def _getPipeline():
select_stage = [("split", Stage('read', process=read))]
join_stage = [("group_label",
Stage('join', init=join_init, process=join,
done=join_done))]
def combine_row(state, k, v, func):
if k not in state:
state[k] = 0
state[k] = state[k] + func(v)
node_combine_stage = [
("group_node_label",
Stage('node_combine',
init=partial(combine_init, init=lambda: {}),
process=partial(combine,
func=partial(combine_row, func=lambda v: 1)),
done=combine_done))
]
combine_all_stage = [
("group_label",
Stage('combine_all',
init=partial(combine_init, init=lambda: {}),
process=partial(combine,
func=partial(combine_row, func=lambda v: v)),
done=combine_done))
]
return select_stage + join_stage + node_combine_stage + combine_all_stage
def measure(test_data,
predictions,
measure="ca",
save_results=True,
show=False):
from disco.worker.pipeline.worker import Worker, Stage
from disco.core import Job, result_iterator
from disco.worker.task_io import task_input_stream, chain_reader
if measure not in ["ca", "mse"]:
raise Exception("measure should be ca or mse.")
if test_data.params["id_index"] == -1:
raise Exception("ID index should be defined.")
if predictions == []:
return "No predictions", None
# define a job and set save of results to ddfs
job = Job(worker=Worker(save_results=save_results))
job = Job(worker=Worker(save_results=save_results))
job.pipeline = [("split",
Stage("map",
input_chain=test_data.params["input_chain"],
init=simple_init,
process=map_test_data))]
job.params = test_data.params
job.run(name="ma_parse_testdata", input=test_data.params["data_tag"])
parsed_testdata = job.wait(show=show)
reduce_proces = reduce_ca if measure == "ca" else reduce_mse
job = Job(worker=Worker(save_results=save_results))
job.pipeline = [("split",
Stage("map",
init=simple_init,
input_chain=[task_input_stream, chain_reader],
process=map_predictions)),
('group_all',
Stage("reduce",
init=simple_init,
process=reduce_proces,
sort=True,
combine=True))]
job.run(name="ma_measure_accuracy", input=parsed_testdata + predictions)
measure, acc = [
(measure, acc) for measure, acc in result_iterator(job.wait(show=show))
][0]
return measure, acc
class CountWords(Job):
def __init__(self):
from disco.worker.pipeline.worker import Worker
super(CountWords, self).__init__(worker=Worker())
pipeline = [("split",
Stage("map",
process=map,
input_chain=[task_input_stream, chain_reader])),
("group_label",
Stage("reduce",
process=reduce,
combine=True,
output_chain=(task_output_stream, plain_output_stream),
sort=True))]
def predict(dataset, fitmodel_url, save_results=True, show=False):
"""
Predict the closest clusters for the datapoints in input.
"""
from disco.job import Job
from disco.worker.pipeline.worker import Worker, Stage
from disco.core import result_iterator
if "kmeans_fitmodel" not in fitmodel_url:
raise Exception("Incorrect fit model.")
job = Job(worker=Worker(save_results=save_results))
job.params = dict(dataset.params.items() + mean_point_center.items())
job.params["centers"] = [
(i, c) for i, c in result_iterator(fitmodel_url["kmeans_fitmodel"])
]
job.pipeline = [("split",
Stage("kmeans_predict",
input_chain=dataset.params["input_chain"],
init=simple_init,
process=predict_map))]
job.run(input=dataset.params["data_tag"], name="kmeans_predict")
return job.wait(show=show)
class SimplePipe(TestPipe):
def map(interface, state, label, inp):
out = interface.output(0)
for e in inp:
out.add(int(e), (bytes_to_str(e)).strip())
def reduce(interface, state, label, inp):
for k, v in sorted(inp):
state.append((k, v))
pipeline = [("split", Stage("map", process=map)),
("group_all",
Stage("reduce",
init=reduce_init,
process=reduce,
done=reduce_done))]
def fit(dataset, nu=0.1, save_results=True, show=False):
"""
Function starts a job for calculation of model parameters
Parameters
----------
input - dataset object with input urls and other parameters
nu - parameter to adjust the classifier
save_results - save results to ddfs
show - show info about job execution
Returns
-------
Urls of fit model results on ddfs
"""
from disco.worker.pipeline.worker import Worker, Stage
from disco.core import Job
if dataset.params["y_map"] == []:
raise Exception(
"Linear proximal SVM requires a target label mapping parameter.")
