Mercurial > ift6266

comparison deep/stacked_dae/v_sylvain/nist_sda.py @ 327:4306796d60a8

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
Rajout de la capacite de faire decroitre le taux d'apprentissage si demande
author SylvainPL <sylvain.pannetier.lebeuf@umontreal.ca>
date Sun, 11 Apr 2010 19:52:28 -0400
parents 7a12d2c3d06b
children ffc06af1c543
comparison
equal deleted inserted replaced
326:b762ac18a2d7 327:4306796d60a8
48 rtt = None 48 rtt = None
49 if state.has_key('reduce_train_to'): 49 if state.has_key('reduce_train_to'):
50 rtt = state['reduce_train_to'] 50 rtt = state['reduce_train_to']
51 elif REDUCE_TRAIN_TO: 51 elif REDUCE_TRAIN_TO:
52 rtt = REDUCE_TRAIN_TO 52 rtt = REDUCE_TRAIN_TO
53
54 if state.has_key('decrease_lr'):
55 decrease_lr = state['decrease_lr']
56 else :
57 decrease_lr = 0
53 58
54 n_ins = 32*32 59 n_ins = 32*32
55 n_outs = 62 # 10 digits, 26*2 (lower, capitals) 60 n_outs = 62 # 10 digits, 26*2 (lower, capitals)
56 61
57 examples_per_epoch = NIST_ALL_TRAIN_SIZE 62 examples_per_epoch = NIST_ALL_TRAIN_SIZE
114 #Decide how the finetune is done 119 #Decide how the finetune is done
115 120
116 if finetune_choice == 0: 121 if finetune_choice == 0:
117 print('\n\n\tfinetune with NIST\n\n') 122 print('\n\n\tfinetune with NIST\n\n')
118 optimizer.reload_parameters('params_pretrain.txt') 123 optimizer.reload_parameters('params_pretrain.txt')
119 optimizer.finetune(datasets.nist_all(),datasets.nist_P07(),max_finetune_epoch_NIST,ind_test=1) 124 optimizer.finetune(datasets.nist_all(),datasets.nist_P07(),max_finetune_epoch_NIST,ind_test=1,decrease=decrease_lr)
120 channel.save() 125 channel.save()
121 if finetune_choice == 1: 126 if finetune_choice == 1:
122 print('\n\n\tfinetune with P07\n\n') 127 print('\n\n\tfinetune with P07\n\n')
123 optimizer.reload_parameters('params_pretrain.txt') 128 optimizer.reload_parameters('params_pretrain.txt')
124 optimizer.finetune(datasets.nist_P07(),datasets.nist_all(),max_finetune_epoch_P07,ind_test=0) 129 optimizer.finetune(datasets.nist_P07(),datasets.nist_all(),max_finetune_epoch_P07,ind_test=0,decrease=decrease_lr)
125 channel.save() 130 channel.save()
126 if finetune_choice == 2: 131 if finetune_choice == 2:
127 print('\n\n\tfinetune with P07 followed by NIST\n\n') 132 print('\n\n\tfinetune with P07 followed by NIST\n\n')
128 optimizer.reload_parameters('params_pretrain.txt') 133 optimizer.reload_parameters('params_pretrain.txt')
129 optimizer.finetune(datasets.nist_P07(),datasets.nist_all(),max_finetune_epoch_P07,ind_test=20) 134 optimizer.finetune(datasets.nist_P07(),datasets.nist_all(),max_finetune_epoch_P07,ind_test=20,decrease=decrease_lr)
130 optimizer.finetune(datasets.nist_all(),datasets.nist_P07(),max_finetune_epoch_NIST,ind_test=21) 135 optimizer.finetune(datasets.nist_all(),datasets.nist_P07(),max_finetune_epoch_NIST,ind_test=21,decrease=decrease_lr)
131 channel.save() 136 channel.save()
132 if finetune_choice == 3: 137 if finetune_choice == 3:
133 print('\n\n\tfinetune with NIST only on the logistic regression on top (but validation on P07).\n\ 138 print('\n\n\tfinetune with NIST only on the logistic regression on top (but validation on P07).\n\
134 All hidden units output are input of the logistic regression\n\n') 139 All hidden units output are input of the logistic regression\n\n')
135 optimizer.reload_parameters('params_pretrain.txt') 140 optimizer.reload_parameters('params_pretrain.txt')
136 optimizer.finetune(datasets.nist_all(),datasets.nist_P07(),max_finetune_epoch_NIST,ind_test=1,special=1) 141 optimizer.finetune(datasets.nist_all(),datasets.nist_P07(),max_finetune_epoch_NIST,ind_test=1,special=1,decrease=decrease_lr)
137 142
138 143
139 if finetune_choice==-1: 144 if finetune_choice==-1:
140 print('\nSERIE OF 4 DIFFERENT FINETUNINGS') 145 print('\nSERIE OF 4 DIFFERENT FINETUNINGS')
141 print('\n\n\tfinetune with NIST\n\n') 146 print('\n\n\tfinetune with NIST\n\n')
142 sys.stdout.flush() 147 sys.stdout.flush()
143 optimizer.reload_parameters('params_pretrain.txt') 148 optimizer.reload_parameters('params_pretrain.txt')
144 optimizer.finetune(datasets.nist_all(),datasets.nist_P07(),max_finetune_epoch_NIST,ind_test=1) 149 optimizer.finetune(datasets.nist_all(),datasets.nist_P07(),max_finetune_epoch_NIST,ind_test=1,decrease=decrease_lr)
145 channel.save() 150 channel.save()
146 print('\n\n\tfinetune with P07\n\n') 151 print('\n\n\tfinetune with P07\n\n')
147 sys.stdout.flush() 152 sys.stdout.flush()
148 optimizer.reload_parameters('params_pretrain.txt') 153 optimizer.reload_parameters('params_pretrain.txt')
149 optimizer.finetune(datasets.nist_P07(),datasets.nist_all(),max_finetune_epoch_P07,ind_test=0) 154 optimizer.finetune(datasets.nist_P07(),datasets.nist_all(),max_finetune_epoch_P07,ind_test=0,decrease=decrease_lr)
150 channel.save() 155 channel.save()
151 print('\n\n\tfinetune with P07 (done earlier) followed by NIST (written here)\n\n') 156 print('\n\n\tfinetune with P07 (done earlier) followed by NIST (written here)\n\n')
152 sys.stdout.flush() 157 sys.stdout.flush()
153 optimizer.reload_parameters('params_finetune_P07.txt') 158 optimizer.reload_parameters('params_finetune_P07.txt')
154 optimizer.finetune(datasets.nist_all(),datasets.nist_P07(),max_finetune_epoch_NIST,ind_test=21) 159 optimizer.finetune(datasets.nist_all(),datasets.nist_P07(),max_finetune_epoch_NIST,ind_test=21,decrease=decrease_lr)
155 channel.save() 160 channel.save()
156 print('\n\n\tfinetune with NIST only on the logistic regression on top.\n\ 161 print('\n\n\tfinetune with NIST only on the logistic regression on top.\n\
157 All hidden units output are input of the logistic regression\n\n') 162 All hidden units output are input of the logistic regression\n\n')
158 sys.stdout.flush() 163 sys.stdout.flush()
159 optimizer.reload_parameters('params_pretrain.txt') 164 optimizer.reload_parameters('params_pretrain.txt')
160 optimizer.finetune(datasets.nist_all(),datasets.nist_P07(),max_finetune_epoch_NIST,ind_test=1,special=1) 165 optimizer.finetune(datasets.nist_all(),datasets.nist_P07(),max_finetune_epoch_NIST,ind_test=1,special=1,decrease=decrease_lr)
161 channel.save() 166 channel.save()
162 167
163 channel.save() 168 channel.save()
164 169
165 return channel.COMPLETE 170 return channel.COMPLETE