def predict(self, user_ids, item_ids):
max_seq_len = self._max_user_seq_len
self._net.train(False)
item_seqs = self._interactions.get_batch_seqs(user_ids, item_ids, max_seq_len=max_seq_len)
item_seqs = Variable(my_utils.gpu(my_utils.numpy2tensor(item_seqs).type(torch.LongTensor),
self._use_cuda), requires_grad=False)
out = self._net(item_seqs,
Variable(my_utils.gpu(my_utils.numpy2tensor(user_ids), self._use_cuda)),
Variable(my_utils.gpu(my_utils.numpy2tensor(item_ids), self._use_cuda)))
return my_utils.cpu(out).detach().data.numpy().flatten()
def _reset_transform_identity(self):
self._W1.weight.data.copy_(my_utils.numpy2tensor(
np.concatenate(
(np.identity(self._embedding_size), -np.identity(self._embedding_size))
, axis=0).T
)
)
self._W2.weight.data.copy_(my_utils.numpy2tensor(
np.concatenate(
(np.identity(self._embedding_size), np.identity(self._embedding_size))
, axis=0).T)
)
def _make_mask(self, x):
mask = np.asarray(my_utils.tensor2numpy(x.data.cpu().clone()), dtype=np.float64)
# mask = my_utils.tensor2numpy(x != 0)
mask[mask != gc.PADDING_IDX] = 1.0
# mask[mask <= 0] = float('inf')
mask[mask <= 0] = 65535
return my_utils.gpu(Variable(my_utils.numpy2tensor(mask)).type(torch.FloatTensor), use_cuda=my_utils.is_cuda(x))
def predict(self, user_ids, item_ids, playlist_ids):
max_seq_len = self._max_user_seq_len
self._net.train(False)
#last item id is the target item
item_seqs = self._interactions.get_batch_seqs(user_ids,
item_ids,
playlist_ids,
max_seq_len=max_seq_len,
type='all')
item_seqs = Variable(my_utils.gpu(
my_utils.numpy2tensor(item_seqs).type(torch.LongTensor),
self._use_cuda),
requires_grad=False)
if self._args.data_type == 'upt':
out = self._net(
item_seqs,
Variable(
my_utils.gpu(my_utils.numpy2tensor(user_ids),
self._use_cuda)),
Variable(
my_utils.gpu(my_utils.numpy2tensor(item_ids),
self._use_cuda)),
Variable(
my_utils.gpu(my_utils.numpy2tensor(playlist_ids),
self._use_cuda)),
)
elif self._args.data_type == 'ut':
out = self._net(
item_seqs,
Variable(
my_utils.gpu(my_utils.numpy2tensor(user_ids),
self._use_cuda)),
Variable(
my_utils.gpu(my_utils.numpy2tensor(item_ids),
self._use_cuda)), None)
elif self._args.data_type == 'pt':
out = self._net(
item_seqs,
Variable(
my_utils.gpu(my_utils.numpy2tensor(playlist_ids),
self._use_cuda)),
Variable(
my_utils.gpu(my_utils.numpy2tensor(item_ids),
self._use_cuda)), None)
else:
import sys
print 'not supporting data type:', self._args.data_type
sys.exit(1)
return my_utils.cpu(out).detach().data.numpy().flatten()
def _reset_transform_identity(self):
self._W1.weight.data.copy_(my_utils.numpy2tensor
(
np.concatenate(
(-np.identity(self._embedding_size),
np.identity(self._embedding_size)
)
, axis=0
).T # initialy, it is subtractions of target user u and target item j, u-j
)
)
if self._n_playlists:
self._W2.weight.data.copy_(my_utils.numpy2tensor
(
np.concatenate(
(-np.identity(self._embedding_size),
np.identity(self._embedding_size)
)
, axis=0
).T # initialy, it is subtractions of target user u and target item j, u-j
)
)
def _reset_weights(self):
self._user_embeddings.weight.data.normal_(0, 1.0 / self._embedding_size)
self._user_embeddings.weight.data[gc.PADDING_IDX].fill_(0)
self._item_embeddings.weight.data.normal_(0, 1.0 / self._embedding_size)
self._item_embeddings.weight.data[gc.PADDING_IDX].fill_(0)
if self._sum_func: initializer.lecun_uniform_initialization(self._sum_func.weight)
