def populate_table(table_fname, features, labels, camNames, actNames,
partiNames):
n_samples = labels.shape[0]
pbar = start_progressbar(n_samples, '%i features to Pytable' % (n_samples))
h5 = ta.openFile(table_fname, mode='a')
table = h5.root.input_output_data.readout
pp = table.row
for i in xrange(n_samples):
pp['frame_index'] = i
pp['features'] = features[i, :]
pp['label'] = labels[i]
#pp['aviNames'] = aviNames[i][0:-4]
pp['camNames'] = camNames[i]
pp['actNames'] = actNames[i]
pp['partiNames'] = partiNames[i]
pp.append()
update_progressbar(pbar, i)
end_progressbar(pbar)
# save everything in the file and close it
table.cols.camNames.createIndex()
table.cols.actNames.createIndex()
table.cols.partiNames.createIndex()
table.flush()
h5.close()
def get_bfast_splits(table_fname, settings, n_samples = N_SAMPLES, n_features = N_FEATURES, n_lab = N_LAB, contig_labels = True):
h5_all = ta.openFile(table_fname, mode = 'r')
table_all = h5_all.root.input_output_data.readout
train_p = settings['train_p']
test_p = settings['test_p']
cams = settings['cameras']
#import ipdb; ipdb.set_trace()
#row_train = [table_all.where("(partiNames == '%s') & (camNames == '%s')" % (ll, nn) ) for ll in train_p for nn in cams ]
labels_train = []
features_train = []
labels_test = []
features_test = []
pbar = start_progressbar(len(train_p), str(len(train_p))+ ' training participants' )
for jj,pat in enumerate(train_p):
for cam in cams:
labels_train = labels_train + [row['label'] for row in
table_all.readWhere("(partiNames == '%s') & (camNames == '%s')" % (pat, cam)) ]
features_train = features_train+ [row['features'] for row in table_all.readWhere("(partiNames == '%s') & (camNames == '%s')" % (pat, cam)) ]
update_progressbar(pbar, jj)
end_progressbar(pbar)
pbar = start_progressbar(len(train_p), str(len(train_p))+ ' testing participants' )
for jj,pat in enumerate(test_p):
for cam in cams:
labels_test = labels_test + [row['label'] for row in
table_all.readWhere("(partiNames == '%s') & (camNames == '%s')" % (pat, cam)) ]
features_test = features_test+ [row['features'] for row in table_all.readWhere("(partiNames == '%s') & (camNames == '%s')" % (pat, cam)) ]
update_progressbar(pbar, jj)
end_progressbar(pbar)
tic = time.time()
#uniqLabels = np.intersect1d(np.unique(labels_train), np.unique(labels_test))
uniqLabels = np.unique(labels_train)
#KILL UNUSED
uniqLabels=uniqLabels[uniqLabels!='SIL']
uniqLabels = uniqLabels[:n_lab]
print 'using ',str(len(uniqLabels)),' labels in total'
labels_train = np.array(labels_train)
selector = np.zeros_like(labels_train, dtype= 'bool')
for uL in uniqLabels:
selector = np.squeeze(selector|[labels_train == uL])
labels_train = labels_train[selector]
features_train = np.array(features_train)[selector,:n_features]
labels_test = np.array(labels_test)
selector = np.zeros_like(labels_test, dtype= 'bool')
for uL in uniqLabels:
selector = np.squeeze(selector|[labels_test == uL])
labels_test = labels_test[selector]
features_test = np.array(features_test)[selector,:n_features]
print "Loaded features converted in ", round(time.time() - tic) , "seconds"
table_all.flush()
h5_all.close()
return features_train , labels_train, features_test, labels_test
def train_adaboost(features, labels, learning_rate, n_lab, n_runs, n_estim, n_samples):
uniqLabels = np.unique(labels)
print 'Taking ', str(n_lab), ' labels'
uniqLabels = uniqLabels[:n_lab]
used_labels = uniqLabels
pbar = start_progressbar(len(uniqLabels), 'training adaboost for %i labels' %len(uniqLabels))
allLearners = []
for yy ,targetLab in enumerate(uniqLabels):
runs=[]
for rrr in xrange(n_runs):
#import ipdb;ipdb.set_trace()
feats,labs = get_binary_sets(features, labels, targetLab, n_samples)
#print 'fitting stump'
#import ipdb;ipdb.set_trace()
baseClf = DecisionTreeClassifier(max_depth=4, min_samples_leaf=10, min_samples_split=10)
baseClf.fit(feats, labs)
ada_real = AdaBoostClassifier( base_estimator=baseClf, learning_rate=learning_rate,
n_estimators=n_estim,
algorithm="SAMME.R")
#import ipdb;ipdb.set_trace()
runs.append(ada_real.fit(feats, labs))
allLearners.append(runs)
update_progressbar(pbar, yy)
end_progressbar(pbar)
return allLearners, used_labels
def populate_table(table_fname, features, labels, camNames, actNames, partiNames):
n_samples = labels.shape[0]
pbar = start_progressbar(n_samples, '%i features to Pytable' % (n_samples))
h5 = ta.openFile(table_fname, mode = 'a')
table = h5.root.input_output_data.readout
pp = table.row
for i in xrange(n_samples):
pp['frame_index'] = i
pp['features'] = features[i, :]
pp['label'] = labels[i]
