def syntetic_test():
# test model on different train set size & on different train sets
results = np.zeros((18, 5))
full_labeled = np.array([2, 4, 10, 25, 100])
train_size = 400
for dataset in xrange(1, 19):
X, Y = load_syntetic(dataset)
for j, nfull in enumerate(full_labeled):
crf = EdgeCRF(n_states=10, n_features=10, n_edge_features=2,
inference_method='qpbo')
clf = OneSlackSSVM(crf, max_iter=10000, C=0.01, verbose=0,
tol=0.1, n_jobs=4, inference_cache=100)
x_train = X[:nfull]
y_train = Y[:nfull]
x_test = X[(train_size + 1):]
y_test = Y[(train_size + 1):]
try:
clf.fit(x_train, y_train)
y_pred = clf.predict(x_test)
results[dataset - 1, j] = compute_error(y_test, y_pred)
print 'dataset=%d, nfull=%d, error=%f' % (dataset, nfull,
results[dataset - 1, j])
except ValueError:
print 'dataset=%d, nfull=%d: Failed' % (dataset, nfull)
np.savetxt('results/syntetic/full_labeled.txt', results)
def test_one_slack_constraint_caching():
# testing cutting plane ssvm on easy multinomial dataset
X, Y = generate_blocks_multinomial(n_samples=10, noise=0.5, seed=0,
size_x=9)
n_labels = len(np.unique(Y))
exact_inference = get_installed([('ad3', {'branch_and_bound': True}), "lp"])[0]
crf = GridCRF(n_states=n_labels, inference_method=exact_inference)
clf = OneSlackSSVM(model=crf, max_iter=150, C=1,
check_constraints=True, break_on_bad=True,
inference_cache=50, inactive_window=0)
clf.fit(X, Y)
Y_pred = clf.predict(X)
assert_array_equal(Y, Y_pred)
assert_equal(len(clf.inference_cache_), len(X))
# there should be 13 constraints, which are less than the 94 iterations
# that are done
# check that we didn't change the behavior of how we construct the cache
constraints_per_sample = [len(cache) for cache in clf.inference_cache_]
if exact_inference == "lp":
assert_equal(len(clf.inference_cache_[0]), 18)
assert_equal(np.max(constraints_per_sample), 18)
assert_equal(np.min(constraints_per_sample), 18)
else:
assert_equal(len(clf.inference_cache_[0]), 13)
assert_equal(np.max(constraints_per_sample), 20)
assert_equal(np.min(constraints_per_sample), 11)
def test_latent_node_boxes_edge_features():
# learn the "easy" 2x2 boxes dataset.
# smoketest using a single constant edge feature
X, Y = make_simple_2x2(seed=1, n_samples=40)
latent_crf = EdgeFeatureLatentNodeCRF(n_labels=2, n_hidden_states=2, n_features=1)
base_svm = OneSlackSSVM(latent_crf)
base_svm.C = 10
latent_svm = LatentSSVM(base_svm,
latent_iter=10)
G = [make_grid_edges(x) for x in X]
# make edges for hidden states:
edges = make_edges_2x2()
G = [np.vstack([make_grid_edges(x), edges]) for x in X]
# reshape / flatten x and y
X_flat = [x.reshape(-1, 1) for x in X]
Y_flat = [y.ravel() for y in Y]
#X_ = zip(X_flat, G, [2 * 2 for x in X_flat])
# add edge features
X_ = [(x, g, np.ones((len(g), 1)), 4) for x, g in zip(X_flat, G)]
latent_svm.fit(X_[:20], Y_flat[:20])
assert_array_equal(latent_svm.predict(X_[:20]), Y_flat[:20])
assert_equal(latent_svm.score(X_[:20], Y_flat[:20]), 1)
# test that score is not always 1
assert_true(.98 < latent_svm.score(X_[20:], Y_flat[20:]) < 1)
def test_standard_svm_blobs_2d_class_weight():
# no edges, reduce to crammer-singer svm
X, Y = make_blobs(n_samples=210,
centers=3,
random_state=1,
cluster_std=3,
shuffle=False)
X = np.hstack([X, np.ones((X.shape[0], 1))])
X, Y = X[:170], Y[:170]
X_graphs = [(x[np.newaxis, :], np.empty((0, 2), dtype=np.int)) for x in X]
pbl = GraphCRF(n_features=3, n_states=3, inference_method='unary')
svm = OneSlackSSVM(pbl, check_constraints=False, C=1000)
svm.fit(X_graphs, Y[:, np.newaxis])
weights = 1. / np.bincount(Y)
weights *= len(weights) / np.sum(weights)
pbl_class_weight = GraphCRF(n_features=3,
n_states=3,
class_weight=weights,
inference_method='unary')
svm_class_weight = OneSlackSSVM(pbl_class_weight,
C=10,
check_constraints=False,
break_on_bad=False)
svm_class_weight.fit(X_graphs, Y[:, np.newaxis])
assert_greater(f1_score(Y, np.hstack(svm_class_weight.predict(X_graphs))),
f1_score(Y, np.hstack(svm.predict(X_graphs))))
def test_class_weights_rescale_C():
# check that our crammer-singer implementation with class weights and
# rescale_C=True is the same as LinearSVC's c-s class_weight implementation
raise SkipTest("class weight test needs update")
from sklearn.svm import LinearSVC
X, Y = make_blobs(n_samples=210,
centers=3,
random_state=1,
cluster_std=3,
shuffle=False)
X = np.hstack([X, np.ones((X.shape[0], 1))])
X, Y = X[:170], Y[:170]
weights = len(Y) / (np.bincount(Y) * len(np.unique(Y)))
pbl_class_weight = MultiClassClf(n_features=3,
n_classes=3,
class_weight=weights,
rescale_C=True)
svm_class_weight = OneSlackSSVM(pbl_class_weight, C=10, tol=1e-5)
svm_class_weight.fit(X, Y)
try:
linearsvm = LinearSVC(multi_class='crammer_singer',
fit_intercept=False,
class_weight='balanced',
C=10)
linearsvm.fit(X, Y)
assert_array_almost_equal(svm_class_weight.w, linearsvm.coef_.ravel(),
3)
except TypeError:
# travis has a really old sklearn version that doesn't support
# class_weight in LinearSVC
pass
def msrc():
models_basedir = 'models/msrc/'
crf = EdgeCRF(n_states=24, n_features=2028, n_edge_features=4,
inference_method='gco')
clf = OneSlackSSVM(crf, max_iter=10000, C=0.01, verbose=2,
tol=0.1, n_jobs=4,
inference_cache=100)
X, Y = load_msrc('train')
Y = remove_areas(Y)
start = time()
clf.fit(X, Y)
stop = time()
np.savetxt(models_basedir + 'msrc_full.csv', clf.w)
with open(models_basedir + 'msrc_full' + '.pickle', 'w') as f:
pickle.dump(clf, f)
X, Y = load_msrc('test')
Y = remove_areas(Y)
Y_pred = clf.predict(X)
print('Error on test set: %f' % compute_error(Y, Y_pred))
print('Score on test set: %f' % clf.score(X, Y))
print('Norm of weight vector: |w|=%f' % np.linalg.norm(clf.w))
print('Elapsed time: %f s' % (stop - start))
return clf
def test_class_weights_rescale_C():
# check that our crammer-singer implementation with class weights and
# rescale_C=True is the same as LinearSVC's c-s class_weight implementation
from sklearn.svm import LinearSVC
X, Y = make_blobs(n_samples=210, centers=3, random_state=1, cluster_std=3,
shuffle=False)
X = np.hstack([X, np.ones((X.shape[0], 1))])
X, Y = X[:170], Y[:170]
weights = 1. / np.bincount(Y)
weights *= len(weights) / np.sum(weights)
pbl_class_weight = MultiClassClf(n_features=3, n_classes=3,
class_weight=weights, rescale_C=True)
svm_class_weight = OneSlackSSVM(pbl_class_weight, C=10, tol=1e-5)
svm_class_weight.fit(X, Y)
try:
linearsvm = LinearSVC(multi_class='crammer_singer',
fit_intercept=False, class_weight='auto', C=10)
linearsvm.fit(X, Y)
assert_array_almost_equal(svm_class_weight.w, linearsvm.coef_.ravel(),
3)
except TypeError:
# travis has a really old sklearn version that doesn't support
# class_weight in LinearSVC
pass
def test_one_slack_constraint_caching():
# testing cutting plane ssvm on easy multinomial dataset
X, Y = generate_blocks_multinomial(n_samples=10, noise=0.5, seed=0,
size_x=9)
n_labels = len(np.unique(Y))
exact_inference = get_installed([('ad3', {'branch_and_bound': True}), "lp"])[0]
crf = GridCRF(n_states=n_labels, inference_method=exact_inference)
clf = OneSlackSSVM(model=crf, max_iter=150, C=1,
check_constraints=True, break_on_bad=True,
inference_cache=50, inactive_window=0)
clf.fit(X, Y)
Y_pred = clf.predict(X)
assert_array_equal(Y, Y_pred)
assert_equal(len(clf.inference_cache_), len(X))
# there should be 13 constraints, which are less than the 94 iterations
# that are done
# check that we didn't change the behavior of how we construct the cache
constraints_per_sample = [len(cache) for cache in clf.inference_cache_]
if exact_inference == "lp":
assert_equal(len(clf.inference_cache_[0]), 18)
assert_equal(np.max(constraints_per_sample), 18)
assert_equal(np.min(constraints_per_sample), 18)
else:
assert_equal(len(clf.inference_cache_[0]), 13)
assert_equal(np.max(constraints_per_sample), 20)
assert_equal(np.min(constraints_per_sample), 11)
def test_latent_node_boxes_edge_features():