for i , fc in enumerate(self._fcs):
if i == 0:
fc.weight.data.copy_(my_utils.numpy2tensor(
np.concatenate(
(np.identity(self._embedding_size), -np.identity(self._embedding_size))
, axis=0).T # initialy, it is subtractions of target user u and target item j, u-j
))
else:
fc.weight.data.normal_(0, 1.0 / self._embedding_size)
def init_parameters(self):
'''
init the parameters in here
:return:
'''
position_enc = np.array([[
pos / np.power(10000, 2 * (j // 2) / self._embedding_dim)
for j in range(self._embedding_dim)
] if pos != gc.PADDING_IDX else np.zeros(self._embedding_dim)
for pos in range(self._num_embeddings)])
position_enc[1:, 0::2] = np.sin(position_enc[1:, 0::2]) # dim 2i
position_enc[1:, 1::2] = np.cos(position_enc[1:, 1::2]) # dim 2i+1
# return my_utils.numpy2tensor(position_enc).type(torch.FloatTensor)
self.weight.data = my_utils.numpy2tensor(position_enc).type(
torch.FloatTensor)
def _get_neg_pred(self,
item_seqs_var,
batch_user_var,
batch_playlist_var,
batch_negative_items,
adv=False):
'''
user_ids are numpy data
:param user_ids:
:param user_seqs:
:return:
'''
negative_prediction = None
for i in range(batch_negative_items.shape[1]):
negative_items = batch_negative_items[:, i]
batch_neg_item_var = Variable(
my_utils.gpu(my_utils.numpy2tensor(negative_items).type(
torch.LongTensor),
use_cuda=self._use_cuda))
if self._args.data_type == 'upt':
tmp_negative_prediction = self._net(item_seqs_var,
batch_user_var,
batch_neg_item_var,
batch_playlist_var,
adv=adv)
elif self._args.data_type == 'pt':
tmp_negative_prediction = self._net(item_seqs_var,
batch_playlist_var,
batch_neg_item_var,
adv=adv)
else: #ut
tmp_negative_prediction = self._net(item_seqs_var,
batch_user_var,
batch_neg_item_var,
adv=adv)
if negative_prediction is None:
negative_prediction = tmp_negative_prediction
else:
negative_prediction = torch.max(negative_prediction,
tmp_negative_prediction)
return negative_prediction
def load_checkpoint(self, args):
lst_models = ['sdm', 'sdp']
if args.eval:
lst_models = ['sdm', 'sdp', 'dmmr']
if args.model in lst_models:
best_hits = 0.0
best_ndcg = 0.0
best_saved_file = ''
saved_file_pattern = '%s_%s_%s*' % (args.dataset, args.model,
self._make_model_desc(args))
for filepath in glob.glob(
os.path.join(args.saved_path, saved_file_pattern)):
# filepath = os.path.join(args.saved_path, fname)
if os.path.isfile(filepath):
print("=> loading checkpoint '{}'".format(filepath))
checkpoint = torch.load(filepath)
hits = float(checkpoint['best_hits'])
ndcg = float(checkpoint['best_ndcg'])
if hits > best_hits or (hits == best_hits
and ndcg > best_ndcg):
# if ndcg > best_ndcg or (ndcg == best_ndcg and hits > best_hits):
best_saved_file = filepath
best_hits = hits
best_ndcg = ndcg
if best_saved_file != '':
checkpoint = torch.load(best_saved_file)
self._net.load_state_dict(checkpoint['model'])
self._optimizer.load_state_dict(checkpoint['optimizer'])
print("=> loaded checkpoint '{}' (epoch {})".format(
best_saved_file, checkpoint['epoch']))
return (best_hits, best_ndcg)
else:
#sdmr : Deep Metric memory Recommender
#load best sdp checkpoint
best_sdp_file = ''
best_sdp_hits, best_sdp_ndcgs = 0, 0
sdp_files_pattern = '%s_sdp_%s*' % (
args.dataset, self._make_model_desc(args, 'sdp'))
# print sdp_files_pattern
for filepath in glob.glob(
os.path.join(args.saved_path, sdp_files_pattern)):
print filepath
if os.path.isfile(filepath):
