def test_partial_fit_regression(self):
data = make_regression(100, 2, 1)
df = pd.DataFrame(data[0])
df["label"] = data[1]
with df.model() as m:
fit_partial = m.fit(
SkModel(
MLPRegressor(max_iter=1, random_state=42),
FeaturesAndLabels(features=[0, 1], labels=['label'])
),
FittingParameter(
naive_splitter(0.3),
batch_size=10,
fold_epochs=10
)
)
with df.model() as m:
fit = m.fit(
SkModel(
MLPRegressor(max_iter=10, random_state=42),
FeaturesAndLabels(features=[0, 1], labels=['label'])
),
FittingParameter(naive_splitter(0.3))
)
self.assertAlmostEqual(df.model.predict(fit.model).iloc[0,-1], df.model.predict(fit_partial.model).iloc[0,-1], 4)
def test_regressor(self):
"""given some toy regression data"""
df = pd.DataFrame({
"a": [-1.0, 0.0, 1.0, 2.0, 3.0, 4.0],
"b": [-2.0, 1.0, 4.0, 7.0, 10.0, 13.0]
})
"""and a model"""
model = self.provide_regression_model(
FeaturesAndLabels(features=["a"], labels=["b"]))
"""when we fit the model"""
batch_size, epochs = self.provide_batch_size_and_epoch()
with df.model() as m:
fit = m.fit(model,
FittingParameter(splitter=naive_splitter(0.3),
batch_size=batch_size,
epochs=epochs),
verbose=0)
print(fit.training_summary.df)
self.assertEqual(4, len(fit.training_summary.df))
self.assertEqual(2, len(fit.test_summary.df))
"""then we can predict"""
prediction = df.model.predict(fit.model)
np.testing.assert_array_almost_equal(prediction.iloc[:, 0].values,
df["b"].values, 1)
"""and save and load the model"""
temp = os.path.join(tempfile.gettempdir(), str(uuid.uuid4()))
try:
fit.model.save(temp)
copy = Model.load(temp)
pd.testing.assert_frame_equal(df.model.predict(fit.model),
df.model.predict(copy),
check_less_precise=True)
finally:
os.remove(temp)
def test_linear_model(self):
df = DF_NOTES.copy()
with df.model() as m:
fit = m.fit(
SkModel(
Lasso(),
FeaturesAndLabels(
features=[
lambda df: df["variance"],
lambda df: (df["skewness"] / df["kurtosis"]).rename("engineered")
],
labels=[
'authentic'
]
)
),
FittingParameter(naive_splitter())
)
print(fit)
prediction = df.model.predict(fit.model)
print(prediction)
backtest = df.model.backtest(fit.model)
self.assertLess(backtest.model.sk_model.coef_[0], 1e-5)
def test_simple_sample_split_multiindex_row(self):
def check_test(test_data):
self.assertIn("A", test_data[0].x.index)
self.assertIn("B", test_data[0].x.index)
sampler = Sampler(XYWeight(TEST_MUTLI_INDEX_ROW_DF),
splitter=naive_splitter(
0.5, partition_row_multi_index=True),
after_fold=check_test,
epochs=1)
samples = list(sampler.sample_for_training())
self.assertEqual(1, len(samples))
self.assertEqual(4, len(samples[0].x))
self.assertIn("A", samples[0].x.index)
self.assertIn("B", samples[0].x.index)
def test_mult_epoch_cross_validation(self):
df = pd.DataFrame({
"a": [
1,
0,
1,
0,
1,
0,
1,
0,
],
"b": [
0,
1,
0,
1,
1,
0,
1,
0,
],
})
with df.model() as m:
class NN(PytorchNN):
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
self.nn = nn.Sequential(
nn.Linear(1, 2),
nn.ReLU(),
nn.Linear(2, 1),
)
def forward_training(self, x):
return self.nn(x)
fit = m.fit(
PytorchModel(NN, FeaturesAndLabels(["a"], ["b"]), nn.MSELoss,
Adam),
FittingParameter(splitter=naive_splitter(0.5),
epochs=2,
fold_epochs=10,
batch_size=2))
print(fit)
def test_no_test_data(self):
"""given some toy regression data"""
df = pd.DataFrame({
"a": [-1.0, 0.0, 1.0, 2.0, 3.0, 4.0],
"b": [-2.0, 1.0, 4.0, 7.0, 10.0, 13.0]
})
"""and a model"""
model = self.provide_regression_model(
FeaturesAndLabels(features=["a"], labels=["b"]))
"""when we fit the model"""
batch_size, epochs = self.provide_batch_size_and_epoch()
with df.model() as m:
fit = m.fit(model,
FittingParameter(splitter=naive_splitter(0),
batch_size=batch_size,
epochs=epochs),
verbose=0)
# print(fit.training_summary.df)
print(fit.test_summary.df)
"""then we have an empty test data frame"""
self.assertEqual(len(fit.training_summary.df), len(df))
self.assertEqual(len(fit.test_summary.df), 0)
def test_multi_sample_regressor(self):
"""given some toy regression data"""
df = pd.DataFrame({
"a": [-1.0, 0.0, 1.0, 2.0, 3.0, 4.0],
"b": [-2.0, 1.0, 4.0, 7.0, 10.0, 13.0]
})
"""and a model"""