# learn the "easy" 2x2 boxes dataset.
# smoketest using a single constant edge feature
X, Y = make_simple_2x2(seed=1, n_samples=40)
latent_crf = EdgeFeatureLatentNodeCRF(n_labels=2,
n_hidden_states=2,
n_features=1)
base_svm = OneSlackSSVM(latent_crf)
base_svm.C = 10
latent_svm = LatentSSVM(base_svm, latent_iter=10)
G = [make_grid_edges(x) for x in X]
# make edges for hidden states:
edges = make_edges_2x2()
G = [np.vstack([make_grid_edges(x), edges]) for x in X]
# reshape / flatten x and y
X_flat = [x.reshape(-1, 1) for x in X]
Y_flat = [y.ravel() for y in Y]
#X_ = zip(X_flat, G, [2 * 2 for x in X_flat])
# add edge features
X_ = [(x, g, np.ones((len(g), 1)), 4) for x, g in zip(X_flat, G)]
latent_svm.fit(X_[:20], Y_flat[:20])
assert_array_equal(latent_svm.predict(X_[:20]), Y_flat[:20])
assert_equal(latent_svm.score(X_[:20], Y_flat[:20]), 1)
# test that score is not always 1
assert_true(.98 < latent_svm.score(X_[20:], Y_flat[20:]) < 1)
def test_one_slack_constraint_caching():
#testing cutting plane ssvm on easy multinomial dataset
X, Y = generate_blocks_multinomial(n_samples=10,
noise=0.5,
seed=0,
size_x=9)
n_labels = len(np.unique(Y))
crf = GridCRF(n_states=n_labels, inference_method='lp')
clf = OneSlackSSVM(model=crf,
max_iter=150,
C=1,
check_constraints=True,
break_on_bad=True,
inference_cache=50,
inactive_window=0,
verbose=10)
clf.fit(X, Y)
Y_pred = clf.predict(X)
assert_array_equal(Y, Y_pred)
assert_equal(len(clf.inference_cache_), len(X))
# there should be 11 constraints, which are less than the 94 iterations
# that are done
assert_equal(len(clf.inference_cache_[0]), 11)
# check that we didn't change the behavior of how we construct the cache
constraints_per_sample = [len(cache) for cache in clf.inference_cache_]
assert_equal(np.max(constraints_per_sample), 19)
assert_equal(np.min(constraints_per_sample), 11)
def test_binary_blocks_one_slack_graph():
#testing cutting plane ssvm on easy binary dataset
# generate graphs explicitly for each example
X, Y = generate_blocks(n_samples=3)
crf = GraphCRF(inference_method=inference_method)
clf = OneSlackSSVM(model=crf, max_iter=100, C=1,
check_constraints=True, break_on_bad=True,
n_jobs=1, tol=.1)
x1, x2, x3 = X
y1, y2, y3 = Y
n_states = len(np.unique(Y))
# delete some rows to make it more fun
x1, y1 = x1[:, :-1], y1[:, :-1]
x2, y2 = x2[:-1], y2[:-1]
# generate graphs
X_ = [x1, x2, x3]
G = [make_grid_edges(x) for x in X_]
# reshape / flatten x and y
X_ = [x.reshape(-1, n_states) for x in X_]
Y = [y.ravel() for y in [y1, y2, y3]]
X = list(zip(X_, G))
clf.fit(X, Y)
Y_pred = clf.predict(X)
for y, y_pred in zip(Y, Y_pred):
assert_array_equal(y, y_pred)
def test_binary_blocks_one_slack_graph():
#testing cutting plane ssvm on easy binary dataset
# generate graphs explicitly for each example
for inference_method in ["dai", "lp"]:
print("testing %s" % inference_method)
X, Y = toy.generate_blocks(n_samples=3)
crf = GraphCRF(inference_method=inference_method)
clf = OneSlackSSVM(problem=crf, max_iter=100, C=100, verbose=100,
check_constraints=True, break_on_bad=True,
n_jobs=1)
x1, x2, x3 = X
y1, y2, y3 = Y
n_states = len(np.unique(Y))
# delete some rows to make it more fun
x1, y1 = x1[:, :-1], y1[:, :-1]
x2, y2 = x2[:-1], y2[:-1]
# generate graphs
X_ = [x1, x2, x3]
G = [make_grid_edges(x) for x in X_]
# reshape / flatten x and y
X_ = [x.reshape(-1, n_states) for x in X_]
Y = [y.ravel() for y in [y1, y2, y3]]
X = zip(X_, G)
clf.fit(X, Y)
Y_pred = clf.predict(X)
for y, y_pred in zip(Y, Y_pred):
assert_array_equal(y, y_pred)
def __init__(self,
prob_estimator,
C_ssvm=1.,
inference='ad3',
inference_cache=50,
tol=1.,
max_iter=200,
n_jobs=1):
"""
Called when initializing the classifier
"""
#self.C_logreg = C_logreg
self.C_ssvm = C_ssvm
self.inference = inference
self.inference_cache = inference_cache
self.tol = tol
self.max_iter = max_iter
self.n_jobs = n_jobs
self.prob_estimator = prob_estimator
self.crf = EdgeFeatureGraphCRF(inference_method=inference)
self.ssvm = OneSlackSSVM(self.crf,
inference_cache=inference_cache,
C=C_ssvm,
tol=tol,
max_iter=max_iter,
n_jobs=n_jobs)
def syntetic_test():
# test model on different train set size & on different train sets
results = np.zeros((18, 5))
full_labeled = np.array([2, 4, 10, 25, 100])
train_size = 400
for dataset in range(1, 19):
X, Y = load_syntetic(dataset)
for j, nfull in enumerate(full_labeled):
crf = EdgeCRF(n_states=10, n_features=10, n_edge_features=2,
inference_method='qpbo')
clf = OneSlackSSVM(crf, max_iter=10000, C=0.01, verbose=0,
tol=0.1, n_jobs=4, inference_cache=100)
x_train = X[:nfull]
y_train = Y[:nfull]
x_test = X[(train_size + 1):]
y_test = Y[(train_size + 1):]
try:
clf.fit(x_train, y_train)
y_pred = clf.predict(x_test)
results[dataset - 1, j] = compute_error(y_test, y_pred)
print('dataset=%d, nfull=%d, error=%f' % (dataset, nfull,
results[dataset - 1, j]))
except ValueError:
print('dataset=%d, nfull=%d: Failed' % (dataset, nfull))
np.savetxt('results/syntetic/full_labeled.txt', results)
def losocv_CRF_prepro(xtrain, y, C=0.5, weight_shift=0, max_iter=1000, fs=128):
"""Leave one subject out cross validation for the CRF model becasuse it requires
special datahandling. Input should be a Pandas Dataframe."""
