#!/usr/bin/env python
# coding: utf-8

# In[1]:


# Setup logging
from timeseria import logger
logger.setup(level='INFO')

# Set default plot type as image
import os
os.environ["DEFAULT_PLOT_TYPE"] = "image"


# ⚠️ In this notebook, plots are configured to be rendered as images because otherwise they won't display correctly on GitHub or when the notebook is loaded (usually because not [trusted](https://stackoverflow.com/questions/44943646/jupyter-notebook-not-trusted)). To get interactive plots, remove the line above (or change it to "interactive") and re-run the notebook.
# 
# # Temperature and humidity forecasting with LSTM
# 
# This notebook showcases the LSTM forecatser of [Timeseria](https://github.com/sarusso/Timeseria).
# 
# Let's start by loading an example time series, and resample it to one hour:

# In[2]:


from timeseria import TEST_DATASETS_PATH
from timeseria.datastructures import TimeSeries
timeseries = TimeSeries.from_csv(TEST_DATASETS_PATH + 'humitemp_long.csv').resample('1h')


# Have a look at the time series

# In[3]:


timeseries


# Plot the time series

# In[4]:


timeseries.plot()


# ## Model fit, usage and evaluation
# 
# Instantiate and fit the model:

# In[5]:


from timeseria.models import LSTMForecaster
forecaster = LSTMForecaster(window=24, neurons=64, features=['values', 'hours'])
forecaster.fit(timeseries, epochs=30, verbose=True, reproducible=True)


# Call the `predict()` function of the model. This returns key-value data with the prediction values

# In[6]:


forecaster.predict(timeseries)


# The timeseries on which to call the predict has to be at least of 24 hours (the window length)

# In[7]:


try:
    forecaster.predict(timeseries[0:23])
except Exception as e:
    print(e)


# The `forecats()` method returns not only data but also their data points, with the timestamp

# In[8]:


forecaster.forecast(timeseries)


# The `apply()` method allows instead to directly apply the model on a time series (by combining the forecasts with the real observations).

# In[9]:


forecaster.apply(timeseries, steps=12)[-1000:].plot()


# Evaluate the forecaster. By default, forecasters get evaluated with RMSE and MAE error metrics:

# In[10]:


forecaster.evaluate(timeseries)


# For more advanced modes of evaluation, including using other error metrics, plotting the error distribution and/or the predictions, check out the "Forecasting - Advanced Evaluation" notebook.