try:
nu = float(nu)
if nu <= 0:
raise Exception("Parameter nu should be greater than 0")
except ValueError:
raise Exception("Parameter should be numerical.")
job = Job(worker=Worker(save_results=save_results))
# job parallelizes mappers and joins them with one reducer
job.pipeline = [("split",
Stage("map",
input_chain=dataset.params["input_chain"],
init=simple_init,
process=map_fit)),
('group_all',
Stage("reduce",
init=simple_init,
process=reduce_fit,
combine=True))]
job.params = dataset.params
job.params["nu"] = nu
job.run(name="linearsvm_fit", input=dataset.params["data_tag"])
fitmodel_url = job.wait(show=show)
return {"linsvm_fitmodel": fitmodel_url} # return results url
class WordCount(Job):
def __init__(self):
from disco.worker.pipeline.worker import Worker
super(WordCount, self).__init__(worker=Worker())
pipeline = [("split",
Stage(
"map",
process=map,
input_chain=[task_input_stream, chain_reader],
output_chain=[
partial(redis_inter_stream_out,
redis_server=redis_server)
],
)),
("group_label",
Stage("reduce",
process=reduce,
input_chain=[task_input_stream,
redis_inter_stream_in]))]
def fit(dataset, save_results=True, show=False):
from disco.worker.pipeline.worker import Worker, Stage
from disco.core import Job
job = Job(worker=Worker(save_results=save_results))
job.pipeline = [("split",
Stage("map",
input_chain=dataset.params["input_chain"],
init=simple_init,
process=map_fit)),
('group_all',
Stage("reduce",
init=simple_init,
process=reduce_fit,
combine=True))]
job.params = dataset.params
job.run(name="linreg_fit", input=dataset.params["data_tag"])
fitmodel_url = job.wait(show=show)
return {"linreg_fitmodel": fitmodel_url} # return results url
class SimplerPipe(SimplePipe):
def reduce_init(interface, params):
return []
def reduce(interface, state, label, inp):
for rec in sorted(inp):
state.append((int(rec), (bytes_to_str(rec).strip())))
pipeline = [("group_all",
Stage("reduce",
init=reduce_init,
process=reduce,
done=reduce_done))]
def estimate(master, input, center, k, iterations):
"""
Optimize k-clustering for `iterations` iterations with cluster
center definitions as given in `center`.
"""
from kclustering_pipeline import Estimate
job = Estimate()
job.pipeline = [("split",
Stage("k_cluster_init_map", input_chain =
[task_input_stream, reader], init = map_init,
process = random_init_map)),
('group_label',
Stage("k_cluster_init_reduce", process = estimate_reduce, init = simple_init))]
job.params = center
job.params['seed'] = 0
job.params['k'] = k
job.run(input = input)
centers = [(i,c) for i,c in result_iterator(job.wait())]
job.purge()
for j in range(iterations):
job = Estimate()
job.params = center
job.params['k'] = k
job.params['centers'] = centers
job.pipeline = [('split', Stage("kcluster_map_iter_%s" %(j,),
input_chain = [task_input_stream, reader],
process=estimate_map, init = simple_init)),
('group_label', Stage("kcluster_reduce_iter_%s" %(j,),
process=estimate_reduce, init = simple_init))]
job.run(input = input)
centers = [(i,c) for i,c in result_iterator(job.wait())]
job.purge()
return centers
def fit(dataset, save_results=True, show=False):
"""
Function builds a model for Naive Bayes. It executes multiple map functions and one reduce function which aggregates intermediate results and returns a model.
