sizes for each spatial dimension of the input tensor. All values must be >= 1.

Outputs:

• 0: A tensor of the same OperandCode as input0.

Available since API level 28.

Concatenates the input tensors along the given dimension.

The input tensors must have identical OperandCode and the same dimensions except the dimension along the concatenation axis.

Supported tensor OperandCode:

• ANEURALNETWORKS_TENSOR_FLOAT32
• ANEURALNETWORKS_TENSOR_QUANT8_ASYMM

Supported tensor rank: up to 4

Inputs:

• 0 ~ n-1: The list of n input tensors, of shape [D0, D1, ..., Daxis(i), ..., Dm]. For inputs of ANEURALNETWORKS_TENSOR_QUANT8_ASYMM, all input tensors must have the same scale and zeroPoint.
• n: An ANEURALNETWORKS_INT32 scalar, specifying the concatenation axis.

Outputs:

• 0: The output, a tensor of the same OperandCode as the input tensors. The output shape is [D0, D1, ..., sum(Daxis(i)), ..., Dm].

Available since API level 27.

Performs an 2-D convolution operation.

The CONV_2D op sweeps a 2-D filter that can mix channels together over a batch of images, applying the filter to each window of each image of the appropriate size.

The output dimensions are functions of the filter dimensions, stride, and padding.

The values in the output tensor are computed as:

output[batch, row, col, channel] =
    sum_{i, j} (
        input[batch, row + i, col + j, k] *
        filter[channel, row + i, col + j, k] +
        bias[channel]
    )

Supported tensor OperandCode:

• ANEURALNETWORKS_TENSOR_FLOAT32
• ANEURALNETWORKS_TENSOR_QUANT8_ASYMM

Supported tensor rank: 4, with "NHWC" data layout.

Both explicit padding and implicit padding are supported.

Inputs (explicit padding):

• 0: A 4-D tensor, of shape [batches, height, width, depth_in], specifying the input.
• 1: A 4-D tensor, of shape [depth_out, filter_height, filter_width, depth_in], specifying the filter.
• 2: A 1-D tensor, of shape [depth_out], specifying the bias. For input tensor of ANEURALNETWORKS_TENSOR_FLOAT32, the bias should also be of ANEURALNETWORKS_TENSOR_FLOAT32. For input tensor of ANEURALNETWORKS_TENSOR_QUANT8_ASYMM, the bias should be of ANEURALNETWORKS_TENSOR_INT32, with zeroPoint of 0 and bias_scale == input_scale * filter_scale.
• 3: An ANEURALNETWORKS_INT32 scalar, specifying the padding on the left, in the ‘width’ dimension.
• 4: An ANEURALNETWORKS_INT32 scalar, specifying the padding on the right, in the ‘width’ dimension.
• 5: An ANEURALNETWORKS_INT32 scalar, specifying the padding on the top, in the ‘height’ dimension.
• 6: An ANEURALNETWORKS_INT32 scalar, specifying the padding on the bottom, in the ‘height’ dimension.
• 7: An ANEURALNETWORKS_INT32 scalar, specifying the stride when walking through input in the ‘width’ dimension.
• 8: An ANEURALNETWORKS_INT32 scalar, specifying the stride when walking through input in the ‘height’ dimension.
• 9: An ANEURALNETWORKS_INT32 scalar, and has to be one of the FuseCode values. Specifies the activation to invoke on the result.

Inputs (implicit padding):

• 0: A 4-D tensor, of shape [batches, height, width, depth_in], specifying the input.
• 1: A 4-D tensor, of shape [depth_out, filter_height, filter_width, depth_in], specifying the filter.
• 2: A 1-D tensor, of shape [depth_out], specifying the bias. For input tensor of ANEURALNETWORKS_TENSOR_FLOAT32, the bias should also be of ANEURALNETWORKS_TENSOR_FLOAT32. For input tensor of ANEURALNETWORKS_TENSOR_QUANT8_ASYMM, the bias should be of ANEURALNETWORKS_TENSOR_INT32, with zeroPoint of 0 and bias_scale == input_scale * filter_scale.
• 3: An ANEURALNETWORKS_INT32 scalar, specifying the implicit padding scheme, has to be one of the PaddingCode values.
• 4: An ANEURALNETWORKS_INT32 scalar, specifying the stride when walking through input in the ‘width’ dimension.
• 5: An ANEURALNETWORKS_INT32 scalar, specifying the stride when walking through input in the ‘height’ dimension.
• 6: An ANEURALNETWORKS_INT32 scalar, and has to be one of the FuseCode values. Specifies the activation to invoke on the result.

Outputs:

• 0: The output 4-D tensor, of shape [batches, out_height, out_width, depth_out]. For output tensor of ANEURALNETWORKS_TENSOR_QUANT8_ASYMM, the following condition must be satisfied: output_scale > input_scale * filter_scale.

Available since API level 27.

Performs a depthwise 2-D convolution operation.

Given an input tensor of shape [batches, height, width, depth_in] and a filter tensor of shape [1, filter_height, filter_width, depth_out] containing depth_out convolutional filters of depth 1, DEPTHWISE_CONV applies a different filter to each input channel (expanding from 1 channel to channel_multiplier channels for each), then concatenates the results together.

The output has depth_out = depth_in * depth_multiplier channels. The output dimensions are functions of the filter dimensions, stride, and padding.

The values in the output tensor are computed as:

output[b, i, j, k * channel_multiplier + q] =
    sum_{di, dj} (
        input[b, strides[1] * i + di, strides[2] * j + dj, k] *
        filter[1, di, dj, k * channel_multiplier + q]
    )

Supported tensor OperandCode:

• ANEURALNETWORKS_TENSOR_FLOAT32
• ANEURALNETWORKS_TENSOR_QUANT8_ASYMM

Supported tensor rank: 4, with "NHWC" data layout.

Both explicit padding and implicit padding are supported.

