def find_scores_Z(N,old_seeds,degrees): scores = [] diffs = [] timespents = [] avgdegs = [] for i in range(len(N)): seeds = old_seeds + [i] diffscore = 0.0 timespentscore = 0.0 for j in range(len(seeds)): seed = seeds[j] for other_seed in seeds[j:]: diffscore += float(abs(degrees[seed] - degrees[other_seed])) for other_seed in seeds: if seed == other_seed: continue timespentscore += N[seed][other_seed] diffs.append(diffscore) timespents.append(timespentscore) avgdegsum = 0 for seed in seeds: avgdegsum += degrees[seed] avgdeg = avgdegsum/len(seeds) avgdegs.append(avgdeg) print timespents[76] print timespents[58] diffs = stats.zscore(diffs) timespents = stats.zscore(timespents) avgdegs = stats.zscore(avgdegs) for i in range(len(diffs)): scores.append(diffs[i] + timespents[i] - avgdegs[i]) return scores
def collaspe_fclusters(data=None, t=None, row_labels=None, col_labels=None, linkage='average', pdist='euclidean', standardize=3, log=False): """a function to collaspe flat clusters by averaging the vectors within each flat clusters achieved from hierarchical clustering""" ## preprocess data if log: data = np.log2(data + 1.0) if standardize == 1: # Standardize along the columns of data data = zscore(data, axis=0) elif standardize == 2: # Standardize along the rows of data data = zscore(data, axis=1) if row_labels is not None and col_labels is None: ## only get fclusters for rows d = dist.pdist(data, metric=pdist) axis = 1 ##!!! haven't checked whether this is correct yet elif row_labels is None and col_labels is not None: ## only get fclusters for cols d = dist.pdist(data.T, metric=pdist) axis = 0 D = dist.squareform(d) Y = sch.linkage(D, method=linkage, metric=pdist) fclusters = sch.fcluster(Y, t, 'distance') fcluster_set = set(fclusters) data_cf = [] for fc in fcluster_set: mask = np.where(fclusters==fc) data_t = data.T vector_avg = np.average(data_t[mask],axis=axis) data_cf.append(vector_avg) data_cf = np.array(data_cf).T return data_cf
def main(argv):
category_predictions = {}
for cat in categories:
file_reg = classifier_reg + "/" + cat + "/predictions"
fo = open(file_reg)
reg_predictions = [float(x) for x in fo.read().split('\n')[:-1]]
reg_zscores = stats.zscore(reg_predictions)
file_red = classifier_red + "/" + cat + "/predictions"
fo = open(file_red)
red_predictions = [float(x) for x in fo.read().split('\n')[:-1]]
red_zscores = stats.zscore(red_predictions)
max_zscore = []
# Take zscore that has the greatest deviation out of the two classifiers
for i in range(len(reg_predictions)):
reg_greater = abs(reg_zscores[i]) >= abs(red_zscores[i])
val = reg_zscores[i] if reg_greater else red_zscores[i]
max_zscore.append(val)
category_predictions[cat] = max_zscore
results_list = output_predictions(categories, category_predictions)
calculate_accuracies(results_list)
def algorithm(w1,w2,w3,w4,G1,G2,G3,G4):
try:
cc=np.array([nx.average_clustering(G1,weight='weight'),nx.average_clustering(G2,weight='weight'),nx.average_clustering(G3,weight='weight'),nx.average_clustering(G4,weight='weight')])
spl=np.array([nx.average_shortest_path_length(G1,weight='weight'),nx.average_shortest_path_length(G2,weight='weight'),nx.average_shortest_path_length(G3,weight='weight'),nx.average_shortest_path_length(G4,weight='weight')])
nds=np.array([nx.number_of_nodes(G1),nx.number_of_nodes(G2),nx.number_of_nodes(G3),nx.number_of_nodes(G4)])
edgs= np.array([nx.number_of_edges(G1),nx.number_of_edges(G2),nx.number_of_edges(G3),nx.number_of_edges(G4)])
if valid(cc):
cc=stats.zscore(cc)
else:
cc=np.array([.1,.1,.1,.1])
cc= cc-min(cc)+.1
if valid(spl):
spl=stats.zscore(spl)
else:
spl=np.array([.1,.1,.1,.1])
spl= spl-min(spl)+.1
if valid(nds):
nds=stats.zscore(nds)
else:
nds=np.array([.1,.1,.1,.1])
nds = nds-min(nds)+.1
if valid(edgs):
edgs=stats.zscore(edgs)
else:
edgs=np.array([.1,.1,.1,.1])
edgs=edgs-min(edgs)+.1
r1=(w1*cc[0]+w2*spl[0]+w3*nds[0]+w4*edgs[0])*1000
r2=(w1*cc[1]+w2*spl[1]+w3*nds[1]+w4*edgs[1])*1000
r3=(w1*cc[2]+w2*spl[2]+w3*nds[2]+w4*edgs[2])*1000
r4=(w1*cc[3]+w2*spl[3]+w3*nds[3]+w4*edgs[3])*1000
d={'Player 1:': r1, 'Player 2:': r2,'Player 3:': r3, 'Player 4:': r4}
rank = sorted(d.items(), key=lambda x: x[1], reverse=True)
return ["USAU RANKINGS",str(rank[0][0])+ " " + str(int(rank[0][1])),str(rank[1][0])+" "+ str(int(rank[1][1])),str(rank[2][0])+" "+ str(int(rank[2][1])),str(rank[3][0])+" "+str(int(rank[3][1]))]
except:
return ["Unable to compute rankings! Need data","Player 1","Player 2","Player 3","Player 4"]
def correct_covariates(Dtrait, Dcov, variables):
Dcomb = pd.merge(Dtrait.T, Dcov.T, left_index=True, right_index=True).T
Dcorr = Dtrait.copy()
traits = Dtrait.columns.values.tolist()
print 'Correcting for %s' % variables
for idx, i in enumerate(Dtrait.index):
sys.stdout.write('\rTrait %d of %d' % (idx, Dtrait.shape[0]))
sys.stdout.flush()
if len(variables) == 1:
rlm_model = sm.RLM(Dcomb.loc[i,:], zscore(array(Dcomb.loc[variables,:]).T))
else:
rlm_model = sm.RLM(Dcomb.loc[i,:], zscore(array(Dcomb.loc[variables,:]).T, axis=0))
rlm_results = rlm_model.fit()
Dcorr.loc[i,:] = rlm_results.resid
"""
if idx > 1:
f, axarr = subplots(3,2)
axarr[0,0].scatter(Dtrait.loc['EIF1AY',:], Dtrait.loc['OSBP', :])
axarr[0,1].scatter(Dcorr.loc['EIF1AY',:], Dcorr.loc['OSBP', :])
axarr[1,0].hist([x for x in Dtrait.loc['EIF1AY',:] if not isnan(x)])
axarr[1,1].hist([x for x in Dcorr.loc['EIF1AY',:] if not isnan(x)])
axarr[2,0].hist([x for x in Dtrait.loc['OSBP',:] if not isnan(x)])
axarr[2,1].hist([x for x in Dcorr.loc['OSBP',:] if not isnan(x)])
f2, axarr2 = subplots(3,2)
axarr2[0,0].scatter(Dcomb.loc['gender',:], Dcomb.loc['OSBP', :])
axarr2[1,0].scatter(Dcomb.loc['age',:], Dcomb.loc['OSBP', :])
axarr2[2,0].scatter(Dcomb.loc['site',:], Dcomb.loc['OSBP', :])
axarr2[0,1].scatter(Dcomb.loc['gender',:], Dcomb.loc['EIF1AY', :])
axarr2[1,1].scatter(Dcomb.loc['age',:], Dcomb.loc['EIF1AY', :])
axarr2[2,1].scatter(Dcomb.loc['site',:], Dcomb.loc['EIF1AY', :])
show()
exit()
"""
return Dcorr
def predict_loo(transformed_data, args):
print 'mysseg loo',
sys.stdout.flush()
(ndim, nsample , nsubjs) = transformed_data.shape
tst_subj = args.loo
win_size = args.winsize
nseg = nsample - win_size
# mysseg prediction prediction
trn_data = np.zeros((ndim*win_size, nseg))
# the trn data also include the tst data, but will be subtracted when
# calculating A
for m in range(nsubjs):
for w in range(win_size):
trn_data[w*ndim:(w+1)*ndim,:] += transformed_data[:,w:(w+nseg),m]
tst_data = np.zeros((ndim*win_size, nseg))
for w in range(win_size):
tst_data[w*ndim:(w+1)*ndim,:] = transformed_data[:,w:(w+nseg),tst_subj]
A = stats.zscore((trn_data - tst_data),axis=0, ddof=1)
B = stats.zscore(tst_data,axis=0, ddof=1)
corr_mtx = B.T.dot(A)
for i in range(nseg):
for j in range(nseg):
if abs(i-j)<win_size and i != j :
corr_mtx[i,j] = -np.inf
rank = np.argmax(corr_mtx, axis=1)
accu = sum(rank == range(nseg)) / float(nseg)
return accu
def hist_and_smooth_data(spike_data):
max_spike_ts = 0
for i in range(len(spike_data)):
if np.amax(spike_data[i]) > max_spike_ts:
max_spike_ts = np.amax(spike_data[i])
max_bin_num = int(np.ceil(max_spike_ts) / float(bin_size) * 1000)
hist_data = np.zeros((len(spike_data),max_bin_num))
hist_bins = np.zeros((len(spike_data),max_bin_num))
for i in range(len(spike_data)):
total_bin_range = np.arange(0,int(np.