Parameters
----------
input - dataset object with input urls and other parameters
save_results - save results to ddfs
show - show info about job execution
Returns
-------
Urls of fit model results on ddfs
"""
from disco.worker.pipeline.worker import Worker, Stage
from disco.core import Job
# define a job and set save of results to ddfs
job = Job(worker=Worker(save_results=save_results))
# job parallelizes mappers, sorts intermediate pairs and joins them with one reducer
job.pipeline = [("split",
Stage("map",
input_chain=dataset.params["input_chain"],
init=simple_init,
process=map_fit)),
('group_all',
Stage("reduce",
init=simple_init,
process=reduce_fit,
sort=True,
combine=True))]
job.params = dataset.params # job parameters (dataset object)
# define name of a job and input data urls
job.run(name="naivebayes_fit", input=dataset.params["data_tag"])
fitmodel_url = job.wait(show=show)
return {"naivebayes_fitmodel": fitmodel_url} # return results url
def _fit_predict(fit_data, samples, tau, save_results, show):
from disco.worker.pipeline.worker import Worker, Stage
from disco.core import Job
job = Job(worker=Worker(save_results=save_results))
job.pipeline = [("split",
Stage("map",
input_chain=fit_data.params["input_chain"],
init=simple_init,
process=map_fit)),
('group_all',
Stage("reduce",
init=simple_init,
process=reduce_fit,
sort=True,
combine=True))]
job.params = fit_data.params
job.params["tau"] = tau
job.params["samples"] = samples
job.run(name="lwlr_fit_predict", input=fit_data.params["data_tag"])
return job.wait(show=show)
def predict(master, input, center, centers):
"""
Predict the closest clusters for the datapoints in input.
"""
from kclustering_pipeline import Estimate
job = Estimate()
job.pipeline = [("split",
Stage("k_cluster_predict", input_chain =
[task_input_stream, reader], init = simple_init,
process = predict_map))]
job.params = center
job.params['centers'] = centers
job.run(input = input)
return job.wait()
def predict(dataset, fitmodel_url, save_results=True, show=False):
from disco.worker.pipeline.worker import Worker, Stage
from disco.core import Job, result_iterator
if "linreg_fitmodel" not in fitmodel_url:
raise Exception("Incorrect fit model.")
job = Job(worker=Worker(save_results=save_results))
job.pipeline = [("split",
Stage("map",
input_chain=dataset.params["input_chain"],
init=simple_init,
process=map_predict))]
job.params = dataset.params
job.params["thetas"] = [
v for _, v in result_iterator(fitmodel_url["linreg_fitmodel"])
][0]
job.run(name="linreg_predict", input=dataset.params["data_tag"])
return job.wait(show=show)
def predict(dataset, fitmodel_url, save_results=True, show=False):
"""
Function starts a job that makes predictions to input data with a given model
Parameters
----------
input - dataset object with input urls and other parameters
fitmodel_url - model created in fit phase
save_results - save results to ddfs
show - show info about job execution
Returns
-------
Urls with predictions on ddfs
"""
from disco.worker.pipeline.worker import Worker, Stage
from disco.core import Job, result_iterator
if dataset.params["y_map"] == []:
raise Exception(
"Logistic regression requires a target label mapping parameter.")
if "logreg_fitmodel" not in fitmodel_url:
raise Exception("Incorrect fit model.")
job = Job(worker=Worker(save_results=save_results))
# job parallelizes execution of mappers
job.pipeline = [("split",
Stage("map",
input_chain=dataset.params["input_chain"],
init=simple_init,
process=map_predict))]
job.params = dataset.params # job parameters (dataset object)
job.params["thetas"] = [
v for k, v in result_iterator(fitmodel_url["logreg_fitmodel"])
if k == "thetas"
][0] # thetas are loaded from ddfs
job.run(name="logreg_predict", input=dataset.params["data_tag"])
results = job.wait(show=show)
return results
def predict(dataset, fitmodel_url, coeff=0.5, save_results=True, show=False):
from disco.worker.pipeline.worker import Worker, Stage
from disco.core import Job, result_iterator
import discomll
path = "/".join(
discomll.__file__.split("/")[:-1] + ["ensemble", "core", ""])
job = Job(worker=Worker(save_results=save_results))
job.pipeline = [("split",
Stage("map",
input_chain=dataset.params["input_chain"],
init=simple_init,
process=map_predict))]
if "dwf_fitmodel" not in fitmodel_url:
raise Exception("Incorrect fit model.")
try:
coeff = float(coeff)
if coeff < 0:
raise Exception("Parameter coeff should be greater than 0.")
except ValueError:
raise Exception("Parameter coeff should be numerical.")
job.params = dataset.params
job.params["coeff"] = coeff
for k, v in result_iterator(fitmodel_url["dwf_fitmodel"]):
job.params[k] = v
if len(job.params["forest"]) == 0:
print "Warning: There is no decision trees in forest"
return []
job.run(name="distributed_weighted_forest_predict",
input=dataset.params["data_tag"],
required_files=[path + "decision_tree.py"])
return job.wait(show=show)
def predict(dataset,
fitmodel_url,
voting=False,
save_results=True,
show=False):
from disco.worker.pipeline.worker import Worker, Stage
from disco.core import Job, result_iterator
import discomll
path = "/".join(
discomll.__file__.split("/")[:-1] + ["ensemble", "core", ""])
if "drf_fitmodel" not in fitmodel_url:
raise Exception("Incorrect fit model.")
job = Job(worker=Worker(save_results=save_results))
job.pipeline = [
("split",
Stage("map",
input_chain=dataset.params["input_chain"],
init=simple_init,
process=map_predict_voting if voting else map_predict_dist))
]
job.params = dataset.params
for k, v in result_iterator(fitmodel_url["drf_fitmodel"]):
job.params[k] = v
if len(job.params["forest"]) == 0:
print "Warning: There is no decision trees in forest"
return []
job.run(name="distributed_random_forest_predict",
input=dataset.params["data_tag"],
required_files=[path + "decision_tree.py"])
return job.wait(show=show)
def fit(dataset,
trees_per_chunk=1,
bootstrap=True,
max_tree_nodes=50,
min_samples_leaf=10,
min_samples_split=5,
class_majority=1,
separate_max=True,
measure="info_gain",
accuracy=1,
random_state=None,
save_results=True,
show=False):
from disco.worker.pipeline.worker import Worker, Stage
from disco.core import Job
import discomll
path = "/".join(
discomll.__file__.split("/")[:-1] + ["ensemble", "core", ""])
try:
trees_per_chunk = int(trees_per_chunk)
max_tree_nodes = int(
max_tree_nodes) if max_tree_nodes != None else max_tree_nodes
min_samples_leaf = int(min_samples_leaf)
min_samples_split = int(min_samples_split)
class_majority = float(class_majority)
accuracy = int(accuracy)
separate_max = separate_max
if trees_per_chunk > 1 and bootstrap == False:
raise Exception(
"Parameter trees_per_chunk (or Trees per subset) should be 1 to disable bootstrap."
)
if trees_per_chunk <= 0 or min_samples_leaf <= 0 or class_majority <= 0 or min_samples_split <= 0 and accuracy < 0 or type(
bootstrap) != bool:
raise Exception("Parameters should be greater than 0.")
except ValueError:
raise Exception("Parameters should be numerical.")
if measure not in ["info_gain", "mdl"]:
raise Exception("measure should be set to info_gain or mdl.")
job = Job(worker=Worker(save_results=save_results))
job.pipeline = [
("split",
Stage("map",
input_chain=dataset.params["input_chain"],
init=map_init,
process=map_fit_bootstrap if bootstrap else map_fit)),
('group_all',
Stage("reduce", init=simple_init, process=reduce_fit, combine=True))
]
job.params = dataset.params
job.params["trees_per_chunk"] = trees_per_chunk
job.params["max_tree_nodes"] = max_tree_nodes
job.params["min_samples_leaf"] = min_samples_leaf
job.params["min_samples_split"] = min_samples_split
job.params["class_majority"] = class_majority
job.params["measure"] = measure
job.params["bootstrap"] = bootstrap
job.params["accuracy"] = accuracy
job.params["separate_max"] = separate_max
job.params['seed'] = random_state
job.run(name="forest_distributed_decision_trees_fit",
input=dataset.params["data_tag"],
required_files=[path + "decision_tree.py", path + "measures.py"])
fitmodel_url = job.wait(show=show)
return {"fddt_fitmodel": fitmodel_url} # return results url
def predict(dataset, fitmodel_url, m=1, save_results=True, show=False):
"""
Function starts a job that makes predictions to input data with a given model
Parameters
----------
input - dataset object with input urls and other parameters
fitmodel_url - model created in fit phase
m - m estimate is used with discrete features
save_results - save results to ddfs
show - show info about job execution
Returns
-------
Urls of predictions on ddfs
"""
from disco.worker.pipeline.worker import Worker, Stage
from disco.core import Job, result_iterator
import numpy as np
try:
m = float(m)
except ValueError:
raise Exception("Parameter m should be numerical.")