Inputs (explicit padding):

• 0: A 4-D tensor, of shape [batches, height, width, depth_in], specifying the input.
• 1: A 4-D tensor, of shape [1, filter_height, filter_width, depth_out], specifying the filter.
• 2: A 1-D tensor, of shape [depth_out], specifying the bias. For input tensor of ANEURALNETWORKS_TENSOR_FLOAT32, the bias should also be of ANEURALNETWORKS_TENSOR_FLOAT32. For input tensor of ANEURALNETWORKS_TENSOR_QUANT8_ASYMM, the bias should be of ANEURALNETWORKS_TENSOR_INT32, with zeroPoint of 0 and bias_scale == input_scale * filter_scale.
• 3: An ANEURALNETWORKS_INT32 scalar, specifying the padding on the left, in the ‘width’ dimension.
• 4: An ANEURALNETWORKS_INT32 scalar, specifying the padding on the right, in the ‘width’ dimension.
• 5: An ANEURALNETWORKS_INT32 scalar, specifying the padding on the top, in the ‘height’ dimension.
• 6: An ANEURALNETWORKS_INT32 scalar, specifying the padding on the bottom, in the ‘height’ dimension.
• 7: An ANEURALNETWORKS_INT32 scalar, specifying the stride when walking through input in the ‘width’ dimension.
• 8: An ANEURALNETWORKS_INT32 scalar, specifying the stride when walking through input in the ‘height’ dimension.
• 9: An ANEURALNETWORKS_INT32 scalar, specifying the depthwise multiplier.
• 10: An ANEURALNETWORKS_INT32 scalar, and has to be one of the FuseCode values. Specifies the activation to invoke on the result.

Inputs (implicit padding):

• 0: A 4-D tensor, of shape [batches, height, width, depth_in], specifying the input.
• 1: A 4-D tensor, of shape [1, filter_height, filter_width, depth_out], specifying the filter.
• 2: A 1-D tensor, of shape [depth_out], specifying the bias. For input tensor of ANEURALNETWORKS_TENSOR_FLOAT32, the bias should also be of ANEURALNETWORKS_TENSOR_FLOAT32. For input tensor of ANEURALNETWORKS_TENSOR_QUANT8_ASYMM, the bias should be of ANEURALNETWORKS_TENSOR_INT32, with zeroPoint of 0 and bias_scale == input_scale * filter_scale.
• 3: An ANEURALNETWORKS_INT32 scalar, specifying the implicit padding scheme, has to be one of the PaddingCode values.
• 4: An ANEURALNETWORKS_INT32 scalar, specifying the stride when walking through input in the ‘width’ dimension.
• 5: An ANEURALNETWORKS_INT32 scalar, specifying the stride when walking through input in the ‘height’ dimension.
• 6: An ANEURALNETWORKS_INT32 scalar, specifying the depthwise multiplier.
• 7: An ANEURALNETWORKS_INT32 scalar, and has to be one of the FuseCode values. Specifies the activation to invoke on the result.

Outputs:

• 0: The output 4-D tensor, of shape [batches, out_height, out_width, depth_out]. For output tensor of ANEURALNETWORKS_TENSOR_QUANT8_ASYMM, the following condition must be satisfied: output_scale > input_scale * filter_scale.

Available since API level 27.

Rearranges data from depth into blocks of spatial data.

More specifically, this op outputs a copy of the input tensor where values from the depth dimension are moved in spatial blocks to the height and width dimensions. The value block_size indicates the input block size and how the data is moved.

Chunks of data of size block_size * block_size from depth are rearranged into non-overlapping blocks of size block_size x block_size.

The width of the output tensor is input_depth * block_size, whereas the height is input_height * block_size. The depth of the input tensor must be divisible by block_size * block_size

Supported tensor OperandCode:

• ANEURALNETWORKS_TENSOR_FLOAT32
• ANEURALNETWORKS_TENSOR_QUANT8_ASYMM

Supported tensor rank: 4, with "NHWC" data layout.

Inputs:

• 0: A 4-D tensor, of shape [batches, height, width, depth_in], specifying the input.
• 1: An ANEURALNETWORKS_INT32 scalar, specifying the block_size. block_size must be >=1 and block_size * block_size must be a divisor of the input depth.

Outputs:

• 0: The output 4-D tensor, of shape [batch, height*block_size, width*block_size, depth/(block_size*block_size)].

Available since API level 27.

Dequantizes the input tensor.

The formula is:

output = (input - zeroPoint) * scale.

Supported tensor OperandCode:

• ANEURALNETWORKS_TENSOR_QUANT8_ASYMM

Supported tensor rank: up to 4

Inputs:

• 0: A tensor of ANEURALNETWORKS_TENSOR_QUANT8_ASYMM.

Outputs:

• 0: The output tensor of same shape as input0, but with ANEURALNETWORKS_TENSOR_FLOAT32.

Available since API level 27.

Element-wise division of two tensors.

Takes two input tensors of identical OperandCode and compatible dimensions. The output is the result of dividing the first input tensor by the second, optionally modified by an activation function.

Two dimensions are compatible when:

1. they are equal, or
2. one of them is 1

The size of the output is the maximum size along each dimension of the input operands. It starts with the trailing dimensions, and works its way forward.

Example: input1.dimension = {4, 1, 2} input2.dimension = {5, 4, 3, 1} output.dimension = {5, 4, 3, 2}

Supported tensor OperandCode:

• ANEURALNETWORKS_TENSOR_FLOAT32

Supported tensor rank: up to 4

Inputs:

• 0: An n-D tensor, specifying the first input.
• 1: A tensor of the same OperandCode, and compatible dimensions as input0.
• 2: An ANEURALNETWORKS_INT32 scalar, and has to be one of the FuseCode values. Specifies the activation to invoke on the result.

Outputs:

• 0: A tensor of the same OperandCode as input0.

Available since API level 28.

Looks up sub-tensors in the input tensor.

This operator takes for input a tensor of values (Values) and a one-dimensional tensor of selection indices (Lookups). The output tensor is the concatenation of sub-tensors of Values as selected by Lookups.

Think of Values as being sliced along its first dimension: The entries in Lookups select which slices are concatenated together to create the output tensor.

For example, if Values has shape of [40, 200, 300] and Lookups has shape of [3], all three values found in Lookups are expected to be between 0 and 39. The resulting tensor must have shape of [3, 200, 300].

If a value in Lookups is out of bounds, the operation must fail and an error must be reported.

Inputs:

• 0: Lookups. A 1-D tensor of ANEURALNETWORKS_TENSOR_INT32. The values are indices into the first dimension of Values.
• 1: Values. An n-D tensor, where n >= 2, from which sub-tensors are extracted.

Output:

• 0: A n-D tensor with the same rank and shape as the Values tensor, except for the first dimension which has the same size as Lookups' only dimension.

Available since API level 27.

Computes element-wise floor() on the input tensor.

Supported tensor OperandCode:

• ANEURALNETWORKS_TENSOR_FLOAT32

Supported tensor rank: up to 4

Inputs:

• 0: A tensor.

Outputs:

• 0: The output tensor, of the same OperandCode and dimensions as the input tensor.

Available since API level 27.

Denotes a fully (densely) connected layer, which connects all elements in the input tensor with each element in the output tensor.

