def ConvMF(res_dir,
train_user,
train_item,
valid_user,
test_user,
R,
CNN_X,
vocab_size,
init_W=None,
give_item_weight=True,
max_iter=50,
lambda_u=1,
lambda_v=100,
dimension=50,
dropout_rate=0.2,
emb_dim=200,
max_len=300,
num_kernel_per_ws=100):
# explicit setting
a = 1
b = 0
num_user = R.shape[0]
num_item = R.shape[1]
PREV_LOSS = 1e-50
if not os.path.exists(res_dir):
os.makedirs(res_dir)
f1 = open(res_dir + '/state.log', 'w')
Train_R_I = train_user[1]
Train_R_J = train_item[1]
Test_R = test_user[1]
Valid_R = valid_user[1]
if give_item_weight is True:
item_weight = np.array([math.sqrt(len(i)) for i in Train_R_J],
dtype=float)
item_weight = (float(num_item) / item_weight.sum()) * item_weight
else:
item_weight = np.ones(num_item, dtype=float)
pre_val_eval = 1e10
cnn_module = CNN_module(dimension, vocab_size, dropout_rate, emb_dim,
max_len, num_kernel_per_ws, init_W)
theta = cnn_module.get_projection_layer(CNN_X)
np.random.seed(133)
U = np.random.uniform(size=(num_user, dimension))
V = theta
endure_count = 5
count = 0
for iteration in range(max_iter):
loss = 0
tic = time.time()
print("%d iteration\t(patience: %d)" % (iteration, count))
VV = b * (V.T.dot(V)) + lambda_u * np.eye(dimension)
sub_loss = np.zeros(num_user)
for i in range(num_user):
idx_item = train_user[0][i]
V_i = V[idx_item]
R_i = Train_R_I[i]
A = VV + (a - b) * (V_i.T.dot(V_i))
B = (a * V_i * (np.tile(R_i, (dimension, 1)).T)).sum(0)
U[i] = np.linalg.solve(A, B)
sub_loss[i] = -0.5 * lambda_u * np.dot(U[i], U[i])
loss = loss + np.sum(sub_loss)
sub_loss = np.zeros(num_item)
UU = b * (U.T.dot(U))
for j in range(num_item):
idx_user = train_item[0][j]
U_j = U[idx_user]
R_j = Train_R_J[j]
tmp_A = UU + (a - b) * (U_j.T.dot(U_j))
A = tmp_A + lambda_v * item_weight[j] * np.eye(dimension)
B = (a * U_j * (np.tile(R_j, (dimension, 1)).T)
).sum(0) + lambda_v * item_weight[j] * theta[j]
V[j] = np.linalg.solve(A, B)
sub_loss[j] = -0.5 * np.square(R_j * a).sum()
sub_loss[j] = sub_loss[j] + a * np.sum((U_j.dot(V[j])) * R_j)
sub_loss[j] = sub_loss[j] - 0.5 * np.dot(V[j].dot(tmp_A), V[j])
loss = loss + np.sum(sub_loss)
seed = np.random.randint(100000)
history = cnn_module.train(CNN_X, V, item_weight, seed)
theta = cnn_module.get_projection_layer(CNN_X)
cnn_loss = history.history['loss'][-1]
loss = loss - 0.5 * lambda_v * cnn_loss * num_item
tr_eval = eval_RMSE(Train_R_I, U, V, train_user[0])
val_eval = eval_RMSE(Valid_R, U, V, valid_user[0])
te_eval = eval_RMSE(Test_R, U, V, test_user[0])
toc = time.time()
elapsed = toc - tic
converge = abs((loss - PREV_LOSS) / PREV_LOSS)
if (val_eval < pre_val_eval):
cnn_module.save_model(res_dir + '/CNN_weights.hdf5')
np.savetxt(res_dir + '/U.dat', U)
np.savetxt(res_dir + '/V.dat', V)
np.savetxt(res_dir + '/theta.dat', theta)
else:
count = count + 1
pre_val_eval = val_eval
print(
"Loss: %.5f Elpased: %.4fs Converge: %.6f Tr: %.5f Val: %.5f Te: %.5f"
% (loss, elapsed, converge, tr_eval, val_eval, te_eval))
f1.write(
"Loss: %.5f Elpased: %.4fs Converge: %.6f Tr: %.5f Val: %.5f Te: %.5f\n"
% (loss, elapsed, converge, tr_eval, val_eval, te_eval))
if (count == endure_count):
break
PREV_LOSS = loss
f1.close()
def DRMF(res_dir,
train_user,
train_item,
valid_user,
test_user,
R,
DNN_X,
DNN_Y,
vocab_size,
init_W=None,
give_weight=True,
max_iter=50,
lambda_u=1,
lambda_v=100,
dimension=50,
dropout_rate=0.2,
emb_dim=200,
num_kernel_per_ws=50,
dnn_type='CNN_GRU',
reg_schema='Dual',
gru_outdim=50,
maxlen_doc=[10, 10],
maxlen_sent=[30, 30]):
# explicit setting
num_user = R.shape[0]
num_item = R.shape[1]
PREV_LOSS = 1e-50
if not os.path.exists(res_dir):
os.makedirs(res_dir)
f1 = open(res_dir + '/state.log', 'a')
f1.write("### DRMF-%s-%s ###\n\n" % (reg_schema, dnn_type))
f1.write("===Configuration===\n")
f1.write("lambda_u=%f, lambda_v=%f\n" % (lambda_u, lambda_v))
f1.write("maxlen_doc=[%d,%d], maxlen_sent=[%d,%d]\n" %
(maxlen_doc[0], maxlen_doc[1], maxlen_sent[0], maxlen_sent[1]))
f1.write(
"emb_dim=%d, dimension=%d, num_kernel_per_ws=%d, dropout_rate=%.2f, gru_outdim=%d\n\n"
% (emb_dim, dimension, num_kernel_per_ws, dropout_rate, gru_outdim))
f1.write("Tr:Training, Val:Validation, Te:Test, []: [MAE, MSE, RMSE]\n")
Train_R_I = train_user[1]
Train_R_J = train_item[1]
Test_R = test_user[1]
Valid_R = valid_user[1]
if give_weight is True:
item_weight = np.array([math.sqrt(len(i)) for i in Train_R_J],
dtype=float)
item_weight = (float(num_item) / item_weight.sum()) * item_weight
user_weight = np.array([math.sqrt(len(u)) for u in Train_R_I],
dtype=float)
user_weight = (float(num_user) / user_weight.sum()) * user_weight
else:
