class BPRMF(Algo, BPRMFOption, Evaluable, Serializable, Optimizable,
TensorboardExtention):
"""Python implementation for C-BPRMF.
"""
def __init__(self, opt_path=None, *args, **kwargs):
Algo.__init__(self, *args, **kwargs)
BPRMFOption.__init__(self, *args, **kwargs)
Evaluable.__init__(self, *args, **kwargs)
Serializable.__init__(self, *args, **kwargs)
Optimizable.__init__(self, *args, **kwargs)
if opt_path is None:
opt_path = BPRMFOption().get_default_option()
self.logger = log.get_logger('BPRMF')
self.opt, self.opt_path = self.get_option(opt_path)
self.obj = CyBPRMF()
assert self.obj.init(bytes(self.opt_path, 'utf-8')),\
'cannot parse option file: %s' % opt_path
self.data = None
data = kwargs.get('data')
data_opt = self.opt.get('data_opt')
data_opt = kwargs.get('data_opt', data_opt)
if data_opt:
self.data = buffalo.data.load(data_opt)
self.data.create()
elif isinstance(data, Data):
self.data = data
self.logger.info('BPRMF(%s)' % json.dumps(self.opt, indent=2))
if self.data:
self.logger.info(self.data.show_info())
assert self.data.data_type in ['matrix']
@staticmethod
def new(path, data_fields=[]):
return BPRMF.instantiate(BPRMFOption, path, data_fields)
def set_data(self, data):
assert isinstance(data, aux.data.Data), 'Wrong instance: {}'.format(
type(data))
self.data = data
def normalize(self, group='item'):
if group == 'item' and not self.opt._nrz_Q:
self.Q = self._normalize(self.Q)
self.opt._nrz_Q = True
elif group == 'user' and not self.opt._nrz_P:
self.P = self._normalize(self.P)
self.opt._nrz_P = True
def initialize(self):
super().initialize()
assert self.data, 'Data is not setted'
if self.opt.random_seed:
np.random.seed(self.opt.random_seed)
self.buf = BufferedDataMatrix()
self.buf.initialize(self.data)
self.init_factors()
self.prepare_sampling()
def init_factors(self):
header = self.data.get_header()
for attr_name in ['P', 'Q', 'Qb']:
setattr(self, attr_name, None)
self.P = np.abs(np.random.normal(scale=1.0 / (self.opt.d**2),
size=(header['num_users'],
self.opt.d)).astype("float32"),
order='C')
self.Q = np.abs(np.random.normal(scale=1.0 / (self.opt.d**2),
size=(header['num_items'],
self.opt.d)).astype("float32"),
order='C')
self.Qb = np.abs(np.random.normal(scale=1.0 / (self.opt.d**2),
size=(header['num_items'],
1)).astype("float32"),
order='C')
self.obj.initialize_model(self.P, self.Q, self.Qb, header['num_nnz'])
def prepare_sampling(self):
self.logger.info('Preparing sampling ...')
header = self.data.get_header()
self.sampling_table_ = np.zeros(header['num_items'], dtype=np.int64)
if self.opt.sampling_power > 0.0:
for sz in self.buf.fetch_batch():
*_, keys, vals = self.buf.get()
for i in range(sz):
self.sampling_table_[keys[i]] += 1
self.sampling_table_ **= int(self.opt.sampling_power)
for i in range(1, header['num_items']):
self.sampling_table_[i] += self.sampling_table_[i - 1]
self.obj.set_cumulative_table(self.sampling_table_,
header['num_items'])
def _get_topk_recommendation(self, rows, topk, pool=None):
p = self.P[rows]
topks = super()._get_topk_recommendation(p, self.Q, pool, topk,
self.opt.num_workers)
return zip(rows, topks)
def _get_most_similar_item(self, col, topk, pool):
return super()._get_most_similar_item(col, topk, self.Q,
self.opt._nrz_Q, pool)
def get_scores(self, row_col_pairs):
rets = {(r, c): self.P[r].dot(self.Q[c]) + self.Qb[c][0]
for r, c in row_col_pairs}
return rets
def sampling_loss_samples(self):
users, positives, negatives = [], [], []
if self.opt.compute_loss_on_training:
self.logger.info('Sampling loss samples...')