#pp['aviNames'] = aviNames[i][0:-4]
pp['camNames'] = camNames[i]
pp['actNames'] = actNames[i]
pp['partiNames'] = partiNames[i]
pp.append()
update_progressbar(pbar, i)
end_progressbar(pbar)
# save everything in the file and close it
table.cols.camNames.createIndex()
table.cols.actNames.createIndex()
table.cols.partiNames.createIndex()
table.flush()
h5.close()
def compute_confidence(allLearners, dada, classifier_type):
#import ipdb;ipdb.set_trace()
tic = time.time()
#import ipdb;ipdb.set_trace()
if classifier_type == 'adaboost':
lab_confidence = np.zeros([dada.shape[0], len(allLearners)], dtype='float64')
pbar = start_progressbar(len(allLearners), '%i producing weighted outputs' % len(allLearners))
for ii,thisLab in enumerate(allLearners):
res = np.zeros([dada.shape[0]], dtype='float64')
for jj, thisLearner in enumerate(thisLab):
my_weights = thisLearner.estimator_weights_
#tic = time.time()
for hh, thisEstimator in enumerate(thisLearner):
res = res+thisEstimator.predict(dada)*my_weights[hh]
#import ipdb;ipdb.set_trace()
lab_confidence[:,ii] = np.float64(res)
update_progressbar(pbar, ii)
end_progressbar(pbar)
if classifier_type == 'randomforest':
#import ipdb;ipdb.set_trace()
lab_confidence = np.zeros((dada.shape[0],len(allLearners[0].classes_)), dtype='float64')
pbar = start_progressbar(len(allLearners), '%i producing weighted outputs' % len(allLearners[0].classes_))
for ii, thisRun in enumerate(allLearners):
lab_confidence += thisRun.predict_proba(dada)
update_progressbar(pbar, ii)
end_progressbar(pbar)
return lab_confidence
def train_adaboost(features, labels, learning_rate, n_lab, n_runs, n_estim,
n_samples):
uniqLabels = np.unique(labels)
print 'Taking ', str(n_lab), ' labels'
uniqLabels = uniqLabels[:n_lab]
used_labels = uniqLabels
pbar = start_progressbar(
len(uniqLabels), 'training adaboost for %i labels' % len(uniqLabels))
allLearners = []
for yy, targetLab in enumerate(uniqLabels):
runs = []
for rrr in xrange(n_runs):
#import ipdb;ipdb.set_trace()
feats, labs = get_binary_sets(features, labels, targetLab,
n_samples)
#print 'fitting stump'
#import ipdb;ipdb.set_trace()
baseClf = DecisionTreeClassifier(max_depth=1,
min_samples_leaf=4,
min_samples_split=4)
baseClf.fit(feats, labs)
ada_real = AdaBoostClassifier(base_estimator=baseClf,
learning_rate=learning_rate,
n_estimators=n_estim,
algorithm="SAMME.R")
#import ipdb;ipdb.set_trace()
runs.append(ada_real.fit(feats, labs))
allLearners.append(runs)
update_progressbar(pbar, yy)
end_progressbar(pbar)
return allLearners, used_labels
def getMonkeySplits_lim_old(table_fname,
splitNo,
n_samples=N_SAMPLES,
n_features=N_FEATURES):
h5_tr = ta.openFile(table_fname + str(splitNo) + '_train.h5', mode='r')
table_tr = h5_tr.root.input_output_data.readout
h5_te = ta.openFile(table_fname + str(splitNo) + '_test.h5', mode='r')
table_te = h5_te.root.input_output_data.readout
print 'Converting arrays to sp'
uniqLabels = np.unique(table_tr.cols.label)
#KILL UNUSED
uniqLabels = uniqLabels[uniqLabels != 'unused']
labels_train = []
features_train = []
exctCnt = 0
pbar = start_progressbar(len(uniqLabels),
[int(len(uniqLabels)), ' training labels'])
for i, thisLab in enumerate(uniqLabels):
tempLabels = [
row['label'] for row in table_tr.where("label == thisLab")
]
try:
selInd = random.sample(range(0, len(tempLabels)), n_samples)
except ValueError:
selInd = range(0, len(tempLabels))
exctCnt = exctCnt + 1
labels_train = labels_train + [tempLabels[gg] for gg in selInd]
tempFeatures = [
row['features'][:][:n_features]
for row in table_tr.where("label == thisLab")
]
features_train = features_train + [tempFeatures[gg] for gg in selInd]
update_progressbar(pbar, i)
end_progressbar(pbar)
print '%d exceptions occured' % (exctCnt)
features_train = sp.array(features_train)[:, :n_features]
labels_train = sp.array(labels_train)
labels_test = sp.array(table_te.cols.label)
#import ipdb; ipdb.set_trace()
features_test = sp.array(
table_te.cols.features)[labels_test != 'unused', :n_features]
labels_test = labels_test[labels_test != 'unused']
print 'Converted'
#import ipdb; ipdb.set_trace()
table_tr.flush()
table_te.flush()
h5_tr.close()
h5_te.close()
print "feature loading completed"
return features_train, labels_train, features_test, labels_test
def split_data_from_table(table_fname, n_samples = N_SAMPLES, n_features = N_FEATURES):
h5 = ta.openFile(table_fname, mode = 'r')
table = h5.root.input_output_data.readout
l_features = table.cols.features
l_index = table.cols.frame_index
l_labels = table.cols.label
import ipdb; ipdb.set_trace()
n_samples_total = len(l_labels)
