NextCloud/dlib
1
0
Fork 0
mirror of https://github.com/davisking/dlib.git synced 2024-11-01 10:14:53 +08:00
Code Issues Packages Projects Releases Wiki Activity
7,767 Commits 4 Branches 61 Tags 147 MiB
9191ebc776
Commit Graph

7767 Commits

This Branch
This Branch
All Branches
Author SHA1 Message Date
Juha Reunanen
9191ebc776 Keep estimating steps_without_progress, even if steps_since_last_learning_rate_shrink < iter_without_progress_thresh 2019-11-12 14:37:43 +02:00
Juha Reunanen
3746b3c1d3 Simplify code as per review comments 2019-11-12 14:31:17 +02:00
Juha Reunanen
d48b406a11 Revert "If available, use constexpr if" 
This reverts commit 503d4dd335.
 2019-11-12 13:55:57 +02:00
Juha Reunanen
503d4dd335 If available, use constexpr if 2019-11-12 07:22:56 +02:00
Juha Reunanen
7cf7152993 Add operator== for the other pixel types as well; remove the inline 2019-11-11 16:44:31 +02:00
Juha Reunanen
0b4a0bc2ad Review fixes 2019-11-11 16:05:42 +02:00
Juha Reunanen
a8eef91ac7 Clarify the intention and the code 2019-11-09 09:28:46 +02:00
Juha Reunanen
e79a3cc0fb Add helpful logging 2019-11-09 08:51:01 +02:00
Juha Reunanen
d4f6e5e538 Detect cats instead of aeroplanes 2019-11-09 08:47:40 +02:00
Juha Reunanen
e76ea537f1 Fix maintaining the dumped (now "effectively disregarded") loss values count 2019-11-09 08:42:04 +02:00
Juha Reunanen
4b99bff7a7 Keep enough previous loss values, until the disk sync 2019-11-09 07:54:21 +02:00
Juha Reunanen
2610f728f0 Revert "Revert some of the earlier change" 
This reverts commit 6c852124ef.
 2019-11-08 17:07:59 +02:00
Juha Reunanen
125ae0a142 Disregard dumped loss values only when deciding if learning rate should be shrunk, but *not* when deciding if loss has been going up since last disk sync 2019-11-08 12:58:18 +02:00
Juha Reunanen
6c852124ef Revert some of the earlier change 2019-11-08 12:32:42 +02:00
Juha Reunanen
2b6dbff8bb Improve behaviour when loss actually increased after disk sync 2019-11-06 07:56:33 +02:00
Juha Reunanen
dfaf715e49 Add previous_loss_values_dump_amount to previous_loss_values.size() when deciding if loss has been increasing 2019-11-05 18:13:20 +02:00
Juha Reunanen
eeff592ed2 Loss may have increased, even if prob_loss_increasing_thresh > prob_loss_increasing_thresh_max_value 2019-11-05 17:03:52 +02:00
Juha Reunanen
acafd0cabd Minor parameters tweaking 2019-11-04 21:31:55 +02:00
Juha Reunanen
f561801f8a Problem: CUDA errors not reported properly to console 
Solution: stop and join data loader threads even in case of exceptions
 2019-11-04 19:03:12 +02:00
Juha Reunanen
73a1ba8a86 Simplify the segmentation network structure; make the object detection network more complex in turn 2019-11-04 18:45:14 +02:00
Juha Reunanen
316d62b254 Ignore large aspect ratios; refactor the code; tweak some network parameters 2019-11-03 21:21:57 +02:00
Juha Reunanen
6d18308c0a Add helpful comments 2019-10-30 16:01:31 +02:00
Juha Reunanen
865f2ae10e Improve bilinear interpolation: if output type is an integer, round instead of truncating 2019-10-30 12:31:29 +02:00
Juha Reunanen
b442a3dc8e Fix interpolating the predicted mask 2019-10-30 12:18:53 +02:00
Juha Reunanen
c9efc4da76 Fix sorting criterion 2019-10-30 12:12:15 +02:00
Juha Reunanen
b66a68d8f4 Allow more overlap 2019-10-29 16:26:15 +02:00
Juha Reunanen
030c468003 Use separate synchronization file for each seg net of each class 2019-10-29 14:50:15 +02:00
Juha Reunanen
2f17289803 Train a separate instance segmentation network for each classlabel 2019-10-29 14:32:29 +02:00
Juha Reunanen
9b20ea1015 Write detector window classes to ostream; also group detection windows by class (when ostreaming) 2019-10-29 13:28:46 +02:00
Juha Reunanen
4442546437 Draw rect in the color of the corresponding class 2019-10-29 12:27:10 +02:00
Juha Reunanen
24aaa40603 Allow to choose multiple classes 2019-10-29 12:19:56 +02:00
Juha Reunanen
eef863aa2e Fix indentation 2019-10-29 10:15:57 +02:00
Juha Reunanen
7a87f648a9 Fix class index problem 2019-10-29 10:15:34 +02:00
Juha Reunanen
e996bbf35f Print to console what was found 2019-10-29 10:06:30 +02:00
Juha Reunanen
5ce0522869 Automatically skip images with no detections 2019-10-28 17:34:08 +02:00
Juha Reunanen
773d3dd81d Remove outdated TODO remark 2019-10-28 07:19:12 +02:00
Juha Reunanen
6025fa3ffd Really train only instances of the selected class 2019-10-28 07:16:36 +02:00
Juha Reunanen
a841eff8a9 Learn only one class 2019-10-28 07:07:16 +02:00
Juha Reunanen
edb083709a Tweak default minibatch size 2019-10-28 06:48:12 +02:00
Juha Reunanen
414b5543f6 Read even class labels 2019-10-28 06:43:02 +02:00
Juha Reunanen
9fbec6abbb Simplify ground-truth reading; fix random cropping 2019-10-27 20:02:19 +02:00
Juha Reunanen
2b28aac83f Add some perturbation to the inputs 2019-10-27 19:26:59 +02:00
Juha Reunanen
31b8cee36e Improve masking of the current instance 2019-10-27 19:13:02 +02:00
Juha Reunanen
a554081a4e Draw the largest findings last 2019-10-27 18:47:40 +02:00
Juha Reunanen
d0cb04246e Fix the cropping rect's aspect ratio; also slightly expand the rect 2019-10-27 18:46:51 +02:00
Juha Reunanen
507f2b325c Take ignore-color into account 2019-10-27 17:41:26 +02:00
Juha Reunanen
276e763df0 Set label image size 2019-10-27 17:07:34 +02:00
Juha Reunanen
0fc72cc00e Set result size 2019-10-26 23:16:06 +03:00
Juha Reunanen
9d62178df8 Revert "Set overlaps_ignore" 
This reverts commit 65adeff1f8.
 2019-10-26 21:39:25 +03:00
Juha Reunanen
476da8fde8 Add TODO remarks 2019-10-26 21:08:21 +03:00