item_weight = np.ones(num_item, dtype=float)
user_weight = np.ones(num_user, dtype=float)
pre_val_rmse = 1e10
if dnn_type == 'CNN':
if reg_schema == 'Item' or reg_schema == 'Dual':
dnn_module_x = CNN_module(dimension, vocab_size, dropout_rate,
emb_dim, maxlen_doc[0] * maxlen_sent[0],
num_kernel_per_ws, init_W)
if reg_schema == 'User' or reg_schema == 'Dual':
dnn_module_y = CNN_module(dimension, vocab_size, dropout_rate,
emb_dim, maxlen_doc[1] * maxlen_sent[1],
num_kernel_per_ws, init_W)
if dnn_type == 'CNN_GRU':
if reg_schema == 'Item' or reg_schema == 'Dual':
dnn_module_x = CNN_GRU_module(dimension, vocab_size, dropout_rate,
emb_dim, gru_outdim, maxlen_doc[0],
maxlen_sent[0], num_kernel_per_ws,
init_W)
if reg_schema == 'User' or reg_schema == 'Dual':
dnn_module_y = CNN_GRU_module(dimension, vocab_size, dropout_rate,
emb_dim, gru_outdim, maxlen_doc[1],
maxlen_sent[1], num_kernel_per_ws,
init_W)
if reg_schema == 'Item' or reg_schema == 'Dual':
theta = dnn_module_x.get_projection_layer(DNN_X)
if reg_schema == 'User' or reg_schema == 'Dual':
phi = dnn_module_y.get_projection_layer(DNN_Y)
np.random.seed(133)
if reg_schema == 'User' or reg_schema == 'Dual':
U = phi
else:
U = np.random.uniform(size=(num_user, dimension))
if reg_schema == 'Item' or reg_schema == 'Dual':
V = theta
else:
V = np.random.uniform(size=(num_item, dimension))
count = 0
for iteration in range(max_iter):
loss = 0
tic = time.time()
print("%d iteration\t(patience: %d)" % (iteration, count))
f1.write("%d iteration\t(patience: %d)\n" % (iteration, count))
VV = b * (V.T.dot(V))
sub_loss = np.zeros(num_user)
# update U
for i in range(num_user):
idx_item = train_user[0][i]
V_i = V[idx_item]
R_i = Train_R_I[i]
tmp_A = VV + (a - b) * (V_i.T.dot(V_i))
A = tmp_A + lambda_u * user_weight[i] * np.eye(dimension)
B = (a * V_i * (np.tile(R_i, (dimension, 1)).T)).sum(0)
if reg_schema == 'User' or reg_schema == 'Dual':
B = B + lambda_u * user_weight[i] * phi[i]
U[i] = np.linalg.solve(A, B)
# -\frac{\lambda_u}{2}\sum_i u_i^Tu_i
if reg_schema == 'Item':
sub_loss[i] = -0.5 * lambda_u * np.dot(U[i], U[i])
loss += np.sum(sub_loss)
sub_loss_dev = np.zeros(num_item)
sub_loss = np.zeros(num_item)
# update V
UU = b * (U.T.dot(U))
for j in range(num_item):
idx_user = train_item[0][j]
U_j = U[idx_user]
R_j = Train_R_J[j]
tmp_A = UU + (a - b) * (U_j.T.dot(U_j))
A = tmp_A + lambda_v * item_weight[j] * np.eye(dimension)
B = (a * U_j * (np.tile(R_j, (dimension, 1)).T)).sum(0)
if reg_schema == 'Item' or reg_schema == 'Dual':
B = B + lambda_v * item_weight[j] * theta[j]
V[j] = np.linalg.solve(A, B)
# -\sum_i\sum_j\frac{c_{i,j}}{2}(r_{ij}-u_i^T v_j)^2
sub_loss_dev[j] = -0.5 * a * np.square(R_j).sum()
sub_loss_dev[j] += a * np.sum((U_j.dot(V[j])) * R_j)
sub_loss_dev[j] += -0.5 * np.dot(V[j].dot(tmp_A), V[j])
# -\frac{\lambda_v}{2}\sum_jv_j^Tv_j
if reg_schema == 'User':
sub_loss[j] = -0.5 * lambda_v * np.dot(V[j], V[j])
loss += np.sum(sub_loss_dev)
loss += np.sum(sub_loss)
seed = np.random.randint(100000)
if reg_schema == 'Item' or reg_schema == 'Dual':
history_x = dnn_module_x.train(DNN_X, V, item_weight, seed)
theta = dnn_module_x.get_projection_layer(DNN_X)
# -\frac{\lambda_v}{2}\sum_j(v_j-\theta_j)^T(v_j-\theta_j)
cnn_loss_x = history_x.history['loss'][-1]
loss += -0.5 * lambda_v * cnn_loss_x * num_item
if reg_schema == 'User' or reg_schema == 'Dual':
history_y = dnn_module_y.train(DNN_Y, U, user_weight, seed)
phi = dnn_module_y.get_projection_layer(DNN_Y)
# -\frac{\lambda_u}{2}\sum_i (u_i-\phi_i)^T(u_i-\phi_i)
cnn_loss_y = history_y.history['loss'][-1]
loss += -0.5 * lambda_u * cnn_loss_y * num_user
tr_mae, tr_mse, tr_rmse = eval_RATING(Train_R_I, U, V, train_user[0])
val_mae, val_mse, val_rmse = eval_RATING(Valid_R, U, V, valid_user[0])
te_mae, te_mse, te_rmse = eval_RATING(Test_R, U, V, test_user[0])
toc = time.time()
elapsed = toc - tic
if iteration == 0:
converge = -1
else:
converge = abs((loss - PREV_LOSS) / PREV_LOSS)
