def hotStartTrainRMSE(model_name, K, L, X1, X2, train_I, train_J, train_Y, train_op):
M = X1.shape[0]
N = X2.shape[0]
Xs = train_op['params']['Xs']
betas = train_op['params']['betas']
r = train_op['params']['r']
if model_name.upper().split('_')[1] == 'LINEAR':
predictions = mmbae_linear_predict(train_I, train_J, M, N, Xs, betas, r, K, L)
Z = sp.csr_matrix((train_Y, (train_I,train_J)), shape=(M,N))
nonzero_Z = np.array(Z[(train_I,train_J)]).ravel()
hotStartTrainRMSE = np.sqrt(np.mean((predictions-nonzero_Z)**2))
return hotStartTrainRMSE
def hotStartValRMSE(model_name, K, L, X1, X2, val_I, val_J, val_Y, train_op):
M = X1.shape[0]
N = X2.shape[0]
Xs = train_op['params']['Xs']
betas = train_op['params']['betas']
r = [np.zeros((M,K)),np.zeros((N,L))]
ones = np.ones((len(train_op['train_I']),))
mu = sp.csr_matrix((ones, (train_op['train_I'],train_op['train_J'])), shape=(M,N)).sum(1)
mv = sp.csr_matrix((ones, (train_op['train_I'],train_op['train_J'])), shape=(M,N)).sum(0)
mu[mu<1] = 1
mv[mv<1] = 1
for k in range(K):
r[0][:,k] = np.array(np.divide(sp.csr_matrix((train_op['params']['r'][0][:,k], (train_op['train_I'],train_op['train_J'])), shape=(M,N)).sum(1),mu).flatten())[0]
for l in range(L):
r[1][:,l] = np.array(np.divide(sp.csr_matrix((train_op['params']['r'][1][:,l], (train_op['train_I'],train_op['train_J'])), shape=(M,N)).sum(0),mv).flatten())[0]
if model_name.upper().split('_')[1] == 'LINEAR':
r1_obs = r[0][val_I,:]
r2_obs = r[1][val_J,:]
predictions = mmbae_linear_predict(val_I, val_J, M, N, Xs, betas, [r1_obs,r2_obs], K, L)
Z = sp.csr_matrix((val_Y, (val_I,val_J)), shape=(M,N))
nonzero_Z = np.array(Z[(val_I,val_J)]).ravel()
hotStartTrainRMSE = np.sqrt(np.mean((predictions-nonzero_Z)**2))
return hotStartTrainRMSE