ceil(spike_data[i].max())),bin_size/1000.0)
hist,bins = np.histogram(spike_data[i],bins=total_bin_range,range=(0,int(np.ceil(spike_data[i].max()))),normed=False,density=False)
#pdb.set_trace()
hist_data[i,0:len(hist)] = hist
hist_bins[i,0:len(bins)] = bins
#TODO fix so gaus divide by bin size and -> fr before smoothing
#TODO make option for zscore and gaus togethre
if zscore_bool and gaussian_bool:
smoothed = stats.zscore(hist_data,axis=1)
smoothed = ndimage.filters.gaussian_filter1d(smoothed,gauss_sigma,axis=1)
elif zscore_bool:
smoothed = stats.zscore(hist_data,axis=1)
elif gaussian_bool:
smoothed = ndimage.filters.gaussian_filter1d(hist_data,gauss_sigma,axis=1)
else:
smoothed = {}
return_dict = {'hist_data':hist_data,'hist_bins':hist_bins,'smoothed':smoothed}
return(return_dict)
def plot_features_distribution(feature_set,
feature_set_permutation,
save_path,
prename='features',
n_features=90,
n_bins=20):
plt.figure()
h_values_p, _ = np.histogram(feature_set_permutation.flatten(),
bins=np.arange(0, n_features+1))
plt.hist(zscore(h_values_p), bins=n_bins)
fname = "%s_features_set_permutation_distribution.png" % (prename)
plt.savefig(os.path.join(save_path,
fname))
plt.figure()
h_values_, _ = np.histogram(feature_set.flatten(),
bins=np.arange(0, n_features+1))
plt.plot(zscore(h_values_))
fname = "%s_features_set_cross_validation.png" % (prename)
plt.savefig(os.path.join(save_path,
fname))
plt.close('all')
def knnClassifier(training_data, test_data, training_target, test_target, k=5):
#normalize the data
#calculate the z-score of the data
#print training_data
training_data = training_data
new_training_data = stats.zscore(training_data.astype(int), axis=0)
new_test_data = stats.zscore(test_data.astype(int), axis=0)
#find the k nearest neighbors for each test data
#print 'test', new_test_data
predictions = []
for test in new_test_data:
#print test
# find the euclidean distance between the test case and all training cases
distances = []
neighbors = []
neighbor_predictions = []
for train in new_training_data:
#print train
distances.append(np.linalg.norm(train-test))
#print distances
for i in range(k):
neighb_i = distances.index(min(distances))
neighbors.append(neighb_i)
distances[neighb_i] = 1000000
#print neighbors
for neighb in neighbors:
neighbor_predictions.append(training_target[neighb])
predictions.append(stats.mode(neighbor_predictions)[0][0])
return predictions
def ExtractDataVer2(all_relevant_channels, marker_positions, target, ms_before, ms_after):
target_idx = marker_positions[np.where(target == 1)[0]] - 1
all_target_transpose = np.asarray(all_relevant_channels).T
number_positive_of_samples = len(target_idx)
before_trigger = int((ms_before * 1.0) / 5)
after_trigger = int((ms_after * 1.0) / 5)
all_target_data = extractTimeWindowFast(all_target_transpose, target_idx, before_trigger, after_trigger)
non_target_idx = marker_positions[np.where(target == 0)[0]] - 1
number_positive_of_samples = len(non_target_idx)
all_non_target_data = extractTimeWindowFast(all_target_transpose, non_target_idx, before_trigger, after_trigger)
# normalize the data over the time axe
all_data = np.vstack((stats.zscore(all_target_data, axis=1).astype('float32'),
stats.zscore(all_non_target_data, axis=1).astype('float32')))
all_tags = np.vstack((np.ones((all_target_data.shape[0], 1), dtype='int8'),
np.zeros((all_non_target_data.shape[0], 1), dtype='int8')))
return all_data, all_tags
def data_parser(theta,kappa,tt,ch,tt_ch):
theta_r = np.array([[resample(theta.values.squeeze()[i,950:1440],50)] for i in range(0,theta.shape[0])])
theta_r = zscore(theta_r.squeeze(),axis=None)
kappa_r = np.array([[resample(kappa.values.squeeze()[i,950:1440],50)] for i in range(0,kappa.shape[0])])
kappa_r = zscore(kappa_r.squeeze(),axis=None)
kappa_df = pd.DataFrame(kappa_r)
theta_df = pd.DataFrame(theta_r)
both_df = pd.concat([theta_df,kappa_df],axis=1)
if tt_ch == 'tt':
# trial type
clean = np.nan_to_num(tt) !=0
tt_c = tt[clean.squeeze()].values
else :
# choice
clean = np.nan_to_num(ch) !=0
tt_c = ch[clean.squeeze()].values
# tt_c = tt[tt.values !=0|3].values
both = both_df.values
# both_c = both[clean.squeeze(),:]
both_c = both[clean.squeeze(),:]
# keeping one hot vector for now (incase we want it later)
# labs = np.eye(3)[tt_c.astype(int)-1]
# y[np.arange(3), a] = 1
# labs = labs.squeeze()
return both_c, tt_c, clean
def significantly_unenriched(xs, ys, zthresh=2., scale='linear'):
assert scale in ['linear', 'log']
if scale =='log':
xs = np.log2(xs)
ys = np.log2(ys)
xs = stats.zscore(xs)
ys = stats.zscore(ys)
return [x < -zthresh or y < -zthresh for x, y in zip(xs, ys)]
def simulate(self, beta0, beta, x):
if(self.distr=='poisson'):
y = np.random.poisson(self.lmb(beta0, beta, zscore(x)))
if(self.distr=='normal'):
y = np.random.normal(self.lmb(beta0, beta, zscore(x)))
if(self.distr=='binomial'):
y = np.random.binomial(1, self.lmb(beta0, beta, zscore(x)))
return y
def _normalize_network(self, x, square=True):
if square:
norm_col = zscore(x, axis=0)
return (norm_col + norm_col.T) / math.sqrt(2)
else:
norm_col = zscore(x, axis=0)
norm_row = zscore(x, axis=1)
return (norm_col + norm_row) / math.sqrt(2)
def combineTT(data1,data2):
iA,iB = findOLIndices(data1.geneList,data2.geneList)
newGeneList = [data1.geneList[i] for i in iA]
newSampleList = data1.cleanSampIDs + data2.cleanSampIDs
newCnData = np.column_stack((data1.cnData[iA,:],data2.cnData[iB,:]))
newRnaData = np.column_stack((ss.zscore(data1.rnaData[iA,:],1),ss.zscore(data2.rnaData[iB,:],1)))
newData = combinedData(newGeneList,newSampleList,newRnaData,newCnData)
return newData
def five_second_z_score(start):
for interval in range(len(matrix[start])/5000 + 1):
j = 5000*interval
# print interval, j
if interval == len(matrix[start])/5000 + 1:
z_score_five_sec.append(stats.zscore(matrix[start][j:]))
else:
z_score_five_sec.append(stats.zscore(matrix[start][j:j+5000]))
def np_fisherZ(x,y,r):
z = 0.5 * (np.log(1.0 + r) - np.log(1.0 - r))
w = np.sqrt(len(x)) * z
x_ = zscore(x)
y_ = zscore(y)
t2 = moment22(x_,y_)
t = np.sqrt(t2)
p = 2. * (1. - norm.cdf(np.abs(w), 0.0, t))
return p
def simulate(self, beta0, beta, X):
"""Simulate data."""
if self.distr == 'poisson':
y = np.random.poisson(self.lmb(beta0, beta, zscore(X)))
if self.distr == 'normal':
y = np.random.normal(self.lmb(beta0, beta, zscore(X)))
if self.distr == 'binomial':
y = np.random.binomial(1, self.lmb(beta0, beta, zscore(X)))
return y
def euclid(v1,v2):
from scipy.stats import zscore
'''Squared Euclidean Distance between two scalars or equally matched vectors
USAGE: d = euclid(v1,v2)'''
v1 = zscore(v1.flatten())
v2 = zscore(v2.flatten())
d2= np.sqrt(np.sum((v1-v2)**2))
return d2
def plot(pred, y):
predm, ym = stats.zscore(pred, axis=0, ddof=1), stats.zscore(y, axis=0, ddof=1)
# predm, ym = pp.normalize(pred), pp.normalize(y)
print predm.shape, ym.shape
times = range(0, len(pred))
# print times.shape
print len(times)
plt.plot(times, predm, "r-", times, ym, "b-")
plt.xlabel("Time (s)")
plt.ylabel("BOLD Response")
plt.savefig("../figure/lassoplot.jpg")
def iClustergram(data=None, row_labels=None, col_labels=None,
row_groups=None, col_groups=None,
row_linkage='average', col_linkage='average',
row_pdist='euclidean', col_pdist='euclidean',
standardize=None, log=False,
display_range=3, username='******', apikey='fmnoxd2t2u'):
## preprocess data
if log:
data = np.log2(data + 1.0)
if standardize == 1: # Standardize along the columns of data
data = zscore(data, axis=0)
elif standardize == 2: # Standardize along the rows of data
data = zscore(data, axis=1)
## cluster data:
## compute pdist for rows
d1 = dist.pdist(data, metric=row_pdist)
D1 = dist.squareform(d1)
Y1 = sch.linkage(D1, method=row_linkage, metric=row_pdist)
Z1 = sch.dendrogram(Y1, orientation='right')
idx1 = Z1['leaves']
## compute pdist for cols
d2 = dist.pdist(data.T, metric=col_pdist)
D2 = dist.squareform(d2)
Y2 = sch.linkage(D2, method=col_linkage, metric=col_pdist)
Z2 = sch.dendrogram(Y2)
idx2 = Z2['leaves']
## transform the orders of data to clustered data
data_clustered = data
data_clustered = data_clustered[:,idx2]
data_clustered = data_clustered[idx1,:]
data_to_plot = data_clustered.tolist()
## transform the orders of row and col labels
new_row_labels = []
new_col_labels = []
for i in range(data.shape[0]):
new_row_labels.append(row_labels[idx1[i]])
for i in range(data.shape[1]):
new_col_labels.append(col_labels[idx2[i]])
## plot clustered data using plotly
py = plotly.plotly(username, apikey)
d = {}
d['x'] = new_row_labels
d['y'] = new_col_labels
d['z'] = data_to_plot
d['type'] = 'heatmap'
py.plot([d])
return
def plot_fclusters(data=None, row_labels=None, col_labels=None,
linkage='average', pdist='euclidean', standardize=3, log=False):
"""a function to plot the relationship between thresholds and number of
flat clusters achieved from hierarchical clustering, aims to find the optimal
threshold for forming clusters"""
## preprocess data
if log:
data = np.log2(data + 1.0)
if standardize == 1: # Standardize along the columns of data
data = zscore(data, axis=0)
elif standardize == 2: # Standardize along the rows of data
data = zscore(data, axis=1)
fig = plt.figure()
ax1 = fig.add_subplot(121)
ax2 = fig.add_subplot(122)
if row_labels is not None and col_labels is None: ## only get fclusters for rows
d = dist.pdist(data, metric=pdist)
elif row_labels is None and col_labels is not None: ## only get fclusters for cols
d = dist.pdist(data.T, metric=pdist)
D = dist.squareform(d)
Y = sch.linkage(D, method=linkage, metric=pdist)
space1 = np.linspace(d.min(), d.max(), num=5, endpoint=False)
space2 = np.linspace(d.max(),1.,num=30,endpoint=True)
thresholds = np.concatenate((space1,space2))
num_clusters = []
num_singles = []
for t in thresholds:
fclusters = sch.fcluster(Y, t,'distance')
c = Counter(fclusters)
num_cluster = len(c.keys())
num_single = c.values().count(1)
num_clusters.append(num_cluster)
num_singles.append(num_single)
print 'threshold=', t, 'clusters:', num_cluster, 'singles:',num_single
if num_cluster < 290:
print c
ax1.plot(thresholds, num_clusters,label='# of flat clusters')
ax1.plot(thresholds, num_singles,label='# of singles',c='r')
ax1.plot(thresholds, np.array(num_clusters)-np.array(num_singles),label='# of non-singles',c='g')
ax1.legend(loc='upper right')
ax1.set_xlabel('threshold for forming flat clusters')
ax2.plot(thresholds, num_clusters,label='# of flat clusters')
ax2.plot(thresholds, num_singles,label='# of singles',c='r')
ax2.plot(thresholds, np.array(num_clusters)-np.array(num_singles),label='# of non-singles',c='g')
ax2.legend(loc='upper right')
ax2.set_xlabel('threshold for forming flat clusters')
ax2.set_yscale('log')
plt.show()
return
def _divide_to_regions(info, add_stim=True):
"""Divide channels to regions by positions."""
from scipy.stats import zscore
picks = _pick_data_channels(info, exclude=[])
chs_in_lobe = len(picks) // 4
pos = np.array([ch['loc'][:3] for ch in info['chs']])
x, y, z = pos.T
frontal = picks[np.argsort(y[picks])[-chs_in_lobe:]]
picks = np.setdiff1d(picks, frontal)
occipital = picks[np.argsort(y[picks])[:chs_in_lobe]]
picks = np.setdiff1d(picks, occipital)
temporal = picks[np.argsort(z[picks])[:chs_in_lobe]]
picks = np.setdiff1d(picks, temporal)
lt, rt = _divide_side(temporal, x)
lf, rf = _divide_side(frontal, x)
lo, ro = _divide_side(occipital, x)
lp, rp = _divide_side(picks, x) # Parietal lobe from the remaining picks.
