Spaces:
Running
Running
Update app.py
Browse files
app.py
CHANGED
|
@@ -1,5 +1,4 @@
|
|
| 1 |
import os
|
| 2 |
-
import inspect
|
| 3 |
import numpy as np
|
| 4 |
import pandas as pd
|
| 5 |
import gradio as gr
|
|
@@ -8,54 +7,29 @@ import torch
|
|
| 8 |
|
| 9 |
from chronos import Chronos2Pipeline
|
| 10 |
|
|
|
|
| 11 |
# =========================
|
| 12 |
# Config
|
| 13 |
# =========================
|
| 14 |
MODEL_ID_DEFAULT = os.getenv("CHRONOS_MODEL_ID", "amazon/chronos-2")
|
| 15 |
DATA_DIR = "data"
|
| 16 |
|
|
|
|
| 17 |
# =========================
|
| 18 |
-
#
|
| 19 |
# =========================
|
| 20 |
def available_test_csv():
|
| 21 |
if not os.path.isdir(DATA_DIR):
|
| 22 |
return []
|
| 23 |
return sorted(f for f in os.listdir(DATA_DIR) if f.lower().endswith(".csv"))
|
| 24 |
|
|
|
|
| 25 |
def pick_device(ui_choice: str) -> str:
|
| 26 |
if (ui_choice or "").startswith("cuda") and torch.cuda.is_available():
|
| 27 |
return "cuda"
|
| 28 |
return "cpu"
|
| 29 |
|
| 30 |
-
# =========================
|
| 31 |
-
# Model cache
|
| 32 |
-
# =========================
|
| 33 |
-
_PIPELINE = None
|
| 34 |
-
_PIPELINE_META = {}
|
| 35 |
|
| 36 |
-
def get_pipeline(model_id: str, device: str):
|
| 37 |
-
"""
|
| 38 |
-
Caches the pipeline across calls to avoid re-downloading and re-loading.
|
| 39 |
-
"""
|
| 40 |
-
global _PIPELINE, _PIPELINE_META
|
| 41 |
-
|
| 42 |
-
model_id = (model_id or MODEL_ID_DEFAULT).strip()
|
| 43 |
-
device = "cuda" if device == "cuda" and torch.cuda.is_available() else "cpu"
|
| 44 |
-
|
| 45 |
-
if (
|
| 46 |
-
_PIPELINE is None
|
| 47 |
-
or _PIPELINE_META.get("model_id") != model_id
|
| 48 |
-
or _PIPELINE_META.get("device") != device
|
| 49 |
-
):
|
| 50 |
-
# Chronos-2 pipeline
|
| 51 |
-
_PIPELINE = Chronos2Pipeline.from_pretrained(model_id, device_map=device)
|
| 52 |
-
_PIPELINE_META = {"model_id": model_id, "device": device}
|
| 53 |
-
|
| 54 |
-
return _PIPELINE
|
| 55 |
-
|
| 56 |
-
# =========================
|
| 57 |
-
# Data generation/loading
|
| 58 |
-
# =========================
|
| 59 |
def make_sample_series(n, seed, trend, season_period, season_amp, noise):
|
| 60 |
rng = np.random.default_rng(int(seed))
|
| 61 |
t = np.arange(int(n))
|
|
@@ -64,30 +38,30 @@ def make_sample_series(n, seed, trend, season_period, season_amp, noise):
|
|
| 64 |
+ float(season_amp) * np.sin(2 * np.pi * t / max(1, int(season_period)))
|
| 65 |
+ rng.normal(0.0, float(noise), size=len(t))
|
| 66 |
)
|
| 67 |
-
# shift up if negative
|
| 68 |
mn = float(np.min(y))
|
| 69 |
if mn < 0:
|
| 70 |
y = y - mn
|
| 71 |
return y.astype(np.float32)
|
| 72 |
|
|
|
|
| 73 |
def load_series_from_csv(path_or_file, column=None):
|
| 74 |
df = pd.read_csv(path_or_file)
|
| 75 |
-
|
| 76 |
if df.shape[1] == 0:
|
| 77 |
raise ValueError("CSV vuoto o non leggibile.")
