深度学习 AI入门-1

初体验

path = untar_data(URLs.PETS)  # 自动下载并解压数据
fnames = get_image_files(path_img) # 得到路径下所有的图片

data.show_batch(rows=3, figsize=(7,6))  # 查看数据3行，显示尺寸7,6
print(data.classes)  # 数据种类

手写数据集识别：

import torch
from fastai import *
from fastai.vision import *

path = untar_data(URLs.MNIST_SAMPLE)
data = ImageDataBunch.from_folder(path)
learn = cnn_learner(data, models.resnet18, metrics=accuracy)
learn.fit_one_cycle(1)

learn.save('stage-1')  # 保存训练的模型

interp = ClassificationInterpretation.from_learner(learn)
interp.plot_top_losses(9,figsize = (7,7)) # 打印损失最大的九个
interp.plot_confusion_matrix(figsize = (7,7))  # 绘制混淆矩阵

interp.most_confused(min_val=2)  # [('3', '7', 50), ('7', '3', 37)]

'''
most_confused
will simply grab out of the confusion matrix the particular 
combinations of predicted and actual that got wrong the most often. 
简单地从混乱矩阵中找出预测和实际的特定组合，这些组合是最容易出错的。
上面的3和7最容易出错 （实际的，预测的，这种情况发生的次数）上面第一个表示实际上是7，
预测是3，错了22次
'''
# 微调
'''
我们有时候需要微调，重新加了一些层进去后，如果你不解冻，它只会训练新加的一些层，
#而不会训练整个层，这是为了让训练更快，而你要想训练整个层的话，就要先解冻再训练
'''
learn.unfreeze()
learn.fit_one_cycle(1)

learn.lr_find()  # 寻找最适合的学习率
learn.recorder.plot() # 并且画图表示出来

learn.unfreeze()
learn.fit_one_cycle(2, max_lr=slice(1e-6,1e-4))  # 这个学习率区间是看图得出来的

---

下面我们用Titanic号这个例子来说：

from fastai import *
from fastai.tabular import *
import pandas as pd
import numpy as np

train_df = pd.read_csv("train.csv")
test_df = pd.read_csv('test.csv')

for df in [train_df, test_df]:
# 这个是提取名字中的性别特征，这个是csv文件中的一个例子 Braund, Mr. Owen Harris
# 我们最终提取到 Mr
    df['Title'] = df['Name'].str.split(',').str[1].str.split(' ').str[1] 
# 船窗号 eg：C103 取到C
    df['Deck'] = df['Cabin'].str[0]

all_df = pd.concat([train_df, test_df], sort=False)
# 这里按照Title进行分组，对每组进行求年龄平均值
mean_age_by_title = all_df.groupby('Title').mean()['Age']

# 填充年龄缺失值，如果乘客是男性，那就用男性年龄的平均值填充，反之，女性也是
for df in [train_df, test_df]:
    for title, age in mean_age_by_title.iteritems():
        df.loc[df['Age'].isnull() & (df['Title'] == title), 'Age'] = age

test_df.Fare.fillna(0,inplace=True) # 用0填充空白的Fare

dep_var = 'Survived'
cat_names = ['Pclass', 'Sex', 'Embarked', 'Title', 'Deck']
cont_names = ['Age', 'Fare', 'SibSp', 'Parch']
procs = [FillMissing, Categorify, Normalize]

test = TabularList.from_df(test_df, cat_names=cat_names, cont_names=cont_names, procs=procs)
data = (TabularList.from_df(train_df, path='.', cat_names=cat_names, cont_names=cont_names, procs=procs)
                           .split_by_idx(list(range(0,200)))
                           #.split_by_idx(valid_idx=range(200,400))
                           .label_from_df(cols=dep_var)
                           .add_test(test, label=0)
                           .databunch())
#learn = tabular_learner(data, layers=[60, 20], metrics=accuracy)
#learn = tabular_learner(data, layers=[ 300, 150], metrics=accuracy, emb_drop=0.1)
#learn.fit(10)
learn = tabular_learner(data, layers=[1000,500], metrics=accuracy)
learn.fit_one_cycle(5, 2.5e-2)
learn.lr_find()
learn.recorder.plot()

learn.unfreeze()
learn.fit_one_cycle(20, slice(1e-3))

preds, _ = learn.get_preds(ds_type=DatasetType.Test)
pred_prob, pred_class = preds.max(1)

submission = pd.DataFrame({'PassengerId':test_df['PassengerId'],'Survived':pred_class})
submission.to_csv('submission-fastai.csv', index=False)