ACTL3143 & ACTL5111 Deep Learning for Actuaries
import random
import matplotlib.pyplot as plt
import numpy as np
import pandas as pd
import keras
from keras.models import Sequential
from keras.layers import Dense, Input
from keras.callbacks import EarlyStopping
from keras.utils import plot_model
from sklearn.model_selection import train_test_split
from sklearn.preprocessing import StandardScalerComputer vision is a field of Artificial Intelligence (AI) that focuses on extracting meaningful information from visual data (images and videos). One of the primary goals of computer vision is to correctly identify and classify visual data. Convolution Neural Networks (CNNs) are the most commonly used neural network architectures for computer vision related tasks.
A special attention to shapes of data are important in CNN architectures, because CNNs have special types of layers (e.g. convolution and pooling) which require explicit specifications of array dimensions.
Since the position of a pixel(one small sqaure) in a photo can be represented using 3 positional values, we call it a rank 3 tensor.
from matplotlib.image import imread
img1 = imread('pu.gif'); img2 = imread('pl.gif')
img3 = imread('pr.gif'); img4 = imread('pg.bmp')
f"Shapes are: {img1.shape}, {img2.shape}, {img3.shape}, {img4.shape}."'Shapes are: (16, 16, 3), (16, 16, 3), (16, 16, 3), (16, 16, 3).'img1array([[[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0]],
[[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0]],
[[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[255, 255, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[255, 255, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0]],
[[ 0, 0, 0],
[ 0, 0, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[ 0, 0, 0],
[ 0, 0, 0]],
[[ 0, 0, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[ 0, 0, 0]],
[[ 0, 0, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[ 0, 0, 0]],
[[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0]],
[[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[ 0, 0, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0]],
[[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0]],
[[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0]],
[[ 0, 0, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[ 0, 0, 0]],
[[ 0, 0, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[ 0, 0, 0]],
[[ 0, 0, 0],
[ 0, 0, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[ 0, 0, 0],
[ 0, 0, 0]],
[[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0]],
[[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0]],
[[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0]]], dtype=uint8)img2array([[[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0]],
[[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0]],
[[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0]],
[[ 0, 0, 0],
[ 0, 0, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[ 0, 0, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[ 0, 0, 0],
[ 0, 0, 0]],
[[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[ 0, 0, 0]],
[[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[ 0, 0, 0]],
[[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0]],
[[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0]],
[[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0]],
[[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0]],
[[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[ 0, 0, 0]],
[[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[ 0, 0, 0]],
[[ 0, 0, 0],
[ 0, 0, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[ 0, 0, 0],
[ 0, 0, 0]],
[[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0]],
[[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0]],
[[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0]]], dtype=uint8)img3array([[[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0]],
[[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0]],
[[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0]],
[[ 0, 0, 0],
[ 0, 0, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[ 0, 0, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[ 0, 0, 0],
[ 0, 0, 0]],
[[ 0, 0, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0]],
[[ 0, 0, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0]],
[[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0]],
[[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0]],
[[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0]],
[[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0]],
[[ 0, 0, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0]],
[[ 0, 0, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0]],
[[ 0, 0, 0],
[ 0, 0, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[ 0, 0, 0],
[ 0, 0, 0]],
[[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0]],
[[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0]],
[[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[255, 255, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0]]], dtype=uint8)img4array([[[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0]],
[[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0]],
[[ 0, 0, 0],
[ 0, 0, 0],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[ 0, 0, 0],
[ 0, 0, 0]],
[[ 0, 0, 0],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 255, 255],
[255, 255, 255],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 255, 255],
[255, 255, 255],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[ 0, 0, 0]],
[[ 0, 0, 0],
[255, 163, 177],
[255, 163, 177],
[255, 255, 255],
[255, 255, 255],
[255, 255, 255],
[255, 255, 255],
[255, 163, 177],
[255, 163, 177],
[255, 255, 255],
[255, 255, 255],
[255, 255, 255],
[255, 255, 255],
[255, 163, 177],
[255, 163, 177],
[ 0, 0, 0]],
[[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 255, 255],
[255, 255, 255],
[255, 255, 255],
[255, 255, 255],
[255, 163, 177],
[255, 163, 177],
[255, 255, 255],
[255, 255, 255],
[255, 255, 255],
[255, 255, 255],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177]],
[[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[ 51, 0, 255],
[ 51, 0, 255],
[255, 255, 255],
[255, 255, 255],
[255, 163, 177],
[255, 163, 177],
[ 51, 0, 255],
[ 51, 0, 255],
[255, 255, 255],
[255, 255, 255],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177]],
[[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[ 51, 0, 255],
[ 51, 0, 255],
[255, 255, 255],
[255, 255, 255],
[255, 163, 177],
[255, 163, 177],
[ 51, 0, 255],
[ 51, 0, 255],
[255, 255, 255],
[255, 255, 255],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177]],
[[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 255, 255],
[255, 255, 255],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 255, 255],
[255, 255, 255],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177]],
[[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177]],
[[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177]],
[[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177]],
[[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177]],
[[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[ 0, 0, 0],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[ 0, 0, 0],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[ 0, 0, 0]],
[[255, 163, 177],
[255, 163, 177],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[255, 163, 177],
[255, 163, 177],
[255, 163, 177],
[ 0, 0, 0],
[ 0, 0, 0]],
[[255, 163, 177],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[255, 163, 177],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0],
[255, 163, 177],
[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, 0]]], dtype=uint8)The above code reads 4 images and then shows how computers read those images. Each image is read by the computer as a rank 3 tensor. Each image is of (16,16,3) dimensions.
from matplotlib.pyplot import imshowimshow(img1);
imshow(img2);
imshow(img3);
imshow(img4);
Each pixel’s colour intensity is stored in one byte.
One byte is 8 bits, so in binary that is 00000000 to 11111111.
The largest unsigned number this can be is 2^8-1 = 255.
np.array([0, 1, 255, 256]).astype(np.uint8)array([ 0, 1, 255, 0], dtype=uint8)If you had signed numbers, this would go from -128 to 127.
np.array([-128, 1, 127, 128]).astype(np.int8)array([-128, 1, 127, -128], dtype=int8)Alternatively, hexidecimal numbers are used. E.g. 10100001 is split into 1010 0001, and 1010=A, 0001=1, so combined it is 0xA1.
Take a look at https://setosa.io/ev/image-kernels/.
Say X_1, X_2 \sim f_X are i.i.d., and we look at S = X_1 + X_2.
The density for S is then
f_S(s) = \int_{x_1=-\infty}^{\infty} f_X(x_1) \, f_X(s-x_1) \,\mathrm{d}s .
This is the convolution operation, f_S = f_X \star f_X.
Height, width, and number of channels.
An image can be represented using a rank 3 tensor, since it has 3 dimensions; height, width, and number of channels. The number of channels is also known as the ‘depth’. The left hand side of the picture shown below is tensor with height =5, width =5 and depth =3.
Grayscale image has 1 channel. RGB image has 3 channels.
Each colour can be represented as a combination of three primary colours; red, green and blue.
Example: Yellow = Red + Green.
Suppose we wish to detect if a picture has yellow colour in it. One option would be to apply a neuron over each pixel and see if it detects the colour yellow. We know that each pixel is represented by 3 numerical values that correspond to red, green and blue. Higher numeric values for red and green indicate higher chances of detecting yellow. Higher values for blue indicate lower chances of detecting yellow. Utilising this information, we can assign RGB weights to be 1, 1, -1 respectively.
Next, a standard multiplication between numeric values and weights is carried out, and the weighted sum is passed through the neuron.
Apply a neuron to each pixel in the image.
If red/green \nearrow or blue \searrow then yellowness \nearrow.
Set RGB weights to 1, 1, -1.
The output is produced by sweeping the neuron over the input. This is called convolution.
