from __future__ import print_function
from __future__ import absolute_import
from collections import namedtuple
import uuid
from .utils.shared_utils import urlopen
from .h2o import H2OConnection, _quoted, H2OFrame
[docs]class H2OAssembly:
"""Extension class of Pipeline implementing additional methods:
* to_pojo: Exports the assembly to a self-contained Java POJO used in a per-row, high-throughput environment.
* union: Combine two H2OAssembly objects, the resulting row from each H2OAssembly are joined with simple concatenation.
"""
# static properties pointing to H2OFrame methods
divide = H2OFrame.__truediv__
plus = H2OFrame.__add__
multiply= H2OFrame.__mul__
minus = H2OFrame.__sub__
less_than = H2OFrame.__lt__
less_than_equal = H2OFrame.__le__
equal_equal = H2OFrame.__eq__
not_equal = H2OFrame.__ne__
greater_than = H2OFrame.__gt__
greater_than_equal = H2OFrame.__ge__
def __init__(self, steps):
"""Build a new H2OAssembly.
Parameters
----------
steps : list
A list of steps that sequentially transforms the input data.
Returns
-------
H2OFrame
"""
self.id = None
self.steps = steps
self.fuzed = []
self.in_colnames = None
self.out_colnames = None
@property
[docs] def names(self):
return list(zip(*self.steps))[0][:-1]
[docs] def to_pojo(self, pojo_name="", path="", get_jar=True):
if pojo_name=="": pojo_name = unicode("AssemblyPOJO_" + str(uuid.uuid4()))
java = H2OConnection.get("Assembly.java/" + self.id + "/" + pojo_name, _rest_version=99)
file_path = path + "/" + pojo_name + ".java"
if path == "": print(java.text)
else:
with open(file_path, 'wb') as f:
f.write(java.text.encode("utf-8")) # this had better be utf-8 ?
if get_jar and path!="":
url = H2OConnection.make_url("h2o-genmodel.jar")
filename = path + "/" + "h2o-genmodel.jar"
response = urlopen()(url)
with open(filename, "wb") as f:
f.write(response.read())
# def union(self, assemblies):
# # fuse the assemblies onto this one, each is added to the end going left -> right
# # assemblies must be a list of namedtuples.
# # [(H2OAssembly, X, y, {params}), ..., (H2OAssembly, X, y, {params})]
# for i in assemblies:
# if not isinstance(i, namedtuple):
# raise ValueError("Not a namedtuple. Assembly must be of type collections.namedtuple with fields [assembly, x, params].")
# if i._fields != ('assembly','x','params'):
# raise ValueError("Assembly must be a namedtuple with fields ('assembly', 'x', 'params').")
# self.fuzed.append(i)
[docs] def fit(self, fr, **fit_params):
res = []
for step in self.steps:
res.append(step[1].to_rest(step[0]))
res = "[" + ",".join([_quoted(r.replace('"',"'")) for r in res]) + "]"
j = H2OConnection.post_json(url_suffix="Assembly", steps=res, frame=fr.frame_id, _rest_version=99)
self.id = j["assembly"]["name"]
return H2OFrame.get_frame(j["result"]["name"])
class H2OCol:
"""Wrapper class for H2OBinaryOp step's left/right args.
Use if you want to signal that a column actually comes from the train to be fitted on.
"""
def __init__(self, column):
self.col = column
# TODO: handle arbitrary (non H2OFrame) inputs -- sql, web, file, generated