def test_auto_rechunk():
clf = ParallelPostFit(GradientBoostingClassifier())
X, y = make_classification(n_samples=1000, n_features=20, chunks=100)
X = X.rechunk({0: 100, 1: 10})
clf.fit(X, y)
assert clf.predict(X).compute().shape == (1000, )
assert clf.predict_proba(X).compute().shape == (1000, 2)
assert clf.score(X, y) == clf.score(X.compute(), y.compute())
X, y = make_classification(n_samples=1000, n_features=20, chunks=100)
X = X.rechunk({0: 100, 1: 10})
X._chunks = (tuple(np.nan for _ in X.chunks[0]), X.chunks[1])
clf.predict(X)
def test_predict(kind):
X, y = make_classification(chunks=100)
if kind == "numpy":
X, y = dask.compute(X, y)
elif kind == "dask.dataframe":
X = dd.from_dask_array(X)
y = dd.from_dask_array(y)
base = LogisticRegression(random_state=0, n_jobs=1, solver="lbfgs")
wrap = ParallelPostFit(
LogisticRegression(random_state=0, n_jobs=1, solver="lbfgs"))
base.fit(*dask.compute(X, y))
wrap.fit(*dask.compute(X, y))
assert_estimator_equal(wrap.estimator, base)
result = wrap.predict(X)
expected = base.predict(X)
assert_eq_ar(result, expected)
result = wrap.predict_proba(X)
expected = base.predict_proba(X)
assert_eq_ar(result, expected)
result = wrap.predict_log_proba(X)
expected = base.predict_log_proba(X)
assert_eq_ar(result, expected)
def test_multiclass():
X, y = sklearn.datasets.make_classification(n_classes=3, n_informative=4)
X = da.from_array(X, chunks=50)
y = da.from_array(y, chunks=50)
clf = ParallelPostFit(
LogisticRegression(random_state=0,
n_jobs=1,
solver="lbfgs",
multi_class="auto"))
clf.fit(*dask.compute(X, y))
result = clf.predict(X)
expected = clf.estimator.predict(X)
assert isinstance(result, da.Array)
assert_eq_ar(result, expected)
result = clf.predict_proba(X)
expected = clf.estimator.predict_proba(X)
assert isinstance(result, da.Array)
assert_eq_ar(result, expected)
result = clf.predict_log_proba(X)
expected = clf.estimator.predict_log_proba(X)
assert_eq_ar(result, expected)
def test_multiclass():
X, y = sklearn.datasets.make_classification(n_classes=3, n_informative=4)
X = da.from_array(X, chunks=50)
y = da.from_array(y, chunks=50)
if SK_GE_020:
kwargs = {"multi_class": "auto"}
else:
kwargs = {}
clf = ParallelPostFit(
LogisticRegression(random_state=0, n_jobs=1, solver="lbfgs", **kwargs)
)
clf.fit(X, y)
result = clf.predict(X)
expected = clf.estimator.predict(X)
assert isinstance(result, da.Array)
assert_eq_ar(result, expected)
result = clf.predict_proba(X)
expected = clf.estimator.predict_proba(X)
assert isinstance(result, da.Array)
assert_eq_ar(result, expected)
def test_regressor_sg_train_mg_predict(datatype, keys, data_size,
fit_intercept, client):
# Just testing for basic compatibility w/ dask-ml's ParallelPostFit.
# Refer to test_pickle.py for more extensive testing of single-GPU
# model serialization.
nrows, ncols, n_info = data_size
X_train, y_train, _ = make_dataset(datatype, nrows, ncols, n_info)
X_train_local = X_train.compute()
y_train_local = y_train.compute()
local_model = cuml.linear_model.LinearRegression(
fit_intercept=fit_intercept)
local_model.fit(X_train_local, y_train_local)
dist_model = ParallelPostFit(estimator=local_model)
predictions = dist_model.predict(X_train).compute()
assert isinstance(predictions, cupy.ndarray)
