|
<html><head><title>dlib C++ Library - tensor_tools.h</title></head><body bgcolor='white'><pre> |
|
<font color='#009900'>// Copyright (C) 2015 Davis E. King ([email protected]) |
|
</font><font color='#009900'>// License: Boost Software License See LICENSE.txt for the full license. |
|
</font><font color='#0000FF'>#ifndef</font> DLIB_TeNSOR_TOOLS_H_ |
|
<font color='#0000FF'>#define</font> DLIB_TeNSOR_TOOLS_H_ |
|
|
|
<font color='#0000FF'>#include</font> "<a style='text-decoration:none' href='tensor.h.html'>tensor.h</a>" |
|
<font color='#0000FF'>#include</font> "<a style='text-decoration:none' href='cudnn_dlibapi.h.html'>cudnn_dlibapi.h</a>" |
|
<font color='#0000FF'>#include</font> "<a style='text-decoration:none' href='cublas_dlibapi.h.html'>cublas_dlibapi.h</a>" |
|
<font color='#0000FF'>#include</font> "<a style='text-decoration:none' href='cusolver_dlibapi.h.html'>cusolver_dlibapi.h</a>" |
|
<font color='#0000FF'>#include</font> "<a style='text-decoration:none' href='curand_dlibapi.h.html'>curand_dlibapi.h</a>" |
|
<font color='#0000FF'>#include</font> "<a style='text-decoration:none' href='cpu_dlib.h.html'>cpu_dlib.h</a>" |
|
<font color='#0000FF'>#include</font> "<a style='text-decoration:none' href='cuda_dlib.h.html'>cuda_dlib.h</a>" |
|
<font color='#0000FF'>#include</font> "<a style='text-decoration:none' href='../rand.h.html'>../rand.h</a>" |
|
<font color='#0000FF'>#include</font> <font color='#5555FF'><</font>memory<font color='#5555FF'>></font> |
|
<font color='#0000FF'>#include</font> "<a style='text-decoration:none' href='../geometry/rectangle.h.html'>../geometry/rectangle.h</a>" |
|
<font color='#0000FF'>#include</font> "<a style='text-decoration:none' href='../test_for_odr_violations.h.html'>../test_for_odr_violations.h</a>" |
|
|
|
<font color='#0000FF'>namespace</font> dlib |
|
<b>{</b> |
|
<font color='#0000FF'><u>bool</u></font> <b><a name='dnn_prefer_fastest_algorithms'></a>dnn_prefer_fastest_algorithms</b><font face='Lucida Console'>(</font><font face='Lucida Console'>)</font>; |
|
<font color='#0000FF'><u>void</u></font> <b><a name='set_dnn_prefer_fastest_algorithms'></a>set_dnn_prefer_fastest_algorithms</b><font face='Lucida Console'>(</font><font face='Lucida Console'>)</font>; |
|
<font color='#0000FF'><u>void</u></font> <b><a name='set_dnn_prefer_smallest_algorithms'></a>set_dnn_prefer_smallest_algorithms</b><font face='Lucida Console'>(</font><font face='Lucida Console'>)</font>; |
|
<b>}</b> |
|
|
|
<font color='#0000FF'>namespace</font> dlib <b>{</b> <font color='#0000FF'>namespace</font> tt |
|
<b>{</b> |
|
|
|
<font color='#009900'>// ---------------------------------------------------------------------------------------- |
|
</font> |
|
<font color='#0000FF'><u>void</u></font> <b><a name='inverse_norms'></a>inverse_norms</b> <font face='Lucida Console'>(</font> |
|
resizable_tensor<font color='#5555FF'>&</font> invnorms, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> data, |
|
<font color='#0000FF'>const</font> <font color='#0000FF'><u>double</u></font> eps |
|
<font face='Lucida Console'>)</font>; |
|
<font color='#009900'>/*! |
|
ensures |
|
- #invnorms == reciprocal(sqrt(sum_cols(squared(mat(data))) + eps)) |
|
!*/</font> |
|
|
|
<font color='#0000FF'><u>void</u></font> <b><a name='dot_prods'></a>dot_prods</b> <font face='Lucida Console'>(</font> |
|
resizable_tensor<font color='#5555FF'>&</font> out, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> lhs, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> rhs |
|
<font face='Lucida Console'>)</font>; |
|
<font color='#009900'>/*! |
|
requires |
|
- have_same_dimensions(lhs,rhs) == true |
|
ensures |
|
- #out.num_samples() == lhs.num_samples() |
|
- #out.k() == #out.nr() == #out.nc() == 1 |
|
- #out == sum_cols(pointwise_multiply(mat(lhs), mat(rhs))); |
|
!*/</font> |
|
|
|
<font color='#0000FF'><u>void</u></font> <b><a name='dot_prods'></a>dot_prods</b> <font face='Lucida Console'>(</font> |
|
<font color='#0000FF'><u>bool</u></font> add_to, |
|
tensor<font color='#5555FF'>&</font> out, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> lhs, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> rhs |
|
<font face='Lucida Console'>)</font>; |
|
<font color='#009900'>/*! |
|
requires |
|
- have_same_dimensions(lhs,rhs) == true |
|
- out.size() == lhs.num_samples() |
|
- out.k() == out.nr() == out.nc() == 1 |
|
ensures |
|
- if (add_to) then |
|
- #out == mat(out) + sum_cols(pointwise_multiply(mat(lhs), mat(rhs))); |
|
- else |
|
- #out == sum_cols(pointwise_multiply(mat(lhs), mat(rhs))); |
|
!*/</font> |
|
|
|
<font color='#0000FF'><u>void</u></font> <b><a name='scale_columns'></a>scale_columns</b> <font face='Lucida Console'>(</font> |
|
tensor<font color='#5555FF'>&</font> out, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> m, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> v |
|
<font face='Lucida Console'>)</font>; |
|
<font color='#009900'>/*! |
|
requires |
|
- have_same_dimensions(out,m) == true |
|
- is_vector(v) == true |
|
- v.size() == mat(m).nc() |
|
ensures |
|
- performs: out = scale_columns(mat(m),mat(v)); |
|
!*/</font> |
|
|
|
<font color='#0000FF'><u>void</u></font> <b><a name='scale_rows'></a>scale_rows</b> <font face='Lucida Console'>(</font> |
|
tensor<font color='#5555FF'>&</font> out, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> m, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> v |
|
<font face='Lucida Console'>)</font>; |
|
<font color='#009900'>/*! |
|
requires |
|
- have_same_dimensions(out,m) == true |
|
- is_vector(v) == true |
|
- v.size() == m.num_samples() |
|
ensures |
|
- performs: out = scale_rows(mat(m),mat(v)); |
|
!*/</font> |
|
|
|
<font color='#0000FF'><u>void</u></font> <b><a name='scale_rows2'></a>scale_rows2</b> <font face='Lucida Console'>(</font> |
|
<font color='#0000FF'><u>float</u></font> beta, |
|
tensor<font color='#5555FF'>&</font> out, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> m1, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> m2, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> v1, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> v2 |
|
<font face='Lucida Console'>)</font>; |
|
<font color='#009900'>/*! |
|
requires |
|
- have_same_dimensions(out,m1) == true |
|
- have_same_dimensions(out,m2) == true |
|
- have_same_dimensions(v1,v2) == true |
|
- is_vector(v1) == true |
|
- v1.size() == m1.num_samples() |
|
ensures |
|
- performs: |
|
out = beta*out + scale_rows(mat(m1) - scale_rows(mat(m2),mat(v1)), mat(v2)); |
|
!*/</font> |
|
|
|
<font color='#009900'>// ---------------------------------------------------------------------------------------- |
|
</font> |
|
<font color='#0000FF'><u>void</u></font> <b><a name='exp'></a>exp</b> <font face='Lucida Console'>(</font> |
|
tensor<font color='#5555FF'>&</font> dest, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> src |
|
<font face='Lucida Console'>)</font>; |
|
<font color='#009900'>/*! |
|
requires |
|
- dest.size() == src.size() |
|
ensures |
|
- performs: dest = exp(mat(src)) |
|
!*/</font> |
|
|
|
<font color='#009900'>// ---------------------------------------------------------------------------------------- |
|
</font> |
|
<font color='#0000FF'><u>void</u></font> <b><a name='log'></a>log</b> <font face='Lucida Console'>(</font> |
|
tensor<font color='#5555FF'>&</font> dest, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> src |
|
<font face='Lucida Console'>)</font>; |
|
<font color='#009900'>/*! |
|
requires |
|
- dest.size() == src.size() |
|
ensures |
|
- performs: dest = log(mat(src)) |
|
!*/</font> |
|
|
|
<font color='#009900'>// ---------------------------------------------------------------------------------------- |
|
</font> |
|
<font color='#0000FF'><u>void</u></font> <b><a name='log10'></a>log10</b> <font face='Lucida Console'>(</font> |
|
tensor<font color='#5555FF'>&</font> dest, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> src |
|
<font face='Lucida Console'>)</font>; |
|
<font color='#009900'>/*! |
|
requires |
|
- dest.size() == src.size() |
|
ensures |
|
- performs: dest = log10(mat(src)) |
|
!*/</font> |
|
|
|
<font color='#009900'>// ---------------------------------------------------------------------------------------- |
|
</font> |
|
<font color='#0000FF'><u>void</u></font> <b><a name='gemm'></a>gemm</b> <font face='Lucida Console'>(</font> |
|
<font color='#0000FF'><u>float</u></font> beta, |
|
tensor<font color='#5555FF'>&</font> dest, |
|
<font color='#0000FF'><u>float</u></font> alpha, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> lhs, |
|
<font color='#0000FF'><u>bool</u></font> trans_lhs, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> rhs, |
|
<font color='#0000FF'><u>bool</u></font> trans_rhs |
|
<font face='Lucida Console'>)</font>; |
|
<font color='#009900'>/*! |
|
requires |
|
- dest does not alias the memory of lhs or rhs |
|
- The dimensions of lhs and rhs must be compatible for matrix multiplication. |
|
In particular: |
|
- Let L == trans_lhs ? trans(mat(lhs)) : mat(lhs) |
|
- Let R == trans_rhs ? trans(mat(rhs)) : mat(rhs) |
|
- Let D == mat(dest) |
|
- D.nr() == L.nr() && D.nc() == R.nc() |
|
(i.e. dest must be preallocated and have the correct output dimensions) |
|
- L.nc() == R.nr() |
|
ensures |
|
- performs: dest = alpha*L*R + beta*mat(dest) |
|
!*/</font> |
|
|
|
<font color='#009900'>// ---------------------------------------------------------------------------------------- |
|
</font> |
|
<font color='#0000FF'>class</font> <b><a name='inv'></a>inv</b> |
|
<b>{</b> |
|
<font color='#009900'>/*! |
|
WHAT THIS OBJECT REPRESENTS |
|
This is a functor for doing matrix inversion on the GPU. The only |
|
reason it's an object is to avoid the reallocation of some GPU memory |
|
blocks if you want to do a bunch of matrix inversions in a row. |
|
!*/</font> |
|
<font color='#0000FF'>public</font>: |
|
|
|
<font color='#0000FF'><u>void</u></font> <b><a name='operator'></a>operator</b><font face='Lucida Console'>(</font><font face='Lucida Console'>)</font> <font face='Lucida Console'>(</font> |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> m, |
|
resizable_tensor<font color='#5555FF'>&</font> out |
|
<font face='Lucida Console'>)</font>; |
|
<font color='#009900'>/*! |
|
requires |
|
- m.size() == m.num_samples()*m.num_samples() |
|
(i.e. mat(m) must be a square matrix) |
|
ensures |
|
- out == inv(mat(m)); |
|
!*/</font> |
|
|
|
<font color='#0000FF'>private</font>: |
|
<font color='#0000FF'>#ifdef</font> DLIB_USE_CUDA |
|
cuda::inv finv; |
|
<font color='#0000FF'>#endif</font> |
|
<b>}</b>; |
|
|
|
<font color='#009900'>// ---------------------------------------------------------------------------------------- |
|
</font> |
|
<font color='#0000FF'>class</font> <b><a name='tensor_rand'></a>tensor_rand</b> |
|
<b>{</b> |
|
<font color='#009900'>/*! |
|
WHAT THIS OBJECT REPRESENTS |
|
This is a tool for filling a tensor with random numbers. |
|
|
|
Note that the sequence of random numbers output by this object is different |
|
when dlib is compiled with DLIB_USE_CUDA. So you should not write code |
|
that depends on any specific sequence of numbers coming out of a |
|
tensor_rand. |
|
|
|
!*/</font> |
|
|
|
<font color='#0000FF'>public</font>: |
|
<font color='#009900'>// not copyable |
|
</font> <b><a name='tensor_rand'></a>tensor_rand</b><font face='Lucida Console'>(</font><font color='#0000FF'>const</font> tensor_rand<font color='#5555FF'>&</font><font face='Lucida Console'>)</font> <font color='#5555FF'>=</font> <font color='#0000FF'>delete</font>; |
|
tensor_rand<font color='#5555FF'>&</font> <b><a name='operator'></a>operator</b><font color='#5555FF'>=</font><font face='Lucida Console'>(</font><font color='#0000FF'>const</font> tensor_rand<font color='#5555FF'>&</font><font face='Lucida Console'>)</font> <font color='#5555FF'>=</font> <font color='#0000FF'>delete</font>; |
|
|
|
<b><a name='tensor_rand'></a>tensor_rand</b><font face='Lucida Console'>(</font><font face='Lucida Console'>)</font> : tensor_rand<font face='Lucida Console'>(</font><font color='#979000'>0</font><font face='Lucida Console'>)</font> <b>{</b><b>}</b> |
|
<b><a name='tensor_rand'></a>tensor_rand</b><font face='Lucida Console'>(</font><font color='#0000FF'><u>unsigned</u></font> <font color='#0000FF'><u>long</u></font> <font color='#0000FF'><u>long</u></font> seed<font face='Lucida Console'>)</font>; |
|
|
|
<font color='#0000FF'><u>void</u></font> <b><a name='fill_gaussian'></a>fill_gaussian</b> <font face='Lucida Console'>(</font> |
|
tensor<font color='#5555FF'>&</font> data, |
|
<font color='#0000FF'><u>float</u></font> mean <font color='#5555FF'>=</font> <font color='#979000'>0</font>, |
|
<font color='#0000FF'><u>float</u></font> stddev <font color='#5555FF'>=</font> <font color='#979000'>1</font> |
|
<font face='Lucida Console'>)</font>; |
|
<font color='#009900'>/*! |
