Kaggle——NFL Big Data Bowl
2024-10-08 15:01:09
neural networks + feature engineering for the win
导入需要的库
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import matplotlib as mpl
import seaborn as sns
import datetime
from kaggle.competitions import nflrush
import tqdm
import re
from string import punctuation
from sklearn.ensemble import RandomForestRegressor
from sklearn.preprocessing import StandardScaler
import keras
from keras.callbacks import ReduceLROnPlateau, ModelCheckpoint, EarlyStopping
from keras.utils import plot_model
import keras.backend as K
import tensorflow as tf
sns.set_style('darkgrid')
mpl.rcParams['figure.figsize'] = [15,10]
env = nflrush.make_env()
train = pd.read_csv('../input/nfl-big-data-bowl-2020/train.csv', dtype={'WindSpeed': 'object'})
全面分析
train.head()
特征工程
#from https://www.kaggle.com/prashantkikani/nfl-starter-lgb-feature-engg
train['DefendersInTheBox_vs_Distance'] = train['DefendersInTheBox'] / train['Distance']
分类特征
cat_features = []
for col in train.columns:
if train[col].dtype =='object':
cat_features.append((col, len(train[col].unique())))
cat_features
对其中一些特性进行预处理。
球场类型
train['StadiumType'].value_counts()
已经看到一些拼写错误了，下面来改正它们。
def clean_StadiumType(txt):
if pd.isna(txt):
return np.nan
txt = txt.lower()
txt = ''.join([c for c in txt if c not in punctuation])
txt = re.sub(' +', ' ', txt)
txt = txt.strip()
txt = txt.replace('outside', 'outdoor')
txt = txt.replace('outdor', 'outdoor')
txt = txt.replace('outddors', 'outdoor')
txt = txt.replace('outdoors', 'outdoor')
txt = txt.replace('oudoor', 'outdoor')
txt = txt.replace('indoors', 'indoor')
txt = txt.replace('ourdoor', 'outdoor')
txt = txt.replace('retractable', 'rtr.')
return txt
train['StadiumType'] = train['StadiumType'].apply(clean_StadiumType)
根据pareto's原则，我们只关注这些词: outdoor, indoor, closed and open.。
def transform_StadiumType(txt):
if pd.isna(txt):
return np.nan
if 'outdoor' in txt or 'open' in txt:
return 1
if 'indoor' in txt or 'closed' in txt:
return 0
return np.nan
train['StadiumType'] = train['StadiumType'].apply(transform_StadiumType)
草坪
#from https://www.kaggle.com/c/nfl-big-data-bowl-2020/discussion/112681#latest-649087
Turf = {'Field Turf':'Artificial',
'A-Turf Titan':'Artificial',
'Grass':'Natural',
'UBU Sports Speed S5-M':'Artificial',
'Artificial':'Artificial',
'DD GrassMaster':'Artificial',
'Natural Grass':'Natural',
'UBU Speed Series-S5-M':'Artificial',
'FieldTurf':'Artificial',
'FieldTurf 360':'Artificial',
'Natural grass':'Natural',
'grass':'Natural',
'Natural':'Natural',
'Artifical':'Artificial',
'FieldTurf360':'Artificial',
'Naturall Grass':'Natural',
'Field turf':'Artificial',
'SISGrass':'Artificial',
'Twenty-Four/Seven Turf':'Artificial',
'natural grass':'Natural'}
train['Turf'] = train['Turf'].map(Turf)
train['Turf'] = train['Turf'] == 'Natural'
拥有的队伍
train[(train['PossessionTeam']!=train['HomeTeamAbbr']) & (train['PossessionTeam']!=train['VisitorTeamAbbr'])][['PossessionTeam', 'HomeTeamAbbr', 'VisitorTeamAbbr']]
在BLT和BAL，ARZ，ARI这样的球队上有一些问题。下面来改正一下。
sorted(train['HomeTeamAbbr'].unique()) == sorted(train['VisitorTeamAbbr'].unique())
diff_abbr = []
for x,y in zip(sorted(train['HomeTeamAbbr'].unique()), sorted(train['PossessionTeam'].unique())):
if x!=y:
print(x + " " + y)
这里有三个问题，让我们来解决它。
map_abbr = {'ARI': 'ARZ', 'BAL': 'BLT', 'CLE': 'CLV', 'HOU': 'HST'}
for abb in train['PossessionTeam'].unique():
map_abbr[abb] = abb
train['PossessionTeam'] = train['PossessionTeam'].map(map_abbr)
train['HomeTeamAbbr'] = train['HomeTeamAbbr'].map(map_abbr)
train['VisitorTeamAbbr'] = train['VisitorTeamAbbr'].map(map_abbr)
train['HomePossesion'] = train['PossessionTeam'] == train['HomeTeamAbbr']
train['Field_eq_Possession'] = train['FieldPosition'] == train['PossessionTeam']
train['HomeField'] = train['FieldPosition'] == train['HomeTeamAbbr']
进攻方式
off_form = train['OffenseFormation'].unique()
train['OffenseFormation'].value_counts()
由于我没有任何关于这方面的知识，我只对这个特性进行一次热编码
train = pd.concat([train.drop(['OffenseFormation'], axis=1), pd.get_dummies(train['OffenseFormation'], prefix='Formation')], axis=1)
dummy_col = train.columns
比赛时间
Game clock is supposed to be a numerical feature.
train['GameClock'].value_counts()
已经有了四分之一的功能，可以将游戏时间除以15分钟，这样就可以得到四分之一的正常时间。
def strtoseconds(txt):
txt = txt.split(':')
ans = int(txt[0])*60 + int(txt[1]) + int(txt[2])/60
return ans
train['GameClock'] = train['GameClock'].apply(strtoseconds)
sns.distplot(train['GameClock'])
球员的身高
train['PlayerHeight‘]
1 ft =12 in, 因此:
train['PlayerHeight'] = train['PlayerHeight'].apply(lambda x: 12*int(x.split('-')[0])+int(x.split('-')[1]))
train['PlayerBMI'] = 703*(train['PlayerWeight']/(train['PlayerHeight'])**2)
Time handoff 和 snap 和参赛者生日
train['TimeHandoff']
train['TimeHandoff'] = train['TimeHandoff'].apply(lambda x: datetime.datetime.strptime(x, "%Y-%m-%dT%H:%M:%S.%fZ"))
train['TimeSnap'] = train['TimeSnap'].apply(lambda x: datetime.datetime.strptime(x, "%Y-%m-%dT%H:%M:%S.%fZ"))
train['TimeDelta'] = train.apply(lambda row: (row['TimeHandoff'] - row['TimeSnap']).total_seconds(), axis=1)
train['PlayerBirthDate'] = train['PlayerBirthDate'].apply(lambda x: datetime.datetime.strptime(x, "%m/%d/%Y"))
让我们用time handoff来计算参赛者的年龄
seconds_in_year = 60*60*24*365.25
train['PlayerAge'] = train.apply(lambda row: (row['TimeHandoff']-row['PlayerBirthDate']).total_seconds()/seconds_in_year, axis=1)
train = train.drop(['TimeHandoff', 'TimeSnap', 'PlayerBirthDate'], axis=1)
风速及风向
train['WindSpeed'].value_counts()
可以看到有些值是不标准化的(例如，12mph)，将从所有的值中删除mph。
train['WindSpeed'] = train['WindSpeed'].apply(lambda x: x.lower().replace('mph', '').strip() if not pd.isna(x) else x)
train['WindSpeed'].value_counts()
#let's replace the ones that has x-y by (x+y)/2
# and also the ones with x gusts up to y
train['WindSpeed'] = train['WindSpeed'].apply(lambda x: (int(x.split('-')[0])+int(x.split('-')[1]))/2 if not pd.isna(x) and '-' in x else x)
train['WindSpeed'] = train['WindSpeed'].apply(lambda x: (int(x.split()[0])+int(x.split()[-1]))/2 if not pd.isna(x) and type(x)!=float and 'gusts up to' in x else x)
def str_to_float(txt):
try:
return float(txt)
except:
return -1
train['WindSpeed'] = train['WindSpeed'].apply(str_to_float)
train['WindDirection'].value_counts()
def clean_WindDirection(txt):
if pd.isna(txt):
return np.nan
txt = txt.lower()
txt = ''.join([c for c in txt if c not in punctuation])
txt = txt.replace('from', '')
txt = txt.replace(' ', '')
txt = txt.replace('north', 'n')
txt = txt.replace('south', 's')
txt = txt.replace('west', 'w')
txt = txt.replace('east', 'e')
return txt
train['WindDirection'] = train['WindDirection'].apply(clean_WindDirection)
train['WindDirection'].value_counts()
def transform_WindDirection(txt):
if pd.isna(txt):
return np.nan
if txt=='n':
return 0
if txt=='nne' or txt=='nen':
return 1/8
if txt=='ne':
return 2/8
if txt=='ene' or txt=='nee':
return 3/8
if txt=='e':
return 4/8
if txt=='ese' or txt=='see':
return 5/8
if txt=='se':
return 6/8
if txt=='ses' or txt=='sse':
return 7/8
if txt=='s':
return 8/8
if txt=='ssw' or txt=='sws':
return 9/8
if txt=='sw':
return 10/8
if txt=='sww' or txt=='wsw':
return 11/8
if txt=='w':
return 12/8
if txt=='wnw' or txt=='nww':
return 13/8
if txt=='nw':
return 14/8
if txt=='nwn' or txt=='nnw':
return 15/8
return np.nan
train['WindDirection'] = train['WindDirection'].apply(transform_WindDirection)
PlayDirection
train['PlayDirection'].value_counts()
train['PlayDirection'] = train['PlayDirection'].apply(lambda x: x.strip() == 'right')
队伍
train['Team'] = train['Team'].apply(lambda x: x.strip()=='home')
比赛时天气
train['GameWeather'].unique()
将应用以下预处理:
- 转成小写
- N/A Indoor, N/A (Indoors) and Indoor => indoor 把它们聚在一起.·
- coudy and clouidy => cloudy
- party => partly
- sunny and clear => clear and sunny
- skies and mostly => ""
train['GameWeather'] = train['GameWeather'].str.lower()
indoor = "indoor"
train['GameWeather'] = train['GameWeather'].apply(lambda x: indoor if not pd.isna(x) and indoor in x else x)
train['GameWeather'] = train['GameWeather'].apply(lambda x: x.replace('coudy', 'cloudy').replace('clouidy', 'cloudy').replace('party', 'partly') if not pd.isna(x) else x)
train['GameWeather'] = train['GameWeather'].apply(lambda x: x.replace('clear and sunny', 'sunny and clear') if not pd.isna(x) else x)
train['GameWeather'] = train['GameWeather'].apply(lambda x: x.replace('skies', '').replace("mostly", "").strip() if not pd.isna(x) else x)
train['GameWeather'].unique()
查看天气描述中最常见的词汇
from collections import Counter
weather_count = Counter()
for weather in train['GameWeather']:
if pd.isna(weather):
continue
for word in weather.split():
weather_count[word]+=1
weather_count.most_common()[:15]
要对我们的天气进行编码，我们要做下面的映射:
- climate controlled or indoor => 3, sunny or sun => 2, clear => 1, cloudy => -1, rain => -2, snow => -3, others => 0
- partly => multiply by 0.5
def map_weather(txt):
ans = 1
if pd.isna(txt):
return 0
if 'partly' in txt:
ans*=0.5
if 'climate controlled' in txt or 'indoor' in txt:
return ans*3
if 'sunny' in txt or 'sun' in txt:
return ans*2
if 'clear' in txt:
return ans
if 'cloudy' in txt:
return -ans
if 'rain' in txt or 'rainy' in txt:
return -2*ans
if 'snow' in txt:
return -3*ans
return 0
train['GameWeather'] = train['GameWeather'].apply(map_weather)
NflId NflIdRusher
train['IsRusher'] = train['NflId'] == train['NflIdRusher']
train.drop(['NflId', 'NflIdRusher'], axis=1, inplace=True)
PlayDirection problems
有一个问题，有一些特征，如X和Y，因为比赛方向，下面来修复这些问题
X, 定位和方向
train['X'] = train.apply(lambda row: row['X'] if row['PlayDirection'] else 120-row['X'], axis=1)
#from https://www.kaggle.com/scirpus/hybrid-gp-and-nn
def new_orientation(angle, play_direction):
if play_direction == 0:
new_angle = 360.0 - angle
if new_angle == 360.0:
new_angle = 0.0
return new_angle
else:
return angle
train['Orientation'] = train.apply(lambda row: new_orientation(row['Orientation'], row['PlayDirection']), axis=1)
train['Dir'] = train.apply(lambda row: new_orientation(row['Dir'], row['PlayDirection']), axis=1)
YardsLeft
我们来计算一下距离终点还有多少码。
train['YardsLeft'] = train.apply(lambda row: 100-row['YardLine'] if row['HomeField'] else row['YardLine'], axis=1)
train['YardsLeft'] = train.apply(lambda row: row['YardsLeft'] if row['PlayDirection'] else 100-row['YardsLeft'], axis=1)
((train['YardsLeft']<train['Yards'])|(train['YardsLeft']-100>train['Yards'])).mean()
Clearly: Yards<=YardsLeft and YardsLeft-100<=Yards,因此，将删除这些错误的行.
train.drop(train.index[(train['YardsLeft']<train['Yards']) | (train['YardsLeft']-100>train['Yards'])], inplace=True)
Baseline model
放弃分类特性，在模型中运行一个简单的随机森林
train = train.sort_values(by=['PlayId', 'Team', 'IsRusher', 'JerseyNumber']).reset_index()
train.drop(['GameId', 'PlayId', 'index', 'IsRusher', 'Team'], axis=1, inplace=True)
cat_features = []
for col in train.columns:
if train[col].dtype =='object':
cat_features.append(col)
train = train.drop(cat_features, axis=1)
现在要为每一场比赛做一个大的排，其中最后一个是冲锋者
train.fillna(-999, inplace=True)
players_col = []
for col in train.columns:
if train[col][:22].std()!=0:
players_col.append(col)
X_train = np.array(train[players_col]).reshape(-1, len(players_col)*22)
play_col = train.drop(players_col+['Yards'], axis=1).columns
X_play_col = np.zeros(shape=(X_train.shape[0], len(play_col)))
for i, col in enumerate(play_col):
X_play_col[:, i] = train[col][::22]
X_train = np.concatenate([X_train, X_play_col], axis=1)
y_train = np.zeros(shape=(X_train.shape[0], 199))
for i,yard in enumerate(train['Yards'][::22]):
y_train[i, yard+99:] = np.ones(shape=(1, 100-yard))
scaler = StandardScaler()
X_train = scaler.fit_transform(X_train)
batch_size=64
class RAdam(keras.optimizers.Optimizer):
"""RAdam optimizer.
# Arguments
learning_rate: float >= 0. Learning rate.
beta_1: float, 0 < beta < 1. Generally close to 1.
beta_2: float, 0 < beta < 1. Generally close to 1.
epsilon: float >= 0. Fuzz factor. If `None`, defaults to `K.epsilon()`.
decay: float >= 0. Learning rate decay over each update.
weight_decay: float >= 0. Weight decay for each param.
amsgrad: boolean. Whether to apply the AMSGrad variant of this
algorithm from the paper "On the Convergence of Adam and
Beyond".
total_steps: int >= 0. Total number of training steps. Enable warmup by setting a positive value.
warmup_proportion: 0 < warmup_proportion < 1. The proportion of increasing steps.
min_lr: float >= 0. Minimum learning rate after warmup.
# References
- [Adam - 一种随机优化方法](https://arxiv.org/abs/1412.6980v8)
- [关于Adam和Beyond的收敛 ](https://openreview.net/forum?id=ryQu7f-RZ)
- [研究了自适应学习速率的方差及其影响因素](https://arxiv.org/pdf/1908.03265v1.pdf)
"""
def __init__(self, learning_rate=0.001, beta_1=0.9, beta_2=0.999,
epsilon=None, decay=0., weight_decay=0., amsgrad=False,
total_steps=0, warmup_proportion=0.1, min_lr=0., **kwargs):
learning_rate = kwargs.pop('lr', learning_rate)
super(RAdam, self).__init__(**kwargs)
with K.name_scope(self.__class__.__name__):
self.iterations = K.variable(0, dtype='int64', name='iterations')
self.learning_rate = K.variable(learning_rate, name='learning_rate')
self.beta_1 = K.variable(beta_1, name='beta_1')
self.beta_2 = K.variable(beta_2, name='beta_2')
self.decay = K.variable(decay, name='decay')
self.weight_decay = K.variable(weight_decay, name='weight_decay')
self.total_steps = K.variable(total_steps, name='total_steps')
self.warmup_proportion = K.variable(warmup_proportion, name='warmup_proportion')
self.min_lr = K.variable(min_lr, name='min_lr')
if epsilon is None:
epsilon = K.epsilon()
self.epsilon = epsilon
self.initial_decay = decay
self.initial_weight_decay = weight_decay
self.initial_total_steps = total_steps
self.amsgrad = amsgrad
def get_updates(self, loss, params):
grads = self.get_gradients(loss, params)
self.updates = [K.update_add(self.iterations, 1)]
lr = self.lr
if self.initial_decay > 0:
lr = lr * (1. / (1. + self.decay * K.cast(self.iterations, K.dtype(self.decay))))
t = K.cast(self.iterations, K.floatx()) + 1
if self.initial_total_steps > 0:
warmup_steps = self.total_steps * self.warmup_proportion
decay_steps = K.maximum(self.total_steps - warmup_steps, 1)
decay_rate = (self.min_lr - lr) / decay_steps
lr = K.switch(
t <= warmup_steps,
lr * (t / warmup_steps),
lr + decay_rate * K.minimum(t - warmup_steps, decay_steps),
)
ms = [K.zeros(K.int_shape(p), dtype=K.dtype(p), name='m_' + str(i)) for (i, p) in enumerate(params)]
vs = [K.zeros(K.int_shape(p), dtype=K.dtype(p), name='v_' + str(i)) for (i, p) in enumerate(params)]
if self.amsgrad:
vhats = [K.zeros(K.int_shape(p), dtype=K.dtype(p), name='vhat_' + str(i)) for (i, p) in enumerate(params)]
else:
vhats = [K.zeros(1, name='vhat_' + str(i)) for i in range(len(params))]
self.weights = [self.iterations] + ms + vs + vhats
beta_1_t = K.pow(self.beta_1, t)
beta_2_t = K.pow(self.beta_2, t)
sma_inf = 2.0 / (1.0 - self.beta_2) - 1.0
sma_t = sma_inf - 2.0 * t * beta_2_t / (1.0 - beta_2_t)
for p, g, m, v, vhat in zip(params, grads, ms, vs, vhats):
m_t = (self.beta_1 * m) + (1. - self.beta_1) * g
v_t = (self.beta_2 * v) + (1. - self.beta_2) * K.square(g)
m_corr_t = m_t / (1.0 - beta_1_t)
if self.amsgrad:
vhat_t = K.maximum(vhat, v_t)
v_corr_t = K.sqrt(vhat_t / (1.0 - beta_2_t))
self.updates.append(K.update(vhat, vhat_t))
else:
v_corr_t = K.sqrt(v_t / (1.0 - beta_2_t))
r_t = K.sqrt((sma_t - 4.0) / (sma_inf - 4.0) *
(sma_t - 2.0) / (sma_inf - 2.0) *
sma_inf / sma_t)
p_t = K.switch(sma_t >= 5, r_t * m_corr_t / (v_corr_t + self.epsilon), m_corr_t)
if self.initial_weight_decay > 0:
p_t += self.weight_decay * p
p_t = p - lr * p_t
self.updates.append(K.update(m, m_t))
self.updates.append(K.update(v, v_t))
new_p = p_t
# Apply constraints.
if getattr(p, 'constraint', None) is not None:
new_p = p.constraint(new_p)
self.updates.append(K.update(p, new_p))
return self.updates
@property
def lr(self):
return self.learning_rate
@lr.setter
def lr(self, learning_rate):
self.learning_rate = learning_rate
def get_config(self):
config = {
'learning_rate': float(K.get_value(self.learning_rate)),
'beta_1': float(K.get_value(self.beta_1)),
'beta_2': float(K.get_value(self.beta_2)),
'decay': float(K.get_value(self.decay)),
'weight_decay': float(K.get_value(self.weight_decay)),
'epsilon': self.epsilon,
'amsgrad': self.amsgrad,
'total_steps': float(K.get_value(self.total_steps)),
'warmup_proportion': float(K.get_value(self.warmup_proportion)),
'min_lr': float(K.get_value(self.min_lr)),
}
base_config = super(RAdam, self).get_config()
return dict(list(base_config.items()) + list(config.items()))
#from https://www.kaggle.com/davidcairuz/nfl-neural-network-w-softmax
def crps(y_true, y_pred):
return K.mean(K.square(y_true - K.cumsum(y_pred, axis=1)), axis=1)
def get_model():
x = keras.layers.Input(shape=[X_train.shape[1]])
fc1 = keras.layers.Dense(units=450, input_shape=[X_train.shape[1]])(x)
act1 = keras.layers.PReLU()(fc1)
bn1 = keras.layers.BatchNormalization()(act1)
dp1 = keras.layers.Dropout(0.55)(bn1)
gn1 = keras.layers.GaussianNoise(0.15)(dp1)
concat1 = keras.layers.Concatenate()([x, gn1])
fc2 = keras.layers.Dense(units=600)(concat1)
act2 = keras.layers.PReLU()(fc2)
bn2 = keras.layers.BatchNormalization()(act2)
dp2 = keras.layers.Dropout(0.55)(bn2)
gn2 = keras.layers.GaussianNoise(0.15)(dp2)
concat2 = keras.layers.Concatenate()([concat1, gn2])
fc3 = keras.layers.Dense(units=400)(concat2)
act3 = keras.layers.PReLU()(fc3)
bn3 = keras.layers.BatchNormalization()(act3)
dp3 = keras.layers.Dropout(0.55)(bn3)
gn3 = keras.layers.GaussianNoise(0.15)(dp3)
concat3 = keras.layers.Concatenate([concat2, gn3])
output = keras.layers.Dense(units=199, activation='softmax')(concat2)
model = keras.models.Model(inputs=[x], outputs=[output])
return model
def train_model(X_train, y_train, X_val, y_val):
model = get_model()
model.compile(optimizer=RAdam(warmup_proportion=0.1, min_lr=1e-7), loss=crps)
er = EarlyStopping(patience=20, min_delta=1e-4, restore_best_weights=True, monitor='val_loss')
model.fit(X_train, y_train, epochs=200, callbacks=[er], validation_data=[X_val, y_val], batch_size=batch_size)
return model
from sklearn.model_selection import RepeatedKFold
rkf = RepeatedKFold(n_splits=5, n_repeats=5)
models = []
for tr_idx, vl_idx in rkf.split(X_train, y_train):
x_tr, y_tr = X_train[tr_idx], y_train[tr_idx]
x_vl, y_vl = X_train[vl_idx], y_train[vl_idx]
model = train_model(x_tr, y_tr, x_vl, y_vl)
models.append(model)
def make_pred(df, sample, env, models):
df['StadiumType'] = df['StadiumType'].apply(clean_StadiumType)
df['StadiumType'] = df['StadiumType'].apply(transform_StadiumType)
df['DefendersInTheBox_vs_Distance'] = df['DefendersInTheBox'] / df['Distance']
df['OffenseFormation'] = df['OffenseFormation'].apply(lambda x: x if x in off_form else np.nan)
df = pd.concat([df.drop(['OffenseFormation'], axis=1), pd.get_dummies(df['OffenseFormation'], prefix='Formation')], axis=1)
missing_cols = set( dummy_col ) - set( df.columns )-set('Yards')
for c in missing_cols:
df[c] = 0
df = df[dummy_col]
df.drop(['Yards'], axis=1, inplace=True)
df['Turf'] = df['Turf'].map(Turf)
df['Turf'] = df['Turf'] == 'Natural'
df['PossessionTeam'] = df['PossessionTeam'].map(map_abbr)
df['HomeTeamAbbr'] = df['HomeTeamAbbr'].map(map_abbr)
df['VisitorTeamAbbr'] = df['VisitorTeamAbbr'].map(map_abbr)
df['HomePossesion'] = df['PossessionTeam'] == df['HomeTeamAbbr']
df['Field_eq_Possession'] = df['FieldPosition'] == df['PossessionTeam']
df['HomeField'] = df['FieldPosition'] == df['HomeTeamAbbr']
df['GameClock'] = df['GameClock'].apply(strtoseconds)
df['PlayerHeight'] = df['PlayerHeight'].apply(lambda x: 12*int(x.split('-')[0])+int(x.split('-')[1]))
df['PlayerBMI'] = 703*(df['PlayerWeight']/(df['PlayerHeight'])**2)
df['TimeHandoff'] = df['TimeHandoff'].apply(lambda x: datetime.datetime.strptime(x, "%Y-%m-%dT%H:%M:%S.%fZ"))
df['TimeSnap'] = df['TimeSnap'].apply(lambda x: datetime.datetime.strptime(x, "%Y-%m-%dT%H:%M:%S.%fZ"))
df['TimeDelta'] = df.apply(lambda row: (row['TimeHandoff'] - row['TimeSnap']).total_seconds(), axis=1)
df['PlayerBirthDate'] = df['PlayerBirthDate'].apply(lambda x: datetime.datetime.strptime(x, "%m/%d/%Y"))
seconds_in_year = 60*60*24*365.25
df['PlayerAge'] = df.apply(lambda row: (row['TimeHandoff']-row['PlayerBirthDate']).total_seconds()/seconds_in_year, axis=1)
df['WindSpeed'] = df['WindSpeed'].apply(lambda x: x.lower().replace('mph', '').strip() if not pd.isna(x) else x)
df['WindSpeed'] = df['WindSpeed'].apply(lambda x: (int(x.split('-')[0])+int(x.split('-')[1]))/2 if not pd.isna(x) and '-' in x else x)
df['WindSpeed'] = df['WindSpeed'].apply(lambda x: (int(x.split()[0])+int(x.split()[-1]))/2 if not pd.isna(x) and type(x)!=float and 'gusts up to' in x else x)
df['WindSpeed'] = df['WindSpeed'].apply(str_to_float)
df['WindDirection'] = df['WindDirection'].apply(clean_WindDirection)
df['WindDirection'] = df['WindDirection'].apply(transform_WindDirection)
df['PlayDirection'] = df['PlayDirection'].apply(lambda x: x.strip() == 'right')
df['Team'] = df['Team'].apply(lambda x: x.strip()=='home')
indoor = "indoor"
df['GameWeather'] = df['GameWeather'].apply(lambda x: indoor if not pd.isna(x) and indoor in x else x)