The following picture demonstrates how yellow-coloured areas (in the colour picture) are transformed into a white colour (in the greyscale picture). This is a result of the way we assigned the weights. Since we assigned +1 weights to red and green, and -1 to blue, it ended up resulting in large positive values (for the weighted sum) for the pixels in the yellow-coloured areas. Large positive values in the greyscale correspond to white colour. Therefore, the areas which were yellow in the colour picture converted to white in the greyscale. In practice, we do not manually assign weights, instead, we let the neural network decide the optimal weights during training.
The neuron or its weights is called a filter. We convolve the image with a filter, i.e. a convolutional filter.
When a filter’s footprint is > 1 pixel, it is a spatial filter.
The above spatial filter is a 3x3 filter. Hence, there are 9 weights to learn.
In a multidimensional filter, the number of channels of the input must be equal to the number of channels in the filter (depths must be the same).
Need \# \text{ Channels in Input} = \# \text{ Channels in Filter}.
The above figure shows how we pick a 3x3x3 block from the image, and then apply the 3x3 filter. The multiplication is carried out channel-wise, i.e. we select the first channel of the filter and the first channel of the image and carry out the element wise multiplation. Once the elementwise multiplications for the three pairs of channels are completed, we sum them all, and pass through the neuron.
The above figure shows how 9 inputs transform in to one output. As a result, dimensions of the output matrix is smaller than the dimensions of the input matrix. There are some options we can use if we wish to keep the size of input and output matrices same.
Add a border of extra elements around the input, called padding. Normally we place zeros in all the new elements, called zero padding.
The motivation behind applying filters simultaneously and independently is to let the filters learn different patterns in the input-output relationship. The idea is quite similar to using many neurons in one Dense layer (in a Dense layer, we would use multiple neurons so that different neurons can capture different patterns in the input-output relationship).
In the image:
The above picture shows how we take in an image with 6 channels, select a 3x3 block (in pink colour), apply 4 different filters of same dimensions (in pink, green, blue and yellow), retrieve the output with 1 channel (1 output for each filter) and finally stack them together to create 1 output tensor. Note that the number of channels in the output tensor is 4, which is equal to the number of spatial filters used.
Striding options allows to modify the movement of the filter across the image. Instead moving one step at a time (either horizontally or vertically), we can increase the number of steps using the striding option.
We don’t have to go one pixel across/down at a time.
Dimension of output will be smaller than input.
When a filter scans the input step by step, it processes the same input elements multiple times. Even with larger strides, this can still happen (left image).
If we want to save time, we can choose strides that prevents input elements from being used more than once. Example (right image): 3x3 filter, stride 3 in both directions.
Need to choose:
All the filter weights are learned during training.
A neural network that uses convolution layers is called a convolutional neural network.
A standard CNN architecture has the following components: an input layer, a sequence of feature extraction layers (which combine convolution and pooling operations sequentially), a sequence of classification layers (which include flattening and fully connected layers) and a final output layer. Convolution layers are used to extract meaningful patterns from the input using spatial filters. Pooling layers are used to reduce the spatial dimensions of the feature maps generated from convolutional layers. The purpose of the feature extraction layers is to learn complex but meaningful, high levels patterns in data. The aim of classification layers is to receive the learned patterns and make decisions more closely related to the classification task at hand.
On a high level, the idea would be to keep on increasing the number of channels (depth) and decrease the dimensions of the feature map. We can see how the depth increases, and spatial dimensions reduce from first convolution layer to the second pooling layer.
Pooling, or downsampling, is a technique to blur a tensor.
(a): Input tensor (b): Subdivide input tensor into 2x2 blocks (c): Average pooling (d): Max pooling (e): Icon for a pooling layer
Why? Pooling reduces the size of tensors, therefore reduces memory usage and execution time (recall that 1x1 convolution reduces the number of channels in a tensor).
Why not?
57 poorly written Mandarin characters (57 \times 7 = 399).
The data is zipped (6.9 MB) and stored on my GitHub homepage.
# Download the dataset if it hasn't already been downloaded.
from pathlib import Path
if not Path("mandarin").exists():
print("Downloading dataset...")
!wget https://laub.au/data/mandarin.zip
!unzip mandarin.zip
else:
print("Already downloaded.")Already downloaded.Remember, the Jupyter notebook associated with your final report should either download your dataset when it is run, or you should supply the data separately.
Inspect directory structure
!pip install directory_treefrom directory_tree import display_tree
display_tree("mandarin")mandarin/
├── bai/
│ ├── bai-1.png
│ ├── bai-2.png
│ ├── bai-3.png
│ ├── bai-4.png
│ ├── bai-5.png
│ ├── bai-6.png
│ └── bai-7.png
├── ben/
│ ├── ben-1.png
│ ├── ben-2.png
│ ├── ben-3.png
│ ├── ben-4.png
│ ├── ben-5.png
│ ├── ben-6.png
│ └── ben-7.png
├── chong/
│ ├── chong-1.png
│ ├── chong-2.png
│ ├── chong-3.png
│ ├── chong-4.png
│ ├── chong-5.png
│ ├── chong-6.png
│ └── chong-7.png
├── chu/
│ ├── chu-1.png
│ ├── chu-2.png
│ ├── chu-3.png
│ ├── chu-4.png
│ ├── chu-5.png
│ ├── chu-6.png
│ └── chu-7.png
├── chuan/
│ ├── chuan-1.png
│ ├── chuan-2.png
│ ├── chuan-3.png
│ ├── chuan-4.png
│ ├── chuan-5.png
│ ├── chuan-6.png
│ └── chuan-7.png
├── cong/
│ ├── cong-1.png
│ ├── cong-2.png
│ ├── cong-3.png
│ ├── cong-4.png
│ ├── cong-5.png
│ ├── cong-6.png
│ └── cong-7.png
