/* * Copyright (C) 2017 The Android Open Source Project * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ package android.hardware.neuralnetworks@1.0; /** * Operand types. * * The type of an operand in a model. * * Types prefaced with TENSOR_* must be used for tensor data (i.e., tensors * with at least one dimension). Types not prefaced by TENSOR_* represent * scalar values and must have no dimensions. * * Although we define many types, most operators accept just a few * types. Most used are {@link OperandType::TENSOR_FLOAT32}, * {@link OperandType::TENSOR_QUANT8_ASYMM}, * and {@link OperandType::INT32}. */ enum OperandType : int32_t { /** A 32 bit floating point scalar value. */ FLOAT32 = 0, /** A signed 32 bit integer scalar value. */ INT32 = 1, /** An unsigned 32 bit integer scalar value. */ UINT32 = 2, /** A tensor of 32 bit floating point values. */ TENSOR_FLOAT32 = 3, /** A tensor of 32 bit integer values. */ TENSOR_INT32 = 4, /** * A tensor of 8 bit unsigned 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. */ TENSOR_QUANT8_ASYMM = 5, /** * DEPRECATED. Since HAL version 1.2, extensions are the preferred * alternative to OEM operation and data types. * * OEM specific scalar value. */ OEM = 10000, /** * DEPRECATED. Since HAL version 1.2, extensions are the preferred * alternative to OEM operation and data types. * * A tensor of OEM specific values. */ TENSOR_OEM_BYTE = 10001, }; /** * Operation types. * * The type of an operation in a model. */ enum OperationType : int32_t { /** * Adds two tensors, element-wise. * * Takes two input tensors of identical {@link OperandType} 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 {@link OperandType}: * * {@link OperandType::TENSOR_FLOAT32} * * {@link OperandType::TENSOR_QUANT8_ASYMM} * * Supported tensor rank: up to 4 * * Inputs: * * 0: A tensor. * * 1: A tensor of the same {@link OperandType}, and compatible dimensions * as input0. * For a {@link OperandType::TENSOR_QUANT8_ASYMM} tensor, * the scales and zeroPoint can be different from input0 scale and zeroPoint. * * 2: An {@link OperandType::INT32} scalar, and has to be one of the * {@link FusedActivationFunc} values. Specifies the activation to * invoke on the result. * * Outputs: * * 0: The sum, a tensor of the same {@link OperandType} as input0. * For a {@link OperandType::TENSOR_QUANT8_ASYMM} tensor, * the scale and zeroPoint can be different from inputs' scale and zeroPoint. */ ADD = 0, /** * 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[b, i, j, channel] = * sum_{di, dj}( * input[b, strides[1] * i + di, strides[2] * j + dj, channel] * ) / sum(1) * * Supported tensor {@link OperandType}: * * {@link OperandType::TENSOR_FLOAT32} * * {@link OperandType::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 {@link OperandType::INT32} scalar, specifying the padding on * the left, in the ‘width’ dimension. * * 2: An {@link OperandType::INT32} scalar, specifying the padding on * the right, in the ‘width’ dimension. * * 3: An {@link OperandType::INT32} scalar, specifying the padding on * the top, in the ‘height’ dimension. * * 4: An {@link OperandType::INT32} scalar, specifying the padding on * the bottom, in the ‘height’ dimension. * * 5: An {@link OperandType::INT32} scalar, specifying the stride when * walking through input in the ‘width’ dimension. * * 6: An {@link OperandType::INT32} scalar, specifying the stride when * walking through input in the ‘height’ dimension. * * 7: An {@link OperandType::INT32} scalar, specifying the filter * width. * * 8: An {@link OperandType::INT32} scalar, specifying the filter * height. * * 9: An {@link OperandType::INT32} scalar, and has to be one of the * {@link FusedActivationFunc} 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 {@link OperandType::INT32} scalar, specifying the implicit * padding scheme, has to be one of the * following values: {0 (NONE), 1 (SAME), 2 (VALID)}. * * 2: An {@link OperandType::INT32} scalar, specifying the stride when * walking through input in the ‘width’ dimension. * * 3: An {@link OperandType::INT32} scalar, specifying the stride when * walking through input in the ‘height’ dimension. * * 4: An {@link OperandType::INT32} scalar, specifying the filter * width. * * 5: An {@link OperandType::INT32} scalar, specifying the filter * height. * * 6: An {@link OperandType::INT32} scalar, and has to be one of the * {@link FusedActivationFunc} values. Specifies the activation to * invoke on the result. * * Outputs: * * 0: The output 4-D tensor, of shape * [batches, out_height, out_width, depth]. * For a {@link OperandType::TENSOR_QUANT8_ASYMM} tensor, * the scale and zeroPoint must be the same as input0. */ AVERAGE_POOL_2D = 1, /** * Concatenates the input tensors along the given dimension. * * The input tensors must have identical {@link OperandType} and the same * dimensions except the dimension along the concatenation axis. * * Supported tensor {@link OperandType}: * * {@link OperandType::TENSOR_FLOAT32} * * {@link OperandType::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]. * All input tensors of * {@link OperandType::TENSOR_QUANT8_ASYMM} * must have the same scale and zeroPoint as the output tensor. * * n: An {@link OperandType::INT32} scalar, specifying the * concatenation axis. * * Outputs: * * 0: The output, a tensor of the same {@link OperandType} as the input * tensors. The output shape is [D0, D1, ..., sum(Daxis(i)), ..., Dm]. * For a {@link OperandType::TENSOR_QUANT8_ASYMM} tensor, the scale and zeroPoint * values must be the same as the input tensors'. */ CONCATENATION = 2, /** * Performs a 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[b, i, j, channel] = * sum_{di, dj, k} ( * input[b, strides[1] * i + di, strides[2] * j + dj, k] * * filter[channel, di, dj, k] * ) + bias[channel] * * Supported tensor {@link OperandType} configurations: * * 32 bit floating point: * * * {@link OperandType::TENSOR_FLOAT32} for input, filter, output, and bias. * * * Quantized: * * * {@link OperandType::TENSOR_QUANT8_ASYMM} for input, filter, and output. * * * {@link OperandType::TENSOR_INT32} for bias (with scale set to * * * input.scale * filter.scale). * * 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_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 type {@link OperandType::TENSOR_FLOAT32} * the bias must be of the same type. * For filter tensor of {@link OperandType::TENSOR_QUANT8_ASYMM}, * the bias should be of {@link OperandType::TENSOR_INT32}, with zeroPoint * of 0 and bias_scale == input_scale * filter_scale. * * 3: An {@link OperandType::INT32} scalar, specifying the padding on * the left, in the ‘width’ dimension. * * 4: An {@link OperandType::INT32} scalar, specifying the padding on * the right, in the ‘width’ dimension. * * 5: An {@link OperandType::INT32} scalar, specifying the padding on * the top, in the ‘height’ dimension. * * 6: An {@link OperandType::INT32} scalar, specifying the padding on * the bottom, in the ‘height’ dimension. * * 7: An {@link OperandType::INT32} scalar, specifying the stride when * walking through input in the ‘width’ dimension. * * 8: An {@link OperandType::INT32} scalar, specifying the stride when * walking through input in the ‘height’ dimension. * * 9: An {@link OperandType::INT32} scalar, and has to be one of the * {@link FusedActivationFunc} 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 type {@link OperandType::TENSOR_FLOAT32} * the bias must be of the same * type. * For filter tensor of {@link OperandType::TENSOR_QUANT8_ASYMM}, * the bias should be of {@link OperandType::TENSOR_INT32}, with zeroPoint * of 0 and bias_scale == input_scale * filter_scale. * * 3: An {@link OperandType::INT32} scalar, specifying the implicit * padding scheme, has to be one of the * following values: {0 (NONE), 1 (SAME), 2 (VALID)}. * * 4: An {@link OperandType::INT32} scalar, specifying the stride when * walking through input in the ‘width’ dimension. * * 5: An {@link OperandType::INT32} scalar, specifying the stride when * walking through input in the ‘height’ dimension. * * 6: An {@link OperandType::INT32} scalar, and has to be one of the * {@link FusedActivationFunc} 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 * {@link OperandType::TENSOR_QUANT8_ASYMM}, the following condition must * be satisfied: output_scale > input_scale * filter_scale */ CONV_2D = 3, /** * 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] * ) + bias[k * channel_multiplier + q] * * Supported tensor {@link OperandType} configurations: * * 32 bit floating point: * * * {@link OperandType::TENSOR_FLOAT32} for input, filter, output, and bias. * * * Quantized: * * * {@link OperandType::TENSOR_QUANT8_ASYMM} for input, filter, and output. * * * {@link OperandType::TENSOR_INT32} for bias (with scale set to * * * input.scale * filter.scale). * * 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_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 type {@link OperandType::TENSOR_FLOAT32} * the bias must be of the same type. * For filter tensor of {@link OperandType::TENSOR_QUANT8_ASYMM}, * the bias should be of {@link OperandType::TENSOR_INT32}, with zeroPoint * of 0 and bias_scale == input_scale * filter_scale. * * 3: An {@link OperandType::INT32} scalar, specifying the padding on * the left, in the ‘width’ dimension. * * 4: An {@link OperandType::INT32} scalar, specifying the padding on * the right, in the ‘width’ dimension. * * 5: An {@link OperandType::INT32} scalar, specifying the padding on * the top, in the ‘height’ dimension. * * 6: An {@link OperandType::INT32} scalar, specifying the padding on * the bottom, in the ‘height’ dimension. * * 7: An {@link OperandType::INT32} scalar, specifying the stride when * walking through input in the ‘width’ dimension. * * 8: An {@link OperandType::INT32} scalar, specifying the stride when * walking through input in the ‘height’ dimension. * * 9: An {@link OperandType::INT32} scalar, specifying the depthwise * multiplier. * * 10: An {@link OperandType::INT32} scalar, and has to be one of the * {@link FusedActivationFunc} 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 type {@link OperandType::TENSOR_FLOAT32} * the bias must be of the same type. * For filter tensor of {@link OperandType::TENSOR_QUANT8_ASYMM}, * the bias should be of {@link OperandType::TENSOR_INT32}, with zeroPoint * of 0 and bias_scale == input_scale * filter_scale. * * 3: An {@link OperandType::INT32} scalar, specifying the implicit * padding scheme, has to be one of the * following values: {0 (NONE), 1 (SAME), 2 (VALID)}. * * 4: An {@link OperandType::INT32} scalar, specifying the stride when * walking through input in the ‘width’ dimension. * * 5: An {@link OperandType::INT32} scalar, specifying the stride when * walking through input in the ‘height’ dimension. * * 6: An {@link