/*
* NiuTrans.Tensor - an open-source tensor library
* Copyright (C) 2018, Natural Language Processing Lab, Northestern University.
* All rights reserved.
*
* 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.
*/
/*
* $Created by: XIAO Tong (email: xiaotong@mail.neu.edu.cn) 2018-05-08
*/
#include <math.h>
#include "SetData.h"
#include "SetData.cuh"
#include "../../XUtility.h"
#include "../movement/CopyValues.h"
#if !defined( WIN32 ) && !defined( _WIN32 )
#include "sys/time.h"
#include "time.h"
#include "iconv.h"
#else
#include "time.h"
#include "windows.h"
#include "process.h"
#endif
namespace nts{ // namespace nts(NiuTrans.Tensor)
/*
Fills the input Tensor or Variable with values according to the method described in
"Understanding the difficulty of training deep feedforward neural networks" - Glorot, X. & Bengio, Y. (2010),
using a uniform distribution. The resulting tensor will have values sampled from :math:`U(-a, a)`
where :math:`a = gain \times \sqrt{2 / (fan\_in + fan\_out)} \times \sqrt{3}`. Also known as Glorot initialisation.
>> tensor - the tensor whose data array would be initialized
>> gain - an optional scaling factor
*/
void _SetDataFanInOut(XTensor * tensor, DTYPE gain)
{
CheckNTErrors(tensor->dataType == X_FLOAT, "the tensor must be in X_FLOAT!");
CheckNTErrors(tensor->order >= 2, "the tensor dimension must be no less than 2!");
int fanIn = 1;
int fanOut = 1;
int order = tensor->order;
if (order == 2) {
fanIn = tensor->dimSize[1];
fanOut = tensor->dimSize[0];
}
else {
int numInputFmaps = tensor->dimSize[1];
int numOutputFmaps = tensor->dimSize[0];
int receptiveFieldSize = 0;
for (int i = 2; i < order; i++)
receptiveFieldSize += tensor->dimSize[i];
fanIn = numInputFmaps * receptiveFieldSize;
fanOut = numOutputFmaps * receptiveFieldSize;
}
DTYPE std = gain * (float)sqrt(2.0/(fanIn + fanOut));
DTYPE a = (DTYPE)sqrt(3.0) * std;
_SetDataRand(tensor, -a, a);
}
/*
generate data items with a fixed value p
>> tensor - the tensor whose data array would be initialized
>> p - pointer to the number for initializing the tensor
*/
void _SetDataFixed(XTensor * tensor, void * valuePointer)
{
int num = tensor->unitNum;
if(tensor->dataType == X_INT){
int p = *(int*)valuePointer;
if(tensor->devID < 0){
int * d = (int*)tensor->data;
if(num % 4 == 0){
for(int i = 0; i < num; i += 4){
d[i] = p;
d[i + 1] = p;
d[i + 2] = p;
d[i + 3] = p;
}
}
else{
for(int i = 0; i < num; i++)
d[i] = p;
}
}
else{
#ifdef USE_CUDA
_CudaSetDataFixedInt(tensor, p);
#endif
}
}
else if(tensor->dataType == X_FLOAT){
float p = *(float*)valuePointer;
if(tensor->devID < 0){
float * d = (float*)tensor->data;
if(num % 4 == 0){
for(int i = 0; i < num; i += 4){
d[i] = p;
d[i + 1] = p;
d[i + 2] = p;
d[i + 3] = p;
}
}
else{
for(int i = 0; i < num; i++)
d[i] = p;
}
}
else{
#ifdef USE_CUDA
_CudaSetDataFixedFloat(tensor, p);
#endif
}
}
else if(tensor->dataType == X_DOUBLE){
double p = *(double*)valuePointer;
if(tensor->devID < 0){
double * d = (double*)tensor->data;
if(num % 4 == 0){
for(int i = 0; i < num; i += 4){
d[i] = p;
d[i + 1] = p;
d[i + 2] = p;
d[i + 3] = p;
}
}
else{
for(int i = 0; i < num; i++)
d[i] = p;
}
}
else{
#ifdef USE_CUDA
_CudaSetDataFixedDouble(tensor, p);
#endif
}
}
else{
ShowNTErrors("TODO");
}
}
/*
generate data items with a fixed value p (in default type)
>> tensor - the tensor whose data array would be initialized
>> p - number in default type
*/
void SetDataFixed(XTensor &tensor, DTYPE p)
{
_SetDataFixed(&tensor, &p);
}
/*
generate data items with a fixed value p (in integer)
>> tensor - the tensor whose data array would be initialized
>> p - an int-valued number
*/
void _SetDataFixedInt(XTensor * tensor, int p)
