You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.

256 lines
8.6 KiB

/*
* Copyright (C) 2018 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.
*/
#include "utils/tflite/string_projection_base.h"
#include <cmath>
#include <cstdint>
#include <cstring>
#include <memory>
#include <string>
#include <unordered_map>
#include <vector>
#include "utils/hash/cityhash.h"
#include "utils/hash/farmhash.h"
#include "flatbuffers/flexbuffers.h"
#include "tensorflow/lite/context.h"
#include "tensorflow_models/seq_flow_lite/tflite_ops/quantization_util.h"
namespace tflite {
namespace ops {
namespace custom {
namespace libtextclassifier3 {
namespace string_projection {
namespace {
const int32_t kMaxInputChars = 300;
const int kOutputLabel = 0;
const char kFastHash[] = "[DEV] FastHash";
const char kAXB[] = "[DEV] AXB";
const int kSeedSize = sizeof(float);
const int kInputItemBytes = sizeof(int32_t);
const int kKeyBytes = sizeof(float) + sizeof(int32_t);
} // namespace
StringProjectionOpBase::StringProjectionOpBase(
const flexbuffers::Map& custom_options)
: hash_function_(custom_options["hash_function"].AsTypedVector()),
num_hash_(custom_options["num_hash"].AsInt32()),
num_bits_(custom_options["num_bits"].AsInt32()),
binary_projection_(custom_options["binary_projection"].AsBool()),
hash_method_(custom_options["hash_method"].ToString()),
axb_scale_(custom_options["axb_scale"].AsFloat()) {
skip_gram_params_ = {
.ngram_size = custom_options["ngram_size"].AsInt32(),
.max_skip_size = custom_options["max_skip_size"].AsInt32(),
.include_all_ngrams = custom_options["include_all_ngrams"].AsBool(),
.preprocess = custom_options["preprocess"].AsBool(),
.char_level = custom_options["char_level"].AsBool(),
.remove_punctuation = custom_options["remove_punctuation"].AsBool(),
.max_input_chars = kMaxInputChars,
};
}
void StringProjectionOpBase::GetFeatureWeights(
const std::unordered_map<std::string, int>& feature_counts,
std::vector<std::vector<int64_t>>* batch_ids,
std::vector<std::vector<float>>* batch_weights) {
std::vector<int64_t> ids;
std::vector<float> weights;
for (const auto& iter : feature_counts) {
if (hash_method_ == kFastHash || hash_method_ == kAXB) {
int32_t feature_id =
tc3farmhash::CityHash64(iter.first.c_str(), iter.first.size());
ids.push_back(feature_id);
weights.push_back(iter.second);
} else {
int64_t feature_id =
tc3farmhash::Fingerprint64(iter.first.c_str(), iter.first.size());
ids.push_back(feature_id);
weights.push_back(iter.second);
}
}
batch_ids->push_back(ids);
batch_weights->push_back(weights);
}
void StringProjectionOpBase::DenseLshProjection(
const int batch_size, const std::vector<std::vector<int64_t>>& batch_ids,
const std::vector<std::vector<float>>& batch_weights,
TfLiteTensor* output) {
auto key = std::unique_ptr<char[]>(
new char[kKeyBytes]); // NOLINT: modernize-make-unique
if (output->type == kTfLiteFloat32) {
for (int batch = 0; batch < batch_size; ++batch) {
const std::vector<int64_t>& input = batch_ids[batch];
const std::vector<float>& weight = batch_weights[batch];
for (int i = 0; i < num_hash_; i++) {
for (int j = 0; j < num_bits_; j++) {
int hash_bit = i * num_bits_ + j;
float seed = hash_function_[hash_bit].AsFloat();
float bit = running_sign_bit(input, weight, seed, key.get());
output->data.f[batch * num_hash_ * num_bits_ + hash_bit] = bit;
}
}
}
} else if (output->type == kTfLiteUInt8) {
const float inverse_scale = 1.0 / output->params.scale;
for (int batch = 0; batch < batch_size; ++batch) {
const std::vector<int64_t>& input = batch_ids[batch];
const std::vector<float>& weight = batch_weights[batch];
for (int i = 0; i < num_hash_; i++) {
for (int j = 0; j < num_bits_; j++) {
int hash_bit = i * num_bits_ + j;
float seed = hash_function_[hash_bit].AsFloat();
float bit = running_sign_bit(input, weight, seed, key.get());
output->data.uint8[batch * num_hash_ * num_bits_ + hash_bit] =
seq_flow_lite::PodQuantize(bit, output->params.zero_point,
inverse_scale);
}
}
}
}
}
namespace {
int32_t hash32(int32_t value, uint32_t seed) {
uint32_t hash = value;
hash = (hash ^ 61) ^ (hash >> 16);
hash = hash + (hash << 3);
hash = hash ^ (hash >> 4);
hash = hash * seed;
hash = hash ^ (hash >> 15);
return static_cast<int32_t>(hash);
}
double axb(int32_t value, float seed, float scale) {
// Convert seed to a larger scale of range, multiplier is 1e5 to avoid
// precision difference on different hardware.