if "naivebayes_fitmodel" in fitmodel_url:
# fit model is loaded from ddfs
fit_model = dict(
(k, v)
for k, v in result_iterator(fitmodel_url["naivebayes_fitmodel"]))
if len(fit_model["y_labels"]) < 2:
print "There is only one class in training data."
return []
else:
raise Exception("Incorrect fit model.")
if dataset.params["X_meta"].count(
"d") > 0: # if there are discrete features in the model
# code calculates logarithms to optimize predict phase as opposed to calculation by every mapped.
np.seterr(divide='ignore')
for iv in fit_model["iv"]:
dist = [
fit_model.pop((y, ) + iv, 0) for y in fit_model["y_labels"]
]
fit_model[iv] = np.nan_to_num(
np.log(
np.true_divide(
np.array(dist) + m * fit_model["prior"],
np.sum(dist) + m))) - fit_model["prior_log"]
del (fit_model["iv"])
# define a job and set save of results to ddfs
job = Job(worker=Worker(save_results=save_results))
# job parallelizes execution of mappers
job.pipeline = [("split",
Stage("map",
input_chain=dataset.params["input_chain"],
init=simple_init,
process=map_predict))]
job.params = dataset.params # job parameters (dataset object)
job.params["fit_model"] = fit_model
# define name of a job and input data urls
job.run(name="naivebayes_predict", input=dataset.params["data_tag"])
results = job.wait(show=show)
return results
def getPipeline(count, type):
intermediates = [(type, Stage("inter_%d" % i, process=intermediate)) for i in range(count)]
pipeline = [("split", Stage("map", process=map))] + intermediates + [("group_label", Stage("reduce",
init=reduce_init,
process=reduce, done=reduce_done))]
return pipeline
class SortJob(TestPipe):
scheduler = {'max_cores': 7}
pipeline = [("split", Stage("Map", process=Map)),
("group_label",
Stage("Reduce", process=Reduce, combine=True, sort=True))]
def fit_predict(training_data,
fitting_data,
tau=1,
samples_per_job=0,
save_results=True,
show=False):
from disco.worker.pipeline.worker import Worker, Stage
from disco.core import Job, result_iterator
from disco.core import Disco
"""
training_data - training samples
fitting_data - dataset to be fitted to training data.
tau - controls how quickly the weight of a training sample falls off with distance of its x(i) from the query point x.
samples_per_job - define a number of samples that will be processed in single mapreduce job. If 0, algorithm will calculate number of samples per job.
"""
try:
tau = float(tau)
if tau <= 0:
raise Exception("Parameter tau should be >= 0.")
except ValueError:
raise Exception("Parameter tau should be numerical.")
if fitting_data.params["id_index"] == -1:
raise Exception("Predict data should have id_index set.")
job = Job(worker=Worker(save_results=save_results))
job.pipeline = [("split",
Stage("map",
input_chain=fitting_data.params["input_chain"],
init=simple_init,
process=map_predict))]
job.params = fitting_data.params
job.run(name="lwlr_read_data", input=fitting_data.params["data_tag"])
samples = {}
results = []
tau = float(2 * tau**2) # calculate tau once
counter = 0
for test_id, x in result_iterator(job.wait(show=show)):
if samples_per_job == 0:
# calculate number of samples per job
if len(x) <= 100: # if there is less than 100 attributes
samples_per_job = 100 # 100 samples is max per on job
else:
# there is more than 100 attributes
samples_per_job = len(x) * -25 / 900. + 53 # linear function
samples[test_id] = x
if counter == samples_per_job:
results.append(
_fit_predict(training_data, samples, tau, save_results, show))
counter = 0
samples = {}
counter += 1
if len(samples) > 0: # if there is some samples left in the the dictionary
results.append(
_fit_predict(training_data, samples, tau, save_results, show))
# merge results of every iteration into a single tag
ddfs = Disco().ddfs
ddfs.tag(job.name, [[list(ddfs.blobs(tag))[0][0]] for tag in results])
return ["tag://" + job.name]
def fit(dataset,
n_clusters=5,
max_iterations=10,
random_state=None,
save_results=True,
show=False):
"""
Optimize k-clustering for `iterations` iterations with cluster
center definitions as given in `center`.
