Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
|
@@ -1,276 +1,220 @@
|
|
| 1 |
-
from time import time
|
| 2 |
-
|
| 3 |
-
import numpy as np
|
| 4 |
-
import pandas as pd
|
| 5 |
-
import plotly.express as px
|
| 6 |
-
import plotly.graph_objects as go
|
| 7 |
import streamlit as st
|
| 8 |
-
|
| 9 |
-
|
| 10 |
-
|
| 11 |
-
from
|
| 12 |
-
from
|
| 13 |
-
|
| 14 |
-
|
| 15 |
-
|
| 16 |
-
|
| 17 |
-
|
| 18 |
-
|
|
|
|
|
|
|
|
|
|
| 19 |
}
|
| 20 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 21 |
|
| 22 |
-
|
| 23 |
-
|
| 24 |
-
|
| 25 |
-
|
| 26 |
-
|
| 27 |
-
|
| 28 |
-
|
| 29 |
-
figs = []
|
| 30 |
-
figs += [
|
| 31 |
-
go.Scatter(
|
| 32 |
-
x=df["ds"],
|
| 33 |
-
y=df["y"],
|
| 34 |
-
mode="lines",
|
| 35 |
-
marker=dict(color="#236796"),
|
| 36 |
-
legendrank=1,
|
| 37 |
-
name=uid,
|
| 38 |
-
),
|
| 39 |
-
]
|
| 40 |
-
if df_forecast is not None:
|
| 41 |
-
ds_f = df_forecast["ds"].to_list()
|
| 42 |
-
lo = df_forecast["forecast_lo_90"].to_list()
|
| 43 |
-
hi = df_forecast["forecast_hi_90"].to_list()
|
| 44 |
-
figs += [
|
| 45 |
-
go.Scatter(
|
| 46 |
-
x=ds_f + ds_f[::-1],
|
| 47 |
-
y=hi + lo[::-1],
|
| 48 |
-
fill="toself",
|
| 49 |
-
fillcolor="#E7C4C0",
|
| 50 |
-
mode="lines",
|
| 51 |
-
line=dict(color="#E7C4C0"),
|
| 52 |
-
name="Prediction Intervals (90%)",
|
| 53 |
-
legendrank=5,
|
| 54 |
-
opacity=0.5,
|
| 55 |
-
hoverinfo="skip",
|
| 56 |
-
),
|
| 57 |
-
go.Scatter(
|
| 58 |
-
x=ds_f,
|
| 59 |
-
y=df_forecast["forecast"],
|
| 60 |
-
mode="lines",
|
| 61 |
-
legendrank=4,
|
| 62 |
-
marker=dict(color="#E7C4C0"),
|
| 63 |
-
name=f"Forecast {uid}",
|
| 64 |
-
),
|
| 65 |
-
]
|
| 66 |
-
fig = go.Figure(figs)
|
| 67 |
-
fig.update_layout(
|
| 68 |
-
{"plot_bgcolor": "rgba(0, 0, 0, 0)", "paper_bgcolor": "rgba(0, 0, 0, 0)"}
|
| 69 |
-
)
|
| 70 |
-
fig.update_layout(
|
| 71 |
-
title=f"Forecasts for {uid} using Transfer Learning (from {model})",
|
| 72 |
-
legend=dict(orientation="h", yanchor="bottom", y=1.02, xanchor="right", x=1),
|
| 73 |
-
margin=dict(l=20, b=20),
|
| 74 |
-
xaxis=dict(rangeslider=dict(visible=True)),
|
| 75 |
-
)
|
| 76 |
-
initial_range = [df.tail(200)["ds"].iloc[0], ds_f[-1]]
|
| 77 |
-
fig["layout"]["xaxis"].update(range=initial_range)
|
| 78 |
-
return fig
|
| 79 |
-
|
| 80 |
-
|
| 81 |
-
def st_transfer_learning():
|
| 82 |
-
st.set_page_config(
|
| 83 |
-
page_title="Time Series Forecasting",
|
| 84 |
-
page_icon="🔮",
|
| 85 |
-
layout="wide",
|
| 86 |
-
initial_sidebar_state="expanded",
|
| 87 |
-
)
|
| 88 |
-
|
| 89 |
-
st.title(
|
| 90 |
-
"Transfer Learning: Revolutionizing Time Series"
|
| 91 |
-
)
|
| 92 |
-
st.write(
|
| 93 |
-
"<style>div.block-container{padding-top:2rem;}</style>", unsafe_allow_html=True
|
| 94 |
-
)
|
| 95 |
-
|
| 96 |
-
intro = """
|
| 97 |
-
The success of startups like Open AI and Stability highlights the potential for transfer learning (TL) techniques to have a similar impact on the field of time series forecasting.