print("=> loading checkpoint '{}'".format(filepath))
if args.cuda:
checkpoint = torch.load(filepath)
else:
checkpoint = torch.load(
filepath,
map_location=lambda storage, loc: storage)
hits = float(checkpoint['best_hits'])
ndcg = float(checkpoint['best_ndcg'])
if hits > best_sdp_hits or (hits == best_sdp_hits
and ndcg > best_sdp_ndcgs):
best_sdp_file = filepath
best_sdp_hits = hits
best_sdp_ndcgs = ndcg
#load best sdm checkpoint
best_sdm_file = ''
best_sdm_hits, best_sdm_ndcgs = 0, 0
sdm_files_pattern = '%s_sdm_%s*' % (
args.dataset, self._make_model_desc(args, 'sdm'))
for filepath in glob.glob(
os.path.join(args.saved_path, sdm_files_pattern)):
if os.path.isfile(filepath):
print("=> loading checkpoint '{}'".format(filepath))
if args.cuda:
checkpoint = torch.load(filepath)
else:
checkpoint = torch.load(
filepath,
map_location=lambda storage, loc: storage)
hits = float(checkpoint['best_hits'])
ndcg = float(checkpoint['best_ndcg'])
if hits > best_sdm_hits or (hits == best_sdm_hits
and ndcg > best_sdm_ndcgs):
# if ndcg > best_sdm_ndcgs or (ndcg == best_sdm_ndcgs and hits > best_sdm_hits):
best_sdm_file = filepath
best_sdm_hits = hits
best_sdm_ndcgs = ndcg
#now loading best checkpoints from sdp and sdm for sdmr:
if best_sdp_file != '':
#load checkpoint into cpu if not using cuda
if args.cuda:
checkpoint = torch.load(best_sdp_file)
else:
checkpoint = torch.load(
best_sdp_file,
map_location=lambda storage, loc: storage)
self._net._sdp.load_state_dict(checkpoint['model'])
print("=> loaded checkpoint '{}' (epoch {})".format(
best_sdp_file, checkpoint['epoch']))
# no train or re-train the sdm and sdp weights?
for params in self._net._sdp.parameters():
params.requires_grad = bool(args.sdmr_retrain)
if best_sdm_file != '':
#load checkpoint into gpu or in cpu if not using cuda
if args.cuda:
checkpoint = torch.load(best_sdm_file)
else:
checkpoint = torch.load(
best_sdm_file,
map_location=lambda storage, loc: storage)
self._net._sdm.load_state_dict(checkpoint['model'])
print("=> loaded checkpoint '{}' (epoch {})".format(
best_sdm_file, checkpoint['epoch']))
#no train the sdm and sdp?
for params in self._net._sdm.parameters():
params.requires_grad = bool(args.sdmr_retrain)
sum_weights = np.concatenate(
((1 - args.beta) *
my_utils.tensor2numpy(self._net._sdp._sum_func.weight.data),
args.beta *
my_utils.tensor2numpy(self._net._sdm._sum_func.weight.data)),
axis=1)
sum_bias = (1-args.beta) * my_utils.tensor2numpy(self._net._sdp._sum_func.bias.data) + \
args.beta * my_utils.tensor2numpy(self._net._sdm._sum_func.bias.data)
self._net._sum_func.weight.data.copy_(
my_utils.numpy2tensor(sum_weights))
self._net._sum_func.bias.data.copy_(
my_utils.numpy2tensor(sum_bias))
return 0, 0
def fit(self,
interactions,
verbose=False,
topN=10,
vadRatings=None,
vadNegatives=None,
testRatings=None,
testNegatives=None,
max_seq_len=100,
args=None):
if not self._is_initialized():
self._initialize(interactions, max_seq_len=max_seq_len)
user_ids = interactions._user_ids.astype(np.int64)
item_ids = interactions._item_ids.astype(np.int64)
playlist_ids = interactions._playlists.astype(np.int64)
# self._check_input(sequences)
if verbose:
best_hit = 0.0
best_ndcg = 0.0
best_epoch = 0
test_hit, test_ndcg = 0.0, 0.0
if args.load_best_chkpoint > 0:
# print self._net.parameters
best_hit, best_ndcg = self.load_checkpoint(args)
print 'Results from best checkpoints ...'
