Conv1d and batchnorm1d layers #209
Conversation
|
|
Anirudh0707
force-pushed
the
master
branch
2 times, most recently
from
12d4184 to
b85cda2
Compare
c_reference/src/conv1d.c
Outdated
|
|
||
| const ConvLayers_LR_Params* tparams= (ConvLayers_LR_Params*) params; | ||
|
|
||
| if (padding == -1) { |
There was a problem hiding this comment.
the header files doesn't say much about what setting padding to -1 (or any other value) means
There was a problem hiding this comment.
I added some documentation for padding. I removed this -1 part altogether to avoid confusions
There was a problem hiding this comment.
what was the -1 for?
There was a problem hiding this comment.
I originally thought of keeping a case to calculate the padding automatically to maintain the time steps.
So if it was -1, the padding would be calculated such that input time steps would be equal to the output time steps. It was similar to the "same" convolution in tensorflow
But since it would be confusing for the user, I removed those lines
| @@ -0,0 +1,272 @@ | |||
| // Copyright (c) Microsoft Corporation. All rights reserved. | |||
| // Licensed under the MIT license. | |||
|
|
|||
There was a problem hiding this comment.
- We need some description of the overall architecture, and clarification over the form of input being tested and the nature of output.
- can this code be extended for streaming data easily?
- The file name could also be more specific as to which model arch for keyword spotting tested here.
There was a problem hiding this comment.
I've added the description and a note on what to cater to when porting the code to the microcontroller.
The code needs minor modifications for handling the streaming data. These changes are mostly variable assignments, for example fixing input streaming buffer it steps, feature length
I've changed the file to phoneme_det_cnn_rnn to signify the use of the combined arch
| memcpy(rnn_out + ((out_index++)*RNN_O_F), temp_hiddenstate, rnn_hidden*sizeof(float)); | ||
| } | ||
|
|
||
| // Backward Pass |
There was a problem hiding this comment.
forward and backward passes could be explained a bit in text here so the reader can follow the intent.
There was a problem hiding this comment.
I've moved the bricking to a separate file altogether. This bricking module has been added to the src folder, so that it can be used as function for other applications as well.
I've added a detailed documentation and line comments explaining how the code works. I've even included notes on how to use the bi-directional option and have highlighted the constrains for the bi-directional application
|
Also can we have more expressive macro names than |
I've incorporated all the modifications as instructed. I'll work on unifying the test benches and will rectify the macro expressiveness . I'll make these changes and push the commit for the same shortly |
I've changed the macros to a more expressive name. I have also combined the test-bench files as mentioned above |
Anirudh0707
changed the title
|
Sorry for the delay in the new commit. All the changes have been made. I have even added a couple of extra lines of comments for the error metrics for better clarity. |
No description provided.