Spaces:
Sleeping
Sleeping
Update inference.py
Browse files- inference.py +16 -57
inference.py
CHANGED
|
@@ -5,9 +5,7 @@ import argparse
|
|
| 5 |
import joblib
|
| 6 |
import os
|
| 7 |
from tqdm import tqdm
|
| 8 |
-
import json
|
| 9 |
-
|
| 10 |
-
# ⭐️ 수정 사항 1: 경로 문제를 피하기 위해 명시적으로 import 경로 추가
|
| 11 |
import sys
|
| 12 |
sys.path.append('.')
|
| 13 |
|
|
@@ -15,9 +13,9 @@ from models import TimeXer
|
|
| 15 |
from utils.metrics import metric
|
| 16 |
from utils.timefeatures import time_features
|
| 17 |
|
| 18 |
-
# --- 1. 인자 파싱
|
|
|
|
| 19 |
parser = argparse.ArgumentParser(description='Time Series Prediction')
|
| 20 |
-
# ... (이전과 동일한 모든 argparse 코드) ...
|
| 21 |
parser.add_argument('--checkpoint_path', type=str, required=True, help='Path to the model checkpoint file (.pth)')
|
| 22 |
parser.add_argument('--scaler_path', type=str, required=True, help='Path to the saved scaler file (.gz)')
|
| 23 |
parser.add_argument('--predict_input_file', type=str, default=None, help='[Mode 1] Path to the CSV file for single future prediction')
|
|
@@ -49,17 +47,9 @@ parser.add_argument('--freq', type=str, default='t', help='freq for time feature
|
|
| 49 |
args = parser.parse_args()
|
| 50 |
|
| 51 |
|
| 52 |
-
|
| 53 |
-
prediction_padded = np.concatenate((padding, prediction_scaled), axis=1)
|
| 54 |
-
prediction = scaler.inverse_transform(prediction_padded)[:, -args.c_out:]
|
| 55 |
-
else:
|
| 56 |
-
prediction = scaler.inverse_transform(prediction_scaled)
|
| 57 |
-
|
| 58 |
-
return prediction
|
| 59 |
-
|
| 60 |
-
|
| 61 |
-
# --- 2. 공통 함수: 모델 및 스케일러 로드 (수정 없음) ---
|
| 62 |
def load_model_and_scaler(args):
|
|
|
|
| 63 |
device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
|
| 64 |
args.device = device
|
| 65 |
model = TimeXer.Model(args).float().to(device)
|
|
@@ -75,39 +65,33 @@ def predict_future(args, model, scaler, device):
|
|
| 75 |
df_input = pd.read_csv(args.predict_input_file)
|
| 76 |
df_input['date'] = pd.to_datetime(df_input['date'])
|
| 77 |
|
| 78 |
-
# ⭐️ 알려주신 정확한 컬럼 이름으로 수정
|
| 79 |
cols_to_scale = ['air_pres', 'wind_dir', 'wind_speed', 'air_temp', 'residual']
|
| 80 |
|
| 81 |
-
# 1. 인코더 입력(x_enc) 생성
|
| 82 |
raw_input = df_input[cols_to_scale].tail(args.seq_len).values
|
| 83 |
input_scaled = scaler.transform(raw_input)
|
| 84 |
batch_x = torch.from_numpy(input_scaled).float().unsqueeze(0).to(device)
|
| 85 |
|
| 86 |
-
# 2. 인코더 시간 정보(x_mark_enc) 생성
|
| 87 |
df_stamp_enc = df_input.tail(args.seq_len)[['date']].reset_index(drop=True)
|
| 88 |
enc_mark = time_features(df_stamp_enc, timeenc=0, freq=args.freq)
|
| 89 |