# print self._net.parameters
if verbose:
t1 = time.time()
hits, ndcg = my_evaluator.evaluate(self, vadRatings,
vadNegatives, topN)
for topN in range(10, 11, 1):
hits_test, ndcg_test = my_evaluator.evaluate(
self, testRatings, testNegatives, topN)
t2 = time.time()
eval_time = t2 - t1
print(
'| Eval time: %d '
'| Vad hits@%d = %.3f | Vad ndcg@%d = %.3f '
'| Test hits@%d = %.3f | Test ndcg@%d = %.3f |' %
(eval_time, topN, hits, topN, ndcg, topN, hits_test,
topN, ndcg_test))
if topN == 10:
test_hit = hits_test
test_ndcg = ndcg_test
best_hit, best_ndcg = hits, ndcg
topN = 10
print 'End!'
if args.eval:
print 'Evaluation using the saved checkpoint done!'
return
if self._has_params():
for epoch in range(self._n_iters):
self._lr_schedule.step(epoch)
self._net.train() # set training environment
t1 = time.time()
users, items, playlists = my_utils.shuffle(
user_ids,
item_ids,
playlist_ids,
random_state=self._random_state)
users = np.asarray(users)
items = np.asarray(items)
playlists = np.asarray(playlists)
#neg_items = self._sampler.random_neg_items(users, num_neg=args.num_neg)
neg_items = self._sampler.random_neg_items(
playlists, num_neg=args.num_neg)
epoch_loss = 0.0
epoch_noise_loss = 0.0
t1 = time.time()
total_interactions = 0
for (minibatch_idx,
(batch_user, batch_item, batch_playlists,
batch_neg_items)) in enumerate(
my_utils.minibatch(users,
items,
playlists,
neg_items,
batch_size=self._batch_size)):
total_interactions += len(batch_user) # or batch_size
if args.model == 'mass' or args.model == 'masr':
item_seqs = interactions.get_batch_seqs(
batch_user,
batch_item,
batch_playlists,
max_seq_len=self._max_user_seq_len,
type='all')
item_seqs_var = Variable(my_utils.gpu(
my_utils.numpy2tensor(item_seqs).type(
torch.LongTensor), self._use_cuda),
requires_grad=False)
else:
item_seqs, item_seqs_var = None, None
batch_user_var = Variable(
my_utils.gpu(my_utils.numpy2tensor(batch_user).type(
torch.LongTensor),
use_cuda=self._use_cuda))
batch_item_var = Variable(
my_utils.gpu(my_utils.numpy2tensor(batch_item).type(
torch.LongTensor),
use_cuda=self._use_cuda))
if args.adv:
########################### construct adversarial pertubations:#################################
if args.model == 'mdr' or args.model == 'mass':
if args.data_type == 'upt' or args.data_type == 'pt':
batch_playlist_var = Variable(
my_utils.gpu(my_utils.numpy2tensor(
batch_playlists).type(
torch.LongTensor),
use_cuda=self._use_cuda))
else:
batch_playlist_var = None
if args.data_type == 'upt':
positive_prediction_adv = self._net(
item_seqs_var,
batch_user_var,
batch_item_var,
batch_playlist_var,
adv=True)
negative_prediction_adv = self._get_neg_pred(
item_seqs_var,
batch_user_var,
batch_playlist_var,
batch_neg_items,
adv=True)
elif args.data_type == 'pt':
positive_prediction_adv = self._net(
item_seqs_var,
batch_playlist_var,
batch_item_var,
None,
adv=True)
negative_prediction_adv = self._get_neg_pred(
item_seqs_var,
batch_user_var,
batch_playlist_var,
batch_neg_items,
adv=True)
elif args.data_type == 'ut':
positive_prediction_adv = self._net(
item_seqs_var,
batch_user_var,
batch_item_var,
None,
adv=True)
negative_prediction_adv = self._get_neg_pred(
