d38c3cbfdb
target shape of CTC loss function updated. [before] sum(target_lengths) [after] batch (N,S). and minor update
2020-06-05 14:38:52 +00:00
8645108824
update best_norm_ED setting
2019-12-28 14:29:11 +09:00
307e776a11
FT start iter update
2019-12-27 11:01:01 +00:00
8920de93e8
log_dataset.txt, IC19-NED, minor updates
2019-12-27 10:31:47 +00:00
603d463e83
comment update
2019-11-04 04:52:51 +00:00
a19dc0f169
ctc fix..
2019-10-22 17:49:52 +00:00
8749902367
ctc update
2019-10-22 17:26:58 +00:00
ec5319296b
arrange
2019-10-22 13:58:23 +00:00
6dc16df598
(ctc_a) (ctc_b) and fine tuning
2019-10-07 05:33:44 +00:00
1c6efa5218
'fix_ctc_loss_issue'
2019-08-07 16:42:42 +09:00
8f7255fb4b
.cuda() to .to(device)
2019-08-03 08:03:46 +00:00
7498f48452
no_grad instead of requires_grad False
2019-08-03 06:55:32 +00:00
533fdf01ba
fix encode with batch_max_length
2019-07-31 08:04:56 +00:00
a239f46df3
'fix'
2019-07-26 15:02:05 +09:00
7b938741bd
.
2019-07-24 16:20:25 +00:00
2d45ba2fab
use amp grad clipping
2019-07-24 14:42:25 +02:00
a0f43fa0c0
Nvidia Apex for FP16 calculations
...
Included Compatibility with the Nvidia's Apex library, which can do Floating Point16 calculations. This gives significant speedup in training. This code has been tested on a single RTX2070. If the Nvidia Apex library is not found the code should run as normal.
To install Apex: https://github.com/NVIDIA/apex#quick-start
Known bugs:
-Does not work with adam parameter
-Gradient overflow keeps happening at the start, however it automatically reduces loss scale to 8192 after which this notification disappears
examples:
Loading: https://i.imgur.com/3nZROJz.png
Training: https://i.imgur.com/Q2w52m7.png
2019-07-23 15:51:26 +02:00
dd18436360
--data_filtering_off
2019-07-16 10:04:20 +00:00
c3e07b538a
fix minor
2019-07-16 03:30:44 +00:00
f15051a493
add data_filtering mode
2019-07-15 05:33:37 +00:00
c665fce337
comment and demo.py update
2019-05-17 13:44:38 +00:00
ce837aba3c
add PAD option
2019-05-10 02:11:06 +00:00
cf390a0873
upgrade to PyTorch 1.1.0 (use torch.nn.CTCLoss) and test.py update
2019-05-09 03:21:18 +00:00
58494a8503
comment and sensitive mode update
2019-04-17 13:42:17 +00:00
c2e28f5c0d
comment and multi-gpu setting update
2019-04-16 07:28:53 +00:00
78f175cdd9
log write update and fix minor
2019-04-14 12:54:22 +00:00
73f42c7207
comment update for multi-gpu
2019-04-12 08:12:46 +00:00
7da666fc3a
update model arg
2019-04-09 09:06:32 +00:00
d322f4ff00
fix argment name
2019-04-09 06:06:26 +00:00
34e6316856
fix minor
2019-04-07 12:41:11 +00:00
2b38eb2574
'init'
2019-04-05 19:45:29 +09:00
Jeonghun Baek