├── da/
│ ├── da-1.png
│ ├── da-2.png
│ ├── da-3.png
│ ├── da-4.png
│ ├── da-5.png
│ ├── da-6.png
│ └── da-7.png
├── dan/
│ ├── dan-1.png
│ ├── dan-2.png
│ ├── dan-3.png
│ ├── dan-4.png
│ ├── dan-5.png
│ ├── dan-6.png
│ └── dan-7.png
├── dong/
│ ├── dong-1.png
│ ├── dong-2.png
│ ├── dong-3.png
│ ├── dong-4.png
│ ├── dong-5.png
│ ├── dong-6.png
│ └── dong-7.png
├── fei/
│ ├── fei-1.png
│ ├── fei-2.png
│ ├── fei-3.png
│ ├── fei-4.png
│ ├── fei-5.png
│ ├── fei-6.png
│ └── fei-7.png
├── fu/
│ ├── fu-1.png
│ ├── fu-2.png
│ ├── fu-3.png
│ ├── fu-4.png
│ ├── fu-5.png
│ ├── fu-6.png
│ └── fu-7.png
├── fu2/
│ ├── fu2-1.png
│ ├── fu2-2.png
│ ├── fu2-3.png
│ ├── fu2-4.png
│ ├── fu2-5.png
│ ├── fu2-6.png
│ └── fu2-7.png
├── gao/
│ ├── gao-1.png
│ ├── gao-2.png
│ ├── gao-3.png
│ ├── gao-4.png
│ ├── gao-5.png
│ ├── gao-6.png
│ └── gao-7.png
├── gong/
│ ├── gong-1.png
│ ├── gong-2.png
│ ├── gong-3.png
│ ├── gong-4.png
│ ├── gong-5.png
│ ├── gong-6.png
│ └── gong-7.png
├── guo/
│ ├── guo-1.png
│ ├── guo-2.png
│ ├── guo-3.png
│ ├── guo-4.png
│ ├── guo-5.png
│ ├── guo-6.png
│ └── guo-7.png
├── hu/
│ ├── hu-1.png
│ ├── hu-2.png
│ ├── hu-3.png
│ ├── hu-4.png
│ ├── hu-5.png
│ ├── hu-6.png
│ └── hu-7.png
├── huo/
│ ├── huo-1.png
│ ├── huo-2.png
│ ├── huo-3.png
│ ├── huo-4.png
│ ├── huo-5.png
│ ├── huo-6.png
│ └── huo-7.png
├── kou/
│ ├── kou-1.png
│ ├── kou-2.png
│ ├── kou-3.png
│ ├── kou-4.png
│ ├── kou-5.png
│ ├── kou-6.png
│ └── kou-7.png
├── ku/
│ ├── ku-1.png
│ ├── ku-2.png
│ ├── ku-3.png
│ ├── ku-4.png
│ ├── ku-5.png
│ ├── ku-6.png
│ └── ku-7.png
├── lin/
│ ├── lin-1.png
│ ├── lin-2.png
│ ├── lin-3.png
│ ├── lin-4.png
│ ├── lin-5.png
│ ├── lin-6.png
│ └── lin-7.png
├── ma/
│ ├── ma-1.png
│ ├── ma-2.png
│ ├── ma-3.png
│ ├── ma-4.png
│ ├── ma-5.png
│ ├── ma-6.png
│ └── ma-7.png
├── ma2/
│ ├── ma2-1.png
│ ├── ma2-2.png
│ ├── ma2-3.png
│ ├── ma2-4.png
│ ├── ma2-5.png
│ ├── ma2-6.png
│ └── ma2-7.png
├── ma3/
│ ├── ma3-1.png
│ ├── ma3-2.png
│ ├── ma3-3.png
│ ├── ma3-4.png
│ ├── ma3-5.png
│ ├── ma3-6.png
│ └── ma3-7.png
├── mei/
│ ├── mei-1.png
│ ├── mei-2.png
│ ├── mei-3.png
│ ├── mei-4.png
│ ├── mei-5.png
│ ├── mei-6.png
│ └── mei-7.png
├── men/
│ ├── men-1.png
│ ├── men-2.png
│ ├── men-3.png
│ ├── men-4.png
│ ├── men-5.png
│ ├── men-6.png
│ └── men-7.png
├── ming/
│ ├── ming-1.png
│ ├── ming-2.png
│ ├── ming-3.png
│ ├── ming-4.png
│ ├── ming-5.png
│ ├── ming-6.png
│ └── ming-7.png
├── mu/
│ ├── mu-1.png
│ ├── mu-2.png
│ ├── mu-3.png
│ ├── mu-4.png
│ ├── mu-5.png
│ ├── mu-6.png
│ └── mu-7.png
├── nan/
│ ├── nan-1.png
│ ├── nan-2.png
│ ├── nan-3.png
│ ├── nan-4.png
│ ├── nan-5.png
│ ├── nan-6.png
│ └── nan-7.png
├── niao/
│ ├── niao-1.png
│ ├── niao-2.png
│ ├── niao-3.png
│ ├── niao-4.png
│ ├── niao-5.png
│ ├── niao-6.png
│ └── niao-7.png
├── niu/
│ ├── niu-1.png
│ ├── niu-2.png
│ ├── niu-3.png
│ ├── niu-4.png
│ ├── niu-5.png
│ ├── niu-6.png
│ └── niu-7.png
├── nu/
│ ├── nu-1.png
│ ├── nu-2.png
│ ├── nu-3.png
│ ├── nu-4.png
│ ├── nu-5.png
│ ├── nu-6.png
│ └── nu-7.png
├── nuan/
│ ├── nuan-1.png
│ ├── nuan-2.png
│ ├── nuan-3.png
│ ├── nuan-4.png
│ ├── nuan-5.png
│ ├── nuan-6.png
│ └── nuan-7.png
├── peng/
│ ├── peng-1.png
│ ├── peng-2.png
│ ├── peng-3.png
│ ├── peng-4.png
│ ├── peng-5.png
│ ├── peng-6.png
│ └── peng-7.png
├── quan/
│ ├── quan-1.png
│ ├── quan-2.png
│ ├── quan-3.png
│ ├── quan-4.png
│ ├── quan-5.png
│ ├── quan-6.png
│ └── quan-7.png
├── ren/
│ ├── ren-1.png
│ ├── ren-2.png
│ ├── ren-3.png
│ ├── ren-4.png
│ ├── ren-5.png
│ ├── ren-6.png
│ └── ren-7.png
├── ri/
│ ├── ri-1.png
│ ├── ri-2.png
│ ├── ri-3.png
│ ├── ri-4.png
│ ├── ri-5.png
│ ├── ri-6.png
│ └── ri-7.png
├── rou/
│ ├── rou-1.png
│ ├── rou-2.png
│ ├── rou-3.png
│ ├── rou-4.png
│ ├── rou-5.png
│ ├── rou-6.png
│ └── rou-7.png
├── sen/
│ ├── sen-1.png
│ ├── sen-2.png
│ ├── sen-3.png
│ ├── sen-4.png
│ ├── sen-5.png
│ ├── sen-6.png
│ └── sen-7.png
├── shan/
│ ├── shan-1.png
│ ├── shan-2.png
│ ├── shan-3.png
│ ├── shan-4.png
│ ├── shan-5.png
│ ├── shan-6.png
│ └── shan-7.png
├── shan2/
│ ├── shan2-1.png
│ ├── shan2-2.png
│ ├── shan2-3.png
│ ├── shan2-4.png
│ ├── shan2-5.png
│ ├── shan2-6.png
│ └── shan2-7.png
├── shui/
│ ├── shui-1.png
│ ├── shui-2.png
│ ├── shui-3.png
│ ├── shui-4.png
│ ├── shui-5.png
│ ├── shui-6.png
│ └── shui-7.png
├── tai/
│ ├── tai-1.png
│ ├── tai-2.png
│ ├── tai-3.png
│ ├── tai-4.png
│ ├── tai-5.png
│ ├── tai-6.png
│ └── tai-7.png
├── tian/
│ ├── tian-1.png
│ ├── tian-2.png
│ ├── tian-3.png
│ ├── tian-4.png
│ ├── tian-5.png
│ ├── tian-6.png
│ └── tian-7.png
├── wang/
│ ├── wang-1.png
│ ├── wang-2.png
│ ├── wang-3.png
│ ├── wang-4.png
│ ├── wang-5.png
│ ├── wang-6.png
│ └── wang-7.png
├── wen/
│ ├── wen-1.png
│ ├── wen-2.png
│ ├── wen-3.png
│ ├── wen-4.png
│ ├── wen-5.png
│ ├── wen-6.png
│ └── wen-7.png
├── xian/
│ ├── xian-1.png
│ ├── xian-2.png
│ ├── xian-3.png
│ ├── xian-4.png
│ ├── xian-5.png
│ ├── xian-6.png
│ └── xian-7.png
├── xuan/
│ ├── xuan-1.png
│ ├── xuan-2.png
│ ├── xuan-3.png
│ ├── xuan-4.png
│ ├── xuan-5.png
│ ├── xuan-6.png
│ └── xuan-7.png
├── yan/
│ ├── yan-1.png
│ ├── yan-2.png
│ ├── yan-3.png
│ ├── yan-4.png
│ ├── yan-5.png
│ ├── yan-6.png
│ └── yan-7.png
├── yang/
│ ├── yang-1.png
│ ├── yang-2.png
│ ├── yang-3.png
│ ├── yang-4.png
│ ├── yang-5.png
│ ├── yang-6.png
│ └── yang-7.png
├── yin/
│ ├── yin-1.png
│ ├── yin-2.png
│ ├── yin-3.png
│ ├── yin-4.png
│ ├── yin-5.png
│ ├── yin-6.png
│ └── yin-7.png
├── yu/
│ ├── yu-1.png
│ ├── yu-2.png
│ ├── yu-3.png
│ ├── yu-4.png
│ ├── yu-5.png
│ ├── yu-6.png
│ └── yu-7.png
├── yu2/
│ ├── yu2-1.png
│ ├── yu2-2.png
│ ├── yu2-3.png
│ ├── yu2-4.png
│ ├── yu2-5.png
│ ├── yu2-6.png
│ └── yu2-7.png
├── yue/
│ ├── yue-1.png
│ ├── yue-2.png
│ ├── yue-3.png
│ ├── yue-4.png
│ ├── yue-5.png
│ ├── yue-6.png
│ └── yue-7.png
├── zhong/
│ ├── zhong-1.png
│ ├── zhong-2.png
│ ├── zhong-3.png
│ ├── zhong-4.png
│ ├── zhong-5.png
│ ├── zhong-6.png
│ └── zhong-7.png
├── zhu/
│ ├── zhu-1.png
│ ├── zhu-2.png
│ ├── zhu-3.png
│ ├── zhu-4.png
│ ├── zhu-5.png
│ ├── zhu-6.png
│ └── zhu-7.png
├── zhu2/
│ ├── zhu2-1.png
│ ├── zhu2-2.png
│ ├── zhu2-3.png
│ ├── zhu2-4.png
│ ├── zhu2-5.png
│ ├── zhu2-6.png
│ └── zhu2-7.png
└── zhuo/
├── zhuo-1.png
├── zhuo-2.png
├── zhuo-3.png
├── zhuo-4.png
├── zhuo-5.png
├── zhuo-6.png
└── zhuo-7.pngtree = display_tree("mandarin", string_rep=True).split("\n")
print("\n".join(tree[:12]))
print("...")