item_seqs_var,
batch_user_var,
batch_playlist_var,
batch_neg_items,
adv=True)
else:
import sys
print 'error, donot support data_type', args.data_type, ' please select: upt, pt, or ut'
sys.exit(1)
loss_adv = self._loss_func(positive_prediction_adv,
negative_prediction_adv)
self._optimizer.zero_grad()
loss_adv.backward(retain_graph=True)
#update adversarial noise
self._net._update_noise_params()
epoch_noise_loss += my_utils.cpu(loss_adv).data.numpy()
###########################update model's params:#################################
if args.model == 'mdr' or args.model == 'mass':
if args.data_type == 'upt' or args.data_type == 'pt':
batch_playlist_var = Variable(
my_utils.gpu(my_utils.numpy2tensor(
batch_playlists).type(
torch.LongTensor),
use_cuda=self._use_cuda))
else:
batch_playlist_var = None
if args.data_type == 'upt':
positive_prediction_adv = self._net(
item_seqs_var,
batch_user_var,
batch_item_var,
batch_playlist_var,
adv=True)
negative_prediction_adv = self._get_neg_pred(
item_seqs_var,
batch_user_var,
batch_playlist_var,
batch_neg_items,
adv=True)
elif args.data_type == 'pt':
positive_prediction_adv = self._net(
item_seqs_var,
batch_playlist_var,
batch_item_var,
None,
adv=True)
negative_prediction_adv = self._get_neg_pred(
item_seqs_var,
batch_user_var,
batch_playlist_var,
batch_neg_items,
adv=True)
elif args.data_type == 'ut':
positive_prediction_adv = self._net(
item_seqs_var,
batch_user_var,
batch_item_var,
None,
adv=True)
negative_prediction_adv = self._get_neg_pred(
item_seqs_var,
batch_user_var,
batch_playlist_var,
batch_neg_items,
adv=True)
else:
import sys
print 'error, donot support data_type', args.data_type, ' please select: upt, pt, or ut'
sys.exit(1)
loss_adv = self._loss_func(positive_prediction_adv,
negative_prediction_adv)
#normal loss
if args.model == 'mdr' or args.model == 'mass':
if args.data_type == 'upt' or args.data_type == 'pt':
batch_playlist_var = Variable(
my_utils.gpu(my_utils.numpy2tensor(
batch_playlists).type(
torch.LongTensor),
use_cuda=self._use_cuda))
else:
batch_playlist_var = None
if args.data_type == 'upt':
positive_prediction = self._net(
item_seqs_var,
batch_user_var,
batch_item_var,
batch_playlist_var,
adv=False)
negative_prediction = self._get_neg_pred(
item_seqs_var,
batch_user_var,
batch_playlist_var,
batch_neg_items,
adv=False)
elif args.data_type == 'pt':
positive_prediction = self._net(
item_seqs_var,
batch_playlist_var,
batch_item_var,
None,
adv=False)
negative_prediction = self._get_neg_pred(
item_seqs_var,
batch_user_var,
batch_playlist_var,
batch_neg_items,
adv=False)
elif args.data_type == 'ut':
positive_prediction = self._net(item_seqs_var,
batch_user_var,
batch_item_var,
None,
adv=False)
negative_prediction = self._get_neg_pred(
item_seqs_var,
batch_user_var,
batch_playlist_var,
batch_neg_items,
adv=False)
else:
import sys
print 'error, donot support data_type', args.data_type, ' please select: upt, pt, or ut'
sys.exit(1)
loss = self._loss_func(positive_prediction,
negative_prediction)
loss_total = loss + args.reg_noise * loss_adv
epoch_loss += my_utils.cpu(loss_total).data.numpy()
self._optimizer.zero_grad() # clear previous grad
loss_total.backward()
