def run(n_obs, loss, reg, amount):
print 'Starting run of MNIST.py'
mnist_train = sio.loadmat('./data/mnist/MNIST_train.mat')
mnist_test = sio.loadmat('./data/mnist/MNIST_test.mat')
X_train = mnist_train['X'][0][0][2].transpose()
y_train = mnist_train['y']
X_test = mnist_test['Xtest'].transpose()
y_test = mnist_test['ytest']
dim = X_train.shape[1]
## Dictionary
if loss == "l1":
dictmod = adadictl1.AdaDictL1(dim, DICT_ACC, reg, amount)
elif loss == "nl":
dictmod = adadictnolearning.AdaDictNL(dim, DICT_ACC, reg, amount)
else:
dictmod = adadictl2.AdaDictL2(dim, DICT_ACC, reg, amount)
# Train model
dictmod.batchtrain(X_train[range(n_obs)])
# Find reconstructions
al1_train, err = dictmod.batchreconstruction(X_train[range(n_obs)],
'mnist_train')
al1_test, err = dictmod.batchreconstruction(X_test,
'mnist_test')
print dictmod.getnatoms()
## Classification
ogd = multiOGD(10, dictmod.getnatoms(), 0.001)
ogd.train(al1_train , y_train[range(n_obs)])
ogd.train(al1_train , y_train[range(n_obs)])
ogd.seteta(0.0005)
ogd.train(al1_train , y_train[range(n_obs)])
ogd.train(al1_train , y_train[range(n_obs)])
ogd.seteta(0.0001)
ogd.train(al1_train , y_train[range(n_obs)])
ogd.train(al1_train , y_train[range(n_obs)])
ogd.seteta(0.00001)
ogd.train(al1_train , y_train[range(n_obs)])
ogd.train(al1_train , y_train[range(n_obs)])
ogd.train(al1_train , y_train[range(n_obs)])
ogd.seteta(0.000001)
ogd.train(al1_train , y_train[range(n_obs)])
ogd.train(al1_train , y_train[range(n_obs)])
ogd.train(al1_train , y_train[range(n_obs)])
ogd.train(al1_train , y_train[range(n_obs)])
ogd.train(al1_train , y_train[range(n_obs)])
ogd.predict(al1_test, y_test)
imagenm = "mnist_%d" % n_obs
#dictmod.dimagesave((28,28), imagenm)
print 'Run of MNIST.py is complete!\n\n'
""" X_train = mnist_train["X"][0][0][2].transpose() X_test = mnist_test["Xtest"].transpose() n_obs = 60000 X_train = X_train[range(n_obs)] y_train = y_train[range(n_obs)] print "Loading complete" print "\t\t\t====== Dictionary ======" # train model dictmod = multiOGD(10, 784, 0.001) dictmod.train(X_train, y_train) dictmod.seteta(0.0005) dictmod.train(X_train, y_train) dictmod.seteta(0.0001) dictmod.train(X_train, y_train) dictmod.seteta(0.00001) dictmod.train(X_train, y_train) dictmod.train(X_train, y_train) dictmod.seteta(0.000001) dictmod.train(X_train, y_train) dictmod.train(X_train, y_train) # predict dictmod.predict(X_train, y_train)
y_train = mnist_train['y'] X_test = mnist_test['Xtest'].transpose() y_test = mnist_test['ytest'] dim = X_train.shape[1] ## Dictionary lasso_d = dictionary.Dictionary(dim, DICT_ATOMS, DICT_FIT, DICT_REG, \ DICT_INIT) lasso_d.batchtrain(X_train # Save dictionary atoms as images #lasso_d.dimagesave((28, 28), 'mnist') # Find reconstructions alphas_train = lasso_d.batchreconstruction(X_train, \ 'mnist_train_s') alphas_test = lasso_d.batchreconstruction(X_test, \ 'mnist_test_s') ## Classification ogd_m = multiOGD(10, DICT_ATOMS, MOD_REG) ogd_m.train(alphas_train, y_train) ogd_m.predict(alphas_test, y_test) print 'Run of MNIST.py is complete!' ''' Atoms: 200 Reg: 0.05 too much '''