"""
from disco.job import Job
from disco.worker.pipeline.worker import Worker, Stage
from disco.core import result_iterator
try:
n_clusters = int(n_clusters)
max_iterations = int(max_iterations)
if n_clusters < 2:
raise Exception("Parameter n_clusters should be greater than 1.")
if max_iterations < 1:
raise Exception(
"Parameter max_iterations should be greater than 0.")
except ValueError:
raise Exception("Parameters should be numerical.")
job = Job(worker=Worker(save_results=save_results))
job.pipeline = [("split",
Stage("kmeans_init_map",
input_chain=dataset.params["input_chain"],
init=map_init,
process=random_init_map)),
('group_label',
Stage("kmeans_init_reduce",
process=estimate_reduce,
init=simple_init,
combine=True))]
job.params = dict(dataset.params.items() + mean_point_center.items())
job.params['seed'] = random_state
job.params['k'] = n_clusters
job.run(input=dataset.params["data_tag"], name="kmeans_init")
init = job.wait(show=show)
centers = [(i, c) for i, c in result_iterator(init)]
for j in range(max_iterations):
job = Job(worker=Worker(save_results=save_results))
job.params = dict(dataset.params.items() + mean_point_center.items())
job.params['k'] = n_clusters
job.params['centers'] = centers
job.pipeline = [('split',
Stage("kmeans_map_iter_%s" % (j + 1, ),
input_chain=dataset.params["input_chain"],
process=estimate_map,
init=simple_init)),
('group_label',
Stage("kmeans_reduce_iter_%s" % (j + 1, ),
process=estimate_reduce,
init=simple_init,
combine=True))]
job.run(input=dataset.params["data_tag"],
name='kmeans_iter_%d' % (j + 1, ))
fitmodel_url = job.wait(show=show)
centers = [(i, c) for i, c in result_iterator(fitmodel_url)]
return {"kmeans_fitmodel": fitmodel_url} # return results url
class SortJob(TestPipe):
pipeline = [("split", Stage("Map", process=Map)),
("group_label",
Stage("Reduce", process=Reduce, combine=True, sort=True))]
class RawJob(TestPipe):
pipeline = [("split", Stage("map", process=map))]
def fit(dataset, alpha=1e-8, max_iterations=10, save_results=True, show=False):
"""
Function starts a job for calculation of theta parameters
Parameters
----------
input - dataset object with input urls and other parameters
alpha - convergence value
max_iterations - define maximum number of iterations
save_results - save results to ddfs
show - show info about job execution
Returns
-------
Urls of fit model results on ddfs
"""
from disco.worker.pipeline.worker import Worker, Stage
from disco.core import Job, result_iterator
import numpy as np
if dataset.params["y_map"] == []:
raise Exception(
"Logistic regression requires a target label mapping parameter.")
try:
alpha = float(alpha)
max_iterations = int(max_iterations)
if max_iterations < 1:
raise Exception(
"Parameter max_iterations should be greater than 0.")
except ValueError:
raise Exception("Parameters should be numerical.")
# initialize thetas to 0 and add intercept term
thetas = np.zeros(len(dataset.params["X_indices"]) + 1)
J = [0] # J cost function values for every iteration
for i in range(max_iterations):
job = Job(worker=Worker(save_results=save_results))
# job parallelizes mappers and joins them with one reducer
job.pipeline = [("split",
Stage("map",
input_chain=dataset.params["input_chain"],
init=simple_init,
process=map_fit)),
('group_all',
Stage("reduce",
init=simple_init,
process=reduce_fit,
combine=True))]
job.params = dataset.params # job parameters (dataset object)
job.params["thetas"] = thetas # every iteration set new thetas
job.run(name="logreg_fit_iter_%d" % (i + 1),
input=dataset.params["data_tag"])
fitmodel_url = job.wait(show=show)
for k, v in result_iterator(fitmodel_url):
if k == "J": #
J.append(v) # save value of J cost function
else:
thetas = v # save new thetas
if np.abs(J[-2] - J[-1]) < alpha: # check for convergence
if show:
print("Converged at iteration %d" % (i + 1))
break
return {"logreg_fitmodel": fitmodel_url} # return results url