9c2b0f6 | Gagandeep Singh | 08 April 2020, 14:07:30 UTC | Update opt_pk_resize.c | 08 April 2020, 14:07:30 UTC |
6bb10c3 | Gagandeep Singh | 19 March 2020, 16:39:45 UTC | added support for not creating new noise symbol per relu for deepz | 19 March 2020, 16:39:45 UTC |
62ccc06 | Gagandeep Singh | 14 March 2020, 16:18:48 UTC | Merge pull request #55 from b-3-n/master Fix double free. | 14 March 2020, 16:18:48 UTC |
54d9c6a | Benjamin Mularczyk | 04 March 2020, 12:21:40 UTC | Fix double free. | 04 March 2020, 12:21:40 UTC |
3fc0d03 | Gagandeep Singh | 07 February 2020, 17:15:12 UTC | Update README.md | 07 February 2020, 17:15:12 UTC |
162d740 | Gagandeep Singh | 07 February 2020, 17:13:14 UTC | Update README.md | 07 February 2020, 17:13:14 UTC |
20efe61 | Christoph Müller | 07 February 2020, 17:05:45 UTC | Changed licensing year | 07 February 2020, 17:06:42 UTC |
02f057f | Christoph Müller | 07 February 2020, 16:58:49 UTC | Fixed license text in Makefile | 07 February 2020, 16:58:49 UTC |
2524ecc | Christoph Müller | 07 February 2020, 16:37:37 UTC | Removed support for sub and mul layer in GPU version | 07 February 2020, 16:37:37 UTC |
14ee57d | Christoph Müller | 07 February 2020, 16:29:02 UTC | Added Contributors list to README | 07 February 2020, 16:32:30 UTC |
856bcfb | Christoph Müller | 07 February 2020, 15:39:49 UTC | Use traditional include guards, put default parameter in python interface | 07 February 2020, 15:39:49 UTC |
8b11d18 | Christoph Müller | 03 February 2020, 13:28:12 UTC | Exposing device_id and backstep_depth to python interface | 03 February 2020, 13:28:12 UTC |
de1c8ec | Christoph Müller | 03 February 2020, 01:15:33 UTC | Added copyright disclaimers and author, filenames, summaries | 03 February 2020, 01:15:33 UTC |
2238094 | Christoph Müller | 03 February 2020, 01:00:38 UTC | Update of the Python interface | 03 February 2020, 01:00:38 UTC |
8bcb954 | Christoph Müller | 29 January 2020, 23:44:46 UTC | Added support for sub and mul layers | 29 January 2020, 23:44:46 UTC |
25c7fda | Christoph Müller | 29 January 2020, 19:13:18 UTC | Merge branch 'master' of https://github.com/eth-sri/ELINA into fppoly_gpu | 29 January 2020, 19:13:18 UTC |
b96b684 | Christoph Müller | 19 December 2019, 21:34:43 UTC | Edited Makefile and configure file to make CUDA optional | 19 December 2019, 21:34:43 UTC |
5eadd82 | Christoph Müller | 18 December 2019, 14:52:18 UTC | Cleanup | 18 December 2019, 14:52:18 UTC |
b30637a | Christoph Müller | 18 December 2019, 14:17:00 UTC | Added support for update_bounds_for_neuron | 18 December 2019, 14:17:00 UTC |
e9d5576 | Gagandeep Singh | 18 December 2019, 10:01:14 UTC | fixed bug in sub | 18 December 2019, 10:01:14 UTC |
0b35f0f | Christoph Müller | 17 December 2019, 21:55:45 UTC | Switched to floating point sound rounding | 17 December 2019, 21:55:45 UTC |
e6d47e8 | Gagandeep Singh | 17 December 2019, 18:45:31 UTC | added add and subtract at the first layer | 17 December 2019, 18:45:31 UTC |
3bdc0b0 | Christoph Müller | 17 December 2019, 12:09:11 UTC | Separated Gpupoly into separate files - interval.h for all basic interval computations - affine_transform.h for all linear functions - relu_approx.h for the relu approximation kernels | 17 December 2019, 12:09:11 UTC |
d40c96e | Gagandeep Singh | 16 December 2019, 22:05:18 UTC | added support for intermediate mult and sub with constant | 16 December 2019, 22:05:18 UTC |
056082a | Christoph Müller | 13 December 2019, 17:12:10 UTC | Fixed incorrect indents in fppoly_gpu.h | 13 December 2019, 17:12:10 UTC |
0a25fb6 | Christoph Müller | 13 December 2019, 17:01:50 UTC | Defined type variables in python interface | 13 December 2019, 17:01:50 UTC |
d0c7679 | Christoph Müller | 13 December 2019, 15:35:43 UTC | Introduced cuda convenience methods - Substituted all calls to cudaMalloc and cudaMemcpy by calls to more convenient templated functions - Bugfixes | 13 December 2019, 15:35:43 UTC |
