graphengine/third_party/fwkacllib/inc/ops/candidate_sampling_ops.h

/**
 * Copyright 2019-2020 Huawei Technologies Co., Ltd
 *
 * 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.
 */

#ifndef GE_OP_CANDIDATE_SAMPLING_OPS_H_
#define GE_OP_CANDIDATE_SAMPLING_OPS_H_

#include "graph/operator_reg.h"

namespace ge {

/**
*@brief Generates labels for candidate sampling with a learned unigram distribution.

*@par Inputs:
*The input true_classes must be two-dimensional matrices. Inputs include: \n
*true_classes:A batch_size * num_true matrix, in which each row contains the IDs of the num_true target_classes in the corresponding original label.

*@par Attributes:
*@li num_true:Number of true labels per context.
*@li num_sampled:Number of candidates to randomly sample.
*@li unique:If unique is true, we sample with rejection, so that all sampled candidates in a batch are unique. This requires some approximation to estimate the post-rejection sampling probabilities.
*@li range_max:The sampler will sample integers from the interval [0, range_max).
*@li seed:If either seed or seed2 are set to be non-zero.
*@li seed2:An second seed to avoid seed collision.

*@par Outputs:
*sampled_candidates:A vector of length num_sampled, in which each element is the ID of a sampled candidate.
*true_expected_count:A batch_size * num_true matrix, representing the number of times each candidate is expected to occur in a batch of sampled candidates. If unique=true, then this is a probability.
*sampled_expected_count:A vector of length num_sampled, for each sampled candidate representing the number of times the candidate is expected to occur in a batch of sampled candidates. If unique=true, then this is a probability.

*@attention Constraints: \n
*-The implementation for ThreadUnsafeUnigramCandidateSampler on Ascend uses AI CPU, with bad performance. \n

*@par Quantization supported or not
*Not supported
*@par Quantized inference supported or not
*Supported
*@par L2 convergence supported or not
*@par Multiple batches supported or not
*/

REG_OP(ThreadUnsafeUnigramCandidateSampler)
    .INPUT(true_classes, TensorType({ DT_INT64 }))
    .OUTPUT(sampled_candidates, TensorType({ DT_INT64 }))
    .OUTPUT(true_expected_count, TensorType({ DT_FLOAT }))
    .OUTPUT(sampled_expected_count, TensorType({ DT_FLOAT }))
    .REQUIRED_ATTR(num_true, Int)
    .REQUIRED_ATTR(num_sampled, Int)
    .REQUIRED_ATTR(unique, Bool)
    .REQUIRED_ATTR(range_max, Int)
    .ATTR(seed, Int, 0)
    .ATTR(seed2, Int, 0)
    .OP_END_FACTORY_REG(ThreadUnsafeUnigramCandidateSampler)

/**
*@brief Generates labels for candidate sampling with a learned unigram distribution.

*@par Inputs:
*The input true_classes must be two-dimensional matrices. Inputs include: \n
*true_classes:A batch_size * num_true matrix, in which each row contains the IDs of the num_true target_classes in the corresponding original label.

*@par Attributes:
*@li num_true:Number of true labels per context.
*@li num_sampled:Number of candidates to randomly sample.
*@li unique:If unique is true, we sample with rejection, so that all sampled candidates in a batch are unique. This requires some approximation to estimate the post-rejection sampling probabilities.
*@li range_max:The sampler will sample integers from the interval [0, range_max).
*@li seed:If either seed or seed2 are set to be non-zero.
*@li seed2:An second seed to avoid seed collision.

*@par Outputs:
*@li sampled_candidates:A vector of length num_sampled, in which each element is the ID of a sampled candidate.
*@li true_expected_count:A batch_size * num_true matrix, representing the number of times each candidate is expected to occur in a batch of sampled candidates. If unique=true, then this is a probability.
*@li sampled_expected_count:A vector of length num_sampled, for each sampled candidate representing the number of times the candidate is expected to occur in a batch of sampled candidates. If unique=true, then this is a probability.

