def main():
args = get_args()
data_path = os.path.join(args.iobasedir, 'processed/downloads',
args.data_set)
log_path = os.path.join(args.iobasedir, 'logs')
log_file = os.path.join(args.iobasedir, 'logs', 'UB.log')
mkdirp(log_path)
set_logger(log_file)
for filename in os.listdir(data_path):
data_file = os.path.join(data_path, filename)
topic = filename[:-5]
docs, refs = load_data(data_file)
if not refs:
continue
if not args.summary_size:
summary_size = len(' '.join(refs[0]).split(' '))
else:
summary_size = int(args.summary_size)
logger.info('Topic ID: %s ', topic)
logger.info('###')
logger.info('Summmary_len: %d', summary_size)
algos = ['UB1', 'UB2']
for algo in algos:
get_summary_scores(algo, docs, refs, summary_size, language, rouge)
logger.info('###')
def main():
args = get_args()
rouge_dir = os.path.join(
os.path.dirname(os.path.dirname(os.path.abspath(__file__))),
'rouge/RELEASE-1.5.5/')
data_path = os.path.join(args.iobasedir, args.data_setpath)
log_path = os.path.join(args.iobasedir, 'logs')
log_file = os.path.join(
args.iobasedir, 'logs',
'baselines_%s_%s.log' % (args.data_set, args.summary_size))
mkdirp(log_path)
set_logger(log_file)
for filename in os.listdir(data_path):
data_file = os.path.join(data_path, filename)
topic = filename[:-5]
try:
docs, refs = load_data(data_file)
except:
pass
if not refs:
continue
if not args.summary_size:
summary_size = len(" ".join(refs[0]).split(' '))
else:
summary_size = int(args.summary_size)
logger.info('Topic ID: %s', topic)
logger.info('###')
logger.info('Summmary_len: %d', summary_size)
rouge = Rouge(rouge_dir)
algos = ['UB1', 'UB2', 'ICSI', 'Luhn', 'LexRank', 'LSA', 'KL']
for algo in algos:
get_summary_scores(algo, docs, refs, summary_size, args.language,
rouge)
rouge._cleanup()
logger.info('###')
def test():
# Load data
logger.info("✔︎ Loading data...")
logger.info("✔︎ Training data processing...")
train_data = dh.load_data(Config().TRAININGSET_DIR)
logger.info("✔︎ Test data processing...")
test_data = dh.load_data(Config().TESTSET_DIR)
logger.info("✔︎ Load negative sample...")
with open(Config().NEG_SAMPLES, 'rb') as handle:
neg_samples = pickle.load(handle)
# Load model
dr_model = torch.load(MODEL_DIR)
dr_model.eval()
item_embedding = dr_model.encode.weight
hidden = dr_model.init_hidden(Config().batch_size)
hitratio_numer = 0
hitratio_denom = 0
ndcg = 0.0
for i, x in enumerate(
dh.batch_iter(train_data,
Config().batch_size,
Config().seq_len,
shuffle=False)):
uids, baskets, lens = x
dynamic_user, _ = dr_model(baskets, lens, hidden)
for uid, l, du in zip(uids, lens, dynamic_user):
scores = []
du_latest = du[l - 1].unsqueeze(0)
# calculating <u,p> score for all test items <u,p> pair
positives = test_data[test_data['userID'] == uid].baskets.values[
0] # list dim 1
p_length = len(positives)
positives = torch.LongTensor(positives)
# Deal with positives samples
scores_pos = list(
torch.mm(du_latest,
item_embedding[positives].t()).data.numpy()[0])
for s in scores_pos:
scores.append(s)
# Deal with negative samples
negtives = random.sample(list(neg_samples[uid]), Config().neg_num)
negtives = torch.LongTensor(negtives)
scores_neg = list(
torch.mm(du_latest,
item_embedding[negtives].t()).data.numpy()[0])
for s in scores_neg:
scores.append(s)
# Calculate hit-ratio
index_k = []
for k in range(Config().top_k):
index = scores.index(max(scores))
index_k.append(index)
scores[index] = -9999
hitratio_numer += len((set(np.arange(0, p_length)) & set(index_k)))
hitratio_denom += p_length
# Calculate NDCG
u_dcg = 0
u_idcg = 0
for k in range(Config().top_k):
if index_k[k] < p_length: # 长度 p_length 内的为正样本
u_dcg += 1 / math.log(k + 1 + 1, 2)
u_idcg += 1 / math.log(k + 1 + 1, 2)
ndcg += u_dcg / u_idcg
hitratio = hitratio_numer / hitratio_denom
ndcg = ndcg / len(train_data)
print('Hit ratio[{0}]: {1}'.format(Config().top_k, hitratio))
print('NDCG[{0}]: {1}'.format(Config().top_k, ndcg))
def train():
# Load data
logger.info("✔︎ Loading data...")