5cbcdb3 | Christoph Müller | 12 December 2019, 13:12:07 UTC | Merged in training branch | 12 December 2019, 13:12:07 UTC |
1423111 | Christoph Müller | 06 December 2019, 11:22:09 UTC | Edited Python Interface | 23 January 2020, 09:14:44 UTC |
aad768e | Christoph Müller | 19 November 2019, 17:02:23 UTC | Changed GPUPoly to be compatible with new ERAN | 19 November 2019, 17:02:23 UTC |
f20d2e3 | Christoph Müller | 19 November 2019, 14:16:55 UTC | Merge branch 'master' into fppoly_gpu | 19 November 2019, 14:16:55 UTC |
62d5de3 | Christoph Müller | 19 November 2019, 14:02:34 UTC | Merge branch 'master' of https://github.com/eth-sri/ELINA | 19 November 2019, 14:02:34 UTC |
54eb9c3 | Christoph Müller | 13 November 2019, 14:00:41 UTC | Simplified area_heuristic code in all relu_bounds functions | 13 November 2019, 14:12:32 UTC |
4c6c85d | Gagandeep Singh | 05 November 2019, 13:10:49 UTC | fixed bug in compute lb for geometric specs | 05 November 2019, 13:10:49 UTC |
fd612d4 | Gagandeep Singh | 04 November 2019, 21:28:31 UTC | fixed memory leak in elina_abstract0_from_zonotope | 04 November 2019, 21:28:31 UTC |
86db404 | Gagandeep Singh | 04 November 2019, 21:27:30 UTC | fixed memory leak in elina_abstract0_from_zonotope | 04 November 2019, 21:27:30 UTC |
f1eb893 | Gagandeep Singh | 04 November 2019, 21:15:48 UTC | fixed memory leak in elina_abstract0_from_zonotope | 04 November 2019, 21:15:48 UTC |
b4363b3 | Gagandeep Singh | 04 November 2019, 21:10:33 UTC | added elina_abstract0_from_zonotope | 04 November 2019, 21:10:33 UTC |
dced645 | Gagandeep Singh | 04 November 2019, 20:57:42 UTC | added elina_abstract0_from_zonotope | 04 November 2019, 20:57:42 UTC |
b081e07 | Christoph Müller | 31 October 2019, 00:58:30 UTC | Switch back to double | 31 October 2019, 00:58:30 UTC |
325b620 | Christoph Müller | 29 October 2019, 15:15:58 UTC | Bugfixes in Python interface for training | 29 October 2019, 16:45:55 UTC |
1609132 | Christoph Müller | 28 October 2019, 03:50:59 UTC | Fixed pointers not being set to nullptr after free and wrong typing | 28 October 2019, 04:22:59 UTC |
5738fc7 | Christoph Müller | 26 October 2019, 04:51:40 UTC | Removed all printouts from training branch | 26 October 2019, 04:51:40 UTC |
c088632 | Christoph Müller | 26 October 2019, 04:05:42 UTC | Fixed booleans for training data retention not being passed correctly | 26 October 2019, 04:06:55 UTC |
820c17c | Christoph Müller | 26 October 2019, 02:57:14 UTC | Changed input data type of network data from Python from double to float - A more flexible method on the side of python would be desirable | 26 October 2019, 04:07:11 UTC |
9d456de | Christoph Müller | 26 October 2019, 02:46:37 UTC | Implementing training interface towards Python | 26 October 2019, 04:07:06 UTC |
5d6defb | Christoph Müller | 25 October 2019, 19:48:44 UTC | Putting the sparse coeffs into CPU arrays | 26 October 2019, 04:07:03 UTC |
c620e39 | Christoph Müller | 25 October 2019, 19:48:35 UTC | Printing the sparse coeffs | 26 October 2019, 04:06:55 UTC |
e951763 | Christoph Müller | 23 October 2019, 20:37:25 UTC | Added consts wherever possible and fixed variable names in merge method | 26 October 2019, 03:57:51 UTC |
b80f3e4 | Christoph Müller | 23 October 2019, 18:49:09 UTC | Replaced size_t and long int by int in all GPU functions | 26 October 2019, 03:58:20 UTC |
6b7bbe4 | Christoph Müller | 23 October 2019, 16:58:22 UTC | Cleanup - Resorting methods - Now number of threads is always num_threads - Removed unnecessary function calls and functions | 26 October 2019, 03:57:51 UTC |
708f09c | Christoph Müller | 23 October 2019, 16:53:10 UTC | Merged relu_bounds methods | 26 October 2019, 03:58:03 UTC |
79b9926 | Christoph Müller | 23 October 2019, 16:19:12 UTC | Cleanup of csts methods | 26 October 2019, 03:57:51 UTC |