epochs = 21600
num_sub = 3
# Indices of the subjects
sub_indices = [
np.arange(0, epochs),
np.arange(epochs, epochs * 2),
np.arange(epochs * 2, epochs * 3)
]
res = []
for i in range(len(sub_indices)):
# For the ith iteration, select as trainin the sub_indices other than those at index i for train_index
train_index = np.concatenate(
[sub_indices[(i + 1) % num_sub], sub_indices[(i + 2) % num_sub]])
xtrain_ = xtrain[train_index]
y_train = y.values[train_index]
ytrain_ = y_train
# The test subject is the one at index i
test_index = sub_indices[i]
xtest_ = xtrain[test_index]
y_test = y.values[test_index]
ytest_ = y_test
# CRF Model Preprocessing
ytrain_classes = np.reshape(y_train, (y_train.shape[0], ))
xtrain_crf = np.reshape(
xtrain_,
(2, -1, xtrain_.shape[1])) # Reshape so that it works with CRF
ytrain_crf = np.reshape(
ytrain_, (2, -1)) - 1 # Reshape so that it works with CRF
# CRF Model fitting:
classes = np.unique(ytrain_)
weights_crf = compute_class_weight("balanced", list(classes),
list(ytrain_classes))
weights_crf[0] = weights_crf[0] + (2.5 * weight_shift)
weights_crf[1] = weights_crf[1] + (1.5 * weight_shift)
model = ChainCRF(class_weight=weights_crf)
ssvm = OneSlackSSVM(model=model, C=C, max_iter=max_iter)
ssvm.fit(xtrain_crf, ytrain_crf)
# Test on the third guy
xtest_crf = np.reshape(xtest_, (1, -1, xtest_.shape[1]))
ytest_crf = np.reshape(ytest_, (1, -1)) - 1
y_pred_crf = ssvm.predict(xtest_crf)
y_pred_crf = np.asarray(y_pred_crf).reshape(-1) + 1
resy = sklearn.metrics.balanced_accuracy_score(ytest_, y_pred_crf)
print("Iteration, result:", i, resy)
res.append(resy)
return res
def CRF_pred(eeg1,
eeg2,
emg,
y,
eeg1test,
eeg2test,
emgtest,
C=0.9,
weight_shift=0,
max_iter=1000,
fs=128):
# For the ith iteration, select as trainin the sub_indices other than those at index i for train_index
eeg1_train = eeg1.values
eeg2_train = eeg2.values
emg_train = emg.values
y_train = y.values
# The test subject is the one at index i
eeg1_test = eeg1test.values
eeg2_test = eeg2test.values
emg_test = emgtest.values
# CRF Model Preprocessing
eeg1_ = process_EEG(eeg1_train)
eeg2_ = process_EEG(eeg2_train)
emg_ = process_EMG(emg_train)
xtrain_ = np.concatenate((eeg1_, eeg2_, emg_), axis=1)
ytrain_classes = np.reshape(y_train, (y_train.shape[0], ))
ytrain_ = y_train
eeg1_ = process_EEG(eeg1_test)
eeg2_ = process_EEG(eeg2_test)
emg_ = process_EMG(emg_test)
xtest_ = np.concatenate((eeg1_, eeg2_, emg_), axis=1)
xtrain_crf = np.reshape(
xtrain_,
(3, -1, xtrain_.shape[1])) # Reshape so that it works with CRF
ytrain_crf = np.reshape(ytrain_,
(3, -1)) - 1 # Reshape so that it works with CRF
# CRF Model fitting:
classes = np.unique(ytrain_)
weights_crf = compute_class_weight("balanced", list(classes),
list(ytrain_classes))
weights_crf[0] = weights_crf[0] + (2.5 * weight_shift)
weights_crf[1] = weights_crf[1] + (1.5 * weight_shift)
model = ChainCRF(class_weight=weights_crf)
ssvm = OneSlackSSVM(model=model, C=C, max_iter=max_iter)
ssvm.fit(xtrain_crf, ytrain_crf)
# Test on the third guy
xtest_crf = np.reshape(xtest_, (2, -1, xtest_.shape[1]))
y_pred_crf = ssvm.predict(xtest_crf)
y_pred_crf = np.asarray(y_pred_crf).reshape(-1) + 1
return y_pred_crf
def __init__(self):
# the model
self.model = ChainCRF(directed=True)
# the learner
self.learner = OneSlackSSVM(model=self.model,
C=.1,
inference_cache=50,
tol=0.1,
n_jobs=1)
def test(nfiles):
X = []
Y = []
X_tst = []
Y_tst = []
ntrain = nfiles
ntest = 5*nfiles
print("Training/testing with %d/%d files." % (ntrain,ntest))
start = time.clock()
filename = '../maps/MAPS_AkPnCGdD_2/AkPnCGdD/MUS'
#filename = '../maps/MAPS_AkPnCGdD_1/AkPnCGdD/ISOL/NO'
files = dirs.get_files_with_extension(filename, '.mid')
train_files = files[:ntrain]
print("\t" + str(train_files))
#test_files = files[ntrain:ntest+ntrain]
# for legit testing
test_files = files[-ntest:]
map(per_file, train_files,
it.repeat(X, ntrain), it.repeat(Y, ntrain))
map(per_file, test_files,
it.repeat(X_tst, ntest), it.repeat(Y_tst, ntest))
end = time.clock()
print("\tRead time: %f" % (end - start))
print("\tnWindows train: " + str([X[i].shape[0] for i in range(len(X))]))
start = time.clock()
crf = ChainCRF(n_states=2)
clf = OneSlackSSVM(model=crf, C=100, n_jobs=-1,
inference_cache=100, tol=.1)
clf.fit(np.array(X), np.array(Y))
end = time.clock()
print("\tTrain time: %f" % (end - start))
start = time.clock()
Y_pred = clf.predict(X_tst)
comp = []
for i in range(len(Y_tst)):
for j in range(len(Y_tst[i])):
comp.append((Y_tst[i][j], Y_pred[i][j]))
print Y_tst[i][j],
print
for j in range(len(Y_tst[i])):
print Y_pred[i][j],
print
print
print("\tTrue positives: %d" % comp.count((1,1)))
print("\tTrue negatives: %d" % comp.count((0,0)))
print("\tFalse positives: %d" % comp.count((0,1)))
print("\tFalse negatives: %d" % comp.count((1,0)))
end = time.clock()
print("\tTest time: %f" % (end - start))
def train(self, trajid_list, n_jobs=4):
if self.poi_info is None:
self.poi_info = self.dat_obj.calc_poi_info(trajid_list)
# build POI_ID <--> POI__INDEX mapping for POIs used to train CRF
# which means only POIs in traj such that len(traj) >= 2 are included
poi_set = {p for tid in trajid_list for p in self.dat_obj.traj_dict[tid]
if len(self.dat_obj.traj_dict[tid]) >= 2}
self.poi_list = sorted(poi_set)
self.poi_id_dict, self.poi_id_rdict = dict(), dict()
for idx, poi in enumerate(self.poi_list):
self.poi_id_dict[poi] = idx
self.poi_id_rdict[idx] = poi
# generate training data
train_traj_list = [self.dat_obj.traj_dict[k] for k in trajid_list if len(self.dat_obj.traj_dict[k]) >= 2]
node_features_list = Parallel(n_jobs=n_jobs)(delayed(calc_node_features)(
tr[0], len(tr), self.poi_list, self.poi_info, self.dat_obj) for tr in train_traj_list)
edge_features = calc_edge_features(trajid_list, self.poi_list, self.poi_info, self.dat_obj)