|
requires |
|
- data.size()%2 == 0 |
|
ensures |
|
- Fills data with random numbers drawn from a Gaussian distribution |
|
with the given mean and standard deviation. |
|
!*/</font> |
|
|
|
<font color='#0000FF'><u>void</u></font> <b><a name='fill_uniform'></a>fill_uniform</b> <font face='Lucida Console'>(</font> |
|
tensor<font color='#5555FF'>&</font> data |
|
<font face='Lucida Console'>)</font>; |
|
<font color='#009900'>/*! |
|
ensures |
|
- Fills data with uniform random numbers in the range (0.0, 1.0]. |
|
!*/</font> |
|
|
|
<font color='#0000FF'>#ifdef</font> DLIB_USE_CUDA |
|
cuda::curand_generator rnd; |
|
<font color='#0000FF'>#else</font> |
|
dlib::rand rnd; |
|
<font color='#0000FF'>#endif</font> |
|
<b>}</b>; |
|
|
|
<font color='#009900'>// ---------------------------------------------------------------------------------------- |
|
</font> |
|
<font color='#0000FF'><u>void</u></font> <b><a name='multiply'></a>multiply</b> <font face='Lucida Console'>(</font> |
|
<font color='#0000FF'><u>bool</u></font> add_to, |
|
tensor<font color='#5555FF'>&</font> dest, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> src1, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> src2 |
|
<font face='Lucida Console'>)</font>; |
|
<font color='#009900'>/*! |
|
requires |
|
- dest.k() == src1.k() == src2.k() |
|
- dest.nr() == src1.nr() == src2.nr() |
|
- dest.nc() == src1.nc() == src2.nc() |
|
- dest.num_samples(), src1.num_samples(), and src2.num_samples() must each |
|
either be 1 or whichever ones aren't equal to 1 must have the same values. |
|
ensures |
|
- let MD = max(dest.num_samples(), src1.num_samples(), src2.num_samples) |
|
- This function pointwise multiplies src1 with src2 and stores the result into |
|
#dest. However, how the multiplication happens depends on the dimensions of |
|
the tensors. First, when src1 and src2 are multiplied together, if either |
|
has a num_samples() dimension that is != MD, then it is first replicated to |
|
produce a tensor with num_samples()==MD dimensions and then they are |
|
pointwise multiplied together. |
|
|
|
Second, if dest.num_samples()==1, then after the pointwise multiplication of |
|
src1 with src2, the result has its samples summed to produce an output tensor |
|
with num_samples()==1 which is then assigned to #dest. |
|
- if (add_to) then |
|
- Instead of assigning the result to dest, this function adds the result to dest. |
|
!*/</font> |
|
|
|
<font color='#0000FF'><u>void</u></font> <b><a name='scale_channels'></a>scale_channels</b> <font face='Lucida Console'>(</font> |
|
<font color='#0000FF'><u>bool</u></font> add_to, |
|
tensor<font color='#5555FF'>&</font> dest, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> src, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> scales |
|
<font face='Lucida Console'>)</font>; |
|
<font color='#009900'>/*! |
|
requires |
|
- have_same_dimensions(dest, src) == true |
|
- scales.num_samples() == src.num_samples() |
|
- scales.k() == src.k() |
|
- scales.nr() == 1 |
|
- scales.nc() == 1 |
|
ensures |
|
- Scales each channel of src by the corresponding value in scales. To be |
|
precise, we will have: |
|
- #dest(n,k,r,c) == src(n,k,r,c)*scales(n,k,1,1) |
|
- if (add_to) then |
|
- Instead of assigning the result to dest, this function adds the result to dest. |
|
!*/</font> |
|
|
|
<font color='#0000FF'><u>void</u></font> <b><a name='multiply_conv'></a>multiply_conv</b> <font face='Lucida Console'>(</font> |
|
<font color='#0000FF'><u>bool</u></font> add_to, |
|
tensor<font color='#5555FF'>&</font> dest, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> src1, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> src2 |
|
<font face='Lucida Console'>)</font>; |
|
<font color='#009900'>/*! |
|
requires |
|
- if (have_same_dimensions(dest, src1) == true) then |
|
- src2.num_samples() == 1 |
|
- src2.nr() == 1 |
|
- src2.nc() == 1 |
|
- src2.k() == src1.k() |
|
- else |
|
- have_same_dimensions(src1, src2) == true) |
|
- dest.num_samples() == 1 |
|
- dest.nr() == 1 |
|
- dest.nc() == 1 |
|
- dest.k() == src1.k() |
|
ensures |
|
- Performs #dest == src1*src2 |
|
In particular, if the elements of dest, src1, and src2 were indexed by (n,k,r,c) then |
|
we would have: |
|
- if (have_same_dimensions(dest,src1)) then |
|
#dest(n,k,r,c) == src1(n,k,r,c)*src2(k) |
|
- else |
|
#dest(k) == sum over {n,r,c} of src1(n,k,r,c)*src2(n,k,r,c) |
|
- if (add_to) then |
|
- Instead of assigning the result to dest, this function adds the result to dest. |
|
!*/</font> |
|
|
|
<font color='#0000FF'><u>void</u></font> <b><a name='multiply_zero_padded'></a>multiply_zero_padded</b> <font face='Lucida Console'>(</font> |
|
<font color='#0000FF'><u>bool</u></font> add_to, |
|
tensor<font color='#5555FF'>&</font> dest, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> src1, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> src2 |
|
<font face='Lucida Console'>)</font>; |
|
<font color='#009900'>/*! |
|
ensures |
|
- if (add_to) then |
|
- performs: dest += src1 * src2 |
|
- else |
|
- performs: dest = src1 * src2 |
|
- In either case, the multiplication happens pointwise according to 4D tensor |
|
arithmetic. If the dimensions don't match then missing elements are presumed |
|
to be equal to 0. |
|
!*/</font> |
|
|
|
<font color='#009900'>// ---------------------------------------------------------------------------------------- |
|
</font> |
|
<font color='#0000FF'><u>void</u></font> <b><a name='affine_transform'></a>affine_transform</b><font face='Lucida Console'>(</font> |
|
tensor<font color='#5555FF'>&</font> dest, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> src, |
|
<font color='#0000FF'>const</font> <font color='#0000FF'><u>float</u></font> A, |
|
<font color='#0000FF'>const</font> <font color='#0000FF'><u>float</u></font> B |
|
<font face='Lucida Console'>)</font>; |
|
<font color='#009900'>/*! |
|
requires |
|
- dest.size()==src.size() |
|
ensures |
|
- #dest == A*src + B |
|
!*/</font> |
|
|
|
<font color='#0000FF'><u>void</u></font> <b><a name='affine_transform'></a>affine_transform</b><font face='Lucida Console'>(</font> |
|
tensor<font color='#5555FF'>&</font> dest, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> src, |
|
<font color='#0000FF'>const</font> <font color='#0000FF'><u>float</u></font> A |
|
<font face='Lucida Console'>)</font>; |
|
<font color='#009900'>/*! |
|
requires |
|
- dest.size()==src.size() |
|
ensures |
|
- #dest == A*src |
|
!*/</font> |
|
|
|
<font color='#0000FF'><u>void</u></font> <b><a name='affine_transform'></a>affine_transform</b><font face='Lucida Console'>(</font> |
|
tensor<font color='#5555FF'>&</font> dest, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> src1, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> src2, |
|
<font color='#0000FF'>const</font> <font color='#0000FF'><u>float</u></font> A, |
|
<font color='#0000FF'>const</font> <font color='#0000FF'><u>float</u></font> B, |
|
<font color='#0000FF'>const</font> <font color='#0000FF'><u>float</u></font> C |
|
<font face='Lucida Console'>)</font>; |
|
<font color='#009900'>/*! |
|
requires |
|
- dest.size()==src1.size() |
|
- dest.size()==src2.size() |
|
ensures |
|
- #dest == A*src1 + B*src2 + C |
|
!*/</font> |
|
|
|
<font color='#0000FF'><u>void</u></font> <b><a name='affine_transform'></a>affine_transform</b><font face='Lucida Console'>(</font> |
|
tensor<font color='#5555FF'>&</font> dest, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> src1, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> src2, |
|
<font color='#0000FF'>const</font> <font color='#0000FF'><u>float</u></font> A, |
|
<font color='#0000FF'>const</font> <font color='#0000FF'><u>float</u></font> B |
|
<font face='Lucida Console'>)</font>; |
|
<font color='#009900'>/*! |
|
requires |
|
- dest.size()==src1.size() |
|
- dest.size()==src2.size() |
|
ensures |
|
- #dest == A*src1 + B*src2 |
|
!*/</font> |
|
|
|
<font color='#0000FF'><u>void</u></font> <b><a name='affine_transform'></a>affine_transform</b><font face='Lucida Console'>(</font> |
|
tensor<font color='#5555FF'>&</font> dest, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> src1, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> src2, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> src3, |
|
<font color='#0000FF'>const</font> <font color='#0000FF'><u>float</u></font> A, |
|
<font color='#0000FF'>const</font> <font color='#0000FF'><u>float</u></font> B, |
|
<font color='#0000FF'>const</font> <font color='#0000FF'><u>float</u></font> C, |
|
<font color='#0000FF'>const</font> <font color='#0000FF'><u>float</u></font> D |
|
<font face='Lucida Console'>)</font>; |
|
<font color='#009900'>/*! |
|
requires |
|
- dest.size()==src1.size() |
|
- dest.size()==src2.size() |
|
- dest.size()==src3.size() |
|
ensures |
|
- #dest == A*src1 + B*src2 + C*src3 + D |
|
!*/</font> |
|
|
|
<font color='#0000FF'><u>void</u></font> <b><a name='affine_transform'></a>affine_transform</b><font face='Lucida Console'>(</font> |
|
tensor<font color='#5555FF'>&</font> dest, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> src1, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> src2, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> src3, |
|
<font color='#0000FF'>const</font> <font color='#0000FF'><u>float</u></font> A, |
|
<font color='#0000FF'>const</font> <font color='#0000FF'><u>float</u></font> B, |
|
<font color='#0000FF'>const</font> <font color='#0000FF'><u>float</u></font> C |
|
<font face='Lucida Console'>)</font>; |
|
<font color='#009900'>/*! |
|
requires |
|
- dest.size()==src1.size() |
|
- dest.size()==src2.size() |
|
- dest.size()==src3.size() |
|
ensures |
|
- #dest == A*src1 + B*src2 + C*src3 |
|
!*/</font> |
|
|
|
<font color='#0000FF'><u>void</u></font> <b><a name='affine_transform_range'></a>affine_transform_range</b><font face='Lucida Console'>(</font> |
|
<font color='#0000FF'><u>size_t</u></font> begin, |
|
<font color='#0000FF'><u>size_t</u></font> end, |
|
tensor<font color='#5555FF'>&</font> dest, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> src1, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> src2, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> src3, |
|
<font color='#0000FF'>const</font> <font color='#0000FF'><u>float</u></font> A, |
|
<font color='#0000FF'>const</font> <font color='#0000FF'><u>float</u></font> B, |
|
<font color='#0000FF'>const</font> <font color='#0000FF'><u>float</u></font> C |
|
<font face='Lucida Console'>)</font>; |
|
<font color='#009900'>/*! |
|
requires |
|
- dest.size()==src1.size() |
|
- dest.size()==src2.size() |
|
- dest.size()==src3.size() |
|
- begin <= end <= dest.size() |
|
ensures |
|
- This function operates much like |
|
affine_transform(dest,src1,src2,src3,A,B,C,0), except that it runs over only |
|
the half open range [begin,end) rather than processing the entire tensor. |
|
Specifically, it does this: |
|
- for i in the range [begin, end): |
|
- #dest.host()[i] == A*src1.host()[i] + B*src2.host()[i] + C*src3.host()[i] |
|
!*/</font> |
|
|
|
<font color='#0000FF'><u>void</u></font> <b><a name='affine_transform'></a>affine_transform</b><font face='Lucida Console'>(</font> |
|
<font color='#0000FF'>const</font> rectangle<font color='#5555FF'>&</font> rect, |
|
tensor<font color='#5555FF'>&</font> dest, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> src1, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> src2, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> src3, |
|
<font color='#0000FF'><u>float</u></font> A, |
|
<font color='#0000FF'><u>float</u></font> B, |
|
<font color='#0000FF'><u>float</u></font> C |
|
<font face='Lucida Console'>)</font>; |
|
<font color='#009900'>/*! |
|
requires |
|
- dest.size()==src1.size() |
|
- dest.size()==src2.size() |
|
- dest.size()==src3.size() |
|
- dest.num_samples()==src1.num_samples() |
|
- dest.num_samples()==src2.num_samples() |
|
- dest.num_samples()==src3.num_samples() |
|
- get_rect(mat(dest)).contains(rect) == true |
|
(i.e. rect must be entirely contained within dest) |
|
ensures |
|
- This function operates much like |
|
affine_transform(dest,src1,src2,src3,A,B,C,0), except that it runs over only |
|
the sub-rectangle indicated by rect. In particular, this function is equivalent |
|
to: |
|
set_subm(dest,rect) = A*subm(mat(src1),rect) + B*subm(mat(src2),rect) + C*subm(mat(src3),rect) |
|
!*/</font> |
|
|
|
<font color='#009900'>// ---------------------------------------------------------------------------------------- |
|
</font> |
|
<font color='#0000FF'><u>void</u></font> <b><a name='affine_transform'></a>affine_transform</b><font face='Lucida Console'>(</font> |
|
tensor<font color='#5555FF'>&</font> dest, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> src, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> A, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> B |
|