OperandType::INT32} scalar, specifying the depthwise * multiplier. * * 7: An {@link OperandType::INT32} scalar, and has to be one of the * {@link FusedActivationFunc} 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 {@link OperandType::TENSOR_QUANT8_ASYMM}, * the following condition must be satisfied: * output_scale > input_scale * filter_scale */ DEPTHWISE_CONV_2D = 4, /** * 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 {@link OperandType}: * * {@link OperandType::TENSOR_FLOAT32} * * {@link OperandType::TENSOR_QUANT8_ASYMM} * * Supported tensor rank: 4, with "NHWC" (i.e., Num_samples, Height, Width, * and Channels) data layout. * * Inputs: * * 0: A 4-D tensor, of shape [batches, height, width, depth_in], * specifying the input. * * 1: An {@link OperandType::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)]. * For a {@link OperandType::TENSOR_QUANT8_ASYMM} tensor, * the scale and zeroPoint must be the same as input0. */ DEPTH_TO_SPACE = 5, /** * Dequantizes the input tensor. * * The formula is: * * output = (input - zeroPoint) * scale. * * Supported input tensor {@link OperandType}: * * {@link OperandType::TENSOR_QUANT8_ASYMM} * * Supported output tensor {@link OperandType}: * * {@link OperandType::TENSOR_FLOAT32}. * * Supported tensor rank: up to 4 * * Inputs: * * 0: A tensor. * * Outputs: * * 0: A tensor with the same shape as input0. */ DEQUANTIZE = 6, /** * 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. * * Supported value tensor {@link OperandType}: * * {@link OperandType::TENSOR_FLOAT32} * * Supported value tensor rank: from 2 * * Inputs: * * 0: Lookups. A 1-D tensor of {@link OperandType::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. * For a {@link OperandType::TENSOR_QUANT8_ASYMM} tensor, * the scale and zeroPoint must be the same as input1. */ EMBEDDING_LOOKUP = 7, /** * Computes element-wise floor() on the input tensor. * * Supported tensor {@link OperandType}: * * {@link OperandType::TENSOR_FLOAT32} * * Supported tensor rank: up to 4 * * Inputs: * * 0: A tensor. * * Outputs: * * 0: The output tensor, of the same {@link OperandType} and dimensions as * the input tensor. */ FLOOR = 8, /** * 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 {@link OperandType}: * * {@link OperandType::TENSOR_FLOAT32} * * {@link OperandType::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 {@link OperandType::TENSOR_FLOAT32}, the bias should * also be of {@link OperandType::TENSOR_FLOAT32}. * For input tensor of {@link OperandType::TENSOR_QUANT8_ASYMM}, * the bias should be of {@link OperandType::TENSOR_INT32}, * with zeroPoint of 0 and bias_scale == input_scale * filter_scale. * * 3: An {@link OperandType::INT32} scalar, and has to be one of the * {@link FusedActivationFunc} values. Specifies the activation to * invoke on the result. * * Outputs: * * 0: The output tensor, of shape [batch_size, num_units]. For * output tensor of {@link OperandType::TENSOR_QUANT8_ASYMM}, the following * condition must be satisfied: output_scale > input_scale * filter_scale. */ FULLY_CONNECTED = 9, /** * 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. * * Supported value tensor {@link OperandType}: * * {@link OperandType::TENSOR_FLOAT32} * * {@link OperandType::TENSOR_INT32} * * {@link OperandType::TENSOR_QUANT8_ASYMM} * * Supported value tensor rank: from 2 * * Inputs: * * 0: Lookups. A 1-D {@link OperandType::TENSOR_INT32} tensor with * shape [ k ]. * * 1: Keys. A 1-D {@link OperandType::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 …]. * For a {@link OperandType::TENSOR_QUANT8_ASYMM} tensor, * the scale and zeroPoint must be the same as input2. * * 1: Hits. A boolean tensor with shape [ k ] indicates whether the lookup * hits (True) or not (False). * Stored as {@link OperandType::TENSOR_QUANT8_ASYMM} with offset 0 * and scale 1.0f. * A non-zero byte represents True, a hit. A zero indicates otherwise. */ HASHTABLE_LOOKUP = 10, /** * Applies L2 normalization along the axis 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)) * * Supported tensor {@link OperandType}: * * {@link OperandType::TENSOR_FLOAT32} * * Supported tensor rank: 4, with "NHWC" data layout (i.e., Num_samples, * Height, Width, and Channels). * * Inputs: * * 0: A 4-D tensor, specifying the tensor to be normalized. * * Outputs: * * 0: A tensor of the same {@link OperandType} and same shape as input0. */ L2_NORMALIZATION = 11, /** * 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[b, i, j, c] = * sqrt(sum_{di, dj} pow(input[b, strides[1] * i + di, strides[2] * j + dj, c], 2) / * sum(1)) * * Supported tensor {@link OperandType}: * * {@link OperandType::TENSOR_FLOAT32} * * 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 {@link OperandType::INT32} scalar, specifying the padding on * the left, in the ‘width’ dimension. * * 2: An {@link OperandType::INT32} scalar, specifying the padding on * the right, in the ‘width’ dimension. * * 3: An {@link OperandType::INT32} scalar, specifying the padding on * the top, in the ‘height’ dimension. * * 4: An {@link OperandType::INT32} scalar, specifying the padding on * the bottom, in the ‘height’ dimension. * * 5: An {@link OperandType::INT32} scalar, specifying the stride when * walking through input in the ‘width’ dimension. * * 6: An {@link OperandType::INT32} scalar, specifying the stride when * walking through input in the ‘height’ dimension. * * 7: An {@link