model = self.provide_regression_model(
FeaturesAndLabels(features=["a"], labels=["b"]))
"""when we fit the model"""
batch_size, epochs = self.provide_batch_size_and_epoch()
with df.model() as m:
fit = m.fit(model,
FittingParameter(splitter=naive_splitter(0.3),
batch_size=batch_size,
epochs=epochs),
verbose=0)
print(fit.training_summary.df)
"""then we can predict"""
prediction = df.model.predict(fit.model, samples=2)
np.testing.assert_array_almost_equal(
prediction.iloc[:, 0]._.values,
np.concatenate([df[["b"]].values, df[["b"]].values], axis=1), 1)
def test_auto_encoder(self):
"""given the implementation can handle auto encoders"""
model = self.provide_auto_encoder_model(
FeaturesAndLabels(features=["a", "b"],
labels=["a", "b"],
latent=["x"]))
if model is None:
return
"""and some toy classification data"""
df = pd.DataFrame({
"a": [1, 0] * 10,
"b": [0, 1] * 10,
})
"""when we fit the model"""
batch_size, epochs = self.provide_batch_size_and_epoch()
with df.model() as m:
fit = m.fit(model,
FittingParameter(splitter=naive_splitter(0.49),
batch_size=batch_size,
epochs=epochs),
verbose=0)
print(fit.training_summary.df)
"""then we can predict Autoencoded"""
auto_encoded_prediction = df.model.predict(fit.model)
self.assertEqual((20, 2), auto_encoded_prediction["prediction"].shape)
"""and we can encode"""
encoded_prediction = df.model.predict(fit.model.as_encoder())
print(encoded_prediction)
self.assertEqual((20, 1), encoded_prediction["prediction"].shape)
"""and we can decode"""
decoded_prediction = encoded_prediction["prediction"].model.predict(
fit.model.as_decoder())
print(decoded_prediction)
np.testing.assert_array_almost_equal(
decoded_prediction["prediction"].values > 0.5, df[["a",
"b"]].values)
"""and we can encoder and decode after safe and load"""
temp = os.path.join(tempfile.gettempdir(), str(uuid.uuid4()))
try:
fit.model.save(temp)
copy = Model.load(temp)
pd.testing.assert_frame_equal(df.model.predict(
fit.model.as_encoder()),
df.model.predict(copy.as_encoder()),
check_less_precise=True)
pd.testing.assert_frame_equal(
encoded_prediction.model.predict(fit.model.as_decoder()),
encoded_prediction.model.predict(copy.as_decoder()),
check_less_precise=True)
finally:
os.remove(temp)
# try to save only as encoder model
try:
fit.model.as_encoder().save(temp)
copy = Model.load(temp)
finally:
os.remove(temp)
def test_classifier(self):
"""given some toy classification data"""
df = pd.DataFrame({
"a": [
1,
0,
1,
0,
1,
0,
1,
0,
],
"b": [
0,
0,
1,
1,
0,
0,
1,
1,
],
"c": [
1,
0,
0,
1,
1,
0,
0,
1,
]
})
"""and a model"""
model = self.provide_classification_model(
FeaturesAndLabels(features=["a", "b"],
labels=["c"],
label_type=int))
temp = os.path.join(tempfile.gettempdir(), str(uuid.uuid4()))
"""when we fit the model"""
batch_size, epochs = self.provide_batch_size_and_epoch()
with df.model(temp) as m:
fit = m.fit(model,
FittingParameter(splitter=naive_splitter(0.49),
batch_size=batch_size,
epochs=epochs),
verbose=0)
print(fit.training_summary.df)
# fit.training_summary.df.to_pickle('/tmp/classifier.df')
# print(fit._repr_html_())
"""then we get a html summary and can predict"""
self.assertIn('<style>', fit.training_summary._repr_html_())
prediction = df.model.predict(fit.model)
binary_prediction = prediction.iloc[:, 0] >= 0.5
np.testing.assert_array_equal(
binary_prediction,
np.array([
True,
False,
False,
True,
True,
False,
False,
True,
]))
"""and load the model"""
try:
copy = Model.load(temp)
pd.testing.assert_frame_equal(df.model.predict(fit.model),
df.model.predict(copy),
check_less_precise=True)
# test using context manager and ForecastProvider
pd.testing.assert_frame_equal(
df.model(temp).predict(forecast_provider=Forecast).df,
df.model.predict(copy),
check_less_precise=True)
finally:
os.remove(temp)
def test_probabilistic_model_with_callback(self):
try:
pandas_ml_quant_data_provider = importlib.import_module(
"pandas_ml_quant")
from pandas_ml_quant import PricePredictionSummary
from pandas_ml_quant.model.summary.price_prediction_summary import PriceSampledSummary
except:
print("pandas_ml_quant not found, skipping!")
