def load_mnist(lazy):
"""
Load mnist dataset from a hdf5 file and test if it matches mlpython's one.
"""
dataset_name = 'mnist'
start = time.time()
import mlpython.datasets.store as mlstore
mldatasets = mlstore.get_classification_problem(dataset_name, load_to_memory= (not lazy))
print "mlpython version loaded ({0:.2f}sec).".format(time.time() - start)
start = time.time()
dataset_name = os.path.join(os.environ['MLPYTHON_DATASET_REPO'], dataset_name + ".h5")
dataset = mldata.dataset_store.load(dataset_name, lazy=lazy)
print "mldata version loaded ({0:.2f}sec).".format(time.time() - start)
print "Comparing first 1000..."
count = 0
for (e1, t1), (e2, t2) in itertools.izip(dataset, itertools.chain(*mldatasets)):
#print t1, t2
assert_array_almost_equal(e1, e2)
assert_equal(t1, t2)
count += 1
if count >= 1000:
break
def train():
sys.argv.pop(0); # Remove first argument
# Check if every option(s) from parent's script are here.
if 5 != len(sys.argv):
print "Usage: python run_stacked_autoencoders_nnet.py lr hidden_size n_epochs n_cdk seed"
print ""
print "Ex.: python run_stacked_autoencoders_nnet.py 0.01 50 10 10 1234"
sys.exit()
# Set the constructor
str_ParamOption = "lr=" + sys.argv[0] + ", " + "hidden_size=" + sys.argv[1] + ", " + "n_epochs=" + sys.argv[2] + ", " +\
"CDk=" + sys.argv[3] + ", " + "seed=" + sys.argv[4]
try:
objectString = 'myObject = RBM(' + str_ParamOption + ')'
exec objectString
#code = compile(objectString, '<string>', 'exec')
#exec code
except Exception as inst:
print "Error while instantiating RBM (required hyper-parameters are probably missing)"
print inst
print "Loading dataset..."
trainset,validset,testset = dataset_store.get_classification_problem('ocr_letters')
print "Training..."
myObject.train(trainset)
#Store the trained dictionary and the parameters to a file.
pickle.dump((myObject.W, myObject.b, myObject.hidden_size), open("Models/RBM/model%d.pkl"%experiment_number, 'wb'))
def get_representation():
# Load the dictionary and corresponding args.
(W, b, hidden_size) = pickle.load(open("Models/RBM/model%d.pkl"%experiment_number,'rb'))
# Set the constructor
myObject = RBM(hidden_size=hidden_size)
print "Loading dataset..."
trainset,validset,testset = dataset_store.get_classification_problem('ocr_letters')
encoded_trainset = []
encoded_validset = []
encoded_testset = []
print "Initializing..."
myObject.initialize(W,b)
print "Encoding the trainset..."
counter = 0 #Inelegant, I know! I use this to only use the first 1000 values.
for input,target in trainset:
#Encode the sample.
h = myObject.encode(input)
encoded_trainset.append(h)
# counter +=1
# if counter == 1000:
# break
# Save the datasets to files.
filename = "Models/RBM/trainset%d.pkl"%(experiment_number)
pickle.dump( np.asarray(encoded_trainset) , open(filename, 'wb'))
counter = 0
print "Encoding the validset..."
for input,target in validset:
#Encode the sample.
h = myObject.encode(input)
encoded_validset.append(h)
# counter +=1
# if counter == 1000:
# break
filename = "Models/RBM/validset%d.pkl"%(experiment_number)
pickle.dump( np.asarray(encoded_validset) , open(filename, 'wb'))
#Note: only need to do it for the best hyper-params at the end.
print "Encoding the testset..."
for input,target in testset:
#Encode the sample.