OperandType::INT32} scalar, specifying the filter * width. * * 8: An {@link OperandType::INT32} scalar, specifying the filter * height. * * 9: An {@link OperandType::INT32} scalar, and has to be one of the * {@link FusedActivationFunc} 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 {@link OperandType::INT32} scalar, specifying the implicit * padding scheme, has to be one of the * following values: {0 (NONE), 1 (SAME), 2 (VALID)}. * * 2: An {@link OperandType::INT32} scalar, specifying the stride when * walking through input in the ‘width’ dimension. * * 3: An {@link OperandType::INT32} scalar, specifying the stride when * walking through input in the ‘height’ dimension. * * 4: An {@link OperandType::INT32} scalar, specifying the filter * width. * * 5: An {@link OperandType::INT32} scalar, specifying the filter * height. * * 6: An {@link OperandType::INT32} scalar, and has to be one of the * {@link FusedActivationFunc} values. Specifies the activation to * invoke on the result. * * Outputs: * * 0: The output 4-D tensor, of shape * [batches, out_height, out_width, depth]. */ L2_POOL_2D = 12, /** * 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 {@link OperandType}: * * {@link OperandType::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 {@link OperandType::INT32} scalar, specifying the radius of * the normalization window. * * 2: A scalar, specifying the bias, must not be zero. * For input tensor of {@link OperandType::TENSOR_FLOAT32}, the bias * value must be of {@link OperandType::FLOAT32}. * * 3: A scalar, specifying the scale factor, alpha. * For input tensor of {@link OperandType::TENSOR_FLOAT32}, the * alpha value must be of {@link OperandType::FLOAT32}. * * 4: A scalar, specifying the exponent, beta. * For input tensor of {@link OperandType::TENSOR_FLOAT32}, the beta * value must be of {@link OperandType::FLOAT32}. * * Outputs: * * 0: The output tensor of same shape as input0. */ LOCAL_RESPONSE_NORMALIZATION = 13, /** * Computes sigmoid activation on the input tensor element-wise. * * The output is calculated using this formula: * * output = 1 / (1 + exp(-input)) * * Supported tensor {@link OperandType}: * * {@link OperandType::TENSOR_FLOAT32} * * {@link OperandType::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 {@link OperandType::TENSOR_QUANT8_ASYMM}, * the scale must be 1.f / 256 and the zeroPoint must be 0. */ LOGISTIC = 14, /** * Projects an input to a bit vector via locality senstive hashing. * * Supported input tensor {@link OperandType}: * * {@link OperandType::TENSOR_FLOAT32} * * {@link OperandType::TENSOR_INT32} * * {@link OperandType::TENSOR_QUANT8_ASYMM} * * Supported input tensor rank: from 1 * * Inputs: * * 0: Hash functions. Dim.size == 2, DataType: Float. * Tensor[0].Dim[0]: Number of hash functions. * Tensor[0].Dim[1]: Number of projected output bits generated by each * hash function. * If the projection type is Sparse: * Tensor[0].Dim[1] + ceil(log2(Tensor[0].Dim[0])) <= 32 * * * 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. */ LSH_PROJECTION = 15, /** * Performs a single time step in a Long Short-Term Memory (LSTM) layer * * The LSTM operation is described by the following equations. * * \f{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. \\ * \f} * Where: * * \f$x_t\f$ is the input, * * \f$i_t\f$ is the input gate, * * \f$f_t\f$ is the forget gate, * * \f$C_t\f$ is the cell state, * * \f$o_t\f$ is the output, * * \f$h_t\f$ is the output state, * * \f$\sigma\f$ is the logistic sigmoid function, * * \f$g\f$ is the cell input and cell output activation function, usually * \f$tahn\f$, * * \f$W_{xi}\f$ is the input-to-input weight matrix, * * \f$W_{hi}\f$ is the recurrent to input weight matrix, * * \f$W_{ci}\f$ is the cell-to-input weight matrix, * * \f$b_i\f$ is the input gate bias, * * \f$W_{xf}\f$ is the input-to-forget weight matrix, * * \f$W_{hf}\f$ is the recurrent-to-forget weight matrix, * * \f$W_{cf}\f$ is the cell-to-forget weight matrix, * * \f$b_f\f$ is the forget gate bias, * * \f$W_{xc}\f$ is the input-to-cell weight matrix, * * \f$W_{hc}\f$ is the recurrent-to-cell weight matrix, * * \f$b_c\f$ is the cell bias, * * \f$W_{xo}\f$ is the input-to-output weight matrix, * * \f$W_{ho}\f$ is the recurrent-to-output weight matrix, * * \f$W_{co}\f$ is the cell-to-output weight matrix, * * \f$b_o\f$ is the output gate bias, * * \f$W_{proj}\f$ is the projection weight matrix, * * \f$b_{proj}\f$ is the projection bias, * * \f$t_{cell}\f$ is the threshold for clipping the cell state, and * * \f$t_{proj}\f$ is the threshold for clipping the projected output. * * \f$\odot\f$ 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 cell-to-input weights (\f$W_{ci}\f$), cell-to-forget weights * (\f$W_{cf}\f$) and cell-to-output weights (\f$W_{co}\f$) either all * have values or neither of them have values (i.e., all set to null). If * they have values, the peephole optimization is used. * * The input-to-input weights (\f$W_{xi}\f$), recurrent-to-input weights * (\f$W_{hi}\f$) and input gate bias (\f$b_i\f$) either all have values, * or none of them have values. If they have no values, coupling of input * and forget gates (CIFG) is used, in which case the input gate * (\f$i_t\f$) is calculated using the following equation instead. * \f{eqnarray*}{ * i_t = 1 - f_t * \f} * In case peephole optimization is used and CIFG is not used * cell-to-input (\f$W_{ci}\f$) weights