<font face='Lucida Console'>)</font>; |
|
<font color='#009900'>/*! |
|
requires |
|
- have_same_dimensions(dest,src) == true |
|
- if (A.num_samples() == 1) then |
|
- B.num_samples() == 1 |
|
- else |
|
- A.num_samples() == src.num_samples() |
|
- B.num_samples() == src.num_samples() |
|
- A.nr() == B.nr() == src.nr() |
|
- A.nc() == B.nc() == src.nc() |
|
- A.k() == B.k() == src.k() |
|
ensures |
|
- if (A.num_samples() == 1) then |
|
- #dest == A*src + B |
|
(done for each sample in src) |
|
- else |
|
- for all valid i: |
|
- #dest.host()[i] == A.host()[i]*src.host()[i] + B.host()[i] |
|
!*/</font> |
|
|
|
<font color='#009900'>// ---------------------------------------------------------------------------------------- |
|
</font> |
|
<font color='#0000FF'><u>void</u></font> <b><a name='affine_transform_conv'></a>affine_transform_conv</b><font face='Lucida Console'>(</font> |
|
tensor<font color='#5555FF'>&</font> dest, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> src, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> A, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> B |
|
<font face='Lucida Console'>)</font>; |
|
<font color='#009900'>/*! |
|
requires |
|
- have_same_dimensions(dest,src) == true |
|
- have_same_dimensions(A, B) == true |
|
- A.num_samples() == 1 |
|
- A.nr() == 1 |
|
- A.nc() == 1 |
|
- A.k() == src.k() |
|
ensures |
|
- Performs #dest == A*src + B |
|
In particular, if the elements of dest and src were indexed by (n,k,r,c) then |
|
we would have: |
|
#dest(n,k,r,c) == A(k)*src(n,k,r,c) + B(k). |
|
!*/</font> |
|
|
|
<font color='#009900'>// ---------------------------------------------------------------------------------------- |
|
</font> |
|
<font color='#0000FF'><u>void</u></font> <b><a name='compute_adam_update'></a>compute_adam_update</b> <font face='Lucida Console'>(</font> |
|
<font color='#0000FF'><u>size_t</u></font> begin, |
|
<font color='#0000FF'><u>size_t</u></font> end, |
|
tensor<font color='#5555FF'>&</font> s, |
|
tensor<font color='#5555FF'>&</font> m, |
|
tensor<font color='#5555FF'>&</font> v, |
|
<font color='#0000FF'>const</font> <font color='#0000FF'><u>float</u></font> t, |
|
<font color='#0000FF'>const</font> <font color='#0000FF'><u>float</u></font> learning_rate, |
|
<font color='#0000FF'>const</font> <font color='#0000FF'><u>float</u></font> weight_decay, |
|
<font color='#0000FF'>const</font> <font color='#0000FF'><u>float</u></font> momentum1, |
|
<font color='#0000FF'>const</font> <font color='#0000FF'><u>float</u></font> momentum2, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> params, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> params_grad |
|
<font face='Lucida Console'>)</font>; |
|
<font color='#009900'>/*! |
|
requires |
|
- s.size() == m.size() = v.size() == params.size() == params_grad.size() |
|
- t > 0 |
|
- learning_rate > 0 |
|
- weight_decay >= 0 |
|
- 0 <= momentum1 < 1 |
|
- 0 <= momentum2 < 1 |
|
- begin <= end <= params.size() |
|
ensures |
|
- This function implements the ADAM parameter update method described in the paper: |
|
Kingma, Diederik P., and Jimmy Ba Adam. "A method for stochastic |
|
optimization." International Conference on Learning Representation. 2015. |
|
Specifically, it implements the method shown as Algorithm 1. |
|
- #s is the update vector that should be added to the parameters. |
|
- The function only operates in the half open range [begin,end) of the memory |
|
blocks of each tensor. E.g. to make this function run on the entire tensor |
|
set begin to 0 and end to params.size(). |
|
!*/</font> |
|
|
|
<font color='#009900'>// ---------------------------------------------------------------------------------------- |
|
</font> |
|
<font color='#0000FF'><u>void</u></font> <b><a name='batch_normalize_inference'></a>batch_normalize_inference</b> <font face='Lucida Console'>(</font> |
|
<font color='#0000FF'>const</font> <font color='#0000FF'><u>double</u></font> eps, |
|
resizable_tensor<font color='#5555FF'>&</font> dest, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> src, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> gamma, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> beta, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> running_means, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> running_variances |
|
<font face='Lucida Console'>)</font>; |
|
<font color='#009900'>/*! |
|
requires |
|
- eps > 0 |
|
- gamma.num_samples() == 1 |
|
- gamma.nr() == src.nr() |
|
- gamma.nc() == src.nc() |
|
- gamma.k() == src.k() |
|
- have_same_dimensions(gamma, beta) |
|
- have_same_dimensions(gamma, running_means) |
|
- have_same_dimensions(gamma, running_variances) |
|
ensures |
|
- Linearly transforms src as a call to batch_normalize() would if src had means |
|
and variances as given by running_means and running_variances. That is, this |
|
function performs: |
|
dest = gamma*(src-running_means)/sqrt(running_variances+eps) + beta |
|
Note that it does it in a pointwise fashion over the samples in src. |
|
!*/</font> |
|
|
|
<font color='#0000FF'><u>void</u></font> <b><a name='batch_normalize'></a>batch_normalize</b> <font face='Lucida Console'>(</font> |
|
<font color='#0000FF'>const</font> <font color='#0000FF'><u>double</u></font> eps, |
|
resizable_tensor<font color='#5555FF'>&</font> dest, |
|
resizable_tensor<font color='#5555FF'>&</font> means, |
|
resizable_tensor<font color='#5555FF'>&</font> invstds, |
|
<font color='#0000FF'>const</font> <font color='#0000FF'><u>double</u></font> averaging_factor, |
|
resizable_tensor<font color='#5555FF'>&</font> running_means, |
|
resizable_tensor<font color='#5555FF'>&</font> running_variances, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> src, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> gamma, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> beta |
|
<font face='Lucida Console'>)</font>; |
|
<font color='#009900'>/*! |
|
requires |
|
- eps > 0 |
|
- src.num_samples() > 1 |
|
- gamma.num_samples() == 1 |
|
- beta.num_samples() == 1 |
|
- gamma.nr() == beta.nr() == src.nr() |
|
- gamma.nc() == beta.nc() == src.nc() |
|
- gamma.k() == beta.k() == src.k() |
|
- 0 <= averaging_factor <= 1 |
|
- if (averaging_factor != 1) |
|
- have_same_dimensions(running_means, means) == true |
|
- have_same_dimensions(running_variances, invstds) == true |
|
ensures |
|
- have_same_dimensions(#dest, src) == true |
|
- #means.num_samples() == 1 |
|
- #invstds.num_samples() == 1 |
|
- means.nr() == invstds.nr() == src.nr() |
|
- means.nc() == invstds.nc() == src.nc() |
|
- means.k() == invstds.k() == src.k() |
|
- #src == the batch normalized version of src. |
|
- #means == the mean values of the contents of src. |
|
- #invstds == 1/(the standard deviation values of the contents of src). |
|
- #running_means = (1-averaging_factor)*mat(#running_means) + averaging_factor*mat(#means); |
|
- #running_variances = (1-averaging_factor)*mat(#running_variances) + averaging_factor*(variance of contents of src); |
|
!*/</font> |
|
|
|
<font color='#0000FF'><u>void</u></font> <b><a name='batch_normalize_gradient'></a>batch_normalize_gradient</b> <font face='Lucida Console'>(</font> |
|
<font color='#0000FF'>const</font> <font color='#0000FF'><u>double</u></font> eps, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> gradient_input, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> means, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> invstds, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> src, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> gamma, |
|
tensor<font color='#5555FF'>&</font> src_grad, |
|
tensor<font color='#5555FF'>&</font> gamma_grad, |
|
tensor<font color='#5555FF'>&</font> beta_grad |
|
<font face='Lucida Console'>)</font>; |
|
<font color='#009900'>/*! |
|
requires |
|
- eps > 0 |
|
- invstds and means should be the output of a call to |
|
batch_normalize(eps,dest,means,invstds,src,gamma,beta) |
|
- have_same_dimensions(gradient_input, src) == true |
|
- have_same_dimensions(src, src_grad) == true |
|
- src.num_samples() > 1 |
|
- gamma.num_samples() == 1 |
|
- have_same_dimensions(gamma, gamma_grad) == true |
|
- have_same_dimensions(gamma, beta_grad) == true |
|
- gamma.nr() == src.nr() |
|
- gamma.nc() == src.nc() |
|
- gamma.k() == src.k() |
|
- have_same_dimensions(means, gamma) == true |
|
- have_same_dimensions(invstds, gamma) == true |
|
ensures |
|
- Let f(src,gamma,beta) == dot(gradient_input, dest output of |
|
batch_normalize(eps,dest,means,invstds,src,gamma,beta)) |
|
- Adds the gradient of f() with respect to src to #src_grad. |
|
- Assigns the gradient of f() with respect to gamma to #gamma_grad. |
|
- Assigns the gradient of f() with respect to beta to #beta_grad. |
|
!*/</font> |
|
|
|
<font color='#009900'>// ---------------------------------------------------------------------------------------- |
|
</font> |
|
<font color='#0000FF'><u>void</u></font> <b><a name='batch_normalize_conv_inference'></a>batch_normalize_conv_inference</b> <font face='Lucida Console'>(</font> |
|
<font color='#0000FF'>const</font> <font color='#0000FF'><u>double</u></font> eps, |
|
resizable_tensor<font color='#5555FF'>&</font> dest, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> src, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> gamma, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> beta, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> running_means, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> running_variances |
|
<font face='Lucida Console'>)</font>; |
|
<font color='#009900'>/*! |
|
requires |
|
- eps > 0 |
|
- gamma.num_samples() == 1 |
|
- gamma.nr() == 1 |
|
- gamma.nc() == 1 |
|
- gamma.k() == src.k() |
|
- have_same_dimensions(gamma, beta) |
|
- have_same_dimensions(gamma, running_means) |
|
- have_same_dimensions(gamma, running_variances) |
|
ensures |
|
- Linearly transforms src as a call to batch_normalize_conv() would if src had |
|
means and variances as given by running_means and running_variances. That |
|
is, this function performs: |
|
dest = gamma*(src-running_means)/sqrt(running_variances+eps) + beta |
|
Note that it does this in a pointwise fashion over the samples, rows, and |
|
columns in src. |
|
!*/</font> |
|
|
|
<font color='#0000FF'><u>void</u></font> <b><a name='batch_normalize_conv'></a>batch_normalize_conv</b> <font face='Lucida Console'>(</font> |
|
<font color='#0000FF'>const</font> <font color='#0000FF'><u>double</u></font> eps, |
|
resizable_tensor<font color='#5555FF'>&</font> dest, |
|
resizable_tensor<font color='#5555FF'>&</font> means, |
|
resizable_tensor<font color='#5555FF'>&</font> invstds, |
|
<font color='#0000FF'>const</font> <font color='#0000FF'><u>double</u></font> averaging_factor, |
|
resizable_tensor<font color='#5555FF'>&</font> running_means, |
|
resizable_tensor<font color='#5555FF'>&</font> running_variances, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> src, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> gamma, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> beta |
|
<font face='Lucida Console'>)</font>; |
|
<font color='#009900'>/*! |
|
requires |
|
- eps > 0 |
|
- src.num_samples() > 1 |
|
- gamma.num_samples()==gamma.nr()==gamma.nc() == 1 |
|
- beta.num_samples() ==beta.nr() ==gamma.nc() == 1 |
|
- gamma.k() == beta.k() == src.k() |
|
- 0 <= averaging_factor <= 1 |
|
- if (averaging_factor != 1) |
|
- have_same_dimensions(running_means, means) == true |
|
- have_same_dimensions(running_variances, invstds) == true |
|
ensures |
|
- have_same_dimensions(#dest, src) == true |
|
- #means.num_samples()==means.nr()==means.nc() == 1 |
|
- #invstds.num_samples() ==invstds.nr() ==invstds.nc() == 1 |
|
- means.k() == invstds.k() == src.k() |
|
- #src == the batch normalized version of src. |
|
- #means == the mean values of the contents of src. |
|
- #invstds == 1/(the standard deviation values of the contents of src). |
|
- #running_means = (1-averaging_factor)*mat(#running_means) + averaging_factor*mat(#means); |
|
- #running_variances = (1-averaging_factor)*mat(#running_variances) + averaging_factor*(variance of contents of src); |
|
!*/</font> |
|
|
|
<font color='#0000FF'><u>void</u></font> <b><a name='batch_normalize_conv_gradient'></a>batch_normalize_conv_gradient</b> <font face='Lucida Console'>(</font> |
|
<font color='#0000FF'>const</font> <font color='#0000FF'><u>double</u></font> eps, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> gradient_input, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> means, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> invstds, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> src, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> gamma, |
|
tensor<font color='#5555FF'>&</font> src_grad, |
|
tensor<font color='#5555FF'>&</font> gamma_grad, |
|
tensor<font color='#5555FF'>&</font> beta_grad |
|