# if (val_rmse < pre_val_rmse):
# if dnn_type == 'CNN':
# if os.path.exists(res_dir + '/drmf_cnn') is not True:
# os.mkdir(res_dir + '/drmf_cnn')
# if os.path.exists(res_dir + '/drmf_cnn/dual') is not True:
# os.mkdir(res_dir + '/drmf_cnn/dual')
# if os.path.exists(res_dir + '/drmf_cnn/user') is not True:
# os.mkdir(res_dir + '/drmf_cnn/user')
# if os.path.exists(res_dir + '/drmf_cnn/item') is not True:
# os.mkdir(res_dir + '/drmf_cnn/item')
# if reg_schema == 'Dual':
# np.savetxt(res_dir + '/drmf_cnn/dual/U.dat', U)
# np.savetxt(res_dir + '/drmf_cnn/dual/V.dat', V)
# np.savetxt(res_dir + '/drmf_cnn/dual/theta.dat', theta)
# np.savetxt(res_dir + '/drmf_cnn/dual/phi.dat', phi)
# dnn_module_x.save_model(res_dir + '/drmf_cnn/dual/x_weights.hdf5')
# dnn_module_y.save_model(res_dir + '/drmf_cnn/dual/y_weights.hdf5')
# if reg_schema == 'User':
# np.savetxt(res_dir + '/drmf_cnn/user/U.dat', U)
# np.savetxt(res_dir + '/drmf_cnn/user/V.dat', V)
# np.savetxt(res_dir + '/drmf_cnn/user/phi.dat', phi)
# dnn_module_y.save_model(res_dir + '/drmf_cnn/user/y_weights.hdf5')
# if reg_schema == 'Item':
# np.savetxt(res_dir + '/drmf_cnn/item/U.dat', U)
# np.savetxt(res_dir + '/drmf_cnn/item/V.dat', V)
# np.savetxt(res_dir + '/drmf_cnn/item/theta.dat', theta)
# dnn_module_x.save_model(res_dir + '/drmf_cnn/item/x_weights.hdf5')
# if dnn_type == 'CNN_GRU':
# if os.path.exists(res_dir + '/drmf_cnn_gru') is not True:
# os.mkdir(res_dir + '/drmf_cnn_gru')
# if os.path.exists(res_dir + '/drmf_cnn_gru/dual') is not True:
# os.mkdir(res_dir + '/drmf_cnn_gru/dual')
# if os.path.exists(res_dir + '/drmf_cnn_gru/user') is not True:
# os.mkdir(res_dir + '/drmf_cnn_gru/user')
# if os.path.exists(res_dir + '/drmf_cnn_gru/item') is not True:
# os.mkdir(res_dir + '/drmf_cnn_gru/item')
# if reg_schema == 'Dual':
# np.savetxt(res_dir + '/drmf_cnn_gru/dual/U.dat', U)
# np.savetxt(res_dir + '/drmf_cnn_gru/dual/V.dat', V)
# np.savetxt(res_dir + '/drmf_cnn_gru/dual/theta.dat', theta)
# np.savetxt(res_dir + '/drmf_cnn_gru/dual/phi.dat', phi)
# dnn_module_x.save_model(res_dir + '/drmf_cnn_gru/dual/x_weights.hdf5')
# dnn_module_y.save_model(res_dir + '/drmf_cnn_gru/dual/y_weights.hdf5')
# if reg_schema == 'User':
# np.savetxt(res_dir + '/drmf_cnn_gru/user/U.dat', U)
# np.savetxt(res_dir + '/drmf_cnn_gru/user/V.dat', V)
# np.savetxt(res_dir + '/drmf_cnn_gru/user/phi.dat', phi)
# dnn_module_y.save_model(res_dir + '/drmf_cnn_gru/user/y_weights.hdf5')
# if reg_schema == 'Item':
# np.savetxt(res_dir + '/drmf_cnn_gru/item/U.dat', U)
# np.savetxt(res_dir + '/drmf_cnn_gru/item/V.dat', V)
# np.savetxt(res_dir + '/drmf_cnn_gru/item/theta.dat', theta)
# dnn_module_x.save_model(res_dir + '/drmf_cnn_gru/item/x_weights.hdf5')
# else:
# count = count + 1
# for fast running, without saving models
if (val_rmse >= pre_val_rmse):
count = count + 1
pre_val_rmse = val_rmse
print(
"Loss: %.5f Elpased: %.4fs Converge: %.6f Tr: [%.5f, %.5f, %.5f] Val: [%.5f, %.5f, %.5f] Te: [%.5f, %.5f, %.5f] "
% (loss, elapsed, converge, tr_mae, tr_mse, tr_rmse, val_mae,
val_mse, val_rmse, te_mae, te_mse, te_rmse))
f1.write(
"Loss: %.5f Elpased: %.4fs Converge: %.6f Tr: [%.5f, %.5f, %.5f] Val: [%.5f, %.5f, %.5f] Te: [%.5f, %.5f, %.5f]\n"
% (loss, elapsed, converge, tr_mae, tr_mse, tr_rmse, val_mae,
val_mse, val_rmse, te_mae, te_mse, te_rmse))
if (count == ENDURE_COUNT):
break
PREV_LOSS = loss
f1.close()
def ConvMF(res_dir,
train_user,
train_item,
valid_user,
test_user,
R,
CNN_X,
vocab_size,
init_W=None,
give_item_weight=True,
max_iter=50,
lambda_u=1,
lambda_v=100,
dimension=50,
dropout_rate=0.2,
emb_dim=200,
max_len=300,
num_kernel_per_ws=100):
'''
构造并训练卷积矩阵分解模型
:param res_dir:结果文件路径
:param train_user:训练集用户稀疏评分向量（libSVM format）
:param train_item:训练集物品稀疏评分向量（libSVM format）
:param valid_user:测试集用户稀疏评分向量（libSVM format）
:param test_user:测试集物品稀疏评分向量（libSVM format）
:param R:原始评分数据,format: user id::item id::rating
:param CNN_X:物品描述词序列
:param vocab_size:词表大小
:param init_W:如果为None则动态训练词向量权重
:param give_item_weight:如果为True则使用静态词向量，否则动态训练词向量
:param max_iter:最大迭代次数
:param lambda_u:用户端正则惩罚项系数
:param lambda_v:用户端正则惩罚项系数
:param dimension:隐变量维度
:param dropout_rate:丢弃率
:param emb_dim:词嵌入维度
:param max_len:物品文本描述序列最大长度
:param num_kernel_per_ws:CNN的卷积核个数
:return:None
'''
# explicit setting
a = 1
b = 0
num_user = R.shape[0] #6040
num_item = R.shape[1] #3544
PREV_LOSS = 1e-50
if not os.path.exists(res_dir):
os.makedirs(res_dir)
f1 = open(res_dir + '/state.log', 'w')
Train_R_I = train_user[1] #user rating_list
Train_R_J = train_item[1] #item rating_list
Test_R = test_user[1]
Valid_R = valid_user[1]
if give_item_weight is True:
item_weight = np.array([math.sqrt(len(i)) for i in Train_R_J],
dtype=float)
item_weight = (float(num_item) / item_weight.sum()) * item_weight
else:
item_weight = np.ones(num_item, dtype=float)
pre_val_eval = 1e10
## init CNN model
cnn_module = CNN_module(dimension, vocab_size, dropout_rate, emb_dim,
max_len, num_kernel_per_ws, init_W)
theta = cnn_module.get_projection_layer(CNN_X)
np.random.seed(133)
# user-latent matrix
U = np.random.uniform(size=(num_user, dimension))
# item-latent matrix
V = theta
endure_count = 5 #超出５次则退出迭代训练
count = 0
for iteration in range(max_iter):
loss = 0
tic = time.time()