must be present. Otherwise, the * cell-to-input weights must have no value. * * The projection weights (\f$W_{proj}\f$) is required only for the * recurrent projection layer, and should otherwise have no value. * * The projection bias (\f$b_{proj}\f$) 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 {@link OperandType}: * * {@link OperandType::TENSOR_FLOAT32} * * All input and output tensors must be of the same type. * * Inputs: * * 0: The input (\f$x_t\f$). * A 2-D tensor 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 (\f$W_{xi}\f$). Optional. * A 2-D tensor of shape [num_units, input_size], where “num_units” * corresponds to the number of cell units. * * 2: The input-to-forget weights (\f$W_{xf}\f$). * A 2-D tensor of shape [num_units, input_size]. * * 3: The input-to-cell weights (\f$W_{xc}\f$). * A 2-D tensor of shape [num_units, input_size]. * * 4: The input-to-output weights (\f$W_{xo}\f$). * A 2-D tensor of shape [num_units, input_size]. * * 5: The recurrent-to-input weights (\f$W_{hi}\f$). Optional. * A 2-D tensor 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 (\f$W_{hf}\f$). * A 2-D tensor of shape [num_units, output_size]. * * 7: The recurrent-to-cell weights (\f$W_{hc}\f$). * A 2-D tensor of shape [num_units, output_size]. * * 8: The recurrent-to-output weights (\f$W_{ho}\f$). * A 2-D tensor of shape [num_units, output_size]. * * 9: The cell-to-input weights (\f$W_{ci}\f$). Optional. * A 1-D tensor of shape [num_units]. * * 10:The cell-to-forget weights (\f$W_{cf}\f$). Optional. * A 1-D tensor of shape [num_units]. * * 11:The cell-to-output weights (\f$W_{co}\f$). Optional. * A 1-D tensor of shape [num_units]. * * 12:The input gate bias (\f$b_i\f$). Optional. * A 1-D tensor of shape [num_units]. * * 13:The forget gate bias (\f$b_f\f$). * A 1-D tensor of shape [num_units]. * * 14:The cell bias (\f$b_c\f$). * A 1-D tensor of shape [num_units]. * * 15:The output gate bias (\f$b_o\f$). * A 1-D tensor of shape [num_units]. * * 16:The projection weights (\f$W_{proj}\f$). Optional. * A 2-D tensor of shape [output_size, num_units]. * * 17:The projection bias (\f$b_{proj}\f$). Optional. * A 1-D tensor of shape [output_size]. * * 18:The output state (in) (\f$h_{t-1}\f$). * A 2-D tensor of shape [batch_size, output_size]. * * 19:The cell state (in) (\f$C_{t-1}\f$). * A 2-D tensor of shape [batch_size, num_units]. * * 20:The activation function (\f$g\f$). * A value indicating the activation function: * * * 21:The clipping threshold (\f$t_{cell}\f$) 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 (\f$t_{proj}\f$) 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 shape [batch_size, num_units * 3] with CIFG, or * [batch_size, num_units * 4] without CIFG. * * 1: The output state (out) (\f$h_t\f$). * A 2-D tensor of shape [batch_size, output_size]. * * 2: The cell state (out) (\f$C_t\f$). * A 2-D tensor of shape [batch_size, num_units]. * * 3: The output (\f$o_t\f$). * A 2-D tensor of shape [batch_size, output_size]. This is effectively * the same as the current “output state (out)” value. */ LSTM = 16, /** * 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[b, i, j, channel] = * max_{di, dj} ( * input[b, strides[1] * i + di, strides[2] * j + dj, channel] * ) * * Supported tensor {@link OperandType}: * * {@link OperandType::TENSOR_FLOAT32} * * {@link OperandType::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 {@link OperandType::INT32} scalar, specifying the padding on * the left, in the ‘width’ dimension. * * 2: An {@link OperandType::INT32} scalar, specifying the padding on * the right, in the ‘width’ dimension. * * 3: An {@link OperandType::INT32} scalar, specifying the padding on * the top, in the ‘height’ dimension. * * 4: An {@link OperandType::INT32} scalar, specifying the padding on * the bottom, in the ‘height’ dimension. * * 5: An {@link OperandType::INT32} scalar, specifying the stride when * walking through input in the ‘width’ dimension. * * 6: An {@link OperandType::INT32} scalar, specifying the stride when * walking through input in the ‘height’ dimension. * * 7: An {@link OperandType::INT32} scalar, specifying the filter * width. * * 8: An {@link OperandType::INT32} scalar, specifying the filter * height. * * 9: An {@link OperandType::INT32} scalar, and has to be one of the * {@link FusedActivationFunc} 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 {@link OperandType::INT32} scalar, specifying the implicit * padding scheme, has to be one of the * following values: {0 (NONE), 1 (SAME), 2 (VALID)}. * * 2: An {@link OperandType::INT32} scalar, specifying the stride when * walking through input in the ‘width’ dimension. * * 3: An {@link OperandType::INT32} scalar, specifying the stride when * walking through input in the ‘height’ dimension. * * 4: An {@link OperandType::INT32} scalar, specifying the filter * width. * * 5: An {@link OperandType::INT32} scalar, specifying the filter * height. * * 6: An {@link OperandType::INT32} scalar, and has to be one of the * {@link FusedActivationFunc} values. Specifies the activation to * invoke on the result. * * Outputs: * * 0: The output 4-D tensor, of shape * [batches, out_height, out_width, depth]. * For a {@link OperandType::TENSOR_QUANT8_ASYMM} tensor, * the scale and zeroPoint must be the same as input0. */ MAX_POOL_2D = 17, /** * Multiplies two tensors, element-wise. * * Takes two input tensors of identical {@link OperandType} 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 {@link OperandType}: * * {@link OperandType::TENSOR_FLOAT32} * * {@link OperandType::TENSOR_QUANT8_ASYMM} * * Supported tensor rank: up to 4 * * Inputs: * * 0: A tensor. * * 1: A tensor of the same {@link OperandType}, and compatible dimensions * as input0. * * 2: An {@link OperandType::INT32} scalar, and has to be one of the * {@link FusedActivationFunc} values. Specifies the activation to * invoke on the result. * * Outputs: * * 0: The product, a tensor of the same {@link OperandType} as input0. * For output tensor of {@link OperandType::TENSOR_QUANT8_ASYMM}, * the following condition must be satisfied: * output_scale > input1_scale * input2_scale. */ MUL = 18, /** * Computes rectified linear activation on the input tensor element-wise. * * The output is calculated using this formula: * * output = max(0, input) * * Supported tensor {@link OperandType}: * * {@link OperandType::TENSOR_FLOAT32} * * {@link OperandType::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 a {@link OperandType::TENSOR_QUANT8_ASYMM} tensor, * the scale and zeroPoint must be the same as input0. */ RELU = 19, /** * 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 {@link OperandType}: * * {@link OperandType::TENSOR_FLOAT32} * * {@link OperandType::TENSOR_QUANT8_ASYMM} * * Supported tensor rank: up to 4. * * Inputs: * * 0: A tensor, specifying the input. * * Outputs: * * 0: The output tensor of the same shape as input0. * For a {@link OperandType::TENSOR_QUANT8_ASYMM} tensor, * the scale and zeroPoint must be the same as input0. */ RELU1 = 20, /** * 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 {@link OperandType}: * * {@link OperandType::TENSOR_FLOAT32} * * {@link OperandType::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 a {@link OperandType::TENSOR_QUANT8_ASYMM} tensor, * the scale and zeroPoint must be the same as input0. */ RELU6 = 21, /** * 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 {@link OperandType}: * * {@link OperandType::TENSOR_FLOAT32} * * {@link OperandType::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 {@link OperandType::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. * * If one component of shape is the special value -1, the size of that * dimension is computed so that the total size remains constant. In * particular, a shape of [-1] flattens into 1-D. At most one component * of shape can be -1. * * Outputs: * * 0: The output tensor, of shape specified by the input shape. * For a {@link OperandType::TENSOR_QUANT8_ASYMM} tensor, * the scale and zeroPoint must be the same as input0. */ RESHAPE = 22, /** * 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 {@link OperandType}: * * {@link OperandType::TENSOR_FLOAT32} * * Supported tensor rank: 4, with "NHWC" (i.e., Num_samples, Height, Width, * and Channels) data layout. * * Inputs (resizing by shape): * * 0: A 4-D tensor, of shape [batches, height, width, depth], specifying * the input. * * 1: An {@link OperandType::INT32} scalar, specifying the output * width of the output tensor. * * 2: An {@link OperandType::INT32} scalar, specifying the output * height of the output tensor. * * Outputs: * * 0: The output 4-D tensor, of shape * [batches, new_height, new_width, depth]. */ RESIZE_BILINEAR = 23, /** * 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 {@link OperandType}: * * {@link OperandType::TENSOR_FLOAT32} * * The input tensors must all be the same type. * * Inputs: * * 0: input. * A 2-D tensor 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 shape [num_units, input_size], where “num_units” * corresponds to the number of units. * * 2: recurrent_weights. * A 2-D tensor of shape [num_units, num_units], with columns * corresponding to the weights from each unit. * * 3: bias. * A 1-D tensor of shape [num_units]. * * 4: hidden state (in). * A 2-D tensor of shape [batch_size, num_units]. * * 5: fused_activation_function. * An optional {@link FusedActivationFunc} 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 shape [batch_size, num_units]. * * * 1: output. * A 2-D tensor of shape [batch_size, num_units]. This is effectively * the same as the current state value. */ RNN = 24, /** * 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)} * * For input tensor with rank other than 2, the activation will be applied * independently on each 1-D slice along specified dimension. * * Supported tensor {@link OperandType}: * * {@link OperandType::TENSOR_FLOAT32} * * {@link OperandType::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: A scalar, specifying the positive scaling factor for the exponent, * beta. If input0 is of {@link OperandType::TENSOR_FLOAT32} or * {@link OperandType::TENSOR_QUANT8_ASYMM}, the scalar must be of * {@link OperandType::FLOAT32}. * * Outputs: * * 0: The output tensor of same shape as input0. * For {@link OperandType::TENSOR_QUANT8_ASYMM}, * the scale must be 1.f / 256 and the zeroPoint must be 0. */ SOFTMAX = 25, /** * 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 {@link OperandType}: * * {@link OperandType::TENSOR_FLOAT32} * * {@link OperandType::TENSOR_QUANT8_ASYMM} * * Supported tensor rank: 4, with "NHWC" (i.e., Num_samples, Height, Width, * and Channels) data layout. * * Inputs: * * 0: A 4-D tensor, of shape [batches, height, width, depth_in], * specifying the input. * * 1: An {@link