|
| 98 |
-
TL can achieve lightning-fast predictions with a fraction of the computational cost by pre-training a flexible model on a large dataset and then using it on another dataset with little to no additional training.
|
| 99 |
-
In this live demo, you can use pre-trained models by Nixtla (trained on the M4 dataset) to predict your own datasets. You can also see how the models perform on unseen example datasets.
|
| 100 |
-
"""
|
| 101 |
-
st.write(intro)
|
| 102 |
-
|
| 103 |
-
required_cols = ["ds", "y"]
|
| 104 |
-
|
| 105 |
-
with st.sidebar.expander("Dataset", expanded=False):
|
| 106 |
-
data_selection = st.selectbox("Select example dataset", DATASETS.keys())
|
| 107 |
-
data_url = DATASETS[data_selection]
|
| 108 |
-
url_json = st.text_input("Data (you can pass your own url here)", data_url)
|
| 109 |
-
st.write(
|
| 110 |
-
"You can also upload a CSV file like [this one](https://github.com/Nixtla/transfer-learning-time-series/blob/main/datasets/air_passengers.csv)."
|
| 111 |
-
)
|
| 112 |
-
|
| 113 |
-
uploaded_file = st.file_uploader("Upload CSV")
|
| 114 |
-
with st.form("Data"):
|
| 115 |
-
|
| 116 |
-
if uploaded_file is not None:
|
| 117 |
-
df = pd.read_csv(uploaded_file)
|
| 118 |
-
cols = df.columns
|
| 119 |
-
timestamp_col = st.selectbox("Timestamp column", options=cols)
|
| 120 |
-
value_col = st.selectbox("Value column", options=cols)
|
| 121 |
-
else:
|
| 122 |
-
timestamp_col = st.text_input("Timestamp column", value="timestamp")
|
| 123 |
-
value_col = st.text_input("Value column", value="value")
|
| 124 |
-
st.write("You must press Submit each time you want to forecast.")
|
| 125 |
-
submitted = st.form_submit_button("Submit")
|
| 126 |
-
if submitted:
|
| 127 |
-
if uploaded_file is None:
|
| 128 |
-
st.write("Please provide a dataframe.")
|
| 129 |
-
if url_json.endswith("json"):
|
| 130 |
-
df = pd.read_json(url_json)
|
| 131 |
-
else:
|
| 132 |
-
df = pd.read_csv(url_json)
|
| 133 |
-
df = df.rename(
|
| 134 |
-
columns=dict(zip([timestamp_col, value_col], required_cols))
|
| 135 |
-
)
|
| 136 |
-
else:
|
| 137 |
-
df = df.rename(
|
| 138 |
-
columns=dict(zip([timestamp_col, value_col], required_cols))
|
| 139 |
-
)
|
| 140 |
-
else:
|
| 141 |
-
if url_json.endswith("json"):
|
| 142 |
-
df = pd.read_json(url_json)
|
| 143 |
-
else:
|
| 144 |
-
df = pd.read_csv(url_json)
|
| 145 |
-
cols = df.columns
|
| 146 |
-
if "unique_id" in cols:
|
| 147 |
-
cols = cols[-2:]
|
| 148 |
-
df = df.rename(columns=dict(zip(cols, required_cols)))
|
| 149 |
-
|
| 150 |
-
if "unique_id" not in df:
|
| 151 |
-
df.insert(0, "unique_id", "ts_0")
|
| 152 |
-
|
| 153 |
-
df["ds"] = pd.to_datetime(df["ds"])
|
| 154 |
-
df = df.sort_values(["unique_id", "ds"])
|
| 155 |
-
|
| 156 |
-
with st.sidebar:
|
| 157 |
-
st.write("Define the pretrained model you want to use to forecast your data")
|
| 158 |
-
model_name = st.selectbox("Select your model", tuple(MODELS.keys()))
