mirrors
/
fast-reid
mirror of https://github.com/JDAI-CV/fast-reid.git
1
0
Fork 0
Code Issues Projects Releases Wiki Activity
532 Commits
1 Branch
4 Tags
15 MiB
Commit Graph

55 Commits (master)

Author SHA1 Message Date
liaoxingyu 7e652fea2a feat: Add contiguous parameters support 
Support contiguous parameters to train faster. It can split parameters into different contiguous groups by freeze_layer, lr and weight decay.
 2021-07-05 11:10:37 +08:00
liaoxingyu 62ad5e1a8b bugfix for cuda tensor to numpy 
Move cuda tensor to host memory in predictor which can be used for following process.
 2021-06-17 16:05:45 +08:00
liaoxingyu 8f8cbf9411 fix lr scheduler warning when amp training 
Skip lr scheduler when this iteration creates NaN gradients
 2021-06-02 16:35:46 +08:00
liaoxingyu 256721cfde Impl `freezebb` in optimizer's step() 
Make impl. of `freezebb` consistent with impl. of grad clip, and both are implemented through step() in optimizer
 2021-05-31 17:15:26 +08:00
liaoxingyu 2b65882447 change way of layer freezing 
Remove `find_unused_parameters` in DDP and add a new step function in optimizer for freezing backbone. It will accelerate training speed in this way.
 2021-05-25 15:57:09 +08:00
liaoxingyu 0c8e3d9805 update imbalanced sampler 
Summary: add a new sampler, which is useful for imbalanced or long-tail dataset. This refers to ufoym/imbalanced-dataset-sampler.
 2021-04-21 17:05:10 +08:00
liaoxingyu 44cee30dfc update fastreid v1.2 
Summary:
1. refactor dataloader and heads
2. bugfix in fastattr, fastclas, fastface and partialreid
3. partial-fc supported in fastface
 2021-04-02 21:33:13 +08:00
Xingyu Liao 890224f25c
support classification in fastreid (#443) 
Summary: support classification and refactor build_dataloader which can support explicit parameters passing
 2021-03-26 20:17:39 +08:00
Xingyu Liao 15c556c43a
remove apex dependency (#442) 
Summary: Use Pytorch1.6(or above) built-in amp training
 2021-03-23 12:12:35 +08:00
liaoxingyu f57c5764e3 support multi-node training 2021-03-09 20:07:28 +08:00
liaoxingyu a53fd17874 update docs 2021-01-23 15:25:58 +08:00
liaoxingyu e26182e6ec make lr warmup by iter 
Summary: change warmup way by iter not by epoch, which will make it more flexible when training small epochs
 2021-01-22 11:17:21 +08:00
liaoxingyu 15e1729a27 update fastreid V1.0 2021-01-18 11:36:38 +08:00
liaoxingyu 2c17847980 feat: freeze FC 
Summary: update freeze FC in the last stages of training
 2020-12-28 14:46:28 +08:00
liaoxingyu 66941cf27a feat: support flip testing 2020-12-22 15:50:50 +08:00
liaoxingyu a327a70f0d v0.3 update 
Summary:
1. change DPP training in apex way;
2. make warmup scheduler by iter and lr scheduler by epoch;
3. replace random erasing with torchvision implementation;
4. naming modification in config file
 2020-12-07 14:19:20 +08:00
liaoxingyu 2724515fd9 save class number to config (#281) 
Summary: Save the class number calculated based on datasets to the config file. If you hard-code the class number, make it unchanged.
 2020-11-06 16:07:37 +08:00
liaoxingyu 7e9a4775da fixup finetune problem 
Summary: support finetune from the other model with different number of classes, and simplify calling way (#325)