This layer implements the operation:

outputs = activation(inputs * weights’ + bias)

Supported tensor OperandCode:

• ANEURALNETWORKS_TENSOR_FLOAT32
• ANEURALNETWORKS_TENSOR_QUANT8_ASYMM

Supported tensor rank: up to 4.

Inputs:

• 0: A tensor of at least rank 2, specifying the input. If rank is greater than 2, then it gets flattened to a 2-D Tensor. The (flattened) 2-D Tensor is reshaped (if necessary) to [batch_size, input_size], where "input_size" corresponds to the number of inputs to the layer, matching the second dimension of weights, and "batch_size" is calculated by dividing the number of elements by "input_size".
• 1: A 2-D tensor, specifying the weights, of shape [num_units, input_size], where "num_units" corresponds to the number of output nodes.
• 2: A 1-D tensor, of shape [num_units], specifying the bias. For input tensor of ANEURALNETWORKS_TENSOR_FLOAT32, the bias should also be of ANEURALNETWORKS_TENSOR_FLOAT32. For input tensor of ANEURALNETWORKS_TENSOR_QUANT8_ASYMM, the bias should be of ANEURALNETWORKS_TENSOR_INT32, with zeroPoint of 0 and bias_scale == input_scale * filter_scale.
• 3: An ANEURALNETWORKS_INT32 scalar, and has to be one of the FuseCode values. Specifies the activation to invoke on the result.

Outputs:

• 0: The output tensor, of shape [batch_size, num_units]. For output tensor of ANEURALNETWORKS_TENSOR_QUANT8_ASYMM, the following condition must be satisfied: output_scale > input_scale * filter_scale.

Available since API level 27.

Looks up sub-tensors in the input tensor using a key-value map.

This operator takes for input a tensor of values (Values), a one-dimensional tensor of selection values (Lookups) and a one-dimensional tensor that maps these values to Values indexes. The output tensor is the concatenation of sub-tensors of Values as selected by Lookups via Keys.

Think of Values as being sliced along its outer-most dimension. The output is a concatenation of selected slices, with one slice for each entry of Lookups. The slice selected is the one at the same index as the Maps entry that matches the value in Lookups.

For a hit, the corresponding sub-tensor of Values is included in the Output tensor. For a miss, the corresponding sub-tensor in Output must have zero values.

For example, if Values has shape of [40, 200, 300], Keys should have a shape of [40]. If Lookups tensor has shape of [3], three slices are being concatenated, so the resulting tensor must have the shape of [3, 200, 300]. If the first entry in Lookups has the value 123456, that value must be located in Keys tensor. If the sixth entry of Keys contains 123456, the sixth slice of Values must be selected. If no entry in Keys has 123456, a slice of zeroes must be concatenated.

Inputs:

• 0: Lookups. A 1-D ANEURALNETWORKS_TENSOR_INT32 tensor with shape [ k ].
• 1: Keys. A 1-D ANEURALNETWORKS_TENSOR_INT32 tensor with shape [ n ]; Keys and Values pair represent a map, i.e., the ith element in Keys (Keys[i]) is the key to select the ith sub-tensor in Values (Values[i]), where 0 <= i <= n-1. Keys tensor MUST be sorted in ascending order.
• 2: Values. A tensor with shape of [ n, … ]; i.e., the first dimension must be n.

Outputs:

• 0: Output. A tensor with shape [ k …].
• 1: Hits. A boolean tensor with shape [ k ] indicates whether the lookup hits (True) or not (False). Stored as ANEURALNETWORKS_TENSOR_QUANT8_ASYMM with offset 0 and scale 1.0f. A non-zero byte represents True, a hit. A zero indicates otherwise.

Available since API level 27.

Applies L2 normalization along the depth dimension.

The values in the output tensor are computed as:

output[batch, row, col, channel] =
    input[batch, row, col, channel] /
    sqrt(sum_{c} pow(input[batch, row, col, c], 2))

For input tensor with more dimensions, independently normalizes each 1-D slice along dimension dim.

Supported tensor OperandCode:

• ANEURALNETWORKS_TENSOR_FLOAT32

Supported tensor rank: 4, with "NHWC" data layout (i.e., Num_samples, Height, Width, and Channels).

Inputs:

• 0: A 4-D tensor, of shape [batches, height, width, depth].

Outputs:

• 0: The output 4-D tensor, of the same shape as input [batches, height, width, depth].

Available since API level 27.

Performs an 2-D L2 pooling operation.

The output dimensions are functions of the filter dimensions, stride, and padding.

The values in the output tensor are computed as:

output[batch, row, col, channel] =
    sqrt(sum_{i, j} pow(input[batch, row + i, col + j, channel], 2) /
         sum(1))

Supported tensor OperandCode:

• ANEURALNETWORKS_TENSOR_FLOAT32

Supported tensor rank: 4, with "NHWC" data layout.

Both explicit padding and implicit padding are supported.

Inputs (explicit padding):

• 0: A 4-D tensor, of shape [batches, height, width, depth], specifying the input.
• 1: An ANEURALNETWORKS_INT32 scalar, specifying the padding on the left, in the ‘width’ dimension.
• 2: An ANEURALNETWORKS_INT32 scalar, specifying the padding on the right, in the ‘width’ dimension.
• 3: An ANEURALNETWORKS_INT32 scalar, specifying the padding on the top, in the ‘height’ dimension.
• 4: An ANEURALNETWORKS_INT32 scalar, specifying the padding on the bottom, in the ‘height’ dimension.
• 5: An ANEURALNETWORKS_INT32 scalar, specifying the stride when walking through input in the ‘width’ dimension.
• 6: An ANEURALNETWORKS_INT32 scalar, specifying the stride when walking through input in the ‘height’ dimension.
• 7: An ANEURALNETWORKS_INT32 scalar, specifying the filter width.
• 8: An ANEURALNETWORKS_INT32 scalar, specifying the filter height.
• 9: An ANEURALNETWORKS_INT32 scalar, and has to be one of the FuseCode values. Specifies the activation to invoke on the result.

Inputs (implicit padding):

• 0: A 4-D tensor, of shape [batches, height, width, depth], specifying the input.
• 1: An ANEURALNETWORKS_INT32 scalar, specifying the implicit padding scheme, has to be one of the PaddingCode values.
• 2: An ANEURALNETWORKS_INT32 scalar, specifying the stride when walking through input in the ‘width’ dimension.
• 3: An ANEURALNETWORKS_INT32 scalar, specifying the stride when walking through input in the ‘height’ dimension.
• 4: An ANEURALNETWORKS_INT32 scalar, specifying the filter width.
• 5: An ANEURALNETWORKS_INT32 scalar, specifying the filter height.
• 6: An ANEURALNETWORKS_INT32 scalar, and has to be one of the FuseCode values. Specifies the activation to invoke on the result.