|
| 159 |
-
model_file = MODELS[model_name]["model"]
|
| 160 |
-
st.write("Choose how many steps you want to forecast")
|
| 161 |
-
fh = st.number_input("Forecast horizon", value=18)
|
| 162 |
-
st.write(
|
| 163 |
-
"Choose for how many steps the pretrained model will be updated using your data (use 0 for fast computation)"
|
| 164 |
-
)
|
| 165 |
-
max_steps = st.number_input("N-shot inference", value=0)
|
| 166 |
-
|
| 167 |
-
# tabs
|
| 168 |
-
tab_fcst, tab_cv, tab_docs = st.tabs(
|
| 169 |
-
[
|
| 170 |
-
"📈 Forecast",
|
| 171 |
-
"🔎 Cross Validation",
|
| 172 |
-
"📚 Documentation",
|
| 173 |
-
]
|
| 174 |
-
)
|
| 175 |
-
|
| 176 |
-
uids = df["unique_id"].unique()
|
| 177 |
-
fcst_cols = ["forecast_lo_90", "forecast", "forecast_hi_90"]
|
| 178 |
-
|
| 179 |
-
with tab_fcst:
|
| 180 |
-
uid = uids[0]
|
| 181 |
-
col1, col2 = st.columns([2, 4])
|
| 182 |
-
with col1:
|
| 183 |
-
tab_insample, tab_forecast = st.tabs(
|
| 184 |
-
["Modify input data", "Modify forecasts"]
|
| 185 |
-
)
|
| 186 |
-
with tab_insample:
|
| 187 |
-
df_grid = df.query("unique_id == @uid").drop(columns="unique_id")
|
| 188 |
-
grid_table = AgGrid(
|
| 189 |
-
df_grid,
|
| 190 |
-
editable=True,
|
| 191 |
-
theme="alpine",
|
| 192 |
-
fit_columns_on_grid_load=True,
|
| 193 |
-
height=360,
|
| 194 |
-
)
|
| 195 |
-
df.loc[df["unique_id"] == uid, "y"] = (
|
| 196 |
-
grid_table["data"].sort_values("ds")["y"].values
|
| 197 |
-
)
|
| 198 |
-
# forecast code
|
| 199 |
-
init = time()
|
| 200 |
-
df_forecast = forecast_pretrained_model(df, model_file, fh, max_steps)
|
| 201 |
-
end = time()
|
| 202 |
-
df_forecast = df_forecast.rename(
|
| 203 |
-
columns=dict(zip(["y_5", "y_50", "y_95"], fcst_cols))
|
| 204 |
-
)
|
| 205 |
-
with tab_forecast:
|
| 206 |
-
df_fcst_grid = df_forecast.query("unique_id == @uid").filter(
|
| 207 |
-
["ds", "forecast"]
|
| 208 |
-
)
|
| 209 |
-
grid_fcst_table = AgGrid(
|
| 210 |
-
df_fcst_grid,
|
| 211 |
-
editable=True,
|
| 212 |
-
theme="alpine",
|
| 213 |
-
fit_columns_on_grid_load=True,
|
| 214 |
-
height=360,
|
| 215 |
-
)
|
| 216 |
-
changes = (
|
| 217 |
-
df_forecast.query("unique_id == @uid")["forecast"].values
|
| 218 |
-
- grid_fcst_table["data"].sort_values("ds")["forecast"].values
|
| 219 |
-
)
|
| 220 |
-
for col in fcst_cols:
|
| 221 |
-
df_forecast.loc[df_forecast["unique_id"] == uid, col] = (
|
| 222 |
-
df_forecast.loc[df_forecast["unique_id"] == uid, col] - changes
|
| 223 |
-
)
|
| 224 |
-
with col2:
|
| 225 |
-
if uploaded_file is not None:
|
| 226 |
-
fct_name = value_col
|
| 227 |
-
else:
|
| 228 |
-
fct_name=uid
|
| 229 |
-
st.plotly_chart(
|
| 230 |
-
plot(
|
| 231 |
-
df.query("unique_id == @uid"),
|
| 232 |
-
fct_name,
|
| 233 |
-
df_forecast.query("unique_id == @uid"),
|
| 234 |
-
model_file,
|
| 235 |
-
),
|
| 236 |
-
use_container_width=True,
|
| 237 |
-
)
|
| 238 |
-
st.write(f"Done in: {np.round(end-init, 2)} secs.")