# Because of the way the sides are divided, there may be outliers in the
# temporal lobes. Here we switch the sides for these outliers. For other
# lobes it is not a big problem because of the vicinity of the lobes.
with np.errstate(invalid='ignore'): # invalid division, greater compare
zs = np.abs(zscore(x[rt]))
outliers = np.array(rt)[np.where(zs > 2.)[0]]
rt = list(np.setdiff1d(rt, outliers))
with np.errstate(invalid='ignore'): # invalid division, greater compare
zs = np.abs(zscore(x[lt]))
outliers = np.append(outliers, (np.array(lt)[np.where(zs > 2.)[0]]))
lt = list(np.setdiff1d(lt, outliers))
l_mean = np.mean(x[lt])
r_mean = np.mean(x[rt])
for outlier in outliers:
if abs(l_mean - x[outlier]) < abs(r_mean - x[outlier]):
lt.append(outlier)
else:
rt.append(outlier)
if add_stim:
stim_ch = _get_stim_channel(None, info, raise_error=False)
if len(stim_ch) > 0:
for region in [lf, rf, lo, ro, lp, rp, lt, rt]:
region.append(info['ch_names'].index(stim_ch[0]))
return {'Left-frontal': lf, 'Right-frontal': rf, 'Left-parietal': lp,
'Right-parietal': rp, 'Left-occipital': lo, 'Right-occipital': ro,
'Left-temporal': lt, 'Right-temporal': rt}
def clustergram(data, rids, cids,
row_linkage='average', col_linkage='average',
row_pdist='euclidean', col_pdist='euclidean',
standardize=3, log=False):
## preprocess data
if log:
data = np.log2(data + 1.0)
if standardize == 1: # Standardize along the columns of data
data = zscore(data, axis=0)
elif standardize == 2: # Standardize along the rows of data
data = zscore(data, axis=1)
## perform hierarchical clustering for rows and cols
## compute pdist for rows:
d1 = dist.pdist(data, metric=row_pdist)
D1 = dist.squareform(d1)
Y1 = sch.linkage(D1, method=row_linkage, metric=row_pdist)
Z1 = sch.dendrogram(Y1, orientation='right')
idx1 = Z1['leaves']
## compute pdist for cols
d2 = dist.pdist(data.T, metric=col_pdist)
D2 = dist.squareform(d2)
Y2 = sch.linkage(D2, method=col_linkage, metric=col_pdist)
Z2 = sch.dendrogram(Y2)
idx2 = Z2['leaves']
row_nodes = []
rids = np.array(rids)[np.array(idx1)]
for idx, rid in enumerate(rids):
row_nodes.append({'sort': idx, 'name': rid})
col_nodes = []
cids = np.array(cids)[np.array(idx2)]
for idx, cid in enumerate(cids):
col_nodes.append({'sort': idx, 'name': cid})
links = []
for i in range(len(rids)):
for j in range(len(cids)):
links.append({'source': i, 'target': j, 'value': data[i,j]})
json_data = {
'row_nodes':row_nodes,
'col_nodes':col_nodes,
'links': links
}
return json_data
def trend_zscore(self,sym,date,window):
slice = self.trends[sym][-window:]
if slice[-1] == slice[-2]:
z = self.zscores[sym][-1]
else:
z = zscore(slice)[-1]
return z
def find_outliers(X, threshold=3.0):
"""Find outliers based on Gaussian mixture
Parameters
----------
X : np.ndarray of float, shape (n_elemenets,)
The scores for which to find outliers.
threshold : float
The value above which a feature is classified as outlier.
Returns
-------
bad_idx : np.ndarray of int, shape (n_features)
The outlier indices.
"""
max_iter = 2
my_mask = np.zeros(len(X), dtype=np.bool)
X = np.abs(X)
for _ in range(max_iter):
X = np.ma.masked_array(X, my_mask)
this_z = stats.zscore(X)
local_bad = this_z > threshold
my_mask = np.max([my_mask, local_bad], 0)
if not np.any(local_bad):
break
bad_idx = np.where(my_mask)[0]
return bad_idx
def align(movie_data, options, args, lrh):
print 'pPCA(scikit-learn)'
nvoxel = movie_data.shape[0]
nTR = movie_data.shape[1]
nsubjs = movie_data.shape[2]
align_algo = args.align_algo
nfeature = args.nfeature
# zscore the data
bX = np.nan((nsubjs*nvoxel,nTR))
for m in xrange(nsubjs):
bX[m*nvoxel:(m+1)*nvoxel,:] = stats.zscore(movie_data[:, :, m].T, axis=0, ddof=1).T
del movie_data
U, s, VT = np.linalg.svd(bX, full_matrices=False)
bW = np.zeros((nsubjs*nvoxel,nfeature))
for m in xrange(nsubjs):
bW[m*nvoxel:(m+1)*nvoxel,:] = U[m*nvoxel:(m+1)*nvoxel,:nfeature]
niter = 10
# initialization when first time run the algorithm
np.savez_compressed(options['working_path']+align_algo+'_'+lrh+'_'+str(niter)+'.npz',\
bW = bW, niter=niter)
return niter
def _generate_noise_system(dimensions_tr,
):
"""Generate the scanner noise
Generate the noise that is typical of a scanner. This is comprised
of two types of noise, Rician and Gaussian
Parameters
----------
dimensions_tr : n length array, int
What are the dimensions of the volume you wish to insert
noise into. This can be a volume of any size
Returns
----------
system_noise : multidimensional array, float
Create a volume with system noise
"""
# Generate the Rician noise
noise_rician = stats.rice.rvs(1, 1, size=dimensions_tr)
# Apply the gaussian noise
noise_gaussian = np.random.normal(0, 1, size=dimensions_tr)
# Combine these two noise types
noise_system = noise_rician + noise_gaussian
# Normalize
noise_system = stats.zscore(noise_system)
return noise_system
def triplet_data_collection(data, tags, batch_size, select=3, outof=10):
from scipy import stats
stimuli_category_size = 30
number_of_repetition = select
magic_number = number_of_repetition * stimuli_category_size
number_of_samples = data.shape[0]
time_samples_dim_size = data.shape[2]
channel_dim_size = data.shape[3]
all_combination = _get_all_possible_combination(np.arange(number_of_samples), outof, select)
shuffled_combination = np.random.permutation(all_combination)
batch_data = np.zeros((magic_number, batch_size, time_samples_dim_size, channel_dim_size), dtype=np.float32)
counter = 0
for i in range(0,len(shuffled_combination), batch_size):
batch_tags = np.zeros((batch_size, stimuli_category_size), dtype=np.int8)
for single_combination in shuffled_combination[i:min(i +batch_size,len(shuffled_combination) )]:
batch_data[:, counter, :, :] = np.vstack([data[item] for item in single_combination])
batch_tags[counter] = np.mean(np.vstack([tags[item] for item in single_combination]), axis=0)
counter += 1
if counter == batch_size:
input_dict = dict(
[["positive_item_input_{}".format(i), stats.zscore(batch_data[i], axis=1)] for i in
range(90)])
input_dict['triplet_loss'] = batch_tags
return input_dict
def plot_transition_clustermap(data_array, gene_names, pseudotimes, n_clusters=10, gradient=False):
if gradient:
data_to_plot = zscore(np.gradient(data_array)[1].T, axis=0)
scale = None
metric = 'seuclidean'
row_linkage = linkage(pdist(abs(data_to_plot), metric=metric), method='complete')
else:
data_to_plot = data_array.T
scale = 0
metric = 'correlation'
row_linkage = linkage(pdist(data_to_plot, metric=metric), method='complete')
assignments = fcluster(row_linkage, n_clusters, criterion='maxclust')
cm = sns.clustermap(data_to_plot, col_cluster=False, standard_scale=scale,
yticklabels=gene_names, row_linkage=row_linkage,
row_colors=[settings.STATE_COLORS[i] for i in assignments])
r = np.arange(10, data_array.shape[0], data_array.shape[0]/10)
plt.setp(cm.ax_heatmap.get_yticklabels(), fontsize=5)
cm.ax_heatmap.set_xticks(r)
cm.ax_heatmap.set_xticklabels(['%.1f' % x for x in pseudotimes[r]])
cm.ax_heatmap.set_xlabel('Pseudotime')
cm.ax_heatmap.set_ylabel('Gene')
gene_clusters = defaultdict(list)
for i, cl in enumerate(assignments):
gene_clusters[settings.STATE_COLORS[cl]].append(gene_names[i])
return gene_clusters
wine_ds[col] = np.log1p(wine_ds[col]) # In[75]: wine_ds.skew() # In[76]: sns.pairplot(wine_ds) plt.show() # In[77]: from scipy.stats import zscore z_score = abs(zscore(wine_ds)) print(wine_ds.shape) wine_ds_final = wine_ds.loc[(z_score < 3).all(axis=1)] print(wine_ds_final.shape) # In[78]: #Separatomg target and input variables df_x = wine_ds_final.drop(columns=['Class']) y = wine_ds_final['Class'] # In[79]: df_x
np.random.seed(42)
n_splits = 10
# Read data
#print("Start loading data @ %.5f\n" % (time.time()-elapsed))
cwd = os.getcwd()
file_name_feature = cwd + "/../dataset/bank-additional-full_new_features.csv"
file_name_label = cwd + "/../dataset/bank-additional-full_new_labels.csv"
#print("End loading data @ %.5f\n" % (time.time()-elapsed))
#print("Start shuffling and sampling @ %.5f\n" % (time.time()-elapsed))
features, header_ele = readData(file_name_feature)
features = shuffle(features, random_state=41)[:5000]
features[:, :9] = zscore(features[:, :9])
label, label_names = loadLabels(file_name_label)
label = shuffle(label, random_state=41)[:5000]
#print("End shuffling and sampling @ %.5f\n" % (time.time()-elapsed))
#print("Start splitting dataset with 10-folds @ %.5f\n" % (time.time()-elapsed))
Kfold = StratifiedKFold(n_splits=n_splits)
#print("End splitting dataset with 10-folds @ %.5f\n" % (time.time()-elapsed))
accuracy_training_log = np.zeros(n_splits)
accuracy_testing_log = np.zeros(n_splits)
nlpd_t = np.zeros(n_splits)
nlpd_v = np.zeros(n_splits)
for i, (train_index, test_index) in enumerate(Kfold.split(features, label)):
}) # S marker size
# Set title
plt.title('plot between Garage Area and SalPrice')
plt.xlim(-200, 1600)
# Set x-axis label
plt.xlabel('GarageArea')
# Set y-axis label
plt.ylabel('SalePrice')
plt.show()
# Removing the Anamolies using z-score
# if the data is more than 3 standard deviations away then it is considered as outlier
# if the data is less than -3 standard deviations away then it is considered as outlier
df = pd.read_csv('train.csv', sep=',', usecols=(62, 80))
z = np.abs(stats.zscore(df))
threshold = 3
print(np.where(z > 3))
modified_df = df[(z < 3).all(axis=1)]
print(df.shape)
print(modified_df.shape)
# Create Scatterplot of the dataframe after removing the anamolies in the data
sns.lmplot(
'GarageArea', # Horizontal axis
'SalePrice', # Vertical axis
data=modified_df, # Data source
fit_reg=False, # Don't fix a regression line
scatter_kws={
"marker": "o", # Set marker style
"s": 80
import warnings
warnings.filterwarnings('ignore')
df_train = pd.read_csv(
'C:/Users/Sushu/Documents/Python/ICP6/Python_Lesson6/train.csv')
df_train.describe()
var = 'GarageArea'
data = pd.concat([df_train['SalePrice'], df_train[var]], axis=1)
data.plot.scatter(x=var, y='SalePrice', ylim=(0, 800000))
data.shape
data_remana = np.abs(stats.zscore(data))
data_remana[:5, :5]
data1 = data[(data_remana < 3).all(axis=1)]
var = 'GarageArea'
data = pd.concat([df_train['SalePrice'], df_train[var]], axis=1)
data1.plot.scatter(x=var, y='SalePrice', ylim=(0, 800000))
df = data[~data.SalePrice.isin(data[data_remana > 3].SalePrice)]
df.describe()
data1.shape
def runPatterns(actmat,
method='ica',
nullhyp='mp',
nshu=1000,
percentile=99,
tracywidom=False):
'''
INPUTS
actmat: activity matrix - numpy array (neurons, time bins)
nullhyp: defines how to generate statistical threshold for assembly detection.
