FilesExpand file tree

master

/

prediction_batch_example.py

Copy path

Blame

More file actions

Blame

More file actions

Latest commit

 

History

History

History

192 lines (166 loc) · 6.96 KB

master

/

prediction_batch_example.py

Copy path

Top

File metadata and controls

• Code
• Blame

192 lines (166 loc) · 6.96 KB

Raw

Copy raw file

Download raw file

Open symbols panel

Edit and raw actions

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

import argparse

import time

from pathlib import Path

import matplotlib.pyplot as plt

import pandas as pd

from common import data_path, detect_device, get_model_spec, output_path, save_json

from model import Kronos, KronosPredictor, KronosTokenizer

LOOKBACK = 400

PRED_LEN = 120

DEFAULT_BATCH_SIZE = 5

def plot_prediction(kline_df: pd.DataFrame, pred_df: pd.DataFrame, plot_file: Path) -> None:

plot_df = pred_df.copy()

plot_df.index = kline_df.index[-plot_df.shape[0]:]

close_df = pd.concat(

[

kline_df["close"].rename("Ground Truth"),

plot_df["close"].rename("Prediction"),

],

axis=1,

)

volume_df = pd.concat(

[

kline_df["volume"].rename("Ground Truth"),

plot_df["volume"].rename("Prediction"),

],

axis=1,

)

fig, (ax1, ax2) = plt.subplots(2, 1, figsize=(8, 6), sharex=True)

ax1.plot(close_df["Ground Truth"], label="Ground Truth", color="blue", linewidth=1.5)

ax1.plot(close_df["Prediction"], label="Prediction", color="red", linewidth=1.5)

ax1.set_ylabel("Close Price", fontsize=14)

ax1.legend(loc="lower left", fontsize=12)

ax1.grid(True)

ax2.plot(volume_df["Ground Truth"], label="Ground Truth", color="blue", linewidth=1.5)

ax2.plot(volume_df["Prediction"], label="Prediction", color="red", linewidth=1.5)

ax2.set_ylabel("Volume", fontsize=14)

ax2.legend(loc="upper left", fontsize=12)

ax2.grid(True)

plt.tight_layout()

plt.savefig(plot_file, dpi=150)

plt.close(fig)

def main() -> None:

parser = argparse.ArgumentParser(

description="Run batch prediction on multiple time windows from the same instrument."

)

parser.add_argument("--model", choices=["small", "base"], default="small", help="Model size to run.")

parser.add_argument(

"--batch-size",

type=int,

default=DEFAULT_BATCH_SIZE,

help="Number of sequential time windows to batch together.",

)

parser.add_argument(

"--plot-series",

type=int,

default=0,

help="Which window index to render as the representative plot.",

)

args = parser.parse_args()

if args.batch_size < 1:

raise ValueError("--batch-size must be at least 1.")

if args.plot_series < 0:

raise ValueError("--plot-series must be greater than or equal to 0.")

spec = get_model_spec(args.model)

device = detect_device()

data_file = data_path("XSHG_5min_600977.csv")

summary_file = output_path(f"prediction_batch_example_{args.model}_summary.json")

print(

"Running batch prediction on multiple sequential windows from the same source file, "

"not on multiple instruments."

)

print(f"Loading {spec['name']} on device: {device}")

start_time = time.perf_counter()

tokenizer = KronosTokenizer.from_pretrained(spec["tokenizer_id"])

model = Kronos.from_pretrained(spec["model_id"])

predictor = KronosPredictor(model, tokenizer, device=device, max_context=spec["max_context"])

df = pd.read_csv(data_file)

df["timestamps"] = pd.to_datetime(df["timestamps"])

max_windows = (len(df) - PRED_LEN) // LOOKBACK

if max_windows < 1:

raise ValueError(

f"Not enough rows in {data_file} for one window; need at least {LOOKBACK + PRED_LEN} rows."

)

if args.batch_size > max_windows:

raise ValueError(

f"--batch-size={args.batch_size} exceeds the available sequential windows ({max_windows}) for "

f"{len(df)} rows. Reduce the batch size or use a longer dataset."

)

if args.plot_series >= args.batch_size:

raise ValueError("--plot-series must be smaller than --batch-size.")

df_list = []

x_timestamp_list = []

y_timestamp_list = []

window_summaries = []

for idx in range(args.batch_size):

start = idx * LOOKBACK

end = start + LOOKBACK - 1

pred_start = end + 1

pred_end = pred_start + PRED_LEN - 1

df_list.append(df.loc[start:end, ["open", "high", "low", "close", "volume", "amount"]])

x_timestamp_list.append(df.loc[start:end, "timestamps"])

y_timestamp_list.append(df.loc[pred_start:pred_end, "timestamps"])

window_summaries.append(

{

"series_index": idx,

"instrument_note": "Same instrument, different historical time window.",

"history_row_start": start,

"history_row_end": end,

"forecast_row_start": pred_start,

"forecast_row_end": pred_end,

"history_start": df.loc[start, "timestamps"],

"history_end": df.loc[end, "timestamps"],

"forecast_start": df.loc[pred_start, "timestamps"],

"forecast_end": df.loc[pred_end, "timestamps"],

}

)

pred_df_list = predictor.predict_batch(

df_list=df_list,

x_timestamp_list=x_timestamp_list,

y_timestamp_list=y_timestamp_list,

pred_len=PRED_LEN,

verbose=False,

)

elapsed_seconds = time.perf_counter() - start_time

print(f"Generated batch predictions: {len(pred_df_list)} windows")

representative_plot = None

for idx, pred_df in enumerate(pred_df_list):

csv_file = output_path(f"prediction_batch_example_{args.model}_series_{idx}.csv")

pred_df.to_csv(csv_file, index_label="timestamps")

window_summaries[idx]["forecast_csv"] = csv_file

print(f"Series {idx} forecast head:")

print(pred_df.head())

print(f"Saved series {idx} CSV: {csv_file}")

if idx == args.plot_series:

plot_file = output_path(f"prediction_batch_example_{args.model}_representative_series_{idx}.png")

kline_df = df.loc[idx * LOOKBACK : idx * LOOKBACK + LOOKBACK + PRED_LEN - 1]

plot_prediction(kline_df, pred_df, plot_file)

representative_plot = plot_file

window_summaries[idx]["representative_plot"] = plot_file

print(f"Saved representative plot for series {idx}: {plot_file}")

save_json(

summary_file,

{

"experiment": "prediction_batch_example",

"description": (

"Batch prediction on multiple sequential time windows from a single instrument. "

"This example demonstrates parallel inference, not cross-sectional multi-asset prediction."

),

"model_key": args.model,

"model_name": spec["name"],

"model_id": spec["model_id"],

"tokenizer_id": spec["tokenizer_id"],

"device": device,

"elapsed_seconds": elapsed_seconds,

"lookback": LOOKBACK,

"pred_len": PRED_LEN,

"batch_size": args.batch_size,

"plot_series": args.plot_series,

"data_file": data_file,

"representative_plot": representative_plot,

"windows": window_summaries,

},

)

print(f"Saved summary JSON: {summary_file}")

if __name__ == "__main__":

main()