print("%d iteration\t(patience: %d)" % (iteration, count))
##get user-latent matirx loss
VV = b * (V.T.dot(V)) + lambda_u * np.eye(dimension) #theta V
sub_loss = np.zeros(num_user)
for i in range(num_user):
idx_item = train_user[0][i]
V_i = V[idx_item]
R_i = Train_R_I[i]
A = VV + (a - b) * (V_i.T.dot(V_i))
B = (a * V_i * (np.tile(R_i, (dimension, 1)).T)).sum(0)
U[i] = np.linalg.solve(A, B)
sub_loss[i] = -0.5 * lambda_u * np.dot(U[i], U[i])
loss = loss + np.sum(sub_loss)
##get item-latent matirx loss
sub_loss = np.zeros(num_item)
UU = b * (U.T.dot(U))
for j in range(num_item):
idx_user = train_item[0][j]
U_j = U[idx_user]
R_j = Train_R_J[j]
tmp_A = UU + (a - b) * (U_j.T.dot(U_j))
A = tmp_A + lambda_v * item_weight[j] * np.eye(dimension)
B = (a * U_j * (np.tile(R_j, (dimension, 1)).T)
).sum(0) + lambda_v * item_weight[j] * theta[j]
V[j] = np.linalg.solve(A, B)
sub_loss[j] = -0.5 * np.square(R_j * a).sum()
sub_loss[j] = sub_loss[j] + a * np.sum((U_j.dot(V[j])) * R_j)
sub_loss[j] = sub_loss[j] - 0.5 * np.dot(V[j].dot(tmp_A), V[j])
loss = loss + np.sum(sub_loss)
seed = np.random.randint(100000)
# get cnn loss
history = cnn_module.train(CNN_X, V, item_weight, seed)
theta = cnn_module.get_projection_layer(CNN_X)
cnn_loss = history.history['loss'][-1]
loss = loss - 0.5 * lambda_v * cnn_loss * num_item
# get rmse eval
tr_eval = eval_RMSE(Train_R_I, U, V, train_user[0])
val_eval = eval_RMSE(Valid_R, U, V, valid_user[0])
te_eval = eval_RMSE(Test_R, U, V, test_user[0])
toc = time.time()
elapsed = toc - tic
converge = abs((loss - PREV_LOSS) / PREV_LOSS)
#save u,v,w weight
if (val_eval < pre_val_eval):
cnn_module.save_model(res_dir + '/CNN_weights.hdf5')
np.savetxt(res_dir + '/U.dat', U)
np.savetxt(res_dir + '/V.dat', V)
np.savetxt(res_dir + '/theta.dat', theta)
else:
count = count + 1
pre_val_eval = val_eval
print(
"Loss: %.5f Elpased: %.4fs Converge: %.6f Tr: %.5f Val: %.5f Te: %.5f"
% (loss, elapsed, converge, tr_eval, val_eval, te_eval))
f1.write(
"Loss: %.5f Elpased: %.4fs Converge: %.6f Tr: %.5f Val: %.5f Te: %.5f\n"
% (loss, elapsed, converge, tr_eval, val_eval, te_eval))
if (count == endure_count):
break
PREV_LOSS = loss
f1.close()
def Haec(res_dir,
train_user,
train_item,
valid_user,
test_user,
R,
CNN_X,
vocab_size,
init_W=None,
give_item_weight=True,
max_iter=30,
lambda_u=1,
lambda_v=1,
dimension=50,
dropout_rate=0.2,
emb_dim=200,
max_len=300,
num_kernel_per_ws=100):
# explicit setting
a = 1
b = 0
num_user = R.shape[0]
num_item = R.shape[1]
# print '=====R.all========'
# print num_user,num_item
fileU = '../data/pre/ml_1m/User.npy'
# fileR = "../Data/convmf/preprocessed/movielens_100k/R.npy"
Uinfo = getData1(fileUser=fileU)
PREV_LOSS = 1e-50
PREV_TE = 1e-50
if not os.path.exists(res_dir):
os.makedirs(res_dir)
f1 = open(res_dir + '/state.log', 'w')
# Train_R_I按用户进行汇总的评分列表[[用户1的所有评分][用户2的所有评分]...]，长6040
Train_R_I = train_user[1]
# Train_R_I按产品进行汇总的评分列表[[产品1的所有评分][产品2的所有评分]...]，长3544
Train_R_J = train_item[1]
Test_R = test_user[1]
Valid_R = valid_user[1]
if give_item_weight is True:
#原理：评分越多的产品，权重越大
# item_weight:"每个产品对应的用户评论数的开方"
item_weight = np.array([math.sqrt(len(i)) for i in Train_R_J],
dtype=float)
# item_weight=item_weight*(产品数/所有产品对应的评论的开方之和），类似归一化
# 处理后，每个Item对应的权重限制在(0-5）之间，float。
item_weight = (float(num_item) / item_weight.sum()) * item_weight
else:
item_weight = np.ones(num_item, dtype=float)
pre_val_eval = 1e10
cnn_module = CNN_module(dimension, vocab_size, dropout_rate, emb_dim,
max_len, num_kernel_per_ws, init_W)
# theta,cnn输出层，litst,长度3544，每个元素为长50的list，数字，取值（-0.06，0.039）。
theta = cnn_module.get_projection_layer(CNN_X)
V = theta
#随机初始化前，设置种子seed，便于复现
np.random.seed(133)
# # # U：6040list(6040,50)，V：长为3544的list，每个list为含50个元素的子list。
U = np.random.uniform(size=(num_user, dimension))
# g1 = tf.Graph()
# with tf.Session(graph=tf.get_default_graph()) as sess:
config = tf.ConfigProto()
config.gpu_options.allocator_type = 'BFC'
sess = tf.Session(config=config)
model = USDAE1(sess, R.shape, Uinfo.shape, is_training=True, **mlp_args)
print("build model...")
model.build()
u_loss, ulatent = model.train(R.toarray(), Uinfo, U,
mlp_args["learning_rate"])
U = ulatent
endure_count = 5
count = 0
for iteration in xrange(max_iter):
loss = 0
mcount = 0
tic = time.time()
print "%d iteration\t(patience: %d)" % (iteration, count)
#公式7的一部分，b被设置为0，所以VV为：值为lambda_u*I_k
VV = b * (V.T.dot(V)) + lambda_u * np.eye(dimension)
# 令偏导为0，得到所有的U[i]，并求U部分的loss
#初始化U部分的loss，共计num_user（6040）个
sub_loss = np.zeros(num_user)
for i in xrange(num_user):
# idx_item：用户i评论过的电影ID列表
idx_item = train_user[0][i]
#V_i：从V中筛选出用户i所评论的item。长度（主List）：用户i对应的评论数，子list:50.