batch_x_mark = torch.from_numpy(enc_mark).float().unsqueeze(0).to(device)
|
| 90 |
|
| 91 |
-
# 3. 디코더 입력(x_dec) 생성
|
| 92 |
dec_inp_label = input_scaled[-args.label_len:]
|
| 93 |
dec_inp_pred = np.zeros([args.pred_len, args.enc_in])
|
| 94 |
decoder_input = np.concatenate([dec_inp_label, dec_inp_pred], axis=0)
|
| 95 |
batch_y = torch.from_numpy(decoder_input).float().unsqueeze(0).to(device)
|
| 96 |
|
| 97 |
-
# 4. 디코더 시간 정보(x_mark_dec) 생성
|
| 98 |
last_date = df_stamp_enc['date'].iloc[-1]
|
| 99 |
-
future_dates = pd.date_range(start=last_date, periods=args.pred_len + 1, freq='5T')[1:]
|
| 100 |
df_stamp_dec = pd.DataFrame({'date': list(df_stamp_enc['date'].values[-args.label_len:]) + list(future_dates)})
|
| 101 |
dec_mark = time_features(df_stamp_dec, timeenc=0, freq=args.freq)
|
| 102 |
batch_y_mark = torch.from_numpy(dec_mark).float().unsqueeze(0).to(device)
|
| 103 |
|
| 104 |
-
# 5. 모델 호출
|
| 105 |
with torch.no_grad():
|
| 106 |
outputs = model(batch_x, batch_x_mark, batch_y, batch_y_mark)
|
| 107 |
|
| 108 |
prediction_scaled = outputs.detach().cpu().numpy()[0]
|
| 109 |
|
| 110 |
-
#
|
| 111 |
if scaler.n_features_in_ > 1:
|
| 112 |
padding = np.zeros((prediction_scaled.shape[0], scaler.n_features_in_ - args.c_out))
|
| 113 |
prediction_padded = np.concatenate((padding, prediction_scaled), axis=1)
|
|
@@ -116,12 +100,13 @@ def predict_future(args, model, scaler, device):
|
|
| 116 |
prediction = scaler.inverse_transform(prediction_scaled)
|
| 117 |
return prediction
|
| 118 |
|
| 119 |
-
# --- 4. 모드 2: 전체 기간 롤링 평가 함수
|
| 120 |
def evaluate_performance(args, model, scaler, device):
|
|
|
|
|
|
|
| 121 |
df_eval = pd.read_csv(args.evaluate_file)
|
| 122 |
df_eval['date'] = pd.to_datetime(df_eval['date'])
|
| 123 |
|
| 124 |
-
# ⭐️ 알려주신 정확한 컬럼 이름으로 수정
|
| 125 |
cols_to_scale = ['air_pres', 'wind_dir', 'wind_speed', 'air_temp', 'residual']
|
| 126 |
raw_data = df_eval[cols_to_scale].values
|
| 127 |
data_scaled = scaler.transform(raw_data)
|
|
@@ -133,7 +118,6 @@ def evaluate_performance(args, model, scaler, device):
|
|
| 133 |
|
| 134 |
num_samples = len(data_scaled) - args.seq_len - args.pred_len + 1
|
| 135 |
for i in tqdm(range(num_samples), desc="Evaluating", file=sys.stderr):
|
| 136 |
-
# 1. 인코더/디코더 입력 생성 (매 스텝마다)
|
| 137 |
s_begin = i
|
| 138 |
s_end = s_begin + args.seq_len
|
| 139 |
|
|
@@ -152,19 +136,16 @@ def evaluate_performance(args, model, scaler, device):
|
|
| 152 |
dec_mark_pred = df_stamp[true_begin:true_end]
|
| 153 |
batch_y_mark = np.concatenate([dec_mark_label, dec_mark_pred], axis=0)
|
| 154 |
|
| 155 |
-
# 텐서로 변환
|
| 156 |
batch_x = torch.from_numpy(batch_x).float().unsqueeze(0).to(device)