# Dataset should be fairly linear already so the predictions should
# be very close to the training data.
np.testing.assert_allclose(predictions.get(),
y_train.compute().get(),
atol=1e-3,
rtol=1e-3)
def test_it_works():
clf = ParallelPostFit(GradientBoostingClassifier())
X, y = make_classification(n_samples=1000, chunks=100)
clf.fit(X, y)
assert isinstance(clf.predict(X), da.Array)
assert isinstance(clf.predict_proba(X), da.Array)
def test_predict_meta_override():
X = pd.DataFrame({"c_0": [1, 2, 3, 4]})
y = np.array([1, 2, 3, 4])
base = CategoricalNB()
base.fit(pd.DataFrame(X), y)
dd_X = dd.from_pandas(X, npartitions=2)
dd_X._meta = pd.DataFrame({"c_0": [5]})
# Failure when not proving predict_meta
# because of value dependent model
wrap = ParallelPostFit(base)
with pytest.raises(ValueError):
wrap.predict(dd_X)
# Success when providing meta over-ride
wrap = ParallelPostFit(base, predict_meta=np.array([1]))
result = wrap.predict(dd_X)
expected = base.predict(X)
assert_eq_ar(result, expected)
def test_warning_on_dask_array_without_array_function():
X, y = make_classification(n_samples=10, n_features=2, chunks=10)
clf = ParallelPostFit(GradientBoostingClassifier())
clf = clf.fit(X, y)
class FakeArray:
def __init__(self, value):
self.value = value
@property
def ndim(self):
return self.value.ndim
@property
def len(self):
return self.value.len
@property
def dtype(self):
return self.value.dtype
@property
def shape(self):
return self.value.shape
ar = FakeArray(np.zeros(shape=(2, 2)))
fake_dask_ar = da.from_array(ar)
fake_dask_ar._meta = FakeArray(np.zeros(shape=(0, 0)))
with pytest.warns(
UserWarning,
match="provide explicit `predict_meta` to the dask_ml.wrapper"):
clf.predict(fake_dask_ar)
with pytest.warns(
UserWarning,
match=
"provide explicit `predict_proba_meta` to the dask_ml.wrapper",
):
clf.predict_proba(fake_dask_ar)
def test_it_works():
clf = ParallelPostFit(GradientBoostingClassifier())
X, y = make_classification(n_samples=1000, chunks=100)
X_, y_ = dask.compute(X, y)
clf.fit(X_, y_)
assert isinstance(clf.predict(X), da.Array)
assert isinstance(clf.predict_proba(X), da.Array)
result = clf.score(X, y)
expected = clf.estimator.score(X_, y_)
assert result == expected
def test_predict_correct_output_dtype():
X, y = make_classification(chunks=100)
X_ddf = dd.from_dask_array(X)
base = LinearRegression(n_jobs=1)
base.fit(X, y)
wrap = ParallelPostFit(base)
base_output = base.predict(X_ddf.compute())
wrap_output = wrap.predict(X_ddf)
assert wrap_output.dtype == base_output.dtype
def test_sparse_inputs():
X = csr_matrix((3, 4))
y = np.asarray([0, 0, 1], dtype=np.int32)
base = SGDClassifier(tol=1e-3)
base = base.fit(X, y)
wrap = ParallelPostFit(base)
X_da = da.from_array(X, chunks=(1, 4))
result = wrap.predict(X_da).compute()
expected = base.predict(X)
assert_eq_ar(result, expected)
def test_multiclass():
X, y = make_classification(chunks=50, n_classes=3, n_informative=4)
clf = ParallelPostFit(LogisticRegression(random_state=0))
clf.fit(X, y)
result = clf.predict(X)
expected = clf.estimator.predict(X)
assert isinstance(result, da.Array)
assert_eq_ar(result, expected)
result = clf.predict_proba(X)
expected = clf.estimator.predict_proba(X)
assert isinstance(result, da.Array)
assert_eq_ar(result, expected)
def test_predict(kind):
X, y = make_classification(chunks=100)
if kind == 'numpy':
X, y = dask.compute(X, y)
elif kind == 'dask.dataframe':
X = dd.from_dask_array(X)
y = dd.from_dask_array(y)
base = LogisticRegression(random_state=0)
wrap = ParallelPostFit(LogisticRegression(random_state=0))
base.fit(X, y)
wrap.fit(X, y)
assert_estimator_equal(wrap.estimator, base)
result = wrap.predict(X)
expected = base.predict(X)
assert_eq_ar(result, expected)
result = wrap.predict_proba(X)
expected = base.predict_proba(X)
assert_eq_ar(result, expected)
def predict_xr(
model,
input_xr,
chunk_size=None,
persist=True,
proba=False,
clean=False,
return_input=False,
):
"""
Using dask-ml ParallelPostfit(), runs the parallel
predict and predict_proba methods of sklearn
estimators. Useful for running predictions
on a larger-than-RAM datasets.