logger.info("✔︎ Training data processing...")
train_data = dh.load_data(Config().TRAININGSET_DIR)
logger.info("✔︎ Validation data processing...")
validation_data = dh.load_data(Config().VALIDATIONSET_DIR)
logger.info("✔︎ Test data processing...")
test_data = dh.load_data(Config().TESTSET_DIR)
logger.info("✔︎ Load negative sample...")
with open(Config().NEG_SAMPLES, 'rb') as handle:
neg_samples = pickle.load(handle)
# Model config
model = DRModel(Config())
# Optimizer
optimizer = torch.optim.Adam(model.parameters(), lr=Config().learning_rate)
def bpr_loss(uids, baskets, dynamic_user, item_embedding):
"""
Bayesian personalized ranking loss for implicit feedback.
For an intro on BPR: https://towardsdatascience.com/recommender-system-using-bayesian-personalized-ranking-d30e98bba0b9
Args:
uids: batch of users' ID
baskets: batch of users' baskets
dynamic_user: batch of users' dynamic representations
item_embedding: item_embedding matrix
"""
loss = 0
for uid, bks, du in zip(uids, baskets, dynamic_user):
du_p_product = torch.mm(
du, item_embedding.t()) # shape: [pad_len, num_item]
loss_u = [] # loss for user
for t, basket_t in enumerate(bks):
if basket_t[0] != 0 and t != 0: # wht skipping the first item??
pos_idx = torch.LongTensor(basket_t)
# Sample negative products
neg = random.sample(list(neg_samples[uid]), len(basket_t))
neg_idx = torch.LongTensor(neg)
# Score p(u, t, v > v')
score = du_p_product[t -
1][pos_idx] - du_p_product[t -
1][neg_idx]
# Average Negative log likelihood for basket_t
loss_u.append(torch.mean(-torch.nn.LogSigmoid()(score)))
for i in loss_u:
loss = loss + i / len(loss_u)
avg_loss = torch.div(loss, len(baskets))
return avg_loss
def train_model():
model.train() # turn on training mode for dropout
dr_hidden = model.init_hidden(Config().batch_size)
train_loss = 0
start_time = time.time()
num_batches = ceil(len(train_data) / Config().batch_size)
for i, x in enumerate(
dh.batch_iter(train_data,
Config().batch_size,
Config().seq_len,
shuffle=True)):
# baskets are padded to seq_length = 12, with [0]
# lens is a list of length corresponding to how many real baskets are
uids, baskets, lens = x
model.zero_grad() # 如果不置零，Variable 的梯度在每次 backward 的时候都会累加
dynamic_user, _ = model(baskets, lens, dr_hidden)
loss = bpr_loss(uids, baskets, dynamic_user, model.encode.weight)
loss.backward()
# Clip to avoid gradient exploding
torch.nn.utils.clip_grad_norm_(model.parameters(), Config().clip)
# Parameter updating
optimizer.step()
train_loss += loss.data
# Logging
if i % Config().log_interval == 0 and i > 0:
elapsed = (time.time() - start_time) / Config().log_interval
cur_loss = train_loss.item() / Config(
).log_interval # turn tensor into float
train_loss = 0
start_time = time.time()
logger.info(
'[Training]| Epochs {:3d} | Batch {:5d} / {:5d} | ms/batch {:02.2f} | Loss {:05.4f} |'
.format(epoch, i, num_batches, elapsed, cur_loss))
def validate_model():
model.eval()
dr_hidden = model.init_hidden(Config().batch_size)
val_loss = 0
start_time = time.time()
num_batches = ceil(len(validation_data) / Config().batch_size)
for i, x in enumerate(
dh.batch_iter(validation_data,
Config().batch_size,
Config().seq_len,
shuffle=False)):
uids, baskets, lens = x
dynamic_user, _ = model(baskets, lens, dr_hidden)
loss = bpr_loss(uids, baskets, dynamic_user, model.encode.weight)
val_loss += loss.data
# Logging
elapsed = (time.time() - start_time) * 1000 / num_batches
val_loss = val_loss.item() / num_batches
logger.info(
'[Validation]| Epochs {:3d} | Elapsed {:02.2f} | Loss {:05.4f} |'.