'bin' - bin shuffling, will shuffle time bins of each neuron independently
'circ' - circular shuffling, will shift time bins of each neuron independently
obs: mantains (virtually) autocorrelations
'mp' - Marcenko-Pastur distribution - analytical threshold
nshu: defines how many shuffling controls will be done (n/a if nullhyp is 'mp')
percentile: defines which percentile to be used use when shuffling methods are employed.
(n/a if nullhyp is 'mp')
tracywidow: determines if Tracy-Widom is used. See Peyrache et al 2010.
(n/a if nullhyp is NOT 'mp')
OUTPUTS
patterns: co-activation patterns (assemblies) - numpy array (assemblies, neurons)
significance: object containing general information about significance tests
zactmat: returns z-scored actmat
'''
nneurons = np.size(actmat, 0)
nbins = np.size(actmat, 1)
silentneurons = np.var(actmat, axis=1) == 0
actmat_ = actmat[~silentneurons, :]
# z-scoring activity matrix
zactmat_ = stats.zscore(actmat_, axis=1)
# running significance (estimating number of assemblies)
significance = PCA()
significance.fit(zactmat_.T)
significance.nneurons = nneurons
significance.nbins = nbins
significance.nshu = nshu
significance.percentile = percentile
significance.tracywidom = tracywidom
significance.nullhyp = nullhyp
significance = runSignificance(zactmat_, significance)
if np.isnan(significance.nassemblies):
return
if significance.nassemblies < 1:
print('WARNING !')
print(' no assembly detecded!')
patterns = []
else:
# extracting co-activation patterns
patterns_ = extractPatterns(zactmat_, significance, method)
if patterns_ is np.nan:
return
# putting eventual silent neurons back (their assembly weights are defined as zero)
patterns = np.zeros((np.size(patterns_, 0), nneurons))
patterns[:, ~silentneurons] = patterns_
zactmat = np.copy(actmat)
zactmat[~silentneurons, :] = zactmat_
return patterns, significance, zactmat
def demand_graphs(branch='A',
product_line='Electronic accessories',
date=datetime.date(2019, 3, 25)):
pd.options.plotting.backend = "plotly"
fig = make_subplots(rows=1,
cols=2,
shared_xaxes=False,
specs=[[{
"type": "xy"
}, {
"type": "xy"
}]],
subplot_titles=('Day vs Demand',
'Predicted Values for a Week'))
fig['layout'].update(xaxis_title='Day',
yaxis_title='Quantity',
paper_bgcolor='rgba(0,0,0,0)',
plot_bgcolor='rgba(0,0,0,0)')
#Gets the earliest date in the database
with connection.cursor() as crs:
earliest_date_query = 'SELECT min(Date) FROM Sales'
earliest_date = crs.execute(earliest_date_query).fetchval()
df = pd.read_sql(
demand_query % (earliest_date, earliest_date, date, branch,
product_line, earliest_date), connection)
date_difference = date - earliest_date
#fills in days where there are no sales
s = df['Day'].tolist()
for i in range(0, date_difference.days):
if i not in s:
df.loc[-1] = [i, 0]
df.index = df.index + 1
#drops data that has a zscore with an absolute value greater than or equal to 3 then sorts
z_scores = np.abs(stats.zscore(df['Demand']))
df = df[(z_scores < 3)]
df = df.sort_values('Day')
#makes a linear regression model then predicts the values and store them
model = linear_model.LinearRegression()
weight = np.ones(len(df)) * 10
weight[-7:] *= 1.5
x = df[['Day']]
y = df[['Demand']]
model.fit(x, y, weight)
df['bestfit'] = model.predict(df[['Day']])
#makes a prediction of the next weeks demand
days_predicted = list(range(date_difference.days,
date_difference.days + 7))
predicted_values = []
total_predicted = 0
for i in days_predicted:
predicted = float(model.predict([[i]]))
predicted_values.append(predicted)
total_predicted += predicted
predicted_values.append(total_predicted)
#makes a scatter plot with a line for the linear regression
fig.add_trace(
go.Scatter(name='data points',
x=df['Day'],
y=df['Demand'].values,
mode='markers'), 1, 1)
fig.add_trace(
go.Scatter(name='regression line',
x=df['Day'],
y=df['bestfit'],
mode='lines'), 1, 1)
#makes a bar chart for the predicted values for the week
fig.add_trace(
go.Bar(name='predicted values',
x=days_predicted.append('total'),
y=predicted_values), 1, 2)
#finds the explained variance score, r2 score, and mean absolute error
evs = sm.explained_variance_score(df['Demand'], df['bestfit'])
r2 = sm.r2_score(df['Demand'], df['bestfit'])
mae = sm.mean_absolute_error(df['Demand'], df['bestfit'])
new_date = True
with connection.cursor() as crs:
crs.execute(check_prediction_query, (date, product_line, branch))
if crs.fetchone() != None:
new_date = False
if new_date:
with connection.cursor() as crs:
crs.execute(update_prediction_log,
(branch, product_line, date, evs, r2, mae))
return fig, evs, r2, mae
ix = obj1.meta.type.isin(['iPSC'])
obj1.filter_samples(ix)
obj2 = copy(ref_obj)
ix = obj2.meta.type == 'ESC'
obj2.filter_samples(ix)
dend = plot_dendrogram([obj1, obj2], qn_method=quantile_norm, n_by_mad=n_gene_by_mad)
dend['fig'].savefig(os.path.join(outdir, "cluster_ipsc_esc.png"), dpi=200)
# 3. iPSC, ESC, Ruiz signature (only)
the_obj = loader.MultipleBatchLoader([obj1, obj2])
dat_r_z = pd.DataFrame(np.log2(the_obj.data + eps))
dat_r_z = dat_r_z.reindex(gene_sign_ens.values).dropna()
for r in dat_r_z.index:
dat_r_z.loc[r] = zscore(dat_r_z.loc[r])
dat_r_z.index = gene_sign_ens.index[gene_sign_ens.isin(dat_r_z.index)]
cg = clustering.plot_clustermap(dat_r_z, show_gene_labels=True, cmap='RdBu_r')
cg.gs.update(bottom=0.2)
cg.savefig(os.path.join(outdir, "clustermap_ruiz_ipsc_esc_ztrans.png"), dpi=200)
# 4. HipSci, iPSC, ESC, FB
obj1 = copy(obj)
ix = obj1.meta.type.isin(['iPSC', 'FB'])
obj1.filter_samples(ix)
dend = plot_dendrogram([obj1, ref_obj, hip_obj], qn_method=quantile_norm, n_by_mad=n_gene_by_mad)
dend['fig'].savefig(os.path.join(outdir, "cluster_ipsc_esc_fb_with_hipsci%d.png" % n_hipsci), dpi=200)
def Price_Main(data: pd.DataFrame):
# Remove price and term outliers (out of 3 sigmas)
data = data[((np.abs(stats.zscore(data.price)) < 2.5) & (np.abs(stats.zscore(data.term)) < 2.5) & (
np.abs(stats.zscore(data.full_sq)) < 2.5))]
# Fill NaN if it appears after merging
data[['term']] = data[['term']].fillna(data[['term']].mean())
# Fix year
data = data[((data.yyyy_announce == 19) | (data.yyyy_announce == 20))]
# Log Transformation
data["longitude"] = np.log1p(data["longitude"])
data["latitude"] = np.log1p(data["latitude"])
data["full_sq"] = np.log1p(data["full_sq"])
data["life_sq"] = np.log1p(data["life_sq"])
data["kitchen_sq"] = np.log1p(data["kitchen_sq"])
data["to_center"] = np.log1p(data["to_center"])
data["price"] = np.log1p(data["price"])
X = data[['life_sq', 'to_center', 'mm_announce', 'rooms', 'renovation', 'has_elevator', 'longitude', 'latitude', 'full_sq', 'kitchen_sq',
'time_to_metro', 'floor_last', 'floor_first', 'clusters', 'is_rented', 'rent_quarter', 'rent_year']]
y = data[['price']].values.ravel()
print(X.shape, y.shape, flush=True)
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=1)
# GBR model
gbr_model = GradientBoostingRegressor(n_estimators=350, max_depth=8, verbose=1, random_state=42)
print(10*'-', '> GBR Spb started fitting...')
gbr_model.fit(X_train, y_train)
gbr_preds = gbr_model.predict(X_test)
print('Spb GBR R2_score: ', r2_score(y_test, gbr_preds), flush=True)
print('Spb GBR RMSE : ', mean_squared_error(y_test, gbr_preds), flush=True)
print('Train GBR on full spb dataset: ', flush=True)
gbr_model.fit(X, y)
dump(gbr_model, PATH_TO_PRICE_MODEL_GBR_D)
print('GBR Spb model saved !', flush=True)
# RANDOM FOREST REGRESSOR
RF = RandomForestRegressor(n_estimators=300, verbose=1, n_jobs=-1)
print(10*'-', '> Rf Spb started fitting...')
RF.fit(X_train, y_train)
rf_predicts = RF.predict(X_test)
print('Spb RF R2_score: ', r2_score(y_test, rf_predicts), flush=True)
print('Spb RF RMSE: ', mean_squared_error(y_test, rf_predicts), flush=True)
print('Train RF on full spb dataset: ', flush=True)
RF.fit(X, y)
dump(RF, PATH_TO_PRICE_MODEL_RF_D)
print('GBR Spb model saved !', flush=True)
# LGBM model
lgbm_model = LGBMRegressor(objective='regression',
learning_rate=0.05,
n_estimators=1250, max_depth=7, min_child_samples=1, verbose=0)
print(10*'-', '> LGBM Spb started fitting...')
