Conversation
src/cudamatrix/cu-kernels.cu
Outdated
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Outdated
src/cudamatrix/cu-kernels.cu
Outdated
src/cudamatrix/cu-matrix.h
Outdated
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Outdated
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Outdated
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.
There was a problem hiding this comment.
remove 2.5D dimension and just explain block!
remove 'for selecting an representative activation in an area'
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.