-
Notifications
You must be signed in to change notification settings - Fork 120
/
init_MSLapSRN_opts.m
119 lines (102 loc) · 4.27 KB
/
init_MSLapSRN_opts.m
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
function opts = init_MSLapSRN_opts(scales, depth, recursive, gpu)
% -------------------------------------------------------------------------
% Description:
% Generate all options for MS-LapSRN
%
% Input:
% - scale : SR upsampling scale
% - depth : number of conv layers in one pyramid level
% - gpu : GPU ID, 0 for CPU mode
%
% Output:
% - opts : all options for MS-LapSRN
%
% Citation:
% Fast and Accurate Image Super-Resolution with Deep Laplacian Pyramid Networks
% Wei-Sheng Lai, Jia-Bin Huang, Narendra Ahuja, and Ming-Hsuan Yang
% arXiv, 2017
%
% Contact:
% Wei-Sheng Lai
% wlai24@ucmerced.edu
% University of California, Merced
% -------------------------------------------------------------------------
%% network options
opts.scales = scales; % training scales (use vector, e.g., [2, 3, 4])
opts.LRL = 'SS'; % local residual learning [Opts = NS, DS, SS]
opts.depth = depth; % number of conv in a block
opts.recursive = recursive;% number of recurrent blocks
opts.weight_decay = 0.0001;
opts.init_sigma = 0.001;
opts.conv_f = 3;
opts.conv_n = 64;
opts.loss = 'L1';
%% training options
opts.gpu = gpu;
opts.batch_size = 64;
opts.num_train_batch = 1000; % number of training batch in one epoch
opts.num_valid_batch = 100; % number of validation batch in one epoch
opts.lr = 5e-6; % initial learning rate
opts.lr_step = 100; % number of epochs to drop learning rate
opts.lr_drop = 0.5; % learning rate drop ratio
opts.lr_min = 1e-6; % minimum learning rate
opts.patch_size = 128;
opts.data_augmentation = 1;
opts.scale_augmentation = 1;
%% dataset options
opts.train_dataset = {};
opts.train_dataset{end+1} = 'T91';
opts.train_dataset{end+1} = 'BSDS200';
%opts.train_dataset{end+1} = 'General100';
opts.valid_dataset = {};
%opts.valid_dataset{end+1} = 'Set5';
%opts.valid_dataset{end+1} = 'Set14';
opts.valid_dataset{end+1} = 'BSDS100';
%% setup model name
opts.data_name = 'train';
for i = 1:length(opts.train_dataset)
opts.data_name = sprintf('%s_%s', opts.data_name, opts.train_dataset{i});
end
scale_str = '';
for s = 1:length(opts.scales)
scale_str = sprintf('%s%d', scale_str, opts.scales(s));
end
opts.net_name = sprintf('MSLapSRN_x%s_%s_D%d_R%d_fn%d', ...
scale_str, opts.LRL, opts.depth, opts.recursive, ...
opts.conv_n);
opts.model_name = sprintf('%s_%s_%s_pw%d_lr%s_step%d_drop%s_min%s_bs%d', ...
opts.net_name, ...
opts.loss, ...
opts.data_name, opts.patch_size, ...
num2str(opts.lr), opts.lr_step, ...
num2str(opts.lr_drop), num2str(opts.lr_min), ...
opts.batch_size);
%% setup dagnn training parameters
if( opts.gpu == 0 )
opts.train.gpus = [];
else
opts.train.gpus = [opts.gpu];
end
opts.train.batchSize = opts.batch_size;
opts.train.numEpochs = 1000;
opts.train.continue = true;
opts.train.learningRate = learning_rate_policy(opts.lr, opts.lr_step, opts.lr_drop, ...
opts.lr_min, opts.train.numEpochs);
opts.train.expDir = fullfile('models', opts.model_name) ; % model output dir
if( ~exist(opts.train.expDir, 'dir') )
mkdir(opts.train.expDir);
end
opts.train.model_name = opts.model_name;
opts.train.num_train_batch = opts.num_train_batch;
opts.train.num_valid_batch = opts.num_valid_batch;
%% setup loss
opts.train.derOutputs = {};
for k = 1:length(opts.scales)
scale = opts.scales(k);
level = ceil(log(scale) / log(2));
for l = 1 : level
opts.train.derOutputs{end + 1} = sprintf('x%dSR_%dx_%s_loss', scale, 2^l, opts.loss);
opts.train.derOutputs{end + 1} = 1;
end
end
end