loop optimization + pytorch (028-xxx)

[yzbx]

focal_loss028/029

• gamma = 0.5 train failed
• gamma = 1.0 miou=0.377 (for upsamle_layer=3) 0.386 (for upsample_layer=5)
• gamma = 2.0 is slight better than gamma=1.0

python test/pspnet_test.py --backbone_pretrained=True --backbone_name=resnet50 \
--upsample_type=bilinear --midnet_scale=15 --batch_size=4 --note=focal_loss028 \
--test=hyperopt --hyperkey=model.focal_loss_gamma --hyperopt_calls=6 \
--use_momentum=False --upsample_layer=3 --dataset_use_part=320

python test/pspnet_test.py --backbone_pretrained=True --backbone_name=resnet50 \
--upsample_type=bilinear --midnet_scale=15 --batch_size=2 --note=focal_loss029 \
--test=hyperopt --hyperkey=model.focal_loss_gamma --hyperopt_calls=6 \
--use_momentum=True --upsample_layer=5 --dataset_use_part=320

	train/iou	val/iou	n_epoch
('focal_loss028', 0.5)	0.01994	0.0192171	100
('focal_loss028', 1.0)	0.570953	0.377005	100
('focal_loss028', 2.0)	0.556718	0.378915	100

	('val/iou', 'mean')	('val/iou', 'std')	('val/iou', 'amax')
('focal_loss028', 0.5)	0.0192171	0	0.0192171
('focal_loss028', 1.0)	0.377005	0.00549323	0.38089
('focal_loss028', 2.0)	0.378915	0.0059462	0.38312

	train/iou	val/iou	note	focal_loss_gamma	n_epoch
2	0.0191035	0.0192171	focal_loss028	0.5	100
5	0.0207766	0.0192171	focal_loss028	0.5	100
4	0.575547	0.373121	focal_loss028	1	100
0	0.562847	0.374711	focal_loss028	2	100
3	0.566359	0.38089	focal_loss028	1	100
1	0.550588	0.38312	focal_loss028	2	100

	train/iou	val/iou	n_epoch
('focal_loss029', 0.5)	0.0191787	0.0192171	100
('focal_loss029', 1.0)	0.635259	0.386365	100
('focal_loss029', 2.0)	0.624699	0.403327	100

	('val/iou', 'mean')	('val/iou', 'std')	('val/iou', 'amax')
('focal_loss029', 0.5)	0.0192171	0	0.0192171
('focal_loss029', 1.0)	0.386365	0.00394113	0.389151
('focal_loss029', 2.0)	0.403327	0.00978559	0.410246

	train/iou	val/iou	note	focal_loss_gamma	n_epoch
0	0.0188437	0.0192171	focal_loss029	0.5	100
2	0.0195137	0.0192171	focal_loss029	0.5	100
1	0.629185	0.383578	focal_loss029	1	100
5	0.641334	0.389151	focal_loss029	1	100
3	0.616328	0.396407	focal_loss029	2	100
4	0.63307	0.410246	focal_loss029	2	100

[yzbx]

norm_ways030

for times in 1 2 3
do
    for norm_ways in caffe cityscapes -1,1 0,1
    do
        python test/pspnet_test.py --batch_size=2 \
        --backbone_pretrained=True --midnet_scale=15 \
        --upsample_type=bilinear --dataset_use_part=320 \
        --note=norm_ways030 --norm_ways=${norm_ways}
    done
done

	train/iou	val/iou	n_epoch
('norm_ways030', 'pytorch')	0.579341	0.360122	100
('norm_ways030', 'cityscapes')	0.582311	0.364104	100
('norm_ways030', 'caffe')	0.579287	0.367253	100
('norm_ways030', '-1,1')	0.582054	0.368425	100
('norm_ways030', '0,1')	0.58766	0.369743	100

	('val/iou', 'mean')	('val/iou', 'std')	('val/iou', 'amax')
('norm_ways030', '-1,1')	0.368425	0.00802664	0.377687
('norm_ways030', '0,1')	0.369743	0.00674804	0.377483
('norm_ways030', 'caffe')	0.367253	0.00350519	0.370689
('norm_ways030', 'cityscapes')	0.364104	0.00219609	0.36635
('norm_ways030', 'pytorch')	0.360122	0.00480617	0.366818

	train/iou	val/iou	note	norm_ways	n_epoch
1	0.582042	0.350914	norm_ways030	pytorch	100
9	0.585439	0.351199	norm_ways030	pytorch	100
8	0.567812	0.354869	norm_ways030	pytorch	100
4	0.573669	0.355681	norm_ways030	pytorch	100
6	0.591837	0.359229	norm_ways030	pytorch	100
16	0.565224	0.359656	norm_ways030	pytorch	100
2	0.573616	0.359778	norm_ways030	pytorch	100
13	0.590471	0.360076	norm_ways030	pytorch	100
12	0.582143	0.360196	norm_ways030	pytorch	100
14	0.573794	0.360642	norm_ways030	pytorch	100
0	0.568066	0.36083	norm_ways030	pytorch	100
7	0.59699	0.361248	norm_ways030	pytorch	100
26	0.576288	0.361961	norm_ways030	cityscapes	100
23	0.580364	0.363498	norm_ways030	-1,1	100
28	0.574199	0.363682	norm_ways030	caffe	100
5	0.590236	0.363961	norm_ways030	pytorch	100
19	0.586602	0.364003	norm_ways030	cityscapes	100
27	0.593775	0.36409	norm_ways030	-1,1	100
15	0.585286	0.364297	norm_ways030	pytorch	100
20	0.580456	0.365093	norm_ways030	0,1	100
3	0.573443	0.365889	norm_ways030	pytorch	100
22	0.584044	0.36635	norm_ways030	cityscapes	100
25	0.590817	0.366654	norm_ways030	0,1	100
10	0.587569	0.366787	norm_ways030	pytorch	100
11	0.561157	0.366818	norm_ways030	pytorch	100
18	0.578464	0.367388	norm_ways030	caffe	100
21	0.585199	0.370689	norm_ways030	caffe	100
24	0.591707	0.377483	norm_ways030	0,1	100
17	0.572022	0.377687	norm_ways030	-1,1	100

[dishen12]

fl_mult031 (deeplab2)

note new_lr_mult=5

• [x] fl_mult031 best miou=0.27, best change_lr_mult=5

  python test/pspnet_test.py --test=hyperopt --n_epoch=50 \
  --note=fl_mult031 --upsample_type=bilinear \
  --backbone_pretrained=True --midnet_scale=5 \
  --dataset_use_part=320 --focal_loss_gamma=2.0 \
  --changed_lr_mult=2 --new_lr_mult=5 --use_momentum=True \
  --upsample_layer=4 --batch_size=16 \
  --hyperkey=model.changed_lr_mult --hyperopt_calls=8

	train/iou	val/iou	n_epoch
('fl_mult031', 1)	0.45368	0.246161	50
('fl_mult031', 2)	0.498987	0.260186	50
('fl_mult031', 10)	0.489595	0.265164	50
('fl_mult031', 5)	0.506165	0.268122	50

	('val/iou', 'mean')	('val/iou', 'std')	('val/iou', 'amax')
('fl_mult031', 1)	0.246161	0.00065943	0.246627
('fl_mult031', 2)	0.260186	0.000597137	0.260609
('fl_mult031', 5)	0.268122	0.00737557	0.273337
('fl_mult031', 10)	0.265164	0.0109059	0.272875

	train/iou	val/iou	note	changed_lr_mult	n_epoch
0	0.45759	0.245695	fl_mult031	1	50
7	0.44977	0.246627	fl_mult031	1	50
4	0.483519	0.257452	fl_mult031	10	50
3	0.509061	0.259764	fl_mult031	2	50
5	0.488913	0.260609	fl_mult031	2	50
1	0.507072	0.262906	fl_mult031	5	50
6	0.495671	0.272875	fl_mult031	10	50
2	0.505259	0.273337	fl_mult031	5	50