*@attention Constraints: \n
*-The implementation for UniformCandidateSampler on Ascend uses AI CPU, with bad performance. \n

*@par Quantization supported or not
*Not supported
*@par Quantized inference supported or not
*Supported
*@par L2 convergence supported or not
*@par Multiple batches supported or not
*/

REG_OP(UniformCandidateSampler)
    .INPUT(true_classes, TensorType({ DT_INT64 }))
    .OUTPUT(sampled_candidates, TensorType({ DT_INT64 }))
    .OUTPUT(true_expected_count, TensorType({ DT_FLOAT }))
    .OUTPUT(sampled_expected_count, TensorType({ DT_FLOAT }))
    .REQUIRED_ATTR(num_true, Int)
    .REQUIRED_ATTR(num_sampled, Int)
    .REQUIRED_ATTR(unique, Bool)
    .REQUIRED_ATTR(range_max, Int)
    .ATTR(seed, Int, 0)
    .ATTR(seed2, Int, 0)
    .OP_END_FACTORY_REG(UniformCandidateSampler)

/**
*@brief Generates labels for candidate sampling with a learned unigram distribution.

*@par Inputs:
*The input true_classes can be two-dimensional matrices. Inputs include: \n
*true_classes:A batch_size * num_true matrix, in which each row contains the IDs of the num_true target_classes in the corresponding original label.

*@par Attributes:
*@li num_true:Number of true labels per context.
*@li num_sampled:Number of candidates to randomly sample.
*@li unique:If unique is true, we sample with rejection, so that all sampled candidates in a batch are unique. This requires some approximation to estimate the post-rejection sampling probabilities.
*@li range_max:The sampler will sample integers from the interval [0, range_max).
*@li vocab_file:Each valid line in this file (which should have a CSV-like format) corresponds to a valid word ID. IDs are in sequential order, starting from num_reserved_ids.
*@li distortion:The distortion is used to skew the unigram probability distribution. Each weight is first raised to the distortion's power before adding to the internal unigram distribution.
*@li num_reserved_ids:Optionally some reserved IDs can be added in the range [0, ..., num_reserved_ids) by the users. One use case is that a special unknown word token is used as ID 0.
*@li num_shards:A sampler can be used to sample from a subset of the original range in order to speed up the whole computation through parallelism.
*@li shard:A sampler can be used to sample from a subset of the original range in order to speed up the whole computation through parallelism.
*@li unigrams:A list of unigram counts or probabilities, one per ID in sequential order.
*@li seed:If either seed or seed2 are set to be non-zero.
*@li seed2:An second seed to avoid seed collision.

*@par Outputs:
*@li sampled_candidates:A vector of length num_sampled, in which each element is the ID of a sampled candidate.
*@li true_expected_count:A batch_size * num_true matrix, representing the number of times each candidate is expected to occur in a batch of sampled candidates. If unique=true, then this is a probability.
*@li sampled_expected_count:A vector of length num_sampled, for each sampled candidate representing the number of times the candidate is expected to occur in a batch of sampled candidates. If unique=true, then this is a probability.

*@attention Constraints: \n
*-The implementation for FixedUnigramCandidateSampler on Ascend uses AI CPU, with bad performance. \n

*@par Quantization supported or not
*Not supported
*@par Quantized inference supported or not
*Supported
*@par L2 convergence supported or not
*@par Multiple batches supported or not
*/

REG_OP(FixedUnigramCandidateSampler)
    .INPUT(true_classes, TensorType({ DT_INT64 }))
    .OUTPUT(sampled_candidates, TensorType({ DT_INT64 }))
    .OUTPUT(true_expected_count, TensorType({ DT_FLOAT }))
    .OUTPUT(sampled_expected_count, TensorType({ DT_FLOAT }))
    .ATTR(num_true, Int, 0)
    .ATTR(num_sampled, Int, 0)
    .ATTR(unique, Bool, false)
    .ATTR(range_max, Int, 0)
    .ATTR(vocab_file, String, "")
    .ATTR(distortion, Float, 1.0)
    .ATTR(num_reserved_ids, Int, 0)
    .ATTR(num_shards, Int, 1)
    .ATTR(shard, Int, 0)
    .REQUIRED_ATTR(unigrams, ListFloat)
    .ATTR(seed, Int, 0)
    .ATTR(seed2, Int, 0)
    .OP_END_FACTORY_REG(FixedUnigramCandidateSampler)

/**
*@brief Generates labels for candidate sampling with a learned unigram distribution.