|
| 78 |
|
| 79 |
col = (column or "").strip()
|
| 80 |
if col == "":
|
|
|
|
| 81 |
numeric_cols = [c for c in df.columns if pd.api.types.is_numeric_dtype(df[c])]
|
|
|
|
| 82 |
if not numeric_cols:
|
| 83 |
-
# try coercion to numeric on all columns (sometimes dtype is object)
|
| 84 |
-
numeric_cols = []
|
| 85 |
for c in df.columns:
|
| 86 |
coerced = pd.to_numeric(df[c], errors="coerce")
|
| 87 |
if coerced.notna().sum() >= 10:
|
| 88 |
numeric_cols.append(c)
|
| 89 |
-
|
| 90 |
-
|
| 91 |
col = numeric_cols[0]
|
| 92 |
|
| 93 |
if col not in df.columns:
|
|
@@ -99,76 +73,78 @@ def load_series_from_csv(path_or_file, column=None):
|
|
| 99 |
|
| 100 |
return y.astype(np.float32), col
|
| 101 |
|
|
|
|
| 102 |
# =========================
|
| 103 |
-
#
|
| 104 |
# =========================
|
| 105 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 106 |
"""
|
| 107 |
-
Chronos2Pipeline.
|
| 108 |
-
|
| 109 |
-
- dict with 'samples'
|
| 110 |
-
- object with attribute 'samples'
|
| 111 |
-
This returns np.ndarray of shape (n_draws, pred_len) or (pred_len,) if only one draw.
|
| 112 |
"""
|
| 113 |
-
|
| 114 |
-
|
| 115 |
-
|
| 116 |
-
|
| 117 |
-
|
| 118 |
-
if "samples" in pred_out:
|
| 119 |
-
return np.asarray(pred_out["samples"])
|
| 120 |
-
# sometimes "forecast" keys etc.
|
| 121 |
-
for k in ("predictions", "prediction", "outputs"):
|
| 122 |
-
if k in pred_out:
|
| 123 |
-
return np.asarray(pred_out[k])
|
| 124 |
-
return np.asarray(pred_out)
|
| 125 |
-
# object with samples attribute
|
| 126 |
-
if hasattr(pred_out, "samples"):
|
| 127 |
-
return np.asarray(getattr(pred_out, "samples"))
|
| 128 |
-
# last resort
|
| 129 |
-
return np.asarray(pred_out)
|
| 130 |
-
|
| 131 |
-
def chronos2_predict_samples(pipe, y, prediction_length: int, n_draws: int):
|
| 132 |
"""
|
| 133 |
-
|
| 134 |
-
|
| 135 |
-
- Uses `num_predictions=` if supported
|
| 136 |
-
- If not supported, falls back to a single prediction and returns shape (1, pred_len)
|
| 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 |
-
if samples.ndim == 1:
|
| 169 |
-
samples = samples[None, :]
|
| 170 |
-
|
| 171 |
-
return samples
|
| 172 |
|
| 173 |
# =========================
|
| 174 |
# Forecast core
|
|
@@ -185,33 +161,32 @@ def run_forecast(
|
|
| 185 |
season_amp,
|
| 186 |
noise,
|
| 187 |
prediction_length,
|
| 188 |
-
num_draws,
|
| 189 |
q_low,
|
| 190 |
q_high,
|
| 191 |
device_ui,
|
| 192 |
model_id,
|
| 193 |
):
|
| 194 |
-
|
| 195 |
-
|
|
|
|
| 196 |
raise gr.Error("Quantile low deve essere < quantile high.")