lgbm_model.fit(X_train, y_train)
lgbm_preds = lgbm_model.predict(X_test)
print('Spb RF R2_score: ', r2_score(y_test, lgbm_preds), flush=True)
print('Spb LGBM RMSE: ', mean_squared_error(y_test, lgbm_preds), flush=True)
print('Train LGBM on full spb dataset: ', flush=True)
lgbm_model.fit(X, y)
dump(lgbm_model, PATH_TO_PRICE_MODEL_LGBM_D)
print('LGBM Spb model saved !', flush=True)
import numpy as np
import pandas as pd
from scipy import stats
# load predict -ligKi matrix
ki_matrix = np.loadtxt("..//..//02_drug_model//03_predicted_ki_matrix.txt", delimiter=",")
# generate zscore matrix
z_score_matrix = []
# calculate zscore
counter = 1
for prot in ki_matrix:
print(counter)
z_score = stats.zscore(prot)
z_score[z_score < 2] = 0
z_score_matrix.append(z_score)
counter += 1
z_score_matrix = np.array(z_score_matrix)
np.savetxt(".//protein_ki_zscore_2_matrix.txt", z_score_matrix, fmt = "%.5e", delimiter=",")
# result = [result1,result2,result3]
fig = plt.figure()
for j in range(len(result)):
dataset = pd.concat([base,result[j]])
n_comp = 2
pca = PCA(n_components=n_comp)
principalComponents = pca.fit_transform(dataset)
columns = ['principal component '+str(i) for i in range(1,n_comp+1)]
principalDf = pd.DataFrame(data = principalComponents , columns = columns)
finalDf = pd.concat([principalDf, classe], axis = 1)
finalDf = finalDf.dropna()
z = np.abs(stats.zscore(finalDf.iloc[:,:2]))
finalDf = finalDf[(z < 3).all(axis=1)]
finalDf.reset_index(drop=True)
loadings = pca.components_.T * np.sqrt(pca.explained_variance_)
for k in range(n_comp):
loadings_ = loadings[:,k]
# smooth2(smoo, loadings_)
smoothed = smooth2(smoo, loadings_)
peaksf, _ = find_peaks(smoothed,distance=10)
plt.plot(xData,smoothed,label='L'+str(1+k)+' '+label[i]+' vs '+classe[0][dataset.index[-1]]+'(var: %d%%)' %(pca.explained_variance_ratio_[k]*100),color=colors[1:len(colors)][j],alpha=(n_comp-k)/(n_comp))
plt.hlines(0, xmin=xData.min(),xmax=xData.max(),ls='dotted',linewidth=1)
plt.xlabel('Raman Shift (cm$^{-1}$)', fontsize = 16)
plt.ylabel('PCA Loadings', fontsize = 16)
Using vectorized functions When performance is paramount, you should avoid using .apply() and .map() because those constructs perform Python for-loops over the data stored in a pandas Series or DataFrame. By using vectorized functions instead, you can loop over the data at the same speed as compiled code (C, Fortran, etc.)! NumPy, SciPy and pandas come with a variety of vectorized functions (called Universal Functions or UFuncs in NumPy). You can even write your own vectorized functions, but for now we will focus on the ones distributed by NumPy and pandas. In this exercise you're going to import the zscore method from scipy.stats and use it to compute the deviation in voter turnout in Pennsylvania from the mean in fractions of the standard deviation. In statistics, the z-score is the number of standard deviations by which an observation is above the mean - so if it is negative, it means the observation is below the mean. Instead of using .apply() as you did in the earlier exercises, the zscore UFunc will take a pandas Series as input and return a NumPy array. You will then assign the values of the NumPy array to a new column in the DataFrame. You will be working with the election DataFrame - it has been pre-loaded for you. Import zscore from scipy.stats. Call zscore with election['turnout'] as input . Print the output of type(turnout_zscore). This has been done for you. Assign turnout_zscore to a new column in election as 'turnout_zscore'. Print the output of election.head(). This has been done for you, so hit 'Submit Answer' to view the result. ''' # Import zscore from scipy.stats from scipy.stats import zscore # Call zscore with election['turnout'] as input: turnout_zscore turnout_zscore = zscore(election['turnout']) # Print the type of turnout_zscore print(type(turnout_zscore)) # Assign turnout_zscore to a new column: election['turnout_zscore'] election['turnout_zscore'] = turnout_zscore # Print the output of election.head() print(election.head())
def prepareForBoxplots(data, gender=True):
"""
brings data into the correct format for generating a boxplot over all students from all years
in this particular case we use year of university entrance as query_id, psu scores as features and notas as rank
writes one dataset with protection_status "gender" and one with protection_status "highschool_type"
"""
data = data[data['sem'] == 1]
data = data[data['inactivo'] != 1]
# drop all lines where values are missing
data = data.dropna(
subset=['nem', 'psu_mat', 'psu_len', 'psu_cie', 'notas_', 'uds_i_'])
# drop all columns that are not needed
if (gender):
keep_cols = [
'hombre', 'psu_mat', 'psu_len', 'psu_cie', 'nem', 'notas_',
'uds_i_', 'uds_r_', 'uds_e_'
]
else:
keep_cols = [
'highschool_type', 'psu_mat', 'psu_len', 'psu_cie', 'nem',
'notas_', 'uds_i_', 'uds_r_', 'uds_e_'
]
data = data[keep_cols]
# replace NaNs with zeros
data['uds_r_'].fillna(0)
data['uds_e_'].fillna(0)
# add new column for ranking scores
data['score'] = np.zeros(data.shape[0])
# calculate score based on grades and credits
for idx, row in data.iterrows():
grades = row.loc['notas_']
credits_taken = row.loc['uds_i_']
credits_failed = row.loc['uds_r_']
credits_dropped = row.loc['uds_e_']
score = grades * (credits_taken - credits_failed -
credits_dropped) / credits_taken
data.loc[idx, 'score'] = score
# don't need these columns anymore
data = data.drop(columns=['notas_', 'uds_i_', 'uds_r_', 'uds_e_'])
# zscore psu scores and normalize scores
data['psu_mat'] = stats.zscore(data['psu_mat'])
data['psu_len'] = stats.zscore(data['psu_len'])
data['psu_cie'] = stats.zscore(data['psu_cie'])
data['nem'] = stats.zscore(data['nem'])
data['score'] = stats.zscore(data['score'])
# rename protected column to prot_attr
data.columns = [
'prot\_attr', 'psu\_mat', 'psu\_len', 'psu\_cie', 'nem', 'score'
]
return data
def time_segment_matching(
data,
win_size=10,
):
"""
Performs time segment matching experiment
(code inspired from brainiak tutorials at
https://brainiak.org/events/ohbm2018/brainiak_sample_tutorials/10-func-align.html)
Parameters
----------
data: array of shape (n_subjects, n_components, n_timeframes)
Input shared responses
Returns
-------
cv_score: np array of shape (n_subjects)
Per-subject accuracy
"""
# Pull out shape information
n_subjs = len(data)
(n_features, n_TR) = data[0].shape # Voxel/feature by timepoint
# How many segments are there (account for edges)
n_seg = n_TR - win_size
# mysseg prediction prediction
train_data = np.zeros((n_features * win_size, n_seg))
# Concatenate the data across participants
for ppt_counter in range(n_subjs):
for window_counter in range(win_size):
train_data[window_counter * n_features:(window_counter + 1) *
n_features, :, ] += data[
ppt_counter][:,
window_counter:window_counter + n_seg]
# Iterate through the participants, leaving one out
accuracy = np.zeros(shape=n_subjs)
for ppt_counter in range(n_subjs):
# Preset
test_data = np.zeros((n_features * win_size, n_seg))
for window_counter in range(win_size):
test_data[window_counter * n_features:(window_counter + 1) *
n_features, :, ] = data[ppt_counter][:, window_counter:(
window_counter + n_seg)]
# Take this participant data away
train_ppts = stats.zscore((train_data - test_data), axis=0, ddof=1)
test_ppts = stats.zscore(test_data, axis=0, ddof=1)
# Correlate the two data sets
corr_mtx = test_ppts.T.dot(train_ppts)
# If any segments have a correlation difference less than the window size and they aren't the same segments then set the value to negative infinity
for seg_1 in range(n_seg):
for seg_2 in range(n_seg):
if abs(seg_1 - seg_2) < win_size and seg_1 != seg_2:
corr_mtx[seg_1, seg_2] = -np.inf
# Find the segement with the max value
rank = np.argmax(corr_mtx, axis=1)
# Find the number of segments that were matched for this participant
accuracy[ppt_counter] = sum(rank == range(n_seg)) / float(n_seg)
return accuracy
def demo_rsHRF(input_file, mask_file, output_dir, para, p_jobs, file_type=".nii", mode="bids", wiener=False, temporal_mask=[]):
# book-keeping w.r.t parameter values
if 'localK' not in para or para['localK'] == None:
if para['TR']<=2:
para['localK'] = 1
else:
para['localK'] = 2
# creating the output-directory if not already present
if not os.path.isdir(output_dir):
os.mkdir(output_dir)
# for four-dimensional input
if mode != 'time-series':
if mode == 'bids' or mode == 'bids w/ atlas':
name = input_file.filename.split('/')[-1].split('.')[0]
v1 = spm_dep.spm.spm_vol(input_file.filename)
else:
name = input_file.split('/')[-1].split('.')[0]
v1 = spm_dep.spm.spm_vol(input_file)
if mask_file != None:
if mode == 'bids':
mask_name = mask_file.filename.split('/')[-1].split('.')[0]
v = spm_dep.spm.spm_vol(mask_file.filename)
else:
mask_name = mask_file.split('/')[-1].split('.')[0]
v = spm_dep.spm.spm_vol(mask_file)
if file_type == ".nii" or file_type == ".nii.gz":
brain = spm_dep.spm.spm_read_vols(v)
else:
brain = v.agg_data().flatten(order='F')
if ((file_type == ".nii" or file_type == ".nii.gz") and \
v1.header.get_data_shape()[:-1] != v.header.get_data_shape()) or \
((file_type == ".gii" or file_type == ".gii.gz") and \
v1.agg_data().shape[0]!= v.agg_data().shape[0]):
raise ValueError ('Inconsistency in input-mask dimensions' + '\n\tinput_file == ' + name + file_type + '\n\tmask_file == ' + mask_name + file_type)
else:
if file_type == ".nii" or file_type == ".nii.gz" :
data = v1.get_data()
else:
data = v1.agg_data()
else:
print('No atlas provided! Generating mask file...')
if file_type == ".nii" or file_type == ".nii.gz" :
data = v1.get_data()
brain = np.nanvar(data.reshape(-1, data.shape[3]), -1, ddof=0)
else:
data = v1.agg_data()
brain = np.nanvar(data, -1, ddof=0)
print('Done')
voxel_ind = np.where(brain > 0)[0]
mask_shape = data.shape[:-1]
nobs = data.shape[-1]
data1 = np.reshape(data, (-1, nobs), order='F').T
bold_sig = stats.zscore(data1[:, voxel_ind], ddof=1)
# for time-series input
else:
name = input_file.split('/')[-1].split('.')[0]
data1 = (np.loadtxt(input_file, delimiter=","))
if data1.ndim == 1:
data1 = np.expand_dims(data1, axis=1)
nobs = data1.shape[0]
bold_sig = stats.zscore(data1, ddof=1)
if len(temporal_mask) > 0 and len(temporal_mask) != nobs:
raise ValueError ('Inconsistency in temporal_mask dimensions.\n' + 'Size of mask: ' + str(len(temporal_mask)) + '\n' + 'Size of time-series: ' + str(nobs))
bold_sig = np.nan_to_num(bold_sig)
bold_sig_deconv = processing. \
rest_filter. \
rest_IdealFilter(bold_sig, para['TR'], para['passband_deconvolve'])
bold_sig = processing. \
rest_filter. \
rest_IdealFilter(bold_sig, para['TR'], para['passband'])
data_deconv = np.zeros(bold_sig.shape)
event_number = np.zeros((1, bold_sig.shape[1]))
print('Retrieving HRF ...')
