dogs_vs._cats's Introduction

Dogs vs. Cats

A computation from Kaggle and a graduation project for machine learning engineer nanodegree on Udacity

本项目使用 Python 在 Jupyter Notebook 环境下开发，并用到了如下的库：

import os
import random
import cv2
from matplotlib import pyplot
import numpy
from keras.models import Model
from keras.layers import Dense, GlobalAveragePooling2D
from keras.applications.xception import Xception
from IPython.display import SVG
from keras.utils.vis_utils import model_to_dot
from keras.optimizers import Nadam
from keras.callbacks import EarlyStopping

其中, cv2 来自 OpenCV, 用于图片的读取、缩放等操作; keras 用于 CNN 深度学习模型的构建和训练。

用于训练 CNN 的数据集来自猫狗大战

在 AWS p2.xlarge 实例中（输出层微调）运行结果如下：

Train on 15000 samples, validate on 5000 samples
Epoch 1/10
15000/15000 [==============================] - 332s - loss: 0.1624 - acc: 0.9503 - val_loss: 0.8713 - val_acc: 0.7998
Epoch 2/10
15000/15000 [==============================] - 334s - loss: 0.0673 - acc: 0.9759 - val_loss: 1.1316 - val_acc: 0.7266
Epoch 3/10
15000/15000 [==============================] - 334s - loss: 0.0641 - acc: 0.9761 - val_loss: 0.0521 - val_acc: 0.9800
Epoch 4/10
15000/15000 [==============================] - 334s - loss: 0.0500 - acc: 0.9805 - val_loss: 0.0646 - val_acc: 0.9758
Epoch 5/10
15000/15000 [==============================] - 334s - loss: 0.0426 - acc: 0.9833 - val_loss: 0.0731 - val_acc: 0.9740
Epoch 6/10
15000/15000 [==============================] - 334s - loss: 0.0389 - acc: 0.9869 - val_loss: 0.0786 - val_acc: 0.9704
Epoch 7/10
15000/15000 [==============================] - 334s - loss: 0.0328 - acc: 0.9877 - val_loss: 0.0713 - val_acc: 0.9758