header = self.data.get_header()
num_loss_samples = int(header['num_users']**0.5)
_users = np.random.choice(range(self.P.shape[0]),
size=num_loss_samples,
replace=False)
users = []
positives, negatives = [], []
for u in _users:
keys, *_ = self.data.get(u)
if len(keys) == 0:
continue
seen = set(keys)
negs = np.random.choice(range(self.Q.shape[0]),
size=len(seen) + 1,
replace=False)
negs = [n for n in negs if n not in seen]
users.append(u)
positives.append(keys[0])
negatives.append(negs[0])
self.logger.info('Generated %s loss samples.' % len(users))
self._sub_samples = [
np.array(users, dtype=np.int32, order='F'),
np.array(positives, dtype=np.int32, order='F'),
np.array(negatives, dtype=np.int32, order='F')
]
def _get_feature(self, index, group='item'):
if group == 'item':
return self.Q[index]
elif group == 'user':
return self.P[index]
return None
def _iterate(self):
header = self.data.get_header()
end = header['num_users']
update_t, feed_t, updated = 0, 0, 0
self.buf.set_group('rowwise')
with log.pbar(log.DEBUG, total=header['num_nnz'],
mininterval=30) as pbar:
start_t = time.time()
for sz in self.buf.fetch_batch():
updated += sz
feed_t += time.time() - start_t
start_x, next_x, indptr, keys, vals = self.buf.get()
start_t = time.time()
self.obj.add_jobs(start_x, next_x, indptr, keys)
update_t += time.time() - start_t
pbar.update(sz)
pbar.refresh()
self.obj.update_parameters()
self.logger.debug(
f'updated processed({updated}) elapsed(data feed: {feed_t:0.3f} update: {update_t:0.3f}'
)
def compute_loss(self):
return self.obj.compute_loss(self._sub_samples[0],
self._sub_samples[1],
self._sub_samples[2])
def train(self):
rmse, self.validation_result = None, {}
self.prepare_evaluation()
self.initialize_tensorboard(self.opt.num_iters)
self.sampling_loss_samples()
self.obj.launch_workers()
best_loss = 987654321.0
for i in range(self.opt.num_iters):
start_t = time.time()
self._iterate()
self.obj.wait_until_done()
loss = self.compute_loss(
) if self.opt.compute_loss_on_training else 0.0
train_t = time.time() - start_t
metrics = {'train_loss': loss}
if self.opt.validation and self.opt.evaluation_on_learning and self.periodical(
self.opt.evaluation_period, i):
start_t = time.time()
self.validation_result = self.get_validation_results()
vali_t = time.time() - start_t
val_str = ' '.join([
f'{k}:{v:0.5f}' for k, v in self.validation_result.items()
])
self.logger.info(f'Validation: {val_str} Elased {vali_t:0.3f}')
metrics.update({
'val_%s' % k: v
for k, v in self.validation_result.items()
})
self.logger.info('Iteration %s: PR-Loss %.3f Elapsed %.3f secs' %
(i + 1, loss, time.time() - start_t))
self.update_tensorboard_data(metrics)
best_loss = self.save_best_only(loss, best_loss, i)
if self.early_stopping(loss):
break
loss = self.obj.join()
ret = {'train_loss': loss}
ret.update(
{'val_%s' % k: v
for k, v in self.validation_result.items()})
self.finalize_tensorboard()
return ret
def _optimize(self, params):
# TODO: implement
self._optimize_params = params
for name, value in params.items():
assert name in self.opt, 'Unexepcted parameter: {}'.format(name)
setattr(self.opt, name, value)
with open(self._temporary_opt_file, 'w') as fout:
json.dump(self.opt, fout, indent=2)
assert self.obj.init(bytes(self._temporary_opt_file, 'utf-8')),\
'cannot parse option file: %s' % self._temporary_opt_file
self.logger.info(params)
self.init_factors()
loss = self.train()
loss['eval_time'] = time.time()
loss['loss'] = loss.get(self.opt.optimize.loss)
# TODO: deal with failture of training
loss['status'] = HOPT_STATUS_OK
self._optimize_loss = loss
return loss
def _get_data(self):
data = super()._get_data()
data.extend([('opt', self.opt), ('Q', self.Q), ('Qb', self.Qb),
('P', self.P)])
return data
def get_evaluation_metrics(self):
return [
'val_rmse', 'val_ndcg', 'val_map', 'val_accuracy', 'val_error',
'train_loss'
]
class WARP(Algo, WARPOption, Evaluable, Serializable, Optimizable,
TensorboardExtention):
"""Python implementation for C-WARP.
