Mercurial > ift6266

comparison deep/convolutional_dae/salah_exp/config.py @ 377:0b7e64e8e93f

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
branch merge
author Arnaud Bergeron <abergeron@gmail.com>
date Sun, 25 Apr 2010 17:12:03 -0400
parents c05680f8c92f
children
comparison
equal deleted inserted replaced
376:01445a75c702 377:0b7e64e8e93f
1 '''
2 These are parameters used by nist_sda.py. They'll end up as globals in there.
3
4 Rename this file to config.py and configure as needed.
5 DON'T add the renamed file to the repository, as others might use it
6 without realizing it, with dire consequences.
7 '''
8
9 # Set this to True when you want to run cluster tests, ie. you want
10 # to run on the cluster, many jobs, but want to reduce the training
11 # set size and the number of epochs, so you know everything runs
12 # fine on the cluster.
13 # Set this PRIOR to inserting your test jobs in the DB.
14 TEST_CONFIG = False
15
16 NIST_ALL_LOCATION = '/data/lisa/data/nist/by_class/all'
17 NIST_ALL_TRAIN_SIZE = 649081
18 # valid et test =82587 82587
19
20 # change "sandbox" when you're ready
21 JOBDB = 'postgres://ift6266h10@gershwin/ift6266h10_db/rifaisal_csda'
22 EXPERIMENT_PATH = "ift6266.deep.convolutional_dae.salah_exp.nist_csda.jobman_entrypoint"
23
24 ##Pour lancer des travaux sur le cluster: (il faut etre ou se trouve les fichiers)
25 ##python nist_sda.py jobman_insert
26 ##dbidispatch --condor --repeat_jobs=2 jobman sql 'postgres://ift6266h10@gershwin/ift6266h10_db/pannetis_finetuningSDA0' . #C'est le path dans config.py
27
28 # reduce training set to that many examples
29 REDUCE_TRAIN_TO = None
30 # that's a max, it usually doesn't get to that point
31 MAX_FINETUNING_EPOCHS = 1000
32 # number of minibatches before taking means for valid error etc.
33 REDUCE_EVERY = 100
34 #Set the finetune dataset
35 FINETUNE_SET=1
36 #Set the pretrain dataset used. 0: NIST, 1:P07
37 PRETRAIN_CHOICE=1
38
39
40 if TEST_CONFIG:
41 REDUCE_TRAIN_TO = 1000
42 MAX_FINETUNING_EPOCHS = 2
43 REDUCE_EVERY = 10
44
45
46 # This is to configure insertion of jobs on the cluster.
47 # Possible values the hyperparameters can take. These are then
48 # combined with produit_cartesien_jobs so we get a list of all
49 # possible combinations, each one resulting in a job inserted
50 # in the jobman DB.
51
52
53 JOB_VALS = {'pretraining_lr': [0.01],#, 0.001],#, 0.0001],
54 'pretraining_epochs_per_layer': [10],
55 'kernels' : [[[52,5,5], [32,3,3]], [[52,7,7], [52,3,3]]],
56 'mlp_size' : [[1000],[500]],
57 'imgshp' : [[32,32]],
58 'max_pool_layers' : [[[2,2],[2,2]]],
59 'corruption_levels': [[0.2,0.1]],
60 'minibatch_size': [100],
61 'max_finetuning_epochs':[MAX_FINETUNING_EPOCHS],
62 'max_finetuning_epochs_P07':[1000],
63 'finetuning_lr':[0.1,0.01], #0.001 was very bad, so we leave it out
64 'num_hidden_layers':[2],
65 'finetune_set':[1],
66 'pretrain_choice':[1]
67 }
68
69 DEFAULT_HP_NIST = {'pretraining_lr': 0.01,
70 'pretraining_epochs_per_layer': 1,
71 'kernels' : [[4,5,5], [2,3,3]],
72 'mlp_size' : [10],
73 'imgshp' : [32,32],
74 'max_pool_layers' : [[2,2],[2,2]],
75 'corruption_levels': [0.1,0.2],
76 'minibatch_size': 20,
77 'max_finetuning_epochs':MAX_FINETUNING_EPOCHS,
78 'max_finetuning_epochs_P07':1000,
79 'finetuning_lr':0.1, #0.001 was very bad, so we leave it out
80 'num_hidden_layers':2,
81 'finetune_set':1,
82 'pretrain_choice':1,
83 #'reduce_train_to':1000,
84 }
85
86
87
88 ##[pannetis@ceylon test]$ python nist_sda.py test_jobman_entrypoint
89 ##WARNING: untracked file /u/pannetis/IFT6266/ift6266/deep/stacked_dae/v_sylvain/TMP_DBI/configobj.py
90 ##WARNING: untracked file /u/pannetis/IFT6266/ift6266/deep/stacked_dae/v_sylvain/TMP_DBI/utils.py
91 ##WARNING: untracked file /u/pannetis/IFT6266/ift6266/deep/stacked_dae/v_sylvain/config.py
92 ##WARNING: untracked file /u/pannetis/IFT6266/ift6266/deep/stacked_dae/v_sylvain/config2.py
93 ##Creating optimizer with state, DD{'reduce_train_to': 11000, 'pretraining_epochs_per_layer': 2, 'hidden_layers_sizes': 300, 'num_hidden_layers': 2, 'corruption_levels': 0.20000000000000001, 'finetuning_lr': 0.10000000000000001, 'pretrain_choice': 0, 'max_finetuning_epochs': 2, 'version_pylearn': '08b37147dec1', 'finetune_set': -1, 'pretraining_lr': 0.10000000000000001, 'version_ift6266': 'a6b6b1140de9', 'version_theano': 'fb6c3a06cb65', 'minibatch_size': 20}