#Estimate HRF for the fourier / hanning / gamma / cannon basis functions
if not (para['estimation'] == 'sFIR' or para['estimation'] == 'FIR'):
bf = basis_functions.basis_functions.get_basis_function(bold_sig.shape, para)
beta_hrf, event_bold = utils.hrf_estimation.compute_hrf(bold_sig, para, temporal_mask, p_jobs, bf=bf)
hrfa = np.dot(bf, beta_hrf[np.arange(0, bf.shape[1]), :])
#Estimate HRF for FIR and sFIR
else:
beta_hrf, event_bold = utils.hrf_estimation.compute_hrf(bold_sig, para, temporal_mask, p_jobs)
hrfa = beta_hrf[:-1,:]
nvar = hrfa.shape[1]
PARA = np.zeros((3, nvar))
for voxel_id in range(nvar):
hrf1 = hrfa[:, voxel_id]
PARA[:, voxel_id] = \
parameters.wgr_get_parameters(hrf1, para['TR'] / para['T'])
print('Done')
print('Deconvolving HRF ...')
if para['T'] > 1:
hrfa_TR = signal.resample_poly(hrfa, 1, para['T'])
else:
hrfa_TR = hrfa
for voxel_id in range(nvar):
hrf = hrfa_TR[:, voxel_id]
if not wiener:
H = np.fft.fft(
np.append(hrf,
np.zeros((nobs - max(hrf.shape), 1))), axis=0)
M = np.fft.fft(bold_sig_deconv[:, voxel_id])
data_deconv[:, voxel_id] = \
np.fft.ifft(H.conj() * M / (H * H.conj() + .1*np.mean((H * H.conj()))))
else:
data_deconv[:, voxel_id] = iterative_wiener_deconv.rsHRF_iterative_wiener_deconv(bold_sig_deconv[:, voxel_id], hrf)
event_number[:, voxel_id] = np.amax(event_bold[voxel_id].shape)
# setting the output-path
if mode == 'bids' or mode == 'bids w/ atlas':
try:
sub_save_dir = os.path.join(
output_dir, 'sub-' + input_file.subject,
'session-' + input_file.session,
input_file.modality
)
except AttributeError as e:
sub_save_dir = os.path.join(
output_dir, 'sub-' + input_file.subject,
input_file.modality
)
else:
sub_save_dir = output_dir
if not os.path.isdir(sub_save_dir):
os.makedirs(sub_save_dir, exist_ok=True)
dic = {'para': para, 'hrfa': hrfa, 'event_bold': event_bold, 'PARA': PARA}
ext = '_hrf.mat'
if mode == "time-series":
dic["event_number"] = event_number
dic["data_deconv"] = data_deconv
ext = '_hrf_deconv.mat'
sio.savemat(os.path.join(sub_save_dir, name + ext), dic)
HRF_para_str = ['Height', 'Time2peak', 'FWHM']
if mode != "time-series":
mask_data = np.zeros(mask_shape).flatten(order='F')
for i in range(3):
fname = os.path.join(sub_save_dir,
name + '_' + HRF_para_str[i])
mask_data[voxel_ind] = PARA[i, :]
mask_data = mask_data.reshape(mask_shape, order='F')
spm_dep.spm.spm_write_vol(v1, mask_data, fname, file_type)
mask_data = mask_data.flatten(order='F')
fname = os.path.join(sub_save_dir, name + '_event_number.nii')
mask_data[voxel_ind] = event_number
mask_data = mask_data.reshape(mask_shape, order='F')
spm_dep.spm.spm_write_vol(v1, mask_data, fname, file_type)
mask_data = np.zeros(data.shape)
dat3 = np.zeros(data.shape[:-1]).flatten(order='F')
for i in range(nobs):
fname = os.path.join(sub_save_dir, name + '_deconv')
dat3[voxel_ind] = data_deconv[i, :]
dat3 = dat3.reshape(data.shape[:-1], order='F')
if file_type == ".nii" or file_type == ".nii.gz" :
mask_data[:, :, :, i] = dat3
else:
mask_data[:, i] = dat3
dat3 = dat3.flatten(order='F')
spm_dep.spm.spm_write_vol(v1, mask_data, fname, file_type)
pos = 0
while pos < hrfa_TR.shape[1]:
if np.any(hrfa_TR[:,pos]):
break
pos += 1
event_plot = lil_matrix((1, nobs))
if event_bold.size:
event_plot[:, event_bold[pos]] = 1
else:
print("No Events Detected!")
return 0
event_plot = np.ravel(event_plot.toarray())
plt.figure()
plt.plot(para['TR'] * np.arange(1, np.amax(hrfa_TR[:, pos].shape) + 1),
hrfa_TR[:, pos], linewidth=1)
plt.xlabel('time (s)')
plt.savefig(os.path.join(sub_save_dir, name + '_plot_1.png'))
plt.figure()
plt.plot(para['TR'] * np.arange(1, nobs + 1),
np.nan_to_num(stats.zscore(bold_sig[:, pos], ddof=1)),
linewidth=1)
plt.plot(para['TR'] * np.arange(1, nobs + 1),
np.nan_to_num(stats.zscore(data_deconv[:, pos], ddof=1)),
color='r', linewidth=1)
markerline, stemlines, baseline = \
plt.stem(para['TR'] * np.arange(1, nobs + 1), event_plot)
plt.setp(baseline, 'color', 'k', 'markersize', 1)
plt.setp(stemlines, 'color', 'k')
plt.setp(markerline, 'color', 'k', 'markersize', 3, 'marker', 'd')
plt.legend(['BOLD', 'deconvolved', 'events'])
plt.xlabel('time (s)')
plt.savefig(os.path.join(sub_save_dir, name + '_plot_2.png'))
print('Done')
return 0
"""
Created on Mon Nov 27 19:11:38 2017
@author: Jacob
"""
import numpy as np
from matplotlib.pyplot import (figure, imshow, bar, title, xticks, yticks, cm,
subplot, show)
from scipy.stats.kde import gaussian_kde
from toolbox_02450 import gausKernelDensity
from sklearn.neighbors import NearestNeighbors
from scipy import stats
import dataSetup
X = dataSetup.numbersData.values
X = stats.zscore(X)
N,M = X.shape
"""
# OUTLIER DETECTION
# Compute kernel density estimate
kde = gaussian_kde(X.ravel(), 'silverman')
scoresKDE = kde.evaluate(X.ravel())
idxKDE = scoresKDE.argsort()
scoresKDE.sort()
print('The index of the lowest density object: {0}'.format(idxKDE[0]))
# Plot kernel density estimate
figure()
def main():
"""
load data
"""
train_set = pd.read_csv('../data/train.csv')
test_set = pd.read_csv('../data/test.csv')
#Without outlier remover, with basic nanRemover 0.12416413124809748
"""
Remove Outliers
"""
outliers = [197, 523, 691, 854, 1182, 1298]
print(outliers)
z = np.abs(zscore(train_set[get_numeric_columns(train_set)]))
row, col = np.where(z > 4)
df = pd.DataFrame({"row": row, "col": col})
rows_count = df.groupby(['row']).count()
outliers = rows_count[rows_count.col > 2].index
print(outliers)
train_set.drop(outliers, inplace=True)
"""
fix salePrice skewness
"""
train_set["SalePrice"] = np.log1p(train_set["SalePrice"])
y_train_values = train_set["SalePrice"].values
"""
prepare combined data.
"""
train_set_id = train_set['Id']
test_set_id = test_set['Id']
train_set_rows = train_set.shape[0]
test_set_rows = test_set.shape[0]
train_set.drop('Id', axis=1, inplace=True)
test_set.drop('Id', axis=1, inplace=True)
train_set.drop('SalePrice', axis=1, inplace=True)
combined_data = pd.concat((train_set, test_set))
"""
create data transform pipeline
"""
transform_pipeline = Pipeline(steps=[
('OutlierRemover', OutlierRemover()),
('NaNImputer', NaNImputer()),
('NaNRemover', NaNRemover()),
('AdditionalFeatureGenerator', AdditionalFeatureGenerator()),
('TypeTransformer', TypeTransformer()),
('ErrorImputer', ErrorImputer()),
('SkewFixer', SkewFixer()),
('Scaler', Scaler()),
('FeatureDropper', FeatureDropper()),
('Dummyfier', Dummyfier()),
])
transformed_data = transform_pipeline.transform(combined_data)
train_data = transformed_data[:train_set_rows]
predict_data = transformed_data[train_set_rows:]
transformed_data.to_csv('transformed_Data.csv', index=False)
"""
try various regressors
"""
rf_param = {
# 'bootstrap': [True],
'max_depth': [3, 4, 5],
'min_samples_leaf': [3, 4, 5],
'n_estimators': [5, 7, 10]
}
ls_param = {'alpha': [0.0003, 0.0004, 0.0005,
0.0006, 0.0007, 0.0008],
'max_iter': [10000], "normalize": [False]}
elnet_param = {'alpha': [0.0003, 0.0004, 0.0005],
'l1_ratio': [0.9, 0.95, 0.99, 1],
'max_iter': [10000]}
ridge_param = {'alpha': [10, 10.1, 10.2, 10.3, 10.4, 10.5]}
svr_param = {'gamma': [1e-08, 1e-09],
'C': [100000, 110000],
'epsilon': [1, 0.1, 0.01]
}
rf = get_best_estimator(train_data, y_train_values, estimator=RandomForestRegressor(),
params=rf_param, n_jobs=4)
elnet = get_best_estimator(train_data, y_train_values, estimator=ElasticNet(),
params=elnet_param, n_jobs=4)
lso = get_best_estimator(train_data, y_train_values, estimator=Lasso(),
params=ls_param, n_jobs=4)
rdg = get_best_estimator(train_data, y_train_values, estimator=Ridge(),
params=ridge_param, n_jobs=4)
svr = get_best_estimator(train_data, y_train_values, estimator=SVR(),
params=svr_param, n_jobs=4)
def cv_rmse(model):
kfolds = KFold(n_splits=5, shuffle=True, random_state=42)
rmse = np.sqrt(-cross_val_score(model, train_data, y_train_values,
scoring="neg_mean_squared_error",
cv=kfolds))
return (rmse)
"""
print("Randomforest model rmse : ", cv_rmse(rf).mean())
print("elastic model rmse : ", cv_rmse(elnet).mean())
print("lasso model rmse : ", cv_rmse(lso).mean())
print("ridge model rmse : ", cv_rmse(rdg).mean())
print("svr model rmse : ", cv_rmse(svr).mean())
"""
model = StackingRegressor(
regressors=[rf, elnet, lso, rdg, svr],
meta_regressor=Lasso(alpha=0.0005)
# meta_regressor=SVR(kernel='rbf')
)
# Fit the model on our data
model.fit(train_data, y_train_values)
#print("StackingRegressor model rmse : ", cv_rmse(model).mean())
# y_pred = model.predict(train_data)
# print(sqrt(mean_squared_error(y_train_values, y_pred)))
# Predict test set
ensembled = np.expm1(model.predict(predict_data))
# sns.scatterplot(np.expm1(rf.predict(train_data),np.expm1(y_train_values)))
# plt.show()
# ensembled = np.expm1(rf.predict(predict_data))
"""
export submission data
"""
submission = pd.DataFrame({
"Id": test_set_id,
"SalePrice": ensembled
})
submission.to_csv('submission_jiwon.csv', index=False)
import pandas as pd
import seaborn as sns
import numpy as np
from scipy import stats
#%%
accidents = pd.read_csv(
r'C:\Users\rcf004\Documents\Python Scripts\US_Accidents_Dec19.csv')
accidents.dropna(
subset=['Start_Lng', 'Start_Lat', 'Severity', 'Visibility(mi)'],
inplace=True)
#%%
no_out = accidents[
np.abs(stats.zscore(accidents['Visibility(mi)'].astype(int))) < 3]
#%%
df1 = no_out[no_out['Start_Time'].astype(str).str.contains('2016-')]