|
| 197 |
|
| 198 |
-
# Device + pipeline
|
| 199 |
device = pick_device(device_ui)
|
| 200 |
pipe = get_pipeline(model_id, device)
|
| 201 |
|
| 202 |
-
#
|
| 203 |
if input_mode == "Test CSV":
|
| 204 |
if not test_csv_name:
|
| 205 |
-
raise gr.Error("Seleziona un file nella dropdown dei Test CSV
|
| 206 |
-
|
| 207 |
-
if not os.path.exists(
|
| 208 |
-
raise gr.Error(f"Non trovo {
|
| 209 |
-
y, used_col = load_series_from_csv(
|
| 210 |
source = f"Test CSV: {test_csv_name} ({used_col})"
|
| 211 |
|
| 212 |
elif input_mode == "Upload CSV":
|
| 213 |
if upload_csv is None:
|
| 214 |
-
raise gr.Error("Carica un CSV
|
| 215 |
y, used_col = load_series_from_csv(upload_csv.name, csv_column)
|
| 216 |
source = f"Upload CSV ({used_col})"
|
| 217 |
|
|
@@ -219,27 +194,41 @@ def run_forecast(
|
|
| 219 |
y = make_sample_series(n, seed, trend, season_period, season_amp, noise)
|
| 220 |
source = "Sample data"
|
| 221 |
|
| 222 |
-
#
|
| 223 |
-
|
| 224 |
-
|
| 225 |
-
|
|
|
|
|
|
|
|
|
|
| 226 |
prediction_length=int(prediction_length),
|
| 227 |
-
|
|
|
|
|
|
|
|
|
|
| 228 |
)
|
| 229 |
|
| 230 |
-
#
|
| 231 |
-
|
| 232 |
-
|
| 233 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 234 |
|
| 235 |
-
#
|
| 236 |
t_hist = np.arange(len(y))
|
| 237 |
t_fcst = np.arange(len(y), len(y) + int(prediction_length))
|
| 238 |
|
| 239 |
fig, ax = plt.subplots(figsize=(10, 4))
|
| 240 |
ax.plot(t_hist, y, label="history")
|
| 241 |
ax.plot(t_fcst, median, label="forecast (median)")
|
| 242 |
-
ax.fill_between(t_fcst, low, high, alpha=0.25, label=f"band [{
|
| 243 |
ax.axvline(len(y) - 1, linestyle="--", linewidth=1)
|
| 244 |
ax.set_title(source)
|
| 245 |
ax.set_xlabel("t")
|
|
@@ -247,13 +236,14 @@ def run_forecast(
|
|
| 247 |
ax.grid(True, alpha=0.3)
|
| 248 |
ax.legend()
|
| 249 |
|
| 250 |
-
#
|
| 251 |
out_df = pd.DataFrame(
|
| 252 |
{
|
| 253 |
"t": t_fcst,
|
|
|
|
| 254 |
"median": median,
|
| 255 |
-
f"q{
|
| 256 |
-
f"q{
|
| 257 |
}
|
| 258 |
)
|
| 259 |
|
|
@@ -266,25 +256,27 @@ def run_forecast(
|
|
| 266 |
"source": source,
|
| 267 |
"history_points": int(len(y)),
|
| 268 |
"prediction_length": int(prediction_length),
|
| 269 |
-
"
|
| 270 |
-
"
|
| 271 |
}
|
| 272 |
|
| 273 |
return fig, out_df, out_path, info
|
| 274 |
|
|
|
|
| 275 |
# =========================
|
| 276 |
# UI
|
| 277 |
# =========================
|
| 278 |
with gr.Blocks(title="Chronos-2 • HF Spaces Demo") as demo:
|
| 279 |
-
gr.Markdown(
|
| 280 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 281 |
|
| 282 |
with gr.Row():
|
| 283 |
-
input_mode = gr.Radio(
|
| 284 |
-
["Sample", "Test CSV", "Upload CSV"],
|
| 285 |
-
value="Sample",
|
| 286 |
-
label="Input source",
|
| 287 |
-
)
|
| 288 |
device_ui = gr.Dropdown(
|
| 289 |
["cpu", "cuda (se disponibile)"],
|
| 290 |
value="cuda (se disponibile)" if torch.cuda.is_available() else "cpu",
|
|
@@ -293,10 +285,7 @@ with gr.Blocks(title="Chronos-2 • HF Spaces Demo") as demo:
|
|
| 293 |
model_id = gr.Textbox(value=MODEL_ID_DEFAULT, label="Model ID")
|
| 294 |
|
| 295 |
with gr.Row():
|
| 296 |
-
test_csv_name = gr.Dropdown(
|
| 297 |
-
choices=available_test_csv(),
|
| 298 |
-
label="Test CSV disponibili (cartella data/)",
|
| 299 |
-
)
|
| 300 |