assert(2*n_samples < n_samples_total)
import warnings ; warnings.warn("""have something that takes a split depending on classes to keep it balanced""")
#TODO: have a balanced split on each class
ind_total = sp.random.permutation(n_samples_total)
ind_train = ind_total[:n_samples]
ind_test = ind_total[n_samples:2*n_samples]
sp.array([(ind_train == test).any() for test in ind_test]).any()
print "checked that train and test do not overlap"
"""
features_train = features.T[ind_train, :n_features]
labels_train = labels[ind_train]
features_test = features.T[ind_test, :n_features]
labels_test = labels[ind_test]
"""
features_train = sp.zeros((n_samples, n_features), dtype = 'uint8')
features_test = sp.zeros((n_samples, n_features), dtype = 'uint8')
labels_train = []
labels_test = []
pbar = start_progressbar(len(ind_train), '%i train features' % (len(ind_train)))
for i, ind in enumerate(ind_train):
features_train[i] = l_features[ind][:n_features]
labels_train.append(l_labels[ind])
update_progressbar(pbar, i)
end_progressbar(pbar)
pbar = start_progressbar(len(ind_test), '%i test features' % (len(ind_test)))
for i, ind in enumerate(ind_test):
features_test[i] = l_features[ind][:n_features]
labels_test.append(l_labels[i])
update_progressbar(pbar, i)
end_progressbar(pbar)
labels_train = sp.array(labels_train)
labels_test = sp.array(labels_test)
table.flush()
h5.close()
return features_train , labels_train, features_test, labels_test
def getMonkeySplits_lim_old(table_fname, splitNo, n_samples = N_SAMPLES, n_features = N_FEATURES):
h5_tr = ta.openFile(table_fname + str(splitNo) + '_train.h5', mode = 'r')
table_tr = h5_tr.root.input_output_data.readout
h5_te = ta.openFile(table_fname + str(splitNo) + '_test.h5', mode = 'r')
table_te = h5_te.root.input_output_data.readout
print 'Converting arrays to sp'
uniqLabels = np.unique(table_tr.cols.label)
#KILL UNUSED
uniqLabels=uniqLabels[uniqLabels!='unused']
labels_train = []
features_train = []
exctCnt = 0
pbar = start_progressbar(len(uniqLabels), [ int(len(uniqLabels)), ' training labels'] )
for i, thisLab in enumerate(uniqLabels):
tempLabels = [row['label'] for row in table_tr.where("label == thisLab")]
try:
selInd = random.sample(range(0,len(tempLabels)), n_samples)
except ValueError:
selInd = range(0,len(tempLabels))
exctCnt = exctCnt+1
labels_train = labels_train + [tempLabels[gg] for gg in selInd]
tempFeatures = [row['features'][:][:n_features] for row in table_tr.where("label == thisLab")]
features_train = features_train + [tempFeatures[gg] for gg in selInd]
update_progressbar(pbar, i)
end_progressbar(pbar)
print '%d exceptions occured' % (exctCnt)
features_train = sp.array(features_train)[:,:n_features]
labels_train = sp.array(labels_train)
labels_test = sp.array(table_te.cols.label)
#import ipdb; ipdb.set_trace()
features_test = sp.array(table_te.cols.features)[labels_test!='unused',:n_features]
labels_test=labels_test[labels_test!='unused']
print 'Converted'
#import ipdb; ipdb.set_trace()
table_tr.flush()
table_te.flush()
h5_tr.close()
h5_te.close()
print "feature loading completed"
return features_train , labels_train, features_test, labels_test
def populate_table(table_fname, features, labels, names):
n_samples = labels.shape[0]
pbar = start_progressbar(n_samples, '%i features to Pytable' % (n_samples))
h5 = ta.openFile(table_fname, mode='a')
table = h5.root.input_output_data.readout
pp = table.row
for i in xrange(n_samples):
pp['features'] = features[i, :]
pp['label'] = labels[i]
pp.append()
update_progressbar(pbar, i)
end_progressbar(pbar)
# save everything in the file and close it
table.flush()
h5.close()
def populate_table(table_fname, features, labels, names):
n_samples = labels.shape[0]
pbar = start_progressbar(n_samples, '%i features to Pytable' % (n_samples))
h5 = ta.openFile(table_fname, mode = 'a')
table = h5.root.input_output_data.readout
pp = table.row
for i in xrange(n_samples):
pp['features'] = features[i,:]
pp['label'] = labels[i]
pp.append()
update_progressbar(pbar, i)
end_progressbar(pbar)
# save everything in the file and close it
table.flush()
h5.close()
def compute_confidence(allLearners, dada):
#import ipdb;ipdb.set_trace()
lab_confidence = np.zeros([dada.shape[0], len(allLearners)])
tic = time.time()
#import ipdb;ipdb.set_trace()
pbar = start_progressbar(len(allLearners), '%i producing weighted outputs' % len(allLearners))
for ii,thisLab in enumerate(allLearners):
res = np.zeros([dada.shape[0]])
for jj, thisLearner in enumerate(thisLab):
for hh, thisEstimator in enumerate(thisLearner):