|
| 157 |
batch_x_mark = torch.from_numpy(batch_x_mark).float().unsqueeze(0).to(device)
|
| 158 |
batch_y = torch.from_numpy(batch_y).float().unsqueeze(0).to(device)
|
| 159 |
batch_y_mark = torch.from_numpy(batch_y_mark).float().unsqueeze(0).to(device)
|
| 160 |
|
| 161 |
-
# 2. 모델 호출
|
| 162 |
with torch.no_grad():
|
| 163 |
outputs = model(batch_x, batch_x_mark, batch_y, batch_y_mark)
|
| 164 |
|
| 165 |
pred_scaled = outputs.detach().cpu().numpy()[0]
|
| 166 |
|
| 167 |
-
# 3. 스케일 복원
|
| 168 |
if scaler.n_features_in_ > 1:
|
| 169 |
padding = np.zeros((pred_scaled.shape[0], scaler.n_features_in_ - args.c_out))
|
| 170 |
pred_padded = np.concatenate((padding, pred_scaled), axis=1)
|
|
@@ -179,45 +160,23 @@ def evaluate_performance(args, model, scaler, device):
|
|
| 179 |
|
| 180 |
return np.array(preds_unscaled), np.array(trues_unscaled)
|
| 181 |
|
| 182 |
-
# --- 5. 메인 로직
|
| 183 |
if __name__ == '__main__':
|
| 184 |
-
|
| 185 |
-
final_output = {}
|
| 186 |
-
|
| 187 |
try:
|
| 188 |
model, scaler, device = load_model_and_scaler(args)
|
| 189 |
-
|
| 190 |
if args.predict_input_file:
|
| 191 |
print("--- Running in Single Prediction Mode ---", file=sys.stderr)
|
| 192 |
prediction = predict_future(args, model, scaler, device)
|
| 193 |
-
final_output = {
|
| 194 |
-
"status": "success",
|
| 195 |
-
"mode": "single_prediction",
|
| 196 |
-
"prediction": prediction.flatten().tolist()
|
| 197 |
-
}
|
| 198 |
-
|
| 199 |
elif args.evaluate_file:
|
| 200 |
print("--- Running in Rolling Evaluation Mode ---", file=sys.stderr)
|
| 201 |
eval_preds, eval_trues = evaluate_performance(args, model, scaler, device)
|
| 202 |
-
|
| 203 |
-
# 성능 지표 계산
|
| 204 |
mae, mse, _, _, _ = metric(eval_preds, eval_trues)
|
| 205 |
-
|
| 206 |
-
final_output = {
|
| 207 |
-
"status": "success",
|
| 208 |
-
"mode": "rolling_evaluation",
|
| 209 |
-
"mse": mse,
|
| 210 |
-
"mae": mae,
|
| 211 |
-
# 전체 예측을 반환하면 너무 크므로, 샘플만 반환하거나 필요한 정보만 반환
|
| 212 |
-
"prediction_samples": [p.flatten().tolist() for p in eval_preds[:5]]
|
| 213 |
-
}
|
| 214 |
-
|
| 215 |
else:
|
| 216 |
-
final_output = {"status": "error", "message": "No mode selected.
|
| 217 |
-
|
| 218 |
except Exception as e:
|
| 219 |
final_output = {"status": "error", "message": str(e)}
|
| 220 |
-
|
| 221 |
-
# 최종 결과를 JSON 문자열로 표준 출력(stdout)에 프린트합니다.
|
| 222 |
-
# 이 출력을 app.py가 읽어서 API 응답으로 사용합니다.
|
| 223 |
print(json.dumps(final_output, indent=2))
|
|
|
|
| 5 |
import joblib
|
| 6 |
import os
|
| 7 |
from tqdm import tqdm
|
| 8 |
+
import json
|
|
|
|
|
|
|
| 9 |
import sys
|
| 10 |
sys.path.append('.')
