fa58c29 | Runqi Yang | 14 March 2022, 02:30:15 UTC | update SpaCy versions and verify pytorch 1.10 compatibility The old 1.x SpaCy is deprecated. So the version is upgraded to 3.x to facilitate recent experiments. | 14 March 2022, 02:30:15 UTC |
c3ea030 | Runqi Yang | 20 December 2018, 05:28:27 UTC | Merge pull request #25 from brettkoonce/pytorch10 replace async with non_blocking for pytorch 1.0 | 20 December 2018, 05:28:27 UTC |
bfdf787 | Brett Koonce | 14 December 2018, 13:19:53 UTC | remove async for pytorch 1.0 | 17 December 2018, 22:27:12 UTC |
6c3b9e5 | Runqi Yang | 03 October 2018, 06:21:14 UTC | Merge pull request #21 from brettkoonce/cut_fix tweak cut command to remove extra space | 03 October 2018, 06:21:14 UTC |
c550ef6 | Brett Koonce | 30 September 2018, 16:15:42 UTC | tweak cut command to remove extra space | 30 September 2018, 22:50:09 UTC |
d9081fe | Runqi Yang | 30 September 2018, 08:01:05 UTC | update Downbench instructions | 30 September 2018, 08:01:05 UTC |
d820d00 | Runqi Yang | 30 September 2018, 07:56:15 UTC | Merge pull request #20 from brettkoonce/dawn_logging add dawn logging option | 30 September 2018, 07:56:15 UTC |
5913086 | Brett Koonce | 30 September 2018, 01:15:39 UTC | add dawn logging option | 30 September 2018, 07:07:25 UTC |
1b4f9f1 | Runqi Yang | 24 September 2018, 07:44:45 UTC | Update README.md | 24 September 2018, 07:44:45 UTC |
c3075f1 | hitvoice | 25 May 2018, 11:44:55 UTC | add more comments | 25 May 2018, 11:44:55 UTC |
c12ac9b | hitvoice | 25 May 2018, 09:29:10 UTC | Faster implementation and better suited pytorch 0.4 - Faster implementation - implement partial tuning via gradient mask instead of copying the (big) original embedding matrix - sort by len before batching (as the original paper says) - better suited pytorch 0.4 - change ".cuda()" to ".to(device)" (this make the code much more readable) - change "clip_grad_norm" to "clip_grad_norm_" - bug fixes - consistent "best_eval_score" after resuming and sanity check for this | 25 May 2018, 09:29:10 UTC |
4f58cbe | Runqi Yang (FA TALENT) | 24 April 2018, 05:26:42 UTC | support CPU inference | 24 April 2018, 05:26:42 UTC |
89e68ef | hitvoice | 12 April 2018, 05:48:10 UTC | support interaction | 12 April 2018, 05:48:10 UTC |
8a25c14 | Runqi Yang | 10 April 2018, 05:59:56 UTC | update to pytorch 0.4 | 10 April 2018, 05:59:56 UTC |
f878b0e | hitvoice | 10 April 2018, 05:54:30 UTC | compatible with pytorch 0.4 - compatible with pytorch 0.4: ".data[0]" -> ".item()", "volatile=True" -> "with torch.no_grad()", etc. - bug fix: cast optimizer state dict to GPU tensors after resuming | 10 April 2018, 05:54:30 UTC |
4ad4452 | hitvoice | 10 April 2018, 05:43:32 UTC | better logging and code structure - now main() is on the top of "train.py" and previous out-of-function house-keeping code are in setup() - fix a bug in drqa/layers.py: call "contiguous()" before "view()" - logging for stdout will not be overwhelmed by thousands of training logs now and we have a progress bar for evaluation | 10 April 2018, 05:43:32 UTC |
90f8a34 | hitvoice | 09 April 2018, 03:56:08 UTC | more robust downloading - skip files that are already downloaded - check requirements before starting download | 09 April 2018, 03:56:08 UTC |