format(epoch, elapsed, val_loss))
return val_loss
def test_model():
model.eval()
item_embedding = model.encode.weight
dr_hidden = model.init_hidden(Config().batch_size)
hitratio_numer = 0
hitratio_denom = 0
ndcg = 0.0
for i, x in enumerate(
dh.batch_iter(train_data,
Config().batch_size,
Config().seq_len,
shuffle=False)):
uids, baskets, lens = x
dynamic_user, _ = model(baskets, lens, dr_hidden)
for uid, l, du in zip(uids, lens, dynamic_user):
scores = []
# we use the last output as user representation
du_latest = du[l - 1].unsqueeze(0)
# calculating <u,p> score for all test items <u,p> pair
positives = test_data[test_data['userID'] ==
uid].baskets.values[0] # list dim 1
p_length = len(positives)
positives = torch.LongTensor(positives)
# Deal with positives samples
scores_pos = list(
torch.mm(du_latest,
item_embedding[positives].t()).data.numpy()[0])
for s in scores_pos:
scores.append(s)
# Deal with negative samples
negtives = random.sample(list(neg_samples[uid]),
Config().neg_num)
negtives = torch.LongTensor(negtives)
scores_neg = list(
torch.mm(du_latest,
item_embedding[negtives].t()).data.numpy()[0])
for s in scores_neg:
scores.append(s)
# Calculate hit-ratio
index_k = []
for k in range(Config().top_k):
index = scores.index(max(scores))
index_k.append(index)
scores[index] = -9999
hitratio_numer += len(
(set(np.arange(0, p_length)) & set(index_k)))
hitratio_denom += p_length
# Calculate NDCG
u_dcg = 0
u_idcg = 0
for k in range(Config().top_k):
if index_k[k] < p_length: # 长度 p_length 内的为正样本
u_dcg += 1 / math.log(k + 1 + 1, 2)
u_idcg += 1 / math.log(k + 1 + 1, 2)
ndcg += u_dcg / u_idcg
hit_ratio = hitratio_numer / hitratio_denom
ndcg = ndcg / len(train_data)
logger.info(
'[Test]| Epochs {:3d} | Hit ratio {:02.4f} | NDCG {:05.4f} |'.
format(epoch, hit_ratio, ndcg))
return hit_ratio, ndcg
timestamp = str(int(time.time()))
out_dir = os.path.abspath(os.path.join(os.path.curdir, "runs", timestamp))
if not os.path.exists(out_dir):
os.makedirs(out_dir)
logger.info('Save into {0}'.format(out_dir))
checkpoint_dir = out_dir + '/model-{epoch:02d}-{hitratio:.4f}-{ndcg:.4f}.model'
best_hit_ratio = None
try:
# Training
for epoch in range(Config().epochs):
train_model()
logger.info('-' * 89)
val_loss = validate_model()
logger.info('-' * 89)
hit_ratio, ndcg = test_model()
logger.info('-' * 89)
# Checkpoint
if not best_hit_ratio or hit_ratio > best_hit_ratio:
with open(
checkpoint_dir.format(epoch=epoch,
hitratio=hit_ratio,
ndcg=ndcg), 'wb') as f:
torch.save(model, f)
best_hit_ratio = hit_ratio
except KeyboardInterrupt:
logger.info('*' * 89)
logger.info('Early Stopping!')