Outputs:

• 0: The output 4-D tensor, of shape [batches, out_height, out_width, depth].

Available since API level 27.

Applies Local Response Normalization along the depth dimension.

The 4-D input tensor is treated as a 3-D array of 1-D vectors (along the last dimension), and each vector is normalized independently. Within a given vector, each component is divided by the weighted, squared sum of inputs within depth_radius.

The output is calculated using this formula:

sqr_sum[a, b, c, d] = sum(
    pow(input[a, b, c, d - depth_radius : d + depth_radius + 1], 2))
output = input / pow((bias + alpha * sqr_sum), beta)

Supported tensor OperandCode:

• ANEURALNETWORKS_TENSOR_FLOAT32

Supported tensor rank: 4, with "NHWC" data layout.

Inputs:

• 0: A 4-D tensor, of shape [batches, height, width, depth], specifying the input.
• 1: An ANEURALNETWORKS_INT32 scalar, specifying the radius of the normalization window.
• 2: An ANEURALNETWORKS_FLOAT32 scalar, specifying the bias, must not be zero.
• 3: An ANEURALNETWORKS_FLOAT32 scalar, specifying the scale factor, alpha.
• 4: An ANEURALNETWORKS_FLOAT32 scalar, specifying the exponent, beta.

Outputs:

• 0: The output tensor of same shape as input0.

Available since API level 27.

Computes sigmoid activation on the input tensor element-wise.

The output is calculated using this formula:

output = 1 / (1 + exp(-input))

Supported tensor OperandCode:

• ANEURALNETWORKS_TENSOR_FLOAT32
• ANEURALNETWORKS_TENSOR_QUANT8_ASYMM

Supported tensor rank: up to 4.

Inputs:

• 0: A tensor, specifying the input.

Outputs:

• 0: The output tensor of same shape as input0. For ANEURALNETWORKS_TENSOR_QUANT8_ASYMM, the scale must be 1.f / 256 and the zeroPoint must be 0.

Available since API level 27.

Projects an input to a bit vector via locality senstive hashing.

Inputs:

• 0: Hash functions. Dim.size == 2, DataType: Float. Tensor[0].Dim[0]: Number of hash functions. Tensor[0].Dim[1]: Number of seeds per hash functions. Tensor[0].Dim[1] <= 32 in sparse case.
• 1: Input. Dim.size >= 1, no restriction on DataType.
• 2: Weight. Optional. Dim.size == 1, DataType: Float. If not set, each input element is considered to have the same weight of 1.0. Tensor[1].Dim[0] == Tensor[2].Dim[0]
• 3: Type: Sparse: Value LSHProjectionType_SPARSE(=1). Computed bit vector is considered to be sparse. Each output element is an int32 made up of multiple bits computed from hash functions.Dense: Value LSHProjectionType_DENSE(=2). Computed bit vector is considered to be dense. Each output element represents a bit and can take the value of either 0 or 1.

Outputs:

• 0: If the projection type is sparse: Output.Dim == { Tensor[0].Dim[0] } A tensor of int32 that represents hash signatures. If the projection type is Dense: Output.Dim == { Tensor[0].Dim[0] * Tensor[0].Dim[1] } A flattened tensor that represents projected bit vectors.

Available since API level 27.

Performs a single time step in a Long Short-Term Memory (LSTM) layer.

The LSTM operation is described by the following equations.

\begin{eqnarray*} i_t =& \sigma(W_{xi}x_t+W_{hi}h_{t-1}+W_{ci}C_{t-1}+b_i) & \\ f_t =& \sigma(W_{xf}x_t+W_{hf}h_{t-1}+W_{cf}C_{t-1}+b_f) & \\ C_t =& clip(f_t \odot C_{t-1} + i_t \odot g(W_{xc}x_t+W_{hc}h_{t-1}+b_c),\ t_{cell}) & \\ o_t =& \sigma(W_{xo}x_t+W_{ho}h_{t-1}+W_{co}C_t+b_o) & \\ & & \\ & clip(W_{proj}(o_t \odot g(C_t))+b_{proj},\ t_{proj}) & if\ there\ is\ a\ projection; \\ h_t =& & \\ & o_t \odot g(C_t) & otherwise. \\ \end{eqnarray*} Where:

• $x_t$ is the input,
• $i_t$ is the input gate,
• $f_t$ is the forget gate,
• $C_t$ is the cell state,
• $o_t$ is the output,
• $h_t$ is the output state,
• $\sigma$ is the logistic sigmoid function,
• $g$ is the cell input and cell output activation function, usually $tahn$,
• $W_{xi}$ is the input-to-input weight matrix,
• $W_{hi}$ is the recurrent to input weight matrix,
• $W_{ci}$ is the cell-to-input weight matrix,
• $b_i$ is the input gate bias,
• $W_{xf}$ is the input-to-forget weight matrix,
• $W_{hf}$ is the recurrent-to-forget weight matrix,
• $W_{cf}$ is the cell-to-forget weight matrix,
• $b_f$ is the forget gate bias,
• $W_{xc}$ is the input-to-cell weight matrix,
• $W_{hc}$ is the recurrent-to-cell weight matrix,
• $b_c$ is the cell bias,
• $W_{xo}$ is the input-to-output weight matrix,
• $W_{ho}$ is the recurrent-to-output weight matrix,
• $W_{co}$ is the cell-to-output weight matrix,
• $b_o$ is the output gate bias,
• $W_{proj}$ is the projection weight matrix,
• $b_{proj}$ is the projection bias,
• $t_{cell}$ is the threshold for clipping the cell state, and
• $t_{proj}$ is the threshold for clipping the projected output.
• $\odot$ is the Hadamard product that takes two matrices and produces another matrix, each element of which is the product of the corresponding elements of the input matrices.

The operation has the following independently optional inputs:

• The input-to-input weights ( $W_{xi}$), recurrent-to-input weights ( $W_{hi}$), cell-to-input ( $W_{ci}$) weights, and input gate bias ( $b_i$) either all have values, or none of them have values (i.e., all set to null). If they have no values, coupling of input and forget gates (CIFG) is used, in which case the input gate ( $i_t$) is calculated using the following equation instead. \begin{eqnarray*} i_t = 1 - f_t \end{eqnarray*}
• The cell-to-forget weights ( $W_{cf}$) and cell-to-output weights ( $W_{co}$) either both have values or neither of them have values. If they have values, the peephole optimization is used. Additionally, if CIFG is not used, cell-to-input weights ( $W_{ci}$) is also required to have values for peephole optimization.
• The projection weights ( $W_{proj}$) is required only for the recurrent projection layer, and should otherwise have no value.
• The projection bias ( $b_{proj}$) may (but not required to) have a value if the recurrent projection layer exists, and should otherwise have no value.