<font face='Lucida Console'>)</font>; |
|
<font color='#009900'>/*! |
|
requires |
|
- eps > 0 |
|
- invstds and means should be the output of a call to |
|
batch_normalize_conv(eps,dest,means,invstds,src,gamma,beta) |
|
- have_same_dimensions(gradient_input, src) == true |
|
- have_same_dimensions(src, src_grad) == true |
|
- src.num_samples() > 1 |
|
- gamma.num_samples()==gamma.nr()==gamma.nc() == 1 |
|
- have_same_dimensions(gamma, gamma_grad) == true |
|
- have_same_dimensions(gamma, beta_grad) == true |
|
- gamma.k() == src.k() |
|
- have_same_dimensions(means, gamma) == true |
|
- have_same_dimensions(invstds, gamma) == true |
|
ensures |
|
- Let f(src,gamma,beta) == dot(gradient_input, dest output of |
|
batch_normalize_conv(eps,dest,means,invstds,src,gamma,beta)) |
|
- Adds the gradient of f() with respect to src to #src_grad. |
|
- Assigns the gradient of f() with respect to gamma to #gamma_grad. |
|
- Assigns the gradient of f() with respect to beta to #beta_grad. |
|
!*/</font> |
|
|
|
<font color='#009900'>// ----------------------------------------------------------------------------------- |
|
</font> |
|
<font color='#0000FF'><u>void</u></font> <b><a name='layer_normalize'></a>layer_normalize</b> <font face='Lucida Console'>(</font> |
|
<font color='#0000FF'>const</font> <font color='#0000FF'><u>double</u></font> eps, |
|
resizable_tensor<font color='#5555FF'>&</font> dest, |
|
resizable_tensor<font color='#5555FF'>&</font> means, |
|
resizable_tensor<font color='#5555FF'>&</font> invstds, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> src, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> gamma, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> beta |
|
<font face='Lucida Console'>)</font>; |
|
|
|
<font color='#0000FF'><u>void</u></font> <b><a name='layer_normalize_gradient'></a>layer_normalize_gradient</b> <font face='Lucida Console'>(</font> |
|
<font color='#0000FF'>const</font> <font color='#0000FF'><u>double</u></font> eps, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> gradient_input, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> means, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> invstds, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> src, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> gamma, |
|
tensor<font color='#5555FF'>&</font> src_grad, |
|
tensor<font color='#5555FF'>&</font> gamma_grad, |
|
tensor<font color='#5555FF'>&</font> beta_grad |
|
<font face='Lucida Console'>)</font>; |
|
|
|
<font color='#009900'>// ----------------------------------------------------------------------------------- |
|
</font> |
|
<font color='#0000FF'><u>void</u></font> <b><a name='threshold'></a>threshold</b> <font face='Lucida Console'>(</font> |
|
tensor<font color='#5555FF'>&</font> data, |
|
<font color='#0000FF'><u>float</u></font> thresh |
|
<font face='Lucida Console'>)</font>; |
|
<font color='#009900'>/*! |
|
ensures |
|
- Sets all elements of data to 1 or 0 depending on if they are above or below |
|
the given threshold. Specifically, for all valid i: |
|
- #data.host()[i] == data.host()[i]>thresh ? 1 : 0 |
|
!*/</font> |
|
|
|
<font color='#0000FF'><u>void</u></font> <b><a name='dot'></a>dot</b> <font face='Lucida Console'>(</font> |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> a, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> b, |
|
tensor<font color='#5555FF'>&</font> result, |
|
<font color='#0000FF'><u>size_t</u></font> idx |
|
<font face='Lucida Console'>)</font>; |
|
<font color='#009900'>/*! |
|
requires |
|
- a.size() == b.size() |
|
- idx < result.size() |
|
ensures |
|
- #result.host()[idx] == result.host()[idx] + dot(a,b); |
|
I.e. Adds the dot product between a and b into the idx-th element of result. |
|
The reason you might want to use this more complex version of dot() is |
|
because, when using CUDA, it runs by generating asynchronous kernel launches |
|
whereas the version of dot() that returns the result immediately as a scalar |
|
must block the host while we wait for the result to be computed and then |
|
transferred from the GPU do the host for return by dot(). So this version of |
|
dot() might be much faster in some cases. |
|
!*/</font> |
|
|
|
<font color='#009900'>// ---------------------------------------------------------------------------------------- |
|
</font> |
|
<font color='#0000FF'><u>void</u></font> <b><a name='add'></a>add</b><font face='Lucida Console'>(</font> |
|
<font color='#0000FF'><u>float</u></font> beta, |
|
tensor<font color='#5555FF'>&</font> dest, |
|
<font color='#0000FF'><u>float</u></font> alpha, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> src |
|
<font face='Lucida Console'>)</font>; |
|
<font color='#009900'>/*! |
|
requires |
|
- One of the following is true: |
|
- have_same_dimensions(src, dest) |
|
- src.num_samples()==1 && src.k()==dest.k() && src.nr()==1 && src.nc()==1 |
|
- src.num_samples()==1 && src.k()==dest.k() && src.nr()==dest.nr() && src.nc()==dest.nc() |
|
- src.num_samples()==1 && src.k()==1 && src.nr()==dest.nr() && src.nc()==dest.nc() |
|
- src.num_samples()==dest.num_samples() && src.k()==1 && src.nr()==1 && src.nc()==1 |
|
- is_same_object(src,dest) == false |
|
ensures |
|
- performs: dest = beta*dest + alpha*src |
|
However, how the addition happens depends on the dimensions of src. In |
|
particular, this function adds the scaled values of one src tensor to dest. |
|
Each dimension of the src tensor must match the corresponding dimension of |
|
the dest tensor or must be equal to 1. In the latter case, the same value |
|
from the src tensor, for those dimensions, will be used to add into the dest |
|
tensor. |
|
!*/</font> |
|
|
|
<font color='#009900'>// ---------------------------------------------------------------------------------------- |
|
</font> |
|
<font color='#0000FF'><u>void</u></font> <b><a name='add'></a>add</b> <font face='Lucida Console'>(</font> |
|
tensor<font color='#5555FF'>&</font> dest, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> src1, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> src2 |
|
<font face='Lucida Console'>)</font>; |
|
<font color='#009900'>/*! |
|
ensures |
|
- performs: dest = src1 + src2 |
|
The addition happens pointwise according to 4D tensor arithmetic. If the |
|
dimensions don't match then missing elements are presumed to be equal to 0. |
|
!*/</font> |
|
|
|
<font color='#009900'>// ---------------------------------------------------------------------------------------- |
|
</font> |
|
<font color='#0000FF'><u>void</u></font> <b><a name='assign_conv_bias_gradient'></a>assign_conv_bias_gradient</b> <font face='Lucida Console'>(</font> |
|
tensor<font color='#5555FF'>&</font> grad, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> gradient_input |
|
<font face='Lucida Console'>)</font>; |
|
<font color='#009900'>/*! |
|
requires |
|
- grad.num_samples() == 1 |
|
- grad.k() >= 1 |
|
- grad.nr() == 1 |
|
- grad.nc() == 1 |
|
- gradient_input.k() == grad.k() |
|
- gradient_input.size() > 0 |
|
- is_same_object(grad,gradient_input) == false |
|
ensures |
|
- let BIAS be a tensor with the same dimensions as grad. |
|
- let OUT be the output of add(1,OUT,1,BIAS) |
|
- let f(gradient_input,BIAS) == dot(gradient_input,OUT) |
|
- Then this function computes the gradient of f() with respect to BIAS and |
|
assigns it to grad. |
|
!*/</font> |
|
|
|
<font color='#009900'>// ---------------------------------------------------------------------------------------- |
|
</font> |
|
<font color='#0000FF'><u>void</u></font> <b><a name='assign_bias_gradient'></a>assign_bias_gradient</b> <font face='Lucida Console'>(</font> |
|
tensor<font color='#5555FF'>&</font> grad, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> gradient_input |
|
<font face='Lucida Console'>)</font>; |
|
<font color='#009900'>/*! |
|
requires |
|
- grad.num_samples() == 1 |
|
- gradient_input.k() == grad.k() |
|
- gradient_input.nr() == grad.nr() |
|
- gradient_input.nc() == grad.nc() |
|
- gradient_input.size() > 0 |
|
- is_same_object(grad,gradient_input) == false |
|
ensures |
|
- let BIAS be a tensor with the same dimensions as grad. |
|
- let OUT be the output of add(1,OUT,1,BIAS) |
|
- let f(gradient_input,BIAS) == dot(gradient_input,OUT) |
|
- Then this function computes the gradient of f() with respect to BIAS and |
|
assigns it to grad. |
|
!*/</font> |
|
|
|
<font color='#009900'>// ---------------------------------------------------------------------------------------- |
|
</font> |
|
<font color='#0000FF'>class</font> <b><a name='tensor_conv'></a>tensor_conv</b> |
|
<b>{</b> |
|
<font color='#0000FF'>public</font>: |
|
<b><a name='tensor_conv'></a>tensor_conv</b><font face='Lucida Console'>(</font><font color='#0000FF'>const</font> tensor_conv<font color='#5555FF'>&</font><font face='Lucida Console'>)</font> <font color='#5555FF'>=</font> <font color='#0000FF'>delete</font>; |
|
tensor_conv<font color='#5555FF'>&</font> <b><a name='operator'></a>operator</b><font color='#5555FF'>=</font><font face='Lucida Console'>(</font><font color='#0000FF'>const</font> tensor_conv<font color='#5555FF'>&</font><font face='Lucida Console'>)</font> <font color='#5555FF'>=</font> <font color='#0000FF'>delete</font>; |
|
|
|
<b><a name='tensor_conv'></a>tensor_conv</b><font face='Lucida Console'>(</font><font face='Lucida Console'>)</font> <b>{</b><b>}</b> |
|
|
|
<font color='#0000FF'><u>void</u></font> <b><a name='clear'></a>clear</b><font face='Lucida Console'>(</font> |
|
<font face='Lucida Console'>)</font> <b>{</b> impl.<font color='#BB00BB'>clear</font><font face='Lucida Console'>(</font><font face='Lucida Console'>)</font>; <b>}</b> |
|
|
|
<font color='#0000FF'><u>void</u></font> <b><a name='operator'></a>operator</b><font face='Lucida Console'>(</font><font face='Lucida Console'>)</font> <font face='Lucida Console'>(</font> |
|
<font color='#0000FF'>const</font> <font color='#0000FF'><u>bool</u></font> add_to_output, |
|
tensor<font color='#5555FF'>&</font> output, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> data, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> filters |
|
<font face='Lucida Console'>)</font> <b>{</b> <font color='#BB00BB'>impl</font><font face='Lucida Console'>(</font>add_to_output,output,data,filters<font face='Lucida Console'>)</font>; <b>}</b> |
|
<font color='#009900'>/*! |
|
requires |
|
- setup() has been called. Specifically, setup() has been called like this: |
|
this->setup(data, filters, stride_y, stride_x, padding_y, padding_x); |
|
- is_same_object(output,data) == false |
|
- is_same_object(output,filters) == false |
|
- filters.k() == data.k() |
|
- filters.nr() <= src.nr() + 2*padding_y |
|
- filters.nc() <= src.nc() + 2*padding_x |
|
- #output.num_samples() == data.num_samples() |
|
- #output.k() == filters.num_samples() |
|
- #output.nr() == 1+(data.nr() + 2*padding_y - filters.nr())/stride_y |
|
- #output.nc() == 1+(data.nc() + 2*padding_x - filters.nc())/stride_x |
|
ensures |
|
- Convolves filters over data. If add_to_output==true then we add the |
|
results to output, otherwise we assign to output, overwriting the |
|
previous values in output. |
|
- filters contains filters.num_samples() filters. |
|
!*/</font> |
|
|
|
<font color='#0000FF'><u>void</u></font> <b><a name='operator'></a>operator</b><font face='Lucida Console'>(</font><font face='Lucida Console'>)</font> <font face='Lucida Console'>(</font> |
|
<font color='#0000FF'>const</font> <font color='#0000FF'><u>bool</u></font> add_to_output, |
|
resizable_tensor<font color='#5555FF'>&</font> output, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> data, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> filters |
|
<font face='Lucida Console'>)</font> <b>{</b> <font color='#BB00BB'>impl</font><font face='Lucida Console'>(</font>add_to_output,output,data,filters<font face='Lucida Console'>)</font>; <b>}</b> |
|
<font color='#009900'>/*! |
|
requires |
|
- setup() has been called. Specifically, setup() has been called like this: |
|
this->setup(data, filters, stride_y, stride_x, padding_y, padding_x); |
|
- is_same_object(output,data) == false |
|
- is_same_object(output,filters) == false |
|
- filters.k() == data.k() |
|
- filters.nr() <= src.nr() + 2*padding_y |
|
- filters.nc() <= src.nc() + 2*padding_x |
|
ensures |
|
- Convolves filters over data. If add_to_output==true then we add the |
|
results to output, otherwise we assign to output, overwriting the |
|
previous values in output. |
|
- filters contains filters.num_samples() filters. |
|
- #output.num_samples() == data.num_samples() |
|
- #output.k() == filters.num_samples() |
|
- #output.nr() == 1+(data.nr() + 2*padding_y - filters.nr())/stride_y |
|
- #output.nc() == 1+(data.nc() + 2*padding_x - filters.nc())/stride_x |
|
!*/</font> |
|
|
|
<font color='#0000FF'><u>void</u></font> <b><a name='get_gradient_for_data'></a>get_gradient_for_data</b> <font face='Lucida Console'>(</font> |