if __name__ == "__main__":
# LOAD RAW DATA $ WORD VECTORS
EVAL_DATASET = '../../dataset/PMtask_TestSet.xml'
MODE = "eval"
WV_PATH = '../../embeddings/PubMed-w2v.txt'
WV_DIMS = 200
MAX_SENT_LENGTH = 45
MAX_SENTS = 23
print("loading word embeddings...")
word2idx, idx2word, embeddings = load_word_vectors(WV_PATH, WV_DIMS, True)
docs, labels, ids = load_data(EVAL_DATASET, MODE)
# convert strings to lists of tokens
print("Tokenizing...")
docs = [[text_to_word_sequence(sent) for sent in sent_tokenize(doc)]
for doc in docs]
# convert words to word indexes
print("Vectorizing...")
docs = [vectorize_doc(doc, word2idx, MAX_SENTS, MAX_SENT_LENGTH)
for doc in docs]
docs = numpy.array(docs)
# LOAD SAVED MODEL
print("Loading model from disk...", end=" ")
model_name = "../experiments/task1_hGRU_2017-10-14 17:25:22.hdf5"
def test(saved_file):
# Load data
logger.info("✔︎ Loading data...")
logger.info("✔︎ Training data processing...")
train_data = dh.load_data(Config().TRAININGSET_DIR)
logger.info("✔︎ Test data processing...")
test_data = dh.load_data(Config().TESTSET_DIR)
logger.info("✔︎ Load negative sample...")
# with open(Config().NEG_SAMPLES, 'rb') as handle:
# neg_samples = pickle.load(handle)
neg_samples = {}
item_list = [i for i in range(336)]
# Load model
MODEL_DIR = dh.load_model_file(saved_file)
dr_model = torch.load(MODEL_DIR)
dr_model.eval()
item_embedding = dr_model.encode.weight
hidden = dr_model.init_hidden(Config().batch_size)
hitratio_numer = 0
hitratio_denom = 0
hitratio_numer = 0
hitratio_denom = 0
hitratio_numer_10 = 0
hitratio_numer_5 = 0
ndcg = 0.0
ndcg_denom = 0
hitratio_list_5 = []
hitratio_list_10 = []
ndcg_list = []
for i, x in enumerate(
tqdm(
dh.batch_iter(train_data,
Config().batch_size,
Config().seq_len_test,
shuffle=False))):
uids, baskets, lens = x
dynamic_user, _ = dr_model(baskets, lens, hidden)
for uid, l, du in zip(uids, lens, dynamic_user):
scores = []
du_latest = du[l - 1].unsqueeze(0)
# Deal with positives samples
positives = test_data[test_data['userID'] ==
uid].baskets.values[0][:-1] # list dim 1
p_length = len(positives)
positives = torch.LongTensor(positives)
print("positives: ", positives)
# calculating <u,p> score for all test items <u,p> pair
scores_pos = list(
torch.mm(du_latest,
item_embedding[positives].t()).data.cpu().numpy()[0])
for s in scores_pos:
scores.append(s)
print("score_pos: ", score_pos)
# Deal with negative samples
neg_item_list = list(set(item_list).difference(set(positives)))
negtives = random.sample(neg_item_list, Config().neg_num)
negtives = torch.LongTensor(negtives)
scores_neg = list(
torch.mm(du_latest,
item_embedding[negtives].t()).data.cpu().numpy()[0])
for s in scores_neg:
scores.append(s)
print("scores: ", scores)
# Calculate hit-ratio
index_k = []
for k in range(Config().top_k):
index = scores.index(max(scores))
index_k.append(index)
scores[index] = -9999
print("index_k: ", index_k)
hr_5_numer = len((set(np.arange(0, p_length)) & set(index_k[0:5])))
hr_10_numer = len((set(np.arange(0, p_length)) & set(index_k)))
hitratio_numer_10 += hr_10_numer # np.arange()产生等差数列
hitratio_numer_5 += hr_5_numer
hitratio_denom += p_length
hitratio_list_5.append(hr_5_numer / p_length)
hitratio_list_10.append(hr_10_numer / p_length)
