Mercurial > ift6266
comparison deep/rbm/rbm.py @ 375:e36ccffb3870
Changes to cast NIST data to floatX for rbm code
| author | Xavier Glorot <glorotxa@iro.umontreal.ca> |
|---|---|
| date | Sun, 25 Apr 2010 14:53:10 -0400 |
| parents | 1e99dc965b5b |
| children |
comparison
equal
deleted
inserted
replaced
| 374:846f0678ffe8 | 375:e36ccffb3870 |
|---|---|
| 271 f = open(data_path+'all/all_train_data.ft') | 271 f = open(data_path+'all/all_train_data.ft') |
| 272 g = open(data_path+'all/all_train_labels.ft') | 272 g = open(data_path+'all/all_train_labels.ft') |
| 273 h = open(data_path+'all/all_test_data.ft') | 273 h = open(data_path+'all/all_test_data.ft') |
| 274 i = open(data_path+'all/all_test_labels.ft') | 274 i = open(data_path+'all/all_test_labels.ft') |
| 275 | 275 |
| 276 train_set_x = theano.shared(ft.read(f)) | 276 train_set_x_uint8 = theano.shared(ft.read(f)) |
| 277 test_set_x_uint8 = theano.shared(ft.read(h)) | |
| 278 | |
| 279 | |
| 280 train_set_x = T.cast(train_set_x_uint8/255.,theano.config.floatX) | |
| 277 train_set_y = ft.read(g) | 281 train_set_y = ft.read(g) |
| 278 test_set_x = ft.read(h) | 282 test_set_x = T.cast(test_set_x_uint8/255.,theano.config.floatX) |
| 279 test_set_y = ft.read(i) | 283 test_set_y = ft.read(i) |
| 280 | 284 |
| 281 f.close() | 285 f.close() |
| 282 g.close() | 286 g.close() |
| 283 i.close() | 287 i.close() |
| 284 h.close() | 288 h.close() |
| 285 | 289 |
| 286 #t = len(train_set_x) | 290 #t = len(train_set_x) |
| 287 print train_set_x.value.shape | |
| 288 | 291 |
| 289 # revoir la recuperation des donnees | 292 # revoir la recuperation des donnees |
| 290 ## dataset = load_data(dataset) | 293 ## dataset = load_data(dataset) |
| 291 ## | 294 ## |
| 292 ## train_set_x, train_set_y = datasets[0] | 295 ## train_set_x, train_set_y = datasets[0] |
| 294 training_epochs = 1 # a determiner | 297 training_epochs = 1 # a determiner |
| 295 | 298 |
| 296 batch_size = b_size # size of the minibatch | 299 batch_size = b_size # size of the minibatch |
| 297 | 300 |
| 298 # compute number of minibatches for training, validation and testing | 301 # compute number of minibatches for training, validation and testing |
| 299 n_train_batches = train_set_x.value.shape[0] / batch_size | 302 n_train_batches = train_set_x_uint8.value.shape[0] / batch_size |
| 300 | 303 |
| 301 # allocate symbolic variables for the data | 304 # allocate symbolic variables for the data |
| 302 index = T.scalar() # index to a [mini]batch | 305 index = T.lscalar() # index to a [mini]batch |
| 303 x = T.matrix('x') # the data is presented as rasterized images | 306 x = T.matrix('x') # the data is presented as rasterized images |
| 304 | 307 |
| 305 rng = numpy.random.RandomState(123) | 308 rng = numpy.random.RandomState(123) |
| 306 theano_rng = RandomStreams( rng.randint(2**30)) | 309 theano_rng = RandomStreams( rng.randint(2**30)) |
| 307 | 310 |
| 330 os.makedirs(dirname) | 333 os.makedirs(dirname) |
| 331 os.chdir(dirname) | 334 os.chdir(dirname) |
| 332 print 'yes' | 335 print 'yes' |
| 333 # it is ok for a theano function to have no output | 336 # it is ok for a theano function to have no output |
| 334 # the purpose of train_rbm is solely to update the RBM parameters | 337 # the purpose of train_rbm is solely to update the RBM parameters |
| 338 print type(batch_size) | |
| 339 print index.dtype | |
| 335 train_rbm = theano.function([index], cost, | 340 train_rbm = theano.function([index], cost, |
| 336 updates = updates, | 341 updates = updates, |
| 337 givens = { x: train_set_x[index*batch_size:(index+1)*batch_size]}) | 342 givens = { x: train_set_x[index*batch_size:(index+1)*batch_size]}) |
| 338 | 343 |
| 339 print 'yep' | 344 print 'yep' |