OperandType::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]. * For a {@link OperandType::TENSOR_QUANT8_ASYMM} tensor, * the scale and zeroPoint must be the same as input0. */ SPACE_TO_DEPTH = 26, /** * 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, * "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 {@link OperandType}: * * {@link OperandType::TENSOR_FLOAT32} * * All input tensors must be the same type. * * Inputs: * * 0: input. * A 2-D tensor 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 shape [num_units, input_size], where “num_units” * corresponds to the number of units. * * 2: weights_time. * A 2-D tensor 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 shape [num_units]. * * 4: state (in). * A 2-D tensor of shape [batch_size, (memory_size - 1) * num_units * rank]. * * 5: rank. * The rank of the SVD approximation. * * 6: fused_activation_function. * An optional {@link FusedActivationFunc} 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 the same {@link OperandType} as the inputs, with shape * [batch_size, (memory_size - 1) * num_units * rank]. * * 1: output. * A 2-D tensor of the same {@link OperandType} as the inputs, with shape * [batch_size, num_units]. */ SVDF = 27, /** * Computes hyperbolic tangent of input tensor element-wise. * * The output is calculated using this formula: * * output = tanh(input) * * Supported tensor {@link OperandType}: * * {@link OperandType::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. */ TANH = 28, /** * DEPRECATED. Since NNAPI 1.2, extensions are the preferred alternative to * OEM operation and data types. * * This operation is OEM specific. It should only be used for OEM * applications. */ OEM_OPERATION = 10000, }; /** * Fused activation function types. */ enum FusedActivationFunc : int32_t { NONE = 0, RELU = 1, RELU1 = 2, RELU6 = 3, }; /** * How an operand is used. */ enum OperandLifeTime : int32_t { /** * The operand is internal to the model. It's created by an operation and * consumed by other operations. It must be an output operand of * exactly one operation. */ TEMPORARY_VARIABLE, /** * The operand is an input of the model. It must not be an output * operand of any operation. * * An operand can't be both input and output of a model. */ MODEL_INPUT, /** * The operand is an output of the model. It must be an output * operand of exactly one operation. * * An operand can't be both input and output of a model. */ MODEL_OUTPUT, /** * The operand is a constant found in Model.operandValues. It must * not be an output operand of any operation. */ CONSTANT_COPY, /** * The operand is a constant that was specified via a Memory * object. It must not be an output operand of any operation. */ CONSTANT_REFERENCE, /** * The operand does not have a value. This is valid only for optional * arguments of operations. */ NO_VALUE, }; /** * Status of a device. */ enum DeviceStatus : int32_t { AVAILABLE, BUSY, OFFLINE, UNKNOWN, }; /** * Performance information for the reference workload. * * Used by a driver to report its performance characteristics. */ struct PerformanceInfo { /** * Ratio of the time taken by the driver to execute the * workload compared to the time the CPU would take for the * same workload. A lower number is better. */ float execTime; /** * Ratio of the energy used by the driver compared to what * the CPU would use for doing the same workload. A lower number * is better. */ float powerUsage; }; /** * The capabilities of a driver. */ struct Capabilities { /** * Driver performance when operating on float32 data. */ PerformanceInfo float32Performance; /** * Driver performance when operating on asymmetric 8-bit quantized data. */ PerformanceInfo quantized8Performance; }; /** * Describes the location of a data object. */ struct DataLocation { /** * The index of the memory pool where this location is found. */ uint32_t poolIndex; /** * Offset in bytes from the start of the pool. */ uint32_t offset; /** * The length of the data in bytes. */ uint32_t length; }; /** * Describes one operand of the model's graph. */ struct Operand { /** * Data type of the operand. */ OperandType type; /** * Dimensions of the operand. * * For a scalar operand, dimensions.size() must be 0. * * For a tensor operand, dimensions.size() must be at least 1; * however, any of the dimensions may be unspecified. * * A tensor operand with all dimensions specified has "fully * specified" dimensions. Whenever possible (i.e., whenever the * dimensions are known at model construction time), a tensor * operand should have (but is not required to have) fully * specified dimensions, in order to enable the best possible * performance. * * If a tensor operand's dimensions are not fully specified, the * dimensions of the operand are deduced from the operand * dimensions and values of the operation for which that operand * is an output. * * In the following situations, a tensor operand's dimensions must * be fully specified: * * . The operand has lifetime CONSTANT_COPY or * CONSTANT_REFERENCE. * * . The operand has lifetime MODEL_INPUT or MODEL_OUTPUT. Fully * specified dimensions must either be present in the * Operand or they must be provided in the corresponding * RequestArgument. * EXCEPTION: If the input or output is optional and omitted * (by setting the hasNoValue field of the corresponding * RequestArgument to true) then it need not have fully * specified dimensions. * * A tensor operand