#multiply the predictions with the weight of the learner
res = res+thisEstimator.predict(dada)*thisLearner.estimator_weights_[hh]
lab_confidence[:,ii] = res
update_progressbar(pbar, ii)
end_progressbar(pbar)
print "time taken to produce confidence:", round(time.time() - tic,2), "seconds"
#import ipdb;ipdb.set_trace()
return lab_confidence
def compute_confidence(allLearners, dada):
#import ipdb;ipdb.set_trace()
lab_confidence = np.zeros([dada.shape[0], len(allLearners)])
tic = time.time()
#import ipdb;ipdb.set_trace()
pbar = start_progressbar(
len(allLearners), '%i producing weighted outputs' % len(allLearners))
for ii, thisLab in enumerate(allLearners):
res = np.zeros([dada.shape[0]])
for jj, thisLearner in enumerate(thisLab):
for hh, thisEstimator in enumerate(thisLearner):
#multiply the predictions with the weight of the learner
res = res + thisEstimator.predict(
dada) * thisLearner.estimator_weights_[hh]
lab_confidence[:, ii] = res
update_progressbar(pbar, ii)
end_progressbar(pbar)
print "time taken to produce confidence:", round(time.time() - tic,
2), "seconds"
#import ipdb;ipdb.set_trace()
return lab_confidence
def compute_confidence(allLearners, dada, classifier_type):
#import ipdb;ipdb.set_trace()
tic = time.time()
#import ipdb;ipdb.set_trace()
if classifier_type == 'adaboost':
lab_confidence = np.zeros(
[dada.shape[0], len(allLearners)], dtype='float64')
pbar = start_progressbar(
len(allLearners),
'%i producing weighted outputs' % len(allLearners))
for ii, thisLab in enumerate(allLearners):
res = np.zeros([dada.shape[0]], dtype='float64')
for jj, thisLearner in enumerate(thisLab):
my_weights = thisLearner.estimator_weights_
#tic = time.time()
for hh, thisEstimator in enumerate(thisLearner):
res = res + thisEstimator.predict(dada) * my_weights[hh]
#import ipdb;ipdb.set_trace()
lab_confidence[:, ii] = np.float64(res)
update_progressbar(pbar, ii)
end_progressbar(pbar)
if classifier_type == 'randomforest':
#import ipdb;ipdb.set_trace()
lab_confidence = np.zeros(
(dada.shape[0], len(allLearners[0].classes_)), dtype='float64')
pbar = start_progressbar(
len(allLearners),
'%i producing weighted outputs' % len(allLearners[0].classes_))
for ii, thisRun in enumerate(allLearners):
lab_confidence += thisRun.predict_proba(dada)
update_progressbar(pbar, ii)
end_progressbar(pbar)
return lab_confidence
def euclidien(features1, features2=None):
"""
Builds a similarity matrix based ont eh euclidien distance
"""
if features2 is None:
features2 = features1
nfeat1 = len(features1)
nfeat2 = len(features2)
# go
kernelmatrix = sp.empty((nfeat1, nfeat2), dtype="float")
if features1 is features2:
# set up progress bar
n_iter = 0
niter = nfeat1 * (nfeat2 + 1) / 2
pbar = start_progressbar(niter, "Kernel Train")
for ifeat1, feat1 in enumerate(features1):
a_2 = (feat1**2.).sum()
for ifeat2, feat2 in enumerate(features2):
if ifeat1 == ifeat2:
kernelmatrix[ifeat1, ifeat2] = 0
# mattrix symmetric, do only top triangle
elif ifeat1 > ifeat2:
a_b = sp.dot(feat1, feat2.T)
b_2 = (feat2**2.).sum()
dist = (a_2 - 2 * a_b + b_2)
# since kernel matrix is symmetric
kernelmatrix[ifeat1, ifeat2] = dist
kernelmatrix[ifeat2, ifeat1] = dist
update_progressbar(pbar, n_iter + 1)
n_iter += 1
else:
# set up progress bar
n_iter = 0
niter = nfeat1 * nfeat2
pbar = start_progressbar(niter, "Kernel Test")
for ifeat1, feat1 in enumerate(features1):
a_2 = (feat1**2.).sum()
for ifeat2, feat2 in enumerate(features2):
a_b = sp.dot(feat1, feat2.T)
b_2 = (feat2**2.).sum()
dist = (a_2 - 2 * a_b + b_2)
kernelmatrix[ifeat1, ifeat2] = dist
pbar = update_progressbar(pbar, n_iter)
n_iter += 1
end_progressbar(pbar)
return kernelmatrix
def get_monkey_splits_lim(table_fname, splitNo, n_samples = N_SAMPLES, n_features = N_FEATURES, n_lab = N_LAB, contig_labels = True):
h5_tr = ta.openFile(table_fname + str(splitNo) + '_train.h5', mode = 'r')
table_tr = h5_tr.root.input_output_data.readout
h5_te = ta.openFile(table_fname + str(splitNo) + '_test.h5', mode = 'r')
table_te = h5_te.root.input_output_data.readout
# if label_focus == 'test':
# uniqLabels = np.unique(table_te.cols.label)
# else:
# uniqLabels = np.unique(table_tr.cols.label)
uniqLabels = np.intersect1d(np.unique(table_te.cols.label), np.unique(table_tr.cols.label))
#KILL UNUSED
uniqLabels=uniqLabels[uniqLabels!='unused']
uniqLabels = uniqLabels[:n_lab]
labels_train = []
features_train = []
exctCnt = 0
pbar = start_progressbar(len(uniqLabels), 'fetching %i training labels' %len(uniqLabels))
for i, thisLab in enumerate(uniqLabels):
tempLabels = [row['label'] for row in table_tr.where("label == thisLab")]