*@par Inputs:
*The input true_classes can be two-dimensional matrices. Inputs include: \n
*true_classes:A batch_size * num_true matrix, in which each row contains the IDs of the num_true target_classes in the corresponding original label.

*@par Attributes:
*@li num_true:Number of true labels per context.
*@li num_sampled:Number of candidates to randomly sample.
*@li unique:If unique is true, we sample with rejection, so that all sampled candidates in a batch are unique. This requires some approximation to estimate the post-rejection sampling probabilities.
*@li range_max:The sampler will sample integers from the interval [0, range_max).
*@li seed:If either seed or seed2 are set to be non-zero.
*@li seed2:An second seed to avoid seed collision.

*@par Outputs:
*@li sampled_candidates:A vector of length num_sampled, in which each element is the ID of a sampled candidate.
*@li true_expected_count:A batch_size * num_true matrix, representing the number of times each candidate is expected to occur in a batch of sampled candidates. If unique=true, then this is a probability.
*@li sampled_expected_count:A vector of length num_sampled, for each sampled candidate representing the number of times the candidate is expected to occur in a batch of sampled candidates. If unique=true, then this is a probability.

*@attention Constraints: \n
*-The implementation for LearnedUnigramCandidateSampler on Ascend uses AI CPU, with bad performance. \n

*@par Quantization supported or not
*Not supported
*@par Quantized inference supported or not
*Supported
*@par L2 convergence supported or not
*@par Multiple batches supported or not
*/

REG_OP(LearnedUnigramCandidateSampler)
    .INPUT(true_classes, TensorType({ DT_INT64 }))
    .OUTPUT(sampled_candidates, TensorType({ DT_INT64 }))
    .OUTPUT(true_expected_count, TensorType({ DT_FLOAT }))
    .OUTPUT(sampled_expected_count, TensorType({ DT_FLOAT }))
    .REQUIRED_ATTR(num_true, Int)
    .REQUIRED_ATTR(num_sampled, Int)
    .REQUIRED_ATTR(unique, Bool)
    .REQUIRED_ATTR(range_max, Int)
    .ATTR(seed, Int, 0)
    .ATTR(seed2, Int, 0)
    .OP_END_FACTORY_REG(LearnedUnigramCandidateSampler)

/**
*@brief Generates labels for candidate sampling with a log-uniform distribution.

*@par Inputs:
*The input true_classes can be two-dimensional matrices. Inputs include: \n
*true_classes:A batch_size * num_true matrix, in which each row contains the IDs of the num_true target_classes in the corresponding original label.

*@par Attributes:
*@li num_true:Number of true labels per context.
*@li num_sampled:Number of candidates to randomly sample.
*@li unique:If unique is true, we sample with rejection, so that all sampled candidates in a batch are unique. This requires some approximation to estimate the post-rejection sampling probabilities.
*@li range_max:The sampler will sample integers from the interval [0, range_max).
*@li seed:If either seed or seed2 are set to be non-zero.
*@li seed2:An second seed to avoid seed collision.

*@par Outputs:
*@li sampled_candidates:A vector of length num_sampled, in which each element is the ID of a sampled candidate.
*@li true_expected_count:A batch_size * num_true matrix, representing the number of times each candidate is expected to occur in a batch of sampled candidates. If unique=true, then this is a probability.
*@li sampled_expected_count:A vector of length num_sampled, for each sampled candidate representing the number of times the candidate is expected to occur in a batch of sampled candidates. If unique=true, then this is a probability.

*@attention Constraints:\n
*-The implementation for LogUniformCandidateSampler on Ascend uses AI CPU, with bad performance.\n

*@par Quantization supported or not
*Not supported
*@par Quantized inference supported or not
*Supported
*@par L2 convergence supported or not
*@par Multiple batches supported or not
*/

REG_OP(LogUniformCandidateSampler)
    .INPUT(true_classes, TensorType({ DT_INT64 }))
    .OUTPUT(sampled_candidates, TensorType({ DT_INT64 }))
    .OUTPUT(true_expected_count, TensorType({ DT_FLOAT }))
    .OUTPUT(sampled_expected_count, TensorType({ DT_FLOAT }))
    .REQUIRED_ATTR(num_true, Int)
    .REQUIRED_ATTR(num_sampled, Int)
    .REQUIRED_ATTR(unique, Bool)
    .REQUIRED_ATTR(range_max, Int)
    .ATTR(seed, Int, 0)
    .ATTR(seed2, Int, 0)
    .OP_END_FACTORY_REG(LogUniformCandidateSampler)

/**
*@brief Generates labels for candidate sampling with a learned unigram distribution.