|
| 239 |
-
st.write(f"Forecast for {fh} steps ahead.")
|
| 240 |
-
st.write("You can download the forecast for the entire dataframe here:")
|
| 241 |
-
csv = convert_df(
|
| 242 |
-
df_forecast[["unique_id", "ds"] + fcst_cols].sort_values(
|
| 243 |
-
["unique_id", "ds"]
|
| 244 |
-
)
|
| 245 |
-
)
|
| 246 |
-
st.download_button(
|
| 247 |
-
label="Download CSV",
|
| 248 |
-
data=csv,
|
| 249 |
-
file_name="forecast.csv",
|
| 250 |
-
mime="text/csv",
|
| 251 |
-
)
|
| 252 |
-
st.write(df_forecast[["unique_id", "ds"] + fcst_cols].tail(10))
|
| 253 |
-
|
| 254 |
-
with tab_cv:
|
| 255 |
-
st.write(
|
| 256 |
-
"To enable Cross Validation, use the advanced forecasting tool at our [site](https://nixtla.github.io/transfer-learning-time-series/)."
|
| 257 |
-
)
|
| 258 |
-
df_forecast_cv = None
|
| 259 |
|
| 260 |
-
|
| 261 |
-
|
| 262 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 263 |
|
| 264 |
-
|
| 265 |
-
|
| 266 |
-
|
| 267 |
-
|
| 268 |
-
|
| 269 |
-
|
| 270 |
-
|
| 271 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 272 |
)
|
| 273 |
-
|
| 274 |
-
|
| 275 |
-
|
| 276 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
import streamlit as st
|
| 2 |
+
import pandas as pd
|
| 3 |
+
import matplotlib.pyplot as plt
|
| 4 |
+
from neuralforecast.core import NeuralForecast
|
| 5 |
+
from neuralforecast.models import NHITS, TimesNet, LSTM, TFT
|
| 6 |
+
from neuralforecast.losses.pytorch import HuberMQLoss
|
| 7 |
+
import time
|
| 8 |
+
|
| 9 |
+
# Paths for saving models
|
| 10 |
+
nhits_paths = {
|
| 11 |
+
'D': './M4/NHITS/daily',
|
| 12 |
+
'M': './M4/NHITS/monthly',
|
| 13 |
+
'H': './M4/NHITS/hourly',
|
| 14 |
+
'W': './M4/NHITS/weekly',
|
| 15 |
+
'Y': './M4/NHITS/yearly'
|
| 16 |
}
|
| 17 |
|
| 18 |
+
timesnet_paths = {
|
| 19 |
+
'D': './M4/TimesNet/daily',
|
| 20 |
+
'M': './M4/TimesNet/monthly',
|
| 21 |
+
'H': './M4/TimesNet/hourly',
|
| 22 |
+
'W': './M4/TimesNet/weekly',
|
| 23 |
+
'Y': './M4/TimesNet/yearly'
|
| 24 |
+
}
|
| 25 |
|
| 26 |
+
lstm_paths = {
|
| 27 |
+
'D': './M4/LSTM/daily',
|
| 28 |
+
'M': './M4/LSTM/monthly',
|
| 29 |
+
'H': './M4/LSTM/hourly',
|
| 30 |
+
'W': './M4/LSTM/weekly',
|
| 31 |
+
'Y': './M4/LSTM/yearly'
|
| 32 |
+
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 33 |
|
| 34 |
+
tft_paths = {
|
| 35 |
+
'D': './M4/TFT/daily',
|
| 36 |
+
'M': './M4/TFT/monthly',
|
| 37 |
+
'H': './M4/TFT/hourly',
|
| 38 |
+
'W': './M4/TFT/weekly',
|
| 39 |
+
'Y': './M4/TFT/yearly'
|
| 40 |
+
}
|
| 41 |
|
| 42 |
+
@st.cache_resource
|
| 43 |
+
def load_model(path, freq):
|
| 44 |
+
nf = NeuralForecast.load(path=path)
|
| 45 |
+
return nf
|