51a7ce0 | hitvoice | 22 February 2018, 12:37:08 UTC | consistent resuming - change AverageMeter to an exponential moving average meter. - save/load meter states. - better random seed synchronization. | 22 February 2018, 12:37:08 UTC |
ede06a7 | yangrunqi | 06 December 2017, 06:01:07 UTC | update to pytorch 0.3.0 | 06 December 2017, 06:01:07 UTC |
303bbfe | yangrunqi | 02 December 2017, 12:36:43 UTC | specify spacy versions | 02 December 2017, 12:36:43 UTC |
9d407ea | yangrunqi | 02 December 2017, 12:33:24 UTC | add progress bar for annotations | 02 December 2017, 12:33:24 UTC |
37d6717 | Runqi Yang | 25 October 2017, 16:04:57 UTC | Update README.md | 25 October 2017, 16:04:57 UTC |
14cd4c2 | 杨润琦 | 25 October 2017, 15:03:36 UTC | Improve this project to reach the scores reported in the original paper (#5) and make it more efficient and more readable. major modifications: - use one-hot encoding for POS & NER features instead of using seperate embeddings; change dropout to 0.4 accordingly. - if words are identical after normalization, their embeddings are averaged. - only replace one invisible character ('\s') with one space. More training examples are reserved because merging multiple spaces to one results in misalignment of answers. improved efficiency: - Code in "prepro.py" is refactored and made more efficient. Time consumption of "prepro.py" is reduced from 515s to 172s (3x faster) on a machine with 8 i7 CPUs and 16GB RAM. - function "get_answer_index" is simplified and much more readable other improvements: - id of each example is reserved for ease of debugging. - vocabulary of POS and NER tags is saved for ease of debugging. - other tiny improvements to make the code more readable. | 25 October 2017, 15:03:36 UTC |
6886ee7 | 杨润琦 | 26 August 2017, 05:17:07 UTC | save arguments to logs | 26 August 2017, 05:17:11 UTC |
baf3c57 | Runqi Yang | 22 August 2017, 02:05:18 UTC | Update ReadMe | 22 August 2017, 02:05:18 UTC |
33c3ec7 | 杨润琦 | 21 August 2017, 16:01:47 UTC | compatible with pytorch 0.2.0 | 21 August 2017, 16:01:47 UTC |
bc0152c | 杨润琦 | 06 August 2017, 05:35:49 UTC | warning for pytorch versions New version of pytorch is released 2 hours ago… | 06 August 2017, 05:35:49 UTC |
ab25b47 | 杨润琦 | 21 July 2017, 07:35:49 UTC | code clean-up (issue #4) resolve #4 | 21 July 2017, 07:35:49 UTC |
827f2ac | 杨润琦 | 07 July 2017, 03:49:09 UTC | fix inconsistence | 07 July 2017, 03:49:09 UTC |
5da25a3 | 杨润琦 | 06 July 2017, 08:02:53 UTC | add doc string - Add doc strings to some functions which may seem confusing at first sight. - fix inconsistent naming (glove->wv) - fix a bug (CLI argument `threads` was not actually applied to each multiprocessing function) | 06 July 2017, 08:02:53 UTC |
76f3649 | 杨润琦 | 01 July 2017, 03:14:20 UTC | remove redundant files | 01 July 2017, 03:14:20 UTC |
dbec0b0 | 杨润琦 | 30 June 2017, 01:50:47 UTC | info update | 30 June 2017, 01:50:47 UTC |
0e40ac7 | 杨润琦 | 29 June 2017, 09:41:07 UTC | fix typos | 29 June 2017, 09:41:07 UTC |
7c648bd | 杨润琦 | 29 June 2017, 04:07:29 UTC | move SVG to RawGit | 29 June 2017, 04:07:29 UTC |
32bccc5 | 杨润琦 | 29 June 2017, 03:54:23 UTC | initial commit | 29 June 2017, 03:54:23 UTC |
34f43ff | Runqi Yang | 29 June 2017, 03:44:58 UTC | Initial commit | 29 June 2017, 03:44:58 UTC |