return
df = pd.DataFrame({
"Returns":
np.random.normal(-0.02, 0.03, 500) +
np.random.normal(0.03, 0.02, 500)
})
fl = PostProcessedFeaturesAndLabels(
features=["Returns"],
feature_post_processor=lambda df: df.ta.rnn(20),
labels=[
lambda df: df["Returns"].shift(-1).rename("Future_Returns")
],
targets=lambda df: (1 + df["Returns"]).cumprod().rename("Close"))
model_factory = PytorchNNFactory.create(
nn.Sequential(
nn.Linear(20, 10),
nn.Tanh(),
nn.Linear(10, 6),
LambdaSplitter(
lambda x: T.softmax(x[..., :2], dim=1),
lambda x: T.exp(x[..., 2:4]),
# enforce one mean positive and the other negativ
lambda x: T.cat([T.exp(x[..., 4:5]), -T.exp(x[..., 5:6])],
dim=1),
)),
predictor=lambda n, i: T.cat(n(i), dim=1),
trainer=lambda n, i: n(i))
def dist(probs, scales, locs):
return MixtureSameFamily(Categorical(probs=probs),
Normal(loc=locs, scale=scales))
def loss(y_pred):
probs, scales, locs = y_pred
return dist(probs, scales, locs)
def cdf_cb(arg):
probs, scales, locs = arg[..., :2], arg[..., 2:4], arg[..., 4:6]
return dist(probs, scales, locs)
summary_provider = PriceSampledSummary.with_reconstructor(
sampler=wrap_applyable(lambda params, samples: cdf_cb(params).
sample([int(samples.item())]),
nr_args=2),
samples=100,
confidence=0.8)
model = PytorchModel(module_provider=model_factory,
features_and_labels=fl,
criterion_provider=lambda: DistributionNLL(
loss, penalize_toal_variance_lambda=1.1),
optimizer_provider=Adam,
summary_provider=summary_provider)
fit = df.model.fit(
model,
FittingParameter(epochs=10,
batch_size=6,
splitter=naive_splitter(0.25)),
#verbose=1,
callbacks=[
TestConfidenceInterval(
TestConfidenceInterval.CdfConfidenceInterval(
wrap_applyable(
lambda params, val: cdf_cb(params).cdf(val),
nr_args=2),
interval=0.8),
wrap_applyable(lambda params: cdf_cb(params).variance),
early_stopping=True)
])
print(fit.test_summary.calc_scores())
def feature_selection(
self,
features_and_labels: FeaturesAndLabels,
training_data_splitter: Callable = naive_splitter(0.2),
correlated_features_th: float = 0.75,
rfecv_splits: int = 4,
forest_splits: int = 7,
min_features_to_select: int = 1,
is_time_series: bool = False,
**kwargs):
assert features_and_labels.label_type in ('regression', 'classification', int, float, bool), \
"label_type need to be specified: 'regression' | 'classification' !"
# find best parameters
with self() as m:
# extract features and labels
ext = m.extract(features_and_labels)
# first perform a correlation analysis and remove correlating features !!!
if correlated_features_th > 0:
_, pairs = get_correlation_pairs(ext.features)
redundant_correlated_features = {
i[0]: p
for i, p in pairs.items() if p > correlated_features_th
}
_log.warning(
f"drop redundant features: {redundant_correlated_features}"
)
features_and_labels = PostProcessedFeaturesAndLabels.from_features_and_labels(
features_and_labels,
feature_post_processor=lambda df: df.drop(
redundant_correlated_features.keys(), axis=1))
# estimate model type and sample properties
is_classification = 'float' not in (str(
features_and_labels.label_type))
nr_samples = len(self.df)
if is_classification:
nr_classes = len(ext.labels.value_counts())
else:
nr_classes = max(len(self.df) / 3, 100)
# estimate grid search parameters
grid = {
"estimator__n_estimators":
sp_randint(10, 500),
"estimator__max_depth": [2, None],
"estimator__min_samples_split":
sp_randint(2, nr_samples / nr_classes),
"estimator__min_samples_leaf":
sp_randint(2, nr_samples / nr_classes),
"estimator__bootstrap": [True, False],
"estimator__criterion":
["gini", "entropy"] if is_classification else ["mse", "mae"]
}
# build model
cross_validation = TimeSeriesSplit if is_time_series else StratifiedKFold if is_classification else KFold
estimator = RandomForestClassifier(
) if is_classification else RandomForestRegressor()
selector = RFECV(estimator,
step=1,
cv=cross_validation(rfecv_splits),
min_features_to_select=min_features_to_select)
skm = RandomizedSearchCV(selector,
param_distributions=grid,
cv=cross_validation(forest_splits),
n_jobs=-1)
# fit model
fit = m.fit(
SkModel(skm,
features_and_labels=features_and_labels,
summary_provider=FeatureSelectionSummary),
FittingParameter(splitter=training_data_splitter))
# we hide the loss plot from this summary
return fit.with_hidden_loss_plot()