V_i = V[idx_item]
#R_i用户i的所有评分,size(1,len(idx_item))
R_i = Train_R_I[i]
#A,size(1,50),公式7求逆的部分，每个子元素size（1,50）
A = VV + (a - b) * (V_i.T.dot(V_i))
# B,size(1,50),公式7后面的部分，每个子元素为一个数。
# np.tile（A,rep）:重复rep次A来构建array；
B = (a * V_i *
(np.tile(R_i,
(dimension, 1)).T)).sum(0) + lambda_u * ulatent[i].T
# linalg线性代数模块，solve（A，B），求解Ax=B 线性方程组
U[i] = np.linalg.solve(A, B)
# U部分的Loss
sub_loss[i] = -0.5 * lambda_u * (np.sum(
np.square(U[i] - ulatent[i])))
loss = loss + np.sum(sub_loss)
# 令偏导为0，得到所有的V[j]，并求V部分的loss
# 初始化V部分的loss，共计num_item（3544）个
sub_loss = np.zeros(num_item)
# b为0
UU = b * (U.T.dot(U))
for j in xrange(num_item):
# idx_user：产品j被评论的用户ID列表
idx_user = train_item[0][j]
# U_j：从U中筛选出在idx_user出现过的。长度（主List）：产品i对应的评论数，子list:50.
U_j = U[idx_user]
R_j = Train_R_J[j]
# tmp_A =(U_j.T.dot(U_j))
tmp_A = UU + (a - b) * (U_j.T.dot(U_j))
# A=公式8求逆的部分，多加入了item_weight[j]
A = tmp_A + lambda_v * item_weight[j] * np.eye(dimension)
# B=公式8的后半部分，注意cnn部分loss=lambda_v * item_weight[j] * theta[j]
B = (a * U_j * (np.tile(R_j, (dimension, 1)).T)
).sum(0) + lambda_v * item_weight[j] * theta[j]
# print '================num_item=============', num_item
# print len(theta[j])
# print 'A.shape：', A.shape
# print 'B.shape：', B.shape
V[j] = np.linalg.solve(A, B)
#下面三个式子：公式6的第一项，（R-UV）^2
sub_loss[j] = -0.5 * np.square(R_j * a).sum()
sub_loss[j] = sub_loss[j] + a * np.sum((U_j.dot(V[j])) * R_j)
sub_loss[j] = sub_loss[j] - 0.5 * np.dot(V[j].dot(tmp_A), V[j])
loss = loss + np.sum(sub_loss)
seed = np.random.randint(100000)
history = cnn_module.train(CNN_X, V, item_weight, seed)
theta = cnn_module.get_projection_layer(CNN_X)
cnn_loss = history.history['loss'][-1]
u_loss, ulatent = model.train(R.toarray(), Uinfo, U,
mlp_args["learning_rate"])
loss = loss - 0.5 * lambda_v * cnn_loss * num_item - u_loss
tr_eval = eval_MAE(Train_R_I, U, V, train_user[0])
val_eval = eval_MAE(Valid_R, U, V, valid_user[0])
te_eval = eval_MAE(Test_R, U, V, test_user[0])
toc = time.time()
elapsed = toc - tic
converge = abs((loss - PREV_LOSS) / PREV_LOSS)
if te_eval > PREV_TE:
mcount += 1
if mcount > 2:
break
if (val_eval < pre_val_eval):
# cnn_module.save_model(res_dir + '/CNN_weights.hdf5')
np.savetxt(res_dir + '/U.dat', U)
np.savetxt(res_dir + '/V.dat', V)
np.savetxt(res_dir + '/theta.dat', theta)
else:
count = count + 1
pre_val_eval = val_eval
print "Loss: %.5f Elpased: %.4fs Converge: %.6f Tr: %.5f Val: %.5f Te: %.5f" % (
loss, elapsed, converge, tr_eval, val_eval, te_eval)
f1.write(
"Iteration：%d Loss: %.5f Elpased: %.4fs Converge: %.6f Tr: %.5f Val: %.5f Te: %.5f\n"
% (iteration, loss, elapsed, converge, tr_eval, val_eval, te_eval))
# if (count == endure_count):
# break
PREV_LOSS = loss
PREV_TE = te_eval
f1.close()
def ConvMF(res_dir, train_user, train_item, valid_user, test_user,
R, CNN_X, vocab_size, init_W=None, give_item_weight=True,
max_iter=50, lambda_u=1, lambda_v=100, dimension=50,
dropout_rate=0.2, emb_dim=200, max_len=300, num_kernel_per_ws=100):
# explicit setting
a = 1
b = 0
num_user = R.shape[0]
num_item = R.shape[1]
PREV_LOSS = 1e-50
if not os.path.exists(res_dir):
os.makedirs(res_dir)
f1 = open(res_dir + '/state.log', 'w')
Train_R_I = train_user[1]
Train_R_J = train_item[1]
Test_R = test_user[1]
Valid_R = valid_user[1]
if give_item_weight is True:
item_weight = np.array([math.sqrt(len(i))
for i in Train_R_J], dtype=float)
item_weight = (float(num_item) / item_weight.sum()) * item_weight
else:
item_weight = np.ones(num_item, dtype=float)
pre_val_eval = 1e10
cnn_module = CNN_module(dimension, vocab_size, dropout_rate,
emb_dim, max_len, num_kernel_per_ws, init_W)
theta = cnn_module.get_projection_layer(CNN_X)
np.random.seed(133)
U = np.random.uniform(size=(num_user, dimension))
V = theta
endure_count = 5
count = 0
for iteration in xrange(max_iter):
loss = 0
tic = time.time()
print "%d iteration\t(patience: %d)" % (iteration, count)
VV = b * (V.T.dot(V)) + lambda_u * np.eye(dimension)
sub_loss = np.zeros(num_user)
for i in xrange(num_user):
idx_item = train_user[0][i]
V_i = V[idx_item]
R_i = Train_R_I[i]
A = VV + (a - b) * (V_i.T.dot(V_i))
B = (a * V_i * (np.tile(R_i, (dimension, 1)).T)).sum(0)
U[i] = np.linalg.solve(A, B)
sub_loss[i] = -0.5 * lambda_u * np.dot(U[i], U[i])
loss = loss + np.sum(sub_loss)
sub_loss = np.zeros(num_item)
UU = b * (U.T.dot(U))
for j in xrange(num_item):
idx_user = train_item[0][j]
U_j = U[idx_user]
R_j = Train_R_J[j]
tmp_A = UU + (a - b) * (U_j.T.dot(U_j))
A = tmp_A + lambda_v * item_weight[j] * np.eye(dimension)
B = (a * U_j * (np.tile(R_j, (dimension, 1)).T)
).sum(0) + lambda_v * item_weight[j] * theta[j]
V[j] = np.linalg.solve(A, B)
sub_loss[j] = -0.5 * np.square(R_j * a).sum()
sub_loss[j] = sub_loss[j] + a * np.sum((U_j.dot(V[j])) * R_j)
sub_loss[j] = sub_loss[j] - 0.5 * np.dot(V[j].dot(tmp_A), V[j])
loss = loss + np.sum(sub_loss)
seed = np.random.randint(100000)
history = cnn_module.train(CNN_X, V, item_weight, seed)
theta = cnn_module.get_projection_layer(CNN_X)
cnn_loss = history.history['loss'][-1]
loss = loss - 0.5 * lambda_v * cnn_loss * num_item
tr_eval = eval_RMSE(Train_R_I, U, V, train_user[0])
val_eval = eval_RMSE(Valid_R, U, V, valid_user[0])
te_eval = eval_RMSE(Test_R, U, V, test_user[0])
toc = time.time()
elapsed = toc - tic
converge = abs((loss - PREV_LOSS) / PREV_LOSS)
if (val_eval < pre_val_eval):