upload_csv = gr.File(label="Upload CSV", file_types=[".csv"])
|
| 301 |
csv_column = gr.Textbox(label="Colonna numerica (opzionale)", placeholder="es: value")
|
| 302 |
|
|
@@ -310,8 +299,6 @@ with gr.Blocks(title="Chronos-2 • HF Spaces Demo") as demo:
|
|
| 310 |
|
| 311 |
with gr.Accordion("Forecast settings", open=True):
|
| 312 |
prediction_length = gr.Slider(1, 180, 30, step=1, label="Prediction length")
|
| 313 |
-
# UI label stays "Num samples", internally treated as number of prediction draws if supported
|
| 314 |
-
num_draws = gr.Slider(1, 400, 200, step=10, label="Num samples (draws)")
|
| 315 |
q_low = gr.Slider(0.01, 0.49, 0.10, step=0.01, label="Quantile low")
|
| 316 |
q_high = gr.Slider(0.51, 0.99, 0.90, step=0.01, label="Quantile high")
|
| 317 |
|
|
@@ -336,7 +323,6 @@ with gr.Blocks(title="Chronos-2 • HF Spaces Demo") as demo:
|
|
| 336 |
season_amp,
|
| 337 |
noise,
|
| 338 |
prediction_length,
|
| 339 |
-
num_draws,
|
| 340 |
q_low,
|
| 341 |
q_high,
|
| 342 |
device_ui,
|
|
|
|
| 1 |
import os
|
|
|
|
| 2 |
import numpy as np
|
| 3 |
import pandas as pd
|
| 4 |
import gradio as gr
|
|
|
|
| 7 |
|
| 8 |
from chronos import Chronos2Pipeline
|
| 9 |
|
| 10 |
+
|
| 11 |
# =========================
|
| 12 |
# Config
|
| 13 |
# =========================
|
| 14 |
MODEL_ID_DEFAULT = os.getenv("CHRONOS_MODEL_ID", "amazon/chronos-2")
|
| 15 |
DATA_DIR = "data"
|
| 16 |
|
| 17 |
+
|
| 18 |
# =========================
|
| 19 |
+
# Utils
|
| 20 |
# =========================
|
| 21 |
def available_test_csv():
|
| 22 |
if not os.path.isdir(DATA_DIR):
|
| 23 |
return []
|
| 24 |
return sorted(f for f in os.listdir(DATA_DIR) if f.lower().endswith(".csv"))
|
| 25 |
|
| 26 |
+
|
| 27 |
def pick_device(ui_choice: str) -> str:
|
| 28 |
if (ui_choice or "").startswith("cuda") and torch.cuda.is_available():
|
| 29 |
return "cuda"
|
| 30 |
return "cpu"
|
| 31 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 32 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 33 |
def make_sample_series(n, seed, trend, season_period, season_amp, noise):
|
| 34 |
rng = np.random.default_rng(int(seed))
|
| 35 |
t = np.arange(int(n))
|
|
|
|
| 38 |
+ float(season_amp) * np.sin(2 * np.pi * t / max(1, int(season_period)))
|
| 39 |
+ rng.normal(0.0, float(noise), size=len(t))
|
| 40 |
)
|
| 41 |
+
# shift up if negative to keep plots nice
|
| 42 |
mn = float(np.min(y))
|
| 43 |
if mn < 0:
|
| 44 |
y = y - mn
|
| 45 |
return y.astype(np.float32)
|
| 46 |
|
| 47 |
+
|
| 48 |
def load_series_from_csv(path_or_file, column=None):
|
| 49 |
df = pd.read_csv(path_or_file)
|
|
|
|
| 50 |
if df.shape[1] == 0:
|
| 51 |
raise ValueError("CSV vuoto o non leggibile.")
|
| 52 |
|
| 53 |
col = (column or "").strip()
|
| 54 |
if col == "":
|
| 55 |
+
# try native numeric dtypes first
|
| 56 |
numeric_cols = [c for c in df.columns if pd.api.types.is_numeric_dtype(df[c])]
|
| 57 |
+
# fallback: try coercion
|
| 58 |
if not numeric_cols:
|
|
|
|
|
|
|
| 59 |
for c in df.columns:
|
| 60 |
coerced = pd.to_numeric(df[c], errors="coerce")
|
| 61 |
if coerced.notna().sum() >= 10:
|
| 62 |
numeric_cols.append(c)
|
| 63 |
+
if not numeric_cols:
|
| 64 |
+
raise ValueError("Nessuna colonna numerica nel CSV. Specifica la colonna corretta.")