if contig_labels:
toThis = min(len(tempLabels), n_samples)
selInd = range(0,toThis)
else:
try:
selInd = random.sample(range(0,len(tempLabels)), n_samples)
except ValueError:
selInd = range(0,len(tempLabels))
exctCnt = exctCnt+1
labels_train = labels_train + [tempLabels[gg] for gg in selInd]
tempFeatures = [row['features'][:][:n_features] for row in table_tr.where("label == thisLab")]
features_train = features_train + [tempFeatures[gg] for gg in selInd]
update_progressbar(pbar, i)
end_progressbar(pbar)
#import ipdb; ipdb.set_trace()
print '%d exceptions occured' % (exctCnt)
pbar = start_progressbar(len(uniqLabels), 'fetching %i testing labels' %len(uniqLabels))
labels_test = []
features_test = []
for i, thisLab in enumerate(uniqLabels):
tempLabels = [row['label'] for row in table_te.where("label == thisLab")]
labels_test = labels_test + tempLabels
tempFeatures = [row['features'][:][:n_features] for row in table_te.where("label == thisLab")]
features_test = features_test + tempFeatures
update_progressbar(pbar, i)
end_progressbar(pbar)
features_train = sp.array(features_train)[:,:n_features]
labels_train = sp.array(labels_train)
features_test = sp.array(features_test)[:,:n_features]
labels_test = sp.array(labels_test)
print 'Converted'
table_tr.flush()
table_te.flush()
h5_tr.close()
h5_te.close()
print "feature loading completed"
return features_train , labels_train, features_test, labels_test
def get_HMDB_splits(table_fname, vidName, vidMode, n_samples = N_SAMPLES, n_features = N_FEATURES):
h5 = ta.openFile(table_fname, mode = 'r')
table = h5.root.input_output_data.readout
l_features = table.cols.features
l_index = table.cols.frame_index
l_labels = table.cols.label
l_aviNames = table.cols.aviNames
# assert(2*n_samples < n_samples_total)
trVidName = [];
teVidName = [];
noVidName = [];
for j in range(0, len(vidMode)-1):
innerVidMode = vidMode[j]
#import ipdb; ipdb.set_trace()
trVidName = trVidName + [vidName[j][i] for i, x in enumerate(innerVidMode) if x == '1']
teVidName = teVidName + [vidName[j][i] for i, x in enumerate(innerVidMode) if x == '2']
noVidName = noVidName + [vidName[j][i] for i, x in enumerate(innerVidMode) if x == '0']
features_train =[]
labels_train = []
features_test = []
labels_test = []
#import ipdb; ipdb.set_trace()
exctCnt = 0
pbar = start_progressbar(len(trVidName), '%i train features' % (len(trVidName)))
for i, vid in enumerate(trVidName):
tempLabels = [row['label'] for row in table.where("aviNames == vid")]
try:
selInd = random.sample(range(0,len(tempLabels)), n_samples)
except ValueError:
selInd = range(0,len(tempLabels))
exctCnt = exctCnt+1
labels_train = labels_train + [tempLabels[gg] for gg in selInd]
tempFeatures = [row['features'][:][:n_features] for row in table.where("aviNames == vid")]
features_train = features_train + [tempFeatures[gg] for gg in selInd]
update_progressbar(pbar, i)
end_progressbar(pbar)
print'finished with %i exceptions' % (exctCnt)
pbar = start_progressbar(len(teVidName), '%i test features' % (len(teVidName)))
for i, vid in enumerate(teVidName):
labels_test = labels_test + [row['label'] for row in table.where("aviNames == vid")]
features_test = features_test + [row['features'][:][:n_features] for row in table.where("aviNames == vid")]
update_progressbar(pbar, i)
end_progressbar(pbar)
print 'Converting arrays to sp'
features_train = sp.array(features_train, dtype = 'uint8')
features_test = sp.array(features_test, dtype = 'uint8')
labels_train = sp.array(labels_train)
labels_test = sp.array(labels_test)
print 'Converted'
table.flush()
h5.close()
print "feature loading completed"
return features_train , labels_train, features_test, labels_test
def euclidien(features1, features2 = None):
"""
Builds a similarity matrix based ont eh euclidien distance
"""
if features2 is None:
features2 = features1
nfeat1 = len(features1)
nfeat2 = len(features2)
# go
kernelmatrix = sp.empty((nfeat1, nfeat2), dtype="float")
if features1 is features2:
# set up progress bar
n_iter = 0
niter = nfeat1 * (nfeat2+1) / 2
pbar = start_progressbar(niter, "Kernel Train")
for ifeat1, feat1 in enumerate(features1):
a_2 = (feat1**2.).sum()
for ifeat2, feat2 in enumerate(features2):
if ifeat1 == ifeat2:
kernelmatrix[ifeat1, ifeat2] = 0
# mattrix symmetric, do only top triangle
elif ifeat1 > ifeat2:
a_b = sp.dot(feat1, feat2.T)
b_2 = (feat2**2.).sum()
dist = (a_2 - 2 *a_b + b_2)
# since kernel matrix is symmetric
kernelmatrix[ifeat1, ifeat2] = dist
kernelmatrix[ifeat2, ifeat1] = dist
update_progressbar(pbar, n_iter+1)
n_iter += 1
else:
# set up progress bar
n_iter = 0