References:

The default non-peephole non-CIFG implementation is based on: http://www.bioinf.jku.at/publications/older/2604.pdf S. Hochreiter and J. Schmidhuber. "Long Short-Term Memory". Neural Computation, 9(8):1735-1780, 1997.

The peephole implementation and projection layer is based on: https://research.google.com/pubs/archive/43905.pdf Hasim Sak, Andrew Senior, and Francoise Beaufays. "Long short-term memory recurrent neural network architectures for large scale acoustic modeling." INTERSPEECH, 2014. (However, the concept of peephole optimization was introduced in work prior to this paper.)

The coupling of input and forget gate (CIFG) is based on: http://arxiv.org/pdf/1503.04069.pdf Greff et al. "LSTM: A Search Space Odyssey"

Supported tensor OperandCode:

• ANEURALNETWORKS_TENSOR_FLOAT32

Inputs:

• 0: The input ( $x_t$). A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [batch_size, input_size], where “batch_size” corresponds to the batching dimension, and “input_size” is the size of the input.
• 1: The input-to-input weights ( $W_{xi}$). Optional. A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [num_units, input_size], where “num_units” corresponds to the number of cell units.
• 2: The input-to-forget weights ( $W_{xf}$). A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [num_units, input_size].
• 3: The input-to-cell weights ( $W_{xc}$). A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [num_units, input_size].
• 4: The input-to-output weights ( $W_{xo}$). A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [num_units, input_size].
• 5: The recurrent-to-input weights ( $W_{hi}$). Optional. A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [num_units, output_size], where “output_size” corresponds to either the number of cell units (i.e., “num_units”), or the second dimension of the “projection_weights”, if defined.
• 6: The recurrent-to-forget weights ( $W_{hf}$). A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [num_units, output_size].
• 7: The recurrent-to-cell weights ( $W_{hc}$). A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [num_units, output_size].
• 8: The recurrent-to-output weights ( $W_{ho}$). A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [num_units, output_size].
• 9: The cell-to-input weights ( $W_{ci}$). Optional. A 1-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [num_units].
• 10:The cell-to-forget weights ( $W_{cf}$). Optional. A 1-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [num_units].
• 11:The cell-to-output weights ( $W_{co}$). Optional. A 1-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [num_units].
• 12:The input gate bias ( $b_i$). Optional. A 1-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [num_units].
• 13:The forget gate bias ( $b_f$). A 1-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [num_units].
• 14:The cell bias ( $b_c$). A 1-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [num_units].
• 15:The output gate bias ( $b_o$). A 1-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [num_units].
• 16:The projection weights ( $W_{proj}$). Optional. A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [output_size, num_units].
• 17:The projection bias ( $b_{proj}$). Optional. A 1-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [output_size].
• 18:The output state (in) ( $h_{t-1}$). A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [batch_size, output_size].
• 19:The cell state (in) ( $C_{t-1}$). A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [batch_size, num_units].
• 20:The activation function ( $g$). A value indicating the activation function:
  • 0: None;
  • 1: Relu;
  • 3: Relu6;
  • 4: Tanh;
  • 6: Sigmoid.
• 21:The clipping threshold ( $t_{cell}$) for the cell state, such that values are bound within [-cell_clip, cell_clip]. If set to 0.0 then clipping is disabled.
• 22:The clipping threshold ( $t_{proj}$) for the output from the projection layer, such that values are bound within [-proj_clip, proj_clip]. If set to 0.0 then clipping is disabled.

Outputs:

• 0: The scratch buffer. A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [batch_size, num_units * 4] with CIFG, or [batch_size, num_units * 3] without CIFG.
• 1: The output state (out) ( $h_t$). A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [batch_size, output_size].
• 2: The cell state (out) ( $C_t$). A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [batch_size, num_units].
• 3: The output ( $o_t$). A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [batch_size, output_size]. This is effectively the same as the current “output state (out)” value.

Available since API level 27.

Performs an 2-D max pooling operation.

The output dimensions are functions of the filter dimensions, stride, and padding.

The values in the output tensor are computed as:

output[batch, row, col, channel] =
    max_{i, j} (input[batch, row + i, col + j, channel])

Supported tensor OperandCode:

• ANEURALNETWORKS_TENSOR_FLOAT32
• ANEURALNETWORKS_TENSOR_QUANT8_ASYMM

Supported tensor rank: 4, with "NHWC" data layout.

Both explicit padding and implicit padding are supported.

Inputs (explicit padding):

• 0: A 4-D tensor, of shape [batches, height, width, depth], specifying the input.
• 1: An ANEURALNETWORKS_INT32 scalar, specifying the padding on the left, in the ‘width’ dimension.
• 2: An ANEURALNETWORKS_INT32 scalar, specifying the padding on the right, in the ‘width’ dimension.
• 3: An ANEURALNETWORKS_INT32 scalar, specifying the padding on the top, in the ‘height’ dimension.
• 4: An ANEURALNETWORKS_INT32 scalar, specifying the padding on the bottom, in the ‘height’ dimension.
• 5: An ANEURALNETWORKS_INT32 scalar, specifying the stride when walking through input in the ‘width’ dimension.
• 6: An ANEURALNETWORKS_INT32 scalar, specifying the stride when walking through input in the ‘height’ dimension.
• 7: An ANEURALNETWORKS_INT32 scalar, specifying the filter width.
• 8: An ANEURALNETWORKS_INT32 scalar, specifying the filter height.
• 9: An ANEURALNETWORKS_INT32 scalar, and has to be one of the FuseCode values. Specifies the activation to invoke on the result.

Inputs (implicit padding):

• 0: A 4-D tensor, of shape [batches, height, width, depth], specifying the input.
• 1: An ANEURALNETWORKS_INT32 scalar, specifying the implicit padding scheme, has to be one of the PaddingCode values.
• 2: An ANEURALNETWORKS_INT32 scalar, specifying the stride when walking through input in the ‘width’ dimension.
• 3: An ANEURALNETWORKS_INT32 scalar, specifying the stride when walking through input in the ‘height’ dimension.
• 4: An ANEURALNETWORKS_INT32 scalar, specifying the filter width.
• 5: An ANEURALNETWORKS_INT32 scalar, specifying the filter height.
• 6: An ANEURALNETWORKS_INT32 scalar, and has to be one of the FuseCode values. Specifies the activation to invoke on the result.