| 46 |
+
|
| 47 |
+
nhits_models = {freq: load_model(path, freq) for freq, path in nhits_paths.items()}
|
| 48 |
+
timesnet_models = {freq: load_model(path, freq) for freq, path in timesnet_paths.items()}
|
| 49 |
+
lstm_models = {freq: load_model(path, freq) for freq, path in lstm_paths.items()}
|
| 50 |
+
tft_models = {freq: load_model(path, freq) for freq, path in tft_paths.items()}
|
| 51 |
+
|
| 52 |
+
def generate_forecast(model, df):
|
| 53 |
+
forecast_df = model.predict(df=df)
|
| 54 |
+
return forecast_df
|
| 55 |
+
|
| 56 |
+
def determine_frequency(df):
|
| 57 |
+
df['ds'] = pd.to_datetime(df['ds'])
|
| 58 |
+
df = df.set_index('ds')
|
| 59 |
+
freq = pd.infer_freq(df.index)
|
| 60 |
+
return freq
|
| 61 |
+
|
| 62 |
+
def plot_forecasts(forecast_df, train_df, title):
|
| 63 |
+
fig, ax = plt.subplots(1, 1, figsize=(20, 7))
|
| 64 |
+
plot_df = pd.concat([train_df, forecast_df]).set_index('ds')
|
| 65 |
+
historical_col = 'y'
|
| 66 |
+
forecast_col = next((col for col in plot_df.columns if 'median' in col), None)
|
| 67 |
+
lo_col = next((col for col in plot_df.columns if 'lo-90' in col), None)
|
| 68 |
+
hi_col = next((col for col in plot_df.columns if 'hi-90' in col), None)
|
| 69 |
+
if forecast_col is None:
|
| 70 |
+
raise KeyError("No forecast column found in the data.")
|
| 71 |
+
plot_df[[historical_col, forecast_col]].plot(ax=ax, linewidth=2, label=['Historical', 'Forecast'])
|
| 72 |
+
if lo_col and hi_col:
|
| 73 |
+
ax.fill_between(
|
| 74 |
+
plot_df.index,
|
| 75 |
+
plot_df[lo_col],
|
| 76 |
+
plot_df[hi_col],
|
| 77 |
+
color='blue',
|
| 78 |
+
alpha=0.3,
|
| 79 |
+
label='90% Confidence Interval'
|
| 80 |
)
|
| 81 |
+
ax.set_title(title, fontsize=22)
|
| 82 |
+
ax.set_ylabel('Value', fontsize=20)
|
| 83 |
+
ax.set_xlabel('Timestamp [t]', fontsize=20)
|
| 84 |
+
ax.legend(prop={'size': 15})
|
| 85 |
+
ax.grid()
|
| 86 |
+
st.pyplot(fig)
|
| 87 |
+
|
| 88 |
+
def select_model_based_on_frequency(freq, nhits_models, timesnet_models, lstm_models, tft_models):
|
| 89 |
+
if freq == 'D':
|
| 90 |
+
return nhits_models['D'], timesnet_models['D'], lstm_models['D'], tft_models['D']
|
| 91 |
+
elif freq == 'M':
|
| 92 |
+
return nhits_models['M'], timesnet_models['M'], lstm_models['M'], tft_models['M']
|
| 93 |
+
elif freq == 'H':
|
| 94 |
+
return nhits_models['H'], timesnet_models['H'], lstm_models['H'], tft_models['H']
|
| 95 |
+
elif freq in ['W', 'W-SUN']:
|
| 96 |
+
return nhits_models['W'], timesnet_models['W'], lstm_models['W'], tft_models['W']
|
| 97 |
+
elif freq in ['Y', 'Y-DEC']:
|
| 98 |
+
return nhits_models['Y'], timesnet_models['Y'], lstm_models['Y'], tft_models['Y']
