def evaluate(self, parser=rightParse):
noverlap = 0
nmodel = 0
ngolden = 0
for parsed, tagged in zip(self.parsed(), self.tagged()):
if (len(tagged) >= 2):
golden = spannings(parsed, unary=False, root=True)
if parser != uBoundParse:
if parser != randomTreeParse:
model = spannings(parser(tagged), unary=False, root=True)
else:
model = set(parser(len(tagged)))
m = Metric(golden, model)
e = m.evaluation()
else:
e = [ len(golden), len(golden), len(tagged) - 1 ]
noverlap += e[0]
ngolden += e[1]
nmodel += e[2]
else:
if len(parsed.leaves()) != len(tagged):
print tagged, "doesn't match."
print_metrics(noverlap, ngolden, nmodel)
def evaluate(self,parser=forwardDependency, directed = True):
noverlap = 0
nmodel = 0
ngolden = 0
for dep in self.iterator(renum=True):
if ( dep.number_of_nodes() - 1 >= 2 ):
gold = set(dep.edges())
model = set(parser(dep.number_of_nodes() - 1))
#if len(gold) != len(model):
# print gold
# print model
# print [sent.node[i]['word'] for i in sent.nodes()]
# print o
#print "Gold: ",
#print gold
#print "Model: ",
#print model
if len(gold) == len(model):
if directed:
for c in gold:
if c in model:
noverlap += 1
else:
for head,sink in gold:
if (head,sink) in model or (sink,head) in model:
noverlap += 1
ngolden += len(gold)
nmodel += len(model)
print_metrics(noverlap, ngolden, nmodel)
def run_ann(X, y, ep, bs, perc, training_size, seed, out_dir):
sizes = 1
F1 = np.zeros((sizes, 3))
PREC = np.zeros((sizes, 3))
REC = np.zeros((sizes, 3))
# Stratified random data
x_train, y_train, idx_0, idx_1 = strati_training_data(
X, y, size=training_size, p_class1=perc,
seed=int(seed)) #size=20000,p_class1=0.2)
print("x_train, y_train", x_train.shape, y_train.shape)
rus = RandomUnderSampler(random_state=0)
x_res, y_res = rus.fit_sample(x_train, y_train)
#t1=time.time()
#sfm = feat_sel(x_res, y_res)
#x_sel = sfm.transform(x_res)
#t2=time.time()
x_sel = x_res
#print("Feature selection time:", t2-t1)
x_train_2, x_test, y_train_2, y_test = train_test_split(
x_sel, y_res, test_size=test_perc, random_state=0)
#sfm = SelectFromModel(LassoCV())
#sfm.fit(x_train_2, y_train_2)
#train = sfm.transform(x_train_2)
#XX = sfm.transform(X)
XX = X
global input_dim
input_dim = x_train_2.shape[1]
#pred_training, pred_big_im, pred_testing = ann(x_train_2,y_train_2,X)
pred_training, pred_big_im, pred_testing = ann(x_train_2, y_train_2, XX,
ep, bs, x_test)
#preffix = out_dir+"/"+str(ep)+"_"+str(bs)+"_"+ str(perc) + "_"
preffix = out_dir + "/" + img_name + "/ann_strati_0/" + "_".join(
[str(ep), str(bs), str(perc), seed, ""])
print("preffix", preffix)
f1_train = open(preffix + "F1_train.txt", "a")
f1_test = open(preffix + "F1_test.txt", "a")
f1_big = open(preffix + "F1_big.txt", "a")
prec_train = open(preffix + "PREC_train.txt", "a")
prec_test = open(preffix + "PREC_test.txt", "a")
prec_big = open(preffix + "PREC_big.txt", "a")
rec_train = open(preffix + "REC_train.txt", "a")
rec_test = open(preffix + "REC_test.txt", "a")
rec_big = open(preffix + "REC_big.txt", "a")
i = 0
cm_train, F1[i,
0], PREC[i,
0], REC[i,
0] = print_metrics("Trainng", y_train_2,
pred_training)
f1_train.write(
str(trainSize) + ":" + str(seed) + ":" + str(F1[i, 0]) + "\n")
prec_train.write(
str(trainSize) + ":" + str(seed) + ":" + str(PREC[i, 0]) + "\n")
rec_train.write(
str(trainSize) + ":" + str(seed) + ":" + str(REC[i, 0]) + "\n")
cm_test, F1[i, 1], PREC[i,
1], REC[i,
1] = print_metrics("Test", y_test,
pred_testing)
f1_test.write(
str(trainSize) + ":" + str(seed) + ":" + str(F1[i, 1]) + "\n")
prec_test.write(
str(trainSize) + ":" + str(seed) + ":" + str(PREC[i, 1]) + "\n")
rec_test.write(
str(trainSize) + ":" + str(seed) + ":" + str(REC[i, 1]) + "\n")
cm_big, F1[i, 2], PREC[i,