cnn_module.save_model(res_dir + '/CNN_weights.hdf5')
np.savetxt(res_dir + '/U.dat', U)
np.savetxt(res_dir + '/V.dat', V)
np.savetxt(res_dir + '/theta.dat', theta)
else:
count = count + 1
pre_val_eval = val_eval
print "Loss: %.5f Elpased: %.4fs Converge: %.6f Tr: %.5f Val: %.5f Te: %.5f" % (
loss, elapsed, converge, tr_eval, val_eval, te_eval)
f1.write("Loss: %.5f Elpased: %.4fs Converge: %.6f Tr: %.5f Val: %.5f Te: %.5f\n" % (
loss, elapsed, converge, tr_eval, val_eval, te_eval))
if (count == endure_count):
break
PREV_LOSS = loss
f1.close()
def main():
#cnn_cae_transfer
# exp_dir = '/home/wanliz/data/Extended_ctr/convmf/citeulike_a_extended/results/5-8_in-matrix_no-val_0.01-100_w_cnn-100_cae-50_transfer-2_no-sc//fold-4/'
exp_dir = '/home/wanliz/data/Extended_ctr/convmf/citeulike_a_extended/results/6-9_in-matrix-200_no-val_0.01-100-w-cnn-100_cae-25_transfer-noSC/fold-4/'
exp_dir = '/home/wanliz/data/Extended_ctr/convmf/citeulike_a_extended/results/7-9_in-matrix-200_no-val_0.01-100-w-cnn-50_cae-150_transfer-noSC/fold-1'
#cnn_cae_concat
# exp_dir ='/home/wanliz/data/Extended_ctr/convmf/citeulike_a_extended/results/6-8_out-of-matrix_no-val_1-1000-w-cnn-100_cae-50_concat/fold-1/'
#cnn
# exp_dir = '/home/wanliz/data/Extended_ctr/convmf/citeulike_a_extended/results/6-8_in-matrix_no-val_0.01-100_w_cnn/fold-1'
fc_weights_file = os.path.join(exp_dir, 'FC_weights.npy')
module = 'cnn_cae_transfer'
''' Network parameters'''
nb_filters = 50
cae_N_hidden = 150
nb_features = 23
if module == 'cnn_cae_transfer' or module == 'cnn_cae_concat':
weights_file = os.path.join(exp_dir, 'CNN_CAE_weights.hdf5')
elif module == 'cnn':
weights_file = os.path.join(exp_dir, 'CNN_weights.hdf5')
if not os.path.exists(fc_weights_file):
if module == 'cnn_cae_transfer':
model = CNN_CAE_transfer_module(output_dimesion=200,
vocab_size=8001,
dropout_rate=0.2,
emb_dim=200,
max_len=300,
nb_filters=nb_filters,
init_W=None,
cae_N_hidden=cae_N_hidden,
nb_features=nb_features)
elif module == 'cnn_cae_concat':
model = CNN_CAE_module(output_dimesion=200,
vocab_size=8001,
dropout_rate=0.2,
emb_dim=200,
max_len=300,
nb_filters=nb_filters,
init_W=None,
cae_N_hidden=cae_N_hidden,
nb_features=17)
elif module == 'cnn':
model = CNN_module(output_dimesion=200,
vocab_size=8001,
dropout_rate=0.2,
emb_dim=200,
max_len=300,
nb_filters=nb_filters,
init_W=None)
model.load_model(weights_file)
if module == 'cnn_cae_concat':
layer_name = 'joint_output'
else:
layer_name = 'fully_connect'
weights = model.model.get_layer(layer_name).get_weights()
bias = weights[1]
fc_weights = weights[0]
np.save(fc_weights_file, fc_weights)
else:
fc_weights = np.load(fc_weights_file)
# Normalized Data
normalized = (fc_weights - np.min(fc_weights)) / (np.max(fc_weights) -
np.min(fc_weights))
cnn_output = {}
# nb_filters = cae_N_hidden
if module == 'cnn_cae_concat':
cnn_vector_length = 2
else:
cnn_vector_length = 3 #normalized.shape[0] / nb_filters
for i in range(cnn_vector_length):
print(i * nb_filters, i * nb_filters + (nb_filters - 1))
cnn_output[i] = np.sum(normalized[i * nb_filters:i * nb_filters +
nb_filters - 1],
axis=0,
keepdims=True)
agg_output = np.vstack(cnn_output.values())
if fc_weights.shape[0] > cnn_vector_length * nb_filters:
#CAE output
cae_output = np.sum(normalized[cnn_vector_length * nb_filters:],
axis=0,
keepdims=True)
agg_output = np.vstack((agg_output, cae_output))
plot_weights(agg_output, exp_dir)
plot_model(model.model,
to_file='/home/wanliz/model.png',
show_layer_names=True,
show_shapes=True)
print('')
def ConvMF(res_dir, state_log_dir, train_user, train_item, valid_user, test_user,
R, CNN_X, vocab_size, init_W=None, give_item_weight=False,
max_iter=50, lambda_u=1, lambda_v=100, dimension=50,
dropout_rate=0.2, emb_dim=200, max_len=300, num_kernel_per_ws=100):
# explicit settinggit
a = 1
b = 0.01
alpha = 40
num_user = R.shape[0]
num_item = R.shape[1]
PREV_LOSS = -1e-50
if not os.path.exists(res_dir):
os.makedirs(res_dir)
os.chdir(res_dir)
# f1 = open(res_dir + '/state.log', 'w')
if not os.path.exists(state_log_dir):
os.makedirs(state_log_dir)
f1 = open(state_log_dir + '/state.log', 'w')
# log metrics into tf.summary
log_dir_name = os.path.basename(os.path.dirname(state_log_dir + '/'))
log_dir = os.path.join(state_log_dir, log_dir_name)
logger_tb = Tb_Logger(log_dir)
# indicate folder to save, plus other options
tensorboard = TensorBoard(log_dir=log_dir, histogram_freq=0,
write_graph=False, write_images=False)
# save it in your callback list, where you can include other callbacks
callbacks_list = [tensorboard]
# then pass to fit as callback, remember to use validation_data also
Train_R_I = train_user[1]
Train_R_J = train_item[1]
Test_R = test_user[1]
no_validation = False
if valid_user:
Valid_R = valid_user[1]
else:
no_validation = True
if give_item_weight is True:
item_weight = np.array([math.sqrt(len(i))
for i in Train_R_J], dtype=float)
item_weight = (float(num_item) / item_weight.sum()) * item_weight
item_weight[item_weight == 0] = 1
else:
item_weight = np.ones(num_item, dtype=float)
pre_val_eval = 1e10
cnn_module = CNN_module(dimension, vocab_size, dropout_rate,
emb_dim, max_len, num_kernel_per_ws, init_W)
theta = cnn_module.get_projection_layer(CNN_X)
np.random.seed(133)
U = np.random.uniform(size=(num_user, dimension))
V = theta
print ('Training CNN-MF ...')