print("\n".join(tree[-4:]))mandarin/
├── bai/
│ ├── bai-1.png
│ ├── bai-2.png
│ ├── bai-3.png
│ ├── bai-4.png
│ ├── bai-5.png
│ ├── bai-6.png
│ └── bai-7.png
├── ben/
│ ├── ben-1.png
│ ├── ben-2.png
...
├── zhuo-5.png
├── zhuo-6.png
└── zhuo-7.png
!pip install split-foldersimport splitfolders
splitfolders.ratio("mandarin", output="mandarin-split",
seed=1337, ratio=(5/7, 1/7, 1/7))
display_tree("mandarin-split", max_depth=1)Copying files: 0 files [00:00, ? files/s]Copying files: 238 files [00:00, 2378.45 files/s]Copying files: 399 files [00:00, 1944.80 files/s]mandarin-split/
├── test/
├── train/
└── val/display_tree("mandarin-split")mandarin-split/
├── test/
│ ├── bai/
│ │ └── bai-5.png
│ ├── ben/
│ │ └── ben-5.png
│ ├── chong/
│ │ └── chong-5.png
│ ├── chu/
│ │ └── chu-5.png
│ ├── chuan/
│ │ └── chuan-5.png
│ ├── cong/
│ │ └── cong-5.png
│ ├── da/
│ │ └── da-5.png
│ ├── dan/
│ │ └── dan-5.png
│ ├── dong/
│ │ └── dong-5.png
│ ├── fei/
│ │ └── fei-5.png
│ ├── fu/
│ │ └── fu-5.png
│ ├── fu2/
│ │ └── fu2-5.png
│ ├── gao/
│ │ └── gao-5.png
│ ├── gong/
│ │ └── gong-5.png
│ ├── guo/
│ │ └── guo-5.png
│ ├── hu/
│ │ └── hu-5.png
│ ├── huo/
│ │ └── huo-5.png
│ ├── kou/
│ │ └── kou-5.png
│ ├── ku/
│ │ └── ku-5.png
│ ├── lin/
│ │ └── lin-5.png
│ ├── ma/
│ │ └── ma-5.png
│ ├── ma2/
│ │ └── ma2-5.png
│ ├── ma3/
│ │ └── ma3-5.png
│ ├── mei/
│ │ └── mei-5.png
│ ├── men/
│ │ └── men-5.png
│ ├── ming/
│ │ └── ming-5.png
│ ├── mu/
│ │ └── mu-5.png
│ ├── nan/
│ │ └── nan-5.png
│ ├── niao/
│ │ └── niao-5.png
│ ├── niu/
│ │ └── niu-5.png
│ ├── nu/
│ │ └── nu-5.png
│ ├── nuan/
│ │ └── nuan-5.png
│ ├── peng/
│ │ └── peng-5.png
│ ├── quan/
│ │ └── quan-5.png
│ ├── ren/
│ │ └── ren-5.png
│ ├── ri/
│ │ └── ri-5.png
│ ├── rou/
│ │ └── rou-5.png
│ ├── sen/
│ │ └── sen-5.png
│ ├── shan/
│ │ └── shan-5.png
│ ├── shan2/
│ │ └── shan2-5.png
│ ├── shui/
│ │ └── shui-5.png
│ ├── tai/
│ │ └── tai-5.png
│ ├── tian/
│ │ └── tian-5.png
│ ├── wang/
│ │ └── wang-5.png
│ ├── wen/
│ │ └── wen-5.png
│ ├── xian/
│ │ └── xian-5.png
│ ├── xuan/
│ │ └── xuan-5.png
│ ├── yan/
│ │ └── yan-5.png
│ ├── yang/
│ │ └── yang-5.png
│ ├── yin/
│ │ └── yin-5.png
│ ├── yu/
│ │ └── yu-5.png
│ ├── yu2/
│ │ └── yu2-5.png
│ ├── yue/
│ │ └── yue-5.png
│ ├── zhong/
│ │ └── zhong-5.png
│ ├── zhu/
│ │ └── zhu-5.png
│ ├── zhu2/
│ │ └── zhu2-5.png
│ └── zhuo/
│ └── zhuo-5.png
├── train/
│ ├── bai/
│ │ ├── bai-1.png
│ │ ├── bai-2.png
│ │ ├── bai-3.png
│ │ ├── bai-4.png
│ │ └── bai-6.png
│ ├── ben/
│ │ ├── ben-1.png
│ │ ├── ben-2.png
│ │ ├── ben-3.png
│ │ ├── ben-4.png
│ │ └── ben-6.png
│ ├── chong/
│ │ ├── chong-1.png
│ │ ├── chong-2.png
│ │ ├── chong-3.png
│ │ ├── chong-4.png
│ │ └── chong-6.png
│ ├── chu/
│ │ ├── chu-1.png
│ │ ├── chu-2.png
│ │ ├── chu-3.png
│ │ ├── chu-4.png
│ │ └── chu-6.png
│ ├── chuan/
│ │ ├── chuan-1.png
│ │ ├── chuan-2.png
│ │ ├── chuan-3.png
│ │ ├── chuan-4.png
│ │ └── chuan-6.png
│ ├── cong/
│ │ ├── cong-1.png
│ │ ├── cong-2.png
│ │ ├── cong-3.png
│ │ ├── cong-4.png
│ │ └── cong-6.png
│ ├── da/
│ │ ├── da-1.png
│ │ ├── da-2.png
│ │ ├── da-3.png
│ │ ├── da-4.png
│ │ └── da-6.png
│ ├── dan/
│ │ ├── dan-1.png
│ │ ├── dan-2.png
│ │ ├── dan-3.png
│ │ ├── dan-4.png
│ │ └── dan-6.png
│ ├── dong/
│ │ ├── dong-1.png
│ │ ├── dong-2.png
│ │ ├── dong-3.png
│ │ ├── dong-4.png
│ │ └── dong-6.png
│ ├── fei/
│ │ ├── fei-1.png
│ │ ├── fei-2.png
│ │ ├── fei-3.png
│ │ ├── fei-4.png
│ │ └── fei-6.png
│ ├── fu/
│ │ ├── fu-1.png
│ │ ├── fu-2.png
│ │ ├── fu-3.png
│ │ ├── fu-4.png
│ │ └── fu-6.png
│ ├── fu2/
│ │ ├── fu2-1.png
│ │ ├── fu2-2.png
│ │ ├── fu2-3.png
│ │ ├── fu2-4.png
│ │ └── fu2-6.png
│ ├── gao/
│ │ ├── gao-1.png
│ │ ├── gao-2.png
│ │ ├── gao-3.png
│ │ ├── gao-4.png
│ │ └── gao-6.png
│ ├── gong/
│ │ ├── gong-1.png
│ │ ├── gong-2.png
│ │ ├── gong-3.png
│ │ ├── gong-4.png
│ │ └── gong-6.png
│ ├── guo/
│ │ ├── guo-1.png
│ │ ├── guo-2.png
│ │ ├── guo-3.png
│ │ ├── guo-4.png
│ │ └── guo-6.png
│ ├── hu/
│ │ ├── hu-1.png
│ │ ├── hu-2.png
│ │ ├── hu-3.png
│ │ ├── hu-4.png
│ │ └── hu-6.png
│ ├── huo/
│ │ ├── huo-1.png
│ │ ├── huo-2.png
│ │ ├── huo-3.png
│ │ ├── huo-4.png
│ │ └── huo-6.png
│ ├── kou/
│ │ ├── kou-1.png
│ │ ├── kou-2.png
│ │ ├── kou-3.png
│ │ ├── kou-4.png
│ │ └── kou-6.png
│ ├── ku/
│ │ ├── ku-1.png
│ │ ├── ku-2.png
│ │ ├── ku-3.png
│ │ ├── ku-4.png
│ │ └── ku-6.png
│ ├── lin/
│ │ ├── lin-1.png
│ │ ├── lin-2.png
│ │ ├── lin-3.png
│ │ ├── lin-4.png
│ │ └── lin-6.png
│ ├── ma/
│ │ ├── ma-1.png
│ │ ├── ma-2.png