2], REC[i,
2] = print_metrics("Big image", y,
pred_big_im)
f1_big.write(str(trainSize) + ":" + str(seed) + ":" + str(F1[i, 2]) + "\n")
prec_big.write(
str(trainSize) + ":" + str(seed) + ":" + str(PREC[i, 2]) + "\n")
rec_big.write(
str(trainSize) + ":" + str(seed) + ":" + str(REC[i, 2]) + "\n")
scores = str(np.round(F1[i, 2], 2)) + "_" + str(np.round(
PREC[i, 2], 2)) + "_" + str(np.round(REC[i, 2], 2))
plt.imsave(preffix + str(training_size) + "_" + scores + "_pred.png",
np.reshape(pred_big_im, (1000, 1000)))
f1_train.close()
f1_test.close()
f1_big.close()
def run_svm(training_size, out_dir):
sizes = 1
F1 = np.zeros((sizes, 3))
PREC = np.zeros((sizes, 3))
REC = np.zeros((sizes, 3))
#TIME = np.zeros((sizes,1))
x_train, y_train, idx_0, idx_1 = strati_training_data(X,
y,
size=training_size,
p_class1=perc,
seed=int(seed))
rus = RandomUnderSampler(random_state=0)
x_res, y_res = rus.fit_sample(x_train, y_train)
#t1=time.time()
#sfm = feat_sel(x_res, y_res)
#x_sel = sfm.transform(x_res)
#t2=time.time()
#print("Feature selection time:", t2-t1)
x_sel = x_res
x_train_2, x_test, y_train_2, y_test = train_test_split(
x_sel, y_res, test_size=test_perc, random_state=0)
#XX = sfm.transform(X)
XX = X
kernel = 'rbf'
C_r = [-5, 5]
C_step = 0.5
g_r = [-5, 5]
g_step = 0.5
C_range = 10.**np.arange(C_r[0], C_r[1], C_step)
gamma_range = 10.**np.arange(g_r[0], g_r[1], g_step)
# score can be: roc_auc, accuracy, recall, precision, f1, average_precision
#score = 'average_precision'
score = "f1"
print(":::run_svm::: x_train_2.shape:", x_train_2.shape)
pred_training, pred_big_im, pred_testing, best_params = svm(
x_train_2, y_train_2, XX, x_test, kernel, gamma_range, C_range, score)
#print(best_params)
best_C = best_params["C"]
best_gamma = best_params["gamma"]
best_k = best_params["kernel"]
#preffix = out_dir+"/"+str(ep)+"_"+str(bs)+"_"+ str(perc) + "_"
preffix = out_dir + "/" + img_name + "/svm_strati_0/" + "_".join(
[str(best_C), str(best_gamma), best_k, seed, ""])
print("preffix", preffix)
f1_train = open(preffix + "F1_train.txt", "a")
f1_test = open(preffix + "F1_test.txt", "a")
f1_big = open(preffix + "F1_big.txt", "a")
prec_train = open(preffix + "PREC_train.txt", "a")
prec_test = open(preffix + "PREC_test.txt", "a")
prec_big = open(preffix + "PREC_big.txt", "a")
rec_train = open(preffix + "REC_train.txt", "a")
rec_test = open(preffix + "REC_test.txt", "a")
rec_big = open(preffix + "REC_big.txt", "a")
i = 0
cm_train, F1[i,
0], PREC[i,
0], REC[i,
0] = print_metrics("Trainng", y_train_2,
pred_training)
f1_train.write(
str(trainSize) + ":" + str(seed) + ":" + str(F1[i, 0]) + "\n")
prec_train.write(
str(trainSize) + ":" + str(seed) + ":" + str(PREC[i, 0]) + "\n")
rec_train.write(
str(trainSize) + ":" + str(seed) + ":" + str(REC[i, 0]) + "\n")
cm_test, F1[i, 1], PREC[i,
1], REC[i,
1] = print_metrics("Test", y_test,
pred_testing)
f1_test.write(
str(trainSize) + ":" + str(seed) + ":" + str(F1[i, 1]) + "\n")
prec_test.write(
str(trainSize) + ":" + str(seed) + ":" + str(PREC[i, 1]) + "\n")
rec_test.write(
str(trainSize) + ":" + str(seed) + ":" + str(REC[i, 1]) + "\n")
cm_big, F1[i, 2], PREC[i,
2], REC[i,
2] = print_metrics("Big image", y,
pred_big_im)
f1_big.write(str(trainSize) + ":" + str(seed) + ":" + str(F1[i, 2]) + "\n")
prec_big.write(
str(trainSize) + ":" + str(seed) + ":" + str(PREC[i, 2]) + "\n")
rec_big.write(
str(trainSize) + ":" + str(seed) + ":" + str(REC[i, 2]) + "\n")
scores = str(np.round(F1[i, 2], 2)) + "_" + str(np.round(
PREC[i, 2], 2)) + "_" + str(np.round(REC[i, 2], 2))
plt.imsave(preffix + str(training_size) + "_" + scores + "_pred.png",
np.reshape(pred_big_im, (1000, 1000)))
f1_train.close()
f1_test.close()
f1_big.close()