1+ # RegSeg segmentation training example with Cityscapes dataset.
2+ # Reproduction of paper: Rethink Dilated Convolution for Real-time Semantic Segmentation.
3+ #
4+ # Usage RegSeg48:
5+ # python -m torch.distributed.launch --nproc_per_node=4 src/super_gradients/examples/train_from_recipe_example/train_from_recipe.py --config-name=regseg48_cityscapes
6+ #
7+ #
8+ # Validation mIoU - Cityscapes, training time:
9+ # RegSeg48: input-size: [1024, 2048] mIoU: 78.15 using 4 GeForce RTX 2080 Ti with DDP, ~2 minutes / epoch
10+ #
11+ # Official git repo:
12+ # https://github.com/RolandGao/RegSeg
13+ # Paper:
14+ # https://arxiv.org/pdf/2111.09957.pdf
15+ #
16+ #
17+ # Logs, tensorboards and network checkpoints:
18+ # s3://deci-pretrained-models/regseg48_cityscapes/
19+ #
20+ #
21+ # Learning rate and batch size parameters, using 4 GeForce RTX 2080 Ti with DDP:
22+ # RegSeg48: input-size: [1024, 2048] initial_lr: 0.02 batch-size: 4 * 4gpus = 16
23+
24+ defaults :
25+ - training_hyperparams : default_train_params
26+ - dataset_params : cityscapes_dataset_params
27+
28+ hydra :
29+ searchpath :
30+ - pkg://super_gradients.recipes
31+
32+ project_name : RegSeg
33+ architecture : regseg48
34+ experiment_name : ${architecture}_cityscapes
35+ multi_gpu : AUTO
36+
37+ arch_params :
38+ num_classes : 19
39+ sync_bn : True
40+ strict_load : no_key_matching
41+
42+ dataset_params :
43+ _convert_ : all
44+ batch_size : 4
45+ val_batch_size : 4
46+ crop_size : 1024
47+ img_size : 1024
48+ random_scales :
49+ - 0.4
50+ - 1.6
51+ image_mask_transforms_aug :
52+ Compose :
53+ transforms :
54+ - ColorJitterSeg :
55+ brightness : 0.1
56+ contrast : 0.1
57+ saturation : 0.1
58+
59+ - RandomFlipSeg
60+
61+ - RandomRescaleSeg :
62+ scales : ${dataset_params.random_scales}
63+
64+ - PadShortToCropSizeSeg :
65+ crop_size : ${dataset_params.crop_size}
66+ fill_image :
67+ - ${dataset_params.cityscapes_ignored_label}
68+ - 0
69+ - 0
70+ fill_mask : ${dataset_params.cityscapes_ignored_label}
71+
72+ - CropImageAndMaskSeg :
73+ crop_size : ${dataset_params.crop_size}
74+ mode : random
75+
76+ image_mask_transforms :
77+ Compose :
78+ transforms : [ ]
79+
80+ dataset_interface :
81+ cityscapes :
82+ dataset_params : ${dataset_params}
83+
84+ data_loader_num_workers : 8
85+
86+ model_checkpoints_location : local
87+ load_checkpoint : False
88+
89+ training_hyperparams :
90+ max_epochs : 800
91+ lr_mode : poly
92+ initial_lr : 0.02 # for effective batch_size=16
93+ lr_warmup_epochs : 0
94+ optimizer : SGD
95+ optimizer_params :
96+ momentum : 0.9
97+ weight_decay : 5e-4
98+
99+ ema : True
100+
101+ loss : cross_entropy
102+ criterion_params :
103+ ignore_index : ${dataset_params.cityscapes_ignored_label}
104+
105+ train_metrics_list :
106+ - PixelAccuracy :
107+ ignore_label : ${dataset_params.cityscapes_ignored_label}
108+ - IoU :
109+ num_classes : 20
110+ ignore_index : ${dataset_params.cityscapes_ignored_label}
111+
112+ valid_metrics_list :
113+ - PixelAccuracy :
114+ ignore_label : ${dataset_params.cityscapes_ignored_label}
115+ - IoU :
116+ num_classes : 20
117+ ignore_index : ${dataset_params.cityscapes_ignored_label}
118+
119+ metric_to_watch : IoU
120+ greater_metric_to_watch_is_better : True
121+
122+ _convert_ : all
0 commit comments