*@par Inputs:
*The input true_classes can be two-dimensional matrices. Inputs include: \n
*true_classes:A batch_size * num_true matrix, in which each row contains the IDs of the num_true target_classes in the corresponding original label.

*@par Attributes:
*@li num_true:Number of true labels per context.
*@li num_sampled:Number of candidates to randomly sample.
*@li unique:If unique is true, we sample with rejection, so that all sampled candidates in a batch are unique. This requires some approximation to estimate the post-rejection sampling probabilities.
*@li seed:If either seed or seed2 are set to be non-zero.
*@li seed2:An second seed to avoid seed collision.

*@par Outputs:
*@li sampled_candidates:A vector of length num_sampled, in which each element is the ID of a sampled candidate.
*@li true_expected_count:A batch_size * num_true matrix, representing the number of times each candidate is expected to occur in a batch of sampled candidates. If unique=true, then this is a probability.
*@li sampled_expected_count:A vector of length num_sampled, for each sampled candidate representing the number of times the candidate is expected to occur in a batch of sampled candidates. If unique=true, then this is a probability.

*@attention Constraints:\n
*-The implementation for AllCandidateSampler on Ascend uses AI CPU, with bad performance.\n

*@par Quantization supported or not
*Not supported
*@par Quantized inference supported or not
*Supported
*@par L2 convergence supported or not
*@par Multiple batches supported or not
*/

REG_OP(AllCandidateSampler)
    .INPUT(true_classes, TensorType({ DT_INT64 }))
    .OUTPUT(sampled_candidates, TensorType({ DT_INT64 }))
    .OUTPUT(true_expected_count, TensorType({ DT_FLOAT }))
    .OUTPUT(sampled_expected_count, TensorType({ DT_FLOAT }))
    .REQUIRED_ATTR(num_true, Int)
    .REQUIRED_ATTR(num_sampled, Int)
    .REQUIRED_ATTR(unique, Bool)
    .ATTR(seed, Int, 0)
    .ATTR(seed2, Int, 0)
    .OP_END_FACTORY_REG(AllCandidateSampler)

/**
*@brief Computes the ids of the positions in sampled_candidates that match true_labels.

*@par Inputs:
* @li The input true_classes can be two-dimensional matrices. Inputs include: \n
* @li true_classes:The true_classes output of UnpackSparseLabels. \n
* @li sampled_candidates:The sampled_candidates output of CandidateSampler. \n

*@par Attributes:
*@li num_true:Number of true labels per context.
*@li seed:If either seed or seed2 are set to be non-zero.
*@li seed2:An second seed to avoid seed collision.

*@par Outputs:
* @li indices:A vector of indices corresponding to rows of true_candidates.
* @li ids:A vector of IDs of positions in sampled_candidates that match a true_label for the row with the corresponding index in indices.
* @li weights:A vector of the same length as indices and ids, in which each element is -FLOAT_MAX.

*@attention Constraints:\n
*-The implementation for ComputeAccidentalHits on Ascend uses AI CPU, with bad performance.\n

*@par Quantization supported or not
*Not supported
*@par Quantized inference supported or not
*Supported
*@par L2 convergence supported or not
*@par Multiple batches supported or not
*/

REG_OP(ComputeAccidentalHits)
    .INPUT(true_classes, TensorType({ DT_INT64 }))
    .INPUT(sampled_candidates, TensorType({ DT_INT64 }))
    .OUTPUT(indices, TensorType({ DT_INT32 }))
    .OUTPUT(ids, TensorType({ DT_INT64 }))
    .OUTPUT(weights, TensorType({ DT_FLOAT }))
    .REQUIRED_ATTR(num_true, Int)
    .ATTR(seed, Int, 0)
    .ATTR(seed2, Int, 0)
    .OP_END_FACTORY_REG(ComputeAccidentalHits)

}  // namespace ge

#endif  // GE_OP_CANDIDATE_SAMPLING_OPS_H_