[yzbx]

fl_alpha032 (deeplab2 118) :x:

focal loss alpha in [1.0, 5.0 ,10.0]

1. change loss by multiple constant is nonsense ? :x: L_new=alpha L_old then grad_new = alpha grad_old
2. for each parameters, the range of grad_old may be different, so for alpha>1 will benefit some parameter with small grad_old, but for other parameters with big grad_old will cause error! :question:

• [x] fl_alpha032

  python test/pspnet_test.py --test=hyperopt --n_epoch=50 \
  --note=fl_alpha032 --upsample_type=bilinear \
  --backbone_pretrained=True --midnet_scale=15 \
  --dataset_use_part=320 --focal_loss_gamma=2.0 \
  --changed_lr_mult=2 --new_lr_mult=10 --use_momentum=False \
  --upsample_layer=3 --batch_size=6 \
  --hyperkey=model.focal_loss_alpha --hyperopt_calls=6

	train/iou	val/iou	n_epoch
('fl_alpha032', 1.0)	0.515991	0.356931	50
('fl_alpha032', 10.0)	0.557402	0.361924	50
('fl_alpha032', 5.0)	0.535622	0.362146	50

	('val/iou', 'mean')	('val/iou', 'std')	('val/iou', 'amax')
('fl_alpha032', 1.0)	0.356931	0.00116684	0.357756
('fl_alpha032', 5.0)	0.362146	0.00501773	0.365695
('fl_alpha032', 10.0)	0.361924	0.0101221	0.369082

	train/iou	val/iou	note	focal_loss_alpha	n_epoch
1	0.560917	0.354767	fl_alpha032	10	50
3	0.513267	0.356106	fl_alpha032	1	50
2	0.518716	0.357756	fl_alpha032	1	50
4	0.532431	0.358598	fl_alpha032	5	50
5	0.538813	0.365695	fl_alpha032	5	50
0	0.553887	0.369082	fl_alpha032	10	50

[yzbx]

cw_alpha033 (deeplab3)

class weight alpha in [0.2,0.4,0.6,0.8]

• [x] cw_alpha

  python test/pspnet_test.py --test=hyperopt --n_epoch=50 \
  --note=cw_alpha033 --upsample_type=bilinear \
  --backbone_pretrained=True --midnet_scale=15 \
  --dataset_use_part=320 --use_class_weight=True \
  --changed_lr_mult=2 --new_lr_mult=10 --use_momentum=False \
  --upsample_layer=3 --batch_size=6 \
  --hyperkey=model.class_weight_alpha --hyperopt_calls=8

	train/iou	val/iou	n_epoch
('cw_alpha033', 0.8)	0.552976	0.361493	50
('cw_alpha033', 0.6)	0.5612	0.366854	50
('cw_alpha033', 0.4)	0.579964	0.375186	50
('cw_alpha033', 0.2)	0.577416	0.381513	50

	('val/iou', 'mean')	('val/iou', 'std')	('val/iou', 'amax')
('cw_alpha033', 0.2)	0.381513	0.00686519	0.386367
('cw_alpha033', 0.4)	0.375186	0.00442036	0.378311
('cw_alpha033', 0.6)	0.366854	0.0085524	0.372901
('cw_alpha033', 0.8)	0.361493	0.0107269	0.369078

	train/iou	val/iou	note	class_weight_alpha	n_epoch
3	0.553801	0.353908	cw_alpha033	0.8	50
1	0.561204	0.360806	cw_alpha033	0.6	50
4	0.55215	0.369078	cw_alpha033	0.8	50
7	0.57911	0.37206	cw_alpha033	0.4	50
2	0.561197	0.372901	cw_alpha033	0.6	50
6	0.57455	0.376659	cw_alpha033	0.2	50
5	0.580818	0.378311	cw_alpha033	0.4	50
0	0.580282	0.386367	cw_alpha033	0.2	50

[yzbx]

batch size (dl)

for batch_size in 8 16 32
do
    python test/pspnet_test.py --batch_size=${batch_size} \
    --backbone_pretrained=True --midnet_scale=5 \
    --backbone_freeze=False --backbone_name=vgg16_bn \
    --upsample_type=bilinear --dataset_use_part=320 \
    --note=bs${batch_size}
done

	train/iou	val/iou	n_epoch
('bs32', 32)	0.564952	0.319975	100
('bs16', 16)	0.594788	0.329257	100
('bs8', 8)	0.607575	0.329859	100

batch size and momentum for batch normaliation

note caffe momentum = 1 - pytorch momentum small batch size :arrow_right: high momentum (0.9-0.99) big batch size :arrow_right: low momentum (0.6-0.85)

[yzbx]

use_dropout034 (dl)

config->get_midnet->os.environ->model

conclusion

use_dropout = False

code

python test/pspnet_test.py --use_dropout False --use_lr_mult False --dataset_use_part=320 \
--backbone_pretrained=True --midnet_scale=15 --n_epoch=50 --upsample_type=bilinear \
--hyperopt_calls=6 --hyperkey=model.use_dropout --note=use_dropout034 --test=hyperopt

experiment results

	train/iou	val/iou	n_epoch
('use_dropout034', True)	0.516743	0.302228	50
('use_dropout034', False)	0.553649	0.325748	50

	('val/iou', 'mean')	('val/iou', 'std')	('val/iou', 'amax')
('use_dropout034', False)	0.325748	0.0171578	0.34556
('use_dropout034', True)	0.302228	0.00658758	0.306819

	train/iou	val/iou	note	use_dropout	n_epoch
0	0.518732	0.29468	use_dropout034	True	50
2	0.514725	0.305184	use_dropout034	True	50
3	0.516772	0.306819	use_dropout034	True	50
4	0.558763	0.315801	use_dropout034	False	50
5	0.550713	0.315883	use_dropout034	False	50
1	0.551469	0.34556	use_dropout034	False	50

[yzbx]

use_bias035 (dl)

conclusion

use_bias = True

code

python test/pspnet_test.py --use_dropout False --use_lr_mult False --dataset_use_part=320 \
--backbone_pretrained=True --midnet_scale=15 --n_epoch=50 --upsample_type=bilinear \
--hyperopt_calls=6 --hyperkey=model.use_bias --note=use_bias035 --test=hyperopt

experiment results

	train/iou	val/iou	n_epoch
('use_bias035', False)	0.549073	0.316397	50
('use_bias035', True)	0.553332	0.321291	50

	('val/iou', 'mean')	('val/iou', 'std')	('val/iou', 'amax')
('use_bias035', False)	0.316397	0.00797418	0.324374
('use_bias035', True)	0.321291	0.00346007	0.325232

	train/iou	val/iou	note	use_bias	n_epoch
1	0.555032	0.308425	use_bias035	False	50
2	0.548912	0.316391	use_bias035	False	50
4	0.551562	0.318756	use_bias035	True	50
0	0.553671	0.319884	use_bias035	True	50
5	0.543275	0.324374	use_bias035	False	50
3	0.554763	0.325232	use_bias035	True	50

[dishen12]

moment_bias036 (deeplab2 118)

momentum in [0.1, 0.05, 0.01] :x: use_momentum=False in config bias in [True False] batch size in [6,4]

conclusion

best params is** model.momentum 0.01/0.05 model.use_bias True val_miou 0.3657269577250924 / 0.372 best score is 0.366**

python test/pspnet_test.py --use_dropout False --use_lr_mult False --dataset_use_part=320 \
--backbone_pretrained=True --midnet_scale=15 --n_epoch=50 --upsample_type=bilinear \
--hyperopt_calls=12 --hyperkey=model.momentum,model.use_bias --note=moment_bias036 \
--test=hyperopt --batch_size=6

python test/pspnet_test.py --use_dropout False --use_lr_mult False --dataset_use_part=320 \
--backbone_pretrained=True --midnet_scale=15 --n_epoch=50 --upsample_type=bilinear \
--hyperopt_calls=12 --hyperkey=model.momentum,model.use_bias --note=moment_bias036 \
--test=hyperopt --batch_size=4

	train/iou	val/iou	n_epoch
('moment_bias036', 4, False, 0.01)	0.53504	0.341246	50
('moment_bias036', 4, False, 0.05)	0.535422	0.348988	50
('moment_bias036', 6, False, 0.1)	0.542413	0.3517	50
('moment_bias036', 4, False, 0.1)	0.544612	0.351922	50
('moment_bias036', 4, True, 0.05)	0.536309	0.353263	50
('moment_bias036', 6, False, 0.01)	0.541852	0.353327	50
('moment_bias036', 4, True, 0.1)	0.541139	0.355432	50
('moment_bias036', 6, True, 0.1)	0.545223	0.357838	50
('moment_bias036', 6, True, 0.01)	0.540284	0.357885	50
('moment_bias036', 4, True, 0.01)	0.536912	0.358324	50
('moment_bias036', 6, False, 0.05)	0.554138	0.360551	50
('moment_bias036', 6, True, 0.05)	0.540135	0.366356	50