# feature scaling: node features
# should each example be flattened to one vector before scaling?
self.fdim_node = node_features_list[0].shape
X_node_all = np.vstack(node_features_list)
X_node_all = self.scaler_node.fit_transform(X_node_all)
X_node_all = X_node_all.reshape(-1, self.fdim_node[0], self.fdim_node[1])
# feature scaling: edge features
fdim_edge = edge_features.shape
edge_features = self.scaler_edge.fit_transform(edge_features.reshape(fdim_edge[0] * fdim_edge[1], -1))
self.edge_features = edge_features.reshape(fdim_edge)
assert(len(train_traj_list) == X_node_all.shape[0])
X_train = [(X_node_all[k, :, :],
self.edge_features.copy(),
(self.poi_id_dict[train_traj_list[k][0]], len(train_traj_list[k])))
for k in range(len(train_traj_list))]
y_train = [np.array([self.poi_id_dict[k] for k in tr]) for tr in train_traj_list]
assert(len(X_train) == len(y_train))
# train
sm = MyModel(inference_train=self.inference_train, inference_pred=self.inference_pred,
share_params=self.share_params, multi_label=self.multi_label)
if self.debug is True:
print('C:', self.C)
verbose = 1 if self.debug is True else 0
self.osssvm = OneSlackSSVM(model=sm, C=self.C, n_jobs=n_jobs, verbose=verbose)
try:
self.osssvm.fit(X_train, y_train, initialize=True)
self.trained = True
print('SSVM training finished.')
# except ValueError:
except:
self.trained = False
sys.stderr.write('SSVM training FAILED.\n')
# raise
return self.trained
def train(X,y):
X_train_directions = make_directions(X)
Y_train_flat = [y.ravel()]
inference = 'qpbo'
# first, train on X with directions only:
crf = EdgeFeatureGraphCRF(inference_method=inference)
ssvm = OneSlackSSVM(crf, inference_cache=50, C=.1, tol=.1, max_iter=10,
n_jobs=1,show_loss_every=1)
ssvm.fit(X_train_directions, Y_train_flat, warm_start=False)
return ssvm
def test_constraint_removal():
digits = load_digits()
X, y = digits.data, digits.target
y = 2 * (y % 2) - 1 # even vs odd as +1 vs -1
X = X / 16.
pbl = BinaryClf(n_features=X.shape[1])
clf_no_removal = OneSlackSSVM(model=pbl, max_iter=500, C=1,
inactive_window=0, tol=0.01)
clf_no_removal.fit(X, y)
clf = OneSlackSSVM(model=pbl, max_iter=500, C=1, tol=0.01,
inactive_threshold=1e-8)
clf.fit(X, y)
# check that we learned something
assert_greater(clf.score(X, y), .92)
# results are mostly equal
# if we decrease tol, they will get more similar
assert_less(np.mean(clf.predict(X) != clf_no_removal.predict(X)), 0.02)
# without removal, have as many constraints as iterations
assert_equal(len(clf_no_removal.objective_curve_),
len(clf_no_removal.constraints_))
# with removal, there are less constraints than iterations
assert_less(len(clf.constraints_),
len(clf.objective_curve_))
def test_one_slack_attractive_potentials():
# test that submodular SSVM can learn the block dataset
X, Y = generate_blocks(n_samples=10)
crf = GridCRF(inference_method=inference_method)
submodular_clf = OneSlackSSVM(
model=crf, max_iter=200, C=1, check_constraints=True, negativity_constraint=[5], inference_cache=50
)
submodular_clf.fit(X, Y)
Y_pred = submodular_clf.predict(X)
assert_array_equal(Y, Y_pred)
assert_true(submodular_clf.w[5] < 0)
def test_multinomial_blocks_one_slack():
#testing cutting plane ssvm on easy multinomial dataset
X, Y = generate_blocks_multinomial(n_samples=10, noise=0.5, seed=0)
n_labels = len(np.unique(Y))
crf = GridCRF(n_states=n_labels, inference_method=inference_method)
clf = OneSlackSSVM(model=crf, max_iter=150, C=1,
check_constraints=True, break_on_bad=True, tol=.1,
inference_cache=50)
clf.fit(X, Y)
Y_pred = clf.predict(X)
assert_array_equal(Y, Y_pred)
def test_multinomial_blocks_one_slack():
#testing cutting plane ssvm on easy multinomial dataset
X, Y = generate_blocks_multinomial(n_samples=10, noise=0.5, seed=0)
n_labels = len(np.unique(Y))
crf = GridCRF(n_states=n_labels, inference_method=inference_method)
clf = OneSlackSSVM(model=crf, max_iter=150, C=1,
check_constraints=True, break_on_bad=True, tol=.1,
inference_cache=50)
clf.fit(X, Y)
Y_pred = clf.predict(X)
assert_array_equal(Y, Y_pred)
def test_one_slack_attractive_potentials():
# test that submodular SSVM can learn the block dataset
X, Y = toy.generate_blocks(n_samples=10)
crf = GridCRF()
submodular_clf = OneSlackSSVM(model=crf, max_iter=200, C=100, verbose=1,
check_constraints=True,
positive_constraint=[5], n_jobs=-1)
submodular_clf.fit(X, Y)
Y_pred = submodular_clf.predict(X)
assert_array_equal(Y, Y_pred)
assert_true(submodular_clf.w[5] < 0) # don't ask me about signs
def syntetic_train_score_per_iter(result, only_weak=False, plot=True):
w_history = result.data["w_history"]
meta_data = result.meta
n_full = meta_data["n_full"]
n_train = meta_data["n_train"]
n_inference_iter = meta_data["n_inference_iter"]
n_full = meta_data["n_full"]
n_train = meta_data["n_train"]
dataset = meta_data["dataset"]
C = meta_data["C"]
latent_iter = meta_data["latent_iter"]
max_iter = meta_data["max_iter"]
inner_tol = meta_data["inner_tol"]
outer_tol = meta_data["outer_tol"]
alpha = meta_data["alpha"]
min_changes = meta_data["min_changes"]
initialize = meta_data["initialize"]
crf = HCRF(
n_states=10, n_features=10, n_edge_features=2, alpha=alpha, inference_method="gco", n_iter=n_inference_iter
)
base_clf = OneSlackSSVM(crf, max_iter=max_iter, C=C, verbose=0, tol=inner_tol, n_jobs=4, inference_cache=100)
clf = LatentSSVM(base_clf, latent_iter=latent_iter, verbose=2, tol=outer_tol, min_changes=min_changes, n_jobs=4)
X, Y = load_syntetic(dataset)
Xtrain, Ytrain, Ytrain_full, Xtest, Ytest = split_test_train(X, Y, n_full, n_train)
if only_weak:
Xtrain = [x for (i, x) in enumerate(Xtrain) if not Ytrain[i].full_labeled]
Ytrain_full = [y for (i, y) in enumerate(Ytrain_full) if not Ytrain[i].full_labeled]
base_clf.w = None
clf.w_history_ = w_history
clf.iter_done = w_history.shape[0]
train_scores = []
for score in clf.staged_score(Xtrain, Ytrain_full):
train_scores.append(score)
train_scores = np.array(train_scores)
if plot:
x = np.arange(0, train_scores.size)
pl.rc("text", usetex=True)
pl.rc("font", family="serif")
pl.figure(figsize=(10, 10), dpi=96)
pl.title("score on train set")
pl.plot(x, train_scores)
pl.scatter(x, train_scores)
pl.xlabel("iteration")
pl.xlim([-0.5, train_scores.size + 1])
return train_scores
def test_multilabel_yeast_independent():
yeast = fetch_mldata("yeast")
X = yeast.data
y = yeast.target.toarray().T.astype(np.int)
# no edges for the moment
edges = np.zeros((0, 2), dtype=np.int)
pbl = MultiLabelProblem(n_features=X.shape[1], n_labels=y.shape[1],
edges=edges)
ssvm = OneSlackSSVM(pbl, verbose=10)
ssvm.fit(X, y)
from IPython.core.debugger import Tracer
Tracer()()
def test_multinomial_blocks_one_slack():
#testing cutting plane ssvm on easy multinomial dataset
X, Y = toy.generate_blocks_multinomial(n_samples=10, noise=0.3,
seed=0)
n_labels = len(np.unique(Y))
for inference_method in ['lp']:
crf = GridCRF(n_states=n_labels, inference_method=inference_method)
clf = OneSlackSSVM(problem=crf, max_iter=50, C=100, verbose=100,
check_constraints=True, break_on_bad=True)
clf.fit(X, Y)
Y_pred = clf.predict(X)
assert_array_equal(Y, Y_pred)
