Create chronos-gradio.py

chronos-gradio.py

@@ -0,0 +1,59 @@
+import gradio as gr
+import torch
+from chronos import ChronosPipeline
+import numpy as np
+import pandas as pd
+
+# 从 Hugging Face 加载模型
+# model_name = "amazon/chronos-t5-small"  # 替换为你在 Hugging Face 上的模型名称
+# model = AutoModelForConditionalGeneration.from_pretrained(model_name)
+# model.eval()
+model = ChronosPipeline.from_pretrained(
+    "amazon/chronos-t5-small",
+    device_map="cuda",
+    torch_dtype=torch.bfloat16,
+)
+
+def predict_with_chronos(input_data):
+    prediction = model.predict(
+        context=input_data, 
+        prediction_length=24,
+        num_samples=1
+    )
+    return np.round(prediction.mean(axis=0).squeeze().cpu().numpy()).astype(int)
+
+def predict_from_csv(csv_file):
+    df = pd.read_csv(csv_file.name)
+    
+    raw_values = pd.to_numeric(df['value'], errors='coerce').dropna().values
+    print(raw_values)
+    print('输入数据长度为：',len(raw_values))
+    input_data = torch.tensor(
+        raw_values.astype(np.float32)
+    )
+    
+    predictions = predict_with_chronos(input_data)
+    predictions = np.asarray(predictions).ravel()
+
+    forecast_index = range(1, len(predictions)+1)
+    
+    assert len(forecast_index) == len(predictions), "数组长度不一致"
+    
+    output_df = pd.DataFrame({
+        'period': forecast_index,
+        'value': predictions 
+    })
+    
+    output_path = "/tmp/predictions.csv"
+    output_df.to_csv(output_path, index=False)
+    return output_path
+
+iface = gr.Interface(
+    fn=predict_from_csv,
+    inputs=gr.File(label="上传包含时序数据的 CSV 文件"),
+    outputs=gr.File(label="预测结果下载", file_count="single"),
+    title="Chronos时序预测",
+    description="上传包含时序数据的 CSV 文件，获取未来24步预测结果。"
+)
+
+iface.launch()