	('val/iou', 'mean')	('val/iou', 'std')	('val/iou', 'amax')
('moment_bias036', 4, False, 0.01)	0.341246	0.000427727	0.341549
('moment_bias036', 4, False, 0.05)	0.348988	0.00940163	0.355636
('moment_bias036', 4, False, 0.1)	0.351922	0.00593493	0.356119
('moment_bias036', 4, True, 0.01)	0.358324	0.01047	0.365727
('moment_bias036', 4, True, 0.05)	0.353263	0.00611485	0.357587
('moment_bias036', 4, True, 0.1)	0.355432	0.00100109	0.35614
('moment_bias036', 6, False, 0.01)	0.353327	0.0100703	0.360448
('moment_bias036', 6, False, 0.05)	0.360551	0.00555076	0.364476
('moment_bias036', 6, False, 0.1)	0.3517	0.00355831	0.354216
('moment_bias036', 6, True, 0.01)	0.357885	0.00590222	0.362058
('moment_bias036', 6, True, 0.05)	0.366356	0.00855539	0.372405
('moment_bias036', 6, True, 0.1)	0.357838	0.00217849	0.359378

	train/iou	val/iou	note	batch_size	use_bias	momentum	n_epoch
21	0.539802	0.340944	moment_bias036	4	False	0.01	50
13	0.530279	0.341549	moment_bias036	4	False	0.01	50
10	0.528492	0.34234	moment_bias036	4	False	0.05	50
7	0.539671	0.346207	moment_bias036	6	False	0.01	50
6	0.562788	0.347726	moment_bias036	4	False	0.1	50
8	0.53638	0.348939	moment_bias036	4	True	0.05	50
15	0.540547	0.349184	moment_bias036	6	False	0.1	50
17	0.539323	0.35092	moment_bias036	4	True	0.01	50
23	0.542467	0.353711	moment_bias036	6	True	0.01	50
2	0.544279	0.354216	moment_bias036	6	False	0.1	50
9	0.533409	0.354724	moment_bias036	4	True	0.1	50
1	0.542351	0.355636	moment_bias036	4	False	0.05	50
19	0.526435	0.356119	moment_bias036	4	False	0.1	50
16	0.54887	0.35614	moment_bias036	4	True	0.1	50
3	0.540699	0.356297	moment_bias036	6	True	0.1	50
11	0.554961	0.356626	moment_bias036	6	False	0.05	50
14	0.536238	0.357587	moment_bias036	4	True	0.05	50
4	0.549747	0.359378	moment_bias036	6	True	0.1	50
22	0.535799	0.360306	moment_bias036	6	True	0.05	50
5	0.544033	0.360448	moment_bias036	6	False	0.01	50
0	0.538102	0.362058	moment_bias036	6	True	0.01	50
18	0.553315	0.364476	moment_bias036	6	False	0.05	50
12	0.5345	0.365727	moment_bias036	4	True	0.01	50
20	0.544472	0.372405	moment_bias036	6	True	0.05	50

[dishen12]

lr_mult037 (deeplab3 237)

'model.changed_lr_mult':('choices',[1,2,5]), 'model.new_lr_mult':('choices',[1,5,10]),

python test/pspnet_test.py --use_dropout False --use_lr_mult True --dataset_use_part=320 \
--backbone_pretrained=True --midnet_scale=10 --use_momentum=True --upsample_layer=4 \
 --n_epoch=50 --upsample_type=bilinear \
--hyperopt_calls=27 --hyperkey=model.changed_lr_mult,model.new_lr_mult --note=lr_mult037 \
--test=hyperopt --batch_size=16

	train/iou	val/iou	n_epoch
('lr_mult037', 1, 1)	0.464864	0.31209	50
('lr_mult037', 2, 1)	0.614842	0.320799	50
('lr_mult037', 1, 5)	0.623889	0.325164	50
('lr_mult037', 1, 10)	0.58821	0.336011	50
('lr_mult037', 5, 10)	0.572909	0.337367	50
('lr_mult037', 5, 1)	0.629137	0.338745	50
('lr_mult037', 2, 10)	0.623068	0.340549	50
('lr_mult037', 5, 5)	0.648084	0.343118	50
('lr_mult037', 2, 5)	0.644184	0.344695	50

	('val/iou', 'mean')	('val/iou', 'std')	('val/iou', 'amax')
('lr_mult037', 1, 1)	0.31209	0.00835003	0.320658
('lr_mult037', 1, 5)	0.325164	0.0039992	0.329778
('lr_mult037', 1, 10)	0.336011	0.0109767	0.348104
('lr_mult037', 2, 1)	0.320799	0.0107251	0.331006
('lr_mult037', 2, 5)	0.344695	0.00553152	0.350432
('lr_mult037', 2, 10)	0.340549	0.00108809	0.341494
('lr_mult037', 5, 1)	0.338745	0.00387832	0.342901
('lr_mult037', 5, 5)	0.343118	0.0102245	0.354641
('lr_mult037', 5, 10)	0.337367	0.0264767	0.356089

	train/iou	val/iou	note	changed_lr_mult	new_lr_mult	n_epoch
23	0.571152	0.303976	lr_mult037	1	1	50
11	0.609806	0.309621	lr_mult037	2	1	50
10	0.579928	0.311635	lr_mult037	1	1	50
14	0.524534	0.318645	lr_mult037	5	10	50
7	0.588605	0.320658	lr_mult037	1	1	50
1	0.615085	0.321768	lr_mult037	2	1	50
24	0.612793	0.322681	lr_mult037	1	5	50
2	0.635219	0.323034	lr_mult037	1	5	50
3	0.590388	0.326677	lr_mult037	1	10	50
15	0.623654	0.329778	lr_mult037	1	5	50
12	0.619635	0.331006	lr_mult037	2	1	50
0	0.588174	0.333253	lr_mult037	1	10	50
16	0.654349	0.33513	lr_mult037	5	5	50
22	0.622382	0.335222	lr_mult037	5	1	50
18	0.637123	0.338112	lr_mult037	5	1	50
21	0.629705	0.33936	lr_mult037	2	10	50
5	0.621339	0.339395	lr_mult037	2	5	50
17	0.641755	0.339584	lr_mult037	5	5	50
26	0.621337	0.340795	lr_mult037	2	10	50
6	0.618163	0.341494	lr_mult037	2	10	50
13	0.627907	0.342901	lr_mult037	5	1	50
8	0.654889	0.344258	lr_mult037	2	5	50
19	0.586069	0.348104	lr_mult037	1	10	50
25	0.656323	0.350432	lr_mult037	2	5	50
9	0.648148	0.354641	lr_mult037	5	5	50
20	0.621285	0.356089	lr_mult037	5	10	50
4	0.119772	nan	lr_mult037	1	1	50

