OrgLion-ML
/
dlib
mirror of https://github.com/davisking/dlib.git
1
0
Fork 0
Code Issues Releases Wiki Activity
5,801 Commits 4 Branches 61 Tags 140 MiB
Commit Graph

5801 Commits

Author SHA1 Message Date
Davis King 13cc545da3 clarified comments. 2016-05-15 14:31:06 -04:00
Davis King 93bbe5ff15 merged 2016-05-14 20:31:50 -04:00
Davis King 66166c674d Changed the solver interface to take the learning rate and the layer details 
object as an input.  This allows the solvers to exhibit a more complex behavior
that depends on the specific layer.  It also removes the learning rate from the
solver's parameter set and pushes it entirely into the core training code.
This also removes the need for the separate "step size" which previously was
multiplied with the output of the solvers.

Most of the code is still the same, and in the core and trainer the step_size
variables have just been renamed to learning_rate.  The dnn_trainer's relevant
member functions have also been renamed.

The examples have been updated to reflect these API changes.  I also cleaned up
the resnet definition and added better downsampling.
 2016-05-14 20:30:45 -04:00
Davis King 9763c471ac merged 2016-05-13 23:08:47 -04:00
Davis King 8421f213ad Fixed the in-place layers so that they don't interfere with the operation of 
skip layers and add_prev style layers.  In particular, now in-place layers only
overwrite the gradient information in their child layer if they are operating
in in-place mode.  Otherwise, they add their gradients to their child layers.