|
| 65 |
col = numeric_cols[0]
|
| 66 |
|
| 67 |
if col not in df.columns:
|
|
|
|
| 73 |
|
| 74 |
return y.astype(np.float32), col
|
| 75 |
|
| 76 |
+
|
| 77 |
# =========================
|
| 78 |
+
# Pipeline cache
|
| 79 |
# =========================
|
| 80 |
+
_PIPELINE = None
|
| 81 |
+
_PIPELINE_META = {}
|
| 82 |
+
|
| 83 |
+
|
| 84 |
+
def get_pipeline(model_id: str, device: str):
|
| 85 |
+
global _PIPELINE, _PIPELINE_META
|
| 86 |
+
|
| 87 |
+
model_id = (model_id or MODEL_ID_DEFAULT).strip()
|
| 88 |
+
device = "cuda" if device == "cuda" and torch.cuda.is_available() else "cpu"
|
| 89 |
+
|
| 90 |
+
if (
|
| 91 |
+
_PIPELINE is None
|
| 92 |
+
or _PIPELINE_META.get("model_id") != model_id
|
| 93 |
+
or _PIPELINE_META.get("device") != device
|
| 94 |
+
):
|
| 95 |
+
_PIPELINE = Chronos2Pipeline.from_pretrained(model_id, device_map=device)
|
| 96 |
+
_PIPELINE_META = {"model_id": model_id, "device": device}
|
| 97 |
+
|
| 98 |
+
return _PIPELINE
|
| 99 |
+
|
| 100 |
+
|
| 101 |
+
# =========================
|
| 102 |
+
# Chronos-2 predict_df helpers
|
| 103 |
+
# =========================
|
| 104 |
+
def build_context_df(y: np.ndarray, freq: str = "D"):
|
| 105 |
"""
|
| 106 |
+
Build a minimal context DataFrame compatible with Chronos2Pipeline.predict_df().
|
| 107 |
+
We generate a synthetic timestamp index so it works for Sample and numeric-only CSV.
|
|
|
|
|
|
|
|
|
|
| 108 |
"""
|
| 109 |
+
ts = pd.date_range("2000-01-01", periods=len(y), freq=freq)
|
| 110 |
+
return pd.DataFrame({"id": "series_0", "timestamp": ts, "target": y})
|
| 111 |
+
|
| 112 |
+
|
| 113 |
+
def pick_quantile_column(pred_df: pd.DataFrame, q: float) -> str:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 114 |
"""
|
| 115 |
+
Column naming can vary. We robustly find a column representing quantile q.
|
| 116 |
+
Common patterns: "0.1", "0.5", "0.9" OR "q0.1" OR "quantile_0.1" etc.
|
|
|
|
|
|
|
| 117 |
"""
|
| 118 |
+
q = float(q)
|
| 119 |
+
# direct numeric-string match
|
| 120 |
+
for c in pred_df.columns:
|
| 121 |
+
try:
|
| 122 |
+
if abs(float(c) - q) < 1e-9:
|
| 123 |
+
return c
|
| 124 |
+
except Exception:
|
| 125 |
+
pass
|
| 126 |
+
|
| 127 |
+
# prefixed patterns
|
| 128 |
+
candidates = []
|
| 129 |
+
for c in pred_df.columns:
|
| 130 |
+
lc = str(c).lower()
|
| 131 |
+
if "quant" in lc or lc.startswith("q"):
|
| 132 |
+
# try to extract float from tail
|
| 133 |
+
for token in [lc.replace("quantile", "").replace("_", ""), lc.replace("q", "")]:
|
| 134 |
+
try:
|
| 135 |
+
if abs(float(token) - q) < 1e-9:
|
| 136 |
+
candidates.append(c)
|
| 137 |
+
except Exception:
|
| 138 |
+
pass
|
| 139 |
+
|
| 140 |
+
if candidates:
|
| 141 |
+
return candidates[0]
|
| 142 |
+
|
| 143 |
+
raise ValueError(
|
| 144 |
+
f"Non riesco a trovare la colonna del quantile {q}. "
|
| 145 |
+
f"Colonne disponibili: {list(pred_df.columns)}"
|
| 146 |
+
)
|
| 147 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
| 148 |
|
| 149 |
# =========================
|
| 150 |
# Forecast core
|
|
|
|
| 161 |
season_amp,
|
| 162 |
noise,
|
| 163 |
prediction_length,
|
|
|
|
| 164 |
q_low,
|
| 165 |
q_high,
|
| 166 |
device_ui,
|
| 167 |
model_id,
|
| 168 |
):
|
| 169 |
+
q_low = float(q_low)
|
| 170 |
+
q_high = float(q_high)
|
| 171 |
+
if q_low >= q_high:
|
| 172 |
raise gr.Error("Quantile low deve essere < quantile high.")