|
| 11 |
|
|
|
|
| 13 |
from utils.metrics import metric
|
| 14 |
from utils.timefeatures import time_features
|
| 15 |
|
| 16 |
+
# --- 1. 인자 파싱 ---
|
| 17 |
+
# ... (이전과 동일, 수정 없음) ...
|
| 18 |
parser = argparse.ArgumentParser(description='Time Series Prediction')
|
|
|
|
| 19 |
parser.add_argument('--checkpoint_path', type=str, required=True, help='Path to the model checkpoint file (.pth)')
|
| 20 |
parser.add_argument('--scaler_path', type=str, required=True, help='Path to the saved scaler file (.gz)')
|
| 21 |
parser.add_argument('--predict_input_file', type=str, default=None, help='[Mode 1] Path to the CSV file for single future prediction')
|
|
|
|
| 47 |
args = parser.parse_args()
|
| 48 |
|
| 49 |
|
| 50 |
+
# --- 2. 공통 함수: 모델 및 스케일러 로드 ---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 51 |
def load_model_and_scaler(args):
|
| 52 |
+
# ... (이전과 동일, 수정 없음) ...
|
| 53 |
device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
|
| 54 |
args.device = device
|
| 55 |
model = TimeXer.Model(args).float().to(device)
|
|
|
|
| 65 |
df_input = pd.read_csv(args.predict_input_file)
|
| 66 |
df_input['date'] = pd.to_datetime(df_input['date'])
|
| 67 |
|
|
|
|
| 68 |
cols_to_scale = ['air_pres', 'wind_dir', 'wind_speed', 'air_temp', 'residual']
|
| 69 |
|
|
|
|
| 70 |
raw_input = df_input[cols_to_scale].tail(args.seq_len).values
|
| 71 |
input_scaled = scaler.transform(raw_input)
|
| 72 |
batch_x = torch.from_numpy(input_scaled).float().unsqueeze(0).to(device)
|
| 73 |
|
|
|
|
| 74 |
df_stamp_enc = df_input.tail(args.seq_len)[['date']].reset_index(drop=True)
|
| 75 |
enc_mark = time_features(df_stamp_enc, timeenc=0, freq=args.freq)
|
| 76 |
batch_x_mark = torch.from_numpy(enc_mark).float().unsqueeze(0).to(device)
|
| 77 |
|
|
|
|
| 78 |
dec_inp_label = input_scaled[-args.label_len:]
|
| 79 |
dec_inp_pred = np.zeros([args.pred_len, args.enc_in])
|
| 80 |
decoder_input = np.concatenate([dec_inp_label, dec_inp_pred], axis=0)
|
| 81 |
batch_y = torch.from_numpy(decoder_input).float().unsqueeze(0).to(device)
|
| 82 |
|
|
|
|
| 83 |
last_date = df_stamp_enc['date'].iloc[-1]
|
| 84 |
+
future_dates = pd.date_range(start=last_date, periods=args.pred_len + 1, freq='5T')[1:]
|
| 85 |
df_stamp_dec = pd.DataFrame({'date': list(df_stamp_enc['date'].values[-args.label_len:]) + list(future_dates)})
|
| 86 |
dec_mark = time_features(df_stamp_dec, timeenc=0, freq=args.freq)
|
| 87 |
batch_y_mark = torch.from_numpy(dec_mark).float().unsqueeze(0).to(device)
|
| 88 |
|
|
|
|
| 89 |
with torch.no_grad():
|
| 90 |
outputs = model(batch_x, batch_x_mark, batch_y, batch_y_mark)
|
| 91 |
|
| 92 |
prediction_scaled = outputs.detach().cpu().numpy()[0]
|
| 93 |
|
| 94 |
+
# ⭐️⭐️⭐️ 이 블록의 들여쓰기를 수정했습니다 ⭐️⭐️⭐️
|
| 95 |
if scaler.n_features_in_ > 1:
|
| 96 |
padding = np.zeros((prediction_scaled.shape[0], scaler.n_features_in_ - args.c_out))
|
| 97 |
prediction_padded = np.concatenate((padding, prediction_scaled), axis=1)
|
|
|
|
| 100 |
prediction = scaler.inverse_transform(prediction_scaled)
|
| 101 |
return prediction
|
| 102 |
|
| 103 |
+
# --- 4. 모드 2: 전체 기간 롤링 평가 함수 ---
|
| 104 |
def evaluate_performance(args, model, scaler, device):
