print("doing test...", "file :", X_filenames[test_idx], ", net:",
"netname", ", win_size:", win_size, ", sampling: ", sampl_style)
# compute the sensitivity map
#layer_name = net.blobs.keys()[-2] # look at penultimate layer (like in Simonyan et al. (2013))
#sensMap = SA.get_sens_map(net, x_test[np.newaxis], layer_name, np.argmax(target_func(x_test)[-1][0]))
start_time = time.time()
if sampl_style == 'conditional':
sampler = utlS.cond_sampler_imagenet(win_size=win_size,
padding_size=padding_size,
image_dims=[224, 224])
elif sampl_style == 'marginal':
sampler = utlS.marg_sampler_imagenet(X_test)
pda = PredDiffAnalyser(x_test,
target_func,
sampler,
num_samples=num_samples,
batch_size=batch_size)
pred_diff = pda.get_rel_vect(win_size=win_size, overlap=overlapping)
#print(len(pred_diff))
#plot and save the results
utlV.plot_results(x_test, x_test_im, None, pred_diff[0], target_func,
classnames, test_idx, save_path)
np.savez(save_path, *pred_diff)
print("--- Total computation took {:.4f} minutes ---".format(
(time.time() - start_time) / 60))
print("--- Total computation took {:.4f} seconds ---".format(
(time.time() - start_time)))
# plot and save the results
existing_results = glob.glob(path_base + '*.png')
if os.path.exists(save_path + '.png'):
print('Relevance map for test{} exist, will move to the next'
' image.'.format(test_idx))
elif existing_results:
print('Results for test{} exist under a different suffix'.format(
test_idx))
print('Linking {} to {}'.format(existing_results[0],
save_path + '.png'))
os.system('ln -s {} {}'.format(existing_results[0],
save_path + '.png'))
else:
print(
"Plotting...test image: {}, net: {}, window size: {}, sampling: {}"
.format(test_idx, netname, win_size, sampl_style))
salience_dict = pickle.load(open(save_path + '.p', 'rb'))
if mode == 'BNN':
titles = ['epistemic', 'aleatoric', 'predictive']
diffs = [salience_dict[x] for x in titles]
titles.append('pred')
diffs.append(salience_dict['pred'][:, salience_dict['y_pred']])
elif mode == 'DNN':
titles = labels
diffs = [salience_dict['pred'][:, ii] for ii in range(len(labels))]
utlV.plot_results(x_test_im, y_true, salience_dict['y_pred'], diffs,
titles, labels, save_path + '.png')