[yzbx]

optimizer038 (dl)

adam series optimizer benchmark optimizer in [adam, admax, adam+amsgrad] learning rate in [1e-4, 1e-3]

python test/pspnet_test.py --test=hyperopt --use_lr_mult=False --midnet_scale=10 \
--batch_size=4 --learning_rate=1e-4 --hyperopt=loop --hyperkey=model.optimizer \
--hyperopt_calls=9 --momentum=0.01 --note=optimizer038

	train/iou	val/iou	n_epoch
('optimizer038', 0.001, 'adam')	0.568786	0.325539	100
('optimizer038', 0.0001, 'adamax')	0.579982	0.336745	100
('optimizer038', 0.0001, 'amsgrad')	0.638081	0.351792	100
('optimizer038', 0.001, 'amsgrad')	0.678134	0.360234	100
('optimizer038', 0.001, 'adamax')	0.678845	0.36052	100
('optimizer038', 0.0001, 'adam')	0.657064	0.360583	100

	('val/iou', 'mean')	('val/iou', 'std')	('val/iou', 'amax')
('optimizer038', 0.0001, 'adam')	0.360583	0.00617481	0.365915
('optimizer038', 0.0001, 'adamax')	0.336745	0.00285825	0.338813
('optimizer038', 0.0001, 'amsgrad')	0.351792	0.00209968	0.354104
('optimizer038', 0.001, 'adam')	0.325539	0.0112959	0.338499
('optimizer038', 0.001, 'adamax')	0.36052	0.00453578	0.36381
('optimizer038', 0.001, 'amsgrad')	0.360234	0.0102075	0.371674

	train/iou	val/iou	note	learning_rate	optimizer	n_epoch
5	0.569034	0.317782	optimizer038	0.001	adam	100
17	0.565804	0.320336	optimizer038	0.001	adam	100
11	0.590885	0.333483	optimizer038	0.0001	adamax	100
16	0.574335	0.337938	optimizer038	0.0001	adamax	100
14	0.571521	0.338499	optimizer038	0.001	adam	100
0	0.574727	0.338813	optimizer038	0.0001	adamax	100
9	0.633256	0.350004	optimizer038	0.0001	amsgrad	100
15	0.634138	0.35127	optimizer038	0.0001	amsgrad	100
1	0.685483	0.352057	optimizer038	0.001	amsgrad	100
2	0.646411	0.353818	optimizer038	0.0001	adam	100
3	0.64685	0.354104	optimizer038	0.0001	amsgrad	100
10	0.688386	0.355345	optimizer038	0.001	adamax	100
4	0.696042	0.35697	optimizer038	0.001	amsgrad	100
8	0.653214	0.362017	optimizer038	0.0001	adam	100
7	0.674424	0.362403	optimizer038	0.001	adamax	100
13	0.673724	0.36381	optimizer038	0.001	adamax	100
6	0.671568	0.365915	optimizer038	0.0001	adam	100
12	0.652877	0.371674	optimizer038	0.001	amsgrad	100

[yzbx]

upsample_layer039 vs upsample_layer023

backbone_name resnet50 vs vgg16 midnet_scale 10 vs 15 dataset_use_part 640 vs 320 use_lr_mult=True

conclusion

• best val miou=0.47 for upsample_layer=4
• [x] running in deeplab2

  python test/pspnet_test.py --test=hyperopt --use_lr_mult=True --changed_lr_mult=2 \
  --new_lr_mult=10 --midnet_scale=10 --dataset_use_part=640 \
  --batch_size=4 --learning_rate=1e-4 --hyperopt=loop --hyperkey=model.upsample_layer \
  --hyperopt_calls=6 --use_momentum=True --momentum=0.01 --note=upsample_layer039 

	train/iou	val/iou	n_epoch
('upsample_layer039', 3)	0.627103	0.371236	100
('upsample_layer039', 5)	0.793217	0.462083	100
('upsample_layer039', 4)	0.765913	0.470752	100

	('val/iou', 'mean')	('val/iou', 'std')	('val/iou', 'amax')
('upsample_layer039', 3)	0.371236	0.0110116	0.379022
('upsample_layer039', 4)	0.470752	0.00293595	0.472829
('upsample_layer039', 5)	0.462083	0.00485348	0.465515

	train/iou	val/iou	note	upsample_layer	n_epoch
4	0.620665	0.36345	upsample_layer039	3	100
5	0.633541	0.379022	upsample_layer039	3	100
2	0.785606	0.458651	upsample_layer039	5	100
0	0.800828	0.465515	upsample_layer039	5	100
3	0.765821	0.468676	upsample_layer039	4	100
1	0.766004	0.472829	upsample_layer039	4	100

[yzbx]

batch_size040 (dl)

• [x] running

  python test/pspnet_test.py --test=hyperopt --use_lr_mult=True --changed_lr_mult=2 \
  --new_lr_mult=10 --midnet_scale=5 --upsample_layer=4 --dataset_use_part=320 \
  --batch_size=4 --learning_rate=1e-4 --hyperopt=loop --hyperkey=args.batch_size \
  --hyperopt_calls=6 --use_momentum=True --momentum=0.01 --note=batch_size040 \
  --backbone_name=vgg16_bn

	train/iou	val/iou	n_epoch
('batch_size040', 8, 0.01)	0.663733	0.337653	100
('batch_size040', 4, 0.01)	0.67594	0.35008	100

	('val/iou', 'mean')	('val/iou', 'std')	('val/iou', 'amax')
('batch_size040', 4, 0.01)	0.35008	0.00654747	0.35471
('batch_size040', 8, 0.01)	0.337653	0.00232004	0.339294

	train/iou	val/iou	note	batch_size	momentum	n_epoch
3	0.669022	0.336013	batch_size040	8	0.01	100
2	0.658444	0.339294	batch_size040	8	0.01	100
1	0.676273	0.34545	batch_size040	4	0.01	100
0	0.675607	0.35471	batch_size040	4	0.01	100

[yzbx]

res50_fz_ul041 + vgg16_fz_ul041

• backbone_name [resnet50, vgg16]
• upsample_layer [3,4,5]

note: --backbone_freeze=True vgg16 > resnet50 :star: reason: we modify the head of resnet when use_momentum=True, view modify_resnet_head for detail

python test/pspnet_test.py --midnet_scale=10 --backbone_name=resnet50 --backbone_freeze=True \
--batch_size=2 --use_momentum=True --momentum=0.01 --changed_lr_mult=1 --new_lr_mult=2 \
--use_lr_mult=True --note=res50_fz_ul041 --upsample_layer=5 --test=hyperopt \
--hyperkey=model.upsample_layer --hyperopt_calls=6

python test/pspnet_test.py --midnet_scale=10 --backbone_name=vgg16 --backbone_freeze=True \
--batch_size=2 --use_momentum=True --momentum=0.01 --changed_lr_mult=1 --new_lr_mult=2 \
--use_lr_mult=True --note=vgg16_fz_ul041 --upsample_layer=5 --test=hyperopt \
--hyperkey=model.upsample_layer --hyperopt_calls=6

	train/iou	val/iou	n_epoch
('res50_fz_ul041', 'resnet50', 5)	0.276721	0.242904	100
('res50_fz_ul041', 'resnet50', 4)	0.400291	0.292659	100
('res50_fz_ul041', 'resnet50', 3)	0.364319	0.298297	100
('vgg16_fz_ul041', 'vgg16', 3)	0.506028	0.469176	100
('vgg16_fz_ul041', 'vgg16', 5)	0.551127	0.476024	100
('vgg16_fz_ul041', 'vgg16', 4)	0.592125	0.515259	100