df2 = no_out[no_out['Start_Time'].astype(str).str.contains('2017-')]
df3 = no_out[no_out['Start_Time'].astype(str).str.contains('2018-')]
df4 = no_out[no_out['Start_Time'].astype(str).str.contains('2019-')]
#%%
a = 0.5
sz = (1, 10)
lw = 0
kwargs = {'marker': "."}
cmap = sns.cubehelix_palette(start=1.1,
elif seq[nucpos:nucpos + 3] == "---":
nucpos += 3
continue
else:
dnlist[codon] += 1
nucpos += 3
except KeyError:
nucpos += 3
continue
nucpos += 3
calclist=[]
for i in range(0,len(dslist)):
calclist.append(dnlist[i] - dslist[i])
zmin =stats.zscore(calclist)
plt.figure()
for i, z in enumerate(zmin):
if z > 3:
plt.scatter(i,z, color ='orange')
else:
plt.scatter(i, z, color='blue')
plt.ylabel("zscore")
plt.xlabel("codon num")
plt.savefig("scatter.png")
plt.close()
date_bucket = np.hstack((date_bucket, np.asarray(tsd['timestamp'][i])))
date_bucket = np.unique(date_bucket)
date_bucket = date_bucket[date_bucket > datetime.date(2020, 1, 1)]
interpolated_time_series = []
for i in range(0, tsd.__len__()):
series = pd.DataFrame(tsd['data'][i], tsd['timestamp'][i])
series = series[~series.index.duplicated(keep='first')]
series = series[series.index > datetime.date(2020, 1, 1)]
series = series.reindex(
date_bucket,
fill_value=0).sort_index().mask(series == 0).interpolate()
if series.isna().sum()[0] > series.__len__() / 2 or series.var(
)[0] < variance_threshold:
continue
elif series.isna().sum()[0] > 0:
series = series.fillna(0)
series['zscore'] = st.zscore(series)
interpolated_time_series.append({
"path": tsd['path'][i],
"node": tsd['node'][i],
"slot": tsd['slot'][i],
"port": tsd['port'][i],
"pm": tsd['pm'][i],
"raw_data": np.asarray(series[0]),
"z-score": np.asarray(series['zscore']),
"timestamp": np.asarray(series.index)
})
print(i)
f = open("vodafone_data_oct30_not_pm_filtered_interpolated.pkl", "wb")
pickle.dump(pd.DataFrame(interpolated_time_series), f)
f.close()
## comparison to shortest path analysis
ALL = set(INH_PLC_list)
HIT = PET_TG & SPL_TG
pval = stats.hypergeom.sf(len(HIT), len(ALL), len(PET_TG), len(SPL_TG))
print("Petrinet & shortest path", pval, [entz_dic[entz] for entz in HIT])
## get gold standard from CTD database
gene_df = pd.read_table('./CTD/CTD_genes_related_to_atrophy.txt', sep='\t')
## get z-score of inferenece score in CTD
infScore_ds = gene_df['InferenceScore']
scoreList = list(infScore_ds.dropna())
scoreDic = {}
zScoreList = stats.zscore(scoreList)
for ii in range(len(scoreList)):
scoreDic[scoreList[ii]] = zScoreList[ii]
GS_dic = {}
for index, row in gene_df.iterrows():
if pd.isnull(row['DirectEvidence']) == False:
if str(row['GeneID']) == '3479': # use IGF1R instead of IGF1
GS_dic['3480'] = 100
else:
GS_dic[str(row['GeneID'])] = 100
for index, row in gene_df.iterrows():
if pd.isnull(row['DirectEvidence']) == True:
if GS_dic.has_key(str(row['GeneID'])):
continue
import numpy as np
import pandas as pd
from copy import deepcopy
from scipy import stats
from mpl_toolkits.mplot3d import Axes3D
get_ipython().magic('matplotlib inline')
#%matplotlib inline
from matplotlib import pyplot as plt
plt.rcParams['figure.figsize'] = (16, 9)
plt.style.use('ggplot')
# Importing the dataset
data = pd.read_csv('C:\\Users\\himverma\\AnacondaProjects\\KMeans\\xclaraOriginal.csv')
print("Input Data and Shape")
print(data.shape)
data.head()
# Getting the values and plotting it
f1 = data['V1'].values
#f2 = data['V2'].values
#X = np.array(list(zip(f1, f2)))
X = np.array(f1)
stats.zscore(X)
fit = stats.norm.pdf(X, np.mean(X), np.std(X))
plt.plot(X,fit,'-o')
plt.hist(X, 30, normed=True)
plt.show()
print(X.mean())
# "standard representation" for the course, is the number of classes, C:
C = len(classNames)
# Add offset attribute
X = np.concatenate((np.ones((X.shape[0],1)),X),1)
#attributeNames = [u'Offset']+attributeNames
M = M+1
#attributeNames = ('Offset', 'Alcohol', 'Malic acid', 'Ash', 'Alcalinity of ash', 'Magnesium', 'Total phenols', 'Flavanoids', 'Nonflavanoid phenols', 'Proanthocyanins', 'Color intensity', 'Hue', 'OD280/OD315 of diluted wines', 'Proline')
#attributeNames = ('Offset', 'Ash', 'Magnesium', 'Color intensity', 'Hue' 'Proline')
attributeNames = ('Offset', 'Malic acid', 'Ash', 'Alcalinity of ash', 'Magnesium', 'Total phenols', 'Flavanoids', 'Nonflavanoid phenols', 'Proanthocyanins', 'Color intensity', 'Hue', 'OD280/OD315 of diluted wines', 'Proline')
#Standalize the data to zero mean and standard deviation of 1
X_standarized = zscore(X, ddof=1) #Do the standalization
############################################
#Use exercise 8.1.1
###########################################
## Crossvalidation
# Create crossvalidation partition for evaluation
K = 10
CV = model_selection.KFold(K, shuffle=True)
#CV = model_selection.KFold(K, shuffle=False)
# Values of lambda
def create_model(self, train_X, train_y, val_X, val_y):
"""
Args:
train_X (pandas dataframe)
train_y (pandas dataframe)
Returns:
ExtraTreesRegressor
"""
# モデル作成
train_X = train_X[self.feature_columns]
train_X = stats.zscore(train_X)
train_X = train_X.reshape(
(train_X.shape[0], 1, train_X.shape[1]))
val_X = val_X[self.feature_columns]
val_X = stats.zscore(val_X)
val_X = val_X.reshape((val_X.shape[0], 1, val_X.shape[1]))
model = Sequential()
model.add(LSTM(512, input_shape=(train_X.shape[1], train_X.shape[2])))
model.add(BatchNormalization())
model.add(Dropout(.2))
model.add(Dense(256))
model.add(PReLU())
model.add(BatchNormalization())
model.add(Dropout(.1))
model.add(Dense(256))
model.add(PReLU())
model.add(BatchNormalization())
model.add(Dropout(.1))
model.add(Dense(128))
model.add(PReLU())
model.add(BatchNormalization())
model.add(Dropout(.05))
model.add(Dense(64))
model.add(PReLU())
model.add(BatchNormalization())
model.add(Dropout(.05))
model.add(Dense(32))
model.add(PReLU())
model.add(BatchNormalization())
model.add(Dropout(.05))
model.add(Dense(16))
model.add(PReLU())
model.add(BatchNormalization())
model.add(Dropout(.05))
model.add(Dense(1))
# ネットワークのコンパイル
model.compile(loss='mse', optimizer=optimizers.Adam(0.001),
metrics=['mse'])
callbacks = [
EarlyStopping(monitor='val_loss', patience=10, verbose=0),
ReduceLROnPlateau(monitor='val_loss', factor=0.1, patience=7,
verbose=1, epsilon=1e-4, mode='min')
]
model.fit(x=train_X, y=train_y, epochs=80,
validation_data=(val_X, val_y), callbacks=[callbacks])
def impute_markers(model_path,
data_path,
*,
save_path=None,
start_frame=None,
n_frames=None,
stride=1,
markers_to_fix=None,
error_diff_thresh=.25,
model=None):
"""Imputes the position of missing markers.
:param model_path: Path to model to use for prediction.
:param data_path: Path to marker and bad_frames data. Can be hdf5 or
mat -v7.3.
:param save_path: Path to .mat file where predictions will be saved.
:param start_frame: Frame at which to begin imputation.
:param n_frames: Number of frames to impute.
:param stride: stride length between frames for faster imputation.
:param markers_to_fix: Markers for which to override suspicious MoCap
measurements
:param error_diff_thresh: Z-scored difference threshold marking suspicious
frames
:param model: Model to be used in prediction. Overrides model_path.
:return: preds
"""
# Check data extensions
filename, file_extension = os.path.splitext(data_path)
accepted_extensions = {'.h5', '.hdf5', '.mat'}
if file_extension not in accepted_extensions:
raise ValueError('Improper extension: hdf5 or \
mat -v7.3 file required.')
# Load data
print('Loading data')
f = h5py.File(data_path, 'r')
if file_extension in {'.h5', '.hdf5'}:
markers = np.array(f['markers'][:]).T
marker_means = np.array(f['marker_means'][:]).T
marker_stds = np.array(f['marker_stds'][:]).T
bad_frames = np.array(f['bad_frames'][:]).T
else:
# Get the markers data from the struct
dset = 'markers_aligned_preproc'
marker_names = list(f[dset].keys())
n_frames_tot = f[dset][marker_names[0]][:].T.shape[0]
n_dims = f[dset][marker_names[0]][:].T.shape[1]
markers = np.zeros((n_frames_tot, len(marker_names) * n_dims))
for i in range(len(marker_names)):
marker = f[dset][marker_names[i]][:].T
for j in range(n_dims):
markers[:, i * n_dims + j] = marker[:, j]
# Z-score the marker data
marker_means = np.mean(markers, axis=0)
marker_means = marker_means[None, ...]
marker_stds = np.std(markers, axis=0)
marker_stds = marker_stds[None, ...]
print(marker_means)
print(marker_stds)
markers = stats.zscore(markers)
# Get the bad_frames data from the cell
dset = 'bad_frames_agg'
n_markers = f[dset][:].shape[0]
bad_frames = np.zeros((markers.shape[0], n_markers))
for i in range(n_markers):
reference = f[dset][i][0]
bad_frames[np.squeeze(f[reference][:]).astype('int32') - 1, i] = 1
# Set number of frames to impute
if n_frames is None:
n_frames = markers.shape[0]
if start_frame is None:
start_frame = 0
print('Predicting %d frames starting at frame %d.' %
(n_frames, start_frame))
# Exceptions
if n_frames > markers.shape[0]:
raise ValueError("Improper n_frames to predict: likely asked to " +
"predict a greater number of frames than were " +
"available.")
if (n_frames + start_frame) > markers.shape[0]:
raise ValueError('start_frame + n_frames exceeds matrix dimensions.')
if n_frames < 0:
raise ValueError("Improper n_frames to predict: likely too few input" +
" frames.")
if n_frames == 0:
raise ValueError("Improper n_frames to predict: likely asked to " +
"predict zero frames.")
