Liyah, stop here. I can explain what I am doing in the next section another time.
Liyah, stop here. I can explain what I am doing in the next section another time.#
# Dr. Gerkin you left this for me a while back but we never came back to it
from sklearn.feature_selection import RFE, RFECV
svr = SVR(C=10, kernel='linear')
rfe = RFE(svr, n_features_to_select=1, step=10)
#rfe.fit(Xn, y)
rfe.ranking_.max()
svr = SVR(C=10, kernel='rbf')
ns = range(1, 109, 1)
rs = pd.Series(index=ns, dtype=float)
for n in tqdm(ns):
Xn_ = Xn[Xn.columns[rfe.ranking_ <= n]]
y_predict = cross_val_predict(svr, Xn_, y, cv=LeaveOneOut(), n_jobs=-1)
rs[n] = np.corrcoef(y, y_predict)[0, 1]
rs.plot()
Cs = np.logspace(-3, 3, 13)
rs_in = pd.Series(index=Cs, dtype=float)
rs_out = pd.Series(index=Cs, dtype=float)
rhos_out = pd.Series(index=Cs, dtype=float)
Xn_ = Xn[Xn.columns[rfe.ranking_ <= 10]]
for C in tqdm(Cs):
svr = SVR(C=C, kernel='rbf')
clf = svr
#y_predict = cross_val_predict(rfr, Xn, y, cv=LeaveOneOut(), n_jobs=-1)
clf.fit(Xn_, y)
y_predict_in = clf.predict(Xn_)
y_predict_out = cross_val_predict(clf, Xn_, y, cv=LeaveOneOut(), n_jobs=-1)
y_predict_in = np.clip(y_predict_in, 0, np.inf)
y_predict_out = np.clip(y_predict_out, 0, np.inf)
rs_in[C] = pearsonr(y, y_predict_in)[0]
rs_out[C] = pearsonr(y, y_predict_out)[0]
rhos_out[C] = spearmanr(y, y_predict_out)[0]
rs_in.plot(label='In-sample prediction R')
rs_out.plot(label='Out-of-sample prediction R')
rhos_out.plot(label=r'Out-of-sample prediction $\rho$')
plt.xscale('log')
plt.ylabel('Correlation\n(predicted vs observed)')
plt.xlabel('C (SVR hyperparameter)')
plt.legend(fontsize=10)
Xn_ = Xn[Xn.columns[rfe.ranking_ <= 10]]
rfr = RandomForestRegressor(n_estimators=100)
y_predict = cross_val_predict(rfr, Xn_, y, cv=LeaveOneOut(), n_jobs=-1)
np.corrcoef(y, y_predict)[0, 1]
rfr.fit(Xn_, y)
pd.Series(rfr.feature_importances_, Xn_.columns).sort_values(ascending=False).head(25)
from sklearn.model_selection import cross_val_predict
from sklearn.model_selection import LeaveOneOut
sns.set_style('whitegrid')
svr = SVR(C=10, kernel='rbf')
svr.fit(Xn_mordred, y_mordred)
y_predict = cross_val_predict(svr, Xn_mordred, y_mordred, cv=LeaveOneOut(), n_jobs=-1)
y_predict = np.clip(y_predict, 0, np.inf)
plt.figure(figsize=(5, 5))
plt.scatter(y_mordred, y_predict, alpha=0.3)
#maxx = y.max()*1.01
#plt.plot([0, maxx], [0, maxx], '--')
plt.plot([0, 4], [0, 4], '--')
plt.xlim(0, 4)
plt.ylim(0, 4)
#plt.xlim(0.9, maxx)
#plt.ylim(0.9, maxx)
plt.title('R = %.2g' % np.corrcoef(y_mordred, y_predict)[0, 1])
#ticks = range(5)
#plt.xticks(ticks)
#plt.yticks(ticks)
plt.xticks([1,2,3,4], ["1x", "10x", "100x", "1000x"])
plt.yticks([1,2,3,4], ["1x", "10x", "100x", "1000x"])
# plt.xticks(ticks, ['%d' % 10**x for x in ticks])
# plt.yticks(ticks, ['%d' % 10**x for x in ticks])
plt.xlabel('Actual Detection Threshold Ratio')
plt.ylabel('Predicted Detection\nThreshold Ratio')
from sklearn.model_selection import cross_val_predict
from sklearn.model_selection import LeaveOneOut
sns.set_style('whitegrid')
svr = SVR(C=10, kernel='rbf')
svr.fit(Xn_morgan, y_morgan)
y_predict = cross_val_predict(svr, Xn_morgan, y_morgan, cv=LeaveOneOut(), n_jobs=-1)
y_predict = np.clip(y_predict, 0, np.inf)
plt.figure(figsize=(5, 5))
plt.scatter(y_morgan, y_predict, alpha=0.3)
plt.plot([0, 4], [0, 4], '--')
plt.xlim(0, 4)
plt.ylim(0, 4)
plt.title('R = %.2g' % np.corrcoef(y_morgan, y_predict)[0, 1])