	('val/iou', 'mean')	('val/iou', 'std')	('val/iou', 'amax')
('res50_fz_ul041', 'resnet50', 3)	0.298297	0.00339883	0.3007
('res50_fz_ul041', 'resnet50', 4)	0.292659	0.00640006	0.297185
('res50_fz_ul041', 'resnet50', 5)	0.242904	0.00554939	0.246828
('vgg16_fz_ul041', 'vgg16', 3)	0.469176	0.0028797	0.471212
('vgg16_fz_ul041', 'vgg16', 4)	0.515259	0.00222455	0.516832
('vgg16_fz_ul041', 'vgg16', 5)	0.476024	0.00391371	0.478792

	train/iou	val/iou	note	backbone_name	upsample_layer	n_epoch
6	0.274063	0.23898	res50_fz_ul041	resnet50	5	100
7	0.27938	0.246828	res50_fz_ul041	resnet50	5	100
8	0.399848	0.288134	res50_fz_ul041	resnet50	4	100
11	0.365608	0.295894	res50_fz_ul041	resnet50	3	100
10	0.400733	0.297185	res50_fz_ul041	resnet50	4	100
9	0.363031	0.3007	res50_fz_ul041	resnet50	3	100
4	0.506466	0.46714	vgg16_fz_ul041	vgg16	3	100
5	0.50559	0.471212	vgg16_fz_ul041	vgg16	3	100
1	0.551973	0.473257	vgg16_fz_ul041	vgg16	5	100
3	0.55028	0.478792	vgg16_fz_ul041	vgg16	5	100
2	0.592139	0.513686	vgg16_fz_ul041	vgg16	4	100
0	0.592112	0.516832	vgg16_fz_ul041	vgg16	4	100

[yzbx]

fl_grad042 (deeplab)

python test/pspnet_test.py --midnet_scale=10 --backbone_name=resnet50 --batch_size=4 --use_lr_mult=False --note=fl_grad042 --upsample_layer=3 --test=hyperopt --hyperkey=model.focal_loss_grad --focal_loss_gamma=2.0 --hyperopt_calls=6 --dataset_use_part=320

fl_grad042 (dl)

python test/pspnet_test.py --midnet_scale=10 --backbone_name=vgg16 --batch_size=2 --use_lr_mult=False --note=fl_grad042 --upsample_layer=4 --freeze_layer=3 --test=hyperopt --hyperkey=model.focal_loss_grad --focal_loss_gamma=2.0 --hyperopt_calls=6 --dataset_use_part=320

	train/iou	val/iou	n_epoch
('fl_grad042', 'vgg16', False)	0.627771	0.395831	100
('fl_grad042', 'vgg16', True)	0.6495	0.397192	100

	('val/iou', 'mean')	('val/iou', 'std')	('val/iou', 'amax')
('fl_grad042', 'vgg16', False)	0.395831	0.00237601	0.39777
('fl_grad042', 'vgg16', True)	0.397192	0.00515451	0.405031

	train/iou	val/iou	note	backbone_name	focal_loss_grad	n_epoch
9	0.641824	0.391228	fl_grad042	vgg16	True	100
8	0.628734	0.391524	fl_grad042	vgg16	False	100
0	0.650955	0.392208	fl_grad042	vgg16	True	100
4	0.633955	0.39472	fl_grad042	vgg16	False	100
3	0.631577	0.39638	fl_grad042	vgg16	False	100
7	0.654385	0.396876	fl_grad042	vgg16	True	100
11	0.632123	0.397072	fl_grad042	vgg16	False	100
10	0.651506	0.397322	fl_grad042	vgg16	True	100
5	0.61812	0.397518	fl_grad042	vgg16	False	100
6	0.622115	0.39777	fl_grad042	vgg16	False	100
2	0.65806	0.400488	fl_grad042	vgg16	True	100
1	0.640268	0.405031	fl_grad042	vgg16	True	100

[yzbx]

scheduler043 (ts)

sgd

• :x: bugs in poly scheduler, learning rate converge to 0 too fast (epoch=5) python test/pspnet_test.py --midnet_scale=10 --backbone_name=vgg16 --batch_size=2 --use_lr_mult=False --note=scheduler043_poly --upsample_layer=4 --dataset_use_part=320 --learning_rate=1e-2 --optimizer=sgd

• when upsample_layer >3, force use_momentum=True python test/pspnet_test.py --midnet_scale=10 --backbone_name=vgg16 --batch_size=2 --use_lr_mult=False --note=scheduler043_poly --upsample_layer=4 --dataset_use_part=320 --learning_rate=1e-2 --optimizer=sgd

adam

python test/pspnet_test.py --midnet_scale=10 --backbone_name=vgg16 --batch_size=2 --use_lr_mult=False --note=scheduler043_adam --upsample_layer=4 --dataset_use_part=320 --learning_rate=1e-4 --optimizer=adam

pop

python test/pspnet_test.py --midnet_scale=10 --backbone_name=vgg16 --batch_size=2 --use_lr_mult=False --note=scheduler043_pop --upsample_layer=4 --dataset_use_part=320 --learning_rate=1e-2 --optimizer=sgd --scheduler=poly_rop

	train/iou	val/iou	note	optimizer	scheduler	n_epoch
0	0.449149	0.33499	scheduler043_poly	sgd		100
2	0.628333	0.390597	scheduler043_pop	sgd	poly_rop	100
1	0.670823	0.410676	scheduler043_adam	adam		100

conclusion

adam > poly_rop + sgd > poly + sgd

[yzbx]

aug044

python test/pspnet_test.py --backbone_name=vgg16 --backbone_freeze=True --n_epoch=50 --keep_crop_ratio=True --note=aug044 --dataset_use_part=320 --test=hyperopt --hyperkey=args.augmentation --hyperopt_calls=2

	train/iou	val/iou	n_epoch
('aug044', False)	0.757454	0.29444	50
('aug044', True)	0.479841	0.317993	50

	('val/iou', 'mean')	('val/iou', 'std')	('val/iou', 'amax')
('aug044', False)	0.29444	0.00474	0.297792
('aug044', True)	0.317993	0.00294297	0.320074

	train/iou	val/iou	note	augmentation	n_epoch
0	0.758187	0.291088	aug044	False	50
1	0.756721	0.297792	aug044	False	50
3	0.480831	0.315912	aug044	True	50
2	0.478852	0.320074	aug044	True	50

• the head(first layer) of resnet is not changed and not got trained val miou=0.31 python test/pspnet_test.py --midnet_scale=10 --backbone_name=resnet50 --batch_size=2 --use_lr_mult=False --note=w_resnet_head --upsample_layer=4 --dataset_use_part=320 --n_epoch=50 --backbone_freeze=True --modify_resnet_head=False

• the head(first layer) of resnet is changed and got trained val miou=0.168 python test/pspnet_test.py --midnet_scale=10 --backbone_name=resnet50 --batch_size=2 --use_lr_mult=False --note=wo_resnet_head --upsample_layer=4 --dataset_use_part=320 --n_epoch=50 --backbone_freeze=True --modify_resnet_head=True

[yzbx]

freeze_layer045

python test/pspnet_test.py --midnet_scale=10 --backbone_name=resnet50 --batch_size=2 --use_lr_mult=False --note=freeze_layer045 --upsample_layer=4 --dataset_use_part=320 --test=hyperopt --hyperkey=model.freeze_layer --hyperopt_calls=2 --n_epoch=50 --backbone_freeze=False --modify_resnet_head=False

python test/pspnet_test.py --midnet_scale=10 --backbone_name=resnet50 --batch_size=2 --use_lr_mult=False --note=freeze_layer045 --upsample_layer=5 --dataset_use_part=320 --freeze_layer=4 --n_epoch=50 --backbone_freeze=False --modify_resnet_head=False

python test/pspnet_test.py --midnet_scale=10 --backbone_name=resnet50 --batch_size=2 --use_lr_mult=False --note=freeze_layer045 --upsample_layer=5 --dataset_use_part=320 --freeze_layer=3 --n_epoch=50 --backbone_freeze=False --modify_resnet_head=False

	train/iou	val/iou	n_epoch
('freeze_layer045', 5, 4)	0.605244	0.393496	50
('freeze_layer045', 4, 3)	0.634177	0.395575	50
('freeze_layer045', 4, 2)	0.63674	0.400545	50
('freeze_layer045', 4, 0)	0.637557	0.401278	50
('freeze_layer045', 4, 1)	0.638608	0.403595	50
('freeze_layer045', 5, 3)	0.670808	0.428208	50