h = myObject.encode(input)
encoded_testset.append(h)
filename = "Models/RBM/testset%d.pkl"%(experiment_number)
pickle.dump( np.asarray(encoded_testset), open(filename, 'wb'))
def sklearn_convex(classifier,
algorithm,
max_evals=100,
seed=1,
filename='none',
preproc=[],
loss=None):
global suppress_output
if suppress_output:
dump_file = None
else:
dump_file = filename + '.dump'
estim = hyperopt_estimator(classifier=classifier,
algo=algorithm,
preprocessing=preproc,
max_evals=max_evals,
trial_timeout=240,
fit_increment_dump_filename=dump_file,
loss_fn=loss)
filename = filename + '.out'
dataset_store.download('convex')
trainset, validset, testset = dataset_store.get_classification_problem(
'convex')
X_train = trainset.data.mem_data[0]
y_train = trainset.data.mem_data[1]
X_valid = validset.data.mem_data[0]
y_valid = validset.data.mem_data[1]
X_test = testset.data.mem_data[0]
y_test = testset.data.mem_data[1]
X_fulltrain = np.concatenate((X_train, X_valid))
y_fulltrain = np.concatenate((y_train, y_valid))
print(y_train.shape)
print(y_valid.shape)
print(y_test.shape)
#find_model( X_train, y_train, X_test, y_test, estim, filename )
find_model(X_fulltrain, y_fulltrain, X_test, y_test, estim, filename)
def convex():
dataset_store.download('convex')
trainset,validset,testset = dataset_store.get_classification_problem('convex')
X_train = trainset.data.mem_data[0]
y_train = trainset.data.mem_data[1]
X_valid = validset.data.mem_data[0]
y_valid = validset.data.mem_data[1]
X_test = testset.data.mem_data[0]
y_test = testset.data.mem_data[1]
X_fulltrain = np.concatenate((X_train, X_valid))
y_fulltrain = np.concatenate((y_train, y_valid))
pca = PCA()
X_train_pca = pca.fit_transform( X_fulltrain )
X_test_pca = pca.fit_transform( X_test )
clfs = [ MultinomialNB(), SVC(),
KNeighborsClassifier(),
SGDClassifier() ]
pca_clfs = [ SVC(),
KNeighborsClassifier(),
SGDClassifier() ]
print("Convex\n")
with open( "convex_baselines.txt", 'w' ) as f:
for clf in clfs:
clf.fit( X_fulltrain, y_fulltrain )
pred = clf.predict( X_test )
score = metrics.f1_score( y_test, pred )
print( "Classifier: %s\nScore: %f\n" % (clf, score) )
f.write("Classifier: %s\nScore: %f\n\n" % (clf, score))
for clf in pca_clfs:
clf.fit( X_train_pca, y_fulltrain )
pred = clf.predict( X_test_pca )
score = metrics.f1_score( y_test, pred )
print( "Classifier: PCA + %s\nScore: %f\n" % (clf, score) )
f.write("Classifier: PCA + %s\nScore: %f\n\n" % (clf, score))
def convex():
dataset_store.download('convex')
trainset, validset, testset = dataset_store.get_classification_problem(
'convex')
X_train = trainset.data.mem_data[0]
y_train = trainset.data.mem_data[1]
X_valid = validset.data.mem_data[0]
y_valid = validset.data.mem_data[1]
X_test = testset.data.mem_data[0]
y_test = testset.data.mem_data[1]
X_fulltrain = np.concatenate((X_train, X_valid))
y_fulltrain = np.concatenate((y_train, y_valid))
pca = PCA()
X_train_pca = pca.fit_transform(X_fulltrain)
X_test_pca = pca.fit_transform(X_test)
clfs = [MultinomialNB(), SVC(), KNeighborsClassifier(), SGDClassifier()]
pca_clfs = [SVC(), KNeighborsClassifier(), SGDClassifier()]
print("Convex\n")