# print("hitratio_list_5: ", hitratio_list_5)
# print("hitratio_list_10: ", hitratio_list_10)
# hitratio_numer += len((set(np.arange(0, p_length)) & set(index_k)))
# hitratio_denom += p_length
# Calculate NDCG
u_dcg = 0
u_idcg = 0
for k in range(Config().top_k):
if index_k[k] < p_length: # 长度 p_length 内的为正样本
u_dcg += 1 / math.log(k + 1 + 1, 2)
u_idcg += 1 / math.log(k + 1 + 1, 2)
ndcg += u_dcg / u_idcg
ndcg_denom += 1
ndcg_list.append(u_dcg / u_idcg)
# print("ndcg_list: ", ndcg_list)
hit_ratio_5 = hitratio_numer_5 / hitratio_denom
hit_ratio_10 = hitratio_numer_10 / hitratio_denom
ndcg = ndcg / ndcg_denom
print('Hit ratio@5: {1} | Hit ratio@10: {1}'.format(
hit_ratio_5, hit_ratio_10))
print('NDCG[{0}]: {1}'.format(Config().top_k, ndcg))
return hitratio_list_5, hitratio_list_10, ndcg_list
CORPUS = 'dataset/PMtask_Triage_TrainingSet.xml'
WV_PATH = 'embeddings/PubMed-w2v.txt'
WV_DIMS = 200
PERSIST = True # if True, then save the model to disk
####################
MAX_SENT_LENGTH = 45
MAX_SENTS = 23
##############################################
# Prepare Data
##############################################
print("loading word embeddings...")
word2idx, idx2word, embeddings = load_word_vectors(WV_PATH, WV_DIMS, True)
print("loading data...")
docs, labels, ids = load_data(CORPUS, mode=MODE)
# word_vectors = load_word_vectors(args.embeddings)
# convert strings to lists of tokens
print("Tokenizing...")
docs = [[text_to_word_sequence(sent) for sent in sent_tokenize(doc)]
for doc in docs]
# convert words to word indexes
print("Vectorizing...")
docs = [
vectorize_doc(doc, word2idx, MAX_SENTS, MAX_SENT_LENGTH) for doc in docs
]
docs = numpy.array(docs)
if MODE == "train":
if __name__ == '__main__':
out_file = open(args.out, 'w')
algos = ['UB1', 'UB2', 'LexRank', 'TextRank', 'Luhn', 'ICSI']
R1 = {'UB1': .0, 'UB2': .0, 'ICSI': .0, 'LSA': .0, 'KL': .0, 'Luhn': .0, 'LexRank': .0, 'TextRank': .0}
R2 = {'UB1': .0, 'UB2': .0, 'ICSI': .0, 'LSA': .0, 'KL': .0, 'Luhn': .0, 'LexRank': .0, 'TextRank': .0}
Rl = {'UB1': .0, 'UB2': .0, 'ICSI': .0, 'LSA': .0, 'KL': .0, 'Luhn': .0, 'LexRank': .0, 'TextRank': .0}
Rsu = {'UB1': .0, 'UB2': .0, 'ICSI': .0, 'LSA': .0, 'KL': .0, 'Luhn': .0, 'LexRank': .0, 'TextRank': .0}
blog_sum = .0
for t in types:
cur_path = args.path + '/' + t + '/'
file_names = os.listdir(cur_path)
blog_sum += len(file_names)
for filename in tqdm(file_names):
data_file = os.path.join(cur_path, filename)
docs, refs = load_data(data_file)
sum_len = len(' '.join(refs[0]).split(' ')) * args.sum_len
print('####', filename, '####')
out_file.write(filename + '\n')
for algo in algos:
r1, r2, rl, rsu = get_summary_scores(algo, docs, refs, sum_len)
print algo, r1, r2, rl, rsu
out_file.write(algo + ' ' + str(r1) + ' ' + str(r2) + ' ' + str(rl) + ' ' + str(rsu) + '\n')
R1[algo] += r1
R2[algo] += r2
Rl[algo] += rl
Rsu[algo] += rsu
out_file.close()
print('Final Results')
for algo in algos:
R1[algo] /= blog_sum
def train():
# Load data
logger.info("✔︎ Loading data...")
logger.info("✔︎ Training data processing...")
train_data = dh.load_data(Config().TRAININGSET_DIR)
logger.info("✔︎ Validation data processing...")
validation_data = dh.load_data(Config().VALIDATIONSET_DIR)
logger.info("✔︎ Test data processing...")
test_data = dh.load_data(Config().TESTSET_DIR)
logger.info("✔︎ Load negative sample...")
