Mercurial > ift6266

comparison scripts/stacked_dae/nist_sda.py @ 131:5c79a2557f2f

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
Un peu de ménage dans code pour stacked DAE, splitté en fichiers dans un nouveau sous-répertoire.
author savardf
date Fri, 19 Feb 2010 08:43:10 -0500
parents
children 7d8366fb90bf
comparison
equal deleted inserted replaced
130:38929c29b602 131:5c79a2557f2f
1 #!/usr/bin/python
2 # coding: utf-8
3
4 import numpy
5 import theano
6 import time
7 import theano.tensor as T
8 from theano.tensor.shared_randomstreams import RandomStreams
9
10 import os.path
11
12 from sgd_optimization import sgd_optimization
13
14 from jobman import DD
15 from pylearn.io import filetensor
16
17 from utils import produit_croise_jobs
18
19 NIST_ALL_LOCATION = '/data/lisa/data/nist/by_class/all'
20
21 # Just useful for tests... minimal number of epochs
22 DEFAULT_HP_NIST = DD({'finetuning_lr':0.1,
23 'pretraining_lr':0.1,
24 'pretraining_epochs_per_layer':1,
25 'max_finetuning_epochs':1,
26 'hidden_layers_sizes':[1000,1000],
27 'corruption_levels':[0.2,0.2],
28 'minibatch_size':20})
29
30 def jobman_entrypoint_nist(state, channel):
31 sgd_optimization_nist(state)
32
33 def jobman_insert_nist():
34 vals = {'finetuning_lr': [0.00001, 0.0001, 0.001, 0.01, 0.1],
35 'pretraining_lr': [0.00001, 0.0001, 0.001, 0.01, 0.1],
36 'pretraining_epochs_per_layer': [2,5,20],
37 'hidden_layer_sizes': [100,300,1000],
38 'num_hidden_layers':[1,2,3],
39 'corruption_levels': [0.1,0.2,0.4],
40 'minibatch_size': [5,20,100]}
41
42 jobs = produit_croise_jobs(vals)
43
44 for job in jobs:
45 insert_job(job)
46
47
48 class NIST:
49 def __init__(self, minibatch_size, basepath=None):
50 global NIST_ALL_LOCATION
51
52 self.minibatch_size = minibatch_size
53 self.basepath = basepath and basepath or NIST_ALL_LOCATION
54
55 self.set_filenames()
56
57 # arrays of 2 elements: .x, .y
58 self.train = [None, None]
59 self.test = [None, None]
60
61 self.load_train_test()
62
63 self.valid = [[], []]
64 #self.split_train_valid()
65
66
67 def get_tvt(self):
68 return self.train, self.valid, self.test
69
70 def set_filenames(self):
71 self.train_files = ['all_train_data.ft',
72 'all_train_labels.ft']
73
74 self.test_files = ['all_test_data.ft',
75 'all_test_labels.ft']
76
77 def load_train_test(self):
78 self.load_data_labels(self.train_files, self.train)
79 self.load_data_labels(self.test_files, self.test)
80
81 def load_data_labels(self, filenames, pair):
82 for i, fn in enumerate(filenames):
83 f = open(os.path.join(self.basepath, fn))
84 pair[i] = filetensor.read(f)
85 f.close()
86
87 def split_train_valid(self):
88 test_len = len(self.test[0])
89
90 new_train_x = self.train[0][:-test_len]
91 new_train_y = self.train[1][:-test_len]
92
93 self.valid[0] = self.train[0][-test_len:]
94 self.valid[1] = self.train[1][-test_len:]
95
96 self.train[0] = new_train_x
97 self.train[1] = new_train_y
98
99 def test_load_nist():
100 print "Will load NIST"
101
102 import time
103 t1 = time.time()
104 nist = NIST(20)
105 t2 = time.time()
106
107 print "NIST loaded. time delta = ", t2-t1
108
109 tr,v,te = nist.get_tvt()
110
111 print "Lenghts: ", len(tr[0]), len(v[0]), len(te[0])
112
113 raw_input("Press any key")
114
115 # hp for hyperparameters
116 def sgd_optimization_nist(hp=None, dataset_dir='/data/lisa/data/nist'):
117 global DEFAULT_HP_NIST
118 hp = hp and hp or DEFAULT_HP_NIST
119
120 print "Will load NIST"
121
122 import time
123 t1 = time.time()
124 nist = NIST(20)
125 t2 = time.time()
126
127 print "NIST loaded. time delta = ", t2-t1
128
129 train,valid,test = nist.get_tvt()
130 dataset = (train,valid,test)
131
132 print "Lenghts train, valid, test: ", len(train[0]), len(valid[0]), len(test[0])
133
134 n_ins = 32*32
135 n_outs = 62 # 10 digits, 26*2 (lower, capitals)
136
137 sgd_optimization(dataset, hp, n_ins, n_outs)
138
139 if __name__ == '__main__':
140
141 import sys
142
143 args = sys.argv[1:]
144
145 if len(args) > 0 and args[0] == 'load_nist':
146 test_load_nist()
147
148 else:
149 sgd_optimization_nist()
150