|
<font color='#0000FF'>const</font> <font color='#0000FF'><u>bool</u></font> add_to_output, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> gradient_input, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> filters, |
|
tensor<font color='#5555FF'>&</font> data_gradient |
|
<font face='Lucida Console'>)</font> <b>{</b> impl.<font color='#BB00BB'>get_gradient_for_data</font><font face='Lucida Console'>(</font>add_to_output,gradient_input,filters,data_gradient<font face='Lucida Console'>)</font>; <b>}</b> |
|
<font color='#009900'>/*! |
|
requires |
|
- One of the following must be true: |
|
- filters has the same dimensions as the filters object given to the |
|
last call to operator(). Also, data_gradient has the same dimensions |
|
as the data object given to the last call to operator(). |
|
- setup() has been called. Specifically, setup() has been called like this: |
|
this->setup(data_gradient, filters, stride_y, stride_x, padding_y, padding_x); |
|
- gradient_input has the following dimensions: |
|
- gradient_input.num_samples() == data_gradient.num_samples() |
|
- gradient_input.k() == filters.num_samples() |
|
- gradient_input.nr() == 1+(data_gradient.nr() + 2*padding_y - filters.nr())/stride_y |
|
- gradient_input.nc() == 1+(data_gradient.nc() + 2*padding_x - filters.nc())/stride_x |
|
- NOTE, these dimensions are what you would obtain if gradient_input |
|
has the same dimensions as the last output of operator(). |
|
- is_same_object(data_gradient,filters) == false |
|
- is_same_object(data_gradient,gradient_input) == false |
|
ensures |
|
- let OUT be the output of (*this)(OUT,data,filters,sx,sy). |
|
- let f(data,filters) == dot(OUT, gradient_input) |
|
- if (add_to_output) then |
|
- This function finds the gradient of f() with respect to data and adds |
|
this gradient to data_gradient. |
|
- else |
|
- This function finds the gradient of f() with respect to data and |
|
assigns this gradient to data_gradient, overwriting the previous |
|
values in data_gradient. |
|
!*/</font> |
|
|
|
<font color='#0000FF'><u>void</u></font> <b><a name='get_gradient_for_filters'></a>get_gradient_for_filters</b> <font face='Lucida Console'>(</font> |
|
<font color='#0000FF'>const</font> <font color='#0000FF'><u>bool</u></font> add_to_output, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> gradient_input, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> data, |
|
tensor<font color='#5555FF'>&</font> filters_gradient |
|
<font face='Lucida Console'>)</font> <b>{</b> impl.<font color='#BB00BB'>get_gradient_for_filters</font><font face='Lucida Console'>(</font>add_to_output,gradient_input,data,filters_gradient<font face='Lucida Console'>)</font>; <b>}</b> |
|
<font color='#009900'>/*! |
|
requires |
|
- One of the following must be true: |
|
- filters_gradient has the same dimensions as the filters object given |
|
to the last call to operator(). Also, data has the same dimensions |
|
as the data object given to the last call to operator(). |
|
- setup() has been called. Specifically, setup() has been called like this: |
|
this->setup(data, filters_gradient, stride_y, stride_x, padding_y, padding_x); |
|
- gradient_input has the following dimensions: |
|
- gradient_input.num_samples() == data.num_samples() |
|
- gradient_input.k() == filters.num_samples() |
|
- gradient_input.nr() == 1+(data.nr() + 2*padding_y - filters.nr())/stride_y |
|
- gradient_input.nc() == 1+(data.nc() + 2*padding_x - filters.nc())/stride_x |
|
- NOTE, these dimensions are what you would obtain if gradient_input |
|
has the same dimensions as the last output of operator(). |
|
- is_same_object(filters_gradient,data) == false |
|
- is_same_object(filters_gradient,gradient_input) == false |
|
ensures |
|
- let OUT be the output of (*this)(OUT,data,filters,sx,sy). |
|
- let f(data,filters) == dot(OUT, gradient_input) |
|
- if (add_to_output) then |
|
- This function finds the gradient of f() with respect to filters and |
|
adds this gradient to filters_gradient. |
|
- else |
|
- This function finds the gradient of f() with respect to filters and |
|
assigns this gradient to filters_gradient, overwriting the previous |
|
values in filters_gradient. |
|
!*/</font> |
|
|
|
|
|
<font color='#0000FF'><u>void</u></font> <b><a name='setup'></a>setup</b><font face='Lucida Console'>(</font> |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> data, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> filters, |
|
<font color='#0000FF'><u>int</u></font> stride_y, |
|
<font color='#0000FF'><u>int</u></font> stride_x, |
|
<font color='#0000FF'><u>int</u></font> padding_y, |
|
<font color='#0000FF'><u>int</u></font> padding_x |
|
<font face='Lucida Console'>)</font> <b>{</b>impl.<font color='#BB00BB'>setup</font><font face='Lucida Console'>(</font>data,filters,stride_y,stride_x,padding_y,padding_x<font face='Lucida Console'>)</font>; <b>}</b> |
|
<font color='#009900'>/*! |
|
requires |
|
- filters.k() == data.k() |
|
- stride_y > 0 |
|
- stride_x > 0 |
|
- 0 <= padding_y < filters.nr() |
|
- 0 <= padding_x < filters.nc() |
|
ensures |
|
- When operator() is called, the output tensor will have these dimensions: |
|
- output.nr() == 1+(data.nr() + 2*padding_y - filters.nr())/stride_y |
|
- output.nc() == 1+(data.nc() + 2*padding_x - filters.nc())/stride_x |
|
- output.num_samples() == data.num_samples() |
|
- output.k() == filters.num_samples() |
|
- The point of setup() is to allow this object to gather information about |
|
all the tensor sizes and filter layouts involved in the computation. In |
|
particular, the reason the tensors are input into setup() is just to |
|
observe their sizes. setup() doesn't do anything with the contents of |
|
the tensors, or store any kind of references to the data or filter |
|
tensors. |
|
!*/</font> |
|
|
|
<font color='#0000FF'>private</font>: |
|
<font color='#0000FF'>#ifdef</font> DLIB_USE_CUDA |
|
cuda::tensor_conv impl; |
|
<font color='#0000FF'>#else</font> |
|
cpu::tensor_conv impl; |
|
<font color='#0000FF'>#endif</font> |
|
|
|
<b>}</b>; |
|
|
|
<font color='#009900'>// ---------------------------------------------------------------------------------------- |
|
</font> |
|
<font color='#0000FF'>class</font> <b><a name='pooling'></a>pooling</b> |
|
<b>{</b> |
|
<font color='#009900'>/*! |
|
WHAT THIS OBJECT REPRESENTS |
|
The pooling object is a tool for performing spatial pooling over a tensor. |
|
It can be configured to do either max or average pooling. |
|
!*/</font> |
|
<font color='#0000FF'>public</font>: |
|
|
|
<b><a name='pooling'></a>pooling</b><font face='Lucida Console'>(</font><font color='#0000FF'>const</font> pooling<font color='#5555FF'>&</font><font face='Lucida Console'>)</font> <font color='#5555FF'>=</font> <font color='#0000FF'>delete</font>; |
|
pooling<font color='#5555FF'>&</font> <b><a name='operator'></a>operator</b><font color='#5555FF'>=</font><font face='Lucida Console'>(</font><font color='#0000FF'>const</font> pooling<font color='#5555FF'>&</font><font face='Lucida Console'>)</font> <font color='#5555FF'>=</font> <font color='#0000FF'>delete</font>; |
|
|
|
<b><a name='pooling'></a>pooling</b> <font face='Lucida Console'>(</font> |
|
<font face='Lucida Console'>)</font> <font color='#5555FF'>=</font> <font color='#0000FF'>default</font>; |
|
|
|
<font color='#0000FF'><u>void</u></font> <b><a name='clear'></a>clear</b><font face='Lucida Console'>(</font> |
|
<font face='Lucida Console'>)</font> <b>{</b> impl.<font color='#BB00BB'>clear</font><font face='Lucida Console'>(</font><font face='Lucida Console'>)</font>; <b>}</b> |
|
|
|
<font color='#0000FF'><u>void</u></font> <b><a name='setup_max_pooling'></a>setup_max_pooling</b><font face='Lucida Console'>(</font> |
|
<font color='#0000FF'><u>int</u></font> window_height, |
|
<font color='#0000FF'><u>int</u></font> window_width, |
|
<font color='#0000FF'><u>int</u></font> stride_y, |
|
<font color='#0000FF'><u>int</u></font> stride_x, |
|
<font color='#0000FF'><u>int</u></font> padding_y, |
|
<font color='#0000FF'><u>int</u></font> padding_x |
|
<font face='Lucida Console'>)</font> <b>{</b> impl.<font color='#BB00BB'>setup_max_pooling</font><font face='Lucida Console'>(</font>window_height, window_width, stride_y, stride_x, padding_y, padding_x<font face='Lucida Console'>)</font>; <b>}</b> |
|
<font color='#009900'>/*! |
|
requires |
|
- window_height > 0 |
|
- window_width > 0 |
|
- stride_y > 0 |
|
- stride_x > 0 |
|
- 0 <= padding_y < window_height |
|
- 0 <= padding_x < window_width |
|
ensures |
|
- When you call operator() it will do max pooling with the given |
|
parameters. |
|
!*/</font> |
|
|
|
<font color='#0000FF'><u>void</u></font> <b><a name='setup_avg_pooling'></a>setup_avg_pooling</b><font face='Lucida Console'>(</font> |
|
<font color='#0000FF'><u>int</u></font> window_height, |
|
<font color='#0000FF'><u>int</u></font> window_width, |
|
<font color='#0000FF'><u>int</u></font> stride_y, |
|
<font color='#0000FF'><u>int</u></font> stride_x, |
|
<font color='#0000FF'><u>int</u></font> padding_y, |
|
<font color='#0000FF'><u>int</u></font> padding_x |
|
<font face='Lucida Console'>)</font> <b>{</b> impl.<font color='#BB00BB'>setup_avg_pooling</font><font face='Lucida Console'>(</font>window_height, window_width, stride_y, stride_x, padding_y, padding_x<font face='Lucida Console'>)</font>; <b>}</b> |
|
<font color='#009900'>/*! |
|
requires |
|
- window_height > 0 |
|
- window_width > 0 |
|
- stride_y > 0 |
|
- stride_x > 0 |
|
- 0 <= padding_y < window_height |
|
- 0 <= padding_x < window_width |
|
ensures |
|
- When you call operator() it will do average pooling with the given |
|
parameters. |
|
!*/</font> |
|
|
|
<font color='#0000FF'><u>bool</u></font> <b><a name='does_max_pooling'></a>does_max_pooling</b><font face='Lucida Console'>(</font> |
|
<font face='Lucida Console'>)</font> <font color='#0000FF'>const</font> <b>{</b> <font color='#0000FF'>return</font> impl.<font color='#BB00BB'>does_max_pooling</font><font face='Lucida Console'>(</font><font face='Lucida Console'>)</font>; <b>}</b> |
|
|
|
<font color='#0000FF'><u>void</u></font> <b><a name='operator'></a>operator</b><font face='Lucida Console'>(</font><font face='Lucida Console'>)</font> <font face='Lucida Console'>(</font> |
|
resizable_tensor<font color='#5555FF'>&</font> dest, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> src |
|
<font face='Lucida Console'>)</font> <b>{</b> <font color='#BB00BB'>impl</font><font face='Lucida Console'>(</font>dest, src<font face='Lucida Console'>)</font>; <b>}</b> |
|
<font color='#009900'>/*! |
|
requires |
|
- is_same_object(dest,src) == false |
|
- either setup_max_pooling() or setup_avg_pooling() has been called. |
|
- window_width <= src.nc() + 2*padding_x |
|
- window_height <= src.nr() + 2*padding_y |
|
ensures |
|
- #dest.num_samples() == src.num_samples() |
|
- #dest.k() == src.k() |
|
- #dest.nr() == 1 + (src.nr() + 2*padding_y - window_height)/stride_y |
|
- #dest.nc() == 1 + (src.nc() + 2*padding_x - window_width)/stride_x |
|
- WINDOW == centered_rect(x*stride_x + window_width/2 - padding_x, |
|
y*stride_y + window_height/2 - padding_y, |
|
window_width, |
|
window_height) |
|
- for all valid s, k, r, and c: |
|
- if (does_max_pooling()) then |
|
- image_plane(#dest,s,k)(r,c) == max(subm_clipped(image_plane(src,s,k),WINDOW(c,r))) |
|
- else |
|
- image_plane(#dest,s,k)(r,c) == mean(subm_clipped(image_plane(src,s,k),WINDOW(c,r))) |
|
!*/</font> |
|
|
|
<font color='#0000FF'><u>void</u></font> <b><a name='get_gradient'></a>get_gradient</b><font face='Lucida Console'>(</font> |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> gradient_input, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> dest, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> src, |
|
tensor<font color='#5555FF'>&</font> grad |
|
<font face='Lucida Console'>)</font> <b>{</b> impl.<font color='#BB00BB'>get_gradient</font><font face='Lucida Console'>(</font>gradient_input, dest, src, grad<font face='Lucida Console'>)</font>; <b>}</b> |
|
<font color='#009900'>/*! |
|
requires |
|
- have_same_dimensions(gradient_input,dest) == true |
|
- have_same_dimensions(src,grad) == true |
|
- dest contains the result of calling (*this)(dest,src) |
|
- is_same_object(grad,gradient_input) == false |
|
- is_same_object(grad,dest) == false |
|
- is_same_object(grad,src) == false |
|
ensures |
|
- Recalling that dest is the output of (*this)(dest,src), |
|
let f(src) == dot(gradient_input,dest) |
|
- Then this function computes the gradient of f() with respect to src and |
|
adds it to grad. |
|
!*/</font> |
|
|
|
<font color='#0000FF'>private</font>: |
|
<font color='#0000FF'>#ifdef</font> DLIB_USE_CUDA |
|
cuda::pooling impl; |
|
<font color='#0000FF'>#else</font> |
|
cpu::pooling impl; |
|
<font color='#0000FF'>#endif</font> |
|
<b>}</b>; |
|
|
|