# with open(Config().NEG_SAMPLES, 'rb') as handle:
# neg_samples = pickle.load(handle)
neg_samples = {}
if torch.cuda.is_available():
model = FVModel(Config()).cuda()
else:
model = FVModel(Config())
# Optimizer
optimizer = torch.optim.Adam(model.parameters(), lr=Config().learning_rate)
def bpr_loss(uids, baskets, dynamic_user, item_list, item_embedding):
"""
Bayesian personalized ranking loss for implicit feedback.
Args:
uids: batch of users' ID
baskets: batch of users' baskets, baskets = daily food categories of users'
dynamic_user: batch of users' dynamic representations
item_embedding: item_embedding matrix
"""
loss = 0
for uid, bks, du in zip(uids, baskets, dynamic_user):
du_p_product = torch.mm(
du, item_embedding.t()) # shape: [pad_len, num_item]
loss_u = [] # loss for user
for t, basket_t in enumerate((bks)):
if basket_t[0] != 0 and t != 0:
basket_t = basket_t[:
-1] # do not consoder the last number, which is weight changes indicator
pos_idx = torch.LongTensor(basket_t)
# Sample negative products
neg_item_list = list(
set(item_list).difference(set(basket_t)))
neg = random.sample(neg_item_list, len(basket_t))
neg_idx = torch.LongTensor(neg)
# Score p(u, t, v > v')
score = du_p_product[t -
1][pos_idx] - du_p_product[t -
1][neg_idx]
# Average Negative log likelihood for basket_t
loss_u.append(torch.mean(-torch.nn.LogSigmoid()(score)))
for i in loss_u:
loss = loss + i / len(loss_u)
# avg_loss = torch.true_divide(loss, len(baskets))
avg_loss = torch.div(loss, len(baskets))
return avg_loss
def train_model():
model.train() # turn on training mode for dropout
dr_hidden = model.init_hidden(Config().batch_size)
train_loss = 0
start_time = time.clock()
#start_time = time.perf_counter
num_batches = ceil(len(train_data) / Config().batch_size)
for i, x in enumerate(
tqdm(
dh.batch_iter(train_data,
Config().batch_size,
Config().seq_len_train,
shuffle=True))):
uids, baskets, lens = x
model.zero_grad() # 如果不置零，Variable 的梯度在每次 backward 的时候都会累加
dynamic_user, _ = model(baskets, lens, dr_hidden)
loss = bpr_loss(uids, baskets, dynamic_user, item_list,
model.encode.weight)
loss.backward()
# Clip to avoid gradient exploding
torch.nn.utils.clip_grad_norm_(model.parameters(), Config().clip)
# Parameter updating
optimizer.step()
train_loss += loss.data
# Logging
if i % Config().log_interval == 0 and i > 0:
elapsed = (time.clock() - start_time) / Config().log_interval
cur_loss = train_loss.item() / Config(
).log_interval # turn tensor into float
train_loss = 0
start_time = time.clock()
logger.info(
'[Training]| Epochs {:3d} | Batch {:5d} / {:5d} | ms/batch {:02.2f} | Loss {:05.4f} |'
.format(epoch, i, num_batches, elapsed, cur_loss))
def validate_model():
model.eval()
dr_hidden = model.init_hidden(Config().batch_size)
val_loss = 0
start_time = time.clock()
num_batches = ceil(len(validation_data) / Config().batch_size)
for i, x in enumerate(
tqdm(
dh.batch_iter(validation_data,
Config().batch_size,
Config().seq_len_valid,
shuffle=False))):
uids, baskets, lens = x
dynamic_user, _ = model(baskets, lens, dr_hidden)
loss = bpr_loss(uids, baskets, dynamic_user, item_list,
model.encode.weight)
val_loss += loss.data
# Logging
elapsed = (time.clock() - start_time) * 1000 / num_batches
val_loss = val_loss.item() / num_batches
logger.info(
'[Validation]| Epochs {:3d} | Elapsed {:02.2f} | Loss {:05.4f} |'.