def test_one_slack_attractive_potentials():
# test that submodular SSVM can learn the block dataset
X, Y = generate_blocks(n_samples=10)
crf = GridCRF(inference_method=inference_method)
submodular_clf = OneSlackSSVM(model=crf, max_iter=200, C=1,
check_constraints=True,
negativity_constraint=[5],
inference_cache=50)
submodular_clf.fit(X, Y)
Y_pred = submodular_clf.predict(X)
assert_array_equal(Y, Y_pred)
assert_true(submodular_clf.w[5] < 0)
def __init__(self, model, max_iter=10000, C=1.0, check_constraints=False,
verbose=0, negativity_constraint=None, n_jobs=1,
break_on_bad=False, show_loss_every=0, tol=1e-3,
inference_cache=0, inactive_threshold=1e-5,
inactive_window=50, logger=None, cache_tol='auto',
switch_to=None):
Pystruct_OneSlackSSVM.__init__(self, model, max_iter=max_iter, C=C, check_constraints=check_constraints,
verbose=verbose, negativity_constraint=negativity_constraint, n_jobs=n_jobs,
break_on_bad=break_on_bad, show_loss_every=show_loss_every, tol=tol,
inference_cache=inference_cache, inactive_threshold=inactive_threshold,
inactive_window=inactive_window, logger=logger, cache_tol=cache_tol,
switch_to=switch_to)
def test_blobs_2d_one_slack():
# make two gaussian blobs
X, Y = make_blobs(n_samples=80, centers=2, random_state=1)
Y = 2 * Y - 1
# we have to add a constant 1 feature by hand :-/
X = np.hstack([X, np.ones((X.shape[0], 1))])
X_train, X_test, Y_train, Y_test = X[:40], X[40:], Y[:40], Y[40:]
pbl = BinarySVMModel(n_features=3)
svm = OneSlackSSVM(pbl, verbose=30, C=1000)
svm.fit(X_train, Y_train)
assert_array_equal(Y_test, np.hstack(svm.predict(X_test)))
def test_blobs_2d_one_slack():
# make two gaussian blobs
X, Y = make_blobs(n_samples=80, centers=3, random_state=42)
# we have to add a constant 1 feature by hand :-/
X = np.hstack([X, np.ones((X.shape[0], 1))])
X_train, X_test, Y_train, Y_test = X[:40], X[40:], Y[:40], Y[40:]
pbl = MultiClassClf(n_features=3, n_classes=3)
svm = OneSlackSSVM(pbl, check_constraints=True, C=1000)
svm.fit(X_train, Y_train)
assert_array_equal(Y_test, np.hstack(svm.predict(X_test)))
def test_blobs_2d_one_slack():
# make two gaussian blobs
X, Y = make_blobs(n_samples=80, centers=3, random_state=42)
# we have to add a constant 1 feature by hand :-/
X = np.hstack([X, np.ones((X.shape[0], 1))])
X_train, X_test, Y_train, Y_test = X[:40], X[40:], Y[:40], Y[40:]
pbl = MultiClassClf(n_features=3, n_classes=3)
svm = OneSlackSSVM(pbl, check_constraints=True, C=1000)
svm.fit(X_train, Y_train)
assert_array_equal(Y_test, np.hstack(svm.predict(X_test)))
def test_standard_svm_blobs_2d_class_weight():
# no edges, reduce to crammer-singer svm
X, Y = make_blobs(n_samples=210, centers=3, random_state=1, cluster_std=3,
shuffle=False)
X = np.hstack([X, np.ones((X.shape[0], 1))])
X, Y = X[:170], Y[:170]
X_graphs = [(x[np.newaxis, :], np.empty((0, 2), dtype=np.int)) for x in X]
pbl = GraphCRF(n_features=3, n_states=3, inference_method='unary')
svm = OneSlackSSVM(pbl, check_constraints=False, C=1000)
svm.fit(X_graphs, Y[:, np.newaxis])
weights = 1. / np.bincount(Y)
weights *= len(weights) / np.sum(weights)
pbl_class_weight = GraphCRF(n_features=3, n_states=3, class_weight=weights,
inference_method='unary')
svm_class_weight = OneSlackSSVM(pbl_class_weight, C=10,
check_constraints=False,
break_on_bad=False)
svm_class_weight.fit(X_graphs, Y[:, np.newaxis])
assert_greater(f1_score(Y, np.hstack(svm_class_weight.predict(X_graphs))),
f1_score(Y, np.hstack(svm.predict(X_graphs))))
def set_params(self, trajid_list, w, n_jobs=4):
if self.poi_info is None:
self.poi_info = self.dat_obj.calc_poi_info(trajid_list)
poi_set = {p for tid in trajid_list for p in self.dat_obj.traj_dict[tid]
if len(self.dat_obj.traj_dict[tid]) >= 2}
self.poi_list = sorted(poi_set)
self.poi_id_dict, self.poi_id_rdict = dict(), dict()
for idx, poi in enumerate(self.poi_list):
self.poi_id_dict[poi] = idx
self.poi_id_rdict[idx] = poi
# generate training data
train_traj_list = [self.dat_obj.traj_dict[k] for k in trajid_list if len(self.dat_obj.traj_dict[k]) >= 2]
node_features_list = Parallel(n_jobs=n_jobs)(delayed(calc_node_features)(
tr[0], len(tr), self.poi_list, self.poi_info, self.dat_obj) for tr in train_traj_list)
edge_features = calc_edge_features(trajid_list, self.poi_list, self.poi_info, self.dat_obj)
# feature scaling: node features
# should each example be flattened to one vector before scaling?
self.fdim_node = node_features_list[0].shape
X_node_all = np.vstack(node_features_list)
X_node_all = self.scaler_node.fit_transform(X_node_all)
X_node_all = X_node_all.reshape(-1, self.fdim_node[0], self.fdim_node[1])
# feature scaling: edge features
fdim_edge = edge_features.shape
edge_features = self.scaler_edge.fit_transform(edge_features.reshape(fdim_edge[0] * fdim_edge[1], -1))
self.edge_features = edge_features.reshape(fdim_edge)
assert(len(train_traj_list) == X_node_all.shape[0])
X_train = [(X_node_all[k, :, :],
self.edge_features.copy(),
(self.poi_id_dict[train_traj_list[k][0]], len(train_traj_list[k])))
for k in range(len(train_traj_list))]
y_train = [np.array([self.poi_id_dict[k] for k in tr]) for tr in train_traj_list]
assert(len(X_train) == len(y_train))
n_features = X_train[0][0].shape[1]
n_states = len(np.unique(np.hstack([y.ravel() for y in y_train])))
n_edge_features = X_train[0][1].shape[2]
sm = MyModel(inference_train=self.inference_train, inference_pred=self.inference_pred,
share_params=self.share_params, multi_label=self.multi_label,
n_states=n_states, n_features=n_features, n_edge_features=n_edge_features, debug=self.debug)
self.learner = OneSlackSSVM(model=sm, C=self.C, n_jobs=1)
self.learner.w = w
self.trained = True
def test_with_crosses_bad_init():
# use less perfect initialization
rnd = np.random.RandomState(0)
X, Y = generate_crosses(n_samples=20, noise=5, n_crosses=1, total_size=8)
X_test, Y_test = X[10:], Y[10:]
X, Y = X[:10], Y[:10]
crf = LatentGridCRF(n_states_per_label=2)
crf.initialize(X, Y)
H_init = crf.init_latent(X, Y)
mask = rnd.uniform(size=H_init.shape) > .7
H_init[mask] = 2 * (H_init[mask] / 2)
one_slack_ssvm = OneSlackSSVM(crf,
inactive_threshold=1e-8,
cache_tol=.0001,
inference_cache=50,
C=100)
clf = LatentSSVM(one_slack_ssvm)
clf.fit(X, Y, H_init=H_init)
Y_pred = clf.predict(X)
assert_array_equal(np.array(Y_pred), Y)
# test that score is not always 1
assert_true(.98 < clf.score(X_test, Y_test) < 1)
def test_standard_svm_blobs_2d():
# no edges, reduce to crammer-singer svm
X, Y = make_blobs(n_samples=80, centers=3, random_state=42)
# we have to add a constant 1 feature by hand :-/
X = np.hstack([X, np.ones((X.shape[0], 1))])
X_train, X_test, Y_train, Y_test = X[:40], X[40:], Y[:40], Y[40:]
X_train_graphs = [(x[np.newaxis, :], np.empty((0, 2), dtype=np.int)) for x in X_train]
X_test_graphs = [(x[np.newaxis, :], np.empty((0, 2), dtype=np.int)) for x in X_test]
pbl = GraphCRF(n_features=3, n_states=3)
svm = OneSlackSSVM(pbl, verbose=10, check_constraints=True, C=1000)
svm.fit(X_train_graphs, Y_train[:, np.newaxis])
assert_array_equal(Y_test, np.hstack(svm.predict(X_test_graphs)))
def test_latent_node_boxes_standard_latent():