"""
def __init__(self, opt_path=None, *args, **kwargs):
Algo.__init__(self, *args, **kwargs)
WARPOption.__init__(self, *args, **kwargs)
Evaluable.__init__(self, *args, **kwargs)
Serializable.__init__(self, *args, **kwargs)
Optimizable.__init__(self, *args, **kwargs)
if opt_path is None:
opt_path = WARPOption().get_default_option()
self.logger = log.get_logger('WARP')
self.opt, self.opt_path = self.get_option(opt_path)
# TODO:GPU Implementation
if self.opt.accelerator is True:
raise NotImplementedError(
"GPU version WARP is not implemented yet")
self.obj = CyWARP()
assert self.obj.init(bytes(self.opt_path, 'utf-8')),\
'cannot parse option file: %s' % opt_path
self.data = None
data = kwargs.get('data')
data_opt = self.opt.get('data_opt')
data_opt = kwargs.get('data_opt', data_opt)
if data_opt:
self.data = buffalo.data.load(data_opt)
self.data.create()
elif isinstance(data, Data):
self.data = data
self.logger.info('WARP(%s)' % json.dumps(self.opt, indent=2))
if self.data:
self.logger.info(self.data.show_info())
assert self.data.data_type in ['matrix']
@staticmethod
def new(path, data_fields=[]):
return WARP.instantiate(WARPOption, path, data_fields)
def set_data(self, data):
assert isinstance(data, aux.data.Data), 'Wrong instance: {}'.format(
type(data))
self.data = data
def normalize(self, group='item'):
if group == 'item' and not self.opt._nrz_Q:
self.Q = self._normalize(self.Q)
self.opt._nrz_Q = True
elif group == 'user' and not self.opt._nrz_P:
self.P = self._normalize(self.P)
self.opt._nrz_P = True
def initialize(self):
super().initialize()
assert self.data, 'Data is not setted'
self.buf = BufferedDataMatrix()
self.buf.initialize(self.data)
self.init_factors()
def init_factors(self):
header = self.data.get_header()
self.num_nnz = header['num_nnz']
for attr_name in ['P', 'Q', 'Qb']:
setattr(self, attr_name, None)
self.P = np.random.normal(scale=1.0 / (self.opt.d**2),
size=(header['num_users'],
self.opt.d)).astype("float32")
self.Q = np.random.normal(scale=1.0 / (self.opt.d**2),
size=(header['num_items'],
self.opt.d)).astype("float32")
self.Qb = np.random.normal(scale=1.0 / (self.opt.d**2),
size=(header['num_items'],
1)).astype("float32")
if not self.opt.use_bias:
self.Qb *= 0
self.obj.initialize_model(self.P, self.Q, self.Qb, self.num_nnz)
def _get_topk_recommendation(self, rows, topk, pool=None):
p = self.P[rows]
Qb = self.Qb if self.opt.use_bias else None
topks = super()._get_topk_recommendation(
p,
self.Q,
pb=None,
Qb=Qb,
pool=pool,
topk=topk,
num_workers=self.opt.num_workers)
return zip(rows, topks)
def _get_most_similar_item(self, col, topk, pool):
return super()._get_most_similar_item(col, topk, self.Q,
self.opt._nrz_Q, pool)
def get_scores(self, row_col_pairs):
rets = {(r, c): self.P[r].dot(self.Q[c]) + self.Qb[c][0]
for r, c in row_col_pairs}
return rets
def _get_scores(self, row, col):
scores = (self.P[row] * self.Q[col]).sum(axis=1) + self.Qb[col][0]
return scores
def sampling_loss_samples(self):
users, positives, negatives = [], [], []
if self.opt.compute_loss_on_training:
self.logger.info('Sampling loss samples...')
