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

502 Commits (7ed6240e2cb5e56e5ccd61744a3045f24ea7e62d)
 

Author SHA1 Message Date
liaoxingyu 7ed6240e2c fix ReidEvalution too much memory cost 
Move `matches` matrix computation in each iteration to reduce the extra memory cost

#420 #404
 2021-06-08 15:41:43 +08:00
liaoxingyu de81b3dbaa fix ClasEvalution too much memory cost 
Compute total number of correct predictions on each batch avoiding keeping all predicted logits, which will cost too much memory when the number of classes is large

#503
 2021-06-07 15:48:47 +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 2d2279be6a refactor freeze training condition 2021-06-02 15:55:35 +08:00
liaoxingyu 764fa67fe9 Align bagtricks config with v1.2 2021-06-01 11:51:22 +08:00
liaoxingyu 3c2eeb865d update readme 2021-05-31 18:33:48 +08:00
liaoxingyu 0572765085 fix for lint_python 2021-05-31 17:40:33 +08:00
liaoxingyu 54f96ba78a fix for lint_python 2021-05-31 17:36:56 +08:00
liaoxingyu 6300bd756e Bugfix for cls_layer 
In `any_softmax`, all operations are in-place, so pass into the `logits.clone()` to prevent outside logits changed.
 2021-05-31 17:32:24 +08:00
liaoxingyu c3ac4f504c Support amp and resume training in fastface 
AMP in partial-fc needs to be done only on backbone; In order to impl `resume training`, need to save & load different part of classifier weight in each GPU.
 2021-05-31 17:30:43 +08:00
liaoxingyu 91ff631184 Minor changes 
Some minor changes, such as class name changing, remove extra blank line, etc.
 2021-05-31 17:27:14 +08:00
liaoxingyu 8ab3554958 Support self-distill with EMA updated model 2021-05-31 17:17:24 +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 07b8251ccb Support gradient clip 
Follow detectron2's instruction and add gradient clip in step function of optimizer
 2021-05-31 17:11:37 +08:00
liaoxingyu 2cabc3428a Support vision transformer backbone 2021-05-31 17:08:57 +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 dbf1604231 fix model deploy problems 2021-05-25 15:55:43 +08:00
Sherlock Liao ff8a958fff
bugfix for `plain_train_net.py` and lr scheduler step (#484) 2021-05-11 15:46:17 +08:00
Xingyu Liao 46b0681313
fix randompatch (#470) 2021-04-21 17:08:16 +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
Xingyu Liao bb6ddbf8b1
bugfix for visualize and demo (#468) 
Summary minor bugs for visualize and demo caused by revised DataLoader
 2021-04-21 16:24:34 +08:00
Darren 37ccd3683d
unify gen_wts.py and inference.cpp dummy test value. (#457) 
Reviewed by: @L1aoXingyu
 2021-04-19 10:23:15 +08:00
Darren 8276ccf4fd
[v005] set INT8 calibrate set via cmake (#459) 
Reviewed by: @L1aoXingyu
 2021-04-19 10:22:48 +08:00
liaoxingyu e124a9afd3 fix lint_python 2021-04-12 15:08:39 +08:00
Darren e0ad8c70bc
update dockerfile in fastrt (#437) 
Reviewed by: @L1aoXingyu
 2021-04-12 15:06:13 +08:00
慕湮 fc67350e99
Add python interface by pybind11 and Int8 mode 
Reviewed by: @TCHeish
 2021-04-12 15:05:21 +08:00
Xingyu Liao 1dce15efad
faster dataloader with pre-fetch and cuda stream (#456) 
Summary: add a background thread to create a generator with pre-fetch, and create a new cuda stream to copy tensor from cpu to gpu in parallel.

Reviewed by: l1aoxingyu
 2021-04-12 15:03:35 +08:00
liaoxingyu 0da5917064 bugfix for ClasDataset 2021-04-08 11:01:15 +08:00
liaoxingyu 55300730e1 update fastreid v1.2 readme and changelog 2021-04-06 20:09:13 +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
liaoxingyu 9288db6303 add filp in fasttune config 2021-03-31 17:09:34 +08:00
Xingyu Liao fb36b23678
bugfix for attribute project (#450) 
Summary: refactor sample weight in attribute recognition;
change all options to False in defaults.py and modify yaml files
 2021-03-31 17:07:19 +08:00
Xie Jingyi 25cfa88fd9
Fix bug: COMBINEALL should add extra prefix to avoid pid conflictions (#447) 2021-03-30 15:49:52 +08:00
Xingyu Liao be0a089e1f
bugfix & merge classification transforms (#448) 
Summary: change heads definition in project and config file, merge classification transforms into default transforms
 2021-03-30 15:47:14 +08:00
Xingyu Liao 664ba4ae11
save idx2class dict when training classification (#444) 2021-03-26 20:51:52 +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
liaoxingyu 9d83550b67 bugfix 
Summary: make `deploy` adaptation for other models except repvgg
 2021-03-26 10:10:45 +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
Xingyu Liao 883fd4aede
add configurable decorator & linear loss decouple (#441) 
Summary: Add configurable decorator which can call `Baseline` with `Baseline(cfg)` or `Baseline(cfg, heads=heads, ...)`
Decouple linear and loss computation for partial-fc support.

Reviewed By: l1aoxingyu
 2021-03-23 12:10:06 +08:00
Xingyu Liao 41c3d6ff4d
add Dockerfile (#440) 2021-03-23 11:45:03 +08:00
Xingyu Liao 9b5af4166e
support repvgg (#429) 
Summary:
* support repvgg backbone, and verify the consistency of train mode and eval mode
* onnx export logger style modification
 2021-03-11 14:23:24 +08:00
Xingyu Liao cb7a1cb3e1
update deployment toolchain (#428) 
Summary:
Remove tiny-tensorrt dependency and rewrite a new tensorrt inference api.
In the new version of trt infer, it can pad the input to fixed batch automatically, so you don't need to worry about dynamic batch size.
 2021-03-10 16:48:59 +08:00
Darren d7c1294d9e
fastrt patch update 
Summary:
move div255 to gpu
add read/write numpy ndarray, which will make the comparison between torch and trt results more easily.

Reviewed By: l1aoxingyu
 2021-03-10 10:53:18 +08:00
Xingyu Liao 0cc9fb95a6
Merge pull request #425 from JDAI-CV/multi-node 
Summary: Add multiple machine training getting started docs.
Change multiple dataset evaluation logging mode, which will show the testing result of each dataset immediately.

Reviewed by: l1aoxingyu
 2021-03-09 20:13:29 +08:00
liaoxingyu f57c5764e3 support multi-node training 2021-03-09 20:07:28 +08:00
liaoxingyu 68c190b53c replace list in evaluator process with dict 2021-03-09 20:07:13 +08:00
Xingyu Liao 44ad4b83b1
update veriwild training config 
Summary: fix veriwild training config for reproducing.
Reviewed by: l1aoxingyu
 2021-03-04 15:20:39 +08:00
Xingyu Liao fcfa6800bb
Merge pull request #416 from TCHeish 
Reviewed by: l1aoxingyu
 2021-03-04 15:18:47 +08:00
Xinchen Liu 575aeaec3f
Update bagtricks_R50-ibn.yml 2021-03-03 12:07:39 +08:00
darrenhsieh b9bda486f0 add trt api div255 2021-02-27 16:45:29 +08:00