Kaggle X FastAI系列 1 --titanic

数据准备

        直接在Anaconda命令行里用kaggle提供的命令下载，参考fastbook中09_tabula的API方法没有成功。

        用pandas读取一下数据。

path = URLs.path('titanic')
train_path = path/'train.csv'
test_path = path/'test.csv'
df = pd.read_csv(train_path, low_memory=False,skipinitialspace=True)
test_df = pd.read_csv(test_path, low_memory=False,skipinitialspace=True)

        生成DataLoaders

cat_names = [ 'Pclass', 'SibSp',
       'Parch', 'Fare', 'Cabin', 'Embarked', 'Sex']
cont_names = ['Age','PassengerId']
procs = [Categorify, FillMissing, Normalize]
dls = TabularDataLoaders.from_csv(train_path, path, procs=procs, cat_names=cat_names, cont_names=cont_names, 
                                 y_names=dep_var, valid_idx=list(range(600,800)), bs=64,y_block = CategoryBlock)

        其中要特别注意的是预测中的y_name,这个数据集中是dep_var = 'Survived'是不可以加入cat_names 和cont_names中的，第一次用的时候直接对df.columns复制粘贴，因为加入了y_name到离散型变量中引发了莫名其妙的pandas bug，找了将近一个小时才发现是这个原因。

开始训练

learn = tabular_learner(dls,metrics=accuracy)
learn.fit_one_cycle(100)

        这里就采用了默认的tabular_learner，结果如下：

epoch	train_loss	valid_loss	accuracy	time
0	0.687910	0.672002	0.640000	00:01
1	0.679152	0.665265	0.640000	00:00
2	0.670670	0.658679	0.680000	00:00
3	0.663144	0.649106	0.670000	00:00
4	0.654119	0.642387	0.670000	00:00
5	0.642554	0.643897	0.610000	00:00
6	0.627525	0.626449	0.615000	00:00
7	0.608022	0.594794	0.680000	00:00
8	0.579566	0.574257	0.710000	00:00
9	0.540928	0.532365	0.725000	00:00
10	0.495352	0.487120	0.755000	00:00
11	0.449479	0.445067	0.770000	00:00
12	0.407328	0.510866	0.755000	00:00
13	0.372619	0.512942	0.775000	00:00
14	0.344380	0.554082	0.775000	00:00
15	0.316252	0.487315	0.775000	00:00
16	0.290813	0.525596	0.755000	00:00
17	0.266976	0.606796	0.775000	00:00
18	0.248752	0.604665	0.755000	00:00
19	0.232290	0.590609	0.750000	00:00
20	0.218391	0.578665	0.745000	00:00
21	0.203587	0.586581	0.750000	00:00
22	0.189189	0.628336	0.705000	00:00
23	0.180911	0.689595	0.715000	00:00
24	0.170023	0.637040	0.735000	00:00
25	0.161536	0.630965	0.765000	00:00
26	0.152605	0.591396	0.740000	00:00
27	0.146294	0.648590	0.730000	00:00
28	0.142579	0.688091	0.730000	00:00
29	0.135967	0.781510	0.715000	00:00
30	0.127739	0.744347	0.740000	00:00
31	0.123255	0.655497	0.745000	00:00
32	0.119241	0.715305	0.745000	00:00
33	0.114936	0.902263	0.725000	00:00
34	0.112255	0.839763	0.720000	00:00
35	0.106522	0.742290	0.730000	00:00
36	0.102651	0.789978	0.760000	00:00
37	0.100017	0.771653	0.730000	00:00
38	0.096346	0.885213	0.720000	00:00
39	0.094071	0.821706	0.745000	00:00
40	0.089969	0.728642	0.750000	00:00
41	0.086300	0.764377	0.735000	00:00
42	0.080375	0.737520	0.750000	00:00
43	0.075749	0.840410	0.735000	00:00
44	0.070874	0.877053	0.735000	00:00
45	0.070402	0.879526	0.725000	00:00
46	0.069162	0.885928	0.730000	00:00
47	0.065940	0.951272	0.755000	00:00
48	0.064935	0.835828	0.760000	00:00
49	0.062491	0.956926	0.705000	00:00
50	0.062818	0.995981	0.730000	00:00
51	0.060378	0.943567	0.745000	00:00
52	0.061401	1.084070	0.710000	00:00
53	0.061499	0.970715	0.725000	00:00
54	0.061217	1.169610	0.735000	00:00
55	0.061242	0.914683	0.740000	00:00
56	0.058386	0.955398	0.740000	00:00
57	0.055345	0.921871	0.720000	00:00
58	0.052027	0.915062	0.735000	00:00
59	0.047930	0.933616	0.735000	00:00
60	0.043968	0.998685	0.735000	00:00
61	0.042489	0.962408	0.745000	00:00
62	0.041475	1.049256	0.720000	00:00
63	0.040391	1.045698	0.720000	00:00
64	0.042820	1.019700	0.735000	00:00
65	0.043163	0.971371	0.725000	00:00
66	0.039586	1.006120	0.745000	00:00
67	0.039305	1.085823	0.705000	00:00
68	0.038155	1.167134	0.700000	00:00
69	0.035776	1.143821	0.700000	00:00
70	0.035192	1.146205	0.725000	00:00
71	0.034136	1.175187	0.710000	00:00
72	0.032035	1.082435	0.730000	00:00
73	0.031123	1.080388	0.725000	00:00
74	0.030651	1.147222	0.710000	00:00
75	0.030057	1.151547	0.730000	00:00
76	0.028681	1.130530	0.730000	00:00
77	0.027378	1.152099	0.715000	00:00
78	0.028083	1.134053	0.725000	00:00
79	0.027603	1.147027	0.725000	00:00
80	0.026748	1.164077	0.715000	00:00
81	0.027223	1.200746	0.715000	00:00
82	0.025508	1.146767	0.715000	00:00
83	0.024893	1.130284	0.720000	00:00
84	0.025231	1.126982	0.720000	00:00
85	0.025987	1.133423	0.720000	00:00
86	0.025825	1.143805	0.715000	00:00
87	0.025462	1.144619	0.725000	00:00
88	0.024687	1.134826	0.720000	00:00
89	0.025501	1.134439	0.735000	00:00
90	0.024548	1.137204	0.725000	00:00
91	0.023910	1.149608	0.730000	00:00
92	0.023950	1.150210	0.730000	00:00
93	0.024212	1.141803	0.730000	00:00
94	0.025319	1.147405	0.725000	00:00
95	0.024329	1.148225	0.730000	00:00
96	0.023654	1.158221	0.725000	00:00
97	0.024523	1.170079	0.715000	00:00
98	0.024479	1.165972	0.720000	00:00
99	0.023362	1.153834	0.730000	00:00

可以看出很不幸，验证集的loss远大于训练集的loss，标准的OverFitting发生了且很严重，下一步准备纳入L2正则化之类的东西改善一下，准确率达到了73%，勉强可以接受。

模型预测

test_df = pd.read_csv(test_path, low_memory=False,skipinitialspace=True)
#df.head()
dl = learn.dls.test_dl(test_df)
preds=learn.get_preds(dl=dl)

提交比赛结果

 

        得到了0.7的得分（最高1.00000） 

改进计划

        体感修一修OverFitting性能提高的空间还是挺大的，目标0.9分，改进之后下次上船Github贴上Github仓库