def save_params(n, params):
fname = 'params_epoch_'+str(n)+'.ft'
f = open(fname, 'w')
for p in params:
ft.write(f, p.get_value(borrow=True))
f.close()
def frequency_table():
filenames = ['nist_train_class_freq.ft','p07_train_class_freq.ft','pnist_train_class_freq.ft']
iterators = [datasets.nist_all(),datasets.nist_P07(),datasets.PNIST07()]
for dataset,filename in zip(iterators,filenames):
freq_table = numpy.zeros(62)
for x,y in dataset.train(1):
freq_table[int(y)]+=1
f = open(filename,'w')
ft.write(f,freq_table)
f.close()
def save_mat(fname, mat, save_dir=''):
assert isinstance(mat, numpy.ndarray)
print 'save ndarray to file: ', save_dir + fname
file_handle = open(os.path.join(save_dir, fname), 'w')
filetensor.write(file_handle, mat)
writebool = False
while not writebool:
try:
file_handle.close()
writebool = True
except:
print 'save model error'
time.sleep((numpy.random.randint(10)+2)*10)
def save_mat(fname, mat, save_dir=''):
assert isinstance(mat, numpy.ndarray)
print 'save ndarray to file: ', save_dir + fname
file_handle = open(os.path.join(save_dir, fname), 'w')
filetensor.write(file_handle, mat)
writebool = False
while not writebool:
try:
file_handle.close()
writebool = True
except:
print 'save model error'
time.sleep((numpy.random.randint(10) + 2) * 10)
ensemble valid est trainorig[:80000] ensemble test est trainorig[80000:160000] ensemble train est trainorig[160000:] trainorig est deja shuffled ''' from pylearn.io import filetensor as ft import numpy, os dir1 = '/data/lisa/data/ocr_breuel/filetensor/' dir2 = "/data/lisa/data/ift6266h10/" f = open(dir1 + 'unlv-corrected-2010-02-01-shuffled.ft') d = ft.read(f) f = open(dir2 + "ocr_valid_data.ft", 'wb') ft.write(f, d[:80000]) f = open(dir2 + "ocr_test_data.ft", 'wb') ft.write(f, d[80000:160000]) f = open(dir2 + "ocr_train_data.ft", 'wb') ft.write(f, d[160000:]) f = open(dir1 + 'unlv-corrected-2010-02-01-labels-shuffled.ft') d = ft.read(f) f = open(dir2 + "ocr_valid_labels.ft", 'wb') ft.write(f, d[:80000]) f = open(dir2 + "ocr_test_labels.ft", 'wb') ft.write(f, d[80000:160000]) f = open(dir2 + "ocr_train_labels.ft", 'wb') ft.write(f, d[160000:]) for i in ["train", "valid", "test"]:
def generate_ft_file(data,labels,ft_name):
fdata = open(ft_name+'_data.ft','w')
flabels = open(ft_name+'_labels.ft','w')
ft.write(fdata,data)
ft.write(flabels,labels)
fdata.close();flabels.close()
all_x = []
all_y = []
all_count = 0
for mb_x,mb_y in ds.train(1):
if mb_y[0] <= 9:
all_x.append(mb_x[0])
all_y.append(mb_y[0])
if (all_count+1) % 100000 == 0:
print "Done next 100k"
all_count += 1
# data is stored as uint8 on 0-255
merged_x = numpy.asarray(all_x, dtype=numpy.uint8)
merged_y = numpy.asarray(all_y, dtype=numpy.int32)
print "Kept", len(all_x), "(shape ", merged_x.shape, ") examples from", all_count
f = open(output_data_file, 'wb')
ft.write(f, merged_x)
f.close()
f = open(output_labels_file, 'wb')
ft.write(f, merged_y)
f.close()
from pylearn.io import filetensor as ft
import pylab
import random as r
from numpy import *
nombre = 20000 #La grandeur de l'echantillon
f = open('all_train_data.ft') #Le jeu de donnees est en local.
d = ft.read(f)
f.close()
print len(d)
random.seed(3525)
echantillon=r.sample(xrange(len(d)),nombre)
nouveau=d[0:nombre]
for i in xrange(nombre):
nouveau[i]=d[echantillon[i]]
f2 = open('echantillon_occlusion.ft', 'w')
ft.write(f2,nouveau)
f2.close()
##Tester si ca a fonctionne
f3 = open('echantillon_occlusion.ft')
d2=ft.read(f3)
pylab.imshow(d2[0].reshape((32,32)))
pylab.show()
f3.close()
