# XGB
num_round = 400
xgb_params = {'objective':'multi:softprob', 'num_class':38,
'eta':.2, 'max_depth':6, 'colsample_bytree':.4, 'subsample':.8,
'silent':1, 'nthread':8}
pr_nn_test_dict = {60:np.zeros((N/2, 38)), 70:np.zeros((N/2, 38))}
pr_nn_train_dict = {60:np.zeros((N/2, 38)), 70:np.zeros((N/2, 38))}
scores = []
t0 = datetime.now()
for j in range(nModels):
seed = 9317*j + 3173
xgb_params['seed'] = seed
cols = feature_selection(X2[v_train-1], y, xgb_params, num_round)
X = X2.tocsc()[:,cols].tocsr().astype(np.float32)
for params in params_lst:
h1 = params['h1']
h2 = params['h2']
max_epochs = params['max_epochs']
pr_nn_train = pr_nn_train_dict[h1]
pr_nn_test = pr_nn_test_dict[h1]
np.random.seed(seed)
net1 = build_net_sparse_input(h1, h2, p, mm, bs=bs, max_epochs=max_epochs, num_in=X.shape[1],
shuffle=True, eval_size=False)
net1.fit(X[v_train-1], y)
pr = net1.predict_proba(X[v_train-1])
# net1.save_params_to(data_path+'model/nn002_h1_'+str(h1)+'_'+str(j)+'.pkl')
pr_nn_train += pr
pr_nn_test += net1.predict_proba(X[v_test-1])
p = .1
bs = 256
h1h2_lst = [{'h1':60, 'h2':100}, {'h1':70, 'h2':90}]
pr_nn_dict = {(params['h1'], params['h2']):[np.zeros((N/2, 38)) for _ in range(nIter)] \
for params in h1h2_lst}
scores = []
t0 = datetime.now()
for fold, idx in enumerate(kf):
train_idx, valid_idx = idx
params_dict = {'fold':fold}
for j in range(nModels):
seed = 11281*j + 9108
xgb_params['seed'] = seed
cols = feature_selection(X4[train_idx], y[train_idx], xgb_params, nt)
X = X4.tocsc()[:,cols].tocsr().astype(np.float32)
for params in h1h2_lst:
h1 = params['h1']
h2 = params['h2']
params_dict.update(params)
pr_avg = pr_nn_dict[(h1, h2)]
np.random.seed(seed)
net1 = build_net_sparse_input(h1, h2, p, mm, max_epochs=epochs, num_in=X.shape[1],
shuffle=True, eval_size=False)
for k in range(nIter):
net1.fit(X[train_idx], target[train_idx])
pr = net1.predict_proba(X[valid_idx])
pr_avg[k][valid_idx] += pr
sc = params_dict.copy()
sc.update({'epochs':epochs*(k+1), 'nModels':j+1,
