AmbiakaTT commented 1 year ago

My DINO config is as follows from detrex.config import get_config from .models.dino_r50 import model

get default config

dataloader = get_config("common/data/coco_detr.py").dataloader optimizer = get_config("common/optim.py").AdamW lr_multiplier = get_config("common/coco_schedule.py").lr_multiplier_12ep train = get_config("common/train.py").train

modify training config

train.init_checkpoint = "detectron2://ImageNetPretrained/torchvision/R-50.pkl" train.output_dir = "./output/dino_r50_4scale_12ep"

max training iterations

train.max_iter = 90000

run evaluation every 5000 iters

train.eval_period = 5000

log training infomation every 20 iters

train.log_period = 20

save checkpoint every 5000 iters

train.checkpointer.period = 5000

gradient clipping for training

train.clip_grad.enabled = True train.clip_grad.params.max_norm = 0.1 train.clip_grad.params.norm_type = 2

set training devices

train.device = "cuda" model.device = train.device

modify optimizer config

optimizer.lr = 1e-4 optimizer.betas = (0.9, 0.999) optimizer.weight_decay = 1e-4 optimizer.params.lr_factor_func = lambda module_name: 0.1 if "backbone" in module_name else 1

modify dataloader config

dataloader.train.num_workers = 16

please notice that this is total batch size.

surpose you're using 4 gpus for training and the batch size for

each gpu is 16/4 = 4

dataloader.train.total_batch_size = 16

dump the testing results into output_dir for visualization

dataloader.evaluator.output_dir = train.output_dir

The training error is very low [04/17 06:35:07] d2.utils.events INFO: eta: 0:00:00 iter: 19999 total_loss: 13.36 loss_class: 0.3083 loss_bbox: 0.09793 loss_giou: 0.5997 loss_class_0: 0.4523 loss_bbox_0: 0.08994 loss_giou_0: 0.5179 loss_class_1: 0.3937 loss_bbox_1: 0.0916 loss_giou_1: 0.5637 loss_class_2: 0.3323 loss_bbox_2: 0.08917 loss_giou_2: 0.6124 loss_class_3: 0.3066 loss_bbox_3: 0.09503 loss_giou_3: 0.583 loss_class_4: 0.3075 loss_bbox_4: 0.1018 loss_giou_4: 0.6114 loss_class_enc: 0.49 loss_bbox_enc: 0.08078 loss_giou_enc: 0.5312

But the accuracy is like 0 on the test set [04/17 06:48:59] d2.evaluation.coco_evaluation INFO: Evaluation results for bbox:	AP	AP50	AP75	APs	APm	APl
0.137	0.694	0.014	0.097	0.203	0.146

What could be the reason for this?

rentainhe commented 1 year ago

Hello! how many GPUs you're using for running this experiments