int64_t hash_signature =
static_cast<int64_t>(scale) * static_cast<int64_t>(seed * 1e5) * value;
hash_signature %= 0x100000000;
hash_signature = fabs(hash_signature);
if (hash_signature >= 0x80000000) {
hash_signature -= 0x100000000;
}
return hash_signature;
}
} // namespace
// Compute sign bit of dot product of hash(seed, input) and weight.
float StringProjectionOpBase::running_sign_bit(
const std::vector<int64_t>& input, const std::vector<float>& weight,
float seed, char* key) {
double score = 0.0;
memcpy(key, &seed, kSeedSize);
int cnt = 0;
for (int i = 0; i < input.size(); ++i) {
if (weight[i] == 0.0) continue;
cnt++;
const int32_t curr_input = input[i];
memcpy(key + kSeedSize, &curr_input, kInputItemBytes);
// Create running hash id and value for current dimension.
if (hash_method_ == kFastHash) {
int32_t hash_signature =
hash32(input[i], *reinterpret_cast<uint32_t*>(&seed));
score += static_cast<double>(weight[i]) * hash_signature;
} else if (hash_method_ == kAXB) {
score += weight[i] * axb(input[i], seed, axb_scale_);
} else {
int64_t hash_signature = tc3farmhash::Fingerprint64(key, kKeyBytes);
double running_value = static_cast<double>(hash_signature);
score += weight[i] * running_value;
}
}
const double inverse_normalizer = 0.00000000046566129;
if (!binary_projection_) {
if (hash_method_ == kAXB) {
return tanh(score / cnt * inverse_normalizer);
} else {
return tanh(score * inverse_normalizer);
}
}
return (score > 0) ? 1 : 0;
}
void Free(TfLiteContext* context, void* buffer) {
delete reinterpret_cast<StringProjectionOpBase*>(buffer);
}
TfLiteStatus Resize(TfLiteContext* context, TfLiteNode* node) {
auto* op = reinterpret_cast<StringProjectionOpBase*>(node->user_data);
// The shape of the output should be the shape of the input + a new inner
// dimension equal to the number of features.
TfLiteIntArray* input_shape = op->GetInputShape(context, node);
TfLiteIntArray* output_shape = TfLiteIntArrayCreate(input_shape->size + 1);
for (int i = 0; i < input_shape->size; ++i) {
output_shape->data[i] = input_shape->data[i];
}
output_shape->data[input_shape->size] = op->num_hash() * op->num_bits();
context->ResizeTensor(context,
&context->tensors[node->outputs->data[kOutputLabel]],
output_shape);
return kTfLiteOk;
}
TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) {
auto* op = reinterpret_cast<StringProjectionOpBase*>(node->user_data);
TfLiteTensor* label = &context->tensors[node->outputs->data[kOutputLabel]];
TfLiteIntArray* input_shape = op->GetInputShape(context, node);
int input_size = 1;
for (int i = 0; i < input_shape->size; ++i) {
input_size *= input_shape->data[i];
}
TF_LITE_ENSURE_STATUS(op->InitializeInput(context, node));
std::vector<std::vector<int64_t>> batch_ids;
std::vector<std::vector<float>> batch_weights;
for (int i = 0; i < input_size; ++i) {
std::unordered_map<std::string, int> feature_counts =
op->ExtractSkipGrams(i);
op->GetFeatureWeights(feature_counts, &batch_ids, &batch_weights);
}
op->DenseLshProjection(input_size, batch_ids, batch_weights, label);
op->FinalizeInput();
return kTfLiteOk;
}
} // namespace string_projection
} // namespace libtextclassifier3
} // namespace custom
} // namespace ops
} // namespace tflite