Last modified: September 2020
Parameters
----------
model : scikit-learn model or compatible object
Must have a .predict() method that takes numpy arrays.
input_xr : xarray.DataArray or xarray.Dataset.
Must have dimensions 'x' and 'y'
chunk_size : int
The dask chunk size to use on the flattened array. If this
is left as None, then the chunks size is inferred from the
.chunks() method on the `input_xr`
persist : bool
If True, and proba=True, then 'input_xr' data will be
loaded into distributed memory. This will ensure data
is not loaded twice for the prediction of probabilities,
but this will only work if the data is not larger than RAM.
proba : bool
If True, predict probabilities. This only applies if the
model has a .predict_proba() method
clean : bool
If True, remove Infs and NaNs from input and output arrays
return_input : bool
If True, then the data variables in the 'input_xr' dataset will
be appended to the output xarray dataset.
Returns
----------
output_xr : xarray.Dataset
An xarray.Dataset containing the prediction output from model
with input_xr as input, if proba=True then dataset will also contain
the prediciton probabilities. Has the same spatiotemporal structure
as input_xr.
"""
if chunk_size is None:
chunk_size = int(input_xr.chunks["x"][0]) * int(
input_xr.chunks["y"][0])
# convert model to dask predict
model = ParallelPostFit(model)
# with joblib.parallel_backend("dask"):
x, y, crs = input_xr.x, input_xr.y, input_xr.geobox.crs
input_data = []
for var_name in input_xr.data_vars:
input_data.append(input_xr[var_name])
input_data_flattened = []
# TODO: transfer to dask dataframe
for arr in input_data:
data = arr.data.flatten().rechunk(chunk_size)
input_data_flattened.append(data)
# reshape for prediction
input_data_flattened = da.array(input_data_flattened).transpose()
if clean:
input_data_flattened = da.where(da.isfinite(input_data_flattened),
input_data_flattened, 0)
if proba and persist:
# persisting data so we don't require loading all the data twice
input_data_flattened = input_data_flattened.persist()
# apply the classification
print(" predicting...")
out_class = model.predict(input_data_flattened)
# Mask out NaN or Inf values in results
if clean:
out_class = da.where(da.isfinite(out_class), out_class, 0)
# Reshape when writing out
out_class = out_class.reshape(len(y), len(x))
# stack back into xarray
output_xr = xr.DataArray(out_class,
coords={
"x": x,
"y": y
},
dims=["y", "x"])
output_xr = output_xr.to_dataset(name="Predictions")
if proba:
print(" probabilities...")
out_proba = model.predict_proba(input_data_flattened)
# convert to %
out_proba = da.max(out_proba, axis=1) * 100.0
if clean:
out_proba = da.where(da.isfinite(out_proba), out_proba, 0)
out_proba = out_proba.reshape(len(y), len(x))
out_proba = xr.DataArray(out_proba,
coords={
"x": x,
"y": y
},
dims=["y", "x"])
output_xr["Probabilities"] = out_proba
if return_input:
print(" input features...")