with open("convex_baselines.txt", 'w') as f:
for clf in clfs:
clf.fit(X_fulltrain, y_fulltrain)
pred = clf.predict(X_test)
score = metrics.f1_score(y_test, pred)
print("Classifier: %s\nScore: %f\n" % (clf, score))
f.write("Classifier: %s\nScore: %f\n\n" % (clf, score))
for clf in pca_clfs:
clf.fit(X_train_pca, y_fulltrain)
pred = clf.predict(X_test_pca)
score = metrics.f1_score(y_test, pred)
print("Classifier: PCA + %s\nScore: %f\n" % (clf, score))
f.write("Classifier: PCA + %s\nScore: %f\n\n" % (clf, score))
def sklearn_convex( classifier, algorithm, max_evals=100, seed=1,
filename = 'none', preproc=[], loss=None ):
global suppress_output
if suppress_output:
dump_file = None
else:
dump_file = filename+'.dump'
estim = hyperopt_estimator( classifier=classifier, algo=algorithm,
preprocessing=preproc,
max_evals=max_evals, trial_timeout=240,
fit_increment_dump_filename=dump_file,
loss_fn=loss)
filename = filename + '.out'
dataset_store.download('convex')
trainset,validset,testset = dataset_store.get_classification_problem('convex')
X_train = trainset.data.mem_data[0]
y_train = trainset.data.mem_data[1]
X_valid = validset.data.mem_data[0]
y_valid = validset.data.mem_data[1]
X_test = testset.data.mem_data[0]
y_test = testset.data.mem_data[1]
X_fulltrain = np.concatenate((X_train, X_valid))
y_fulltrain = np.concatenate((y_train, y_valid))
print(y_train.shape)
print(y_valid.shape)
print(y_test.shape)
#find_model( X_train, y_train, X_test, y_test, estim, filename )
find_model( X_fulltrain, y_fulltrain, X_test, y_test, estim, filename )
def get_dictionary():
"""
Train the sparse coding model
and save the dictionary and params to
a file.
"""
sys.argv.pop(0); # Remove first argument
# Check if every option(s) from parent's script are here.
if 5 != len(sys.argv):
print "Usage: python run_sparse_code.py lr size L1 n_epochs seed"
print ""
print "Ex.: python run_sparse_code.py 0.1 20 0.1 5 1234"
sys.exit()
# Set the constructor
str_ParamOption = "lr=" + sys.argv[0] + ", " + "size=" + sys.argv[1] + ", " + "L1=" + sys.argv[2] + ", " + "n_epochs=" + sys.argv[3] + ", " + "seed=" + sys.argv[4]
str_ParamOptionValue = sys.argv[0] + "\t" + sys.argv[1] + "\t" + sys.argv[2] + "\t" + sys.argv[3] + "\t" + sys.argv[4]
try:
objectString = 'myObject = SparseCode(' + str_ParamOption + ')'
exec objectString
#code = compile(objectString, '<string>', 'exec')
#exec code
except Exception as inst:
print "Error while instantiating SparseCode (required hyper-parameters are probably missing)"
print inst
print "Loading dataset..."
trainset,validset,testset = dataset_store.get_classification_problem('ocr_letters')
print "Training..."
myObject.train(trainset)
#Store the trained dictionary and the parameters to a file.
pickle.dump((myObject.dictionary, myObject.lr, myObject.hidden_size, myObject.L1), open("Models/SC/dictionary%d.pkl"%experiment_number, 'wb'))
myObject.show_filters()
import os
import itertools
import numpy as np
import fcntl
import copy
from string import Template
import mlpython.datasets.store as dataset_store
import mlpython.mlproblems.generic as mlpb
from rbm import RBM
#from autoencoder import Autoencoder
print "Loading dataset..."