|
| 99 |
+
else:
|
| 100 |
+
raise ValueError(f"Unsupported frequency: {freq}")
|
| 101 |
+
|
| 102 |
+
def select_model(horizon, model_type, max_steps=200):
|
| 103 |
+
if model_type == 'NHITS':
|
| 104 |
+
return NHITS(input_size=5 * horizon,
|
| 105 |
+
h=horizon,
|
| 106 |
+
max_steps=max_steps,
|
| 107 |
+
stack_types=3*['identity'],
|
| 108 |
+
n_blocks=3*[1],
|
| 109 |
+
mlp_units=[[256, 256] for _ in range(3)],
|
| 110 |
+
n_pool_kernel_size=3*[1],
|
| 111 |
+
batch_size=32,
|
| 112 |
+
scaler_type='standard',
|
| 113 |
+
n_freq_downsample=[12, 4, 1],
|
| 114 |
+
loss=HuberMQLoss(level=[90]))
|
| 115 |
+
elif model_type == 'TimesNet':
|
| 116 |
+
return TimesNet(h=horizon,
|
| 117 |
+
input_size=horizon * 5,
|
| 118 |
+
hidden_size=16,
|
| 119 |
+
conv_hidden_size=32,
|
| 120 |
+
loss=HuberMQLoss(level=[90]),
|
| 121 |
+
scaler_type='standard',
|
| 122 |
+
learning_rate=1e-3,
|
| 123 |
+
max_steps=max_steps,
|
| 124 |
+
val_check_steps=200,
|
| 125 |
+
valid_batch_size=64,
|
| 126 |
+
windows_batch_size=128,
|
| 127 |
+
inference_windows_batch_size=512)
|
| 128 |
+
elif model_type == 'LSTM':
|
| 129 |
+
return LSTM(h=horizon,
|
| 130 |
+
input_size=horizon * 5,
|
| 131 |
+
loss=HuberMQLoss(level=[90]),
|
| 132 |
+
scaler_type='standard',
|
| 133 |
+
encoder_n_layers=2,
|
| 134 |
+
encoder_hidden_size=64,
|
| 135 |
+
context_size=10,
|
| 136 |
+
decoder_hidden_size=64,
|
| 137 |
+
decoder_layers=2,
|
| 138 |
+
max_steps=max_steps)
|
| 139 |
+
elif model_type == 'TFT':
|
| 140 |
+
return TFT(h=horizon,
|
| 141 |
+
input_size=horizon,
|
| 142 |
+
hidden_size=16,
|
| 143 |
+
loss=HuberMQLoss(level=[90]),
|
| 144 |
+
learning_rate=0.005,
|
| 145 |
+
scaler_type='standard',
|
| 146 |
+
windows_batch_size=128,
|
| 147 |
+
max_steps=max_steps,
|
| 148 |
+
val_check_steps=200,
|
| 149 |
+
valid_batch_size=64,
|
| 150 |
+
enable_progress_bar=True)
|
| 151 |
+
else:
|
| 152 |
+
raise ValueError(f"Unsupported model type: {model_type}")
|
| 153 |
+
|
| 154 |
+
def forecast_time_series(df, model_type, freq, horizon, max_steps=200):
|
| 155 |
+
start_time = time.time() # Start timing
|
| 156 |
+
if freq:
|
| 157 |
+
df['ds'] = pd.date_range(start='1970-01-01', periods=len(df), freq=freq)
|
| 158 |
+
else:
|
| 159 |
+
freq = determine_frequency(df)
|
| 160 |
+
st.write(f"Determined frequency: {freq}")
|
| 161 |
+
df['ds'] = pd.to_datetime(df['ds'], errors='coerce')
|
| 162 |
+
df = df.dropna(subset=['ds'])
|
| 163 |
+
model = select_model(horizon, model_type, max_steps)
|
| 164 |
+
forecast_results = {}
|
| 165 |
+
st.write(f"Generating forecast using {model_type} model...")