format(epoch, elapsed, val_loss))
return val_loss
def test_model():
model.eval()
item_embedding = model.encode.weight
dr_hidden = model.init_hidden(Config().batch_size)
hitratio_numer = 0
hitratio_denom = 0
hitratio_numer_10 = 0
hitratio_numer_5 = 0
ndcg = 0.0
ndcg_denom = 0
for i, x in enumerate(
tqdm(
dh.batch_iter(train_data,
Config().batch_size,
Config().seq_len_test,
shuffle=False))):
uids, baskets, lens = x
dynamic_user, _ = model(baskets, lens, dr_hidden)
for uid, l, du in zip(uids, lens, dynamic_user):
scores = []
du_latest = du[l - 1].unsqueeze(0)
# calculating <u,p> score for all test items <u,p> pair
positives = test_data[test_data['userID'] ==
uid].baskets.values[0][:-1] # list dim 1
p_length = len(positives)
positives = torch.LongTensor(positives)
# Deal with positives samples
scores_pos = list(
torch.mm(
du_latest,
item_embedding[positives].t()).data.cpu().numpy()[0])
for s in scores_pos:
scores.append(s)
# Deal with negative samples
neg_item_list = list(set(item_list).difference(set(positives)))
negtives = random.sample(neg_item_list, Config().neg_num)
negtives = torch.LongTensor(negtives)
scores_neg = list(
torch.mm(
du_latest,
item_embedding[negtives].t()).data.cpu().numpy()[0])
for s in scores_neg:
scores.append(s)
# Calculate hit-ratio
index_k = [] # ranking list
# k = 5 and k = 10
for k in range(Config().top_k):
index = scores.index(
max(scores)) # score 最高的category的index
index_k.append(index)
scores[index] = -9999
hitratio_numer_10 += len((set(np.arange(0, p_length))
& set(index_k))) # np.arange()产生等差数列
hitratio_numer_5 += len(
(set(np.arange(0, p_length)) & set(index_k[0:5])))
hitratio_denom += p_length
# Calculate NDCG
u_dcg = 0
u_idcg = 0
for k in range(Config().top_k):
if index_k[k] < p_length: # 长度 p_length 内的为正样本
u_dcg += 1 / math.log(k + 1 + 1, 2)
u_idcg += 1 / math.log(k + 1 + 1, 2)
ndcg += u_dcg / u_idcg
ndcg_denom += 1
hit_ratio_5 = hitratio_numer_5 / hitratio_denom
hit_ratio_10 = hitratio_numer_10 / hitratio_denom
ndcg = ndcg / ndcg_denom
logger.info(
'[Test]| Epochs {:3d} | Hit ratio@5 {:02.4f} | Hit ratio@10 {:02.4f} | NDCG {:05.4f} |'
.format(epoch, hit_ratio_5, hit_ratio_10, ndcg))
return hit_ratio_5, hit_ratio_10, ndcg
timestamp = str(int(time.time()))
out_dir = os.path.join(os.path.curdir, "runs", timestamp)
item_list = [i for i in range(336)]
if not os.path.exists(out_dir):
os.makedirs(out_dir)
logger.info('Save into {0}'.format(out_dir))
checkpoint_dir = out_dir + '/model-{epoch:02d}-{hitratio:.4f}-{ndcg:.4f}.model'
best_hit_ratio = None
# ==================== test
# val_loss = validate_model()
try:
# Training
for epoch in range(Config().epochs):
train_model()
logger.info('-' * 89)
val_loss = validate_model()
logger.info('-' * 89)
hit_ratio_5, hit_ratio_10, ndcg = test_model()
logger.info('-' * 89)
# Checkpoint
if not best_hit_ratio or hit_ratio_10 > best_hit_ratio:
with open(
checkpoint_dir.format(epoch=epoch,
hitratio=hit_ratio_10,
ndcg=ndcg), 'wb') as f:
torch.save(model, f)
best_hit_ratio = hit_ratio_10
except KeyboardInterrupt:
logger.info('*' * 89)
logger.info('Early Stopping!')