# Created by Dennis Willsch (d.willsch@fz-juelich.de) # Modified by Gabriele Cavallaro (g.cavallaro@fz-juelich.de) # and Madita Willsch (m.willsch@fz-juelich.de) import sys import re import json import os import numpy as np import matplotlib.pyplot as plt import matplotlib.colors as cols from sklearn.metrics import roc_auc_score,average_precision_score,precision_recall_curve,roc_curve,accuracy_score,auc np.set_printoptions(precision=4, suppress=True) def kernel(xn, xm, gamma=-1): # here (xn.shape: NxD, xm.shape: ...xD) -> Nx... if gamma == -1: return xn @ xm.T xn = np.atleast_2d(xn) xm = np.atleast_2d(xm) return np.exp(-gamma * np.sum((xn[:,None] - xm[None,:])**2, axis=-1)) # (N,1,D) - (1,...,D) -> (N,...,D) -> (N,...); see Hsu guide.pdf for formula # B = base # K = number of qubits per alpha # E = shift of exponent # decode binary -> alpha def decode(binary, B=10, K=3, E=0): N = len(binary) // K Bvec = float(B) ** (np.arange(K)-E) return np.fromiter(binary,float).reshape(N,K) @ Bvec # encode alpha -> binary with B and K (for each n, the binary coefficients an,k such that sum_k an,k B**k is closest to alphan) def encode(alphas, B=10, K=3, E=0): # E allows for encodings with floating point numbers (limited precision of course) N = len(alphas) Bvec = float(B) ** (np.arange(K)-E) # B^(0-E) B^(1-E) B^(2-E) ... B^(K-1-E) allvals = np.array(list(map(lambda n : np.fromiter(bin(n)[2:].zfill(K),float,K), range(2**K)))) @ Bvec # [[0,0,0],[0,0,1],...] @ [1, 10, 100] return ''.join(list(map(lambda n : bin(n)[2:].zfill(K),np.argmin(np.abs(allvals[:,None] - alphas), axis=0)))) def encode_as_vec(alphas, B=10, K=3, E=0): return np.fromiter(encode(alphas,B,K,E), float) def loaddataset(datakey): dataset = np.loadtxt(datakey, dtype=float, skiprows=1) return dataset[:,2:], dataset[:,1] # data, labels def save_json(filename, var): with open(filename,'w') as f: f.write(str(json.dumps(var, indent=4, sort_keys=True, separators=(',', ': '), ensure_ascii=False))) def eval_classifier(x, alphas, data, label, gamma, b=0): # evaluates the distance to the hyper plane according to 16.5.32 on p. 891 (Numerical Recipes); sign is the assigned class; x.shape = ...xD return np.sum((alphas * label)[:,None] * kernel(data, x, gamma), axis=0) + b def eval_on_sv(x, alphas, data, label, gamma, C): return np.sum((alphas * (C-alphas) * label)[:,None] * kernel(data, x, gamma), axis=0) def eval_offset_search(alphas, data, label, gamma, C, useavgforb=True): # search for the best offset maxacc=0 b1=-9 for i in np.linspace(-9,9,500): acc = accuracy_score(label,np.sign(eval_classifier(data, alphas, data, label, gamma, i))) if acc > maxacc: maxacc = acc b1=i maxacc=0 b2=9 reversed_space=np.linspace(-9,9,500)[::-1] for i in reversed_space: acc = accuracy_score(label,np.sign(eval_classifier(data, alphas, data, label, gamma, i))) if acc > maxacc: maxacc = acc b2=i return (b1+b2)/2 def eval_offset_MM(alphas, data, label, gamma, C, useavgforb=True): # evaluates offset b according to 16.5.37 (Mangasarian-Musicant variant) NOTE: does not seem to work with integer/very coarsely spaced alpha! return np.sum(alphas*label) def eval_offset_avg(alphas, data, label, gamma, C, useavgforb=True): # evaluates offset b according to 16.5.33 cross = eval_classifier(data, alphas, data, label, gamma) # cross[i] = sum_j aj yj K(xj, xi) (error in Numerical Recipes) if useavgforb: return np.sum(alphas * (C-alphas) * (label - cross)) / np.sum(alphas * (C-alphas)) #return np.sum(label - cross) / num_sv else: # this is actually not used, but we did a similar-in-spirit implementation in eval_finaltraining_avgscore.py if np.isclose(np.sum(alphas * (C-alphas)),0): print('no support vectors found, discarding this classifer') return np.nan bcandidates = [np.sum(alphas * (C-alphas) * (label - cross)) / np.sum(alphas * (C-alphas))] # average according to NR should be the first candidate crosssorted = np.sort(cross) crosscandidates = -(crosssorted[1:] + crosssorted[:-1])/2 # each value between f(xi) and the next higher f(xj) is a candidate bcandidates += sorted(crosscandidates, key=lambda x:abs(x - bcandidates[0])) # try candidates closest to the average first bnumcorrect = [(label == np.sign(cross + b)).sum() for b in bcandidates] return bcandidates[np.argmax(bnumcorrect)] def eval_acc_auroc_auprc(label, score): # score is the distance to the hyper plane (output from eval_classifier) precision,recall,_ = precision_recall_curve(label, score) return accuracy_score(label,np.sign(score)), roc_auc_score(label,score), auc(recall,precision) ################ This I/O functions are provided by http://hyperlabelme.uv.es/index.html ################ def dataread(filename): lasttag = 'description:' # Open file and locate lasttag f = open(filename, 'r') nl = 1 for line in f: if line.startswith(lasttag): break nl += 1 f.close() # Read data data = np.loadtxt(filename, delimiter=',', skiprows=nl) Y = data[:, 0] X = data[:, 1:] # Separate train/test Xtest = X[Y < 0, :] X = X[Y >= 0, :] Y = Y[Y >= 0, None] return X, Y, Xtest def datawrite(path,method, dataset, Yp): filename = '{0}{1}_predictions.txt'.format(path, dataset) res = True try: with open(filename, mode='w') as f: f.write('{0} {1}'.format(method, dataset)) for v in Yp: f.write(' {0}'.format(str(v))) f.write('\n') except Exception as e: print('Error', e) res = False return res ################ def write_samples(X, Y,path): f = open(path,"w") f.write("id label data \n") for i in range(0,X.shape[0]): f.write(str(i)+" ") if(Y[i]==1): f.write("-1 ") else: f.write("1 ") for j in range(0,X.shape[1]): f.write(str(X[i,j])+" ") f.write("\n") f.close()