endure_count = 5
count = 0
converge_threshold = 1e-4
converge = 1.0
iteration = 0
while (iteration < max_iter and converge > converge_threshold) or iteration < min_iter:
# for iteration in xrange(max_iter):
loss = 0
tic = time.time()
print "%d iteration\t(patience: %d)" % (iteration, count)
# VV = b * (V.T.dot(V)) + lambda_u * np.eye(dimension)
VV = (V.T.dot(V)) + lambda_u * np.eye(dimension)
sub_loss = np.zeros(num_user)
for i in xrange(num_user):
idx_item = train_user[0][i]
V_i = V[idx_item]
R_i = Train_R_I[i]
# A = VV + (a - b) * (V_i.T.dot(V_i))
# B = (a * V_i * (np.tile(R_i, (dimension, 1)).T)).sum(0)
C_i = np.diag(alpha * R_i)
A = VV + V_i.T.dot(C_i).dot(V_i)
B = V_i.T.dot(C_i + np.eye(len(idx_item))).dot(R_i)
U[i] = np.linalg.solve(A, B)
sub_loss[i] = -0.5 * lambda_u * np.dot(U[i], U[i])
loss = loss + np.sum(sub_loss)
sub_loss = np.zeros(num_item)
# UU = b * (U.T.dot(U))
UU = (U.T.dot(U))
for j in xrange(num_item):
idx_user = train_item[0][j]
U_j = U[idx_user]
R_j = Train_R_J[j]
C_j = np.diag(alpha * R_j)
if len(U_j) > 0:
# tmp_A = UU + (a - b) * (U_j.T.dot(U_j))
tmp_A = UU + (U_j.T.dot(C_j).dot(U_j))
A = tmp_A + lambda_v * item_weight[j] * np.eye(dimension)
B = U_j.T.dot(C_j + np.eye(len(idx_user))).dot(R_j) + lambda_v * item_weight[j] * theta[j]
# B = (a * U_j * (np.tile(R_j, (dimension, 1)).T)
# ).sum(0) + lambda_v * item_weight[j] * theta[j]
V[j] = np.linalg.solve(A, B)
# sub_loss[j] = -0.5 * np.square(R_j * a).sum()
# sub_loss[j] = sub_loss[j] + a * np.sum((U_j.dot(V[j])) * R_j)
sub_loss[j] = -0.5 * np.square(R_j * C_j).sum()
sub_loss[j] = sub_loss[j] + np.sum(C_j * (U_j.dot(V[j])) * R_j)
sub_loss[j] = sub_loss[j] - 0.5 * np.dot(V[j].dot(tmp_A), V[j])
else:
V[j] = theta[j]
loss = loss + np.sum(sub_loss)
seed = np.random.randint(100000)
history = cnn_module.train(CNN_X, V, item_weight, seed, callbacks_list)
theta = cnn_module.get_projection_layer(CNN_X)
cnn_loss = history.history['loss'][-1]
loss = loss - 0.5 * lambda_v * cnn_loss * num_item
tr_eval = eval_RMSE(Train_R_I, U, V, train_user[0])
if not no_validation:
val_eval = eval_RMSE(Valid_R, U, V, valid_user[0])
else:
val_eval = -1
te_eval = eval_RMSE(Test_R, U, V, test_user[0])
logger_tb.log_scalar('train_rmse', tr_eval, iteration)
if not no_validation:
logger_tb.log_scalar('eval_rmse', val_eval, iteration)
logger_tb.log_scalar('test_rmse', te_eval, iteration)
logger_tb.writer.flush()
toc = time.time()
elapsed = toc - tic
converge = abs((loss - PREV_LOSS) / PREV_LOSS)
# if (val_eval < pre_val_eval):
if (loss > PREV_LOSS):
# count = 0
print ("likelihood is increasing!")
cnn_module.save_model(res_dir + '/CNN_weights.hdf5')
np.savetxt(res_dir + '/final-U.dat', U)
np.savetxt(res_dir + '/final-V.dat', V)
np.savetxt(res_dir + '/theta.dat', theta)
best_train_rmse = tr_eval
best_test_rmse = te_eval
best_val_rmse = val_eval
else:
count = count + 1
# if (val_eval < pre_val_eval):
# count = 0
# cnn_module.save_model(res_dir + '/CNN_weights.hdf5')
# np.savetxt(res_dir + '/final-U.dat', U)
# np.savetxt(res_dir + '/final-V.dat', V)
# np.savetxt(res_dir + '/theta.dat', theta)
# else:
# count = count + 1
pre_val_eval = val_eval
print "Loss: %.5f Elpased: %.4fs Converge: %.6f Tr: %.5f Val: %.5f Te: %.5f" % (
loss, elapsed, converge, tr_eval, val_eval, te_eval)
f1.write("Loss: %.5f Elpased: %.4fs Converge: %.6f Tr: %.5f Val: %.5f Te: %.5f\n" % (
loss, elapsed, converge, tr_eval, val_eval, te_eval))
if (count >= endure_count and iteration > min_iter):
# if (count == endure_count):
break
elif (iteration < min_iter):
count = 0
PREV_LOSS = loss
iteration += 1
f1.close()
return best_train_rmse, best_test_rmse, best_val_rmse
def Raw_att_CNN_concat(res_dir, state_log_dir, train_user, train_item, valid_user, test_user,
R, attributes_X, CNN_X, vocab_size, init_W, max_iter, lambda_u, lambda_v,
dimension, use_CAE,
dropout_rate=0.2, emb_dim=200, max_len=300, num_kernel_per_ws=100,
a=1, b=0.01, give_item_weight=False):
# explicit setting
# a = 1
# b = 0.01
alpha = 40
# confidence_matrix = get_confidence_matrix(R,'user-dependant',alpha=40)
num_user = R.shape[0]
num_item = R.shape[1]
num_features = attributes_X.shape[1]
'''prepare path to store results and log'''
if not os.path.exists(res_dir):
os.makedirs(res_dir)
os.chdir(res_dir)
if not os.path.exists(state_log_dir):
os.makedirs(state_log_dir)
f1 = open(state_log_dir + '/state.log', 'w')
'''log metrics using tf.summary '''
log_dir_name = os.path.basename(os.path.dirname(state_log_dir + '/'))
log_dir = os.path.join(state_log_dir, log_dir_name)
logger_tb = Tb_Logger(log_dir)
# indicate folder to save, plus other options
tensorboard = TensorBoard(log_dir=log_dir, histogram_freq=0,
write_graph=False, write_images=False)
# save it in your callback list, where you can include other callbacks
callbacks_list = [tensorboard]
# then pass to fit as callback, remember to use validation_data also
Train_R_I = train_user[1]
Train_R_J = train_item[1]
Test_R = test_user[1]
# check if the dataset has validation set
no_validation = False
if valid_user:
Valid_R = valid_user[1]
else:
no_validation = True
# assign weights to each item according to the number of time the item was rated
if give_item_weight is True:
item_weight = np.array([math.sqrt(len(i))
for i in Train_R_J], dtype=float)
item_weight = (float(num_item) / item_weight.sum()) * item_weight
item_weight[item_weight == 0] = 1
else:
item_weight = np.ones(num_item, dtype=float)
'''initialize'''
cnn_output_dim = 150
att_output_dim = dimension - cnn_output_dim
cnn_module = CNN_module(cnn_output_dim, vocab_size, dropout_rate,
emb_dim, max_len, num_kernel_per_ws, init_W)
if use_CAE:
att_module = CAE_module(att_output_dim, cae_N_hidden=att_output_dim, nb_features=num_features)
else:
att_module = Stacking_NN_CNN_CAE(input_dim=num_features,output_dimesion=att_output_dim,
num_layers=1, hidden_dim=num_features * 2)
theta = cnn_module.get_projection_layer(CNN_X)
gamma = att_module.get_projection_layer(attributes_X)
delta = np.concatenate((gamma, theta), axis=1)
if not (theta.shape[1] + gamma.shape[1] == dimension):
sys.exit("theta and gamma shapes are wrong")
np.random.seed(133)
U = np.random.uniform(size=(num_user, dimension))
V = delta
print ('Training CNN-CAE-MF ...')
