def run():
# gpu setting
gpu = torch.device('cuda:' + str(opt.gpu))
# for train
lr, batch_size, num_ns = opt.lr, opt.batch_size, opt.num_ns
reg = opt.reg
# dataset
data_path, dataset, LOO_seed = opt.data_path, opt.dataset, opt.LOO_seed
user_count, item_count, train_mat, train_interactions, valid_sample, test_sample, candidates = read_LOO_settings(
data_path, dataset, LOO_seed)
train_dataset = implicit_CF_dataset(user_count, item_count, train_mat,
train_interactions, num_ns)
test_dataset = implicit_CF_dataset_test(user_count, test_sample,
valid_sample, candidates)
train_loader = data.DataLoader(train_dataset,
batch_size=batch_size,
shuffle=True)
# Read teacher
teacher_dims = 200
teacher_model = BPR(user_count, item_count, teacher_dims, gpu)
with torch.no_grad():
teacher_model_path = opt.teacher_path + dataset + "/" + 'bpr_0.001_200_0.001.model_0'
teacher_model = teacher_model.to(gpu)
teacher_model.load_state_dict(torch.load(teacher_model_path))
teacher_user_emb, teacher_item_emb = teacher_model.get_embedding()
del teacher_model
# Student model
dim = int(teacher_dims * opt.percent)
model = BPR_DE(user_count,
item_count,
teacher_user_emb,
teacher_item_emb,
gpu=gpu,
student_dim=dim,
num_experts=opt.num_expert)
# optimizer
model = model.to(gpu)
optimizer = optim.Adam(model.parameters(), lr=lr, weight_decay=reg)
# start train
LOO_DE_run(opt,
model,
gpu,
optimizer,
train_loader,
test_dataset,
model_save_path=None)
def run():
# gpu setting
gpu = torch.device('cuda:' + str(opt.gpu))
# dataset setting
data_path, dataset, LOO_seed = opt.data_path, opt.dataset, opt.LOO_seed
# for train
model, lr, batch_size, num_ns = opt.model, opt.lr, opt.batch_size, opt.num_ns
reg = opt.reg
# for URRD
user_topk_dict = load_pickle('for_KD/',
'citeULike.bpr.teacher_topk_dict_0')
user_count, item_count, train_mat, train_interactions, valid_sample, test_sample, candidates = read_LOO_settings(
data_path, dataset, LOO_seed)
print("User::", user_count, "Item::", item_count, "Interactions::",
len(train_interactions))
train_dataset = implicit_CF_dataset_URRD(user_count, item_count, train_mat,
train_interactions, num_ns, gpu,
user_topk_dict, opt.U_T, opt.U_K,
opt.U_L)
test_dataset = implicit_CF_dataset_test(user_count, test_sample,
valid_sample, candidates)
train_loader = data.DataLoader(train_dataset,
batch_size=batch_size,
shuffle=True)
# model
teacher_dims = 200
dim = int(teacher_dims * opt.percent)
model = BPR(user_count, item_count, dim, gpu)
# optimizer
model = model.to(gpu)
optimizer = optim.Adam(model.parameters(), lr=lr, weight_decay=reg)
# start train
LOO_URRD_run(opt,
model,
gpu,
optimizer,
train_loader,
test_dataset,
model_save_path=None)
def run():
# gpu setting
gpu = torch.device('cuda:' + str(opt.gpu))
# for training
model, lr, batch_size, num_ns = opt.model, opt.lr, opt.batch_size, opt.num_ns
reg = opt.reg
save = opt.save
# dataset
data_path, dataset, LOO_seed = opt.data_path, opt.dataset, opt.LOO_seed
user_count, item_count, train_mat, train_interactions, valid_sample, test_sample, candidates = read_LOO_settings(
data_path, dataset, LOO_seed)
train_dataset = implicit_CF_dataset(user_count, item_count, train_mat,
train_interactions, num_ns)
test_dataset = implicit_CF_dataset_test(user_count, test_sample,
valid_sample, candidates)
train_loader = data.DataLoader(train_dataset,
batch_size=batch_size,
shuffle=True)
dim = opt.dim
model = BPR(user_count, item_count, dim, gpu)
# optimizer
model = model.to(gpu)
optimizer = optim.Adam(model.parameters(), lr=lr, weight_decay=reg)
print("User::", user_count, "Item::", item_count, "Interactions::",
len(train_interactions))
# to save model
model_save_path = 'None'
if (save == 1):
model_save_path = './Saved_models/' + opt.dataset + "/" + str(
opt.model) + "_" + str(opt.lr) + "_" + str(opt.dim) + "_" + str(
opt.reg) + '.model' + "_" + str(opt.LOO_seed)
# start train
LOO_run(opt, model, gpu, optimizer, train_loader, test_dataset,
model_save_path)