revise grid search python script

Your question is very nice. SVM researchers are not interested in penalty c,

.See my several papers about  c. I compare H-SVM, S-SVM (c=10000, c=1) and other LDFs. My results is C=1 is almost wrong than c=10000.

Thanks all. But my understanding is, values of c and gamma will be obtained from data and grid-search and cross validation ,of the SVM. Also in many papers i read that Linear SVM with C=1 gives the best results. How can it be true?

No, I discriminate many data by soft-margin SVM (s-SVM for C=10000 and C=1), H-SVM, our three LDFs, logistic regression and Fisher's LDF by 100-fold cross validation for small sample method. In many cases, Fisher's LDF is worst and S-SVM with C=1 is second worst by the mean of error rate in the validation samples.

Many researchers are lack of complete comparisons. 

The value of C depends on the range of feature vector. So, comparing the value C is not valid without showing the range of feature.

Use evolutionary algorithms such as GA, PSO and etc. These algorithms can search solution space and find best values for C or others. Your cost function may should be selected as output error of sample classifier.

####################### SVM training ###################################

param = svm_parameter("-q")

problem = svm_problem(y_train, x_train)

model= svm_train(problem, param)

k = param.kernal_type = 2 # 2 means RBF Kernal

results = []

######################### prediction ###################################

for c in range(-1, 10): #here you can change the value of c

for g in range (2, 8): # #here you can change the value of g (gamma)

param.C, param.gamma, param.kernal_type=2**c, 2**g, k

pred_label, pred_accuracy, pred_val = svm_predict(y_test, x_test, model)

results.append([param.C, param.gamma, param.kernal_type, pred_accuracy[0]])

in my above code :

last line append all the output of , gamma, kernal_type and accuracy.

so you van print results

print(results) # it shows the value of c, gamma , kernal_type and accuracy.