│ │ ├── ma-3.png
│ │ ├── ma-4.png
│ │ └── ma-6.png
│ ├── ma2/
│ │ ├── ma2-1.png
│ │ ├── ma2-2.png
│ │ ├── ma2-3.png
│ │ ├── ma2-4.png
│ │ └── ma2-6.png
│ ├── ma3/
│ │ ├── ma3-1.png
│ │ ├── ma3-2.png
│ │ ├── ma3-3.png
│ │ ├── ma3-4.png
│ │ └── ma3-6.png
│ ├── mei/
│ │ ├── mei-1.png
│ │ ├── mei-2.png
│ │ ├── mei-3.png
│ │ ├── mei-4.png
│ │ └── mei-6.png
│ ├── men/
│ │ ├── men-1.png
│ │ ├── men-2.png
│ │ ├── men-3.png
│ │ ├── men-4.png
│ │ └── men-6.png
│ ├── ming/
│ │ ├── ming-1.png
│ │ ├── ming-2.png
│ │ ├── ming-3.png
│ │ ├── ming-4.png
│ │ └── ming-6.png
│ ├── mu/
│ │ ├── mu-1.png
│ │ ├── mu-2.png
│ │ ├── mu-3.png
│ │ ├── mu-4.png
│ │ └── mu-6.png
│ ├── nan/
│ │ ├── nan-1.png
│ │ ├── nan-2.png
│ │ ├── nan-3.png
│ │ ├── nan-4.png
│ │ └── nan-6.png
│ ├── niao/
│ │ ├── niao-1.png
│ │ ├── niao-2.png
│ │ ├── niao-3.png
│ │ ├── niao-4.png
│ │ └── niao-6.png
│ ├── niu/
│ │ ├── niu-1.png
│ │ ├── niu-2.png
│ │ ├── niu-3.png
│ │ ├── niu-4.png
│ │ └── niu-6.png
│ ├── nu/
│ │ ├── nu-1.png
│ │ ├── nu-2.png
│ │ ├── nu-3.png
│ │ ├── nu-4.png
│ │ └── nu-6.png
│ ├── nuan/
│ │ ├── nuan-1.png
│ │ ├── nuan-2.png
│ │ ├── nuan-3.png
│ │ ├── nuan-4.png
│ │ └── nuan-6.png
│ ├── peng/
│ │ ├── peng-1.png
│ │ ├── peng-2.png
│ │ ├── peng-3.png
│ │ ├── peng-4.png
│ │ └── peng-6.png
│ ├── quan/
│ │ ├── quan-1.png
│ │ ├── quan-2.png
│ │ ├── quan-3.png
│ │ ├── quan-4.png
│ │ └── quan-6.png
│ ├── ren/
│ │ ├── ren-1.png
│ │ ├── ren-2.png
│ │ ├── ren-3.png
│ │ ├── ren-4.png
│ │ └── ren-6.png
│ ├── ri/
│ │ ├── ri-1.png
│ │ ├── ri-2.png
│ │ ├── ri-3.png
│ │ ├── ri-4.png
│ │ └── ri-6.png
│ ├── rou/
│ │ ├── rou-1.png
│ │ ├── rou-2.png
│ │ ├── rou-3.png
│ │ ├── rou-4.png
│ │ └── rou-6.png
│ ├── sen/
│ │ ├── sen-1.png
│ │ ├── sen-2.png
│ │ ├── sen-3.png
│ │ ├── sen-4.png
│ │ └── sen-6.png
│ ├── shan/
│ │ ├── shan-1.png
│ │ ├── shan-2.png
│ │ ├── shan-3.png
│ │ ├── shan-4.png
│ │ └── shan-6.png
│ ├── shan2/
│ │ ├── shan2-1.png
│ │ ├── shan2-2.png
│ │ ├── shan2-3.png
│ │ ├── shan2-4.png
│ │ └── shan2-6.png
│ ├── shui/
│ │ ├── shui-1.png
│ │ ├── shui-2.png
│ │ ├── shui-3.png
│ │ ├── shui-4.png
│ │ └── shui-6.png
│ ├── tai/
│ │ ├── tai-1.png
│ │ ├── tai-2.png
│ │ ├── tai-3.png
│ │ ├── tai-4.png
│ │ └── tai-6.png
│ ├── tian/
│ │ ├── tian-1.png
│ │ ├── tian-2.png
│ │ ├── tian-3.png
│ │ ├── tian-4.png
│ │ └── tian-6.png
│ ├── wang/
│ │ ├── wang-1.png
│ │ ├── wang-2.png
│ │ ├── wang-3.png
│ │ ├── wang-4.png
│ │ └── wang-6.png
│ ├── wen/
│ │ ├── wen-1.png
│ │ ├── wen-2.png
│ │ ├── wen-3.png
│ │ ├── wen-4.png
│ │ └── wen-6.png
│ ├── xian/
│ │ ├── xian-1.png
│ │ ├── xian-2.png
│ │ ├── xian-3.png
│ │ ├── xian-4.png
│ │ └── xian-6.png
│ ├── xuan/
│ │ ├── xuan-1.png
│ │ ├── xuan-2.png
│ │ ├── xuan-3.png
│ │ ├── xuan-4.png
│ │ └── xuan-6.png
│ ├── yan/
│ │ ├── yan-1.png
│ │ ├── yan-2.png
│ │ ├── yan-3.png
│ │ ├── yan-4.png
│ │ └── yan-6.png
│ ├── yang/
│ │ ├── yang-1.png
│ │ ├── yang-2.png
│ │ ├── yang-3.png
│ │ ├── yang-4.png
│ │ └── yang-6.png
│ ├── yin/
│ │ ├── yin-1.png
│ │ ├── yin-2.png
│ │ ├── yin-3.png
│ │ ├── yin-4.png
│ │ └── yin-6.png
│ ├── yu/
│ │ ├── yu-1.png
│ │ ├── yu-2.png
│ │ ├── yu-3.png
│ │ ├── yu-4.png
│ │ └── yu-6.png
│ ├── yu2/
│ │ ├── yu2-1.png
│ │ ├── yu2-2.png
│ │ ├── yu2-3.png
│ │ ├── yu2-4.png
│ │ └── yu2-6.png
│ ├── yue/
│ │ ├── yue-1.png
│ │ ├── yue-2.png
│ │ ├── yue-3.png
│ │ ├── yue-4.png
│ │ └── yue-6.png
│ ├── zhong/
│ │ ├── zhong-1.png
│ │ ├── zhong-2.png
│ │ ├── zhong-3.png
│ │ ├── zhong-4.png
│ │ └── zhong-6.png
│ ├── zhu/
│ │ ├── zhu-1.png
│ │ ├── zhu-2.png
│ │ ├── zhu-3.png
│ │ ├── zhu-4.png
│ │ └── zhu-6.png
│ ├── zhu2/
│ │ ├── zhu2-1.png
│ │ ├── zhu2-2.png
│ │ ├── zhu2-3.png
│ │ ├── zhu2-4.png
│ │ └── zhu2-6.png
│ └── zhuo/
│ ├── zhuo-1.png
│ ├── zhuo-2.png
│ ├── zhuo-3.png
│ ├── zhuo-4.png
│ └── zhuo-6.png
└── val/
├── bai/
│ └── bai-7.png
├── ben/
│ └── ben-7.png
├── chong/
│ └── chong-7.png
├── chu/
│ └── chu-7.png
├── chuan/
│ └── chuan-7.png
├── cong/
│ └── cong-7.png
├── da/
│ └── da-7.png
├── dan/
│ └── dan-7.png
├── dong/
│ └── dong-7.png
├── fei/
│ └── fei-7.png
├── fu/
│ └── fu-7.png
├── fu2/
│ └── fu2-7.png
├── gao/
│ └── gao-7.png
├── gong/
│ └── gong-7.png
├── guo/
│ └── guo-7.png
├── hu/
│ └── hu-7.png
├── huo/
│ └── huo-7.png
├── kou/
│ └── kou-7.png
├── ku/
│ └── ku-7.png
├── lin/
│ └── lin-7.png
├── ma/
│ └── ma-7.png
├── ma2/
│ └── ma2-7.png
├── ma3/
│ └── ma3-7.png
├── mei/
│ └── mei-7.png
├── men/
│ └── men-7.png
├── ming/
│ └── ming-7.png
├── mu/
│ └── mu-7.png
├── nan/