|
| 166 |
+
forecast_results[model_type] = generate_forecast(model, df, freq)
|
| 167 |
+
|
| 168 |
+
for model_name, forecast_df in forecast_results.items():
|
| 169 |
+
plot_forecasts(forecast_df, df, f'{model_name} Forecast Comparison')
|
| 170 |
+
|
| 171 |
+
end_time = time.time() # End timing
|
| 172 |
+
time_taken = end_time - start_time
|
| 173 |
+
st.success(f"Time taken for {model_type} forecast: {time_taken:.2f} seconds")
|
| 174 |
+
|
| 175 |
+
# Streamlit App
|
| 176 |
+
st.title("Dynamic and Automatic Time Series Forecasting")
|
| 177 |
+
|
| 178 |
+
# Upload dataset
|
| 179 |
+
uploaded_file = st.file_uploader("Upload your time series data (CSV)", type=["csv"])
|
| 180 |
+
if uploaded_file:
|
| 181 |
+
df = pd.read_csv(uploaded_file)
|
| 182 |
+
else:
|
| 183 |
+
st.warning("Using default data")
|
| 184 |
+
df = AirPassengersDF.copy()
|
| 185 |
+
|
| 186 |
+
# Model selection and forecasting
|
| 187 |
+
st.subheader("Transfer Learning Forecasting")
|
| 188 |
+
model_choice = st.selectbox("Select model", ["NHITS", "TimesNet", "LSTM", "TFT"])
|
| 189 |
+
horizon = st.slider("Forecast horizon", 1, 100, 10)
|
| 190 |
+
|
| 191 |
+
# Determine frequency of data
|
| 192 |
+
frequency = determine_frequency(df)
|
| 193 |
+
st.write(f"Detected frequency: {frequency}")
|
| 194 |
+
|
| 195 |
+
# Load pre-trained models
|
| 196 |
+
nhits_model, timesnet_model, lstm_model, tft_model = select_model_based_on_frequency(frequency, nhits_models, timesnet_models, lstm_models, tft_models)
|
| 197 |
+
forecast_results = {}
|
| 198 |
+
|
| 199 |
+
start_time = time.time() # Start timing
|
| 200 |
+
if model_choice == "NHITS":
|
| 201 |
+
forecast_results['NHITS'] = generate_forecast(nhits_model, df)
|
| 202 |
+
elif model_choice == "TimesNet":
|
| 203 |
+
forecast_results['TimesNet'] = generate_forecast(timesnet_model, df)
|
| 204 |
+
elif model_choice == "LSTM":
|
| 205 |
+
forecast_results['LSTM'] = generate_forecast(lstm_model, df)
|
| 206 |
+
elif model_choice == "TFT":
|
| 207 |
+
forecast_results['TFT'] = generate_forecast(tft_model, df)
|
| 208 |
+
|
| 209 |
+
for model_name, forecast_df in forecast_results.items():
|
| 210 |
+
plot_forecasts(forecast_df, df, f'{model_name} Forecast')
|
| 211 |
+
|
| 212 |
+
end_time = time.time() # End timing
|
| 213 |
+
time_taken = end_time - start_time
|
| 214 |
+
st.success(f"Time taken for {model_choice} forecast: {time_taken:.2f} seconds")
|
| 215 |
+
|
| 216 |
+
# Dynamic forecasting
|
| 217 |
+
st.subheader("Dynamic Forecasting")
|
| 218 |
+
dynamic_model_choice = st.selectbox("Select model for dynamic forecasting", ["NHITS", "TimesNet", "LSTM", "TFT"], key="dynamic_model_choice")
|
| 219 |
+
dynamic_horizon = st.slider("Forecast horizon for dynamic forecasting", 1, 100, 10, key="dynamic_horizon")
|
| 220 |
+
forecast_time_series(df, dynamic_model_choice, frequency, dynamic_horizon)
|