# learn the "easy" 2x2 boxes dataset.
# a 2x2 box is placed randomly in a 4x4 grid
# we add a latent variable for each 2x2 patch
# that should make the model fairly simple
X, Y = make_simple_2x2(seed=1, n_samples=40)
latent_crf = LatentNodeCRF(n_labels=2, n_hidden_states=2, n_features=1)
one_slack = OneSlackSSVM(latent_crf)
n_slack = NSlackSSVM(latent_crf)
subgradient = SubgradientSSVM(latent_crf, max_iter=100)
for base_svm in [one_slack, n_slack, subgradient]:
base_svm.C = 10
latent_svm = LatentSSVM(base_svm, latent_iter=10)
G = [make_grid_edges(x) for x in X]
# make edges for hidden states:
edges = make_edges_2x2()
G = [np.vstack([make_grid_edges(x), edges]) for x in X]
# reshape / flatten x and y
X_flat = [x.reshape(-1, 1) for x in X]
Y_flat = [y.ravel() for y in Y]
X_ = zip(X_flat, G, [2 * 2 for x in X_flat])
latent_svm.fit(X_[:20], Y_flat[:20])
assert_array_equal(latent_svm.predict(X_[:20]), Y_flat[:20])
assert_equal(latent_svm.score(X_[:20], Y_flat[:20]), 1)
# test that score is not always 1
assert_true(.98 < latent_svm.score(X_[20:], Y_flat[20:]) < 1)
def do_comparison(X_train, y_train, X_test, y_test, dataset):
# evaluate both svms on a given datasets, generate plots
Cs = 10.**np.arange(-4, 1)
multisvm = MultiSVM()
svm = OneSlackSSVM(MultiClassClf(), tol=0.01)
accs_pystruct, times_pystruct = eval_on_data(X_train,
y_train,
X_test,
y_test,
svm,
Cs=Cs)
accs_svmstruct, times_svmstruct = eval_on_data(X_train,
y_train,
X_test,
y_test,
multisvm,
Cs=Cs)
plot_curves(times_svmstruct,
times_pystruct,
Cs=Cs,
title="learning time (s) %s" % dataset,
filename="times_%s.pdf" % dataset)
plot_curves(accs_svmstruct,
accs_pystruct,
Cs=Cs,
title="accuracy %s" % dataset,
filename="accs_%s.pdf" % dataset)
def main():
print("Please be patient. Will take 5-20 minutes.")
snakes = load_snakes()
X_train, Y_train = snakes['X_train'], snakes['Y_train']
X_train = [one_hot_colors(x) for x in X_train]
Y_train_flat = [y_.ravel() for y_ in Y_train]
X_train_directions, X_train_edge_features = prepare_data(X_train)
# first, train on X with directions only:
crf = EdgeFeatureGraphCRF(inference_method='qpbo')
ssvm = OneSlackSSVM(crf, inference_cache=50, C=.1, tol=.1, switch_to='ad3',
n_jobs=-1)
ssvm.fit(X_train_directions, Y_train_flat)
# Evaluate using confusion matrix.
# Clearly the middel of the snake is the hardest part.
X_test, Y_test = snakes['X_test'], snakes['Y_test']
X_test = [one_hot_colors(x) for x in X_test]
Y_test_flat = [y_.ravel() for y_ in Y_test]
X_test_directions, X_test_edge_features = prepare_data(X_test)
Y_pred = ssvm.predict(X_test_directions)
print("Results using only directional features for edges")
print("Test accuracy: %.3f"
% accuracy_score(np.hstack(Y_test_flat), np.hstack(Y_pred)))
print(confusion_matrix(np.hstack(Y_test_flat), np.hstack(Y_pred)))
# now, use more informative edge features:
crf = EdgeFeatureGraphCRF(inference_method='qpbo')
ssvm = OneSlackSSVM(crf, inference_cache=50, C=.1, tol=.1, switch_to='ad3',
n_jobs=-1)
ssvm.fit(X_train_edge_features, Y_train_flat)
Y_pred2 = ssvm.predict(X_test_edge_features)
print("Results using also input features for edges")
print("Test accuracy: %.3f"
% accuracy_score(np.hstack(Y_test_flat), np.hstack(Y_pred2)))
print(confusion_matrix(np.hstack(Y_test_flat), np.hstack(Y_pred2)))
# plot stuff
fig, axes = plt.subplots(2, 2)
axes[0, 0].imshow(snakes['X_test'][0], interpolation='nearest')
axes[0, 0].set_title('Input')
y = Y_test[0].astype(np.int)
bg = 2 * (y != 0) # enhance contrast
axes[0, 1].matshow(y + bg, cmap=plt.cm.Greys)
axes[0, 1].set_title("Ground Truth")
axes[1, 0].matshow(Y_pred[0].reshape(y.shape) + bg, cmap=plt.cm.Greys)
axes[1, 0].set_title("Prediction w/o edge features")
axes[1, 1].matshow(Y_pred2[0].reshape(y.shape) + bg, cmap=plt.cm.Greys)
axes[1, 1].set_title("Prediction with edge features")
for a in axes.ravel():
a.set_xticks(())
a.set_yticks(())
plt.show()
from IPython.core.debugger import Tracer
Tracer()()
def synteticTrain(peer):
# train model on a single set
clf = OneSlackSSVM(peer, max_iter=10000, C=0.01, verbose=2,
tol=0.1, n_jobs=4, inference_cache=100)
start = time()
clf.fit()
stop = time()
models_basedir = peer.config.get("models.basedir")
np.savetxt(models_basedir + 'syntetic_full.csv', clf.w[None], delimiter=' ')
#with open(models_basedir + 'syntetic_full' + '.pickle', 'w') as f:
# pickle.dump(clf, f)
peer.log('Elapsed time: %f s' % (stop - start))
return clf
def test_svm_as_crf_pickling():
iris = load_iris()
X, y = iris.data, iris.target
X_ = [(np.atleast_2d(x), np.empty((0, 2), dtype=np.int)) for x in X]
Y = y.reshape(-1, 1)
X_train, X_test, y_train, y_test = train_test_split(X_, Y, random_state=1)
_, file_name = mkstemp()
pbl = GraphCRF(n_features=4, n_states=3, inference_method="unary")
logger = SaveLogger(file_name)
svm = OneSlackSSVM(pbl, check_constraints=True, C=1, n_jobs=1, logger=logger)
svm.fit(X_train, y_train)
assert_less(0.97, svm.score(X_test, y_test))
assert_less(0.97, logger.load().score(X_test, y_test))
def msrc_test():
# test model on different train set sizes
basedir = '../data/msrc/trainmasks/'
models_basedir = 'models/msrc/'
quality = []
Xtest, Ytest = load_msrc('test')
Ytest = remove_areas(Ytest)
Xtrain, Ytrain = load_msrc('train')
Ytrain = remove_areas(Ytrain)
for n_train in [20, 40, 80, 160, 276]:
crf = EdgeCRF(n_states=24, n_features=2028, n_edge_features=4,
inference_method='gco')
clf = OneSlackSSVM(crf, max_iter=1000, C=0.01, verbose=0,
tol=0.1, n_jobs=4, inference_cache=100)
if n_train != 276:
train_mask = np.genfromtxt(basedir + 'trainMaskX%d.txt' % n_train)
train_mask = train_mask[:277].astype(np.bool)
else:
train_mask = np.ones(276).astype(np.bool)
curX = []
curY = []
for (s, x, y) in zip(train_mask, Xtrain, Ytrain):
if s:
curX.append(x)
curY.append(y)
start = time()
clf.fit(curX, curY)
stop = time()
np.savetxt(models_basedir + 'test_model_%d.csv' % n_train, clf.w)
with open(models_basedir + 'test_model_%d' % n_train + '.pickle', 'w') as f:
pickle.dump(clf, f)
Ypred = clf.predict(Xtest)
q = 1 - compute_error(Ytest, Ypred)
print('n_train=%d, quality=%f, time=%f' % (n_train, q, stop - start))
