Maxpool #16
Maxpool #16
Conversation
src/cudamatrix/cu-kernels.cu
Outdated
| for (int32_cuda q = 0; q < num_col_blocks; q++) { | ||
| dst[index] = fmax( | ||
| src[index_src + p * src_stride * d.rows + q * d.cols], | ||
| dst[index]); |
There was a problem hiding this comment.
Don't you need to initialize dst[index]?
There was a problem hiding this comment.
but dst is one of the input.
I am not sure how should initialize it?
I just imitate the function "_add_mat_blocks" in this file(line 722).
| src[index_src + p * src_stride * d.cols + q * d.rows], | ||
| dst[index]); | ||
| } | ||
| } |
There was a problem hiding this comment.
Why do you need alpha in these functions?
You can modify functions by removing alpha.
There was a problem hiding this comment.
OK, I would delete all the alpha.
I was not sure if we need this parameter and since the default value is one, it seems has no impact.
src/cudamatrix/cu-kernels.cu
Outdated
| dst_index = i + j * dst_dim.stride, | ||
| src_index = src_i + src_j * src_dim.stride; | ||
| if (i < dst_dim.cols && j < dst_dim.rows) | ||
| dst[dst_index] += alpha * src[src_index]; |
src/cudamatrix/cu-kernels.cu
Outdated
| void cudaF_max_mat_blocks(dim3 Gr, dim3 Bl, float alpha, const float* src, | ||
| int32_cuda num_row_blocks, int32_cuda num_col_blocks, | ||
| float* dst, MatrixDim d, int src_stride, | ||
| int A_trans) { |
src/cudamatrix/cu-matrix.h
Outdated
| void AddMatBlocks(Real alpha, const CuMatrixBase<Real> &A, | ||
| MatrixTransposeType trans = kNoTrans); | ||
|
|
||
| void MaxMatBlocks(Real alpha, const CuMatrixBase<Real> &A, |
There was a problem hiding this comment.
Add comments containing function and parameter definition
…ion-component)(without cuda functions)
src/cudamatrix/cu-kernels.cu
Outdated
| // loop over all the elements in each pool to find the maximum one, | ||
| // and record its index. | ||
|
|
||
| for (int32_cuda t = 0; t < pool_t_size_; t++) { |
There was a problem hiding this comment.
I think your assumption in this function is that stride in all dimensions are 1.
There was a problem hiding this comment.
Do you mean I should set all the stride as 1?
the stride is represented by pool_{t,h,f}_step. I would add some description.
There was a problem hiding this comment.
pool_{t,h,f}_step is the stride between blocks, but as far as I found out the stride within block is 1.
There was a problem hiding this comment.
Do you mean it is possible that I need to get the maximum of part of the block rather than the whole block?
Like getting 5 from block [1,6,5,7,1] with within_block_stride as 2.
There was a problem hiding this comment.
I mean the rows of block can be something like [1 5 9]!
There was a problem hiding this comment.
I see what you mean, I would fix this! Thank you!
By the way, is this kind of stride only apply for rows? would the column also be like [1 5 9]?
src/cudamatrix/cu-matrix.h
Outdated
| /// function set all the values in &out_deriv whose index is not in | ||
| /// vector(not corresponding to maximum value in each pool of &in_value) | ||
| /// as zero, and keeps those correponding to maximum value as the *in_deriv. | ||
| void MaxMatBlocks(const CuMatrixBase<Real> &A, CuVectorBase<Real> &index_max_, |
There was a problem hiding this comment.
Please add description for all variables and explain how you select pooling block w.r.t parameters (You can add formula)
There was a problem hiding this comment.
I have added the description of variables, but these descriptions are almost the same as in the file nnet-convolution-component.h(line 400), I am not sure if we should write it again.
| MaxPoolingOverBlock component was firstly used in ConvNet for selecting an | ||
| representative activation in an area. It inspired Maxout nonlinearity. | ||
| Each output element of this component is the maximum of a block of | ||
| input elements where the block has a 2.5D dimension (pool_t_size_, |
There was a problem hiding this comment.
remove 2.5D dimension and just explain block!
remove 'for selecting an representative activation in an area'
|
|
||
| This component is designed to be used after a ConvolutionComponent | ||
| so that the input matrix is propagated from a 2d-convolutional layer. | ||
| This component implements 2.5d-maxpooling which performs |
There was a problem hiding this comment.
2.5d is meaningless! You can say it is 3D-maxpooling and explain how the 3d matrix is stored in 2 dim!
|
@qbetterk , please add test code for your function in cu-matrix.*/ |
|
@qbetterk Did you add test code for max function in matrix class? |
|
Hi Pegah,
I did write some test code，but I am not sure I should put it matrix.cc file or write a new cu-matrix-test.cc file like others (cu-math-test.cc).
And I am not in Maryland this week. So, sorry I cannot upload them so far. I would return on this Thursday and would upload the code as soon as Possible.
Best,
Kun
…Sent from my iPhone
On May 15, 2018, at 12:48, pegahgh ***@***.***> wrote:
@qbetterk Did you add test code for max function in matrix class?
—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub, or mute the thread.
|
|
Sure, we can discuss about it when you are back! |
2 participants
No description provided.