# unflatten the input_data_flattened array and append
# to the output_xr containin the predictions
arr = input_xr.to_array()
stacked = arr.stack(z=["y", "x"])
# handle multivariable output
output_px_shape = ()
if len(input_data_flattened.shape[1:]):
output_px_shape = input_data_flattened.shape[1:]
output_features = input_data_flattened.reshape(
(len(stacked.z), *output_px_shape))
# set the stacked coordinate to match the input
output_features = xr.DataArray(
output_features,
coords={
"z": stacked["z"]
},
dims=[
"z",
*[
"output_dim_" + str(idx)
for idx in range(len(output_px_shape))
],
],
).unstack()
# convert to dataset and rename arrays
output_features = output_features.to_dataset(dim="output_dim_0")
data_vars = list(input_xr.data_vars)
output_features = output_features.rename(
{i: j
for i, j in zip(output_features.data_vars, data_vars)} # noqa pylint: disable=unnecessary-comprehension
)
# merge with predictions
output_xr = xr.merge([output_xr, output_features], compat="override")
return assign_crs(output_xr, str(crs))
from dask_ml.wrappers import ParallelPostFit
X, y = sklearn.datasets.make_classification(n_samples=1000)
clf = ParallelPostFit(SVC(gamma='scale'))
clf.fit(X, y)
Ns = [100_000, 200_000, 400_000, 800_000]
timings = []
for n in Ns:
X, y = dask_ml.datasets.make_classification(n_samples=n,
random_state=n,
chunks=n // 20)
t1 = tic()
# Serial scikit-learn version
clf.estimator.predict(X)
timings.append(('Scikit-Learn', n, tic() - t1))
t1 = tic()
# Parallelized scikit-learn version
clf.predict(X).compute()
timings.append(('dask-ml', n, tic() - t1))
df = pd.DataFrame(timings,
columns=['method', 'Number of Samples', 'Predict Time'])
ax = sns.factorplot(x='Number of Samples',
y='Predict Time',
hue='method',
data=df,
aspect=1.5)
import time
import datetime
# Start the computation.
n_samples_classification = 1000
n_samples = 100000000
chunks = n_samples//20
name = f'parallelizing_svm_800k_40k'
client = Client('10.255.23.115:8786')
X, y = sklearn.datasets.make_classification(n_samples=n_samples_classification)
clf = ParallelPostFit(SVC(gamma='scale'))
clf.fit(X, y)
X, y = dask_ml.datasets.make_classification(n_samples = n_samples,
random_state = n_samples,
chunks = chunks)
# Start the computation.
start = datetime.datetime.now()
results = clf.predict(X).compute(scheduler='distributed')
end = datetime.datetime.now()
print(f'Parallelizing svm is done in {end - start}')
clf.predict(X).visualize(filename=f'{name}.png')
def test_sklearn():
from sklearn.feature_extraction.text import CountVectorizer, TfidfTransformer
from sklearn.linear_model import SGDClassifier, LogisticRegressionCV
from sklearn.model_selection import GridSearchCV
from sklearn.pipeline import Pipeline
from sklearn.svm import SVC
from sklearn.externals import joblib
from sklearn.datasets import make_classification, load_digits, fetch_20newsgroups
from dask_ml.wrappers import ParallelPostFit
categories = [
'alt.atheism',
'talk.religion.misc',
]
print("Loading 20 newsgroups dataset for categories:")
print(categories)
data = fetch_20newsgroups(subset='train', categories=categories)
print("%d documents" % len(data.filenames))
print("%d categories" % len(data.target_names))
print()
pipeline = Pipeline([
('vect', CountVectorizer()),
('tfidf', TfidfTransformer()),
('clf', SGDClassifier(max_iter=1000)),
])
parameters = {
'vect__max_df': (0.5, 0.75, 1.0),
# 'vect__max_features': (None, 5000, 10000, 50000),
'vect__ngram_range': ((1, 1), (1, 2)), # unigrams or bigrams
# 'tfidf__use_idf': (True, False),
# 'tfidf__norm': ('l1', 'l2'),
# 'clf__alpha': (0.00001, 0.000001),
# 'clf__penalty': ('l2', 'elasticnet'),
# 'clf__n_iter': (10, 50, 80),
}
grid_search = GridSearchCV(pipeline,
parameters,
n_jobs=-1,
verbose=1,
cv=3,
refit=False,
iid=False)
grid_search.fit(data.data, data.target)
with joblib.parallel_backend('dask'):
grid_search.fit(data.data, data.target)
X, y = load_digits(return_X_y=True)
svc = ParallelPostFit(SVC(random_state=0, gamma='scale'))
param_grid = {