header = self.data.get_header()
num_loss_samples = int(header['num_users']**0.5)
_users = np.random.choice(range(self.P.shape[0]),
size=num_loss_samples,
replace=False)
users = []
positives, negatives = [], []
for u in _users:
keys, *_ = self.data.get(u)
if len(keys) == 0:
continue
seen = set(keys)
negs = np.random.choice(range(self.Q.shape[0]),
size=len(seen) + 1,
replace=False)
negs = [n for n in negs if n not in seen]
users.append(u)
positives.append(keys[0])
negatives.append(negs[0])
self.logger.info('Generated %s loss samples.' % len(users))
self._sub_samples = [
np.array(users, dtype=np.int32, order='F'),
np.array(positives, dtype=np.int32, order='F'),
np.array(negatives, dtype=np.int32, order='F')
]
def _get_feature(self, index, group='item'):
if group == 'item':
return self.Q[index]
elif group == 'user':
return self.P[index]
return None
def _iterate(self):
header = self.data.get_header()
# end = header['num_users']
update_t, feed_t, updated = 0, 0, 0
self.buf.set_group('rowwise')
with log.ProgressBar(log.DEBUG,
total=header['num_nnz'],
mininterval=30) as pbar:
start_t = time.time()
for sz in self.buf.fetch_batch():
updated += sz
feed_t += time.time() - start_t
start_x, next_x, indptr, keys, _ = self.buf.get()
start_t = time.time()
self.obj.add_jobs(start_x, next_x, indptr, keys)
update_t += time.time() - start_t
pbar.update(sz)
pbar.refresh()
self.obj.update_parameters()
self.logger.debug(
f'updated processed({updated}) elapsed(data feed: {feed_t:0.3f} update: {update_t:0.3f}'
)
def compute_loss(self):
return self.obj.compute_loss(self._sub_samples[0],
self._sub_samples[1],
self._sub_samples[2])
def _prepare_train(self):
if self.opt.accelerator:
vdim = self.obj.get_vdim()
for attr in ["P", "Q"]:
F = getattr(self, attr)
if F.shape[1] < vdim:
_F = np.empty(shape=(F.shape[0], vdim), dtype=np.float32)
_F[:, :F.shape[1]] = F
_F[:, self.opt.d:] = 0.0
setattr(self, attr, _F)
indptr, _, batch_size = self.buf.get_indptrs()
self.obj.set_placeholder(indptr, batch_size)
self.obj.initialize_model(self.P, self.Q, self.Qb, self.num_nnz,
True)
else:
self.obj.launch_workers()
def _finalize_train(self):
if self.opt.accelerator:
self.P = self.P[:, :self.opt.d]
self.Q = self.Q[:, :self.opt.d]
return 0.0
else:
return self.obj.join()
def train(self):
self.validation_result = {}
self.initialize_tensorboard(self.opt.num_iters)
self.sampling_loss_samples()
best_loss = 987654321.0
# initialize placeholder in case of running accelerator otherwise launch workers
self._prepare_train()
for i in range(self.opt.num_iters):
start_t = time.time()
self._iterate()
self.obj.wait_until_done()
loss = self.compute_loss(
) if self.opt.compute_loss_on_training else 0.0
metrics = {'train_loss': loss}
if self.opt.validation and \
self.opt.evaluation_on_learning and \
self.periodical(self.opt.evaluation_period, i):
start_t = time.time()
self.validation_result = self.get_validation_results()
vali_t = time.time() - start_t
val_str = ' '.join([
f'{k}:{v:0.5f}' for k, v in self.validation_result.items()
])
self.logger.info(f'Validation: {val_str} Elased {vali_t:0.3f}')
metrics.update({
'val_%s' % k: v
for k, v in self.validation_result.items()
})
self.logger.info('Iteration %s: PR-Loss %.3f Elapsed %.3f secs' %
(i + 1, loss, time.time() - start_t))
self.update_tensorboard_data(metrics)
best_loss = self.save_best_only(loss, best_loss, i)
if self.early_stopping(loss):
break
# reshape factor if using accelerator else join workers
ret = {'train_loss': self._finalize_train()}
ret.update(
{'val_%s' % k: v
for k, v in self.validation_result.items()})
self.finalize_tensorboard()
return ret
def _optimize(self, params):
# TODO: implement
self._optimize_params = params
for name, value in params.items():
assert name in self.opt, 'Unexepcted parameter: {}'.format(name)
if isinstance(value, np.generic):
setattr(self.opt, name, value.item())
else:
setattr(self.opt, name, value)
with open(self._temporary_opt_file, 'w') as fout:
json.dump(self.opt, fout, indent=2)
assert self.obj.init(bytes(self._temporary_opt_file, 'utf-8')),\
'cannot parse option file: %s' % self._temporary_opt_file
self.logger.info(params)
self.init_factors()
loss = self.train()
loss['loss'] = loss.get(self.opt.optimize.loss)
# TODO: deal with failture of training
loss['status'] = HOPT_STATUS_OK
self._optimize_loss = loss
return loss
def _get_data(self):
data = super()._get_data()
data.extend([('opt', self.opt), ('Q', self.Q), ('Qb', self.Qb),
('P', self.P)])
return data
def get_evaluation_metrics(self):
return [
'val_rmse', 'val_ndcg', 'val_map', 'val_accuracy', 'val_error',
'train_loss'
]