quality.append(q)
np.savetxt('results/msrc/msrc_full.txt', quality)
def msrc_test():
# test model on different train set sizes
basedir = '../data/msrc/trainmasks/'
models_basedir = 'models/msrc/'
quality = []
Xtest, Ytest = load_msrc('test')
Ytest = remove_areas(Ytest)
Xtrain, Ytrain = load_msrc('train')
Ytrain = remove_areas(Ytrain)
for n_train in [20, 40, 80, 160, 276]:
crf = EdgeCRF(n_states=24, n_features=2028, n_edge_features=4,
inference_method='gco')
clf = OneSlackSSVM(crf, max_iter=1000, C=0.01, verbose=0,
tol=0.1, n_jobs=4, inference_cache=100)
if n_train != 276:
train_mask = np.genfromtxt(basedir + 'trainMaskX%d.txt' % n_train)
train_mask = train_mask[:277].astype(np.bool)
else:
train_mask = np.ones(276).astype(np.bool)
curX = []
curY = []
for (s, x, y) in zip(train_mask, Xtrain, Ytrain):
if s:
curX.append(x)
curY.append(y)
start = time()
clf.fit(curX, curY)
stop = time()
np.savetxt(models_basedir + 'test_model_%d.csv' % n_train, clf.w)
with open(models_basedir + 'test_model_%d' % n_train + '.pickle', 'w') as f:
cPickle.dump(clf, f)
Ypred = clf.predict(Xtest)
q = 1 - compute_error(Ytest, Ypred)
print 'n_train=%d, quality=%f, time=%f' % (n_train, q, stop - start)
quality.append(q)
np.savetxt('results/msrc/msrc_full.txt', quality)
def train(x_train, y_train, x_test, y_test):
x_train = np.asarray(x_train, dtype=np.float)
y_train = np.asarray(y_train, dtype=np.int64)
# x_test = np.asarray(x_test, dtype=np.float)
# y_test = np.asarray(y_test, dtype=np.int64)
x_test = x_train
y_test = y_train
from pystruct.learners import NSlackSSVM, OneSlackSSVM, SubgradientSSVM, LatentSSVM, SubgradientLatentSSVM, PrimalDSStructuredSVM
from pystruct.models import MultiLabelClf, MultiClassClf
clf = OneSlackSSVM(MultiLabelClf(),
C=1,
show_loss_every=1,
verbose=1,
max_iter=1000)
# print(x_train, y_train)
# input()
clf.fit(x_train, y_train)
result = clf.predict(x_test)
print('Result: \n', result)
print('True label:\n', y_test)
clf.score(x_test, y_test)
print('\n')
count = 0
for i in range(len(result)):
# print(np.sum(np.square(y_test[i]-result[i])))
if np.sum(np.square(y_test[i] - result[i])) != 0:
print('True label: ', y_test[i], 'Predict: ', result[i])
count += 1
print(count)
translate_vector(x_test, y_test)
def syntetic():
# train model on a single set
models_basedir = 'models/syntetic/'
crf = EdgeCRF(n_states=10, n_features=10, n_edge_features=2,
inference_method='gco')
clf = OneSlackSSVM(crf, max_iter=10000, C=0.01, verbose=2,
tol=0.1, n_jobs=4, inference_cache=100)
X, Y = load_syntetic(1)
x_train, x_test, y_train, y_test = train_test_split(X, Y,
train_size=100,
random_state=179)
start = time()
clf.fit(x_train, y_train)
stop = time()
np.savetxt(models_basedir + 'syntetic_full.csv', clf.w)
with open(models_basedir + 'syntetic_full' + '.pickle', 'w') as f:
cPickle.dump(clf, f)
y_pred = clf.predict(x_test)
print 'Error on test set: %f' % compute_error(y_test, y_pred)
print 'Score on test set: %f' % clf.score(x_test, y_test)
print 'Score on train set: %f' % clf.score(x_train, y_train)
print 'Norm of weight vector: |w|=%f' % np.linalg.norm(clf.w)
print 'Elapsed time: %f s' % (stop - start)
return clf
def test_standard_svm_blobs_2d():
# no edges, reduce to crammer-singer svm
X, Y = make_blobs(n_samples=80, centers=3, random_state=42)
# we have to add a constant 1 feature by hand :-/
X = np.hstack([X, np.ones((X.shape[0], 1))])
X_train, X_test, Y_train, Y_test = X[:40], X[40:], Y[:40], Y[40:]
X_train_graphs = [(x[np.newaxis, :], np.empty((0, 2), dtype=np.int))
for x in X_train]
X_test_graphs = [(x[np.newaxis, :], np.empty((0, 2), dtype=np.int))
for x in X_test]
pbl = GraphCRF(n_features=3, n_states=3, inference_method='unary')
svm = OneSlackSSVM(pbl, check_constraints=True, C=1000)
svm.fit(X_train_graphs, Y_train[:, np.newaxis])
assert_array_equal(Y_test, np.hstack(svm.predict(X_test_graphs)))
class PystructSequenceLearner(SequenceLearner):
def __init__(self):
# the model
self.model = ChainCRF(directed=True)
# the learner
self.learner = OneSlackSSVM(model=self.model,
C=.1,
inference_cache=50,
tol=0.1,
n_jobs=1)
# self.learner = StructuredPerceptron(model=self.model, average=True)
def fit(self, X, y):
# Train linear chain CRF
self.learner.fit(X, y)
def predict(self, X):
return self.learner.predict(X)
def test_svm_as_crf_pickling():
iris = load_iris()
X, y = iris.data, iris.target
X_ = [(np.atleast_2d(x), np.empty((0, 2), dtype=np.int)) for x in X]
Y = y.reshape(-1, 1)
X_train, X_test, y_train, y_test = train_test_split(X_, Y, random_state=1)
_, file_name = mkstemp()
pbl = GraphCRF(n_features=4, n_states=3, inference_method='unary')
logger = SaveLogger(file_name)
svm = OneSlackSSVM(pbl, check_constraints=True, C=1, n_jobs=1,
logger=logger)
svm.fit(X_train, y_train)
assert_less(.97, svm.score(X_test, y_test))
assert_less(.97, logger.load().score(X_test, y_test))
def conditional_random_fields(X, y):
"""
"""
X_ = [(np.atleast_2d(x), np.empty((0, 2), dtype=np.int)) for x in X]
Y = y.reshape(-1, 1)
X_train, X_test, y_train, y_test = train_test_split(X_, Y)
pbl = GraphCRF()
svm = OneSlackSSVM(pbl)
svm.fit(X_train, y_train)
y_pred = np.vstack(svm.predict(X_test))
print("Score with pystruct crf svm: %f "
% (np.mean(y_pred == y_test)))
print classification_report(y_test, y_pred)
plot_confusion_matrix(y_test, y_pred)
def __init__(self,
model,
max_iter=10000,
C=1.0,
check_constraints=False,
verbose=0,
negativity_constraint=None,
positivity_constraint=None,
null_constraints=None,
hard_constraints=None,
n_jobs=1,
break_on_bad=False,
show_loss_every=0,
tol=1e-3,
inference_cache=0,
inactive_threshold=1e-5,
inactive_window=50,
logger=None,
cache_tol='auto',
switch_to=None,
generate_hard_constraints=None,
initialize_constraints=None,
qp_eps=1e-5):
OneSlackSSVM.__init__(self, model, max_iter, C, check_constraints,
verbose, negativity_constraint,
positivity_constraint, hard_constraints, n_jobs,
break_on_bad, show_loss_every, tol,
inference_cache, inactive_threshold,
inactive_window, logger, cache_tol, switch_to)
if (hard_constraints, positivity_constraint, negativity_constraint,