Outputs:

• 0: The output 4-D tensor, of shape [batches, out_height, out_width, depth].

Available since API level 27.

Computes the mean of elements across dimensions of a tensor.

Reduces the input tensor along the given dimensions to reduce. Unless keep_dims is true, the rank of the tensor is reduced by 1 for each entry in axis. If keep_dims is true, the reduced dimensions are retained with length 1.

If dimensions to reduce have no entries, all dimensions are reduced, and a tensor with a single element is returned.

Supported tensor OperandCode:

• ANEURALNETWORKS_TENSOR_FLOAT32
• ANEURALNETWORKS_TENSOR_QUANT8_ASYMM

Supported tensor rank: up to 4

Inputs:

• 0: A tensor, specifying the input.
• 1: A 1-D Tensor of ANEURALNETWORKS_TENSOR_INT32. The dimensions to reduce. If None (the default), reduces all dimensions. Must be in the range [-rank(input_tensor), rank(input_tensor)).
• 2: An ANEURALNETWORKS_INT32 scalar, keep_dims. If positive, retains reduced dimensions with length 1.

Outputs:

• 0: A tensor of the same OperandCode as input0.

Available since API level 28.

Multiplies two tensors, element-wise.

Takes two input tensors of identical OperandCode and compatible dimensions. The output is the product of both input tensors, optionally modified by an activation function.

Two dimensions are compatible when:

1. they are equal, or
2. one of them is 1

The size of the resulting output is the maximum size along each dimension of the input operands. It starts with the trailing dimensions, and works its way forward.

Supported tensor OperandCode:

• ANEURALNETWORKS_TENSOR_FLOAT32
• ANEURALNETWORKS_TENSOR_QUANT8_ASYMM

Supported tensor rank: up to 4

Inputs:

• 0: A tensor.
• 1: A tensor of the same OperandCode, and compatible dimensions as input0.
• 2: An ANEURALNETWORKS_INT32 scalar, and has to be one of the FuseCode values. Specifies the activation to invoke on the result.

Outputs:

• 0: The product, a tensor of the same OperandCode as input0. For output tensor of ANEURALNETWORKS_TENSOR_QUANT8_ASYMM, the following condition must be satisfied: output_scale > input1_scale * input2_scale.

Available since API level 27.

Pads a tensor.

This operation pads a tensor according to the specified paddings.

Supported tensor OperandCode:

• ANEURALNETWORKS_TENSOR_FLOAT32
• ANEURALNETWORKS_TENSOR_QUANT8_ASYMM

Supported tensor rank: up to 4

Inputs:

• 0: An n-D tensor, specifying the tensor to be padded.
• 1: A 2-D Tensor of ANEURALNETWORKS_TENSOR_INT32, the paddings for each spatial dimension of the input tensor. The shape of the tensor must be {rank(input0), 2}. padding[i, 0] specifies the number of elements to be padded in the front of dimension i. padding[i, 1] specifies the number of elements to be padded after the end of dimension i.

Outputs:

• 0: A tensor of the same OperandCode as input0. The output tensor has the same rank as input0, and each dimension of the output tensor has the same size as the corresponding dimension of the input tensor plus the size of the padding: output0.dimension[i] = padding[i, 0] + input0.dimension[i] + padding[i, 1]

Available since API level 28.

Computes rectified linear activation on the input tensor element-wise.

The output is calculated using this formula:

output = max(0, input)

Supported tensor OperandCode:

• ANEURALNETWORKS_TENSOR_FLOAT32
• ANEURALNETWORKS_TENSOR_QUANT8_ASYMM

Supported tensor rank: up to 4.

Inputs:

• 0: A tensor, specifying the input.

Outputs:

• 0: The output tensor of same shape as input0.

Available since API level 27.

Computes rectified linear 1 activation on the input tensor element-wise.

The output is calculated using this formula:

output = min(1.f, max(-1.f, input))

Supported tensor OperandCode:

• ANEURALNETWORKS_TENSOR_FLOAT32
• ANEURALNETWORKS_TENSOR_QUANT8_ASYMM

Supported tensor rank: up to 4.

Inputs:

• 0: A tensor, specifying the input.

Outputs:

• 0: The output tensor of same shape as input0.

Available since API level 27.

Computes rectified linear 6 activation on the input tensor element-wise.

The output is calculated using this formula:

output = min(6, max(0, input))

Supported tensor OperandCode:

• ANEURALNETWORKS_TENSOR_FLOAT32
• ANEURALNETWORKS_TENSOR_QUANT8_ASYMM

Supported tensor rank: up to 4.

Inputs:

• 0: A tensor, specifying the input.

Outputs:

• 0: The output tensor of same shape as input0.

Available since API level 27.

Reshapes a tensor.

Given tensor, this operation returns a tensor that has the same values as tensor, but with a newly specified shape.

Supported tensor OperandCode:

• ANEURALNETWORKS_TENSOR_FLOAT32
• ANEURALNETWORKS_TENSOR_QUANT8_ASYMM

Supported tensor rank: up to 4.

Inputs:

• 0: A tensor, specifying the tensor to be reshaped.
• 1: A 1-D tensor of ANEURALNETWORKS_TENSOR_INT32, defining the shape of the output tensor. The number of elements implied by shape must be the same as the number of elements in the input tensor.

Outputs:

• 0: The output tensor, of shape specified by the input shape.

Available since API level 27.

Resizes images to given size using the bilinear interpretation.

Resized images must be distorted if their output aspect ratio is not the same as input aspect ratio. The corner pixels of output may not be the same as corner pixels of input.

Supported tensor OperandCode:

• ANEURALNETWORKS_TENSOR_FLOAT32

Supported tensor rank: 4, with "NHWC" data layout.

Inputs:

• 0: A 4-D tensor, of shape [batches, height, width, depth], specifying the input.
• 1: An ANEURALNETWORKS_INT32 scalar, specifying the output height of the output tensor.
• 2: An ANEURALNETWORKS_INT32 scalar, specifying the output width of the output tensor.

Outputs:

• 0: The output 4-D tensor, of shape [batches, new_height, new_width, depth].

Available since API level 27.

A basic recurrent neural network layer.