pre_val_eval = -1e10
PREV_LOSS = -1e-50
endure_count = 5
count = 0
converge_threshold = 1e-4
converge = 1.0
iteration = 0
while (iteration < max_iter and converge > converge_threshold) or iteration < min_iter:
# for iteration in xrange(max_iter):
loss = 0
tic = time.time()
print "%d iteration\t(patience: %d)" % (iteration, count)
# Update U
# VV = b * (V.T.dot(V)) + lambda_u * np.eye(dimension)
VV = (V.T.dot(V)) + lambda_u * np.eye(dimension)
sub_loss = np.zeros(num_user)
for i in xrange(num_user):
idx_item = train_user[0][i]
V_i = V[idx_item]
R_i = Train_R_I[i]
C_i = np.diag(alpha * R_i)
# A = VV + (a - b) * (V_i.T.dot(V_i))
# B = (a * V_i * (np.tile(R_i, (dimension, 1)).T)).sum(0)
A = VV + V_i.T.dot(C_i).dot(V_i)
B = V_i.T.dot(C_i + np.eye(len(idx_item))).dot(R_i)
U[i] = np.linalg.solve(A, B)
sub_loss[i] = -0.5 * lambda_u * np.dot(U[i], U[i])
loss = loss + np.sum(sub_loss)
# Update V
sub_loss = np.zeros(num_item)
# UU = b * (U.T.dot(U))
UU = (U.T.dot(U))
for j in xrange(num_item):
idx_user = train_item[0][j]
U_j = U[idx_user]
R_j = Train_R_J[j]
C_j = np.diag(alpha * R_j)
if len(U_j) > 0:
# tmp_A = UU + (a - b) * (U_j.T.dot(U_j))
tmp_A = UU + (U_j.T.dot(C_j).dot(U_j))
A = tmp_A + lambda_v * item_weight[j] * np.eye(dimension)
# B = (a * U_j * (np.tile(R_j, (dimension, 1)).T)
# ).sum(0) + lambda_v * item_weight[j] * delta[j]
B = U_j.T.dot(C_j + np.eye(len(idx_user))).dot(R_j) + lambda_v * item_weight[j] * delta[j]
V[j] = np.linalg.solve(A, B)
sub_loss[j] = -0.5 * np.square(R_j * C_j).sum()
sub_loss[j] = sub_loss[j] + np.sum(C_j * ((U_j.dot(V[j])) * R_j))
sub_loss[j] = sub_loss[j] - 0.5 * np.dot(V[j].dot(tmp_A), V[j])
else:
# in case the item has no ratings
V[j] = delta[j]
loss = loss + np.sum(sub_loss)
# Update theta
seed = np.random.randint(100000)
history = cnn_module.train(CNN_X, V[:, att_output_dim:], item_weight=item_weight,
seed=seed, callbacks_list=callbacks_list)
theta = cnn_module.get_projection_layer(CNN_X)
cnn_loss = history.history['loss'][-1]
# update gamma
history = att_module.train(attributes_X, V[:, :att_output_dim], item_weight, seed, callbacks_list)
gamma = att_module.get_projection_layer(attributes_X)
att_loss = history.history['loss'][-1]
# update delta
delta = np.concatenate((gamma, theta), axis=1)
loss = loss - 0.5 * lambda_v * (cnn_loss + att_loss) * num_item
toc = time.time()
elapsed = toc - tic
'''calculate RMSE'''
tr_eval = eval_RMSE(Train_R_I, U, V, train_user[0])
if not no_validation:
val_eval = eval_RMSE(Valid_R, U, V, valid_user[0])
else:
val_eval = -1
te_eval = eval_RMSE(Test_R, U, V, test_user[0])
''' write tf.summary'''
logger_tb.log_scalar('train_rmse', tr_eval, iteration)
if not no_validation:
logger_tb.log_scalar('eval_rmse', val_eval, iteration)
logger_tb.log_scalar('test_rmse', te_eval, iteration)
logger_tb.writer.flush()
'''Calculate converge and stor best values of U,V,theta'''
converge = abs((loss - PREV_LOSS) / PREV_LOSS)
# if (val_eval < pre_val_eval):
if (loss > PREV_LOSS):
# count = 0
print ("likelihood is increasing!")
cnn_module.save_model(res_dir + '/CNN_weights.hdf5')
cnn_module.save_model(res_dir + '/Att_weights.hdf5')
np.savetxt(res_dir + '/final-U.dat', U)
np.savetxt(res_dir + '/final-V.dat', V)
np.savetxt(res_dir + '/theta.dat', theta)
np.savetxt(res_dir + '/gamma.dat', gamma)
best_train_rmse = tr_eval
best_test_rmse = te_eval
best_val_rmse = val_eval
else:
count = count + 1
pre_val_eval = val_eval
print "Loss: %.5f Elpased: %.4fs Converge: %.6f Tr: %.5f Val: %.5f Te: %.5f" % (
loss, elapsed, converge, tr_eval, val_eval, te_eval)
f1.write("Loss: %.5f Elpased: %.4fs Converge: %.6f Tr: %.5f Val: %.5f Te: %.5f\n" % (
loss, elapsed, converge, tr_eval, val_eval, te_eval))
if (count >= endure_count and iteration > min_iter):
# if (count == endure_count):
break
elif (iteration < min_iter):
count = 0
PREV_LOSS = loss
iteration += 1
f1.close()
return best_train_rmse, best_test_rmse, best_val_rmse