-
Notifications
You must be signed in to change notification settings - Fork 0
/
predict.py
200 lines (153 loc) · 7.56 KB
/
predict.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
import pandas as pd
import argparse
import numpy as np
from typing import Any
import torch
from model import LSTMModel
import pickle
from dataloader import TimeSeriesDataset, SolarDataset
from torch.utils.data import DataLoader, Dataset
import glob
import os
import logging
import warnings
from sklearn.exceptions import InconsistentVersionWarning
from sklearn.metrics import mean_absolute_error, mean_squared_error
import math
logging.basicConfig(format='%(asctime)s - %(levelname)s - %(name)s - %(message)s', datefmt='%m/%d/%Y %H:%M:%S',
level=logging.INFO)
logger = logging.getLogger(__name__)
LOAD_CKPT = "./load/best_results/model_20230807_165319.pt"
LOAD_HP = "./load/best_results/20230807_165319.pkl"
LOAD_TEST = "./load/data/merged_data_KW.xlsx"
PV_CKPT = "./pv/checkpoints/*.pt"
PV_HP = "./pv/results/*.pkl"
PV_TEST = "./pv/eval_sequence/*.csv"
DEVICE = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
TARGET_DAY ='0831'
parser = argparse.ArgumentParser(description="Prediction for PV and LOAD")
parser.add_argument('--mode', type=str, required=True,
help="Only 'pv' and 'load' are accepted. Should not be empty.")
args = parser.parse_args()
def load_checkpoint(ckpt_path: str, hp_path: str):
with open(hp_path, 'rb') as f:
results = pickle.load(f)
hyperparameters = results['Hyperparameters']
config = {
'input_size': 17 if args.mode=='pv' else 7,
'hidden_size': hyperparameters['hidden_size'] if args.mode=='pv' else hyperparameters['hidden_dim'],
'num_layers': hyperparameters['num_layers'] if args.mode=='pv' else hyperparameters['n_layers'],
'output_size': 24 if args.mode=='pv' else 24*56
}
model = LSTMModel(
config['input_size'],
config['hidden_size'],
config['num_layers'],
config['output_size']
).to(DEVICE)
checkpoint = torch.load(ckpt_path, map_location=DEVICE)
model.load_state_dict(checkpoint['model_state_dict'] if args.mode=='pv' else checkpoint) # else model.load_state_dict(torch.load(ckpt_path, map_location=torch.device('cpu')))
return model
def pv_predict(datapath: str, model):
model.eval()
df = pd.read_csv(datapath)
dataset = SolarDataset(df)
eval_loader = DataLoader(dataset=dataset, batch_size=1, shuffle=False)
for seq, label in eval_loader:
pred = model(seq.to(DEVICE))
prediction_array = pred.detach().cpu().numpy()
target_array = label.detach().cpu().numpy()
prediction = dataset.y_scaler.inverse_transform(prediction_array)
target = dataset.y_scaler.inverse_transform(target_array)
rounded_data = np.maximum(prediction, 0.0).reshape(-1,1)
return rounded_data.flatten(), target.flatten(), prediction_array.flatten(), target_array.flatten()
def load_predict(datapath: str, model):
df = pd.read_excel(datapath)
dataset = TimeSeriesDataset(df)
test_loader = DataLoader(dataset=dataset, batch_size=1, shuffle=False)
real_sequence, real_target = next(iter(test_loader))
# Assuming df is the original dataset and it includes a 'date' column
building_names = df.columns[-56:]
# 모델을 평가 모드로 전환합니다.
model.eval()
# 예측을 수행합니다.
with torch.no_grad():
prediction = model(real_sequence.to(DEVICE)).detach().cpu()
# 성능 지표를 계산합니다.
mae = mean_absolute_error(real_target, prediction)
mse = mean_squared_error(real_target, prediction)
rmse = math.sqrt(mse)
#print(prediction.shape, real_target.shape)
prediction = prediction.squeeze(0).reshape(24, 56).numpy()
real_target = real_target.view(24, 56).numpy()