plt.xticks([1,2,3,4], ["1x", "10x", "100x", "1000x"])
plt.yticks([1,2,3,4], ["1x", "10x", "100x", "1000x"])
plt.xlabel('Actual Detection Threshold Ratio')
plt.ylabel('Predicted Detection\nThreshold Ratio')
sns.heatmap(svr.support_vectors_, cmap='RdBu_r')
abraham = pyrfume.load_data('abraham_2011/abraham-2011-with-CIDs.csv')
abraham = abraham.dropna(subset=['SMILES'])
from pyrfume.features import smiles_to_mordred
abraham_mordred = smiles_to_mordred(abraham['SMILES'].values)
#X = abraham_mordred.join(abraham.set_index('SMILES').drop(['Substance', 'Group'], axis=1), how='inner', rsuffix='abr_')
abe_ok = ['MW',
'Log (1/ODT)',
'E',
'S',
'A',
'B',
'L',
'V',
'M',
'AL',
'AC',
'ES',
'C1',
'C1AC',
'C1AL',
'HS',
'C2',
'C2Al',
'C2AC']
X = abraham_mordred.join(abraham.set_index('SMILES')[['Log (1/ODT)']], how='inner', rsuffix='abr_')
y = X['Log (1/ODT)']
X = X.astype('float').drop('Log (1/ODT)', axis=1)
#finite_all = X.dropna(axis=1).columns.intersection(finite_mordred)
X = X[finite_mordred].fillna(0)
X = X.dropna(axis=1)
Xn = X.copy()
Xn[:] = StandardScaler().fit_transform(X)
Cs = np.logspace(-3, 3, 13)
rs_in = pd.Series(index=Cs, dtype=float)
rs_out = pd.Series(index=Cs, dtype=float)
rhos_out = pd.Series(index=Cs, dtype=float)
for C in tqdm(Cs):
svr = SVR(C=C, kernel='rbf')
clf = svr
#y_predict = cross_val_predict(rfr, Xn, y, cv=LeaveOneOut(), n_jobs=-1)
clf.fit(Xn, y)
y_predict_in = clf.predict(Xn)
y_predict_out = cross_val_predict(clf, Xn, y, cv=LeaveOneOut(), n_jobs=-1)
y_predict_in = np.clip(y_predict_in, 0, np.inf)
y_predict_out = np.clip(y_predict_out, 0, np.inf)
rs_in[C] = pearsonr(y, y_predict_in)[0]
rs_out[C] = pearsonr(y, y_predict_out)[0]
rhos_out[C] = spearmanr(y, y_predict_out)[0]
rs_in.plot(label='In-sample prediction R')
rs_out.plot(label='Out-of-sample prediction R')
rhos_out.plot(label=r'Out-of-sample prediction $\rho$')
plt.xscale('log')
plt.ylabel('Correlation\n(predicted vs observed)')
plt.xlabel('C (SVR hyperparameter)')
plt.legend(fontsize=10)
svr = SVR(C=1, kernel='rbf')
svr.fit(Xn, y)
X_ = mordred_data[finite_mordred]; y_ = mordred_data['log_abs']
X_ = X_.join(half_coleman_data.set_index('SMILES String')['det'])
X_ = X_[y_ < 10]
y_ = y_[y_ < 10]
Xn_ = pd.DataFrame(StandardScaler().fit_transform(X_), index=X_.index, columns=X_.columns)
X_['det'] = svr.predict(Xn_.drop('det', axis=1))
#y_predict = cross_val_predict(svr, Xn, y, cv=LeaveOneOut(), n_jobs=-1)
Xn_ = pd.DataFrame(StandardScaler().fit_transform(X_), index=X_.index, columns=X_.columns)
Cs = np.logspace(-3, 3, 13)
rs_in = pd.Series(index=Cs, dtype=float)
rs_out = pd.Series(index=Cs, dtype=float)
rhos_out = pd.Series(index=Cs, dtype=float)
for C in tqdm(Cs):
svr = SVR(C=C, kernel='rbf')
clf = svr
#y_predict = cross_val_predict(rfr, Xn, y, cv=LeaveOneOut(), n_jobs=-1)
clf.fit(Xn_, y_)
y_predict_in = clf.predict(Xn_)
y_predict_out = cross_val_predict(clf, Xn_, y_, cv=LeaveOneOut(), n_jobs=-1)
y_predict_in = np.clip(y_predict_in, 0, np.inf)
y_predict_out = np.clip(y_predict_out, 0, np.inf)
rs_in[C] = pearsonr(y_, y_predict_in)[0]
rs_out[C] = pearsonr(y_, y_predict_out)[0]
rhos_out[C] = spearmanr(y_, y_predict_out)[0]
rs_in.plot(label='In-sample prediction R')
rs_out.plot(label='Out-of-sample prediction R')
rhos_out.plot(label=r'Out-of-sample prediction $\rho$')
plt.xscale('log')
plt.ylabel('Correlation\n(predicted vs observed)')
plt.xlabel('C (SVR hyperparameter)')
plt.legend(fontsize=10)