|
| 173 |
|
|
|
|
| 174 |
device = pick_device(device_ui)
|
| 175 |
pipe = get_pipeline(model_id, device)
|
| 176 |
|
| 177 |
+
# 1) pick data
|
| 178 |
if input_mode == "Test CSV":
|
| 179 |
if not test_csv_name:
|
| 180 |
+
raise gr.Error("Seleziona un file nella dropdown dei Test CSV.")
|
| 181 |
+
path = os.path.join(DATA_DIR, test_csv_name)
|
| 182 |
+
if not os.path.exists(path):
|
| 183 |
+
raise gr.Error(f"Non trovo {path}. Assicurati che sia nel repo.")
|
| 184 |
+
y, used_col = load_series_from_csv(path, csv_column)
|
| 185 |
source = f"Test CSV: {test_csv_name} ({used_col})"
|
| 186 |
|
| 187 |
elif input_mode == "Upload CSV":
|
| 188 |
if upload_csv is None:
|
| 189 |
+
raise gr.Error("Carica un CSV per usare la modalità Upload.")
|
| 190 |
y, used_col = load_series_from_csv(upload_csv.name, csv_column)
|
| 191 |
source = f"Upload CSV ({used_col})"
|
| 192 |
|
|
|
|
| 194 |
y = make_sample_series(n, seed, trend, season_period, season_amp, noise)
|
| 195 |
source = "Sample data"
|
| 196 |
|
| 197 |
+
# 2) build context df (single series)
|
| 198 |
+
context_df = build_context_df(y, freq="D")
|
| 199 |
+
|
| 200 |
+
# 3) predict quantiles via predict_df (stable API per chronos-2)
|
| 201 |
+
quantiles = sorted({q_low, 0.5, q_high})
|
| 202 |
+
pred_df = pipe.predict_df(
|
| 203 |
+
context_df,
|
| 204 |
prediction_length=int(prediction_length),
|
| 205 |
+
quantile_levels=quantiles,
|
| 206 |
+
id_column="id",
|
| 207 |
+
timestamp_column="timestamp",
|
| 208 |
+
target="target",
|
| 209 |
)
|
| 210 |
|
| 211 |
+
# 4) extract arrays
|
| 212 |
+
col_low = pick_quantile_column(pred_df, q_low)
|
| 213 |
+
col_med = pick_quantile_column(pred_df, 0.5)
|
| 214 |
+
col_high = pick_quantile_column(pred_df, q_high)
|
| 215 |
+
|
| 216 |
+
# pred_df contains the forecast horizon rows; keep only series_0
|
| 217 |
+
pred_df = pred_df[pred_df["id"] == "series_0"].copy()
|
| 218 |
+
|
| 219 |
+
ts_fcst = pd.to_datetime(pred_df["timestamp"]).to_numpy()
|
| 220 |
+
low = pred_df[col_low].to_numpy(dtype=np.float32)
|
| 221 |
+
median = pred_df[col_med].to_numpy(dtype=np.float32)
|
| 222 |
+
high = pred_df[col_high].to_numpy(dtype=np.float32)
|
| 223 |
|
| 224 |
+
# 5) plot (use integer axis for simplicity)
|
| 225 |
t_hist = np.arange(len(y))
|
| 226 |
t_fcst = np.arange(len(y), len(y) + int(prediction_length))
|
| 227 |
|
| 228 |
fig, ax = plt.subplots(figsize=(10, 4))
|
| 229 |
ax.plot(t_hist, y, label="history")
|
| 230 |
ax.plot(t_fcst, median, label="forecast (median)")
|
| 231 |
+
ax.fill_between(t_fcst, low, high, alpha=0.25, label=f"band [{q_low:.2f}, {q_high:.2f}]")