markers = markers[start_frame:(start_frame + n_frames):stride, :]
bad_frames = bad_frames[start_frame:(start_frame + n_frames):stride, :]
# Load model
if model is None:
print('Loading model')
model = load_model(model_path)
# Check how many outputs the model has, and how many members if returning
# member data.
n_outputs = len(model.output_shape)
if n_outputs == 2:
return_member_data = True
else:
return_member_data = False
member_predsF = [None]
member_predsR = [None]
# Set Markers to fix
if markers_to_fix is None:
markers_to_fix = np.zeros((markers.shape[1])) > 1
# TODO(Skeleton): Automate this by including the skeleton.
markers_to_fix[30:36] = True
markers_to_fix[42:] = True
# If the model can return the member predictions, do so.
if return_member_data:
# Forward predict
print('Imputing markers: forward pass')
predsF, bad_framesF, member_predsF = \
predict_markers(model, markers, bad_frames,
markers_to_fix=markers_to_fix,
error_diff_thresh=error_diff_thresh,
return_member_data=return_member_data)
# Reverse Predict
print('Imputing markers: reverse pass')
predsR, bad_framesR, member_predsR = \
predict_markers(model, markers[::-1, :], bad_frames[::-1, :],
markers_to_fix=markers_to_fix,
error_diff_thresh=error_diff_thresh,
return_member_data=return_member_data)
else:
# Forward predict
print('Imputing markers: forward pass')
predsF, bad_framesF = \
predict_markers(model, markers, bad_frames,
markers_to_fix=markers_to_fix,
error_diff_thresh=error_diff_thresh,
return_member_data=return_member_data)
# Reverse Predict
print('Imputing markers: reverse pass')
predsR, bad_framesR = \
predict_markers(model, markers[::-1, :], bad_frames[::-1, :],
markers_to_fix=markers_to_fix,
error_diff_thresh=error_diff_thresh,
return_member_data=return_member_data)
# Convert to real world coordinates
markers_world = np.zeros((markers.shape))
predsF_world = np.zeros((predsF.shape))
predsR_world = np.zeros((predsR.shape))
for i in range(markers_world.shape[1]):
markers_world[:, i] = \
markers[:, i]*marker_stds[0, i] + marker_means[0, i]
predsF_world[:, i] = \
predsF[:, i]*marker_stds[0, i] + marker_means[0, i]
predsR_world[:, i] = \
predsR[:, i]*marker_stds[0, i] + marker_means[0, i]
predsR_world = predsR_world[::-1, :]
bad_framesR = bad_framesR[::-1, :]
# This is not necessarily all of the error frames from
# multiple_predict_recording_with_replacement, but if they overlap,
# we would just take the weighted average.
for i in range(bad_frames.shape[1]):
bad_frames[:, i] = np.any(bad_framesF[:, (i * 3):(i * 3) + 3]
& bad_framesR[:, (i * 3):(i * 3) + 3],
axis=1)
# Compute the weighted average of the forward and reverse predictions using
# a logistic function
print('Computing weighted average')
preds_world = np.zeros(predsF_world.shape)
for i in range(bad_frames.shape[1] * 3):
is_bad = bad_frames[:, np.floor(i / 3).astype('int32')]
CC = measure.label(is_bad, background=0)
num_CC = len(np.unique(CC)) - 1
preds_world[:, i] = predsF_world[:, i]
for j in range(num_CC):
length_CC = np.sum(CC == (j + 1))
x_0 = np.round(length_CC / 2)
k = 1
weightR = sigmoid(np.arange(length_CC), x_0, k)
weightF = 1 - weightR
preds_world[CC == (j+1), i] = \
(predsF_world[CC == (j+1), i]*weightF) +\
(predsR_world[CC == (j+1), i]*weightR)
# Save predictions to a matlab file.
if save_path is not None:
s = 'Saving to %s' % (save_path)
print(s)
savemat(
save_path, {
'preds': preds_world,
'markers': markers_world,
'badFrames': bad_frames,
'member_predsF': member_predsF,
'member_predsR': member_predsR
})
return preds_world
def prepare_confounders(
sm,
confounders=tuple(),
hcp_confounders=False,
hcp_confounder_software_version=True,
squared_confounders=False,
impute0=True,
#headmotion_features=('movement_AbsoluteRMS_mean',
# 'movement_RelativeRMS_mean')
):
"""Prepare the confounder matrix.
Parameters
----------
sm : pd.DataFrame (n_samples, n_features)
behavioral data matrix
confounders : tuple of str
column-names in ``sm`` to be used as confounders. If some are
not found a warning is issued and the code will continue without the
missing ones.
hcp_confounders : bool
if ``True`` 'Weight', 'Height', 'BPSystolic', 'BPDiastolic', 'HbA1C'
as well as the cubic roots of 'FS_BrainSeg_Vol', 'FS_IntraCranial_Vol'
are included as confounders
hcp_confounder_software_version : bool
if ``True`` and ``hcp_confounders`` is also ``True``, then the feature
'fMRI_3T_ReconVrs' (encoded as a dummy variable) is used as confounder
squared_confounders : bool
if ``True`` the squares of all confounders (except software version, if
used) are used as additional confounders
impute0 : bool
if True, missing confound values are imputed with 0 (after an
inverse normal transformation)
Returns
-------
confounders : np.ndarray (n_samples, n_features)
confounder data matrix, if impute0 is ``False`` it can have ``NaN``s
Raises
------
ValueError
if confounders couldn't be found in ``sm``
"""
_confounders = [f for f in confounders if f in sm]
if len(_confounders) != len(confounders):
missing_confounders = [f for f in confounders if f not in sm]
raise ValueError('Confounders not found: '
'{}'.format(missing_confounders))
confounders_matrix = sm[_confounders].values
if hcp_confounders:
sm_confounders = sm[[
'Weight',
'Height',
'BPSystolic',
'BPDiastolic',
'HbA1C',
]].values
fs_confounders = sm[['FS_BrainSeg_Vol',
'FS_IntraCranial_Vol']].values**(1. / 3)
confounders_matrix = np.hstack(
[confounders_matrix, sm_confounders, fs_confounders])
if squared_confounders:
confounders_matrix = np.hstack(
[confounders_matrix, confounders_matrix**2])
if hcp_confounders and hcp_confounder_software_version:
# software reconstruction version
reconvrs = sm['fMRI_3T_ReconVrs'].values
used_reconvrss = np.unique(reconvrs)
print('used fMRI 3T reconstruction software versions are:',
used_reconvrss)
assert set(used_reconvrss.tolist()) == {'r177', 'r177 r227', 'r227'}
# dummy-coding: r177 -> 0, r227 -> 1, "r177 r227" -> 1
reconvrs = np.where(reconvrs == 'r177', 0, 1).reshape(-1, 1)
confounders_matrix = np.hstack([confounders_matrix, reconvrs])
if confounders_matrix.shape[1] > 0:
# inverse normal transform (this also results in mean 0)
confounders_matrix = \
rank_based_inverse_normal_trafo(confounders_matrix)
if impute0:
# impute 0 for missing values
print('{:.2f}% of values in confounders missing, imputing 0 for '
'these'.format(100 *
(1 - np.isfinite(confounders_matrix).mean())))
confounders_matrix[~np.isfinite(confounders_matrix)] = 0
else:
print('{:.2f}% of values in confounders missing'.format(
100 * (1 - np.isfinite(confounders_matrix).mean())))
# normalise
confounders_matrix = zscore(confounders_matrix, nan_policy='omit')
return confounders_matrix
def train_price_model(self, data: pd.DataFrame):
df = data
df = df[((np.abs(stats.zscore(df.price)) < 2.8) &
(np.abs(stats.zscore(df.term)) < 2.8) &
(np.abs(stats.zscore(df.full_sq)) < 2.8))]
# !!!!!!!! ADD 'was_opened'
# Fix year: only 2019
df = df[(df.yyyy_announce.isin([19, 20]))]
df = df[[
'price', 'to_center', 'full_sq', 'kitchen_sq', 'life_sq', 'rooms',
'is_apartment', 'renovation', 'has_elevator', 'time_to_metro',
'floor_first', 'floor_last', 'is_rented', 'rent_quarter',
'rent_year', 'mm_announce', 'yyyy_announce', 'clusters'
]]
# Save leaved columns to variable
columns = list(df.columns)
# Log transformation
# Log Transformation
df["full_sq"] = np.log1p(df["full_sq"])
df["life_sq"] = np.log1p(df["life_sq"])
df["kitchen_sq"] = np.log1p(df["kitchen_sq"])
df["price"] = np.log1p(df["price"])
df["to_center"] = np.log1p(df["to_center"])
# Create features - predictors
X = df.drop(['price'], axis=1)
# Target feature
y = df[['price']].values.ravel()
# Split for train and test
X_train, X_test, y_train, y_test = train_test_split(X,
y,
test_size=0.2,
random_state=1)
# Define Gradient Boosting Machine model
lgbm_model = LGBMRegressor(objective='regression',
learning_rate=0.07,
n_estimators=1250,
max_depth=10,
min_child_samples=1,
verbose=0)
# RF = RandomForestRegressor(n_estimators=300, verbose=1, n_jobs=-1)
# Train GBR on train dataset
lgbm_model.fit(X_train, y_train)
lgbm_preds = lgbm_model.predict(X_test)
print('The R2_score of the Gradient boost is',
r2_score(y_test, lgbm_preds),
flush=True)
print('RMSE is: \n',
mean_squared_error(y_test, lgbm_preds),
flush=True)
# Train GBR on full dataset
lgbm_model.fit(X, y)
return lgbm_model, columns
# # Preliminary data analysis
# ## missing value
# In[6]:
null_data = seeds[seeds.isnull().any(axis=1)]
display(null_data)
# No missing value here
#
# ## outliers
# In[7]:
"""" clean outliers here by compute Z-score of each value in the column, if abs of Z score bigger than 3 than delete this row"""
data = seeds[(np.abs(stats.zscore(seeds)) < 3).all(axis=1)]
data.info()
# only deleted 2 rows so delete the outlier rows directly is a easy way to process ouelier.
#
# ## Correlation
# In[8]:
plt.figure(figsize=(20, 7))
sns.heatmap(data.corr(), cmap='BrBG', annot=True)
plt.title('Variables Correlation', fontsize=18)
plt.show()
# As can be seen asym is most unrelated to other attributes.
#
plt.savefig('KDE_MINIMUM_PAYMENTS.png')
data['MINIMUM_PAYMENTS'] = data['MINIMUM_PAYMENTS'].fillna(
data['MINIMUM_PAYMENTS'].median())
pd.isnull(data).sum()
#%%
'''
Outliers Treatment
'''
# calculate z-score
from scipy import stats
# drop string feature and features with meaningful range
data1 = data.drop(columns=['CUST_ID', 'TENURE'])
z_score = pd.DataFrame(np.abs(stats.zscore(data1)), columns=data1.columns)
# Find out features with more than 2% outliers (absolute z-score >3)
z_score3 = []
over3_index = []
for i in z_score.columns:
indexs = z_score.index[z_score[i] > 3].tolist()
ans = i, "{:.3f}".format(len(indexs) / len(z_score)), indexs
z_score3.append(ans)
if len(indexs) / len(z_score) > 0.02:
over3_index.append(i)
# remove 'BALANCE' and 'CASH_ADVANCE' since thay are regarded as high
# discriminative features
del over3_index[0]
del over3_index[1]
int(team["L"]),
"is_world_series_winner":
team["WSWin"],
"attendance":
float(team["attendance"]),
"avg_salary":
float(round(average(players_in_team, "salary"), 2)),
"batting_avg":
float(round(average(map(batting_average, players_in_team)), 3)),
"era":
float(round(average(pitchers_in_team, "ERA"), 3))
})
grouped_by_year = group_by(flat_franchise_year, "year")
for year, records in grouped_by_year.iteritems():
z_scores_salary = np.round(stats.zscore(pluck("avg_salary", records)), 2)
z_scores_wins = np.round(stats.zscore(pluck("wins", records)), 2)
z_scores_batting_avg = np.round(
stats.zscore(pluck("batting_avg", records)), 2)
z_scores_losses = np.round(stats.zscore(pluck("losses", records)), 2)
z_scores_attendance = np.round(stats.zscore(pluck("attendance", records)),
2)
z_scores_era = np.round(stats.zscore(pluck("era", records)), 2)
for i, record in enumerate(records):
record["z_avg_salary"] = z_scores_salary[i]
record["z_wins"] = z_scores_wins[i]
record["z_batting_avg"] = z_scores_batting_avg[i]
record["z_losses"] = z_scores_losses[i]
record["z_attendance"] = z_scores_attendance[i]
record["z_era"] = z_scores_era[i]
def removeOutliersZScore(data):
outlierColumns = data[[IncomeColumn]].copy()
z = numpy.abs(stats.zscore(outlierColumns))
newData = data[(z < 12).all(axis=1)]
return newData