niter = nfeat1 * nfeat2
pbar = start_progressbar(niter, "Kernel Test")
for ifeat1, feat1 in enumerate(features1):
a_2 = (feat1**2.).sum()
for ifeat2, feat2 in enumerate(features2):
a_b = sp.dot(feat1, feat2.T)
b_2 = (feat2**2.).sum()
dist = (a_2 - 2 *a_b + b_2)
kernelmatrix[ifeat1, ifeat2] = dist
pbar = update_progressbar(pbar, n_iter)
n_iter += 1
end_progressbar(pbar)
return kernelmatrix
def get_bfast_splits(table_fname,
settings,
n_samples=N_SAMPLES,
n_features=N_FEATURES,
n_lab=N_LAB,
contig_labels=True):
h5_all = ta.openFile(table_fname, mode='r')
table_all = h5_all.root.input_output_data.readout
train_p = settings['train_p']
test_p = settings['test_p']
cams = settings['cameras']
#import ipdb; ipdb.set_trace()
#row_train = [table_all.where("(partiNames == '%s') & (camNames == '%s')" % (ll, nn) ) for ll in train_p for nn in cams ]
labels_train = []
features_train = []
labels_test = []
features_test = []
pbar = start_progressbar(len(train_p),
str(len(train_p)) + ' training participants')
for jj, pat in enumerate(train_p):
for cam in cams:
labels_train = labels_train + [
row['label'] for row in table_all.readWhere(
"(partiNames == '%s') & (camNames == '%s')" % (pat, cam))
]
features_train = features_train + [
row['features'] for row in table_all.readWhere(
"(partiNames == '%s') & (camNames == '%s')" % (pat, cam))
]
update_progressbar(pbar, jj)
end_progressbar(pbar)
pbar = start_progressbar(len(train_p),
str(len(train_p)) + ' testing participants')
for jj, pat in enumerate(test_p):
for cam in cams:
labels_test = labels_test + [
row['label'] for row in table_all.readWhere(
"(partiNames == '%s') & (camNames == '%s')" % (pat, cam))
]
features_test = features_test + [
row['features'] for row in table_all.readWhere(
"(partiNames == '%s') & (camNames == '%s')" % (pat, cam))
]
update_progressbar(pbar, jj)
end_progressbar(pbar)
tic = time.time()
#uniqLabels = np.intersect1d(np.unique(labels_train), np.unique(labels_test))
uniqLabels = np.unique(labels_train)
#KILL UNUSED
uniqLabels = uniqLabels[uniqLabels != 'SIL']
uniqLabels = uniqLabels[:n_lab]
print 'using ', str(len(uniqLabels)), ' labels in total'
labels_train = np.array(labels_train)
selector = np.zeros_like(labels_train, dtype='bool')
for uL in uniqLabels:
selector = np.squeeze(selector | [labels_train == uL])
labels_train = labels_train[selector]
features_train = np.array(features_train)[selector, :n_features]
labels_test = np.array(labels_test)
selector = np.zeros_like(labels_test, dtype='bool')
for uL in uniqLabels:
selector = np.squeeze(selector | [labels_test == uL])
labels_test = labels_test[selector]
features_test = np.array(features_test)[selector, :n_features]
print "Loaded features converted in ", round(time.time() - tic), "seconds"
table_all.flush()
h5_all.close()
return features_train, labels_train, features_test, labels_test
def get_bfast_splits(table_fname, settings, n_samples = N_SAMPLES, n_features = N_FEATURES, n_lab = N_LAB, contig_labels = True):
h5_all = ta.openFile(table_fname, mode = 'r')
table_all = h5_all.root.input_output_data.readout
train_p = settings['train_p']
test_p = settings['test_p']
cams = settings['cameras']
cur_cam = settings['cur_cam']
lab_per_cam = {}
print 'figuring out the shared labels between cams'
for cam in cams:
lab_per_pat = []
for pat in train_p:
rowz = [row['label'] for row in table_all.readWhere("(partiNames == '%s') & (camNames == '%s')" % (pat, cam)) ]
lab_per_pat += list(np.unique(rowz))
lab_per_cam[cam] = np.unique(lab_per_pat)
#import ipdb; ipdb.set_trace()
aaa = lab_per_cam.values() #avoiding a possible bug here by making sure we take only the labels that exist for all cams
labs_for_all = set(aaa[0]).intersection(*aaa) # will use this towards the end to weed out un-shared ones
len_data = 0
print 'figuring out how many training samples we have for the cam ', cur_cam
for pat in train_p:
rowz = [row['label'] for row in table_all.readWhere("(partiNames == '%s') & (camNames == '%s')" % (pat, cur_cam)) ]
len_data += len(rowz)
features_train = np.empty((len_data,n_features), dtype='float64')
labels_train = np.empty(len_data, dtype='|S24')
cnt =0
pbar = start_progressbar(len(train_p), str(len(train_p))+ ' training participants loading for cam '+cur_cam )
for jj,pat in enumerate(train_p):
temp = [row['features'] for row in table_all.readWhere("(partiNames == '%s') & (camNames == '%s')" % (pat, cur_cam)) ]
temp2 = [row['label'] for row in table_all.readWhere("(partiNames == '%s') & (camNames == '%s')" % (pat, cur_cam)) ]
temp = [roww[:n_features] for roww in temp]
