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

# This notebook is part of my [Python data science curriculum](http://www.terran.us/articles/python_curriculum.html)
# 
# There are two major sets of documentation which are relevant for H2O:
# 
# http://docs.h2o.ai/h2o/latest-stable/h2o-docs/welcome.html  
# http://docs.h2o.ai/h2o/latest-stable/h2o-py/docs/intro.html
# 

# In[1]:


import pandas as pd
import altair as alt
alt.renderers.enable('notebook')


# # Cluster Init

# In[2]:


import h2o
h2o.init()


# # Data Loading

# At first I thought you could only load from disk:

# In[3]:


from plotnine.data import diamonds
diamonds.to_csv('/tmp/diamonds.csv')
h2o_diamonds = h2o.import_file('/tmp/diamonds.csv')


# But in fact you __can__ transfer data from Python.  The key piece of information is that you use the h2o.H2OFrame __constructor__ to do it.  See references:
# 
# http://docs.h2o.ai/h2o/latest-stable/h2o-py/docs/data.html#loading-data-from-a-python-object  
# http://docs.h2o.ai/h2o/latest-stable/h2o-py/docs/frame.html#h2oframe  

# In[4]:


# The destination_frame argument is optional, but if you don't use it, you get a horrible hex name.
h2o_diamonds2 = h2o.H2OFrame(diamonds, destination_frame='diamonds2')


# In[5]:


h2o.ls()


# In[6]:


h2o_diamonds2


# # Linear Regression

# ## Failed Attempt

# Let's fit a very simple model:

# In[7]:


h2o_lm = h2o.estimators.H2OGeneralizedLinearEstimator(family='gaussian')


# In[8]:


h2o_lm.train(x=['carat','cut','color','clarity'],y='price',training_frame=h2o_diamonds2)


# In[9]:


h2o_lm


# __This is not right at all.  $R^2$ should be 0.91 for this model, not 0.35!__
# 
# __Aha, it appears the default model is regularized.__  This is not explicitly stated but it is implied by the available arguments.

# ## Successful Attempt

# In[10]:


# Note the lambda_=0 to turn off regularization
h2o_lm = h2o.estimators.H2OGeneralizedLinearEstimator(family='gaussian',lambda_=0)
h2o_lm.train(x=['carat','cut','color','clarity'],y='price',training_frame=h2o_diamonds2)
h2o_lm


# That's more like it!

# In[11]:


h2o_lm.coef()


# In[12]:


h2o.ls()


# # Random Forest

# In[13]:


h2o_rf = h2o.estimators.random_forest.H2ORandomForestEstimator()
h2o_rf.train(x=['carat','cut','color','clarity'],y='price',training_frame=h2o_diamonds2)
h2o_rf


# In[14]:


h2o_rf.varimp()


# # Gradient Boosting

# We will split the data into training and test, use cross-validation on the training data to tune the hyperparameters, and then evaluate on the test data.  This is a standard workflow for high-variance ML models.

# First split the data
# 
# http://docs.h2o.ai/h2o/latest-stable/h2o-docs/data-munging/splitting-datasets.html

# In[15]:


diamonds_split = h2o_diamonds2.split_frame(ratios=[0.75], destination_frames=['diamonds_train','diamonds_test'])


# In[16]:


diamonds_split[0].dim


# Then fit a model with cross-validation by specifying nfolds  
# http://docs.h2o.ai/h2o/latest-stable/h2o-docs/cross-validation.html

# In[17]:


h2o_gb = h2o.estimators.gbm.H2OGradientBoostingEstimator(nfolds=5)                                                       


# In[18]:


h2o_gb.train(x=['carat','cut','color','clarity'],y='price',training_frame=diamonds_split[0])


# In[19]:


h2o_gb


# That didn't print well; let's try this instead:

# In[20]:


h2o_gb.cross_validation_metrics_summary().as_data_frame()


# In[21]:


results=pd.DataFrame()
for lr in [0.02,.05,0.1,0.2,0.5]:
    for ntrees in [5,50,500]:
        h2o_gb = h2o.estimators.gbm.H2OGradientBoostingEstimator(nfolds=5,learn_rate=lr,ntrees=ntrees)
        h2o_gb.train(x=['carat','cut','color','clarity','x','y','z','depth','table'],y='price',training_frame=diamonds_split[0])
        tmp=h2o_gb.cross_validation_metrics_summary().as_data_frame()
        tmp['lr']=lr
        tmp['ntrees']=ntrees
        results=pd.concat([results,tmp])
    


# In[22]:


tmp=results[lambda x: x.iloc[:,0]=='rmse'].copy()
# For some reason all the data is strings
tmp['mean']=tmp['mean'].astype('double')

tmp.sort_values('mean').head()


# In[23]:


c=alt.Chart(tmp[tmp.iloc[:,0]=='rmse'])
c.mark_point().encode(x='lr:Q',color='ntrees:N',y='mean').interactive()

# I didn't end up using this because it was too dense to see on the chart
#tmp['sd']=tmp['sd'].astype('double')
#tmp['low']=tmp['mean'] - 2*tmp.sd
#tmp['high']=tmp['mean'] + 2*tmp.sd


# In[24]:


h2o_gb_best = h2o.estimators.gbm.H2OGradientBoostingEstimator(learn_rate=0.05,ntrees=500)
h2o_gb_best.train(
    x=['carat','cut','color','clarity','x','y','z','depth','table'],y='price',
    training_frame=diamonds_split[0],validation_frame=diamonds_split[1])
h2o_gb_best


# Alternatively, you could try H2OGridSearch  
# http://docs.h2o.ai/h2o/latest-stable/h2o-docs/grid-search.html#grid-search-example-in-python

# # Predictions
# 
# We can also run the models and get the results back in Pandas objects:

# In[31]:


gb_predictions=h2o_gb_best.predict(diamonds_split[1])
lm_predictions=h2o_lm.predict(diamonds_split[1])


# In[43]:


pred=diamonds_split[1].as_data_frame().assign(
  model='gb',pred=gb_predictions.as_data_frame()).append(
  diamonds_split[1].as_data_frame().assign(
  model='lm',pred=lm_predictions.as_data_frame())).sample(1000)
    
alt.Chart(pred).mark_circle(opacity=0.25).encode(
    x='price',y='pred',color='model'
).interactive()


# In[ ]: