viz_features.py

from sklearn.metrics import confusion_matrix
import matplotlib.pyplot as plt
import random
import cv2
import IPython
import numpy as np



class Viz_Feat(object):


    def __init__(self,val_data,train_data, class_labels,sess):

        self.val_data = val_data
        self.train_data = train_data
        self.CLASS_LABELS = class_labels
        self.sess = sess





    def vizualize_features(self,net):

        images = [0,10,100]
        '''
        Compute the response map for the index images
        '''

        for i in images:

            # validation data
            curr_img = self.val_data[i]['features']
            curr_img = np.reshape(curr_img, (1,) + curr_img.shape)
            curr_label = np.reshape(self.val_data[i]['label'], (1, -1))
            response_map = self.sess.run(net.response_map,
                                         feed_dict={net.images: curr_img, net.labels: curr_label})

            class_label = str(self.CLASS_LABELS[np.nonzero(curr_label[0])[0][0]])
            cv2.imwrite('val_' + str(i) + '_' + class_label  + '_raw.png', self.val_data[i]['c_img'])

            for ifilter in range(response_map.shape[-1]):
                cv2.imwrite('val_'  + str(i) + '_' + class_label  + '_filter-' + str(ifilter) + '.png',
                           self.revert_image(response_map[0, :, :, ifilter]))



    def revert_image(self,img):
        '''
        Used to revert images back to a form that can be easily visualized
        '''

        img = (img+1.0)/2.0*255.0

        img = np.array(img,dtype=int)

        blank_img = np.zeros([img.shape[0],img.shape[1],3])

        blank_img[:,:,0] = img
        blank_img[:,:,1] = img
        blank_img[:,:,2] = img

        img = blank_img.astype("uint8")

        return img