Commit Graph

1690 Commits (0f86c55576cd3d9144c377ebd69455d159344ea9)

Author SHA1 Message Date
gaotingquan 0f86c55576 add amp args, use_amp=False 2023-05-29 19:52:09 +08:00
gaotingquan 2d8346cd3b fix _init_amp when export 2023-05-29 19:52:09 +08:00
gaotingquan f67cfe2c2a fix ema: set_value() -> paddle.assign() 2023-05-26 15:40:48 +08:00
gaotingquan 2823e48be5 fix head_init_scale 2023-05-26 15:40:48 +08:00
gaotingquan 042d1e7ef8 fix layer key name for dynamic lr in adamwdl optimizer 2023-05-26 15:40:48 +08:00
gaotingquan 80ae9079cd add clip finetune config 2023-05-26 15:40:48 +08:00
gaotingquan 6d924f85ee fix for clip
1. fix bias_attr to False for conv of PatchEmbed;
2. support return_tokens_mean for Head of CLIP;
3. support remove_cls_token_in_forward for CLIP;
4. support head_init_scale argument for ViT backbone;
5. support get_num_layers() and no_weight_decay() for ViT backbone.
2023-05-26 15:40:48 +08:00
gaotingquan 74e6c8aa33 add fp32 and ampo2 ultra configs 2023-05-25 16:57:16 +08:00
gaotingquan f469dfe8d2 decrease bs 2023-05-25 16:57:16 +08:00
gaotingquan 53ac4675ad warmup 5 epochs 2023-05-25 16:57:16 +08:00
gaotingquan c2802b90aa increase bs, num_workers to speed up 2023-05-25 16:57:16 +08:00
gaotingquan b2ba6994a0 add ultra configs 2023-05-25 16:57:16 +08:00
gaotingquan 14d06fb6bd support AMP.use_amp arg 2023-05-25 16:16:02 +08:00
gaotingquan b0877289f4 disable promote kernel for amp training
compatible with paddle 2.5 and older version.
ref: https://github.com/PaddlePaddle/PaddleClas/pull/2798
2023-05-25 11:58:05 +08:00
gaotingquan 162f013ebe fix: minimize() dont support parameter_list of type dict
there are diffs that step()+update() and minimize().
this will be fixed in https://github.com/PaddlePaddle/Paddle/pull/53773.
2023-05-25 11:58:05 +08:00
gaotingquan 8b218b01ac refactor amp auto_cast context manager & loss scaler 2023-05-25 11:58:05 +08:00
gaotingquan f884f28853 refactor amp 2023-05-25 11:58:05 +08:00
gaotingquan b3678234fe fix bug when update_freq > iter_per_epoch 2023-05-17 15:19:13 +08:00
gaotingquan 377950865c getargspec -> getfullargspec
getargspec dont support param annotations
2023-05-17 15:19:13 +08:00
gaotingquan bb831c3baa code style 2023-05-17 15:19:13 +08:00
gaotingquan a3e9e99fa0 revert: fix bs 2023-05-17 15:19:13 +08:00
gaotingquan f42d6b6204 fix name: w24 -> W24 2023-05-17 15:19:13 +08:00
gaotingquan 07b9162bc0 fix pretrained url 2023-05-17 15:19:13 +08:00
gaotingquan a1fa19cd29 rename: v3 -> V3 2023-05-17 15:19:13 +08:00
gaotingquan 2091a59ff5 fix reference url 2023-05-17 15:19:13 +08:00
gaotingquan 890f77411a fix bs and unset update_freq to adapt to 8 gpus 2023-05-17 15:19:13 +08:00
gaotingquan fc9c59c4b1 update pretrained url 2023-05-17 15:19:13 +08:00
Yang Nie b66ee6384b fix RMSProp one_dim_param_no_weight_decay 2023-05-06 19:04:37 +08:00
Yang Nie c351dac67e add tinynet 2023-05-06 19:04:37 +08:00
zhangting2020 e7bef51f9e fix data dtype for amp training 2023-04-26 18:40:20 +08:00
kangguangli 731006f1fc set seed by configs 2023-04-25 17:39:55 +08:00
kangguangli 293a216a0b fix random seed 2023-04-25 17:39:55 +08:00
zh-hike d7bd275379 update foundation_vit from EVA_vit_huge to EVA_vit_giant 2023-04-23 10:16:08 +08:00
Yang Nie 0af4680f86 set num_workers 8 2023-04-19 21:21:06 +08:00
Yang Nie cdd3c3a05c clear type hint 2023-04-19 21:21:06 +08:00
Yang Nie 692204eee6 fix code style 2023-04-19 21:21:06 +08:00
Yang Nie e7ad3909c8 update configs for 8gpus 2023-04-19 21:21:06 +08:00
Yang Nie deb8e98779 rename v2 to V2 2023-04-19 21:21:06 +08:00
Yang Nie be6a22be18 add MobileViTv2 2023-04-19 21:21:06 +08:00
gaotingquan 9f621279b8 fix infer output 2023-04-17 20:28:40 +08:00
gaotingquan 73f4d8e4ce to avoid cause issues for unset no_weight_decay models.
there seems be a diff for optimizer about using [] and [{"params":}, {"params":}] params
2023-04-12 20:55:38 +08:00
gaotingquan 31ea33c884 revert the cutmix, mixup, fmix fixes
because this change(commit: df31d808fc) will cause other issues, such as a change in the value of QA monitoring, so revert temporary.
2023-04-12 20:55:38 +08:00
parap1uie-s 52f16cc85d Update engine.py 2023-04-11 19:23:57 +08:00
parap1uie-s 6e6586f59b Fixed the incorrect infer outputs 2023-04-11 19:23:57 +08:00
Yang Nie f36ffbc492 fix 2023-04-10 15:02:54 +08:00
Yang Nie a69bc945bf modified batch_size and update_freq & add more tipc_test configs 2023-04-06 15:33:30 +08:00
Yang Nie e135e2cd37 modified batch_size and update_freq
modified batch_size per gpu and update_freq in MobileViTv3_S.yaml for training with 4 gpus
2023-04-06 15:33:30 +08:00
Yang Nie b8a1589377 update data augment and init method for MobileViTv3-v2 2023-04-06 15:33:30 +08:00
Yang Nie c32e2b098a Revert "Speedup EMA"
This reverts commit 35fc732dadac4761852b18512b5c5df8785e36df.
2023-04-06 15:33:30 +08:00
Yang Nie 001cdb0955 update MobileViTv3-v2 configs 2023-04-06 15:33:30 +08:00