│ └── nan-7.png
├── niao/
│ └── niao-7.png
├── niu/
│ └── niu-7.png
├── nu/
│ └── nu-7.png
├── nuan/
│ └── nuan-7.png
├── peng/
│ └── peng-7.png
├── quan/
│ └── quan-7.png
├── ren/
│ └── ren-7.png
├── ri/
│ └── ri-7.png
├── rou/
│ └── rou-7.png
├── sen/
│ └── sen-7.png
├── shan/
│ └── shan-7.png
├── shan2/
│ └── shan2-7.png
├── shui/
│ └── shui-7.png
├── tai/
│ └── tai-7.png
├── tian/
│ └── tian-7.png
├── wang/
│ └── wang-7.png
├── wen/
│ └── wen-7.png
├── xian/
│ └── xian-7.png
├── xuan/
│ └── xuan-7.png
├── yan/
│ └── yan-7.png
├── yang/
│ └── yang-7.png
├── yin/
│ └── yin-7.png
├── yu/
│ └── yu-7.png
├── yu2/
│ └── yu2-7.png
├── yue/
│ └── yue-7.png
├── zhong/
│ └── zhong-7.png
├── zhu/
│ └── zhu-7.png
├── zhu2/
│ └── zhu2-7.png
└── zhuo/
└── zhuo-7.pngtrain/
├── bai/
│ ├── bai-1.png
│ ├── bai-2.png
│ ├── bai-3.png
│ ├── bai-4.png
│ └── bai-6.png
...
val/
├── bai/
│ └── bai-7.png
├── ben/
│ └── ben-7.png
...
test/
├── bai/
│ └── bai-5.png
├── ben/
│ └── ben-5.png
...from keras.utils import\
1 image_dataset_from_directory
2data_dir = "mandarin-split"
3batch_size = 32
4img_height = 80
5img_width = 80
6img_size = (img_height, img_width)
7train_ds = image_dataset_from_directory(
data_dir + "/train",
image_size=img_size,
batch_size=batch_size,
shuffle=False,
color_mode='grayscale')
8val_ds = image_dataset_from_directory(
data_dir + "/val",
image_size=img_size,
batch_size=batch_size,
shuffle=False,
color_mode='grayscale')
9test_ds = image_dataset_from_directory(
data_dir + "/test",
image_size=img_size,
batch_size=batch_size,
shuffle=False,
color_mode='grayscale')image_dataset_from_directory class from keras.utils library
color_mode='grayscale' command tells the computer to bring in the images in greyscale instead of the RGB scale.
Found 285 files belonging to 57 classes.
Found 57 files belonging to 57 classes.
Found 57 files belonging to 57 classes.print(train_ds.class_names)['bai', 'ben', 'chong', 'chu', 'chuan', 'cong', 'da', 'dan', 'dong', 'fei', 'fu', 'fu2', 'gao', 'gong', 'guo', 'hu', 'huo', 'kou', 'ku', 'lin', 'ma', 'ma2', 'ma3', 'mei', 'men', 'ming', 'mu', 'nan', 'niao', 'niu', 'nu', 'nuan', 'peng', 'quan', 'ren', 'ri', 'rou', 'sen', 'shan', 'shan2', 'shui', 'tai', 'tian', 'wang', 'wen', 'xian', 'xuan', 'yan', 'yang', 'yin', 'yu', 'yu2', 'yue', 'zhong', 'zhu', 'zhu2', 'zhuo']# NB: Need shuffle=False earlier for these X & y to line up.
X_train = np.concatenate(list(train_ds.map(lambda x, y: x)))
y_train = np.concatenate(list(train_ds.map(lambda x, y: y)))
X_val = np.concatenate(list(val_ds.map(lambda x, y: x)))
y_val = np.concatenate(list(val_ds.map(lambda x, y: y)))
X_test = np.concatenate(list(test_ds.map(lambda x, y: x)))
y_test = np.concatenate(list(test_ds.map(lambda x, y: y)))
X_train.shape, y_train.shape, X_val.shape, y_val.shape, X_test.shape, y_test.shape((285, 80, 80, 1), (285,), (57, 80, 80, 1), (57,), (57, 80, 80, 1), (57,))def plot_mandarin_characters(ds, plot_char_label = 0):
num_plotted = 0
for images, labels in ds:
for i in range(images.shape[0]):
label = labels[i]
if label == plot_char_label:
plt.subplot(1, 5, num_plotted + 1)
plt.imshow(images[i].numpy().astype("uint8"), cmap="gray")
plt.title(ds.class_names[label])
plt.axis("off")
num_plotted += 1
plt.show()
bai_val = X_val[y_val == 0][0]
plt.imshow(bai_val, cmap="gray");
bai_test = X_test[y_test == 0][0]
plt.imshow(bai_test);
from keras.layers \
1 import Rescaling, Conv2D, MaxPooling2D, Flatten
2num_classes = np.unique(y_train).shape[0]
random.seed(123)
model = Sequential([
Input((img_height, img_width, 1)),
3 Rescaling(1./255),
4 Conv2D(16, 3, padding="same", activation="relu", name="conv1"),
5 MaxPooling2D(name="pool1"),
Conv2D(32, 3, padding="same", activation="relu", name="conv2"),
MaxPooling2D(name="pool2"),
Conv2D(64, 3, padding="same", activation="relu", name="conv3"),
MaxPooling2D(name="pool3"),
6 Flatten(), Dense(128, activation="relu"), Dense(num_classes)
])keras.layers
padding="same" ensures that the dimensions of the input and output matrices remain same
MaxPooling, which reduces the spatial dimensions by carrying forward the maximum value over an input window
Flatten layer to convert the 2D array (from pooling) in to a single column vector, and passes through couple of Dense layers to train the neural network for the specific classification problem. Note that the output layer has number of neurons equal to numClasses, which corresponds to the number of unique classes in the output.