<font color='#009900'>// ---------------------------------------------------------------------------------------- |
|
</font> |
|
<font color='#0000FF'><u>void</u></font> <b><a name='softmax'></a>softmax</b> <font face='Lucida Console'>(</font> |
|
tensor<font color='#5555FF'>&</font> dest, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> src |
|
<font face='Lucida Console'>)</font>; |
|
<font color='#009900'>/*! |
|
requires |
|
- have_same_dimensions(dest, src) == true |
|
ensures |
|
- Note that the softmax function is a vector valued function: |
|
s(x) == exp(x)/sum(exp(x)) |
|
- Computes the softmax function on src and writes the results to dest. The |
|
softmax is computed per spatial location across the different channels at |
|
each location. That is, softmax() outputs a new tensor, #dest, where each of |
|
the spatial locations in dest (i.e. image idx, row idx, and column idx) |
|
contains the output of s() evaluated over the channel values at each |
|
location. |
|
- This function supports in-place operation, i.e. having |
|
is_same_object(dest, src)==true |
|
!*/</font> |
|
|
|
<font color='#0000FF'><u>void</u></font> <b><a name='softmax_gradient'></a>softmax_gradient</b> <font face='Lucida Console'>(</font> |
|
tensor<font color='#5555FF'>&</font> grad, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> dest, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> gradient_input |
|
<font face='Lucida Console'>)</font>; |
|
<font color='#009900'>/*! |
|
requires |
|
- have_same_dimensions(dest,gradient_input) == true |
|
- have_same_dimensions(dest,grad) == true |
|
ensures |
|
- We interpret dest as the output of softmax(dest,SRC) for some SRC tensor. |
|
Then let f(SRC) == dot(gradient_input,dest). Then this function computes the |
|
gradient of f() with respect to SRC and stores it to grad. Moreover, if |
|
is_same_object(grad,gradient_input)==true then the output is assigned to |
|
grad, replacing its previous contents. Otherwise the output is added to |
|
grad. |
|
- This function supports in-place operation, i.e. having |
|
is_same_object(grad, gradient_input)==true |
|
!*/</font> |
|
|
|
<font color='#009900'>// ---------------------------------------------------------------------------------------- |
|
</font> |
|
<font color='#0000FF'><u>void</u></font> <b><a name='softmax_all'></a>softmax_all</b> <font face='Lucida Console'>(</font> |
|
tensor<font color='#5555FF'>&</font> dest, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> src |
|
<font face='Lucida Console'>)</font>; |
|
<font color='#009900'>/*! |
|
requires |
|
- have_same_dimensions(dest, src) == true |
|
ensures |
|
- Note that the softmax function is a vector valued function: |
|
s(x) == exp(x)/sum(exp(x)) |
|
- Computes the softmax function on src and writes the results to dest. The |
|
softmax is computed over the entire tensor with one invocation of s(). So |
|
unlike softmax() which computes many s() evaluations, one for each spatial |
|
location, softmax_all() calls s() once for the entire tensor. |
|
- This function supports in-place operation, i.e. having |
|
is_same_object(dest, src)==true |
|
!*/</font> |
|
|
|
<font color='#0000FF'><u>void</u></font> <b><a name='softmax_all_gradient'></a>softmax_all_gradient</b> <font face='Lucida Console'>(</font> |
|
tensor<font color='#5555FF'>&</font> grad, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> dest, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> gradient_input |
|
<font face='Lucida Console'>)</font>; |
|
<font color='#009900'>/*! |
|
requires |
|
- have_same_dimensions(dest,gradient_input) == true |
|
- have_same_dimensions(dest,grad) == true |
|
- is_same_object(grad, dest)==false |
|
ensures |
|
- We interpret dest as the output of softmax_all(dest,SRC) for some SRC tensor. |
|
Then let f(SRC) == dot(gradient_input,dest) Then this function computes the |
|
gradient of f() with respect to SRC and assigns it to grad. |
|
- This function supports in-place operation, i.e. having |
|
is_same_object(grad, gradient_input)==true |
|
!*/</font> |
|
|
|
<font color='#009900'>// ---------------------------------------------------------------------------------------- |
|
</font> |
|
<font color='#0000FF'><u>void</u></font> <b><a name='sigmoid'></a>sigmoid</b> <font face='Lucida Console'>(</font> |
|
tensor<font color='#5555FF'>&</font> dest, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> src |
|
<font face='Lucida Console'>)</font>; |
|
<font color='#009900'>/*! |
|
requires |
|
- have_same_dimensions(dest, src) == true |
|
ensures |
|
- for all valid i: |
|
- #dest.host()[i] == 1/(1+std::exp(-src.host()[i])) |
|
- This function supports in-place operation, i.e. having |
|
is_same_object(dest, src)==true |
|
!*/</font> |
|
|
|
<font color='#0000FF'><u>void</u></font> <b><a name='sigmoid_gradient'></a>sigmoid_gradient</b> <font face='Lucida Console'>(</font> |
|
tensor<font color='#5555FF'>&</font> grad, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> dest, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> gradient_input |
|
<font face='Lucida Console'>)</font>; |
|
<font color='#009900'>/*! |
|
requires |
|
- have_same_dimensions(dest,gradient_input) == true |
|
- have_same_dimensions(dest,grad) == true |
|
ensures |
|
- Recalling that dest is the output of sigmoid(dest,SRC) for some SRC tensor, |
|
let f(SRC) == dot(gradient_input,dest). Then this function computes the |
|
gradient of f() with respect to SRC and stores it to grad. Moreover, if |
|
is_same_object(grad,gradient_input)==true then the output is assigned to |
|
grad, replacing its previous contents. Otherwise the output is added to |
|
grad. |
|
- This function supports in-place operation, i.e. having |
|
is_same_object(grad, gradient_input)==true |
|
!*/</font> |
|
|
|
<font color='#009900'>// ---------------------------------------------------------------------------------------- |
|
</font> |
|
<font color='#0000FF'><u>void</u></font> <b><a name='mish'></a>mish</b> <font face='Lucida Console'>(</font> |
|
tensor<font color='#5555FF'>&</font> dest, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> src |
|
<font face='Lucida Console'>)</font>; |
|
<font color='#009900'>/*! |
|
requires |
|
- have_same_dimensions(dest, src) == true |
|
ensures |
|
- for all valid i: |
|
- #dest.host()[i] == src.host()[i]*std::tanh(std::log(1+std::exp(src.host()[i]))) |
|
- This function supports in-place operation, i.e. having |
|
is_same_object(dest, src)==true |
|
!*/</font> |
|
|
|
<font color='#0000FF'><u>void</u></font> <b><a name='mish_gradient'></a>mish_gradient</b> <font face='Lucida Console'>(</font> |
|
tensor<font color='#5555FF'>&</font> grad, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> dest, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> gradient_input |
|
<font face='Lucida Console'>)</font>; |
|
<font color='#009900'>/*! |
|
requires |
|
- have_same_dimensions(dest,gradient_input) == true |
|
- have_same_dimensions(dest,grad) == true |
|
ensures |
|
- This function computes the gradient of f() with respect to SRC and stores |
|
it to grad. Moreover, if is_same_object(grad,gradient_input)==true then |
|
the output is assigned to grad, replacing its previous contents. |
|
Otherwise the output is added to grad. |
|
- This function supports in-place operation, i.e. having |
|
is_same_object(grad, gradient_input)==true |
|
!*/</font> |
|
|
|
<font color='#009900'>// ---------------------------------------------------------------------------------------- |
|
</font> |
|
<font color='#0000FF'><u>void</u></font> <b><a name='relu'></a>relu</b> <font face='Lucida Console'>(</font> |
|
tensor<font color='#5555FF'>&</font> dest, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> src |
|
<font face='Lucida Console'>)</font>; |
|
<font color='#009900'>/*! |
|
requires |
|
- have_same_dimensions(dest, src) == true |
|
ensures |
|
- for all valid i: |
|
- #dest.host()[i] == std::max(0,src.host()[i]) |
|
- This function supports in-place operation, i.e. having |
|
is_same_object(dest, src)==true |
|
!*/</font> |
|
|
|
<font color='#0000FF'><u>void</u></font> <b><a name='relu_gradient'></a>relu_gradient</b> <font face='Lucida Console'>(</font> |
|
tensor<font color='#5555FF'>&</font> grad, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> dest, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> gradient_input |
|
<font face='Lucida Console'>)</font>; |
|
<font color='#009900'>/*! |
|
requires |
|
- have_same_dimensions(dest,gradient_input) == true |
|
- have_same_dimensions(dest,grad) == true |
|
ensures |
|
- Recalling that dest is the output of relu(dest,SRC) for some SRC tensor, |
|
let f(SRC) == dot(gradient_input,dest). Then this function computes the |
|
gradient of f() with respect to SRC and stores it to grad. Moreover, if |
|
is_same_object(grad,gradient_input)==true then the output is assigned to |
|
grad, replacing its previous contents. Otherwise the output is added to |
|
grad. |
|
- This function supports in-place operation, i.e. having |
|
is_same_object(grad, gradient_input)==true |
|
!*/</font> |
|
|
|
<font color='#009900'>// ---------------------------------------------------------------------------------------- |
|
</font> |
|
<font color='#0000FF'><u>void</u></font> <b><a name='prelu'></a>prelu</b> <font face='Lucida Console'>(</font> |
|
tensor<font color='#5555FF'>&</font> dest, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> src, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> param |
|
<font face='Lucida Console'>)</font>; |
|
<font color='#009900'>/*! |
|
requires |
|
- have_same_dimensions(dest, src) == true |
|
- param.size() == 1 |
|
ensures |
|
- for all valid i: |
|
- if (src.host()[i] > 0) then |
|
- #dest.host()[i] == src.host()[i] |
|
- else |
|
- #dest.host()[i] == src.host()[i] * param.host()[0] |
|
- This function supports in-place operation, i.e. having |
|
is_same_object(dest, src)==true |
|
!*/</font> |
|
|
|
<font color='#0000FF'><u>void</u></font> <b><a name='prelu_gradient'></a>prelu_gradient</b> <font face='Lucida Console'>(</font> |
|
tensor<font color='#5555FF'>&</font> grad, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> src, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> gradient_input, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> param, |
|
tensor<font color='#5555FF'>&</font> params_grad |
|
<font face='Lucida Console'>)</font>; |
|
<font color='#009900'>/*! |
|
requires |
|
- have_same_dimensions(grad,src) == true |
|
- have_same_dimensions(grad,gradient_input) == true |
|
- param.size() == 1 |
|
- params_grad.size() == 1 |
|
- is_same_object(grad, gradient_input) == false |
|
ensures |
|
- Recalling that dest is the output of prelu(dest,src,param) let |
|
f(src,param) == dot(gradient_input,dest) |
|
- Then this function computes the gradient of f() with respect to src and |
|
param. It assigns the gradient with respect to param to #params_grad and |
|
adds the gradient with respect to src to #grad. |
|
!*/</font> |
|
|
|
<font color='#009900'>// ---------------------------------------------------------------------------------------- |
|
</font> |
|
<font color='#0000FF'><u>void</u></font> <b><a name='leaky_relu'></a>leaky_relu</b> <font face='Lucida Console'>(</font> |
|
tensor<font color='#5555FF'>&</font> dest, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> src, |
|
<font color='#0000FF'>const</font> <font color='#0000FF'><u>float</u></font> alpha |
|
<font face='Lucida Console'>)</font>; |
|
<font color='#009900'>/*! |
|
requires |
|
- have_same_dimensions(dest, src) == true |
|
ensures |
|
- for all valid i: |
|
- if (src.host()[i] > 0) then |
|
- #dest.host()[i] == src.host()[i] |
|
- else |
|
- #dest.host()[i] == src.host()[i] * alpha |
|
!*/</font> |
|
|
|
<font color='#0000FF'><u>void</u></font> <b><a name='leaky_relu_gradient'></a>leaky_relu_gradient</b> <font face='Lucida Console'>(</font> |
|
tensor<font color='#5555FF'>&</font> grad, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> dest, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> gradient_input, |
|
<font color='#0000FF'>const</font> <font color='#0000FF'><u>float</u></font> alpha |
|
<font face='Lucida Console'>)</font>; |
|
<font color='#009900'>/*! |
|
requires |
|
- have_same_dimensions(dest,gradient_input) == true |
|
- have_same_dimensions(dest,grad) == true |
|
ensures |
|
- Recalling that dest is the output of leaky_relu(dest,SRC) for some SRC tensor, |
|
let f(SRC) == dot(gradient_input,dest). Then this function computes the |
|
gradient of f() with respect to SRC and stores it to grad. Moreover, if |
|
is_same_object(grad,gradient_input)==true then the output is assigned to |
|
grad, replacing its previous contents. Otherwise the output is added to |
|
grad. |
|
- This function supports in-place operation, i.e. having |
|
is_same_object(grad, gradient_input)==true |
|
!*/</font> |