dc223c7 | Christoph Müller | 23 October 2019, 15:44:06 UTC | Merged csts_from_previous_layer methods | 26 October 2019, 03:57:51 UTC |
8826f2e | Christoph Müller | 23 October 2019, 15:14:17 UTC | Added more thread-flexible coeff kernel launching | 26 October 2019, 03:49:08 UTC |
53f18e4 | Christoph Müller | 23 October 2019, 01:52:57 UTC | Possible performance optimization in coeffs methods (still needs proof) - Try to only load weights once if there are non-zero coeffs - Not clear if this really improves performance - Possible pipeline stall if all 4 coeffs need to be loaded at the same time | 26 October 2019, 03:49:59 UTC |
6c233d4 | Christoph Müller | 23 October 2019, 01:33:36 UTC | Resorted methods | 23 October 2019, 01:33:36 UTC |
d50524c | Christoph Müller | 23 October 2019, 01:30:53 UTC | Merged coeff_from_previous_layer methods | 23 October 2019, 01:30:53 UTC |
b637379 | Christoph Müller | 17 October 2019, 13:48:09 UTC | Added reduction | 22 October 2019, 15:20:25 UTC |
9a85ae1 | gsingh | 22 October 2019, 11:56:21 UTC | support for get_bounds_linexpr0 residual laers | 22 October 2019, 11:56:21 UTC |
644ed43 | Gagandeep Singh | 21 October 2019, 14:18:21 UTC | fixed free in zonoml maxpool | 21 October 2019, 14:18:21 UTC |
d65cd87 | Gagandeep Singh | 16 October 2019, 22:28:33 UTC | better bounds | 16 October 2019, 22:28:33 UTC |
41cbbe9 | Christoph Müller | 16 October 2019, 18:02:05 UTC | Labeling is now correct for polyhedral input bounds - the update_.._lower_half function did not take the geometric changes into account | 16 October 2019, 18:03:51 UTC |
f8c9923 | Gagandeep Singh | 16 October 2019, 15:27:19 UTC | change in maxpool for zonoml | 16 October 2019, 15:27:19 UTC |
b6d4339 | Gagandeep Singh | 15 October 2019, 19:11:28 UTC | refacted fpppoly code | 15 October 2019, 19:11:28 UTC |
e76a023 | Christoph Müller | 14 October 2019, 15:39:29 UTC | Renamed update methods | 14 October 2019, 15:39:29 UTC |
1be0556 | Christoph Müller | 14 October 2019, 15:19:28 UTC | Fixed wrong typing in update function definitions - Now everything is consistently typed as float_type if it is not a double input coming from python | 14 October 2019, 15:19:28 UTC |
dcaa71b | Christoph Müller | 09 October 2019, 08:54:33 UTC | Now support same API for polyhedral input constraints as CPU version | 14 October 2019, 15:14:39 UTC |
fcc7a88 | Gagandeep Singh | 14 October 2019, 14:44:08 UTC | fixed an issue when expression is null | 14 October 2019, 14:44:08 UTC |
3bb6530 | Christoph Müller | 12 September 2019, 08:34:51 UTC | Now precompute indices for conv backsubstitution - No longer necessary to skip over out-of-bound indices in the for loops | 14 October 2019, 14:05:24 UTC |
6bf5c45 | Gagandeep Singh | 10 October 2019, 09:01:30 UTC | correction | 10 October 2019, 09:01:30 UTC |
141c690 | Gagandeep Singh | 10 October 2019, 06:53:58 UTC | remove a block and factor | 10 October 2019, 06:53:58 UTC |
f93a9a6 | Gagandeep Singh | 07 October 2019, 20:47:08 UTC | added function for accessing affine form | 07 October 2019, 20:47:08 UTC |
7b285e9 | Gagandeep Singh | 04 October 2019, 15:49:35 UTC | made changes to ffn_sub_bias | 04 October 2019, 15:49:35 UTC |
e0e29a6 | Gagandeep Singh | 03 October 2019, 18:44:40 UTC | added support for bias sub and mult | 03 October 2019, 18:44:40 UTC |
b96fd24 | Gagandeep Singh | 01 October 2019, 23:33:22 UTC | anytime backsubstitution | 01 October 2019, 23:33:22 UTC |
174341f | Gagandeep Singh | 01 October 2019, 22:34:10 UTC | fixed | 01 October 2019, 22:34:10 UTC |
043eaef | Gagandeep Singh | 01 October 2019, 22:27:49 UTC | fixed | 01 October 2019, 22:27:49 UTC |
7147092 | Gagandeep Singh | 01 October 2019, 19:57:33 UTC | Merge pull request #52 from jmct/master Add ./configure option to suppress overflow warnings | 01 October 2019, 19:57:33 UTC |
00f83cf | Gagandeep Singh | 01 October 2019, 15:22:20 UTC | fixed bug | 01 October 2019, 15:22:20 UTC |