# 패딩을 추가합니다.
padding = np.zeros((prediction.shape[0], 7))
prediction_pad = np.hstack((padding, prediction))
real_target_pad = np.hstack((padding, real_target))
# print(prediction_pad.shape, real_target_pad.shape)
# 역변환을 적용하여 정규화를 해제합니다.
prediction_inv = dataset.scaler.inverse_transform(prediction_pad)
real_target_inv = dataset.scaler.inverse_transform(real_target_pad)
# 처음 7개의 컬럼을 삭제합니다.
prediction_inv = np.delete(prediction_inv, np.s_[:7], axis=1)
real_target_inv = np.delete(real_target_inv, np.s_[:7], axis=1)
# 원래의 형태로 다시 변형합니다.
prediction = prediction_inv.reshape(prediction.shape)
real_target = real_target_inv.reshape(real_target.shape)
# 에러(실제 목표값과 예측값의 차이)를 계산합니다.
error = real_target - prediction
# 예측값을 위한 DataFrame을 생성합니다.
predicted_df = pd.DataFrame(prediction, columns=building_names)
real_target_df = pd.DataFrame(real_target, columns=building_names)
error_df = pd.DataFrame(error, columns=building_names)
err_total = error_df.values.flatten().sum()
# 성능 지표를 계산합니다.
mae_n = mean_absolute_error(real_target, prediction)
mse_n = mean_squared_error(real_target, prediction)
rmse_n = math.sqrt(mse_n)
logger.info(f'[Unnormalized] MAE: {mae_n:.4f}, MSE: {mse_n:.4f}, RMSE: {rmse_n:.4f}')
logger.info(f'[Normalized] MAE: {mae:.4f}, MSE: {mse:.4f}, RMSE: {rmse:.4f}')
logger.info(f"Total Prediction Error: {err_total}")
# Save to Excel file
output_filepath = f'./{args.mode}_predict_for_'+TARGET_DAY+'.xlsx' # adjust this as necessary
predicted_df.to_excel(output_filepath, index=False)
logger.info(f"Prediction completed. filepath: {output_filepath}")
if __name__=='__main__':
warnings.filterwarnings("ignore", category=InconsistentVersionWarning)
"""
PV_pred: load_checkpoint, pv_predict INSIDE the loop
LOAD_pred: load_checkpoint, load_predict at once
"""
logger.info("Predicting {} for {}...".format(args.mode.upper(), TARGET_DAY))
if args.mode == 'load':
model = load_checkpoint(LOAD_CKPT, LOAD_HP)
load_predict(LOAD_TEST, model)
elif args.mode == 'pv':
models = dict()
predictions = dict()
target = dict()
predictions_norm = dict()
target_norm = dict()
for ckpt, hp, datapath in zip(sorted(glob.glob(PV_CKPT)), sorted(glob.glob(PV_HP)), sorted(glob.glob(PV_TEST))):
b = ckpt.split("/")[-1].split("-")[0]
models[b] = load_checkpoint(ckpt, hp)
predictions[b], target[b], predictions_norm[b], target_norm[b] = pv_predict(datapath, models[b])
pred = pd.DataFrame(predictions)
real = pd.DataFrame(target)
pred_n = pd.DataFrame(predictions_norm)
real_n = pd.DataFrame(target_norm)
mae = mean_absolute_error(real, pred)
mse = mean_squared_error(real, pred)
rmse = math.sqrt(mse)
mae_n = mean_absolute_error(real_n, pred_n)
mse_n = mean_squared_error(real_n, pred_n)
rmse_n = math.sqrt(mse_n)
error = pd.DataFrame(target)-pd.DataFrame(predictions)
logger.info(f"[Unnormalized] MAE: {mae:.4f}, MSE: {mse:.4f}, RMSE: {rmse:.4f}")
logger.info(f'[Normalized] MAE: {mae_n:.4f}, MSE: {mse_n:.4f}, RMSE: {rmse_n:.4f}')
logger.info(f"Total Prediction Error: {error.sum().sum()}")
output_filepath = os.path.join('.',f'{args.mode}_predict_for_'+TARGET_DAY+'.csv')
pd.DataFrame(predictions).to_csv(output_filepath, index=False)
logger.info(f"Prediction completed. filepath: {output_filepath}")
else:
raise ValueError("--mode excepts either 'pv' or 'load'")