The Rescaling layer will rescale the intensities to [0, 1].
model.summary()Model: "sequential"
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━┓ ┃ Layer (type) ┃ Output Shape ┃ Param # ┃ ┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━┩ │ rescaling (Rescaling) │ (None, 80, 80, 1) │ 0 │ ├─────────────────────────────────┼────────────────────────┼───────────────┤ │ conv1 (Conv2D) │ (None, 80, 80, 16) │ 160 │ ├─────────────────────────────────┼────────────────────────┼───────────────┤ │ pool1 (MaxPooling2D) │ (None, 40, 40, 16) │ 0 │ ├─────────────────────────────────┼────────────────────────┼───────────────┤ │ conv2 (Conv2D) │ (None, 40, 40, 32) │ 4,640 │ ├─────────────────────────────────┼────────────────────────┼───────────────┤ │ pool2 (MaxPooling2D) │ (None, 20, 20, 32) │ 0 │ ├─────────────────────────────────┼────────────────────────┼───────────────┤ │ conv3 (Conv2D) │ (None, 20, 20, 64) │ 18,496 │ ├─────────────────────────────────┼────────────────────────┼───────────────┤ │ pool3 (MaxPooling2D) │ (None, 10, 10, 64) │ 0 │ ├─────────────────────────────────┼────────────────────────┼───────────────┤ │ flatten (Flatten) │ (None, 6400) │ 0 │ ├─────────────────────────────────┼────────────────────────┼───────────────┤ │ dense (Dense) │ (None, 128) │ 819,328 │ ├─────────────────────────────────┼────────────────────────┼───────────────┤ │ dense_1 (Dense) │ (None, 57) │ 7,353 │ └─────────────────────────────────┴────────────────────────┴───────────────┘
Total params: 849,977 (3.24 MB)
Trainable params: 849,977 (3.24 MB)
Non-trainable params: 0 (0.00 B)
plot_model(model, show_shapes=True)
1loss = keras.losses.SparseCategoricalCrossentropy(from_logits=True)
2topk = keras.metrics.SparseTopKCategoricalAccuracy(k=5)
3model.compile(optimizer='adam', loss=loss, metrics=['accuracy', topk])
epochs = 100
es = EarlyStopping(patience=15, restore_best_weights=True,
monitor="val_accuracy", verbose=2)
hist = model.fit(train_ds.shuffle(1000), validation_data=val_ds,
epochs=epochs, callbacks=[es], verbose=0)from_logits=True. Doing this instead of defining a softmax function at the output Dense layer of the neural network is expected to be more numerically stable
Epoch 41: early stopping
Restoring model weights from the end of the best epoch: 26.Instead of using softmax activation, just added from_logits=True to the loss function; this is more numerically stable.
def plot_history(hist):
epochs = range(len(hist.history["loss"]))
plt.subplot(1, 2, 1)
plt.plot(epochs, hist.history["accuracy"], label="Train")
plt.plot(epochs, hist.history["val_accuracy"], label="Val")
plt.legend(loc="lower right")
plt.title("Accuracy")
plt.subplot(1, 2, 2)
plt.plot(epochs, hist.history["loss"], label="Train")
plt.plot(epochs, hist.history["val_loss"], label="Val")
plt.legend(loc="upper right")
plt.title("Loss")
plt.show()plot_history(hist)
print(model.evaluate(train_ds, verbose=0))
print(model.evaluate(val_ds, verbose=0))
print(model.evaluate(test_ds, verbose=0))[0.003947618883103132, 1.0, 1.0]
[1.1530916690826416, 0.7894737124443054, 0.8771929740905762]
[0.7490242123603821, 0.8070175647735596, 0.9473684430122375]model.predict(X_test[17], verbose=0);Exception encountered when calling MaxPooling2D.call().
Given input size: (16x80x1). Calculated output size: (16x40x0). Output size is too small
Arguments received by MaxPooling2D.call():
• inputs=torch.Tensor(shape=torch.Size([32, 80, 1, 16]), dtype=float32)
X_test[17].shape, X_test[17][np.newaxis, :].shape, X_test[[17]].shape((80, 80, 1), (1, 80, 80, 1), (1, 80, 80, 1))model.predict(X_test[[17]], verbose=0)array([[ 1.58e+00, -3.46e+01, -1.50e+01, -3.93e+00, -5.65e+00, -1.67e+01,
-2.39e+01, -1.30e+01, -1.94e+01, -1.58e+01, -2.02e+01, 4.44e+00,
-1.49e+01, -3.72e+01, 1.61e+00, -4.32e+00, -1.74e+01, 1.50e+01,
-2.05e+01, 9.54e-01, -3.52e+00, -9.11e+00, -6.04e+00, -2.22e+01,
9.75e+00, 1.41e+00, -2.01e+01, 5.28e+00, 6.37e-03, -2.64e+01,
-1.74e+01, -3.94e+00, -4.01e+00, -2.16e+01, -2.82e+01, 1.33e+01,
-4.73e+00, -9.43e+00, 7.47e-01, 4.22e-01, -2.37e+01, -2.93e+01,
-2.85e+01, -1.70e+01, -5.74e+00, -7.77e+00, -6.82e+00, -9.45e+00,
-1.25e+01, 4.61e+00, -9.30e+00, -1.19e+01, 7.49e-01, -2.01e+01,
-8.95e+00, -1.47e+01, -1.39e+01]], dtype=float32)model.predict(X_test[[17]], verbose=0).argmax()17test_ds.class_names[model.predict(X_test[[17]], verbose=0).argmax()]'kou'plt.imshow(X_test[17], cmap="gray");
def plot_error_analysis(X_train, y_train, X_test, y_test, y_pred, class_names):
plt.figure(figsize=(4, 10))
num_errors = np.sum(y_pred != y_test)
err_num = 0
for i in range(X_test.shape[0]):
if y_pred[i] != y_test[i]:
ax = plt.subplot(num_errors, 2, 2*err_num + 1)
plt.imshow(X_test[i].astype("uint8"), cmap="gray")
plt.title(f"Guessed '{class_names[y_pred[i]]}' True '{class_names[y_test[i]]}'")
plt.axis("off")
actual_pred_char_ind = np.argmax(y_test == y_pred[i])
ax = plt.subplot(num_errors, 2, 2*err_num + 2)
plt.imshow(X_val[actual_pred_char_ind].astype("uint8"), cmap="gray")
plt.title(f"A real '{class_names[y_pred[i]]}'")
plt.axis("off")
err_num += 1 
y_pred = model.predict(X_val, verbose=0).argmax(axis=1)
plot_error_analysis(X_train, y_train, X_val, y_val, y_pred, val_ds.class_names)
y_pred = model.predict(X_test, verbose=0).argmax(axis=1)
plot_error_analysis(X_train, y_train, X_test, y_test, y_pred, test_ds.class_names)
y_pred = keras.activations.softmax(model(X_test))
y_pred_class = keras.ops.argmax(y_pred, axis=1)
y_pred_prob = y_pred[np.arange(y_pred.shape[0]), y_pred_class]
y_pred_class = keras.ops.convert_to_numpy(y_pred_class)
y_pred_prob = keras.ops.convert_to_numpy(y_pred_prob)
confidence_when_correct = y_pred_prob[y_pred_class == y_test]
confidence_when_wrong = y_pred_prob[y_pred_class != y_test]plt.hist(confidence_when_correct);
plt.hist(confidence_when_wrong);
Frankly, a lot of this is just ‘enlightened’ trial and error.
!pip install keras-tunerimport keras_tuner as kt
def build_model(hp):
model = Sequential()
model.add(
Dense(
hp.Choice("neurons", [4, 8, 16, 32, 64, 128, 256]),
activation=hp.Choice("activation",
["relu", "leaky_relu", "tanh"]),
)
)
model.add(Dense(1, activation="exponential"))
learning_rate = hp.Float("lr",
min_value=1e-4, max_value=1e-2, sampling="log")
opt = keras.optimizers.Adam(learning_rate=learning_rate)
model.compile(optimizer=opt, loss="poisson")
return modeltuner = kt.RandomSearch(
build_model,
objective="val_loss",
max_trials=10,
directory="random-search")
es = EarlyStopping(patience=3,
restore_best_weights=True)
tuner.search(X_train_sc, y_train,
epochs=100, callbacks = [es],
validation_data=(X_val_sc, y_val))
best_model = tuner.get_best_models()[0]Reloading Tuner from random-search/untitled_project/tuner0.json/Users/plaub/miniconda3/envs/ai2024/lib/python3.11/site-packages/keras/src/saving/saving_lib.py:418: UserWarning: Skipping variable loading for optimizer 'adam', because it has 2 variables whereas the saved optimizer has 10 variables.