	('val/iou', 'mean')	('val/iou', 'std')	('val/iou', 'amax')
('freeze_layer045', 4, 0)	0.401278	0.0124565	0.410086
('freeze_layer045', 4, 1)	0.403595	0.0174144	0.415909
('freeze_layer045', 4, 2)	0.400545	0.00740236	0.40578
('freeze_layer045', 4, 3)	0.395575	0.00802339	0.401248
('freeze_layer045', 5, 3)	0.428208	nan	0.428208
('freeze_layer045', 5, 4)	0.393496	nan	0.393496

	train/iou	val/iou	note	upsample_layer	freeze_layer	n_epoch
3	0.634262	0.389902	freeze_layer045	4	3	50
7	0.636844	0.391281	freeze_layer045	4	1	50
1	0.63677	0.39247	freeze_layer045	4	0	50
8	0.605244	0.393496	freeze_layer045	5	4	50
0	0.631536	0.395311	freeze_layer045	4	2	50
6	0.634093	0.401248	freeze_layer045	4	3	50
2	0.641945	0.40578	freeze_layer045	4	2	50
5	0.638344	0.410086	freeze_layer045	4	0	50
4	0.640372	0.415909	freeze_layer045	4	1	50
9	0.670808	0.428208	freeze_layer045	5	3	50

[yzbx]

crop_size_step046

python test/pspnet_test.py --backbone_name=vgg16 --backbone_freeze=True --n_epoch=50 --keep_crop_ratio=False --note=crop_size_step046 --min_crop_size=720 --max_crop_size=1024 --test=hyperopt --hyperkey=aug.crop_size_step --hyperopt_calls=2 --dataset_use_part=320

	train/iou	val/iou	n_epoch
('crop_size_step046', 128)	0.413005	0.295337	50
('crop_size_step046', 64)	0.419373	0.296078	50
('crop_size_step046', 32)	0.42527	0.296183	50

	('val/iou', 'mean')	('val/iou', 'std')	('val/iou', 'amax')
('crop_size_step046', 32)	0.296183	8.98149e-05	0.296246
('crop_size_step046', 64)	0.296078	0.000675677	0.296556
('crop_size_step046', 128)	0.295337	0.00698611	0.300276

	train/iou	val/iou	note	crop_size_step	n_epoch
3	0.410183	0.290397	crop_size_step046	128	50
5	0.42376	0.295601	crop_size_step046	64	50
2	0.420888	0.296119	crop_size_step046	32	50
0	0.429652	0.296246	crop_size_step046	32	50
1	0.414986	0.296556	crop_size_step046	64	50
4	0.415828	0.300276	crop_size_step046	128	50

[yzbx]

crop_size_step047

python test/pspnet_test.py --backbone_name=vgg16 --backbone_freeze=True --n_epoch=50 --keep_crop_ratio=False --note=crop_size_step047 --min_crop_size=480 --max_crop_size=1080 --test=hyperopt --hyperkey=aug.crop_size_step --hyperopt_calls=2 --dataset_use_part=320

• deeplab2 python test/pspnet_test.py --backbone_name=vgg19_bn --backbone_freeze=True --n_epoch=50 --keep_crop_ratio=False --note=crop_size_step047 --min_crop_size=480 --max_crop_size=1080 --test=hyperopt --hyperkey=aug.crop_size_step --hyperopt_calls=2 --dataset_use_part=320

experiment results

	train/iou	val/iou	n_epoch
('crop_size_step047', 'vgg19_bn', 128)	0.310905	0.245783	50
('crop_size_step047', 'vgg19_bn', 64)	0.332896	0.259655	50
('crop_size_step047', 'vgg19_bn', 32)	0.334612	0.267079	50
('crop_size_step047', 'vgg16', 64)	0.316341	0.269054	50
('crop_size_step047', 'vgg16', 32)	0.37554	0.294664	50

	('val/iou', 'mean')	('val/iou', 'std')	('val/iou', 'amax')
('crop_size_step047', 'vgg16', 32)	0.294664	0.00315882	0.296898
('crop_size_step047', 'vgg16', 64)	0.269054	0.036287	0.294713
('crop_size_step047', 'vgg19_bn', 32)	0.267079	0.0015325	0.268162
('crop_size_step047', 'vgg19_bn', 64)	0.259655	0.00830487	0.265528
('crop_size_step047', 'vgg19_bn', 128)	0.245783	0.00398913	0.248604

	train/iou	val/iou	note	backbone_name	crop_size_step	n_epoch
6	0.312579	0.242962	crop_size_step047	vgg19_bn	128	50
5	0.26658	0.243395	crop_size_step047	vgg16	64	50
1	0.309232	0.248604	crop_size_step047	vgg19_bn	128	50
4	0.321546	0.253783	crop_size_step047	vgg19_bn	64	50
2	0.344247	0.265528	crop_size_step047	vgg19_bn	64	50
9	0.33452	0.265995	crop_size_step047	vgg19_bn	32	50
0	0.334704	0.268162	crop_size_step047	vgg19_bn	32	50
7	0.37056	0.29243	crop_size_step047	vgg16	32	50
8	0.366101	0.294713	crop_size_step047	vgg16	64	50
3	0.38052	0.296898	crop_size_step047	vgg16	32	50

[dishen12]

freeze_layer048

python test/pspnet_test.py --midnet_scale=10 --backbone_name=resnet50 --batch_size=2 --use_lr_mult=False --note=freeze_layer048 --upsample_layer=5 --dataset_use_part=320 --test=hyperopt --hyperkey=model.freeze_layer --hyperopt_calls=2 --n_epoch=50 --backbone_freeze=False --modify_resnet_head=False

python test/pspnet_test.py --midnet_scale=10 --backbone_name=vgg19_bn --batch_size=2 --use_lr_mult=False --note=freeze_layer048 --upsample_layer=5 --dataset_use_part=320 --test=hyperopt --hyperkey=model.freeze_layer --hyperopt_calls=2 --n_epoch=50 --backbone_freeze=False --modify_resnet_head=False

	train/iou	val/iou	n_epoch
('freeze_layer048', 'resnet50', 4)	0.605886	0.394298	50
('freeze_layer048', 'vgg19_bn', 4)	0.594443	0.398692	50
('freeze_layer048', 'vgg19_bn', 0)	0.643579	0.405965	50
('freeze_layer048', 'vgg19_bn', 3)	0.620446	0.410002	50
('freeze_layer048', 'vgg19_bn', 1)	0.637407	0.413641	50
('freeze_layer048', 'vgg19_bn', 2)	0.63983	0.413768	50
('freeze_layer048', 'resnet50', 1)	0.666989	0.418592	50
('freeze_layer048', 'resnet50', 2)	0.652004	0.41893	50
('freeze_layer048', 'resnet50', 3)	0.655319	0.425352	50
('freeze_layer048', 'resnet50', 0)	0.677817	0.426222	50

	('val/iou', 'mean')	('val/iou', 'std')	('val/iou', 'amax')
('freeze_layer048', 'resnet50', 0)	0.426222	0.0130617	0.439855
('freeze_layer048', 'resnet50', 1)	0.418592	0.00972157	0.425466
('freeze_layer048', 'resnet50', 2)	0.41893	0.00347716	0.421389
('freeze_layer048', 'resnet50', 3)	0.425352	0.00134545	0.426303
('freeze_layer048', 'resnet50', 4)	0.394298	nan	0.394298
('freeze_layer048', 'vgg19_bn', 0)	0.405965	0.00155379	0.407064
('freeze_layer048', 'vgg19_bn', 1)	0.413641	0.00685455	0.418488
('freeze_layer048', 'vgg19_bn', 2)	0.413768	0.0143709	0.42393
('freeze_layer048', 'vgg19_bn', 3)	0.410002	0.0017582	0.411245
('freeze_layer048', 'vgg19_bn', 4)	0.398692	0.010856	0.406369