trainset,validset,testset = dataset_store.get_classification_problem('ocr_letters')
print "Train RBM for 10 iterations... (this might take a few minutes)"
rbm = RBM(n_epochs = 10,
hidden_size = 200,
lr = 0.01,
CDk = 1,
seed=1234
)
rbm.train(mlpb.SubsetFieldsProblem(trainset))
rbm.show_filters()
# Set the constructor
str_ParamOption = "lr=" + sys.argv[0] + ", " + "dc=" + sys.argv[1] + ", " + "sizes=" + sys.argv[2] + ", " + "L2=" + \
sys.argv[3] + ", " + "L1=" + sys.argv[4] + ", " + "seed=" + sys.argv[5] + ", " + "tanh=" + sys.argv[6]
str_ParamOptionValue = sys.argv[0] + "\t" + sys.argv[1] + "\t" + sys.argv[2] + "\t" + sys.argv[3] + "\t" + sys.argv[
4] + "\t" + sys.argv[5] + "\t" + sys.argv[6]
try:
objectString = 'myObject = NeuralNetwork(n_epochs=1,' + str_ParamOption + ')'
exec objectString
# code = compile(objectString, '<string>', 'exec')
# exec code
except Exception as inst:
print "Error while instantiating NeuralNetwork (required hyper-parameters are probably missing)"
print inst
print "Loading dataset..."
trainset, validset, testset = dataset_store.get_classification_problem('ocr_letters')
print "Training..."
# Early stopping code
best_val_error = np.inf
best_it = 0
str_header = 'best_it\t'
look_ahead = 5
n_incr_error = 0
for stage in range(1, 500 + 1, 1):
if not n_incr_error < look_ahead:
break
myObject.n_epochs = stage
myObject.train(trainset)
n_incr_error += 1
outputs, costs = myObject.test(trainset)
errors = np.mean(costs, axis=0)
def get_representation():
"""
Grab the dictionary, convert
the datasets to a sparse representation and
save them to a file.
"""
# Load the dictionary and corresponding args.
(dictionary, lr, hidden_size, L1) = pickle.load(open("Models/SC/dictionary%d.pkl"%experiment_number,'rb'))
# Set the constructor
myObject = SparseCode(lr,hidden_size,L1)
print "Loading dataset..."
trainset,validset,testset = dataset_store.get_classification_problem('ocr_letters')
encoded_trainset = []
trainset_out = []
encoded_validset = []
validset_out = []
encoded_testset = []
testset_out = []
print "Initializing..."
myObject.initialize_dictionary(dictionary)
print "Encoding the trainset..."
#counter = 0 #Inelegant, I know! I use this to only use the first 1000 values.
for input,target in trainset:
#Run ISTA
h = myObject.infer(input)
encoded_trainset.append(h)
trainset_out.append(target)
# counter +=1
# if counter == 1000:
# break
# Save the datasets to files.
filename = "Models/SC/trainset%d.pkl"%(experiment_number)
pickle.dump( np.asarray(encoded_trainset) , open(filename, 'wb'))
filename = "Models/train_outputs.pkl"
pickle.dump( np.asarray(trainset_out) , open(filename, 'wb'))
#counter = 0
print "Encoding the validset..."
for input,target in validset:
#Run ISTA
h = myObject.infer(input)
encoded_validset.append(h)
validset_out.append(target)
# counter +=1
# if counter == 1000:
# break
filename = "Models/SC/validset%d.pkl"%(experiment_number)
pickle.dump( np.asarray(encoded_validset) , open(filename, 'wb'))
filename = "Models/valid_outputs.pkl"
pickle.dump( np.asarray(validset_out) , open(filename, 'wb'))
Note: only need to do it for the best hyper-params at the end.
print "Encoding the testset..."
for input,target in testset:
#Run ISTA
h = myObject.infer(input)
encoded_testset.append(h)
testset_out.append(target)
filename = "Models/SC/testset%d.pkl"%(experiment_number)
pickle.dump( np.asarray(encoded_testset), open(filename, 'wb'))
filename = "Models/test_outputs.pkl"
pickle.dump( np.asarray(testset_out) , open(filename, 'wb'))
