model_test.py

# coding: utf-8
"""
    Test file or on a subset of 2013 CASIA competition data

"""

import os
import codecs
import random
import copy
import argparse
import numpy as np
import tensorflow as tf
from scipy.misc import imread
from keras.utils.np_utils import to_categorical
from keras import backend
from the_model import model_8

parser = argparse.ArgumentParser()
parser.add_argument("-p", "--pred", action="store_true")
parser.add_argument("infile", nargs="?", type=argparse.FileType('rb'))
args = parser.parse_args()

random.seed(888)
np.random.seed(888)
tf.set_random_seed(888)

IMG_SIZE = 96

Pred_Details = False
if args.pred or args.infile:
    Pred_Details = True

TEST_PATH = os.path.join("data", "test")
WEIGHTS_PATH = os.path.join("data", "weights08.h5")
LABELS_PATH = os.path.join("data", "labels.txt")

label_file = codecs.open(LABELS_PATH, "r", "UTF-8")
klasses = [a.strip() for a in label_file.readlines()]
label_file.close()

model = model_8(IMG_SIZE, len(klasses))
model.load_weights(WEIGHTS_PATH)

model.compile(optimizer='adam',
              loss='categorical_crossentropy',
              metrics=['accuracy'])

if args.infile:
    test_data = np.ndarray([1, IMG_SIZE, IMG_SIZE], dtype=np.uint8)
    test_data[0] = imread(args.infile)

else:
    label_pngs = []
    for k, v in enumerate(klasses):
        for png in os.listdir(os.path.join(TEST_PATH, v)):
            label_pngs.append((k, v, png))

    print("Total number of test samples:", len(label_pngs))
    test_data = np.ndarray([len(label_pngs), IMG_SIZE, IMG_SIZE],
                           dtype=np.uint8)
    test_label = np.ndarray([len(label_pngs)], dtype=np.uint32)

    i = 0
    for label_png in label_pngs:
        fimg = open(os.path.join(TEST_PATH, label_png[1], label_png[2]), 'rb')
        test_data[i] = imread(fimg)
        test_label[i] = label_png[0]
        fimg.close()
        i += 1

    y_test = to_categorical(test_label)

x_test = test_data.reshape(test_data.shape[0],
                           test_data.shape[1],
                           test_data.shape[2],
                           1)
x_test = x_test.astype(np.float32)
x_test /= 255.0


def top_predictions(n, pred):
    tops = []
    pred_copy = copy.copy(pred)
    for j in range(n):
        i = np.argmax(pred_copy)
        tops.append((klasses[i], "%.2f" % (pred_copy[i] * 100)))
        pred_copy[i] = 0

    return tops


if args.infile:
    preds = model.predict(x_test)
    print("Prediction: top3:", top_predictions(3, preds[0]))

elif Pred_Details:
    preds = model.predict(x_test)
    for k, v in enumerate(preds):
        p = np.argmax(v)
        if p != test_label[k]:
            print("Wrong prediction: top3:", top_predictions(3, v),
                  "label/file:", label_pngs[k][1], label_pngs[k][2])
else:
    loss, acc = model.evaluate(x_test, y_test, batch_size=64)
    print("Loss:", loss)
    print("Accuracy:", acc)

backend.clear_session()