	train/iou	val/iou	note	backbone_name	freeze_layer	n_epoch
5	0.597991	0.391016	freeze_layer048	vgg19_bn	4	50
2	0.605886	0.394298	freeze_layer048	resnet50	4	50
4	0.635482	0.403606	freeze_layer048	vgg19_bn	2	50
11	0.636239	0.404866	freeze_layer048	vgg19_bn	0	50
8	0.590894	0.406369	freeze_layer048	vgg19_bn	4	50
7	0.65092	0.407064	freeze_layer048	vgg19_bn	0	50
16	0.617055	0.408758	freeze_layer048	vgg19_bn	3	50
19	0.63602	0.408794	freeze_layer048	vgg19_bn	1	50
17	0.623837	0.411245	freeze_layer048	vgg19_bn	3	50
18	0.670538	0.411718	freeze_layer048	resnet50	1	50
15	0.694188	0.413818	freeze_layer048	resnet50	0	50
9	0.658555	0.416471	freeze_layer048	resnet50	2	50
3	0.638793	0.418488	freeze_layer048	vgg19_bn	1	50
12	0.645454	0.421389	freeze_layer048	resnet50	2	50
10	0.644178	0.42393	freeze_layer048	vgg19_bn	2	50
0	0.650143	0.424401	freeze_layer048	resnet50	3	50
14	0.673936	0.424994	freeze_layer048	resnet50	0	50
13	0.66344	0.425466	freeze_layer048	resnet50	1	50
6	0.660495	0.426303	freeze_layer048	resnet50	3	50
1	0.665328	0.439855	freeze_layer048	resnet50	0	50

[dishen12]

pad_for_crop049

python test/pspnet_test.py --backbone_name=vgg19_bn --backbone_freeze=True --n_epoch=50 --keep_crop_ratio=False --note=pad_for_crop049 --min_crop_size=480 --max_crop_size=1080 --test=hyperopt --hyperkey=aug.crop_size_step --hyperopt_calls=2 --dataset_use_part=320 --pad_for_crop=True

	train/iou	val/iou	n_epoch
('pad_for_crop049', 32)	0.328854	0.255305	50
('pad_for_crop049', 64)	0.332352	0.258802	50
('pad_for_crop049', 128)	0.325598	0.264147	50

	('val/iou', 'mean')	('val/iou', 'std')	('val/iou', 'amax')
('pad_for_crop049', 32)	0.255305	0.00339567	0.257706
('pad_for_crop049', 64)	0.258802	0.0035625	0.261321
('pad_for_crop049', 128)	0.264147	0.00337695	0.266535

	train/iou	val/iou	note	crop_size_step	n_epoch
4	0.33402	0.252903	pad_for_crop049	32	50
5	0.332843	0.256283	pad_for_crop049	64	50
1	0.323688	0.257706	pad_for_crop049	32	50
0	0.331861	0.261321	pad_for_crop049	64	50
2	0.323482	0.261759	pad_for_crop049	128	50
3	0.327714	0.266535	pad_for_crop049	128	50

[dishen12]

moment050

python test/pspnet_test.py --use_dropout False --use_lr_mult False --dataset_use_part=320 --backbone_pretrained=True --midnet_scale=5 --n_epoch=50 --upsample_type=bilinear --hyperopt_calls=3 --use_momentum=True --hyperkey=model.momentum --note=moment050 --test=hyperopt --batch_size=16

python test/pspnet_test.py --use_dropout False --use_lr_mult False --dataset_use_part=320 --backbone_pretrained=True --midnet_scale=5 --n_epoch=50 --upsample_type=bilinear --hyperopt_calls=3 --use_momentum=True --hyperkey=model.momentum --note=moment050 --test=hyperopt --batch_size=8

python test/pspnet_test.py --use_dropout False --use_lr_mult False --dataset_use_part=320 --backbone_pretrained=True --midnet_scale=5 --n_epoch=50 --upsample_type=bilinear --hyperopt_calls=3 --use_momentum=True --hyperkey=model.momentum --note=moment050 --test=hyperopt --batch_size=4

	train/iou	val/iou	n_epoch
('moment050', 0.1, 16)	0.387601	0.232016	50
('moment050', 0.05, 16)	0.384924	0.233656	50
('moment050', 0.01, 16)	0.393587	0.234275	50
('moment050', 0.1, 12)	0.398063	0.238977	50
('moment050', 0.01, 8)	0.410852	0.240343	50
('moment050', 0.05, 12)	0.393638	0.241999	50
('moment050', 0.01, 12)	0.397497	0.243363	50
('moment050', 0.05, 8)	0.412782	0.249181	50
('moment050', 0.1, 8)	0.410346	0.250393	50
('moment050', 0.05, 4)	0.427929	0.252875	50
('moment050', 0.01, 4)	0.41603	0.254698	50
('moment050', 0.1, 4)	0.426359	0.258612	50

	('val/iou', 'mean')	('val/iou', 'std')	('val/iou', 'amax')
('moment050', 0.01, 4)	0.254698	0.00623278	0.261015
('moment050', 0.01, 8)	0.240343	0.00138532	0.241815
('moment050', 0.01, 12)	0.243363	0.00265116	0.245827
('moment050', 0.01, 16)	0.234275	0.00719522	0.240275
('moment050', 0.05, 4)	0.252875	0.0048432	0.256612
('moment050', 0.05, 8)	0.249181	0.00545942	0.253982
('moment050', 0.05, 12)	0.241999	0.00957748	0.252978
('moment050', 0.05, 16)	0.233656	0.00586555	0.240343
('moment050', 0.1, 4)	0.258612	0.0016425	0.259569
('moment050', 0.1, 8)	0.250393	0.00448551	0.253317
('moment050', 0.1, 12)	0.238977	0.00284067	0.242145
('moment050', 0.1, 16)	0.232016	0.00432807	0.234779

	train/iou	val/iou	note	momentum	batch_size	n_epoch
12	0.393825	0.226298	moment050	0.01	16	50
20	0.381152	0.227028	moment050	0.1	16	50
15	0.39203	0.229379	moment050	0.05	16	50
19	0.376486	0.231246	moment050	0.05	16	50
2	0.387239	0.234241	moment050	0.1	16	50
4	0.394413	0.234779	moment050	0.1	16	50
32	0.396637	0.23536	moment050	0.05	12	50
8	0.395999	0.236253	moment050	0.01	16	50
27	0.388556	0.236657	moment050	0.1	12	50
29	0.395	0.237659	moment050	0.05	12	50
23	0.406014	0.23813	moment050	0.1	12	50
9	0.413757	0.239065	moment050	0.01	8	50
35	0.405647	0.240149	moment050	0.01	8	50
22	0.390935	0.240275	moment050	0.01	16	50
6	0.386256	0.240343	moment050	0.05	16	50
1	0.392433	0.240557	moment050	0.01	12	50
0	0.413153	0.241815	moment050	0.01	8	50
31	0.39962	0.242145	moment050	0.1	12	50
7	0.410361	0.243243	moment050	0.05	8	50
17	0.40451	0.243704	moment050	0.01	12	50
25	0.416548	0.245229	moment050	0.1	8	50
33	0.395547	0.245827	moment050	0.01	12	50
16	0.433838	0.247404	moment050	0.05	4	50
24	0.421631	0.248553	moment050	0.01	4	50
21	0.420177	0.25032	moment050	0.05	8	50
13	0.404341	0.252633	moment050	0.1	8	50
14	0.389276	0.252978	moment050	0.05	12	50
34	0.41015	0.253317	moment050	0.1	8	50
28	0.407808	0.253982	moment050	0.05	8	50
10	0.42267	0.254527	moment050	0.01	4	50
30	0.409209	0.254611	moment050	0.05	4	50
18	0.440741	0.256612	moment050	0.05	4	50
26	0.42998	0.256715	moment050	0.1	4	50
5	0.428624	0.259552	moment050	0.1	4	50
11	0.420472	0.259569	moment050	0.1	4	50
3	0.403789	0.261015	moment050	0.01	4	50