|
|
|
<font color='#009900'>// ---------------------------------------------------------------------------------------- |
|
</font> |
|
<font color='#0000FF'><u>void</u></font> <b><a name='tanh'></a>tanh</b> <font face='Lucida Console'>(</font> |
|
tensor<font color='#5555FF'>&</font> dest, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> src |
|
<font face='Lucida Console'>)</font>; |
|
<font color='#009900'>/*! |
|
requires |
|
- have_same_dimensions(dest, src) == true |
|
ensures |
|
- for all valid i: |
|
- #dest.host()[i] == std::tanh(src.host()[i]) |
|
- This function supports in-place operation, i.e. having |
|
is_same_object(dest, src)==true |
|
!*/</font> |
|
|
|
<font color='#0000FF'><u>void</u></font> <b><a name='tanh_gradient'></a>tanh_gradient</b> <font face='Lucida Console'>(</font> |
|
tensor<font color='#5555FF'>&</font> grad, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> dest, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> gradient_input |
|
<font face='Lucida Console'>)</font>; |
|
<font color='#009900'>/*! |
|
requires |
|
- have_same_dimensions(dest,gradient_input) == true |
|
- have_same_dimensions(dest,grad) == true |
|
ensures |
|
- Recalling that dest is the output of tanh(dest,SRC) for some SRC tensor, |
|
let f(SRC) == dot(gradient_input,dest). Then this function computes the |
|
gradient of f() with respect to SRC and stores it to grad. Moreover, if |
|
is_same_object(grad,gradient_input)==true then the output is assigned to |
|
grad, replacing its previous contents. Otherwise the output is added to |
|
grad. |
|
- This function supports in-place operation, i.e. having |
|
is_same_object(grad, gradient_input)==true |
|
!*/</font> |
|
|
|
<font color='#009900'>// ---------------------------------------------------------------------------------------- |
|
</font> |
|
<font color='#0000FF'><u>void</u></font> <b><a name='gelu'></a>gelu</b> <font face='Lucida Console'>(</font> |
|
tensor<font color='#5555FF'>&</font> dest, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> src |
|
<font face='Lucida Console'>)</font>; |
|
<font color='#009900'>/*! |
|
requires |
|
- have_same_dimensions(dest, src) == true |
|
ensures |
|
- for all valid i: |
|
- #dest.host()[i] == src.host()[i]/2 * (1 + erf(src.host()[i]/sqrt(2)) |
|
- This function supports in-place operation, i.e. having |
|
is_same_object(dest, src)==true |
|
!*/</font> |
|
|
|
<font color='#0000FF'><u>void</u></font> <b><a name='gelu_gradient'></a>gelu_gradient</b> <font face='Lucida Console'>(</font> |
|
tensor<font color='#5555FF'>&</font> grad, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> dest, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> gradient_input |
|
<font face='Lucida Console'>)</font>; |
|
<font color='#009900'>/*! |
|
requires |
|
- have_same_dimensions(dest,gradient_input) == true |
|
- have_same_dimensions(dest,grad) == true |
|
ensures |
|
- This function computes the gradient of f() with respect to SRC and stores |
|
it to grad. Moreover, if is_same_object(grad,gradient_input)==true then |
|
the output is assigned to grad, replacing its previous contents. |
|
Otherwise the output is added to grad. |
|
- This function supports in-place operation, i.e. having |
|
is_same_object(grad, gradient_input)==true |
|
!*/</font> |
|
|
|
<font color='#009900'>// ---------------------------------------------------------------------------------------- |
|
</font> |
|
<font color='#0000FF'><u>void</u></font> <b><a name='resize_bilinear'></a>resize_bilinear</b> <font face='Lucida Console'>(</font> |
|
tensor<font color='#5555FF'>&</font> dest, |
|
<font color='#0000FF'><u>long</u></font> dest_row_stride, |
|
<font color='#0000FF'><u>long</u></font> dest_channel_stride, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> src, |
|
<font color='#0000FF'><u>long</u></font> src_row_stride, |
|
<font color='#0000FF'><u>long</u></font> src_channel_stride |
|
<font face='Lucida Console'>)</font>; |
|
<font color='#009900'>/*! |
|
requires |
|
- is_same_object(dest, src)==false |
|
- dest.num_samples() == src.num_samples() |
|
- dest.k() == src.k() |
|
ensures |
|
- for all valid i,k: image_plane(dest,i,k) is a copy of image_plane(src,i,k) |
|
that has been bilinearly interpolated to fit into the shape of |
|
image_plane(dest,i,k). |
|
- Instead of supposing the row stride and channel stride in the tensors is |
|
given by tensor::nc() and tensor::nr()*tensor::nc() respectively, we use the |
|
provided stride values to transition from one row and channel to the next. |
|
This is useful in combination with alias_tensor objects since it allows you |
|
to operate on subwindows in an image. |
|
!*/</font> |
|
|
|
<font color='#0000FF'><u>void</u></font> <b><a name='resize_bilinear_gradient'></a>resize_bilinear_gradient</b> <font face='Lucida Console'>(</font> |
|
tensor<font color='#5555FF'>&</font> grad, |
|
<font color='#0000FF'><u>long</u></font> grad_row_stride, |
|
<font color='#0000FF'><u>long</u></font> grad_channel_stride, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> gradient_input, |
|
<font color='#0000FF'><u>long</u></font> gradient_input_row_stride, |
|
<font color='#0000FF'><u>long</u></font> gradient_input_channel_stride |
|
<font face='Lucida Console'>)</font>; |
|
<font color='#009900'>/*! |
|
requires |
|
- is_same_object(grad, gradient_input)==false |
|
- gradient_input.num_samples() == grad.num_samples() |
|
- gradient_input.k() == grad.k() |
|
ensures |
|
- Suppose that DEST is the output of resize_bilinear(DEST,SRC) for some SRC |
|
tensor, let f(SRC) == dot(gradient_input,DEST). Then this function computes |
|
the gradient of f() with respect to SRC and adds it to grad. It should be |
|
noted that we don't need to know the contents of DEST to compute this |
|
gradient. All that matters is that gradient_input have the same dimensions |
|
as DEST. |
|
- Instead of supposing the row stride and channel stride in the tensors is |
|
given by tensor::nc() and tensor::nr()*tensor::nc() respectively, we use the |
|
provided stride values to transition from one row and channel to the next. |
|
This is useful in combination with alias_tensor objects since it allows you |
|
to operate on subwindows in an image. |
|
!*/</font> |
|
|
|
<font color='#0000FF'>inline</font> <font color='#0000FF'><u>void</u></font> <b><a name='resize_bilinear'></a>resize_bilinear</b> <font face='Lucida Console'>(</font> |
|
tensor<font color='#5555FF'>&</font> dest, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> src |
|
<font face='Lucida Console'>)</font> <b>{</b> <font color='#BB00BB'>resize_bilinear</font><font face='Lucida Console'>(</font>dest, dest.<font color='#BB00BB'>nc</font><font face='Lucida Console'>(</font><font face='Lucida Console'>)</font>, dest.<font color='#BB00BB'>nr</font><font face='Lucida Console'>(</font><font face='Lucida Console'>)</font><font color='#5555FF'>*</font>dest.<font color='#BB00BB'>nc</font><font face='Lucida Console'>(</font><font face='Lucida Console'>)</font>, src, src.<font color='#BB00BB'>nc</font><font face='Lucida Console'>(</font><font face='Lucida Console'>)</font>, src.<font color='#BB00BB'>nr</font><font face='Lucida Console'>(</font><font face='Lucida Console'>)</font><font color='#5555FF'>*</font>src.<font color='#BB00BB'>nc</font><font face='Lucida Console'>(</font><font face='Lucida Console'>)</font><font face='Lucida Console'>)</font>; <b>}</b> |
|
<font color='#009900'>/*! |
|
requires |
|
- is_same_object(dest, src)==false |
|
- dest.num_samples() == src.num_samples() |
|
- dest.k() == src.k() |
|
ensures |
|
- for all valid i,k: image_plane(dest,i,k) is a copy of image_plane(src,i,k) |
|
that has been bilinearly interpolated to fit into the shape of |
|
image_plane(dest,i,k). |
|
!*/</font> |
|
|
|
<font color='#0000FF'>inline</font> <font color='#0000FF'><u>void</u></font> <b><a name='resize_bilinear_gradient'></a>resize_bilinear_gradient</b> <font face='Lucida Console'>(</font> |
|
tensor<font color='#5555FF'>&</font> grad, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> gradient_input |
|
<font face='Lucida Console'>)</font> <b>{</b> <font color='#BB00BB'>resize_bilinear_gradient</font><font face='Lucida Console'>(</font>grad, grad.<font color='#BB00BB'>nc</font><font face='Lucida Console'>(</font><font face='Lucida Console'>)</font>, grad.<font color='#BB00BB'>nr</font><font face='Lucida Console'>(</font><font face='Lucida Console'>)</font><font color='#5555FF'>*</font>grad.<font color='#BB00BB'>nc</font><font face='Lucida Console'>(</font><font face='Lucida Console'>)</font>, gradient_input, gradient_input.<font color='#BB00BB'>nc</font><font face='Lucida Console'>(</font><font face='Lucida Console'>)</font>, gradient_input.<font color='#BB00BB'>nr</font><font face='Lucida Console'>(</font><font face='Lucida Console'>)</font><font color='#5555FF'>*</font>gradient_input.<font color='#BB00BB'>nc</font><font face='Lucida Console'>(</font><font face='Lucida Console'>)</font><font face='Lucida Console'>)</font>; <b>}</b> |
|
<font color='#009900'>/*! |
|
requires |
|
- is_same_object(grad, gradient_input)==false |
|
- gradient_input.num_samples() == grad.num_samples() |
|
- gradient_input.k() == grad.k() |
|
ensures |
|
- Suppose that DEST is the output of resize_bilinear(DEST,SRC) for some SRC |
|
tensor, let f(SRC) == dot(gradient_input,DEST). Then this function computes |
|
the gradient of f() with respect to SRC and adds it to grad. It should be |
|
noted that we don't need to know the contents of DEST to compute this |
|
gradient. All that matters is that gradient_input have the same dimensions |
|
as DEST. |
|
!*/</font> |
|
|
|
<font color='#009900'>// ---------------------------------------------------------------------------------------- |
|
</font> |
|
<font color='#0000FF'>class</font> <b><a name='multi_device_tensor_averager'></a>multi_device_tensor_averager</b> |
|
<b>{</b> |
|
<font color='#009900'>/*! |
|
WHAT THIS OBJECT REPRESENTS |
|
This object is a tool for very quickly averaging a bunch of tensors |
|
together. |
|
!*/</font> |
|
<font color='#0000FF'>public</font>: |
|
|
|
<b><a name='multi_device_tensor_averager'></a>multi_device_tensor_averager</b><font face='Lucida Console'>(</font><font color='#0000FF'>const</font> multi_device_tensor_averager<font color='#5555FF'>&</font><font face='Lucida Console'>)</font> <font color='#5555FF'>=</font> <font color='#0000FF'>delete</font>; |
|
multi_device_tensor_averager<font color='#5555FF'>&</font> <b><a name='operator'></a>operator</b><font color='#5555FF'>=</font><font face='Lucida Console'>(</font><font color='#0000FF'>const</font> multi_device_tensor_averager<font color='#5555FF'>&</font><font face='Lucida Console'>)</font> <font color='#5555FF'>=</font> <font color='#0000FF'>delete</font>; |
|
|
|
<b><a name='multi_device_tensor_averager'></a>multi_device_tensor_averager</b><font face='Lucida Console'>(</font><font face='Lucida Console'>)</font> <font color='#5555FF'>=</font> <font color='#0000FF'>default</font>; |
|
|
|
<font color='#0000FF'><u>void</u></font> <b><a name='set'></a>set</b><font face='Lucida Console'>(</font> |
|
std::vector<font color='#5555FF'><</font>tensor<font color='#5555FF'>*</font><font color='#5555FF'>></font> items |
|
<font face='Lucida Console'>)</font> |
|
<font color='#009900'>/*! |
|
requires |
|
- All the tensors in items are the same size |
|
ensures |
|
- When you call average() we will average the tensors in items. |
|
- It's important that the tensors already be allocated to their devices |
|
before you call set(). This is because set() will setup the types of |
|
between device transfers now and use them when you call average(). |
|
!*/</font> |
|
<b>{</b> |
|
<font color='#0000FF'>using</font> <font color='#0000FF'>namespace</font> ::dlib::cuda; |
|
accessible_groups.<font color='#BB00BB'>clear</font><font face='Lucida Console'>(</font><font face='Lucida Console'>)</font>; |
|
epa.<font color='#BB00BB'>clear</font><font face='Lucida Console'>(</font><font face='Lucida Console'>)</font>; |
|
<font color='#0000FF'>if</font> <font face='Lucida Console'>(</font>items.<font color='#BB00BB'>size</font><font face='Lucida Console'>(</font><font face='Lucida Console'>)</font> <font color='#5555FF'><</font> <font color='#979000'>1</font><font face='Lucida Console'>)</font> |
|
<font color='#0000FF'>return</font>; |
|
|
|
scale <font color='#5555FF'>=</font> <font color='#979000'>1.0</font><font color='#5555FF'>/</font>items.<font color='#BB00BB'>size</font><font face='Lucida Console'>(</font><font face='Lucida Console'>)</font>; |
|
|
|
<font color='#009900'>// split item into groups of accessible devices |
|
</font> std::vector<font color='#5555FF'><</font>tensor<font color='#5555FF'>*</font><font color='#5555FF'>></font> group, unused; |
|