close #325

close #325
 2020-11-06 15:58:22 +08:00
liaoxingyu 10cbaab155 support finetuning from trained models 
Summary: add a flag for supporting finetuning model from the trained weights, and it's very useful when performing across domain reid
 2020-09-28 17:10:10 +08:00
liaoxingyu 154a06b875 refactor code 2020-09-23 19:31:46 +08:00
liaoxingyu 4fa3f08a4a fix typro 
close #268
 2020-09-14 11:34:28 +08:00
liaoxingyu 4d573b8107 refactor reid head 
Summary: merge BNneckHead, LinearHead and ReductionHead into EmbeddingHead
because they are highly similar and can be prepared for ClsHead
 2020-09-10 10:57:37 +08:00
liaoxingyu 53fed7451d feat: support amp training 
Summary: support automatic mixed precision training #217
 2020-09-02 18:03:12 +08:00
liaoxingyu d00ce8fc3c refactor model arch 2020-09-01 16:14:45 +08:00
liaoxingyu ac8409a7da updating for pytorch1.6 2020-08-20 15:51:41 +08:00
liaoxingyu 2430b8ed75 pretrain model bugfix 
Fix pretrain model download bugs and testing bugs in multiprocess
 2020-07-31 10:42:38 +08:00
liaoxingyu 3b57dea49f support regnet backbone 2020-07-17 19:13:45 +08:00
liaoxingyu 3f35eb449d minor update 2020-07-14 11:58:06 +08:00
liaoxingyu ea8a3cc534 fix typro 2020-07-10 16:26:35 +08:00
liaoxingyu fec7abc461 finish v0.2 ddp training 2020-07-06 16:57:43 +08:00
liaoxingyu 36c04f0a9f fix resume training problem 
Summary: when resume training, need to reiter dataloader because we update pid_dict in dataset, but dataloader with multiprocess won't do the same update
 2020-05-30 16:44:18 +08:00
liaoxingyu 84c733fa85 fix: remove prefetcher, put normalizer in model 
1. remove messy data prefetcher which will cause  confusion
2. put normliazer in model to accelerate training via GPU computing
 2020-05-25 23:39:11 +08:00
liaoxingyu e990cf3e34 style: fix some typro 2020-05-21 15:55:51 +08:00
liaoxingyu d63bf5facc fix: add syncBN options in defaultTraine 2020-05-16 22:44:53 +08:00
liaoxingyu b28c0032e8 fix: add monkey-patching to enable syncBN 
add a trigger to make syncBN work
 2020-05-15 13:33:33 +08:00
liaoxingyu bf18479541 fix: revise syncBN bug 2020-05-14 14:52:37 +08:00
liaoxingyu 651e6ba9c4 feat: support multiprocess predictor 
add asyncpredictor to support multiprocessing feature extraction with dataloader
 2020-05-09 18:23:36 +08:00
liaoxingyu 4be4cacb73 fix: add a simple way to reset data prefetcher when resume training 
use data prefetcher build-in reset function to reload it rather than
redefining a new data prefetcher, otherwise it will introduce other
problems in eval-only mode.
 2020-05-09 11:58:27 +08:00
liaoxingyu 9fae467adf feat(engine/defaults): add DefaultPredictor to get image reid features 
Add a new predictor interface, and modify demo code to predict image features.
 2020-05-08 19:24:27 +08:00
liaoxingyu 0b15ac4e03 feat(hooks&optim): update stochastic weight averging hooks 
Update swa method which will do after regular training if you
set this option enabled.
 2020-05-08 12:20:04 +08:00
liaoxingyu afac8aad5d Fix(engine): fix preciseBN dataloader bugs 
preciseBN needs to pass data prefetcher, but now a DataLoader is passed
 2020-05-06 14:26:34 +08:00
liaoxingyu 6d96529d4c fix(data): fix resume training bug 
fix dataset pid dictionary loading bug when resume training,
data prefetcher will pre-load a batch of data, this will lead to
misalignment of old pid dict and updated pid dict.
We can address this problem by redefine a prefetcher in resume_or_load
 2020-05-05 23:20:42 +08:00
liaoxingyu a2dcd7b4ab feat(layers/norm): add ghost batchnorm 
add a get_norm fucntion to easily change normalization between batchnorm, ghost bn and group bn
 2020-05-01 09:02:46 +08:00
liaoxingyu d27729c5bb refactor(preciseBN): add preciseBN datasets show 2020-04-29 21:05:53 +08:00
liaoxingyu e38a799b63 fix(engine/defaults): fix precise bn bug 
fix problem in precise bn, which will not use precise bn datasets, and throw some errors
 2020-04-29 16:16:54 +08:00
liaoxingyu 4d2fa28dbb update freeze layer 
update preciseBN
update circle loss with metric learning and cross entropy loss form
update loss call methods
 2020-04-06 23:34:27 +08:00
liaoxingyu 6a8961ce48 1. upload circle loss and arcface 
2. finish freeze training
3. update augmix data augmentation
 2020-04-05 23:54:26 +08:00
liaoxingyu 23bedfce12 update version0.2 code 2020-03-25 10:58:26 +08:00
L1aoXingyu b020c7f0ae Fix data prefetcher minor bug 2020-02-27 12:16:57 +08:00
L1aoXingyu 12957f66aa Change architecture: 
1. delete redundant preprocess
2. add data prefetcher to accelerate data loading
3. fix minor bug of triplet sampler when only one image for one id
 2020-02-18 21:01:23 +08:00