{
CheckNTErrors(tensor->dataType == X_INT, "the tensor must be in X_INT!");
if(p == 0)
tensor->SetZeroAll();
else
_SetDataFixed(tensor, &p);
}
/*
generate data items with a fixed value p (in float)
>> tensor - the tensor whose data array would be initialized
>> p - a float-valued number
*/
void _SetDataFixedFloat(XTensor * tensor, float p)
{
CheckNTErrors(tensor->dataType == X_FLOAT, "the tensor must be in X_FLOAT!");
if(p == 0)
tensor->SetZeroAll();
else
_SetDataFixed(tensor, &p);
}
/*
generate data items with a fixed value p (in double)
>> tensor - the tensor whose data array would be initialized
>> p - a double-valued number
*/
void _SetDataFixedDouble(XTensor * tensor, double p)
{
CheckNTErrors(tensor->dataType == X_DOUBLE, "the tensor must be in X_DOUBLE!");
if(p == 0)
tensor->SetZeroAll();
else
_SetDataFixed(tensor, &p);
}
/*
generate data as lower triangular matrics for last two dimensions
>> tensor - the tensor whose data to be set
>> p - the value for each entry of the lower triangular matrics
>> shift - the offset from diagonal
e.g., for a 3* 3 tensor,
when p = 1 ans shift = 0, we have
1 0 0
1 1 0
1 1 1
when p = 2 and shift = -1, we have
0 0 0
2 0 0
2 2 0
*/
void _SetDataLowTri(XTensor * tensor, DTYPE p, int shift)
{
int n = tensor->order;
CheckNTErrors(tensor->dataType == DEFAULT_DTYPE, "TODO!");
CheckNTErrors(n >= 2, "The tensor must have a order no less than 2!");
CheckNTErrors(tensor->GetDim(n - 1) == tensor->GetDim(n - 2),
"The last two dimensions must be of the same size!");
if(tensor->devID < 0){
int l = tensor->GetDim(-1);
int blockNum = 1;
int blockSize = l * l;
for(int i = 0; i < n - 2; i++)
blockNum *= tensor->GetDim(i);
for(int i = 0; i < blockNum; i++){
DTYPE * d = (DTYPE*)tensor->data + i * blockSize;
for(int row = 0; row < l; row++){
for(int col = 0; col < row + shift; col++){
d[row * l + col] = row;
}
for(int col = row + shift; col < l; col++){
d[row * l + col] = 0;
}
}
}
}
else{
#ifdef USE_CUDA
_CudaSetDataLowTri(tensor, p, shift);
#endif
}
}
/*
generate data items with a uniform distribution in [lower, upper]
>> tensor - the tensor whose data array would be initialized
>> lower - lower value of the range
>> upper - upper value of the range
*/
void _SetDataRand(XTensor * tensor, DTYPE lower, DTYPE upper)
{
CheckNTErrors(upper > lower, "the high value must be greater than low value!");
if(tensor == NULL)
return;
/* GPU code */
if(tensor->devID < 0){
DTYPE variance = upper - lower;
if(tensor->dataType == X_FLOAT){
float * d = (float*)tensor->data;
for(int i = 0; i < tensor->unitNum; i++){
d[i] = variance * ((float)rand()/RAND_MAX) + lower;
}
}
else if(tensor->dataType == X_DOUBLE){
double * d = (double*)tensor->data;
for(int i = 0; i < tensor->unitNum; i++){
d[i] = variance * ((double)rand()/RAND_MAX) + lower;
}
}
else{
ShowNTErrors("TODO");
}
}
/*
GPU code
The trick here is that initialize the data on a temperary tensor on CPU.
The CPU data is then copied to GPU.
TODO: generate data points on GPUs straightforwardly.
*/
else{
#ifdef USE_CUDA
_CudaSetDataRand(tensor, lower, upper);
#endif
//XTensor * t2 = NewTensor(tensor->order, tensor->dimSize, tensor->dataType, tensor->denseRatio, -1);
//_SetDataRand(t2, low, high);
//_CopyValues(t2, tensor);
//delete t2;
}
}
/*
generate data items with a normal distribution with specified mean and standard deviation
>> mean - mean or expectation of the distribution
>> standardDeviation - standard deviation of the distribution
*/
void _SetDataRandN(XTensor * tensor, DTYPE mean, DTYPE standardDeviation)
{
// TODO: rewrite it and add cuda code!!!!!!!
tensor->SetDataRandn(mean, standardDeviation);
}
} // namespace nts(NiuTrans.Tensor)