if temp:
#features_train[cnt:cnt+len(temp),:] = np.array(temp)[:,:n_features]
features_train[cnt:cnt+len(temp),:] = temp
labels_train[cnt:cnt+len(temp)] = temp2
cnt = cnt+len(temp)
update_progressbar(pbar, jj)
end_progressbar(pbar)
len_data = 0
pbar = start_progressbar(len(test_p), ' now figuring out how many test samples we have' )
for jj,pat in enumerate(test_p):
for cam in cams:
len_data += len([row['label'] for row in table_all.readWhere("(partiNames == '%s') & (camNames == '%s')" % (pat, cam)) ])
update_progressbar(pbar, jj)
end_progressbar(pbar)
features_test = np.empty((len_data,n_features), dtype='float64')
labels_test = np.empty(len_data, dtype='|S24')
cnt =0
pbar2 = start_progressbar(len(test_p), str(len(test_p))+ ' testing participants loading' )
for jj,pat in enumerate(test_p):
for cam in cams:
temp = [row['features'] for row in table_all.readWhere("(partiNames == '%s') & (camNames == '%s')" % (pat, cam)) ]
temp2 = [row['label'] for row in table_all.readWhere("(partiNames == '%s') & (camNames == '%s')" % (pat, cam)) ]
temp = [roww[:n_features] for roww in temp]
if temp:
#features_train[cnt:cnt+len(temp),:] = np.array(temp)[:,:n_features]
features_test[cnt:cnt+len(temp),:] = temp
labels_test[cnt:cnt+len(temp)] = temp2
#import ipdb; ipdb.set_trace()
cnt = cnt+len(temp)
update_progressbar(pbar2, jj)
end_progressbar(pbar2)
tic = time.time()
uniqLabels = np.intersect1d(labs_for_all, np.unique(labels_test))
#KILL UNUSED
uniqLabels = uniqLabels[uniqLabels!='SIL']
uniqLabels = uniqLabels[:n_lab]
print 'using ',str(len(uniqLabels)),' labels in total'
labels_train = np.array(labels_train)
selector = np.zeros_like(labels_train, dtype= 'bool')
excpt = 0
for uL in uniqLabels:
label_all = labels_train == uL
label_all_subs = np.where(label_all)[0]
if label_all_subs.shape >= n_samples:
label_some_subs = label_all_subs[:n_samples]
else:
excpt += 1
label_some_subs = label_all_subs
label_lim = np.zeros_like(label_all,dtype='bool')
label_lim[label_some_subs] = True
selector = np.squeeze(selector|[label_lim])
labels_train = labels_train[selector]
features_train = features_train[selector,:n_features]
labels_test = np.array(labels_test)
selector = np.zeros_like(labels_test, dtype= 'bool')
for uL in uniqLabels:
selector = np.squeeze(selector|[labels_test == uL])
labels_test = labels_test[selector]
features_test = features_test[selector,:n_features]
print "Loaded features converted in ", round(time.time() - tic) , "seconds"
print "there were ", str(excpt), " exceptions "
table_all.flush()
h5_all.close()
labels_train = group_labels(labels_train)
labels_test = group_labels(labels_test)
print 'using ',str(len(labels_test)),' labels in total'
return features_train , labels_train, features_test, labels_test
def split_data_from_table(table_fname,
n_samples=N_SAMPLES,
n_features=N_FEATURES):
h5 = ta.openFile(table_fname, mode='r')
table = h5.root.input_output_data.readout
l_features = table.cols.features
l_index = table.cols.frame_index
l_labels = table.cols.label
import ipdb
ipdb.set_trace()
n_samples_total = len(l_labels)
assert (2 * n_samples < n_samples_total)
import warnings
warnings.warn(
"""have something that takes a split depending on classes to keep it balanced"""
)
#TODO: have a balanced split on each class
ind_total = sp.random.permutation(n_samples_total)
ind_train = ind_total[:n_samples]
ind_test = ind_total[n_samples:2 * n_samples]
sp.array([(ind_train == test).any() for test in ind_test]).any()
print "checked that train and test do not overlap"
"""
features_train = features.T[ind_train, :n_features]
labels_train = labels[ind_train]
features_test = features.T[ind_test, :n_features]
labels_test = labels[ind_test]
"""
features_train = sp.zeros((n_samples, n_features), dtype='uint8')
features_test = sp.zeros((n_samples, n_features), dtype='uint8')
labels_train = []
labels_test = []
pbar = start_progressbar(len(ind_train),
'%i train features' % (len(ind_train)))
for i, ind in enumerate(ind_train):
features_train[i] = l_features[ind][:n_features]
labels_train.append(l_labels[ind])
update_progressbar(pbar, i)
end_progressbar(pbar)
pbar = start_progressbar(len(ind_test),
'%i test features' % (len(ind_test)))
for i, ind in enumerate(ind_test):
features_test[i] = l_features[ind][:n_features]
labels_test.append(l_labels[i])
update_progressbar(pbar, i)
end_progressbar(pbar)
labels_train = sp.array(labels_train)
labels_test = sp.array(labels_test)
table.flush()
h5.close()
return features_train, labels_train, features_test, labels_test
def get_monkey_splits_lim(table_fname,
splitNo,
