def compute_linear_predictions(X_train,
X_test_int,
X_test_other,
Y_train,
lin_ranked,
max_features=100,
regularize=0.8,
alpha=10.0):
max_features = 100
descriptors = loading.get_descriptors(format=True)
Y = pd.Panel(items=range(1, 50),
major_axis=X_test_other.index,
minor_axis=pd.Series(descriptors, name='Descriptor'))
for col, descriptor in enumerate(descriptors): # For each descriptor.
prog(col, len(descriptors))
X_test = X_test_int if col == 0 else X_test_other
n_molecules = X_test.shape[0]
est = Ridge(alpha=alpha,
fit_intercept=True,
normalize=False,
copy_X=True,
max_iter=None,
tol=0.001,
solver='auto',
random_state=0)
features = lin_ranked[col, :][:max_features]
for subject in range(1, 50):
# Perceptual data for this descriptor.
observed = Y_train[subject][descriptor]
# Fit the model on the training data with the
# 'max_features' features
est.fit(
X_train.drop('mean_dilution', 1).values[:, features], observed)
predicted = est.predict(
X_test.values[:, features]) # Predict the test data.
Y[subject][descriptor].loc[list(X_test.index)] = predicted
# Regularize each subject to the across-subject mean
Y_mean = Y.mean(axis=0)
for subject in range(1, 50):
Y[subject] = Y_mean * regularize + Y[subject] * (1 - regularize)
return Y
def rfc_final(X,Y_imp,Y_mask,
max_features,min_samples_leaf,max_depth,et,use_mask,trans_weight,
trans_params,X_test_int=None,X_test_other=None,Y_test=None,n_estimators=100,seed=0,quiet=False):
if X_test_int is None:
X_test_int = X
if X_test_other is None:
X_test_other = X
if Y_test is None:
Y_test = Y_mask
def rfc_maker(n_estimators=n_estimators,max_features=max_features,
min_samples_leaf=min_samples_leaf,max_depth=max_depth,et=False):
if not et:
kls = RandomForestRegressor
kwargs = {'oob_score':False}
else:
kls = ExtraTreesRegressor
kwargs = {}
return kls(n_estimators=n_estimators, max_features=max_features,
min_samples_leaf=min_samples_leaf, max_depth=max_depth,
n_jobs=-1, random_state=seed, **kwargs)
rfcs = {}
for col in range(42):
prog(col,42)
rfcs[col] = rfc_maker(n_estimators=n_estimators,
max_features=max_features[col],
min_samples_leaf=min_samples_leaf[col],
max_depth=max_depth[col],
et=et[col])
if use_mask[col]:
rfcs[col].fit(X,Y_mask[:,col])
else:
rfcs[col].fit(X,Y_imp[:,col])
predicted = np.zeros((X_test_int.shape[0],42))
for col in range(42):
if et[col] or not np.array_equal(X,X_test_int):
# Possibly check in-sample fit because there isn't any alternative.
if col in [0,21]:
predicted[:,col] = rfcs[col].predict(X_test_int)
else:
predicted[:,col] = rfcs[col].predict(X_test_other)
else:
try:
predicted[:,col] = rfcs[col].oob_prediction_
except AttributeError:
if col in [0,21]:
predicted[:,col] = rfcs[col].predict(X_test_int)
else:
predicted[:,col] = rfcs[col].predict(X_test_other)
def f_transform(x, k0, k1):
return 100*(k0*(x/100)**(k1*0.5) - k0*(x/100)**(k1*2))
for col in range(21):
tw = trans_weight[col]
k0,k1 = trans_params[col]
p_m = predicted[:,col]
p_s = predicted[:,col+21]
predicted[:,col+21] = tw*f_transform(p_m,k0,k1) + (1-tw)*p_s
observed = Y_test
score = scoring.score2(predicted,observed)
rs = {}
for kind in ['int','ple','dec']:
rs[kind] = {}
for moment in ['mean','sigma']:
rs[kind][moment] = scoring.r2(kind,moment,predicted,observed)
if not quiet:
print("For subchallenge 2:")
print("\tScore = %.2f" % score)
for kind in ['int','ple','dec']:
for moment in ['mean','sigma']:
print("\t%s_%s = %.3f" % (kind,moment,rs[kind][moment]))
return (rfcs,score,rs)
def rfc_final(X,
Y,
max_features,
min_samples_leaf,
max_depth,
use_et,
regularize=np.ones(21) * 0.8,
n_estimators=100,
seed=0):
descriptors = loading.get_descriptors(format=True)
n_subjects = 49
n_obs = X.shape[0]
def rfc_maker(n_estimators=n_estimators,
max_features=max_features,
min_samples_leaf=min_samples_leaf,
max_depth=max_depth,
use_et=False):
if not use_et:
kls = RandomForestRegressor
kwargs = {'oob_score': True}
else:
kls = ExtraTreesRegressor
kwargs = {}
return kls(n_estimators=n_estimators,
max_features=max_features,
min_samples_leaf=min_samples_leaf,
max_depth=max_depth,
n_jobs=-1,
random_state=seed,
**kwargs)
rfcs = {x: {} for x in range(1, n_subjects + 1)}
for d, descriptor in enumerate(descriptors):
for subject in range(1, n_subjects + 1):
rfcs[subject][descriptor] = rfc_maker(
n_estimators=n_estimators,
max_features=max_features[col],
min_samples_leaf=min_samples_leaf[col],
max_depth=max_depth[col],
use_et=use_et[col])
for subject in range(1, n_subjects + 1):
prog(subject, n_subjects + 1)
from time import gmtime, strftime
for d, descriptor in enumerate(descriptors):
rfcs[subject][descriptor].fit(X, Y[subject][descriptor])
predicted = np.zeros((n_obs, len(descriptors), n_subjects))
for col, descriptor in enumerate(descriptors):
for subject in range(1, n_subjects + 1):
if use_et[col]:
# Check in-sample fit because there isn't any alternative.
predicted[:,col,subject-1] = \
rfcs[subject][descriptor].predict(X)
else:
predicted[:,col,subject-1] = \
rfcs[subject][descriptor].oob_prediction_
# Regularize:
predicted_mean = predicted.mean(axis=2, keepdims=True)
predicted_reg = predicted.copy()
for d, descriptor in enumerate(descriptors):
predicted_reg[:,d,:] = regularize[d]*predicted_mean[:,d,:] \
+ (1-regularize[d])*predicted[:,d,:]
predicted = predicted_reg
observed = predicted.copy()
for subject in range(1, n_subjects + 1):
observed[:, :, subject - 1] = Y[subject]
score = scoring.score(predicted, observed)
rs = {}
predictions = {}
print("For subchallenge 1:")
print("\tScore = %.2f" % score)
for kind in ['int', 'ple', 'dec']:
rs[kind] = scoring.r(kind, predicted, observed)
print("\t%s = %.3f" % (kind, rs[kind]))
return (rfcs, score, rs)
def rfc_final(X,Y,
max_features,min_samples_leaf,max_depth,use_et,
regularize=np.ones(21)*0.8,Y_test=None,n_estimators=100,seed=0):
def rfc_maker(n_estimators=n_estimators,max_features=max_features,
min_samples_leaf=min_samples_leaf,max_depth=max_depth,
use_et=False):
if not use_et:
kls = RandomForestRegressor
kwargs = {'oob_score':True}
else:
kls = ExtraTreesRegressor
kwargs = {}
return kls(n_estimators=n_estimators, max_features=max_features,
min_samples_leaf=min_samples_leaf, max_depth=max_depth,
n_jobs=-1, random_state=seed, **kwargs)
rfcs = {}
for col in range(21):
rfcs[col] = {}
for subject in range(1,50):
rfcs[col][subject] = rfc_maker(n_estimators=n_estimators,
max_features=max_features[col],
min_samples_leaf=min_samples_leaf[col],
max_depth=max_depth[col],
use_et=use_et[col])
for subject in range(1,50):
prog(subject,50)
from time import gmtime, strftime
for col in range(21):
rfcs[col][subject].fit(X,Y[subject][:,col])
predicted0 = rfcs[0][1].predict(X)
predicted = np.zeros((predicted0.shape[0],21,49))
for col in range(21):
for subject in range(1,50):
if use_et[col]:
# Check in-sample fit because there isn't any alternative.
predicted[:,col,subject-1] = rfcs[col][subject].predict(X)
else:
predicted[:,col,subject-1] = rfcs[col][subject].oob_prediction_
# Regularize:
predicted_mean = predicted.mean(axis=2,keepdims=True)
predicted_reg = predicted.copy()
for col in range(21):
predicted_reg[:,col,:] = regularize[col]*predicted_mean[:,col,:] \
+ (1-regularize[col])*predicted[:,col,:]
predicted = predicted_reg
observed = predicted.copy()
for subject in range(1,50):
observed[:,:,subject-1] = Y[subject]
score = scoring.score(predicted,observed)
rs = {}
predictions = {}
print("For subchallenge 1:")
print("\tScore = %.2f" % score)
for kind in ['int','ple','dec']:
rs[kind] = scoring.r(kind,predicted,observed)
print("\t%s = %.3f" % (kind,rs[kind]))
return (rfcs,score,rs)