test.py

# Adapted from: Arun Ponnusamy
# website: https://www.arunponnusamy.com

# import necessary packages
from keras.preprocessing.image import img_to_array
from keras.models import load_model
from keras.utils import get_file
import numpy as np
import argparse
import cv2
import os, glob
import cvlib as cv
import matplotlib.pyplot as plt

import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import seaborn as sns
from sklearn.metrics import confusion_matrix
from utils import cm_analysis

# handle command line arguments
ap = argparse.ArgumentParser()
ap.add_argument("-t", "--testdir", required=True, #type=str, default="classification_mary/test",
	help="path to test images' folder")
ap.add_argument("-p", "--preddir", required=True, #type=str, default="predictions_B_softmax",
	help="path to predictions' folder")
ap.add_argument("-m", "--model", required=True,
	help="Path to the trained model")
ap.add_argument("-hq", "--high_quality", type=bool, default=False)

args = ap.parse_args()
model_path = args.model 

if os.path.exists(model_path):
    print ("Model path is :" + str(model_path))
else:
    raise Exception("Model not found")

# load model
print (model_path)
model = load_model(model_path)

# read the test images' directory
image_files = [f for f in glob.glob(args.testdir + "/**/*", recursive=True) if not os.path.isdir(f)]

#print (image.shape)
if image_files is None:
    print("Could not test images' folder")
    exit()

'''
In our case/paper, we have considered Gabriel as male and
Mary as female. Since we are also using the model to fine-tune on
styled data, we use the common labels : male and female. 
'''
classes = ['Mary','Gabriel']
classes_a = ['Gabriel', 'Mary']

all_preds = []
all_groundtruths = []

# Generate predictions_old as well as face bboxes
for imagename in image_files:
    # detect faces in the image
    image = cv2.imread(imagename)
    face, confidence = cv.detect_face(image)
    
    if confidence != []:
         # get corner points of face rectangle
        (startX, startY) = face[0][0], face[0][1]
        (endX, endY) = face[0][2], face[0][3]

        # draw rectangle over face
        cv2.rectangle(image, (startX,startY), (endX,endY), (0,255,0), 2)

    else:
        startY = 10
        startX = 10

    if (np.shape(image)[0] != 0)  and (np.shape(image)[1] != 0):
        image = cv2.resize(image, (224,224))
        image = image.astype("float") / 255.0
        image = img_to_array(image)
        image = np.expand_dims(image, axis=0)
        
        # apply gender detection on the body
        conf = model.predict(image)[0]

        image = np.uint8(np.squeeze(image)*255.0)
        # get label with max accuracy
        idx = np.argmax(conf)
        
        # While training for Model A, we used classes_a as labels.
        if model_path.split('/')[-1] == 'A_VGGFace_resnet_E100_B32_I224_cvfold_2.model':
            classes = classes_a
        
        label = classes[idx]
        possible_true_label = imagename.split('/')

        if 'female' in possible_true_label or 'female' in possible_true_label[-1].split('_'):
            true_label = 'Mary'
        else:
            true_label = 'Gabriel'
                
        all_preds.append(label)
        all_groundtruths.append(true_label)
        
        label = "{}: {:.2f}%".format(label, conf[idx] * 100)

        Y = startY - 10 if startY - 10 > 10 else startY + 10

        # write label and confidence above face rectangle
        cv2.putText(image, label, (startX, Y),  cv2.FONT_HERSHEY_SIMPLEX,
            0.5, (0, 0, 255), 1)

        if not os.path.exists(args.preddir):
            print("Making the output directory")
            os.makedirs(args.preddir)

        out_image_name = '_'.join(imagename.split('/')[2:])
        cv2.imwrite(args.preddir + os.sep + out_image_name, image)

        if args.high_quality == True:
            high_quality_image_name = args.image.split('/')[-1].split('.jpg')[0] + '_gender_detection.eps'
            print("Saving {} image".format(high_quality_image_name))
            plt.imshow(image)
            plt.savefig(high_quality_image_name, format='eps', dpi=1000)
            
cm_folder = 'final_stuff/confusion_matrix/' 
if not os.path.isdir(cm_folder):
    os.makedirs(cm_folder)
    
cm_path = cm_folder + model_path.split('/')[-1].split('.')[0] + '_confusion_matrix.eps'
cm = cm_analysis(all_groundtruths, all_preds, cm_path, ['Mary', 'Gabriel'])
print (cm)