n_samples=N_SAMPLES,
n_features=N_FEATURES,
n_lab=N_LAB,
contig_labels=True):
h5_tr = ta.openFile(table_fname + str(splitNo) + '_train.h5', mode='r')
table_tr = h5_tr.root.input_output_data.readout
h5_te = ta.openFile(table_fname + str(splitNo) + '_test.h5', mode='r')
table_te = h5_te.root.input_output_data.readout
# if label_focus == 'test':
# uniqLabels = np.unique(table_te.cols.label)
# else:
# uniqLabels = np.unique(table_tr.cols.label)
uniqLabels = np.intersect1d(np.unique(table_te.cols.label),
np.unique(table_tr.cols.label))
#KILL UNUSED
uniqLabels = uniqLabels[uniqLabels != 'unused']
uniqLabels = uniqLabels[:n_lab]
labels_train = []
features_train = []
exctCnt = 0
pbar = start_progressbar(len(uniqLabels),
'fetching %i training labels' % len(uniqLabels))
for i, thisLab in enumerate(uniqLabels):
tempLabels = [
row['label'] for row in table_tr.where("label == thisLab")
]
if contig_labels:
toThis = min(len(tempLabels), n_samples)
selInd = range(0, toThis)
else:
try:
selInd = random.sample(range(0, len(tempLabels)), n_samples)
except ValueError:
selInd = range(0, len(tempLabels))
exctCnt = exctCnt + 1
labels_train = labels_train + [tempLabels[gg] for gg in selInd]
tempFeatures = [
row['features'][:][:n_features]
for row in table_tr.where("label == thisLab")
]
features_train = features_train + [tempFeatures[gg] for gg in selInd]
update_progressbar(pbar, i)
end_progressbar(pbar)
#import ipdb; ipdb.set_trace()
print '%d exceptions occured' % (exctCnt)
pbar = start_progressbar(len(uniqLabels),
'fetching %i testing labels' % len(uniqLabels))
labels_test = []
features_test = []
for i, thisLab in enumerate(uniqLabels):
tempLabels = [
row['label'] for row in table_te.where("label == thisLab")
]
labels_test = labels_test + tempLabels
tempFeatures = [
row['features'][:][:n_features]
for row in table_te.where("label == thisLab")
]
features_test = features_test + tempFeatures
update_progressbar(pbar, i)
end_progressbar(pbar)
features_train = sp.array(features_train)[:, :n_features]
labels_train = sp.array(labels_train)
features_test = sp.array(features_test)[:, :n_features]
labels_test = sp.array(labels_test)
print 'Converted'
table_tr.flush()
table_te.flush()
h5_tr.close()
h5_te.close()
print "feature loading completed"
return features_train, labels_train, features_test, labels_test
def get_HMDB_splits(table_fname,
vidName,
vidMode,
n_samples=N_SAMPLES,
n_features=N_FEATURES):
h5 = ta.openFile(table_fname, mode='r')
table = h5.root.input_output_data.readout
l_features = table.cols.features
l_index = table.cols.frame_index
l_labels = table.cols.label
l_aviNames = table.cols.aviNames
# assert(2*n_samples < n_samples_total)
trVidName = []
teVidName = []
noVidName = []
for j in range(0, len(vidMode) - 1):
innerVidMode = vidMode[j]
#import ipdb; ipdb.set_trace()
trVidName = trVidName + [
vidName[j][i] for i, x in enumerate(innerVidMode) if x == '1'
]
teVidName = teVidName + [
vidName[j][i] for i, x in enumerate(innerVidMode) if x == '2'
]
noVidName = noVidName + [
vidName[j][i] for i, x in enumerate(innerVidMode) if x == '0'
]
features_train = []
labels_train = []
features_test = []
labels_test = []
#import ipdb; ipdb.set_trace()
exctCnt = 0
pbar = start_progressbar(len(trVidName),
'%i train features' % (len(trVidName)))
for i, vid in enumerate(trVidName):
tempLabels = [row['label'] for row in table.where("aviNames == vid")]
try:
selInd = random.sample(range(0, len(tempLabels)), n_samples)
except ValueError:
selInd = range(0, len(tempLabels))
exctCnt = exctCnt + 1
labels_train = labels_train + [tempLabels[gg] for gg in selInd]
tempFeatures = [
row['features'][:][:n_features]
for row in table.where("aviNames == vid")
]
features_train = features_train + [tempFeatures[gg] for gg in selInd]
update_progressbar(pbar, i)
end_progressbar(pbar)
print 'finished with %i exceptions' % (exctCnt)
pbar = start_progressbar(len(teVidName),
'%i test features' % (len(teVidName)))
for i, vid in enumerate(teVidName):
labels_test = labels_test + [
row['label'] for row in table.where("aviNames == vid")
]
features_test = features_test + [
row['features'][:][:n_features]
for row in table.where("aviNames == vid")
]
update_progressbar(pbar, i)
end_progressbar(pbar)
print 'Converting arrays to sp'
features_train = sp.array(features_train, dtype='uint8')
features_test = sp.array(features_test, dtype='uint8')
labels_train = sp.array(labels_train)
labels_test = sp.array(labels_test)
print 'Converted'
table.flush()
h5.close()
print "feature loading completed"
return features_train, labels_train, features_test, labels_test