This layer implements the operation: outputs = state = activation(inputs * input_weights + state * recurrent_weights + bias)

Where:

• “input_weights” is a weight matrix that multiplies the inputs;
• “recurrent_weights” is a weight matrix that multiplies the current “state” which itself is the output from the previous time step computation;
• “bias” is a bias vector (added to each output vector in the batch);
• “activation” is the function passed as the “fused_activation_function” argument (if not “NONE”).

Supported tensor OperandCode:

• ANEURALNETWORKS_TENSOR_FLOAT32

Inputs:

• 0: input. A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32 of shape [batch_size, input_size], where “batch_size” corresponds to the batching dimension, and “input_size” is the size of the input.
• 1: weights. A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [num_units, input_size], where “num_units” corresponds to the number of units.
• 2: recurrent_weights. A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [num_units, num_units], with columns corresponding to the weights from each unit.
• 3: bias. A 1-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [num_units].
• 4: hidden state (in). A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [batch_size, num_units].
• 5: fused_activation_function. An optional FuseCode value indicating the activation function. If “NONE” is specified then it results in a linear activation.

Outputs:

• 0: hidden state (out). A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [batch_size, num_units].
• 1: output. A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [batch_size, num_units]. This is effectively the same as the current state value.

Available since API level 27.

Computes the softmax activation on the input tensor element-wise, per batch, by normalizing the input vector so the maximum coefficient is zero.

The output is calculated using this formula:

output[batch, i] =
    exp((input[batch, i] - max(input[batch, :])) * beta) /
    sum_{k}{exp((input[batch, k] - max(input[batch, :])) * beta)}

Supported tensor OperandCode:

• ANEURALNETWORKS_TENSOR_FLOAT32
• ANEURALNETWORKS_TENSOR_QUANT8_ASYMM

Supported tensor rank: 2 or 4.

Inputs:

• 0: A 2-D or 4-D tensor, specifying the tensor to be reshaped.
• 1: An ANEURALNETWORKS_FLOAT32 scalar, specifying the positive scaling factor for the exponent, beta.

Outputs:

• 0: The output tensor of same shape as input0. For ANEURALNETWORKS_TENSOR_QUANT8_ASYMM, the scale must be 1.f / 256 and the zeroPoint must be 0.

Available since API level 27.

SpaceToBatch for N-Dimensional tensors.

This operation divides "spatial" dimensions [1, ..., M] of the input into a grid of blocks of shape block_shape, and interleaves these blocks with the "batch" dimension (0) such that in the output, the spatial dimensions [1, ..., M] correspond to the position within the grid, and the batch dimension combines both the position within a spatial block and the original batch position. Prior to division into blocks, the spatial dimensions of the input are optionally zero padded according to paddings.

Supported tensor OperandCode:

• ANEURALNETWORKS_TENSOR_FLOAT32
• ANEURALNETWORKS_TENSOR_QUANT8_ASYMM

Supported tensor rank: 4

Inputs:

• 0: An n-D tensor, specifying the input.
• 1: A 1-D Tensor of ANEURALNETWORKS_TENSOR_INT32, the block sizes for each spatial dimension of the input tensor. All values must be >= 1.
• 2: A 2-D Tensor of ANEURALNETWORKS_TENSOR_INT32, the paddings for each spatial dimension of the input tensor. All values must be >= 0. The shape of the tensor must be {rank(input0), 2}. padding[i, 0] specifies the number of element to be padded in the front of dimension i. padding[i, 1] specifies the number of element to be padded after the end of dimension i.

Outputs:

• 0: A tensor of the same OperandCode as input0.

Available since API level 28.

Rearranges blocks of spatial data, into depth.

More specifically, this op outputs a copy of the input tensor where values from the height and width dimensions are moved to the depth dimension. The value block_size indicates the input block size and how the data is moved.

Chunks of data of size block_size * block_size from depth are rearranged into non-overlapping blocks of size block_size x block_size.

The depth of the output tensor is input_depth * block_size * block_size. The input tensor's height and width must be divisible by block_size.

Supported tensor OperandCode:

• ANEURALNETWORKS_TENSOR_FLOAT32
• ANEURALNETWORKS_TENSOR_QUANT8_ASYMM

Supported tensor rank: 4, with "NHWC" data layout.

Inputs:

• 0: A 4-D tensor, of shape [batches, height, width, depth_in], specifying the input.
• 1: An ANEURALNETWORKS_INT32 scalar, specifying the block_size. block_size must be >=1 and block_size must be a divisor of both the input height and width.

Outputs:

• 0: The output 4-D tensor, of shape [batches, height/block_size, width/block_size, depth_in*block_size*block_size].

Available since API level 27.

Removes dimensions of size 1 from the shape of a tensor.

Given a tensor input, this operation returns a tensor of the same OperandCode with all dimensions of size 1 removed. If you don't want to remove all size 1 dimensions, you can remove specific size 1 dimensions by specifying the axes (input1).

Supported tensor OperandCode:

• ANEURALNETWORKS_TENSOR_FLOAT32
• ANEURALNETWORKS_TENSOR_QUANT8_ASYMM

Supported tensor rank: up to 4

Inputs:

• 0: An n-D tensor, the tensor to be squeezed.
• 1: An optional 1-D tensor of ANEURALNETWORKS_TENSOR_INT32. The dimensions to squeeze. If specified only squeezes the dimensions listed. Otherwise, squeezes all dimensions. The dimension index starts at 0. An error must be reported if squeezing a dimension that is not 1.

Outputs:

• 0: A tensor of the same OperandCode as input0. Contains the same data as input, but has one or more dimensions of size 1 removed.

Available since API level 28.

Extracts a strided slice of a tensor.

Roughly speaking, this op extracts a slice of size (end - begin) / stride from the given input tensor. Starting at the location specified by begin the slice continues by adding stride to the index until all dimensions are not less than end. Note that a stride can be negative, which causes a reverse slice.

Supported tensor OperandCode:

• ANEURALNETWORKS_TENSOR_FLOAT32
• ANEURALNETWORKS_TENSOR_QUANT8_ASYMM

Supported tensor rank: up to 4

Inputs:

• 0: An n-D tensor, specifying the tensor to be sliced.
• 1: A 1-D Tensor of ANEURALNETWORKS_TENSOR_INT32, the starts of the dimensions of the input tensor to be sliced. The length must be of rank(input0).
• 2: A 1-D Tensor of ANEURALNETWORKS_TENSOR_INT32, the ends of the dimensions of the input tensor to be sliced. The length must be of rank(input0).
• 3: A 1-D Tensor of ANEURALNETWORKS_TENSOR_INT32, the strides of the dimensions of the input tensor to be sliced. The length must be of rank(input0).
• 4: An ANEURALNETWORKS_INT32 scalar, begin_mask. If the ith bit of begin_mask is set, begin[i] is ignored and the fullest possible range in that dimension is used instead.
• 5: An ANEURALNETWORKS_INT32 scalar, end_mask. If the ith bit of end_mask is set, end[i] is ignored and the fullest possible range in that dimension is used instead.
• 6: An ANEURALNETWORKS_INT32 scalar, shrink_axis_mask. An int32 mask. If the ith bit of shrink_axis_mask is set, it implies that the ith specification shrinks the dimensionality by 1. A slice of size 1 starting from begin[i] in the dimension must be preserved.

