func_utils.py

from visualization import *
import numpy as np
import cv2
from keras import backend as K
from keras.preprocessing import image
import os


def ShowAttentionV1(model,image_path):
    file_list = os.listdir(image_path)
    file_list.sort()
    for filename in file_list:
        print (filename)
        filepath=image_path+filename
        image=load_image(filepath)
        image = image/255.0
        gender=1.0
        gender=np.asarray(gender)
        gender=np.expand_dims(gender,axis=0)
        layer=K.function([model.layers[0].input],[model.layers[196].output])
        FeatureMap=layer([image,gender])[0]
        print (FeatureMap.shape)
        FeatureMap = np.squeeze(FeatureMap, axis=0)
        FeatureMap = np.abs(FeatureMap)
        heatmap = np.mean(FeatureMap,axis=2)
        heatmap = heatmap/np.max(heatmap)
        heatmap = np.uint8(255*heatmap)
        print (heatmap.shape)
        heatmap = cv2.applyColorMap(heatmap,cv2.COLORMAP_JET)
        SaveImg(filename,filepath,heatmap)
    print ('********** Done ***********')




def GAPAttention(model,weights,image_path):
    file_list = os.listdir(image_path)
    file_list.sort()
    for filename in file_list:
        filepath=image_path+filename
        print (filepath)
        image=load_image(filepath)
        image = image/255.0
        gender=1.0
        gender=np.asarray(gender)
        gender=np.expand_dims(gender,axis=0)
        layer=K.function([model.layers[0].input],[model.layers[1].get_output_at(-1),model.layers[-1].output])
        GAP,prediction=layer([image,gender])
        GAP=np.squeeze(GAP,axis=0)
        print (GAP.shape)
        index = np.argmax(prediction)
        print (index)
       # weight = weights[:,index]
        weight =np.mean(weights[:,index-5:index+5],axis=1)
        heatmap = np.zeros((GAP.shape[0],GAP.shape[1]))
        for k in range(GAP.shape[2]):
            heatmap = heatmap + weight[k]*GAP[:,:,k]
        heatmap = heatmap/np.max(heatmap)
        heatmap = np.uint8(255*heatmap)
        print (heatmap.shape)
        heatmap = cv2.applyColorMap(heatmap,cv2.COLORMAP_JET)
        SaveImg(filename,filepath,heatmap)
    print ('********** Done ***********')



def SaveImg(filename,filepath,heatmap):
    img = cv2.imread(filepath)
    heatmap = cv2.resize(heatmap,(img.shape[1],img.shape[0]))
    AttentionImg =0.5* heatmap + img
    cv2.imwrite('heatmap/'+filename,heatmap)
    cv2.imwrite('AttentionImg/'+filename,AttentionImg)


def load_image(path):
    img = cv2.imread(path)
    img = cv2.resize(img,(300,300))
    x = np.asarray(img, dtype=np.float32)
   # img = image.load_img(path, target_size=(448, 448))
   # print (img.shape)
   # x = image.img_to_array(img)
    x = np.expand_dims(x, axis=0)
    return x




def softlabel(label,num_class):
    softlabel=np.zeros((len(label),num_class))
    ratio = 1.0/50
    for i in range(len(label)):
        for j in range(num_class):
            softlabel[i,j]=1.0 - ratio*np.abs(j-label[i])
    softlabel = np.maximum(softlabel,0)
    return softlabel



def GaussLabel(label,num_class):
    sigma=15.0
    GaussLabel = np.zeros((len(label),num_class))
    x = np.array(range(num_class))+1
    for k in range(len(label)):
        GaussLabel[k,:]=np.exp(-(x-label[k])**2/(2.0*sigma**2))
    return GaussLabel


def TestMAE(model,test_data,test_label,test_gender):
    test_gender = np.array(test_gender)
    test_gender = np.expand_dims(test_gender,axis=1)
    layer=K.function([model.layers[0].input,model.layers[3].input],[model.layers[-1].output])
    predictions=layer([test_data,test_gender])
    predictions = np.array(predictions)
    predictions = np.squeeze(predictions,axis=0)
    print (predictions.shape)
    predict_label = np.argmax(predictions,axis=1)
    test_label = np.argmax(test_label,axis=1)
    print (predict_label)
    print (test_label)
    TestMAE = np.mean(np.abs(predict_label-test_label))
    return TestMAE



def DataAugment(x_train):
    x_train_Aug = np.zeros(x_train.shape)
    for i in range(x_train.shape[0]):
        for j in range(3):
            img = x_train[i,:,:,j]
            img = RandomMask(img)
            img = RandomMask(img)
            if np.random.random()>-1:
                x_train_Aug[i,:,:,j]=img 
            else:
                x_train_Aug[i,:,:,j]=x_train[i,:,:,j]
    return x_train_Aug


def RandomMask(img):
    m,n=img.shape
    m=int(m/6)
    n=int(n/6)
    i,j = np.random.randint(0,6,2)
    img[i*m:(i+1)*m,j*n:(j+1)*n]=np.random.random()
    return img