def boxcoxTransformation(data=None):
ind = sp.where(data.flatten() <= 0.0)[0]
if ind.shape[0] > 0:
return data
if data.ndim == 1:
[data, boxcox_lambda] = boxcox(data)
elif data.shape[1] == 1:
[tmp_y, boxcox_lambda] = boxcox(data[:, 0])
data = sp.zeros((tmp_y.shape[0], 1))
data[:, 0] = tmp_y
elif data.shape[0] == 1:
[tmp_y, boxcox_lambda] = boxcox(data[0, :])
data = sp.zeros((1, tmp_y.shape[0]))
data[0, :] = tmp_y
return data
def boxcoxTransformation(data=None):
ind = sp.where(data.flatten()<=0.0)[0]
if ind.shape[0]>0:
return data
if data.ndim==1:
[data, boxcox_lambda] = boxcox(data)
elif data.shape[1]==1:
[tmp_y, boxcox_lambda] = boxcox(data[:,0])
data = sp.zeros((tmp_y.shape[0],1))
data[:,0] = tmp_y
elif data.shape[0]==1:
[tmp_y, boxcox_lambda] = boxcox(data[0,:])
data = sp.zeros((1,tmp_y.shape[0]))
data[0,:] = tmp_y
return data
def treatdata(self, data): """ Returns the box-cox transformed data if the flag has been set. otherwise it will return the data unchanged. """ #print 'length', len(data) if self.use_bc: t_data, llmb = boxcox(data) return t_data else: return data
def doBoxCoxTransformation(self):
for i in range(self.n_t):
y = self.Y[:, i]
idx = SP.isfinite(y)
self.Y[idx, i] = MS.boxcox(y[idx])[0]
def doBoxCoxTransformation(self):
for i in range(self.n_t):
y = self.Y[:,i]
idx = SP.isfinite(y)
self.Y[idx,i] = MS.boxcox(y[idx])[0]
def plotHistogram(y=None,phenotype_name=None,transform='sqrt',outdir=None):
ind = sp.where(~sp.isnan(y))
y = y[ind]
pl.figure(figsize=(12,6))
pl.subplot(121)
[test_statistic, p_value] = shapiro(y)
pl.hist(y,bins=30,color=color_t[1],label="Original")
pl.title(phenotype_name.replace("_"," ") + ", Shapiro: %.2e"%(p_value))
leg = pl.legend(fancybox=True)
leg.get_frame().set_alpha(0.2)
#leg.get_frame().set_edgecolor("none")
remove_border()
pl.subplot(122)
if transform=="all":
p_vals = []
transformations = sp.array(['boxcox','sqrt','log','log10'])
for t in transformations:
zeros = sp.where(y==0)[0]
if zeros.shape[0]==0:
if t=='sqrt':
tmpy = sp.sqrt(y)
elif t=="boxcox":
[tmpy,b_lambda] = boxcox(y)
elif t=="log":
tmpy = sp.log(y)
elif t=="log10":
tmpy = sp.log10(y)
[test_statistic, pv] = shapiro(tmpy)
else:
pv = 0.0
p_vals.append(pv)
p_vals = sp.array(p_vals)
ind = sp.argmax(p_vals)
transform = transformations[ind]
ind = sp.where(y==0)[0]
if ind.shape[0]>0:
print "IMPORTANT: y contains 0 -> transformation changed to SQRT"
transform = "sqrt"
if transform=='sqrt':
y = sp.sqrt(y)
elif transform=="boxcox":
[y,b_lambda] = boxcox(y)
elif transform=="log":
y = sp.log(y)
elif transform=="log10":
y = sp.log10(y)
[test_statistic, p_value_t] = shapiro(y)
pl.hist(y,bins=30,color=color_t[4],label=transform)
pl.title(phenotype_name.replace("_"," ") + ", Shapiro: %.2e"%(p_value_t))
leg = pl.legend(fancybox=True)
leg.get_frame().set_alpha(0.2)
#leg.get_frame().set_edgecolor("none")
remove_border()
pl.subplots_adjust(left=0.03,bottom=0.05,right=0.99,top=0.94,wspace=0.07,hspace=0.34)
pl.savefig(os.path.join(outdir,phenotype_name + ".pdf"))