Outputs:

• 0: A tensor of the same OperandCode as input0.

Available since API level 28.

Element-wise subtraction of two tensors.

Takes two input tensors of identical OperandCode and compatible dimensions. The output is the result of subtracting the second input tensor from the first one, optionally modified by an activation function.

Two dimensions are compatible when:

1. they are equal, or
2. one of them is 1

The size of the output is the maximum size along each dimension of the input operands. It starts with the trailing dimensions, and works its way forward.

Example: input1.dimension = {4, 1, 2} input2.dimension = {5, 4, 3, 1} output.dimension = {5, 4, 3, 2}

Supported tensor OperandCode:

• ANEURALNETWORKS_TENSOR_FLOAT32

Supported tensor rank: up to 4

Inputs:

• 0: An n-D tensor, specifying the first input.
• 1: A tensor of the same OperandCode, and compatible dimensions as input0.
• 2: An ANEURALNETWORKS_INT32 scalar, and has to be one of the FuseCode values. Specifies the activation to invoke on the result.

Outputs:

• 0: A tensor of the same OperandCode as input0.

Available since API level 28.

SVDF op is a kind of stateful layer derived from the notion that a densely connected layer that's processing a sequence of input frames can be approximated by using a singular value decomposition of each of its nodes.

The implementation is based on:

https://research.google.com/pubs/archive/43813.pdf

P. Nakkiran, R. Alvarez, R. Prabhavalkar, C. Parada. “Compressing Deep Neural Networks using a Rank-Constrained Topology”. INTERSPEECH, 2015.

It processes the incoming input using a 2-stage filtering mechanism:

• stage 1 performs filtering on the "features" dimension, whose outputs get pushed into a memory of fixed-size memory_size.
• stage 2 performs filtering on the "time" dimension of the memory_size memoized outputs of stage 1.

Specifically, for rank 1, this layer implements the operation:

memory = push(conv1d(inputs, weights_feature, feature_dim,
                     "ANEURALNETWORKS_PADDING_VALID"));
outputs = activation(memory * weights_time + bias);

Where:

• “weights_feature” is a weights matrix that processes the inputs (by convolving the input with every “feature filter”), and whose outputs get pushed, stacked in order, into the fixed-size “memory” (the oldest entry gets dropped);
• “weights_time” is a weights matrix that processes the “memory” (by a batched matrix multiplication on the num_units);
• “bias” is an optional bias vector (added to each output vector in the batch); and
• “activation” is the function passed as the “fused_activation_function” argument (if not “NONE”).

Each rank adds a dimension to the weights matrices by means of stacking the filters.

Supported tensor OperandCode:

• ANEURALNETWORKS_TENSOR_FLOAT32

Inputs:

• 0: input. A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [batch_size, input_size], where “batch_size” corresponds to the batching dimension, and “input_size” is the size of the input.
• 1: weights_feature. A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [num_units, input_size], where “num_units” corresponds to the number of units.
• 2: weights_time. A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [num_units, memory_size], where “memory_size” corresponds to the fixed-size of the memory.
• 3: bias. An optional 1-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [num_units].
• 4: state (in). A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [batch_size, (memory_size - 1) * num_units * rank].
• 5: rank. The rank of the SVD approximation.
• 6: fused_activation_function. An optional FuseCode value indicating the activation function. If “NONE” is specified then it results in a linear activation.

Outputs:

• 0: state (out). A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [batch_size, (memory_size - 1) * num_units * rank].
• 1: output. A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [batch_size, num_units].

Available since API level 27.

Computes hyperbolic tangent of input tensor element-wise.

The output is calculated using this formula:

output = tanh(input)

Supported tensor OperandCode:

• ANEURALNETWORKS_TENSOR_FLOAT32

Supported tensor rank: up to 4.

Inputs:

• 0: A tensor, specifying the input.

Outputs:

• 0: The output tensor of same shape as input0.

Available since API level 27.

Transposes the input tensor, permuting the dimensions according to the perm tensor.

The returned tensor's dimension i corresponds to the input dimension perm[i]. If perm is not given, it is set to (n-1...0), where n is the rank of the input tensor. Hence by default, this operation performs a regular matrix transpose on 2-D input Tensors.

Supported tensor OperandCode:

• ANEURALNETWORKS_TENSOR_FLOAT32
• ANEURALNETWORKS_TENSOR_QUANT8_ASYMM

Supported tensor rank: up to 4

Inputs:

• 0: An n-D tensor, specifying the tensor to be transposed.
• 1: An optional 1-D Tensor of ANEURALNETWORKS_TENSOR_INT32, the permutation of the dimensions of the input tensor.

Outputs:

• 0: A tensor of the same OperandCode as input0.

Available since API level 28.

SAME padding.

Padding on both ends are the "same": padding_to_beginning = total_padding / 2 padding_to_end = (total_padding + 1)/2. i.e., for even number of padding, padding to both ends are exactly the same; for odd number of padding, padding to the ending is bigger than the padding to the beginning by 1.

total_padding is a function of input, stride and filter size. It could be computed as follows: out_size = (input + stride - 1) / stride; needed_input = (out_size - 1) * stride + filter_size total_padding = max(0, needed_input - input_size) The computation is the same for the horizontal and vertical directions.

VALID padding.

No padding. When the input size is not evenly divisible by the filter size, the input at the end that could not fill the whole filter tile will simply be ignored.

Prefer returning a single answer as fast as possible, even if this causes more power consumption.

Prefer executing in a way that minimizes battery drain.

This is desirable for compilations that will be executed often.

Prefer maximizing the throughput of successive frames, for example when processing successive frames coming from the camera.

Failure caused by invalid function arguments, invalid model definition, invalid execution definition or invalid data at execution time.

Failure caused by object being in the wrong state.

Operation was succesful.

Failure caused by failed model execution.

Failure caused by not enough available memory.

Failure caused by unexpected null argument.

Failure caused by not being able to map a file into memory.

This may be caused by a file descriptor not being mappable. Mitigate by reading its content into memory.