Sam Gross
625850c2c2
Check cuDNN version at runtime ( #1586 )
...
* Check cuDNN version at runtime
This checks that the version from cudnn.h matches the version from
libcudnn.so.
Fixes #1476
* Only check major and minor version numbers
2017-05-19 01:55:09 -04:00
Sam Gross
e6c9509a41
Fix call to Tensor.set_ in rnn.py ( #1592 )
2017-05-18 20:28:49 -04:00
Sam Gross
b9379cfab7
Use cuDNN and NCCL symbols from _C library ( #1017 )
...
This ensures that we use the same library at the C++ level and with
Python ctypes. It moves the searching for the correct library from
run-time to compile-time.
2017-03-16 16:10:17 -04:00
Adam Paszke
1487278fdf
Allow backprop through cuDNN RNN in eval mode
...
Handling of dropout descriptors has been improved too.
2017-03-01 19:42:39 +01:00
Adam Paszke
da725830c2
Add support for variable length sequences in RNNs ( #873 )
2017-03-01 17:36:32 +01:00
Christian Sarofeen
04aba1caec
Fix cuDNN dropout desc for multi-gpu ( #772 )
2017-02-17 19:16:12 +01:00
bdfhjk
a217fefee1
Update rnn.py
...
Fixed a problem with outputting the RuntimeError if arguments are incorrect in cudnn/rnn.py
2017-02-15 21:49:42 +01:00
Adam Paszke
72c1982734
Add some more asserts to cuDNN RNN
2017-02-14 21:28:50 +01:00
Adam Paszke
63edca44f2
Add tests for non-contiguous inputs and gradients
2017-02-14 21:28:50 +01:00
ngimel
f096fb6859
adding cudnn V6 support ( #515 )
2017-01-31 02:01:37 +01:00
Adam Paszke
0180e638e5
Remove unnecessary zero_() calls in cuDNN RNN
2017-01-28 14:36:57 +01:00
Adam Paszke
95c6ae04fb
Fix non-contiguous grad handling in cuDNN RNN
2017-01-28 14:36:57 +01:00
Luke Yeager
e7c1e6a8e3
[pep8] Fix most lint automatically with autopep8
...
Here's the command I used to invoke autopep8 (in parallel!):
git ls-files | grep '\.py$' | xargs -n1 -P`nproc` autopep8 -i
Several rules are ignored in setup.cfg. The goal is to let autopep8
handle everything which it can handle safely, and to disable any rules
which are tricky or controversial to address. We may want to come back
and re-enable some of these rules later, but I'm trying to make this
patch as safe as possible.
Also configures flake8 to match pep8's behavior.
Also configures TravisCI to check the whole project for lint.
2017-01-28 01:15:51 +01:00
ngimel
b32dd4a876
add cudnn deb package installation paths to cudnn discovery, add 5.1.10 to load options ( #448 )
2017-01-13 14:32:23 -05:00
ngimel
59b23d79c6
fix cudnn rnn batch_first with tests ( #445 )
...
* fix cudnn rnn batch_first with tests
2017-01-13 13:40:27 -05:00
Adam Lerer
183b3aacd2
Hold CuDNN PRNG state between RNN iterations
2016-12-30 00:14:55 +01:00
Sam Gross
8a29338837
Use cuDNN for Conv3d and ConvTranspose3d ( #359 )
...
I've also updated test_nn.py to run marked tests twice: once with cuDNN
enabled and once with it disabled.
2016-12-28 16:14:47 -05:00
Adam Paszke
cd82b2b869
Implement comparison and logical operators for tensors
2016-12-28 00:04:08 +01:00
soumith
a9c2809ce3
change the order of cudnn libs
2016-12-21 05:44:16 -08:00
Sergey Zagoruyko
5586f48ad5
add cudnn 5.0.5 to supported versions ( #321 )
2016-12-17 07:57:20 -05:00
Adam Paszke
8e09f0590b
Make sure that C extension was compiled with cuDNN before using it
2016-12-15 00:47:55 +01:00
Adam Paszke
0580f5a928
Add __len__ for tensors
2016-12-01 23:14:41 +01:00
Marat Dukhan
e3f440b1d0
Make torch.backends.cudnn work on OSX
2016-11-22 19:06:08 +01:00
Adam Lerer
7f51af7cbc
adding dropout, bidirection, etc. to RNN ( #214 )
2016-11-10 13:25:14 -05:00
Sam Gross
ad2d413c0b
Add C++ bindings for cuDNN ( #167 )
...
The Python ctypes bindings overhead was high enough that it slowed down
multi-gpu training when using 4+ Maxwell GPUs.
2016-10-26 19:51:48 -04:00
Adam Lerer
b5d13296c6
addressing comments
2016-10-23 21:11:22 -07:00
Adam Lerer
86288265ad
Adding rnn cell library
2016-10-23 20:23:48 -07:00
Adam Lerer
1eb6870853
add nobias option to rnn
2016-10-23 20:23:48 -07:00
Adam Lerer
942ca477a6
Copying weights for CUDNN
2016-10-23 20:23:48 -07:00
Adam Lerer
b0e33fb473
cudnn + THNN match with parameters
2016-10-23 20:23:48 -07:00
Adam Lerer
d58b627b98
CUDNN RNN bindings
2016-10-23 20:23:48 -07:00
Sam Gross
a02917f502
Fix typo
2016-10-14 14:07:29 -07:00
Sam Gross
70d8bd04c0
Make cuDNN descriptors extend object
...
Fixes weird double __del__ issue
2016-10-14 13:58:20 -07:00
Soumith Chintala
50326e94b1
try cudnn 5.1.5 and 5.1.3 in that order to load them up. This is needed because cudnn for cuda 7.5 ships with 5.1.3 and cudnn for cuda 8.0 ships with 5.1.5
2016-10-09 22:26:43 -04:00
Soumith Chintala
160723b5b4
fix cudnn lib name
2016-10-09 21:19:50 -04:00
soumith
833bedb46b
cudnn relative check in binary builds
2016-10-02 11:45:46 -07:00
Sam Gross
14965cfce9
Run cuDNN operations on the correct device
2016-09-29 16:27:07 -07:00
Sam Gross
cb5d4e836f
Lazy load CUDA and THNN modules ( #64 )
2016-09-28 19:29:53 -04:00
Soumith Chintala
412019dbe4
fixing CPU builds by making cuda imports optional
2016-09-28 11:56:18 -04:00
Sam Gross
779a460030
Add cuDNN support for convolutions ( #36 )
2016-09-27 17:55:04 -04:00