# clear all the grads for noises now since we are not updating the noise.
self._net._clear_grad(adv=True)
self._optimizer.step()
pass
else:
#sample negative items
#batch_neg_items = self.sample_neg_items(batch_user, num_negs=args.num_neg)
if args.model == 'mdr' or args.model == 'mass':
if args.data_type == 'upt' or args.data_type == 'pt':
batch_playlist_var = Variable(
my_utils.gpu(my_utils.numpy2tensor(
batch_playlists).type(
torch.LongTensor),
use_cuda=self._use_cuda))
else:
batch_playlist_var = None
if args.data_type == 'upt':
positive_prediction = self._net(
item_seqs_var, batch_user_var,
batch_item_var, batch_playlist_var)
negative_prediction = self._get_neg_pred(
item_seqs_var, batch_user_var,
batch_playlist_var, batch_neg_items)
elif args.data_type == 'pt':
positive_prediction = self._net(
item_seqs_var, batch_playlist_var,
batch_item_var, None)
negative_prediction = self._get_neg_pred(
item_seqs_var, batch_user_var,
batch_playlist_var, batch_neg_items)
elif args.data_type == 'ut':
positive_prediction = self._net(
item_seqs_var, batch_user_var,
batch_item_var, None)
negative_prediction = self._get_neg_pred(
item_seqs_var, batch_user_var,
batch_playlist_var, batch_neg_items)
else:
import sys
print 'error, donot support data_type', args.data_type, ' please select: upt, pt, or ut'
sys.exit(1)
self._optimizer.zero_grad()
loss = self._loss_func(positive_prediction,
negative_prediction)
#regularizer = self._net._get_l2_loss()
#if self._model == 'mass': loss += self._reg_mass * regularizer
#elif self._model == 'mdr': loss += self._reg_mdr * regularizer
#else: loss += self._reg_mdr * regularizer[0] + self._reg_mass * regularizer[1]
epoch_loss += my_utils.cpu(loss).data.numpy()
loss.backward()
self._optimizer.step()
if gc.DEBUG:
# print 'W1 :', self._net._W1.weight.data
# print 'W2 :', self._net._W2.weight.data
#
# print 'W1 grad:', self._net._W1.weight.grad.data
# print 'W2 grad:', self._net._W2.weight.grad.data
#
print 'Loss is :', loss
print 'Positive prediction:', positive_prediction
print 'negative prediction:', negative_prediction
if minibatch_idx >= 1: break # debug
epoch_loss = epoch_loss / total_interactions
epoch_noise_loss = epoch_noise_loss / total_interactions
t2 = time.time()
epoch_train_time = t2 - t1
if verbose:
t1 = time.time()
hits, ndcg = my_evaluator.evaluate(self, vadRatings,
vadNegatives, topN)
hits_test, ndcg_test = my_evaluator.evaluate(
self, testRatings, testNegatives, topN)
t2 = time.time()
eval_time = t2 - t1
print(
'|Epoch %d | Train time: %d | Train loss: %.3f | Train Noise loss: %.3f | Eval time: %d '
'| Vad hits@%d = %.3f | Vad ndcg@%d = %.3f '
'| Test hits@%d = %.3f | Test ndcg@%d = %.3f |' %
(epoch, epoch_train_time, epoch_loss, epoch_noise_loss,
eval_time, topN, hits, topN, ndcg, topN, hits_test,
topN, ndcg_test))
if hits > best_hit or (hits == best_hit
and ndcg > best_ndcg):
# if (hits + ndcg) > (best_hit + best_ndcg):
# if (ndcg > best_ndcg) or (ndcg == best_ndcg and hits > best_hit):
best_hit, best_ndcg, best_epoch = hits, ndcg, epoch
test_hit, test_ndcg = hits_test, ndcg_test
#self.save_checkpoint(args, best_hit, best_ndcg, epoch)
if args.out:
self.save_checkpoint(args, hits, ndcg,
epoch) # save best params
else:
print('|Epoch %d | Train time: %d | loss: %.3f' %
(epoch, epoch_train_time, epoch_loss))
if np.isnan(epoch_loss) or epoch_loss == 0.0:
raise ValueError(
'Degenerate epoch loss: {}'.format(epoch_loss))
print(
'Best result: '
'| vad hits@%d = %.3f | vad ndcg@%d = %.3f '
'| test hits@%d = %.3f | test ndcg@%d = %.3f | epoch = %d' %
(topN, best_hit, topN, best_ndcg, topN, test_hit, topN,
test_ndcg, best_epoch))
def fit(self, interactions, topN=10,
testRatings=None, testNegatives=None,
max_seq_len=5, args=None):
if not self._is_initialized():
self._initialize(interactions, max_seq_len=max_seq_len)
user_ids = interactions._user_ids.astype(np.int64)
item_ids = interactions._item_ids.astype(np.int64)
best_hit = 0.0
best_ndcg = 0.0
best_epoch = 0
test_hit, test_ndcg = 0.0, 0.0
if args.load_best_chkpoint > 0:
best_hit, best_ndcg = self.load_checkpoint(args)
print 'Results from best checkpoints ...'
