05315a3567
Merge pull request #2 from augustin64/cuda-backpropagation
2023-03-31 14:22:46 +02:00
a3a803ba40
Fix previous compilation errors
2023-03-31 14:17:41 +02:00
7511856621
Translate backward convolution to CUDA
...
Not working yet, CUDA kernels in `backpropagation.cu` don't have access to activation functions declared in `function.cu` using `get_activation_function_cuda`.
Temporary workaround: copy `backpropagation.cu` parts that don't work to `function.cu` (all the parts using function pointers in kernels
2023-03-30 18:16:41 +02:00
5088c415d6
Add comments to config.h
2023-03-30 18:11:19 +02:00
dd16e34cce
extern "C" get_activation_function_cuda
2023-03-30 18:11:00 +02:00
953c92ac61
Check for errors
2023-03-30 18:08:31 +02:00
2ee1bc4079
Add reset_3d_array function
2023-03-30 18:08:13 +02:00
eeff720ae4
Add CUDA blocksize to config.h
2023-03-28 12:54:49 +02:00
2d6b4fe011
Implement copy_3d_array in CUDA
2023-03-28 12:38:06 +02:00
57954a27c0
Update tests
2023-03-27 18:30:19 +02:00
julienChemillier
5f684d9f91
Implementation of Adam Optimizer
2023-03-27 18:17:50 +02:00
677de892e7
Add function.cu
...
Translate `apply_function_input` and `get_activation_function` to CUDA
2023-03-26 17:14:07 +02:00
e4003aea28
Add config.h
2023-03-24 15:21:57 +01:00
7da6541937
Rename BOUND_RELU to RELU_CLIP_VALUE
2023-03-24 15:12:31 +01:00
b893f11da0
Add random offset to mnist images during learning
2023-03-22 13:03:19 +01:00
julienChemillier
1bd92074ab
Back to multiple bias implementation
2023-03-18 13:25:58 +01:00
00ac29b1d0
Update train.h parameters
2023-03-17 15:46:30 +01:00
bb0c709a50
Remove unusable arguments from dense network
2023-03-15 21:34:30 +01:00
ca06cbe088
Forgot that
2023-03-15 21:34:03 +01:00
68b84b60d1
Change more occurrences
2023-03-15 21:33:45 +01:00
c078d75ca6
Rename mnist network to dense for consistency
2023-03-15 21:27:57 +01:00
8e411e29f4
Update webserver
2023-03-14 15:20:40 +01:00
1f60df0a5b
Revert 04087c3 removal of CUDA compilation
2023-03-13 19:55:21 +01:00
julienChemillier
84e05a732b
Change of a line
2023-03-13 18:47:32 +01:00
julienChemillier
adeee56f83
Merge branch 'main' of https://github.com/julienChemillier/TIPE
2023-03-13 18:46:31 +01:00
04087c3de4
Rename mnist folder to dense
2023-03-13 13:55:09 +01:00
julienChemillier
b3bc4b7787
Removal of useless variables
2023-03-13 13:43:42 +01:00
julienChemillier
2a88621c34
Change bias implementation to fix keras's one
2023-03-12 11:13:26 +01:00
julienChemillier
fa7d8b2dbc
Add identity as activation function
2023-03-12 09:34:06 +01:00
julienChemillier
179a372159
Fix backpropagation error
2023-03-11 19:40:25 +01:00
julienChemillier
21ee3430c5
Update backpropation of max_pooling
2023-03-11 19:36:46 +01:00
3e57054215
Add webserver/guess
2023-03-11 17:51:52 +01:00
81ff4f4d00
Trying to improve train.c readability
2023-03-11 13:45:00 +01:00
7d7cd2e3a7
Fix FLT_MIN misuses
2023-03-10 18:20:10 +01:00
3a50b08179
Add max_pooling backward
2023-03-10 18:19:23 +01:00
julienChemillier
e2d4ea7949
Ajout le compte_rendu du 18 février
2023-03-09 14:35:15 +01:00
julienChemillier
9d2d61703d
Replace min_float by fminf
2023-03-09 14:27:23 +01:00
julienChemillier
fec4651946
Add 3 new types of initialisation
2023-03-09 06:57:51 +01:00
julienChemillier
91e4a1b316
Fix creation.c
2023-03-08 20:59:53 +01:00
julienChemillier
89402b9ee6
Replace numbers by defines
2023-03-08 20:48:34 +01:00
4aea4d6321
Remove unnecessary tests from GH Actions
2023-03-03 22:18:25 +01:00
c9a890a20a
This workaround seems better
...
See https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66425#c34
2023-03-03 22:14:42 +01:00
cdbabd21be
Update test/cnn_structure
2023-03-03 22:06:25 +01:00
6362382e91
Remove warning [-Wunused-result]
2023-03-03 22:06:09 +01:00
9d54b1e4ea
Rename variables
2023-03-03 21:59:51 +01:00
796a94207f
Add comments to cnn.c
2023-03-03 21:58:05 +01:00
177aaa869c
Reduce redundancy in function.c
2023-03-03 21:56:42 +01:00
7f88acf17f
Add float clip(float) to update.c
2023-03-02 10:35:25 +01:00
c45b21e322
Rename MAX_RESEAU to CLIP_VALUE
2023-03-01 19:12:57 +01:00
95ce123587
Fix wrong argument for pooling (forward)
2023-03-01 09:37:40 +01:00