if(p_value>p_value_t):
return "original"
else:
return transform
if len(sys.argv)==4:
transform = sys.argv[3]
[y,phenotype_names,sample_ids,fid] = read_data(sys.argv[1])
output_dir = sys.argv[2]
for i,phenotype in enumerate(phenotype_names):
selected_transform = plotHistogram(y=y[:,i],phenotype_name=phenotype,transform=transform,outdir=output_dir)
ind = sp.where(~sp.isnan(y[:,i]))[0]
if selected_transform=='sqrt':
phenotype_names[i] = "sqrt_" + phenotype
y[ind,i] = sp.sqrt(y[ind,i])
elif selected_transform=="boxcox":
phenotype_names[i] = "boxcox_" + phenotype
tmp = y[ind,i]
[y[ind,i],b_lambda] = boxcox(tmp)
elif selected_transform=="log":
phenotype_names[i] = "log_" + phenotype
y[ind,i] = sp.log(y[ind,i])
elif selected_transform=="log10":
phenotype_names[i] = "log10_" + phenotype
y[ind,i] = sp.log10(y[ind,i])
f = open(os.path.join(output_dir, "transformed_phenotypes.txt"),'w')
f.write("FID IID ")
string = ""
for phenotype in phenotype_names:
string += phenotype + " "
f.write(string[:-1] + "\n")
for i in range(fid.shape[0]):
f.write(fid[i] + " " + sample_ids[i] + " ")
def plotHistogram(y=None, phenotype_name=None, transform='sqrt', outdir=None):
ind = sp.where(~sp.isnan(y))
y = y[ind]
pl.figure(figsize=(12, 6))
pl.subplot(121)
[test_statistic, p_value] = shapiro(y)
pl.hist(y, bins=30, color=color_t[1], label="Original")
pl.title(phenotype_name.replace("_", " ") + ", Shapiro: %.2e" % (p_value))
leg = pl.legend(fancybox=True)
leg.get_frame().set_alpha(0.2)
#leg.get_frame().set_edgecolor("none")
remove_border()
pl.subplot(122)
if transform == "all":
p_vals = []
transformations = sp.array(['boxcox', 'sqrt', 'log', 'log10'])
for t in transformations:
zeros = sp.where(y == 0)[0]
if zeros.shape[0] == 0:
if t == 'sqrt':
tmpy = sp.sqrt(y)
elif t == "boxcox":
[tmpy, b_lambda] = boxcox(y)
elif t == "log":
tmpy = sp.log(y)
elif t == "log10":
tmpy = sp.log10(y)
[test_statistic, pv] = shapiro(tmpy)
else:
pv = 0.0
p_vals.append(pv)
p_vals = sp.array(p_vals)
ind = sp.argmax(p_vals)
transform = transformations[ind]
ind = sp.where(y == 0)[0]
if ind.shape[0] > 0:
print "IMPORTANT: y contains 0 -> transformation changed to SQRT"
transform = "sqrt"
if transform == 'sqrt':
y = sp.sqrt(y)
elif transform == "boxcox":
[y, b_lambda] = boxcox(y)
elif transform == "log":
y = sp.log(y)
elif transform == "log10":
y = sp.log10(y)
[test_statistic, p_value_t] = shapiro(y)
pl.hist(y, bins=30, color=color_t[4], label=transform)
pl.title(
phenotype_name.replace("_", " ") + ", Shapiro: %.2e" % (p_value_t))
leg = pl.legend(fancybox=True)
leg.get_frame().set_alpha(0.2)
#leg.get_frame().set_edgecolor("none")
remove_border()
pl.subplots_adjust(left=0.03,
bottom=0.05,
right=0.99,
top=0.94,
wspace=0.07,
hspace=0.34)
pl.savefig(os.path.join(outdir, phenotype_name + ".pdf"))
if (p_value > p_value_t):
return "original"
else:
return transform
[y, phenotype_names, sample_ids, fid] = read_data(sys.argv[1])
output_dir = sys.argv[2]
for i, phenotype in enumerate(phenotype_names):