with some number of unspecified dimensions is * represented by setting each unspecified dimension to 0. */ vec dimensions; /** * The number of times this operand appears as an operation input. * * (For example, if this operand appears once in one operation's * input list, and three times in another operation's input list, * then numberOfConsumers = 4.) */ uint32_t numberOfConsumers; /** * Quantized scale of the operand. * * Only applicable if the operand is of type TENSOR_QUANT8_ASYMM or * TENSOR_INT32. */ float scale; /** * Quantized zero-point offset of the operand. * * Only applicable if the operand is of type TENSOR_QUANT8_ASYMM. */ int32_t zeroPoint; /** * How the operand is used. */ OperandLifeTime lifetime; /** * Where to find the data for this operand. * If the lifetime is TEMPORARY_VARIABLE, MODEL_INPUT, MODEL_OUTPUT, or * NO_VALUE: * - All the fields must be 0. * If the lifetime is CONSTANT_COPY: * - location.poolIndex is 0. * - location.offset is the offset in bytes into Model.operandValues. * - location.length is set. * If the lifetime is CONSTANT_REFERENCE: * - location.poolIndex is set. * - location.offset is the offset in bytes into the specified pool. * - location.length is set. */ DataLocation location; }; /** * Describes one operation of the model's graph. */ struct Operation { /** * The operation type. */ OperationType type; /** * Describes the table that contains the indexes of the inputs of the * operation. The offset is the index in the operandIndexes table. */ vec inputs; /** * Describes the table that contains the indexes of the outputs of the * operation. The offset is the index in the operandIndexes table. */ vec outputs; }; /** * A Neural Network Model. * * This includes not only the execution graph, but also constant data such as * weights or scalars added at construction time. The only information that * might not be known is the shape of the input tensors. */ struct Model { /** * All operands included in the model. */ vec operands; /** * All operations included in the model. * * The operations are sorted into execution order. Every operand * with lifetime MODEL_OUTPUT or TEMPORARY_VARIABLE must be * written before it is read. */ vec operations; /** * Input indexes of the model. There must be at least one. * * Each value corresponds to the index of the operand in "operands". */ vec inputIndexes; /** * Output indexes of the model. There must be at least one. * * Each value corresponds to the index of the operand in "operands". */ vec outputIndexes; /** * A byte buffer containing operand data that were copied into the model. * * An operand's value must be located here if and only if Operand::lifetime * equals OperandLifeTime::CONSTANT_COPY. */ vec operandValues; /** * A collection of shared memory pools containing operand values. * * An operand's value must be located here if and only if Operand::lifetime * equals OperandLifeTime::CONSTANT_REFERENCE. */ vec pools; }; /** * Metadata information specifying the location of the input or output data and * any updates to the input or output operand. */ struct RequestArgument { /** * If true, the argument does not have a value. This can be used for * operations that take optional arguments. If true, the fields of location * are set to 0 and the dimensions vector is left empty. */ bool hasNoValue; /** * The location within one of the memory pools passed in the Request. */ DataLocation location; /** * Updated dimension information. * * If dimensions.size() > 0, dimension information was provided * along with the argument. This can be the case for models that * accept inputs of varying size. This can't change the rank, just * the value of the dimensions that were unspecified in the * model. If dimensions.size() > 0, then all dimensions must be * specified here; and any dimension that was specified in the * model must have the same value here. * * If the dimensions in the model are not fully specified, then * they must be fully specified here, unless hasNoValue is set to * true. If the dimensions in the model are fully specified, then * either dimensions.size() may be 0, or the dimensions in the * model must be identical to the dimensions here. */ vec dimensions; }; /** * Inputs to be sent to and outputs to be retrieved from a prepared model. * * A Request serves two primary tasks: * 1) Provides the input and output data to be used when executing the model. * 2) Specifies any updates to the input operand metadata that were left * unspecified at model preparation time. * * An output must not overlap with any other output, with an input, or * with an operand of lifetime CONSTANT_REFERENCE. */ struct Request { /** * Input data and information to be used in the execution of a prepared * model. * * The index of the input corresponds to the index in Model.inputIndexes. * E.g., input[i] corresponds to Model.inputIndexes[i]. */ vec inputs; /** * Output data and information to be used in the execution of a prepared * model. * * The index of the output corresponds to the index in Model.outputIndexes. * E.g., output[i] corresponds to Model.outputIndexes[i]. */ vec outputs; /** * A collection of shared memory pools containing operand data for both the * inputs and the outputs to a model. */ vec pools; }; /** * Return status of a function. */ enum ErrorStatus : int32_t { NONE, DEVICE_UNAVAILABLE, GENERAL_FAILURE, OUTPUT_INSUFFICIENT_SIZE, INVALID_ARGUMENT, };