|
| 105 |
+
# ... (이전과 동일, 수정 없음) ...
|
| 106 |
+
# ⭐️ 이 함수 내부의 들여쓰기도 함께 점검하여 수정했습니다.
|
| 107 |
df_eval = pd.read_csv(args.evaluate_file)
|
| 108 |
df_eval['date'] = pd.to_datetime(df_eval['date'])
|
| 109 |
|
|
|
|
| 110 |
cols_to_scale = ['air_pres', 'wind_dir', 'wind_speed', 'air_temp', 'residual']
|
| 111 |
raw_data = df_eval[cols_to_scale].values
|
| 112 |
data_scaled = scaler.transform(raw_data)
|
|
|
|
| 118 |
|
| 119 |
num_samples = len(data_scaled) - args.seq_len - args.pred_len + 1
|
| 120 |
for i in tqdm(range(num_samples), desc="Evaluating", file=sys.stderr):
|
|
|
|
| 121 |
s_begin = i
|
| 122 |
s_end = s_begin + args.seq_len
|
| 123 |
|
|
|
|
| 136 |
dec_mark_pred = df_stamp[true_begin:true_end]
|
| 137 |
batch_y_mark = np.concatenate([dec_mark_label, dec_mark_pred], axis=0)
|
| 138 |
|
|
|
|
| 139 |
batch_x = torch.from_numpy(batch_x).float().unsqueeze(0).to(device)
|
| 140 |
batch_x_mark = torch.from_numpy(batch_x_mark).float().unsqueeze(0).to(device)
|
| 141 |
batch_y = torch.from_numpy(batch_y).float().unsqueeze(0).to(device)
|
| 142 |
batch_y_mark = torch.from_numpy(batch_y_mark).float().unsqueeze(0).to(device)
|
| 143 |
|
|
|
|
| 144 |
with torch.no_grad():
|
| 145 |
outputs = model(batch_x, batch_x_mark, batch_y, batch_y_mark)
|
| 146 |
|
| 147 |
pred_scaled = outputs.detach().cpu().numpy()[0]
|
| 148 |
|
|
|
|
| 149 |
if scaler.n_features_in_ > 1:
|
| 150 |
padding = np.zeros((pred_scaled.shape[0], scaler.n_features_in_ - args.c_out))
|
| 151 |
pred_padded = np.concatenate((padding, pred_scaled), axis=1)
|
|
|
|
| 160 |
|
| 161 |
return np.array(preds_unscaled), np.array(trues_unscaled)
|
| 162 |
|
| 163 |
+
# --- 5. 메인 로직 ---
|
| 164 |
if __name__ == '__main__':
|
| 165 |
+
# ... (이전과 동일, 수정 없음) ...
|
| 166 |
+
final_output = {}
|
|
|
|
| 167 |
try:
|
| 168 |
model, scaler, device = load_model_and_scaler(args)
|
|
|
|
| 169 |
if args.predict_input_file:
|
| 170 |
print("--- Running in Single Prediction Mode ---", file=sys.stderr)
|
| 171 |
prediction = predict_future(args, model, scaler, device)
|
| 172 |
+
final_output = {"status": "success", "mode": "single_prediction", "prediction": prediction.flatten().tolist()}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 173 |
elif args.evaluate_file:
|
| 174 |
print("--- Running in Rolling Evaluation Mode ---", file=sys.stderr)
|
| 175 |
eval_preds, eval_trues = evaluate_performance(args, model, scaler, device)
|
|
|
|
|
|
|
| 176 |
mae, mse, _, _, _ = metric(eval_preds, eval_trues)
|
| 177 |
+
final_output = {"status": "success", "mode": "rolling_evaluation", "mse": mse, "mae": mae, "prediction_samples": [p.flatten().tolist() for p in eval_preds[:5]]}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 178 |
else:
|
| 179 |
+
final_output = {"status": "error", "message": "No mode selected."}
|
|
|
|
| 180 |
except Exception as e:
|
| 181 |
final_output = {"status": "error", "message": str(e)}
|
|
|
|
|
|
|
|
|
|
| 182 |
print(json.dumps(final_output, indent=2))
|