[dishen12]

freeze_ratio051

python test/pspnet_test.py --backbone_name=resnet50 --use_lr_mult=True --freeze_ratio=0.5 --upsample_layer=5 --modify_resnet_head=False --n_epoch=50 --dataset_use_part=320 --test=hyperopt --hyperkey=model.freeze_ratio --hyperopt_calls=2 --note=freeze_ratio051

python test/pspnet_test.py --backbone_name=vgg19_bn --use_lr_mult=True --freeze_ratio=0.5 --upsample_layer=5 --modify_resnet_head=False --n_epoch=50 --dataset_use_part=320 --test=hyperopt --hyperkey=model.freeze_ratio --hyperopt_calls=2 --note=freeze_ratio051

	train/iou	val/iou	n_epoch
('freeze_ratio051', 0.5, 'vgg19_bn')	0.567544	0.344758	50
('freeze_ratio051', 0.3, 'resnet50')	0.566313	0.346822	50
('freeze_ratio051', 0.3, 'vgg19_bn')	0.57419	0.347284	50

	('val/iou', 'mean')	('val/iou', 'std')	('val/iou', 'amax')
('freeze_ratio051', 0.3, 'resnet50')	0.346822	0.00243954	0.348547
('freeze_ratio051', 0.3, 'vgg19_bn')	0.347284	0.00110024	0.348062
('freeze_ratio051', 0.5, 'vgg19_bn')	0.344758	0.00337617	0.347145

	train/iou	val/iou	note	freeze_ratio	backbone_name	n_epoch
0	0.567685	0.34237	freeze_ratio051	0.5	vgg19_bn	50
5	0.567863	0.345097	freeze_ratio051	0.3	resnet50	50
4	0.577913	0.346506	freeze_ratio051	0.3	vgg19_bn	50
2	0.567403	0.347145	freeze_ratio051	0.5	vgg19_bn	50
1	0.570467	0.348062	freeze_ratio051	0.3	vgg19_bn	50
3	0.564764	0.348547	freeze_ratio051	0.3	resnet50	50

[yzbx]

batch norm vs group norm

vgg16 baseline 0.589

python test/pspnet_test.py --backbone_name=vgg16_bn --use_momentum=True --midnet_scale=10 --upsample_layer=4 --use_lr_mult=False --batch_size=4 --note=vgg16_bn

vgg16 + duc 0.623

python test/pspnet_test.py --backbone_name=vgg16_bn --use_momentum=True --midnet_scale=10 --upsample_layer=4 --use_lr_mult=False --batch_size=4 --upsample_type=duc --note=vgg16_duc

vgg16 + gn 0.5802->0.5883

gn4 0.5802 python test/pspnet_test.py --backbone_name=vgg16_gn --use_momentum=True --midnet_scale=10 --upsample_layer=4 --use_lr_mult=False --batch_size=4 --note=vgg16_gn

gn32 0.5883 python test/pspnet_test.py --backbone_name=vgg16_gn --use_momentum=True --midnet_scale=10 --upsample_layer=4 --use_lr_mult=False --batch_size=4 --note=vgg16_gn32

vgg19 + gn 0.5393

python test/pspnet_test.py --backbone_name=vgg19_gn --use_momentum=True --midnet_scale=10 --upsample_layer=3 --use_lr_mult=False --batch_size=4 --note=vgg19_gn

norm052

python test/pspnet_test.py --backbone_name=vgg16_gn --use_lr_mult=False --batch_size=4 --use_momentum=True --n_epoch=50 --dataset_use_part=320 --test=hyperopt --hyperkey=model.upsample_layer --hyperopt_calls=3 --note=norm052

python test/pspnet_test.py --backbone_name=vgg16_bn --use_lr_mult=False --batch_size=4 --use_momentum=True --n_epoch=50 --dataset_use_part=320 --test=hyperopt --hyperkey=model.upsample_layer --hyperopt_calls=3 --note=norm052

	train/iou	val/iou	n_epoch
('norm052', 'vgg16_gn', 3)	0.501118	0.306874	50
('norm052', 'vgg16_bn', 3)	0.501284	0.314949	50
('norm052', 'vgg16_gn', 4)	0.545462	0.324762	50
('norm052', 'vgg16_gn', 5)	0.552894	0.330314	50
('norm052', 'vgg16_bn', 4)	0.556631	0.337607	50
('norm052', 'vgg16_bn', 5)	0.58496	0.350951	50

	('val/iou', 'mean')	('val/iou', 'std')	('val/iou', 'amax')
('norm052', 'vgg16_bn', 3)	0.314949	0.0017261	0.316807
('norm052', 'vgg16_bn', 4)	0.337607	0.00124658	0.338511
('norm052', 'vgg16_bn', 5)	0.350951	0.00147268	0.351882
('norm052', 'vgg16_gn', 3)	0.306874	0.00704153	0.314657
('norm052', 'vgg16_gn', 4)	0.324762	0.00550772	0.330364
('norm052', 'vgg16_gn', 5)	0.330314	0.00513577	0.336072

	train/iou	val/iou	note	backbone_name	upsample_layer	n_epoch
13	0.505019	0.300944	norm052	vgg16_gn	3	50
8	0.49779	0.30502	norm052	vgg16_gn	3	50
7	0.502547	0.313395	norm052	vgg16_bn	3	50
4	0.498631	0.314645	norm052	vgg16_bn	3	50
11	0.500545	0.314657	norm052	vgg16_gn	3	50
5	0.502675	0.316807	norm052	vgg16_bn	3	50
3	0.539528	0.319354	norm052	vgg16_gn	4	50
14	0.556341	0.324568	norm052	vgg16_gn	4	50
6	0.545202	0.326208	norm052	vgg16_gn	5	50
15	0.552308	0.32866	norm052	vgg16_gn	5	50
17	0.540517	0.330364	norm052	vgg16_gn	4	50
16	0.561171	0.336072	norm052	vgg16_gn	5	50
9	0.556136	0.336185	norm052	vgg16_bn	4	50
12	0.564894	0.338125	norm052	vgg16_bn	4	50
1	0.548862	0.338511	norm052	vgg16_bn	4	50
10	0.586208	0.349253	norm052	vgg16_bn	5	50
2	0.580422	0.351717	norm052	vgg16_bn	5	50
0	0.588251	0.351882	norm052	vgg16_bn	5	50

[yzbx]

upsample_type053

python test/pspnet_test.py --backbone_name=vgg16_gn --use_lr_mult=False --upsample_layer=4 --batch_size=4 --use_momentum=True --n_epoch=50 --dataset_use_part=320 --test=hyperopt --hyperkey=model.upsample_type --hyperopt_calls=2 --note=upsample_type053

	train/iou	val/iou	n_epoch
('upsample_type053', 'vgg16_gn', 'fcn')	0.410659	0.251091	50
('upsample_type053', 'vgg16_gn', 'bilinear')	0.55071	0.322475	50
('upsample_type053', 'vgg16_gn', 'duc')	0.578014	0.350003	50

	('val/iou', 'mean')	('val/iou', 'std')	('val/iou', 'amax')
('upsample_type053', 'vgg16_gn', 'bilinear')	0.322475	0.00376715	0.325818
('upsample_type053', 'vgg16_gn', 'duc')	0.350003	0.00689807	0.359094
('upsample_type053', 'vgg16_gn', 'fcn')	0.251091	0.0210701	0.26599

	train/iou	val/iou	note	backbone_name	upsample_type	n_epoch
9	0.373488	0.236192	upsample_type053	vgg16_gn	fcn	50
5	0.44783	0.26599	upsample_type053	vgg16_gn	fcn	50
8	0.548628	0.317766	upsample_type053	vgg16_gn	bilinear	50
2	0.548012	0.321123	upsample_type053	vgg16_gn	bilinear	50
7	0.554652	0.325194	upsample_type053	vgg16_gn	bilinear	50
0	0.55155	0.325818	upsample_type053	vgg16_gn	bilinear	50
3	0.571883	0.342683	upsample_type053	vgg16_gn	duc	50
4	0.581518	0.347543	upsample_type053	vgg16_gn	duc	50
6	0.577644	0.350691	upsample_type053	vgg16_gn	duc	50
1	0.581011	0.359094	upsample_type053	vgg16_gn	duc	50