yzbx commented 6 years ago

hyper params experiment summary

no influence-> obviously influence :no_mouth: < :heavy_plus_sign: < :yum: < :star:

hyper params	value	miou	experiments	others
lr for sgd	0.01	-	hyperopt002, hyperopt	:star: 1e-2
lr for adam	1e-4	-	hyperopt005/012/013/014	:star: 1e-4
norm_ways	-	-	hyperopt004 :question: , hyperopt007 :question: , norm_ways030	:heavy_plus_sign: caffe/cityscapes
l2_reg	[1e-5, 0.1]	-	hyperopt006	:no_mouth:
dataset size	32->320->640->full	0.22 -> 0.39 -> 0.45 -> 0.57	-	:star:
data augmentaion	T/F	-	hyperopt008 :question:	:no_mouth:
backbone freeze	T/F	-	hyperopt009	:heavy_plus_sign: T for datasize=32
backbone pretrained	T/F	-	hyperopt010/011	:star: T for datasize=32
use_lr_mult	T/F	0.35->0.37	hyperopt011/hyperopt021	:no_mouth: F for datasize=32, :star: T for datasize=320
batch norm	T/F	0.378->0.381	[hyperopt015 vs hyperopt007], hyperopt017	:yum: T for batch size=4
edge_weight	0.1,0.2,0.5,1.0	0.03	hyperopt016	:no_mouth: without edge will better
momentum	0.1,0.3,0.5,0.7,0.9	0.381 -> 0.380 -> 0.382 -> 0.388 -> 0.379	hyperopt018	:no_mouth:
upsample layer	3,4,5	0.32->0.38->0.40	hyperopt019	:star: 5 for fcn
use_bias	T/F	0.37->0.36,0.32->0.31	hyperopt020,use_bias035	:heavy_plus_sign: T for fcn32, pspnet

yzbx commented 6 years ago

imagecls: cifar100 classification benchmark

the best acc almost the merged acc

	train/cls_acc	val/cls_acc	note	fcc_block_number	n_epoch	name
2	0.9979	0.4039	sota_cnw_9	9	30	fc_net
0	0.99976	0.5655	sota_cnw_6	6	30	fc_net
1	0.99978	0.6169	sota_cnw_3	3	30	fc_net

	train/cls_acc	val/cls_acc	note	fcc_block_number	n_epoch	name
1	0.99868	0.4465	sota_cw_9	9	30	fc_net
2	0.99968	0.5277	sota_cw_6	6	30	fc_net
0	0.99978	0.5905	sota_cw_3	3	30	fc_net

	train/cls_acc	val/cls_acc	note	csc_number	n_epoch	name
3	0.99988	0.6276	sota_000	8	30	csc_net
4	0.99978	0.6376	sota_000	1	30	csc_net
0	0.99976	0.6392	sota_000	2	30	csc_net
1	0.99978	0.6434	sota_000	4	30	csc_net

	train/cls_acc	val/cls_acc	note	sm_block_number	n_epoch	name
1	0.91898	0.3393	sota_depth_000	9	30	sm_net
0	0.99908	0.4792	sota_depth_000	6	30	sm_net
2	0.99976	0.5931	sota_depth_000	3	30	sm_net

	train/cls_acc	val/cls_acc	note	sm_block_width	n_epoch	name
0	0.99978	0.5983	sota_width_000	3	30	sm_net
1	0.99974	0.6005	sota_width_000	9	30	sm_net
2	0.99978	0.6084	sota_width_000	6	30	sm_net

	train/cls_acc	val/cls_acc	note	mst_depth	n_epoch	name
0	0.9928	0.6029	sota_depth_3	3	30	mst_net
1	0.99742	0.6234	sota_depth_2	2	30	mst_net
2	0.9998	0.6406	sota_depth_1	1	30	mst_net

	train/cls_acc	val/cls_acc	note	mst_mode	n_epoch	name
2	0.99296	0.6013	sota_mode_1	1	30	mst_net
0	0.9994	0.6246	sota_mode_111	111	30	mst_net
1	0.99976	0.6634	sota_mode_321	321	30	mst_net

yzbx commented 5 years ago

no influence-> obviously influence :no_mouth: < :heavy_plus_sign: < :yum: < :star: 1% < 1.5% < 2% < 5%

hyper params	value	miou	experiments	others
class weight	T/F	0.36-0.37, 0.35-0.38	class_weight026/027	:no_mouth:
focal loss gamma	[1,2]	0.377 - 0.378, 0.386 - 0.403	focal_loss028/029	:yum:
use dropout	T/F	0.30 - 0.32	use_dropout034	:yum:
momentum	0.01, 0.05, 0.1	-	moment_bias036	:no_mouth:
bias	T/F	-	moment_bias036	:heavy_plus_sign:
lr_mult	-	0.31-0.34	lr_mult037	:yum:
upsample_layer	3,4,5	0.37,0.47,0.46	upsample_layer039	:start:
batch_size	4,6,8	-	moment_bias036, batch_size040	:no_mouth: T for resnet, F for vgg
focal loss grad	T/F	0.397/0.395	fl_grad_042	:no_mouth: T is better
scheduler	pop, poly , adam	0.39,0.33,0.41	scheduler043	:star: for data size=320
augmentation	T/F	0.318/0.294	aug044	:yum:
freeze_layer	-	0.393 - 0.428	freeze_layer045,048	:yum: 5_3,2,1 may better
freeze_ratio	0.3,0.5	0.347,0.343	freeze_ratio051	:no_mouth:
crop_size_step	32,64,128	0.295-0.296	crop_size_step046,47	:no_mouth: 32 is better
crop_size_step	32,64,128	0.255,0.258,0.264	pad_for_crop049	:no_mouth: 128 is better

yzbx commented 5 years ago

acc (mean/max)

hyper params	value	acc	experiments	others
use_data_aug	T/F	0.72/0.66	aug	:star:
use_ortho_loss	T/F	- / -	ortho	:heavy_plus_sign:
scheduler	poly_pop, rop, none	-/-	poly_rop, rop, pop	:heavy_plus_sign: