始めに
コンペ主催 cf. @solafune (https://solafune.com) にて、「夜間光データから土地価格を予測」というコンペが開催中です。
カラムが5行しかないということで、気軽に参加できます
こういった記事でのシェアリングを許可しているコンペです
(コード含め2時間くらいで書いたので、結構怪しいところ残っているかも。あまりリファクタリングできてません)
データ概要
公式より
把握しておくべきは、IDがtrainとtestで完全にわかれていることくらいです
予測対象は1992年～2013年までのある地域のすべての年の「土地の平均価格」です。
なので、地域ごとに1992年～2013年の時系列データととらえることも出来ます。
EDAはこの辺の記事みれば多分大体わかると思います。
Solafune 夜間光コンペ Baseline(xgb,lgb,cat)
[solafune] 夜間光データから土地価格を予測ベースモデル
コード
コード
LSTM_baseline.ipynb
import os
import pandas as pd
import numpy as np
import matplotlib.pyplot as plt
import seaborn as sns
from sklearn.metrics import mean_squared_error
from sklearn.model_selection import StratifiedKFold, GroupKFold
from copy import deepcopy
from sklearn.preprocessing import StandardScaler
sns.set_theme(style="ticks")
DATA_DIR = './data/'
OUTPUT_DIR = "output"
# データ読み込み
train = pd.read_csv(os.path.join(DATA_DIR, 'TrainDataSet.csv'))
test = pd.read_csv(os.path.join(DATA_DIR, 'EvaluationData.csv'))
submission = pd.read_csv(os.path.join(DATA_DIR, 'UploadFileTemplate.csv'))
# only full data
count_ids = train.PlaceID.value_counts().reset_index()
valid_PlaceID =count_ids[count_ids.PlaceID ==22]["index"]
n_train = len(train)
train = train[train.PlaceID.isin(valid_PlaceID)]
train = train.reset_index().drop(columns = "index")
print(f"Dropped : {round(100 - 100 * len(train)/n_train,1)}%")
from contextlib import contextmanager
from time import time
@contextmanager
def timer(logger=None, format_str='{:.3f}[s]', prefix=None, suffix=None):
if prefix: format_str = str(prefix) + format_str
if suffix: format_str = format_str + str(suffix)
start = time()
yield
d = time() - start
out_str = format_str.format(d)
if logger:
logger.info(out_str)
else:
print(out_str)
feature_cols = ['MeanLight', 'SumLight']
for f in feature_cols:
ss = StandardScaler()
train[f] = ss.fit_transform(train[[f]])
test[f] = ss.transform(test[[f]])
import torch
from torch.utils.data import Dataset
from torch import nn
import torch.optim as optim
class SeqDataset(Dataset):
def __init__(self, df, place_ids, feature_col, target_col, is_log = True, test = False):
self.df = df
self.place_ids = place_ids
self.feature_col = feature_col
self.target_col = target_col
self.is_log = is_log
if test:
self.df[target_col] = 0
def __len__(self):
return len(self.place_ids)
def __getitem__(self, idx):
place_id = self.place_ids[idx]
seq = self.df.loc[self.df.PlaceID == place_id, self.feature_col].values
label = self.df.loc[self.df.PlaceID == place_id, self.target_col].values
seq, label = torch.tensor(seq).float(), torch.tensor(label).float()
if self.is_log:
label = torch.log1p(label)
return seq, label
# dataset = SeqDataset(df = train, place_ids = train.PlaceID.unique(),
# feature_col = feature_cols, target_col = "AverageLandPrice")
testset = SeqDataset(df = test, place_ids = test.PlaceID.unique(),
feature_col = feature_cols, target_col = "AverageLandPrice", test = True)
# dataloader = torch.utils.data.DataLoader(dataset, batch_size = 5)
testloader = torch.utils.data.DataLoader(testset, batch_size = 1)
class BaselineLSTM(nn.Module):
def __init__(
self,
emb_dim=128,