t1 = time.time()
hits, ndcgs = my_evaluator.evaluate(self, testRatings, testNegatives, topN)
t2 = time.time()
eval_time = t2-t1
best_hit, best_ndcg = hits, ndcgs
print('| Eval time: %d '
'| Test hits@%d = %.3f | Test ndcg@%d = %.3f |'
% (eval_time, topN, hits, topN, ndcgs))
topN = 10
print 'End!'
if args.eval:
print 'Evaluation using the saved checkpoint done!'
return
if self._has_params():
for epoch in range(self._n_iters):
self._lr_schedule.step(epoch)
self._net.train() # set training environment
users, items = my_utils.shuffle(user_ids,
item_ids,
random_state=self._random_state)
users = np.asarray(users)
items = np.asarray(items)
neg_items = self._sampler.random_neg_items(users, num_neg=args.num_neg)
# copy to GPU:
user_vars = Variable(my_utils.gpu(my_utils.numpy2tensor(users).type(torch.LongTensor),
use_cuda=self._use_cuda))
item_vars = Variable(my_utils.gpu(my_utils.numpy2tensor(items).type(torch.LongTensor),
use_cuda=self._use_cuda))
neg_item_vars = Variable(my_utils.gpu(my_utils.numpy2tensor(neg_items).type(torch.LongTensor),
use_cuda=self._use_cuda))
context_vars = None
#if args.model != 'sdp':
contexts = interactions.get_batch_seqs(users, items,
max_seq_len=self._max_user_seq_len)
context_vars = Variable(my_utils.gpu(my_utils.numpy2tensor(contexts).type(torch.LongTensor),
self._use_cuda), requires_grad=False)
epoch_loss = 0.0
t1 = time.time()
total_interactions = 0
for (minibatch_idx, (batch_users, batch_items, batch_neg_items, batch_contexts )) in enumerate(
my_utils.minibatch(
user_vars,
item_vars,
neg_item_vars,
context_vars,
batch_size=self._batch_size)):
total_interactions += len(batch_users) # or batch_size
if args.model == 'sdp':
batch_contexts = None
positive_prediction = self._net(batch_contexts, batch_users, batch_items)
negative_prediction = self._get_neg_pred(batch_contexts, batch_users, batch_neg_items)
self._optimizer.zero_grad()
loss = self._loss_func(positive_prediction,
negative_prediction
)
epoch_loss += my_utils.cpu(loss).data.numpy()
loss.backward()
self._optimizer.step()
epoch_loss = epoch_loss / total_interactions
t2 = time.time()
epoch_train_time = t2 - t1
t1 = time.time()
hits, ndcgs = my_evaluator.evaluate(self, testRatings, testNegatives, topN)
t2 = time.time()
eval_time = t2 - t1
print('|Epoch %d | Train time: %d | Train loss: %.3f | Eval time: %d '
'| Test hits@%d = %.3f | Test ndcg@%d = %.3f |'
% (epoch, epoch_train_time, epoch_loss, eval_time, topN, hits, topN, ndcgs))
if hits > best_hit or (hits == best_hit and ndcgs >= best_ndcg):
best_hit, best_ndcg, best_epoch = hits, ndcgs, epoch
if args.out:
self.save_checkpoint(args, hits, ndcgs, epoch) # save best params
if np.isnan(epoch_loss) or epoch_loss == 0.0:
raise ValueError('Degenerate epoch loss: {}'
.format(epoch_loss))
print ('Best result: '
'| test hits@%d = %.3f | test ndcg@%d = %.3f | epoch = %d' % (topN, best_hit, topN, best_ndcg,
best_epoch))