<font color='#0000FF'>while</font><font face='Lucida Console'>(</font>items.<font color='#BB00BB'>size</font><font face='Lucida Console'>(</font><font face='Lucida Console'>)</font> <font color='#5555FF'>></font> <font color='#979000'>0</font><font face='Lucida Console'>)</font> |
|
<b>{</b> |
|
group.<font color='#BB00BB'>push_back</font><font face='Lucida Console'>(</font>items[<font color='#979000'>0</font>]<font face='Lucida Console'>)</font>; |
|
<font color='#0000FF'>for</font><font face='Lucida Console'>(</font><font color='#0000FF'><u>size_t</u></font> i <font color='#5555FF'>=</font> <font color='#979000'>1</font>; i <font color='#5555FF'><</font> items.<font color='#BB00BB'>size</font><font face='Lucida Console'>(</font><font face='Lucida Console'>)</font>; <font color='#5555FF'>+</font><font color='#5555FF'>+</font>i<font face='Lucida Console'>)</font> |
|
<b>{</b> |
|
<font color='#0000FF'>if</font> <font face='Lucida Console'>(</font><font color='#BB00BB'>can_access_peer</font><font face='Lucida Console'>(</font><font color='#5555FF'>*</font>items[<font color='#979000'>0</font>], <font color='#5555FF'>*</font>items[i]<font face='Lucida Console'>)</font><font face='Lucida Console'>)</font> |
|
group.<font color='#BB00BB'>push_back</font><font face='Lucida Console'>(</font>items[i]<font face='Lucida Console'>)</font>; |
|
<font color='#0000FF'>else</font> |
|
unused.<font color='#BB00BB'>push_back</font><font face='Lucida Console'>(</font>items[i]<font face='Lucida Console'>)</font>; |
|
<b>}</b> |
|
accessible_groups.<font color='#BB00BB'>push_back</font><font face='Lucida Console'>(</font>group<font face='Lucida Console'>)</font>; |
|
unused.<font color='#BB00BB'>swap</font><font face='Lucida Console'>(</font>items<font face='Lucida Console'>)</font>; |
|
unused.<font color='#BB00BB'>clear</font><font face='Lucida Console'>(</font><font face='Lucida Console'>)</font>; |
|
group.<font color='#BB00BB'>clear</font><font face='Lucida Console'>(</font><font face='Lucida Console'>)</font>; |
|
<b>}</b> |
|
<font color='#0000FF'>for</font> <font face='Lucida Console'>(</font><font color='#0000FF'>auto</font><font color='#5555FF'>&</font><font color='#5555FF'>&</font> g : accessible_groups<font face='Lucida Console'>)</font> |
|
<b>{</b> |
|
<font color='#0000FF'>for</font> <font face='Lucida Console'>(</font><font color='#0000FF'><u>size_t</u></font> i <font color='#5555FF'>=</font> <font color='#979000'>1</font>; i <font color='#5555FF'><</font> g.<font color='#BB00BB'>size</font><font face='Lucida Console'>(</font><font face='Lucida Console'>)</font>; <font color='#5555FF'>+</font><font color='#5555FF'>+</font>i<font face='Lucida Console'>)</font> |
|
<b>{</b> |
|
epa.<font color='#BB00BB'>emplace_back</font><font face='Lucida Console'>(</font><font color='#0000FF'>new</font> <font color='#BB00BB'>enable_peer_access</font><font face='Lucida Console'>(</font><font color='#5555FF'>*</font>g[<font color='#979000'>0</font>], <font color='#5555FF'>*</font>g[i]<font face='Lucida Console'>)</font><font face='Lucida Console'>)</font>; |
|
<b>}</b> |
|
<b>}</b> |
|
<b>}</b> |
|
|
|
<font color='#0000FF'><u>size_t</u></font> <b><a name='num_device_groups'></a>num_device_groups</b><font face='Lucida Console'>(</font> |
|
<font face='Lucida Console'>)</font> <font color='#0000FF'>const</font> <b>{</b> <font color='#0000FF'>return</font> accessible_groups.<font color='#BB00BB'>size</font><font face='Lucida Console'>(</font><font face='Lucida Console'>)</font>; <b>}</b> |
|
<font color='#009900'>/*! |
|
ensures |
|
- The devices given to set() are grouped together when they can directly |
|
access each other using GPUDirect. This function returns the number of |
|
such groups. For example, if all devices can directly access each other |
|
then the number of groups is 1. |
|
!*/</font> |
|
|
|
<font color='#0000FF'><u>void</u></font> <b><a name='average'></a>average</b><font face='Lucida Console'>(</font><font face='Lucida Console'>)</font> |
|
<font color='#009900'>/*! |
|
requires |
|
- All the devices have stopped writing to the tensors given to set(). So |
|
you should probably call cudaDeviceSynchronize() on each of the relevant |
|
devices before calling average(). |
|
ensures |
|
- Computes the average of all the tensors given to set() and then sets them |
|
all equal to the average. |
|
!*/</font> |
|
<b>{</b> |
|
<font color='#0000FF'>using</font> <font color='#0000FF'>namespace</font> ::dlib::cuda; |
|
|
|
|
|
<font color='#009900'>// First we average things within each group |
|
</font> <font color='#0000FF'>for</font> <font face='Lucida Console'>(</font><font color='#0000FF'>auto</font><font color='#5555FF'>&</font><font color='#5555FF'>&</font> g : accessible_groups<font face='Lucida Console'>)</font> |
|
<b>{</b> |
|
raii_set_device <font color='#BB00BB'>set_dev</font><font face='Lucida Console'>(</font><font color='#5555FF'>*</font>g[<font color='#979000'>0</font>]<font face='Lucida Console'>)</font>; |
|
<font color='#0000FF'>if</font> <font face='Lucida Console'>(</font>g.<font color='#BB00BB'>size</font><font face='Lucida Console'>(</font><font face='Lucida Console'>)</font> <font color='#5555FF'>=</font><font color='#5555FF'>=</font> <font color='#979000'>1</font><font face='Lucida Console'>)</font> |
|
tt::<font color='#BB00BB'>affine_transform</font><font face='Lucida Console'>(</font><font color='#5555FF'>*</font>g[<font color='#979000'>0</font>], <font color='#5555FF'>*</font>g[<font color='#979000'>0</font>], scale<font face='Lucida Console'>)</font>; |
|
<font color='#0000FF'>else</font> |
|
tt::<font color='#BB00BB'>affine_transform</font><font face='Lucida Console'>(</font><font color='#5555FF'>*</font>g[<font color='#979000'>0</font>], <font color='#5555FF'>*</font>g[<font color='#979000'>0</font>], <font color='#5555FF'>*</font>g[<font color='#979000'>1</font>], scale, scale<font face='Lucida Console'>)</font>; |
|
|
|
<font color='#0000FF'>for</font> <font face='Lucida Console'>(</font><font color='#0000FF'><u>size_t</u></font> i <font color='#5555FF'>=</font> <font color='#979000'>2</font>; i <font color='#5555FF'><</font> g.<font color='#BB00BB'>size</font><font face='Lucida Console'>(</font><font face='Lucida Console'>)</font>; <font color='#5555FF'>+</font><font color='#5555FF'>+</font>i<font face='Lucida Console'>)</font> |
|
tt::<font color='#BB00BB'>affine_transform</font><font face='Lucida Console'>(</font><font color='#5555FF'>*</font>g[<font color='#979000'>0</font>], <font color='#5555FF'>*</font>g[<font color='#979000'>0</font>], <font color='#5555FF'>*</font>g[i], <font color='#979000'>1</font>, scale<font face='Lucida Console'>)</font>; |
|
<b>}</b> |
|
|
|
<font color='#0000FF'>if</font> <font face='Lucida Console'>(</font>accessible_groups.<font color='#BB00BB'>size</font><font face='Lucida Console'>(</font><font face='Lucida Console'>)</font> <font color='#5555FF'>></font> <font color='#979000'>1</font><font face='Lucida Console'>)</font> |
|
<b>{</b> |
|
tensor<font color='#5555FF'>&</font> total_avg <font color='#5555FF'>=</font> <font color='#5555FF'>*</font>accessible_groups[<font color='#979000'>0</font>][<font color='#979000'>0</font>]; |
|
raii_set_device <font color='#BB00BB'>set_dev</font><font face='Lucida Console'>(</font>total_avg<font face='Lucida Console'>)</font>; |
|
accum_buffer.<font color='#BB00BB'>copy_size</font><font face='Lucida Console'>(</font>total_avg<font face='Lucida Console'>)</font>; |
|
<font color='#009900'>// now we need to average things across groups |
|
</font> <font color='#0000FF'>for</font> <font face='Lucida Console'>(</font><font color='#0000FF'><u>size_t</u></font> i <font color='#5555FF'>=</font> <font color='#979000'>1</font>; i <font color='#5555FF'><</font> accessible_groups.<font color='#BB00BB'>size</font><font face='Lucida Console'>(</font><font face='Lucida Console'>)</font>; <font color='#5555FF'>+</font><font color='#5555FF'>+</font>i<font face='Lucida Console'>)</font> |
|
<b>{</b> |
|
<font color='#BB00BB'>memcpy</font><font face='Lucida Console'>(</font>accum_buffer, <font color='#5555FF'>*</font>accessible_groups[i][<font color='#979000'>0</font>]<font face='Lucida Console'>)</font>; |
|
tt::<font color='#BB00BB'>add</font><font face='Lucida Console'>(</font>total_avg, total_avg, accum_buffer<font face='Lucida Console'>)</font>; |
|
<b>}</b> |
|
|
|
<font color='#009900'>// Now total_avg has the final average in it. So we need to send |
|
</font> <font color='#009900'>// copies of it back to each of the groups. |
|
</font> <font color='#0000FF'>for</font> <font face='Lucida Console'>(</font><font color='#0000FF'><u>size_t</u></font> i <font color='#5555FF'>=</font> <font color='#979000'>1</font>; i <font color='#5555FF'><</font> accessible_groups.<font color='#BB00BB'>size</font><font face='Lucida Console'>(</font><font face='Lucida Console'>)</font>; <font color='#5555FF'>+</font><font color='#5555FF'>+</font>i<font face='Lucida Console'>)</font> |
|
<b>{</b> |
|
<font color='#BB00BB'>memcpy</font><font face='Lucida Console'>(</font><font color='#5555FF'>*</font>accessible_groups[i][<font color='#979000'>0</font>], total_avg<font face='Lucida Console'>)</font>; |
|
<b>}</b> |
|
<b>}</b> |
|
|
|
|
|
<font color='#009900'>// Now propagate averages back out to each element using point to point |
|
</font> <font color='#009900'>// communication inside a group. |
|
</font> <font color='#0000FF'>for</font> <font face='Lucida Console'>(</font><font color='#0000FF'>auto</font><font color='#5555FF'>&</font><font color='#5555FF'>&</font> g : accessible_groups<font face='Lucida Console'>)</font> |
|
<b>{</b> |
|
raii_set_device <font color='#BB00BB'>set_dev</font><font face='Lucida Console'>(</font><font color='#5555FF'>*</font>g[<font color='#979000'>0</font>]<font face='Lucida Console'>)</font>; |
|
<font color='#0000FF'>for</font> <font face='Lucida Console'>(</font><font color='#0000FF'><u>size_t</u></font> i <font color='#5555FF'>=</font> <font color='#979000'>1</font>; i <font color='#5555FF'><</font> g.<font color='#BB00BB'>size</font><font face='Lucida Console'>(</font><font face='Lucida Console'>)</font>; <font color='#5555FF'>+</font><font color='#5555FF'>+</font>i<font face='Lucida Console'>)</font> |
|
<font color='#BB00BB'>memcpy</font><font face='Lucida Console'>(</font><font color='#5555FF'>*</font>g[i], <font color='#5555FF'>*</font>g[<font color='#979000'>0</font>]<font face='Lucida Console'>)</font>; |
|
<b>}</b> |
|
<b>}</b> |
|
|
|
<font color='#0000FF'>private</font>: |
|
std::vector<font color='#5555FF'><</font>std::unique_ptr<font color='#5555FF'><</font>::dlib::cuda::enable_peer_access<font color='#5555FF'>></font><font color='#5555FF'>></font> epa; |
|
std::vector<font color='#5555FF'><</font>std::vector<font color='#5555FF'><</font>tensor<font color='#5555FF'>*</font><font color='#5555FF'>></font><font color='#5555FF'>></font> accessible_groups; |
|
<font color='#0000FF'><u>float</u></font> scale; |
|
|
|
resizable_tensor accum_buffer; |
|
<b>}</b>; |
|
|
|
<font color='#009900'>// ---------------------------------------------------------------------------------------- |
|
</font> |
|
<font color='#0000FF'><u>void</u></font> <b><a name='copy_tensor'></a>copy_tensor</b><font face='Lucida Console'>(</font> |
|
<font color='#0000FF'><u>bool</u></font> add_to, |
|
tensor<font color='#5555FF'>&</font> dest, |
|
<font color='#0000FF'><u>size_t</u></font> dest_k_offset, |
|
<font color='#0000FF'>const</font> tensor<font color='#5555FF'>&</font> src, |
|
<font color='#0000FF'><u>size_t</u></font> src_k_offset, |
|
<font color='#0000FF'><u>size_t</u></font> count_k |
|
<font face='Lucida Console'>)</font>; |
|
<font color='#009900'>/*! |
|
requires |
|
- dest.nc() == src.nc() |
|
- dest.nr() == src.nr() |
|
- dest.num_samples() == src.num_samples() |
|
- dest.k() - dest_k_offset >= count_k |
|
- src.k() - src_k_offset >= count_k |
|
- is_same_object(dest,src) == false |
|
- The memory areas of src and dest do not overlap. |
|
ensures |
|
- if (add_to) then |
|
- performs: dest[i, k + dest_k_offset, r, c] += src[i, k + src_k_offset, r, c], where k in [0..count_k] |
|
i.e., adds content of each sample from src in to corresponding place of sample at dest. |
|
- else |
|
- performs: dest[i, k + dest_k_offset, r, c] = src[i, k + src_k_offset, r, c], where k in [0..count_k] |
|
i.e., copies content of each sample from src in to corresponding place of sample at dest. |
|
!*/</font> |
|
|
|
<font color='#009900'>// ---------------------------------------------------------------------------------------- |
|
</font> |
|
<b>}</b><b>}</b> |
|
|
|
<font color='#0000FF'>#ifdef</font> NO_MAKEFILE |
|
<font color='#0000FF'>#include</font> "<a style='text-decoration:none' href='tensor_tools.cpp.html'>tensor_tools.cpp</a>" |
|
<font color='#0000FF'>#endif</font> |
|
|
|
<font color='#0000FF'>#endif</font> <font color='#009900'>// DLIB_TeNSOR_TOOLS_H_ |
|
</font> |
|
|
|
|
|
</pre></body></html> |