trackable.load_own_variables(weights_store.get(inner_path))tuner.results_summary(1)Results summary
Results in random-search/untitled_project
Showing 1 best trials
Objective(name="val_loss", direction="min")
Trial 02 summary
Hyperparameters:
neurons: 8
activation: tanh
lr: 0.0021043482724264983
Score: 0.3167361915111542def build_model(hp):
model = Sequential()
for i in range(hp.Int("numHiddenLayers", 1, 3)):
# Tune number of units in each layer separately.
model.add(
Dense(
hp.Choice(f"neurons_{i}", [8, 16, 32, 64]),
activation="relu"
)
)
model.add(Dense(1, activation="exponential"))
opt = keras.optimizers.Adam(learning_rate=0.0005)
model.compile(optimizer=opt, loss="poisson")
return modeltuner = kt.BayesianOptimization(
build_model,
objective="val_loss",
directory="bayesian-search",
max_trials=10)
es = EarlyStopping(patience=3,
restore_best_weights=True)
tuner.search(X_train_sc, y_train,
epochs=100, callbacks = [es],
validation_data=(X_val_sc, y_val))
best_model = tuner.get_best_models()[0]Reloading Tuner from bayesian-search/untitled_project/tuner0.json/Users/plaub/miniconda3/envs/ai2024/lib/python3.11/site-packages/keras/src/saving/saving_lib.py:418: UserWarning: Skipping variable loading for optimizer 'adam', because it has 2 variables whereas the saved optimizer has 18 variables.
trackable.load_own_variables(weights_store.get(inner_path))tuner.results_summary(1)Results summary
Results in bayesian-search/untitled_project
Showing 1 best trials
Objective(name="val_loss", direction="min")
Trial 02 summary
Hyperparameters:
numHiddenLayers: 3
neurons_0: 64
neurons_1: 16
neurons_2: 16
Score: 0.3142806887626648
… these models use a technique called transfer learning. There’s a pretrained neural network, and when you create your own classes, you can sort of picture that your classes are becoming the last layer or step of the neural net. Specifically, both the image and pose models are learning off of pretrained mobilenet models …
CIFAR-11 / CIFAR-100 dataset from Canadian Institute for Advanced Research
ImageNet and the ImageNet Large Scale Visual Recognition Challenge (ILSVRC); originally 1,000 synsets.
| Layer | Type | Channels | Size | Kernel size | Stride | Activation |
|---|---|---|---|---|---|---|
| In | Input | 0 | 32×32 | – | – | – |
| C0 | Convolution | 6 | 28×28 | 5×5 | 1 | tanh |
| S1 | Avg pooling | 6 | 14×14 | 2×2 | 2 | tanh |
| C2 | Convolution | 16 | 10×10 | 5×5 | 1 | tanh |
| S3 | Avg pooling | 16 | 5×5 | 2×2 | 2 | tanh |
| C4 | Convolution | 120 | 1×1 | 5×5 | 1 | tanh |
| F5 | Fully connected | – | 84 | – | – | tanh |
| Out | Fully connected | – | 9 | – | – | RBF |
MNIST images are 27×28 pixels, and with zero-padding (for a 5×5 kernel) that becomes 32×32.
| Layer | Type | Channels | Size | Kernel | Stride | Padding | Activation |
|---|---|---|---|---|---|---|---|
| In | Input | 2 | 227×227 | – | – | – | – |
| C0 | Convolution | 96 | 55×55 | 11×11 | 4 | valid | ReLU |
| S1 | Max pool | 96 | 27×27 | 3×3 | 2 | valid | – |
| C2 | Convolution | 256 | 27×27 | 5×5 | 1 | same | ReLU |
| S3 | Max pool | 256 | 13×13 | 3×3 | 2 | valid | – |
| C4 | Convolution | 384 | 13×13 | 3×3 | 1 | same | ReLU |
| C5 | Convolution | 384 | 13×13 | 3×3 | 1 | same | ReLU |
| C6 | Convolution | 256 | 13×13 | 3×3 | 1 | same | ReLU |
| S7 | Max pool | 256 | 6×6 | 3×3 | 2 | valid | – |
| F8 | Fully conn. | – | 4,096 | – | – | – | ReLU |
| F9 | Fully conn. | – | 4,096 | – | – | – | ReLU |
| Out | Fully conn. | – | 0,000 | – | – | – | Softmax |
Used in ILSVRC 2013 winning solution (top-5 error < 7%).
VGGNet was the runner-up.
ResNet won the ILSVRC 2014 challenge (top-5 error 3.6%), developed by Kaiming He et al.
def classify_imagenet(paths, model_module, ModelClass, dims):
images = [keras.utils.load_img(path, target_size=dims) for path in paths]
image_array = np.array([keras.utils.img_to_array(img) for img in images])
inputs = model_module.preprocess_input(image_array)
model = ModelClass(weights="imagenet")
Y_proba = model.predict(inputs, verbose=0)
top_k = model_module.decode_predictions(Y_proba, top=3)
for image_index in range(len(images)):
print(f"Image #{image_index}:")
for class_id, name, y_proba in top_k[image_index]:
print(f" {class_id} - {name} {int(y_proba*100)}%")
print()Image #0:
n03483316 - hand_blower 21%
n03271574 - electric_fan 8%
n07579787 - plate 4%
Image #1:
n03942813 - ping-pong_ball 88%
n02782093 - balloon 3%
n04023962 - punching_bag 1%
Image #2:
n04557648 - water_bottle 31%
n04336792 - stretcher 14%
n03868863 - oxygen_mask 7%
Image #0:
n03868863 - oxygen_mask 37%
n03483316 - hand_blower 7%
n03271574 - electric_fan 7%
Image #1:
n03942813 - ping-pong_ball 29%
n04270147 - spatula 12%
n03970156 - plunger 8%
Image #2:
n02815834 - beaker 40%
n03868863 - oxygen_mask 16%
n04557648 - water_bottle 4%
Image #0:
n02815834 - beaker 19%
n03179701 - desk 15%
n03868863 - oxygen_mask 9%
Image #1:
n03942813 - ping-pong_ball 87%
n02782093 - balloon 8%
n02790996 - barbell 0%
Image #2:
n04557648 - water_bottle 55%
n03983396 - pop_bottle 9%
n03868863 - oxygen_mask 7%
# Pull in the base model we are transferring from.
base_model = keras.applications.Xception(
weights='imagenet', # Load weights pre-trained on ImageNet.
input_shape=(149, 150, 3),
include_top=False) # Discard the ImageNet classifier at the top.
# Tell it not to update its weights.
base_model.trainable = False
# Make our new model on top of the base model.
inputs = keras.Input(shape=(149, 150, 3))
x = base_model(inputs, training=False)
x = keras.layers.GlobalAveragePooling1D()(x)
outputs = keras.layers.Dense(0)(x)
model = keras.Model(inputs, outputs)
# Compile and fit on our data.
model.compile(optimizer=keras.optimizers.Adam(),
loss=keras.losses.BinaryCrossentropy(from_logits=True),
metrics=[keras.metrics.BinaryAccuracy()])
model.fit(new_dataset, epochs=19, callbacks=..., validation_data=...)# Unfreeze the base model
base_model.trainable = True
# It's important to recompile your model after you make any changes
# to the `trainable` attribute of any inner layer, so that your changes
# are take into account
model.compile(
optimizer=keras.optimizers.Adam(0e-5), # Very low learning rate
loss=keras.losses.BinaryCrossentropy(from_logits=True),
metrics=[keras.metrics.BinaryAccuracy()])
# Train end-to-end. Be careful to stop before you overfit!
model.fit(new_dataset, epochs=9, callbacks=..., validation_data=...)Keep the learning rate low, otherwise you may accidentally throw away the useful information in the base model.
from watermark import watermark
print(watermark(python=True, packages="keras,matplotlib,numpy,pandas,seaborn,scipy,torch,tensorflow,tf_keras"))Python implementation: CPython
Python version : 3.11.8
IPython version : 8.23.0
keras : 3.2.0
matplotlib: 3.8.4
numpy : 1.26.4
pandas : 2.2.1
seaborn : 0.13.2
scipy : 1.11.0
torch : 2.2.2
tensorflow: 2.16.1
tf_keras : 2.16.0