# use estimator__param instead of param
'estimator__C': [0.01, 1.0, 10],
}
grid_search = GridSearchCV(svc, param_grid, iid=False, cv=3)
grid_search.fit(X, y)
big_X = da.concatenate(
[da.from_array(X, chunks=X.shape) for _ in range(10)])
predicted = grid_search.predict(big_X)
#
X_train, y_train = make_classification(n_features=2,
n_redundant=0,
n_informative=2,
random_state=1,
n_clusters_per_class=1,
n_samples=1000)
N = 100
X_large = da.concatenate(
[da.from_array(X_train, chunks=X_train.shape) for _ in range(N)])
y_large = da.concatenate(
[da.from_array(y_train, chunks=y_train.shape) for _ in range(N)])
clf = ParallelPostFit(LogisticRegressionCV(cv=3))
clf.fit(X_train, y_train)
y_pred = clf.predict(X_large)
clf.score(X_large, y_large)
# est.partial_fit(X_train_1, y_train_1)
# from tpot import TPOTClassifier
pass
import numpy as np
import argparse
from joblib import load
parser = argparse.ArgumentParser()
parser.add_argument('--filename', type=str, dest='filename', help='path to the dataset to be scored')
args = parser.parse_args()
if __name__ == "__main__":
from distributed import Client, LocalCluster
from dask_ml.wrappers import ParallelPostFit
import dask.dataframe as dd
cluster = LocalCluster()
client = Client(cluster)
clf = load('wta-matches-model.joblib')
clf = ParallelPostFit(clf)
matches = dd.read_csv(args.filename, assume_missing=True)
point_diff = (matches.winner_rank_points - matches.loser_rank_points).dropna()
X_test = point_diff.compute().values[:, np.newaxis]
y_test_pred = clf.predict(X_test)
np.save("predictions.npy", y_test_pred)
def function_2(file_name, rows_to_parse):
data = pd.read_csv(os.path.join('data', 'nycflights', file_name +'.csv'),nrows=int(rows_to_parse)) #User
data = data.fillna(0)
print(data)
train, test = train_test_split(data,train_size=0.5, test_size=0.5)
train_x = train.drop(['DepDelay','UniqueCarrier','Origin','Dest'], axis=1)
train_y = train['DepDelay']
test_x = test.drop(['DepDelay','UniqueCarrier','Origin','Dest'], axis=1)
test_y = test['DepDelay']
# Support Vector Machines
GridSearch.support_vector_machine(train_x,train_y,test_x,test_y)
import time
start_time = time.time()
GridSearch.sklearn_grid_search(train_x, train_y)
print("--- %s seconds ---" % (time.time() - start_time))
#____DASK____
c = dask.distributed.Client()
client = Client(processes=False, threads_per_worker=4, n_workers=1, memory_limit='2GB')
print(client)
import time
start_time = time.time()
GridSearch.dask_grid_search(train_x, train_y)
print(f"--- {time.time() - start_time}seconds ---")
#DASK DELAY
output = []
#for x in data:
a = dask.delayed(GridSearch.support_vector_machine)(train_x,train_y,test_x,test_y)
print(a)
start_time = time.time()
a.compute()
print("--- %s seconds ---" % (time.time() - start_time))
output.append(a)
b = dask.delayed(GridSearch.sklearn_grid_search)(train_x, train_y)
print(b)
output.append(b)
start_time = time.time()
b.compute()
print("--- %s seconds ---" % (time.time() - start_time))
c = dask.delayed(GridSearch.dask_grid_search)(train_x, train_y)
print(c)
output.append(c)
start_time = time.time()
c.compute()
print("--- %s seconds ---" % (time.time() - start_time))
total = dask.delayed(sum)(output)
#Visaualize
total.visualize()
#Other Code:
clean_dataset(train_x)
train_x = train_x.values
train_x
train_y = train_y.values
train_y
from sklearn.preprocessing import Normalizer
x = train_x
transformer = Normalizer().fit(x)
transformer
transformer.transform(x)
train_x = transformer.transform(x)
train_x = train_x.round(decimals=2)
train_x
train_x, train_y = make_classification(
n_features=2, n_redundant=0, n_informative=2,
random_state=1, n_clusters_per_class=1, n_samples=1000)
train_x[:5]
train_y
# Scale up: increase N, the number of times we replicate the data.
N = 2
X_large = da.concatenate([da.from_array(train_x, chunks=train_x.shape) for _ in range(N)])
y_large = da.concatenate([da.from_array(train_y, chunks=train_y.shape) for _ in range(N)])
print(X_large)
clf = ParallelPostFit(LogisticRegressionCV(cv=3), scoring="r2")
y_pred = clf.predict(X_large)
print(y_pred)