df['GameWeather'] = df['GameWeather'].apply(lambda x: x.lower().replace('coudy', 'cloudy').replace('clouidy', 'cloudy').replace('party', 'partly').replace('clear and sunny', 'sunny and clear').replace('skies', '').replace("mostly", "").strip() if not pd.isna(x) else x)
df['GameWeather'] = df['GameWeather'].apply(map_weather)
df['IsRusher'] = df['NflId'] == df['NflIdRusher']
df['X'] = df.apply(lambda row: row['X'] if row['PlayDirection'] else 120-row['X'], axis=1)
df['Orientation'] = df.apply(lambda row: new_orientation(row['Orientation'], row['PlayDirection']), axis=1)
df['Dir'] = df.apply(lambda row: new_orientation(row['Dir'], row['PlayDirection']), axis=1)
df['YardsLeft'] = df.apply(lambda row: 100-row['YardLine'] if row['HomeField'] else row['YardLine'], axis=1)
df['YardsLeft'] = df.apply(lambda row: row['YardsLeft'] if row['PlayDirection'] else 100-row['YardsLeft'], axis=1)
df = df.sort_values(by=['PlayId', 'Team', 'IsRusher', 'JerseyNumber']).reset_index()
df = df.drop(['TimeHandoff', 'TimeSnap', 'PlayerBirthDate', 'NflId', 'NflIdRusher', 'GameId', 'PlayId', 'index', 'IsRusher', 'Team'], axis=1)
cat_features = []
for col in df.columns:
if df[col].dtype =='object':
cat_features.append(col)
df = df.drop(cat_features, axis=1)
df.fillna(-999, inplace=True)
X = np.array(df[players_col]).reshape(-1, len(players_col)*22)
play_col = df.drop(players_col, axis=1).columns
X_play_col = np.zeros(shape=(X.shape[0], len(play_col)))
for i, col in enumerate(play_col):
X_play_col[:, i] = df[col][::22]
X = np.concatenate([X, X_play_col], axis=1)
X = scaler.transform(X)
y_pred = np.mean([np.cumsum(model.predict(X), axis=1) for model in models], axis=0)
yardsleft = np.array(df['YardsLeft'][::22])
for i in range(len(yardsleft)):
y_pred[i, :yardsleft[i]-1] = 0
y_pred[i, yardsleft[i]+100:] = 1
env.predict(pd.DataFrame(data=y_pred.clip(0,1),columns=sample.columns))
return y_pred
for test, sample in tqdm.tqdm(env.iter_test()):
make_pred(test, sample, env, models)
env.write_submission_file()
env.write_submission_file()
最新文章
- [转]Asp.net mvc 网站之速度优化 -- 页面缓存
- 背包dp整理
- 一个js（javascript）使用案例
- Linux使用常见错误集锦
- Bootstrap3 formテキストフィールド横幅の指定の仕方
- 基于视觉信息的网页分块算法（VIPS） - yysdsyl的专栏 - 博客频道 - CSDN.NET
- [Andriod官方API指南]连接之蓝牙
- MQ通道搭建以及连通性检查
- Python爬虫番外篇之关于登录
- (Lesson2)根据类名称和属性获得元素-JavaScript面向对象
- python+selenium实现登录账户
- 第24章 退出 - Identity Server 4 中文文档(v1.0.0)
- Linux中Tomcat设置开机启动
- flutter 底部按钮切换页面
- Windows 10 & change DNS
- CSS3媒体查询的部分重要属性
- Fence Repair（poj3253）
- Here We Go(relians) Again HDU2722
- 吴恩达讲了干货满满的一节全新AI课，全程手写板书充满诚意非常干货
- SVG 学习<七> SVG的路径——path(1)直线命令、弧线命令
热门文章
- hibernate注解 笔记
- 18 —— node 热部署工具 — supervisor /nodemon 。
- bzoj1415 NOI2005聪聪和可可
- Toolbar中系统自带返回键（setDisplayHomeAsUpEnabled）与Toolbar本身的inflateMenu冲突问题
- stm32h7 开发板学习
- element穿梭框el-transfer增加拖拽排序和shift多选checkbox功能
- <style scoped >中使用深度选择器影响子组件
- Bootstrap-模态框 modal.js
- CTF -攻防世界-web高手区-mfw
- VMware-Workstation-Full-12.5.9