selected_transform = plotHistogram(y=y[:, i],
phenotype_name=phenotype,
transform=transform,
outdir=output_dir)
ind = sp.where(~sp.isnan(y[:, i]))[0]
if selected_transform == 'sqrt':
phenotype_names[i] = "sqrt_" + phenotype
y[ind, i] = sp.sqrt(y[ind, i])
elif selected_transform == "boxcox":
phenotype_names[i] = "boxcox_" + phenotype
tmp = y[ind, i]
[y[ind, i], b_lambda] = boxcox(tmp)
elif selected_transform == "log":
phenotype_names[i] = "log_" + phenotype
y[ind, i] = sp.log(y[ind, i])
elif selected_transform == "log10":
phenotype_names[i] = "log10_" + phenotype
y[ind, i] = sp.log10(y[ind, i])
f = open(os.path.join(output_dir, "transformed_phenotypes.txt"), 'w')
f.write("FID IID ")
string = ""
for phenotype in phenotype_names:
string += phenotype + " "
f.write(string[:-1] + "\n")
for i in range(fid.shape[0]):
f.write(fid[i] + " " + sample_ids[i] + " ")
def train(data, dist_choices, FORCE_APPART = False, bc_transform = True): """ Given a vector of data and list of distribution types the trainer will find the best fit for the mixture distribution. """ param_vec = N.array([]) bounds = [] if bc_transform: t_data, llm = boxcox(data) else: t_data = data #make initial guesses based on kmeans num_clust = len(dist_choices) init_weight = float(1)/float(num_clust) if num_clust == 1: centroids = [N.mean(t_data)] elif (num_clust == 2) & FORCE_APPART: c1 = N.max(t_data) c2 = N.min(t_data) centroids = N.array([c1, c2]) else: (centroids, distortion) = kmeans(t_data, num_clust) centroids.sort() min_val = N.min(t_data) max_val = N.max(t_data) #create an "emperical" pdf bin_divisor = get_bin_divisor(len(t_data)) n, bins = N.histogram(t_data, new=True, bins = len(t_data)/bin_divisor, normed = True) if len(dist_choices) > 1: max_width = (bins[1]-bins[0])*bin_divisor else: max_width = None #create param_vec and bounds vector for this_dist, cent in izip(dist_choices, centroids): if this_dist == 'norm': try: param_vec = N.concatenate((param_vec, N.array([cent,1,init_weight]))) except: param_vec = N.concatenate((param_vec, N.array([cent[0],1,init_weight]))) bounds += [(min_val, max_val), (0, max_width), (0,1)] elif this_dist == 'uniform': param_vec = N.concatenate((param_vec, N.array([min_val,max_val,init_weight]))) bounds += [(min_val, max_val), (min_val, max_val), (0,1)] elif this_dist == 'skewnorm': try: param_vec = N.concatenate((param_vec, N.array([cent,1,1,init_weight]))) except: param_vec = N.concatenate((param_vec, N.array([cent[0],1,1,init_weight]))) bounds += [(min_val, max_val), (0, max_width), (0, max_width), (0,1)] else: raise KeyError, 'Unknown distribution %s' % this_dist #do the actual training param_val, like, d = fmin_tnc(score, param_vec, args = (dist_choices, bins[1:], n), approx_grad = True, bounds = bounds, messages = 0) #make the trained distribution t_dist = m_modal() t_dist.use_bc = bc_transform t_dist.dists, t_w = unpack(param_val, dist_choices) #save normalized weights t_dist.weights = N.array(t_w)/N.array(t_w).sum() #TF, pval = t_dist.pval(data) return t_dist