|
| 232 |
ax.axvline(len(y) - 1, linestyle="--", linewidth=1)
|
| 233 |
ax.set_title(source)
|
| 234 |
ax.set_xlabel("t")
|
|
|
|
| 236 |
ax.grid(True, alpha=0.3)
|
| 237 |
ax.legend()
|
| 238 |
|
| 239 |
+
# 6) output table + downloadable csv
|
| 240 |
out_df = pd.DataFrame(
|
| 241 |
{
|
| 242 |
"t": t_fcst,
|
| 243 |
+
"timestamp": ts_fcst,
|
| 244 |
"median": median,
|
| 245 |
+
f"q{q_low:.2f}": low,
|
| 246 |
+
f"q{q_high:.2f}": high,
|
| 247 |
}
|
| 248 |
)
|
| 249 |
|
|
|
|
| 256 |
"source": source,
|
| 257 |
"history_points": int(len(y)),
|
| 258 |
"prediction_length": int(prediction_length),
|
| 259 |
+
"quantile_levels": quantiles,
|
| 260 |
+
"pred_df_columns": list(out_df.columns),
|
| 261 |
}
|
| 262 |
|
| 263 |
return fig, out_df, out_path, info
|
| 264 |
|
| 265 |
+
|
| 266 |
# =========================
|
| 267 |
# UI
|
| 268 |
# =========================
|
| 269 |
with gr.Blocks(title="Chronos-2 • HF Spaces Demo") as demo:
|
| 270 |
+
gr.Markdown(
|
| 271 |
+
"# ⏱️ Chronos-2 Forecast Demo (HF Spaces)\n"
|
| 272 |
+
"- **Sample**: genera una serie sintetica\n"
|
| 273 |
+
"- **Test CSV**: usa file in `data/`\n"
|
| 274 |
+
"- **Upload CSV**: carica un tuo CSV\n\n"
|
| 275 |
+
"Questa versione usa **predict_df()** (API consigliata per Chronos-2) e calcola direttamente i **quantili**. "
|
| 276 |
+
)
|
| 277 |
|
| 278 |
with gr.Row():
|
| 279 |
+
input_mode = gr.Radio(["Sample", "Test CSV", "Upload CSV"], value="Sample", label="Input source")
|
|
|
|
|
|
|
|
|
|
|
|
|
| 280 |
device_ui = gr.Dropdown(
|
| 281 |
["cpu", "cuda (se disponibile)"],
|
| 282 |
value="cuda (se disponibile)" if torch.cuda.is_available() else "cpu",
|
|
|
|
| 285 |
model_id = gr.Textbox(value=MODEL_ID_DEFAULT, label="Model ID")
|
| 286 |
|
| 287 |
with gr.Row():
|
| 288 |
+
test_csv_name = gr.Dropdown(choices=available_test_csv(), label="Test CSV disponibili (data/)")
|
|
|
|
|
|
|
|
|
|
| 289 |
upload_csv = gr.File(label="Upload CSV", file_types=[".csv"])
|
| 290 |
csv_column = gr.Textbox(label="Colonna numerica (opzionale)", placeholder="es: value")
|
| 291 |
|
|
|
|
| 299 |
|
| 300 |
with gr.Accordion("Forecast settings", open=True):
|
| 301 |
prediction_length = gr.Slider(1, 180, 30, step=1, label="Prediction length")
|
|
|
|
|
|
|
| 302 |
q_low = gr.Slider(0.01, 0.49, 0.10, step=0.01, label="Quantile low")
|
| 303 |
q_high = gr.Slider(0.51, 0.99, 0.90, step=0.01, label="Quantile high")
|
| 304 |
|
|
|
|
| 323 |
season_amp,
|
| 324 |
noise,
|
| 325 |
prediction_length,
|
|
|
|
| 326 |
q_low,
|
| 327 |
q_high,
|
| 328 |
device_ui,
|