rnn_dim=128,
hidden_size=128,
num_layers=2,
dropout=0.3,
rnn_dropout=0.3,
):
super().__init__()
self.emb_dim = emb_dim
self.rnn_dim = rnn_dim
self.hidden_size = hidden_size
self.num_layers = num_layers
self.dropout = dropout
self.rnn_dropout = rnn_dropout
self.lstm = nn.LSTM(
input_size=2,
hidden_size=hidden_size,
num_layers=num_layers,
dropout=rnn_dropout,
bidirectional=False,
batch_first=True,
)
self.fc = nn.Linear(hidden_size, 1)
def forward(self,x):
x, _ = self.lstm(x)
x = self.fc(x)
return x
EPOCHS = 10
LEARNING_LATE = 0.01
def fit_LSTM(X, y, cv=None,):
models = []
oof_pred = np.zeros_like(y, dtype=np.float)
for i, (idx_train, idx_valid) in enumerate(cv.split(X, y, train['PlaceID'])):
trainloader = torch.utils.data.DataLoader(
SeqDataset(
df = X.iloc[idx_train],
place_ids = X.iloc[idx_train].PlaceID.unique(),
feature_col = feature_cols,
target_col = "AverageLandPrice"
)
, batch_size = 4)
validloader = torch.utils.data.DataLoader(
SeqDataset(
df = X.iloc[idx_valid],
place_ids = X.iloc[idx_valid].PlaceID.unique(),
feature_col = feature_cols,
target_col = "AverageLandPrice"
)
, batch_size = 1)
model = BaselineLSTM()
criterion = nn.MSELoss()
#optimizer = optim.SGD(model.parameters(), lr=LEARNING_LATE)
optimizer = optim.Adam(model.parameters(), lr=LEARNING_LATE)
with timer(prefix='fit fold={} '.format(i + 1)):
for epoch in range(EPOCHS):
for data in trainloader:
seq_data,label = data
optimizer.zero_grad()
pred = model(seq_data)
loss = torch.sqrt(criterion(pred.squeeze(2), label))
loss.backward()
optimizer.step()
print(f"\repoch:{epoch} loss:{loss.item()}", end = "")
pred_i = []
for data in validloader:
seq_data, _ = data
with torch.no_grad():
model.eval()
pred_i.append(model(seq_data).squeeze(2).detach().numpy())
pred_i = np.array(pred_i).reshape(-1)
oof_pred[idx_valid] = pred_i
models.append(model)
print(f'Fold {i} RMSLE: {mean_squared_error(np.log1p(X.iloc[idx_valid].AverageLandPrice.values), pred_i):.4f}')
score = mean_squared_error(np.log1p(y), oof_pred)
print('FINISHED \ whole score: {:.4f}'.format(score))
return oof_pred, models
def create_predict(models, testloader):
pred = []
for data in testloader:
seq_data, _ = data
with torch.no_grad():
p = []
for model in models:
model.eval()
p.append(
model(seq_data).detach().numpy().reshape(-1)
)
pred.append(p)
pred = np.mean(pred, axis=1)
return pred
def fit_and_predict(train_df,
target_df, ):
target_name = "AverageLandPrice"
print('-' * 20 + ' start {} '.format(target_name) + '-' * 20)
y = target_df.values
cv = GroupKFold(n_splits=5)
# モデルの学習.
oof, models = fit_LSTM(train_df, y, cv=cv)
return oof, models
oof, models = fit_and_predict(train_df=train,
target_df=train.AverageLandPrice,)
# 予測モデルで推論実行
with timer(prefix='predict'):
pred = create_predict(
models,
testloader
)
submission.LandPrice = np.expm1(pred.reshape(-1))
submission.to_csv(os.path.join(OUTPUT_DIR, 'submission_lstm.csv'), index=False)
# LB 0.784676
本当はipynbなんですが、ipynbはれないので、中のコードだけバシバシはっておきます。
ipynbはsorafuneのディスコードに投下しておきます
To Do
モデルパラメータはすごく適当。FCやhidden sizeなど
学習のパラメータ。学習率やロスなどもわりとそのまま
入力の特徴量は2つのみ。もっといろいろ入れたら精度あがりそう
アンサンブルとかに使えるかも？