c6dff3c | José Manuel Calderón Trilla | 01 October 2019, 14:35:59 UTC | Add ./configure option to suppress overflow warnings This addresses part of the concerns in issue #39. If we know the overflow is sound, then we can choose to skip printing to `stderr` when the sound overflow occurs. The implementation is straightforward: * We added an option to `./configure`: `-no-warn-overflow` * When enabled, this adds `-DNO_WARN_OVERFLOW` to `extra_elina_options` * The printing to `stderr` in `elina_poly/opt_pk/vector.c:570` is now under a CPP macro that checks whether `-DNO_WARN_OVERFLOW` is defined. | 01 October 2019, 14:40:54 UTC |
f4d7fca | Gagandeep Singh | 01 October 2019, 14:23:29 UTC | updates for get_bounds_for_linexpr0 | 01 October 2019, 14:23:29 UTC |
50e509b | Gagandeep Singh | 30 September 2019, 20:04:21 UTC | fixed bug in understanding of gather | 30 September 2019, 20:04:21 UTC |
cce35e3 | Gagandeep Singh | 30 September 2019, 19:58:49 UTC | added support for gather in zonoml | 30 September 2019, 19:58:49 UTC |
eacc37c | Gagandeep Singh | 27 September 2019, 18:09:42 UTC | changed interface for convolution in zonoml and fpppoly | 27 September 2019, 18:09:42 UTC |
4fb9be0 | Gagandeep Singh | 23 September 2019, 17:19:06 UTC | added python interface for constructing abstract element from zonotope | 23 September 2019, 17:19:06 UTC |
fefbc28 | Christoph Müller | 06 September 2019, 14:10:35 UTC | Refactored the two update methods back together - Only one update method remaining, which works for both conv and residual - Also only one size predictor remaining | 06 September 2019, 14:10:35 UTC |
858c15f | Christoph Müller | 03 August 2019, 20:41:33 UTC | Commented out one of the interval multiplication methods | 03 August 2019, 20:41:33 UTC |
963ccdd | Christoph Müller | 03 August 2019, 20:28:25 UTC | Now do not calculate new lengths in kernels but before | 03 August 2019, 20:28:25 UTC |
13fbf81 | Christoph Müller | 03 August 2019, 19:55:49 UTC | Size predictors now also work with limited backpropagation | 03 August 2019, 19:55:49 UTC |
67d20d8 | Christoph Müller | 03 August 2019, 19:28:39 UTC | Removed multi-threading and atomics from ffn relu functions | 03 August 2019, 19:28:39 UTC |
a9e7c07 | Christoph Müller | 03 August 2019, 00:53:03 UTC | Implemented counter for how many residual layers to go back | 03 August 2019, 03:10:09 UTC |
8de553e | Christoph Müller | 02 August 2019, 21:12:09 UTC | Size prediction now works for residual layers - Some more small fixes | 02 August 2019, 21:12:09 UTC |
d1b05c7 | Christoph Müller | 02 August 2019, 19:08:33 UTC | Bugfix - accidentally added biases twice | 02 August 2019, 19:08:33 UTC |
7682184 | Gagandeep Singh | 01 August 2019, 18:41:15 UTC | replaced mno-fused-add option with -ffp-contract=off | 01 August 2019, 18:41:15 UTC |
1f6dc4b | Gagandeep Singh | 01 August 2019, 18:33:18 UTC | fixed a bug for substitute linexpr, thanks to Jorge Navas for spotting | 01 August 2019, 18:33:18 UTC |
d6b7ac6 | Gagandeep Singh | 31 July 2019, 00:44:19 UTC | made a commit for get_ub_from_previous layuers | 31 July 2019, 00:44:19 UTC |
9efc7ef | Christoph Müller | 30 July 2019, 23:32:18 UTC | Support richer set of residual layers - Now can handle if there is no layer on one branch - Now can deal with relu-activation directly before residual layer | 30 July 2019, 23:32:18 UTC |
ccdfb39 | Christoph Müller | 30 July 2019, 21:58:34 UTC | Treat residual layers like a convolutional layer [1, 1, #filters] - the res_add_layer methods uses the same dimensions as for the previous layer (output_size, input_size), but filter_size is set to [1, 1], strides are set to [1, 1] and padding to [0, 0]. - code now also can handle convolutional layers without relu activation | 30 July 2019, 22:09:38 UTC |
ba325ab | Christoph Müller | 30 July 2019, 05:14:33 UTC | Merge branch 'master' into fppoly_gpu | 30 July 2019, 05:14:33 UTC |