It should also be noted that it's safe for in-place layers to overwrite
gradients when in in-place mode since their child layers are inaccessible when
in-place layers operate in in-place mode.  This prevents any other layers from
trying to add to the child layer, thereby avoiding the potability of layer
interference.  So the bug this change fixes is that, when not in in-place mode
the child layers are still accessible but in-place layers were *still*
overwriting child gradients.
 2016-05-13 23:08:03 -04:00
Davis King ead735d0b0 Made printing to cout or cerr check for matlab ctrl+c and throw an exception 
if so.  This way, if you have a long running mex file it will be killable if
it is periodically printing.
 2016-05-13 09:01:24 -04:00
Davis King 81eb18a40f Added intel architecture detection so SSE/AVX isn't enabled when it shouldn't be. 2016-05-13 08:50:25 -04:00
Davis King 3124ec4b81 Made dropout layer initialize its random number generator from std::rand() rather than always 
using the same seed.
 2016-05-13 08:35:22 -04:00
Davis King da28fe3b3e Fixed a compiler error on some systems. 2016-05-13 06:40:15 -04:00
Davis King c79d412c32 slightly relaxed tests to avoid false failure 2016-05-13 06:32:18 -04:00
Davis King 022dce43c2 I accidentally turned the max pooling layers into average pooling a few days 
ago.  This fixes that and turns them back into max pooling layers.
 2016-05-12 19:40:24 -04:00
Davis King 6f970e1854 Fixed network constructors so you can give an input layer for initialization 
without getting a compiler error.
 2016-05-12 19:22:53 -04:00
Davis E. King 0846daabdc Merge pull request #108 from SoapGentoo/master 
Augment build-system, allow for multilib installation, support pkg-config discovery
 2016-05-10 06:37:44 -04:00
David Seifert f3eae43897
* Add pkg-config file, allowing for build-system agnostic dependencies 
* Define LIB_INSTALL_DIR cache variable, allowing for multilib installations
* Discover BLAS and LAPACK via pkg-config if possible
* Fix incorrect toolchain variables in "dlib/test/makefile"
 2016-05-10 09:18:27 +02:00
Davis King cad73fae9d Fixed compiler warnings 2016-05-09 21:46:04 -04:00
Davis King 1e70c721a4 Made example use the "everything" version of avg pooling. 2016-05-07 14:30:42 -04:00
Davis King f669552186 Added max_pool_everything and avg_pool_everything. 2016-05-07 14:30:18 -04:00
Davis King 6d2495a20a Changed extract_fhog_features() back to returning matrix<double> since that's 
what it's contract says needs to happen (and what existing code expects).
 2016-05-07 07:53:27 -04:00
Davis King 51065b9808 Made test error reporting more reliable 2016-05-05 06:58:52 -04:00
Davis King 4a7633056c Fixed avg pooling filter sizes to avoid errors with the new rules about 
non-one based strides.
 2016-05-04 21:40:29 -04:00
Davis King e4d70f8d3d Improved error messages 2016-05-04 21:39:31 -04:00
Davis King 4ef5908b0a Pushed the padding parameters into the con_, max_pool_, and avg_pool_ 
interfaces.  Also changed the default behavior when the stride isn't 1.  Now
the filters will be applied only to the "valid" part of the image.
 2016-05-04 21:32:59 -04:00
Davis King 6bab1f50ad Added missing ; to DLIB_TEST and DLIB_TEST_MSG macros now that they are required. 2016-05-04 21:27:10 -04:00
Davis King a17e7f6b3e fixed typo in comment 2016-05-04 21:13:28 -04:00
Davis King 73d7835590 Added a padding parameter to the pooling and convolution classes. Still need 
to expose it in the final layer interface.
 2016-05-03 22:55:46 -04:00
Davis King 08c8778479 Made test less likely to false alarm. 2016-05-03 21:44:39 -04:00
Davis King 6e9b9d035d Improved macros 2016-05-03 20:36:09 -04:00
Davis King 707a6b4c88 fixed typo in spec 2016-05-02 18:53:00 -04:00
Davis King 4649193f9c Made dlib work with libgif v5 2016-05-01 06:56:34 -04:00
Davis King 9260243c21 Fixed bugs in multi-gpu training code. 2016-04-30 12:53:53 -04:00
Davis King 26d2d889df Added comments 2016-04-30 12:28:00 -04:00
Davis King d161d0bada Added a num_device_groups() member function. 2016-04-30 11:09:50 -04:00
Davis King 909ca040fe Made launch_kernel() not generate CUDA errors if you ask it to launch a kernel of size 0. 2016-04-30 10:42:16 -04:00
Davis King f5abfca18c Added missing #include. 2016-04-29 09:53:05 -04:00
Davis King 4ee88457ca Fixed namespace lookup problem due to a name clash. 2016-04-29 09:40:19 -04:00
Davis King 893824a18c Fixed compiler warnings 2016-04-29 08:04:54 -04:00
Davis King aba68d6f0c merged 2016-04-29 06:56:25 -04:00
Davis King 428b0bb856 Made multi-gpu mode use GPUDirect rather than copying through the CPU. 2016-04-29 06:55:53 -04:00
Davis King 0d6e3f12d6 Made the multi_device_tensor_averager not assume the size of the tensors is 
known at set() time.
 2016-04-29 06:55:04 -04:00
Davis King b85688acec clarified spec 2016-04-29 06:54:30 -04:00
Davis King 1f0705ae92 clarified example 2016-04-28 19:41:27 -04:00
Davis King 7770225a88 Fixed a bug in gaussian_blur() that caused messed up outputs when big sigma 
values were used on some pixel types.
 2016-04-28 18:57:15 -04:00
Davis King 482d82db41 Added multi_device_tensor_averager 2016-04-27 21:50:31 -04:00
Davis King d3f12d832c Added more overloads of affine_transform() 2016-04-27 20:20:06 -04:00
Davis King 7d5fb9c60c Added device_id() methods to gpu_data and tensor objects. These functions 
allow you to find out which device owns the memory inside these objects.
 2016-04-27 20:18:57 -04:00
Davis King 8ece542852 Made the cout and cerr redirection in mex files restore the state of the 
streambufs before exiting.
 2016-04-27 12:36:50 -04:00
Davis King 9f2f6146ec merged 2016-04-27 08:50:23 -04:00
Davis King 339ac50d21 Fixed double counting of mini-batches for the purposes of solver termination 
when multiple GPUs are used.
 2016-04-27 08:49:02 -04:00
Davis King 0f180a68d6 Fixed bad random number generation in layer initialization. 2016-04-27 08:48:23 -04:00
Davis King dfd9543cdb Made the elastic_net inputs be in terms of trans(X)*X and trans(X)*Y rather 
than raw X and Y matrices.
 2016-04-26 10:24:19 -04:00