generate_hard_constraints) is (None, None, None, None):
self.hard_satisfied = True
else:
# there are hard constraints to satisfy
self.hard_satisfied = False
self.null_constraints = null_constraints
self.generate_hard_constraints = generate_hard_constraints
self.initialize_constraints = initialize_constraints
self.cutting_constraints = []
self._inference_calls = 0
self.qp_eps = qp_eps
self.converged = False
def test_one_slack_constraint_caching():
#testing cutting plane ssvm on easy multinomial dataset
X, Y = toy.generate_blocks_multinomial(n_samples=10, noise=0.5,
seed=0, size_x=9)
n_labels = len(np.unique(Y))
crf = GridCRF(n_states=n_labels)
clf = OneSlackSSVM(model=crf, max_iter=150, C=1,
check_constraints=True, break_on_bad=True,
inference_cache=50, inactive_window=0)
clf.fit(X, Y)
Y_pred = clf.predict(X)
assert_array_equal(Y, Y_pred)
assert_equal(len(clf.inference_cache_), len(X))
# there should be 21 constraints, which are less than the 94 iterations
# that are done
assert_equal(len(clf.inference_cache_[0]), 21)
# check that we didn't change the behavior of how we construct the cache
constraints_per_sample = [len(cache) for cache in clf.inference_cache_]
assert_equal(np.max(constraints_per_sample), 21)
assert_equal(np.min(constraints_per_sample), 21)
def train_seq(X, y, crf_params):
X_ = [X[k] for k in sorted(X.keys())]
y_ = [y[k] for k in sorted(y.keys())]
class_sizes = np.bincount(np.hstack(y_))
cw = 1. / class_sizes
cw = cw / cw.sum()
return OneSlackSSVM(model=ChainCRF(inference_method='max-product',
class_weight=cw),
max_iter=100000,
verbose=False,
**crf_params).fit(X_, y_)
def test_one_slack_constraint_caching():
#testing cutting plane ssvm on easy multinomial dataset
X, Y = toy.generate_blocks_multinomial(n_samples=10, noise=0.3,
seed=0)
n_labels = len(np.unique(Y))
crf = GridCRF(n_states=n_labels, inference_method='lp')
clf = OneSlackSSVM(problem=crf, max_iter=50, C=100, verbose=100,
check_constraints=True, break_on_bad=True,
inference_cache=50)
clf.fit(X, Y)
Y_pred = clf.predict(X)
assert_array_equal(Y, Y_pred)
assert_equal(len(clf.inference_cache_), len(X))
# there should be 9 constraints, which are less than the 16 iterations
# that are done
assert_equal(len(clf.inference_cache_[0]), 9)
# check that we didn't change the behavior of how we construct the cache
constraints_per_sample = [len(cache) for cache in clf.inference_cache_]
assert_equal(np.max(constraints_per_sample), 10)
assert_equal(np.min(constraints_per_sample), 8)
def test_equal_class_weights():
# test that equal class weight is the same as no class weight
X, Y = make_blobs(n_samples=80, centers=3, random_state=42)
X = np.hstack([X, np.ones((X.shape[0], 1))])
X_train, X_test, Y_train, Y_test = X[:40], X[40:], Y[:40], Y[40:]
pbl = MultiClassClf(n_features=3, n_classes=3)
svm = OneSlackSSVM(pbl, C=10)
svm.fit(X_train, Y_train)
predict_no_class_weight = svm.predict(X_test)
pbl_class_weight = MultiClassClf(n_features=3,
n_classes=3,
class_weight=np.ones(3))
svm_class_weight = OneSlackSSVM(pbl_class_weight, C=10)
svm_class_weight.fit(X_train, Y_train)
predict_class_weight = svm_class_weight.predict(X_test)
assert_array_equal(predict_no_class_weight, predict_class_weight)
assert_array_almost_equal(svm.w, svm_class_weight.w)
def train_structured_svm(observations,targets):
"""
:param observations: our train dataset
:param targets: multiple target variables.
:return: the structured svm model
"""
# ideally you can say the edges that are connected. For now, we use full.
n_labels = len(targets[0])
full = np.vstack([x for x in itertools.combinations(range(n_labels), 2)])
#tree = chow_liu_tree(targets)
# Choose the best model...
full_model = MultiLabelClf(edges=full, inference_method='lp')
#tree_model = MultiLabelClf(edges=tree, inference_method="max-product")
full_ssvm = OneSlackSSVM(full_model, inference_cache=50, C=.1, tol=0.01)
full_ssvm.fit(np.array(observations), np.array(targets))
return full_ssvm
def test_one_slack_repellent_potentials():
# test non-submodular problem with and without submodularity constraint
# dataset is checkerboard
X, Y = generate_checker()
crf = GridCRF(inference_method=inference_method)
clf = OneSlackSSVM(model=crf, max_iter=10, C=0.01, check_constraints=True)
clf.fit(X, Y)
Y_pred = clf.predict(X)
# standard crf can predict perfectly
assert_array_equal(Y, Y_pred)
submodular_clf = OneSlackSSVM(
model=crf, max_iter=10, C=0.01, check_constraints=True, negativity_constraint=[4, 5, 6]
)
submodular_clf.fit(X, Y)
Y_pred = submodular_clf.predict(X)
assert_less(submodular_clf.score(X, Y), 0.99)
# submodular crf can not do better than unaries
for i, x in enumerate(X):
y_pred_unaries = crf.inference(x, np.array([1, 0, 0, 1, 0, 0, 0]))
assert_array_equal(y_pred_unaries, Y_pred[i])
def test_equal_class_weights():
# test that equal class weight is the same as no class weight
X, Y = make_blobs(n_samples=80, centers=3, random_state=42)
X = np.hstack([X, np.ones((X.shape[0], 1))])
X_train, X_test, Y_train, Y_test = X[:40], X[40:], Y[:40], Y[40:]
pbl = MultiClassClf(n_features=3, n_classes=3)
svm = OneSlackSSVM(pbl, C=10)
svm.fit(X_train, Y_train)
predict_no_class_weight = svm.predict(X_test)
pbl_class_weight = MultiClassClf(n_features=3, n_classes=3,
class_weight=np.ones(3))
svm_class_weight = OneSlackSSVM(pbl_class_weight, C=10)
svm_class_weight.fit(X_train, Y_train)
predict_class_weight = svm_class_weight.predict(X_test)
assert_array_equal(predict_no_class_weight, predict_class_weight)
assert_array_almost_equal(svm.w, svm_class_weight.w)
def test_class_weights():
X, Y = make_blobs(n_samples=210, centers=3, random_state=1, cluster_std=3,
shuffle=False)
X = np.hstack([X, np.ones((X.shape[0], 1))])
X, Y = X[:170], Y[:170]
pbl = MultiClassClf(n_features=3, n_classes=3)
svm = OneSlackSSVM(pbl, C=10)
svm.fit(X, Y)
weights = 1. / np.bincount(Y)
weights *= len(weights) / np.sum(weights)
pbl_class_weight = MultiClassClf(n_features=3, n_classes=3,
class_weight=weights)
svm_class_weight = OneSlackSSVM(pbl_class_weight, C=10)
svm_class_weight.fit(X, Y)
assert_greater(f1_score(Y, svm_class_weight.predict(X)),
f1_score(Y, svm.predict(X)))
