mirrors
/
deit
mirror of https://github.com/facebookresearch/deit
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity
127 Commits
2 Branches
0 Tags
9.1 MiB
Commit Graph

21 Commits (e2c74433f9868cdb77dd6f4bf2634ddac65adc92)

Author SHA1 Message Date
Mehdi Yazdani e2c74433f9
Update main.py 
Adding the distributed parameter to the main function for the data loader to work as expected.
 2023-01-17 16:30:58 -05:00
Byeongho Heo 012a1007cc Add eval-crop-ratio to args 2022-09-22 11:27:03 +09:00
Hugo Touvron d64e7e927b
Update main.py 2022-07-29 08:17:08 +02:00
Zhaokai Wang d48dde3a4d
Fix img_size for DeiT-III 2022-06-24 01:59:42 +08:00
Hugo Touvron b9a8c27970
fix scaling lr 2022-06-14 10:27:37 +02:00
Hugo Touvron 4da66916c7
Add files via upload 2022-05-08 20:06:44 +02:00
Hugo Touvron 5d5fbbd753
Update main.py 2022-01-29 11:57:22 +01:00
Hugo Touvron f512394620
Merge pull request #118 from developer0hye/patch-1 
save best checkpoint
 2021-11-21 15:24:12 +01:00
Yonghye Kwon fdafaf0b8c
save best checkpoint 2021-10-01 17:12:25 +09:00
Maxwell Aladago 2f76ec64fd BUG Fix: ensure the right loss function is selected when mixtup or cutmix is enabeled 2021-05-02 10:27:25 -04:00
Maxwell Aladago 686507a6a5 BUG Fix: ensure the right loss function is selected when mixtup or cutmix is enabeled 2021-05-02 10:09:58 -04:00
Francisco Massa ab5715372d
Add deit_base_distilled_patch16_384 (#55) 2021-01-18 12:19:47 +01:00
Francisco Massa 1f961b060a
Fix no-weight-decay in optimizer (#52) 2021-01-16 11:53:29 +01:00
Francisco Massa adc7f1e849
Add loss_scaler to checkpoints (#49) 
Ensure bit-wise reproducibility when rescheduling jobs
 2021-01-15 10:52:41 +01:00
Francisco Massa a8e90967a3
Add option to finetune on larger resolution (#43) 
* Add option for finetuning a model

* Fixes

* Keep model in eval mode during finetuning

* Only skip head weights if size mismatch

* Remove finetune-epochs

Might not be needed

* Raise error if distillation + finetune are enabled
 2021-01-15 10:13:52 +01:00
Francisco Massa 8eae3269da
Add Knowledge-Distillation (#42) 
* Add knowledge distillation

* Bugfix

* Bugfix

* Make names more readable and use single torch.cat call

* Remove criterion.train() in engine

The teacher should stay in eval mode

* Change default argument for teacher-model

* Return the average of classifiers during inference

* Cleanup unused code

* Add docstring for DistillationLoss

* Remove warnings from newer PyTorch

Also uses more stable variant, instead of using softmax + log, use directly log_softmax
 2021-01-13 14:19:23 +01:00
Changlin Li b06fcba6bf
Support parallelized evaluation (#24) 
* support parallelized evaluation

* remove shuffle arg of loader val, add sampler val in non-dist branch

* replace timm eval sampler with torch sampler

* add logger synchronizing to support parallelized evaluation

* add command line argument dist-eval and warning
 2021-01-08 11:05:39 +01:00
Francisco Massa 0c4b8f60bd Change LICENSE to Apache 2.0 2021-01-08 10:51:58 +01:00
Francisco Massa dcd888df20
Update default data-path (#25) 
Points to the new valid location
 2021-01-07 22:18:54 +01:00
Zhiyuan Chen 0282d2a175
Remove drop block args (#23) 
timm's Vision Transformer does not support drop block
 2021-01-07 17:28:41 +01:00
Francisco Massa 1d38fa4c37 Initial commit 2020-12-23 10:47:58 -08:00