94 ##SdaSgdOptimizer, max_minibatches = 11000
95 ##C##n_outs 62
96 ##pretrain_lr 0.1
97 ##finetune_lr 0.1
98 ##----
99 ##
100 ##pretraining with NIST
101 ##
102 ##STARTING PRETRAINING, time = 2010-03-29 15:07:43.945981
103 ##Pre-training layer 0, epoch 0, cost 113.562562494
104 ##Pre-training layer 0, epoch 1, cost 113.410032944
105 ##Pre-training layer 1, epoch 0, cost 98.4539954687
106 ##Pre-training layer 1, epoch 1, cost 97.8658966686
107 ##Pretraining took 9.011333 minutes
108 ##
109 ##SERIE OF 3 DIFFERENT FINETUNINGS
110 ##
111 ##
112 ##finetune with NIST
113 ##
114 ##
115 ##STARTING FINETUNING, time = 2010-03-29 15:16:46.512235
116 ##epoch 1, minibatch 4999, validation error on P07 : 29.511250 %
117 ## epoch 1, minibatch 4999, test error on dataset NIST (train data) of best model 40.408509 %
118 ## epoch 1, minibatch 4999, test error on dataset P07 of best model 96.700000 %
119 ##epoch 1, minibatch 9999, validation error on P07 : 25.560000 %
120 ## epoch 1, minibatch 9999, test error on dataset NIST (train data) of best model 34.778969 %
121 ## epoch 1, minibatch 9999, test error on dataset P07 of best model 97.037500 %
122 ##
123 ##Optimization complete with best validation score of 25.560000 %,with test performance 34.778969 % on dataset NIST
124 ##The test score on the P07 dataset is 97.037500
125 ##The finetuning ran for 3.281833 minutes
126 ##
127 ##
128 ##finetune with P07
129 ##
130 ##
131 ##STARTING FINETUNING, time = 2010-03-29 15:20:06.346009
132 ##epoch 1, minibatch 4999, validation error on NIST : 65.226250 %
133 ## epoch 1, minibatch 4999, test error on dataset P07 (train data) of best model 84.465000 %
134 ## epoch 1, minibatch 4999, test error on dataset NIST of best model 65.965237 %
135 ##epoch 1, minibatch 9999, validation error on NIST : 58.745000 %
136 ## epoch 1, minibatch 9999, test error on dataset P07 (train data) of best model 80.405000 %
137 ## epoch 1, minibatch 9999, test error on dataset NIST of best model 61.341923 %
138 ##
139 ##Optimization complete with best validation score of 58.745000 %,with test performance 80.405000 % on dataset P07
140 ##The test score on the NIST dataset is 61.341923
141 ##The finetuning ran for 3.299500 minutes
142 ##
143 ##
144 ##finetune with NIST (done earlier) followed by P07 (written here)
145 ##
146 ##
147 ##STARTING FINETUNING, time = 2010-03-29 15:23:27.947374
148 ##epoch 1, minibatch 4999, validation error on NIST : 83.975000 %
149 ## epoch 1, minibatch 4999, test error on dataset P07 (train data) of best model 83.872500 %
150 ## epoch 1, minibatch 4999, test error on dataset NIST of best model 43.170010 %
151 ##epoch 1, minibatch 9999, validation error on NIST : 79.775000 %
152 ## epoch 1, minibatch 9999, test error on dataset P07 (train data) of best model 80.971250 %
153 ## epoch 1, minibatch 9999, test error on dataset NIST of best model 49.017468 %
154 ##
155 ##Optimization complete with best validation score of 79.775000 %,with test performance 80.971250 % on dataset P07
156 ##The test score on the NIST dataset is 49.017468
157 ##The finetuning ran for 2.851500 minutes
158 ##
159 ##
160 ##finetune with NIST only on the logistic regression on top.
161 ## All hidden units output are input of the logistic regression
162 ##
163 ##
164 ##STARTING FINETUNING, time = 2010-03-29 15:26:21.430557
165 ##epoch 1, minibatch 4999, validation error on P07 : 95.223750 %
166 ## epoch 1, minibatch 4999, test error on dataset NIST (train data) of best model 93.268765 %
167 ## epoch 1, minibatch 4999, test error on dataset P07 of best model 96.535000 %
168 ##epoch 1, minibatch 9999, validation error on P07 : 95.223750 %
169 ##
170 ##Optimization complete with best validation score of 95.223750 %,with test performance 93.268765 % on dataset NIST
171 ##The test score on the P07 dataset is 96.535000
172 ##The finetuning ran for 2.013167 minutes
173 ##Closing remaining open files: /u/pannetis/IFT6266/test/series.h5... done
174 ##[pannetis@ceylon test]$
175
176
177