def predict_level(self, level, test_x, k, labels_num):
data_cnf, model_cnf = self.data_cnf, self.model_cnf
model = self.models.get(level, None)
if level == 0:
logger.info(F'Predicting Level-{level}, Top: {k}')
if model is None:
model = Model(AttentionRNN, labels_num=labels_num, model_path=F'{self.model_path}-Level-{level}',
emb_init=self.emb_init, **data_cnf['model'], **model_cnf['model'])
test_loader = DataLoader(MultiLabelDataset(test_x), model_cnf['predict']['batch_size'],
num_workers=4)
return model.predict(test_loader, k=k)
else:
if level == self.level - 1:
groups = np.load(F'{self.groups_path}-Level-{level-1}.npy')
else:
groups = self.get_inter_groups(labels_num)
group_scores, group_labels = self.predict_level(level - 1, test_x, self.top, len(groups))
torch.cuda.empty_cache()
logger.info(F'Predicting Level-{level}, Top: {k}')
if model is None:
model = XMLModel(network=FastAttentionRNN, labels_num=labels_num,
model_path=F'{self.model_path}-Level-{level}',
emb_init=self.emb_init, **data_cnf['model'], **model_cnf['model'])
test_loader = DataLoader(XMLDataset(test_x, labels_num=labels_num,
groups=groups, group_labels=group_labels, group_scores=group_scores),
model_cnf['predict']['batch_size'], num_workers=4)
return model.predict(test_loader, k=k)
def default_eval(
data_cnf, model_cnf, data_name, model_name, model_path, emb_init,
tree_id, output_suffix, dry_run,
):
logger.info('Loading Test Set')
mlb = get_mlb(data_cnf['labels_binarizer'])
labels_num = len(mlb.classes_)
test_x, _ = get_data(data_cnf['test']['texts'], None)
logger.info(F'Size of Test Set: {len(test_x):,}')
logger.info('Predicting')
model_cnf['model'].pop('load_model', None)
if 'cluster' not in model_cnf:
test_loader = DataLoader(
MultiLabelDataset(test_x),
model_cnf['predict']['batch_size'],
num_workers=4)
if 'loss' in model_cnf:
gamma = model_cnf['loss'].get('gamma', 1.0)
loss_name = model_cnf['loss']['name']
else:
gamma = None
loss_name = 'bce'
model = Model(
network=AttentionRNN, labels_num=labels_num,
model_path=model_path, emb_init=emb_init,
load_model=True, loss_name=loss_name, gamma=gamma,
**data_cnf['model'], **model_cnf['model'])
scores, labels = model.predict(test_loader, k=model_cnf['predict'].get('k', 100))
labels = mlb.classes_[labels]
else:
model = FastAttentionXML(labels_num, data_cnf, model_cnf,
tree_id, output_suffix)
scores, labels = model.predict(test_x, model_cnf['predict'].get('k', 100))
labels = mlb.classes_[labels]
logger.info('Finish Predicting')
score_path, label_path = output_res(data_cnf['output']['res'],
f'{model_name}-{data_name}{tree_id}',
scores, labels, output_suffix)
log_results(score_path, label_path, dry_run)
def main(data_cnf, model_cnf, mode, reg):
yaml = YAML(typ='safe')
data_cnf, model_cnf = yaml.load(Path(data_cnf)), yaml.load(Path(model_cnf))
model, model_name, data_name = None, model_cnf['name'], data_cnf['name']
model_path = os.path.join(model_cnf['path'], F'{model_name}-{data_name}')
emb_init = get_word_emb(data_cnf['embedding']['emb_init'])
logger.info(F'Model Name: {model_name}')
if mode is None or mode == 'train':
logger.info('Loading Training and Validation Set')
train_x, train_labels = get_data(data_cnf['train']['texts'],
data_cnf['train']['labels'])
if 'size' in data_cnf['valid']:
random_state = data_cnf['valid'].get('random_state', 1240)
train_x, valid_x, train_labels, valid_labels = train_test_split(
train_x,
train_labels,
test_size=data_cnf['valid']['size'],
random_state=random_state)
else:
valid_x, valid_labels = get_data(data_cnf['valid']['texts'],
data_cnf['valid']['labels'])
mlb = get_mlb(data_cnf['labels_binarizer'],
np.hstack((train_labels, valid_labels)))
train_y, valid_y = mlb.transform(train_labels), mlb.transform(
valid_labels)
labels_num = len(mlb.classes_)
logger.info(F'Number of Labels: {labels_num}')
logger.info(F'Size of Training Set: {len(train_x)}')
logger.info(F'Size of Validation Set: {len(valid_x)}')
edges = set()
if reg:
classes = mlb.classes_.tolist()
with open(data_cnf['hierarchy']) as fin:
for line in fin:
data = line.strip().split()
p = data[0]
if p not in classes:
continue
p_id = classes.index(p)
for c in data[1:]:
if c not in classes:
continue
c_id = classes.index(c)
edges.add((p_id, c_id))
logger.info(F'Number of Edges: {len(edges)}')
logger.info('Training')
train_loader = DataLoader(MultiLabelDataset(train_x, train_y),
model_cnf['train']['batch_size'],
shuffle=True,
num_workers=4)
valid_loader = DataLoader(MultiLabelDataset(valid_x,
valid_y,
training=True),
model_cnf['valid']['batch_size'],
num_workers=4)
model = Model(network=MATCH,
labels_num=labels_num,
model_path=model_path,
emb_init=emb_init,
mode='train',
reg=reg,
hierarchy=edges,
**data_cnf['model'],
**model_cnf['model'])
opt_params = {
'lr': model_cnf['train']['learning_rate'],
'betas':
(model_cnf['train']['beta1'], model_cnf['train']['beta2']),
'weight_decay': model_cnf['train']['weight_decay']
}
model.train(train_loader,
valid_loader,
opt_params=opt_params,
**model_cnf['train']) # CHANGE: inserted opt_params
logger.info('Finish Training')
if mode is None or mode == 'eval':
logger.info('Loading Test Set')
mlb = get_mlb(data_cnf['labels_binarizer'])
labels_num = len(mlb.classes_)
test_x, _ = get_data(data_cnf['test']['texts'], None)
logger.info(F'Size of Test Set: {len(test_x)}')
logger.info('Predicting')
test_loader = DataLoader(MultiLabelDataset(test_x),
model_cnf['predict']['batch_size'],
num_workers=4)
if model is None:
model = Model(network=MATCH,
labels_num=labels_num,
model_path=model_path,
emb_init=emb_init,
mode='eval',
**data_cnf['model'],
**model_cnf['model'])
scores, labels = model.predict(test_loader,
k=model_cnf['predict'].get('k', 100))
logger.info('Finish Predicting')
labels = mlb.classes_[labels]
output_res(data_cnf['output']['res'], F'{model_name}-{data_name}',
scores, labels)
def train_level(self, level, train_x, train_y, valid_x, valid_y):
model_cnf, data_cnf = self.model_cnf, self.data_cnf
if level == 0:
while not os.path.exists(F'{self.groups_path}-Level-{level}.npy'):
time.sleep(30)
groups = np.load(F'{self.groups_path}-Level-{level}.npy')
train_y, valid_y = self.get_mapping_y(groups, self.labels_num,
train_y, valid_y)
labels_num = len(groups)
train_loader = DataLoader(MultiLabelDataset(train_x, train_y),
model_cnf['train'][level]['batch_size'],
num_workers=4,
shuffle=True)
valid_loader = DataLoader(MultiLabelDataset(valid_x,
valid_y,
training=False),
model_cnf['valid']['batch_size'],
num_workers=4)
model = Model(AttentionRNN,
labels_num=labels_num,
model_path=F'{self.model_path}-Level-{level}',
emb_init=self.emb_init,
**data_cnf['model'],
**model_cnf['model'])
if not os.path.exists(model.model_path):
logger.info(
F'Training Level-{level}, Number of Labels: {labels_num}')
model.train(train_loader, valid_loader,
**model_cnf['train'][level])
model.optimizer = None
logger.info(F'Finish Training Level-{level}')
self.models[level] = model
logger.info(F'Generating Candidates for Level-{level+1}, '
F'Number of Labels: {labels_num}, Top: {self.top}')
train_loader = DataLoader(MultiLabelDataset(train_x),
model_cnf['valid']['batch_size'],
num_workers=4)
return train_y, model.predict(
train_loader, k=self.top), model.predict(valid_loader,
k=self.top)
else:
train_group_y, train_group, valid_group = self.train_level(
level - 1, train_x, train_y, valid_x, valid_y)
torch.cuda.empty_cache()
logger.info('Getting Candidates')
_, group_labels = train_group
group_candidates = np.empty((len(train_x), self.top), dtype=np.int)
for i, labels in tqdm(enumerate(group_labels),
leave=False,
desc='Parents'):
ys, ye = train_group_y.indptr[i], train_group_y.indptr[i + 1]
positive = set(train_group_y.indices[ys:ye])
if self.top >= len(positive):
candidates = positive
for la in labels:
if len(candidates) == self.top:
break
if la not in candidates:
candidates.add(la)
else:
candidates = set()
for la in labels:
if la in positive:
candidates.add(la)
if len(candidates) == self.top:
break
if len(candidates) < self.top:
candidates = (list(candidates) +
list(positive - candidates))[:self.top]
group_candidates[i] = np.asarray(list(candidates))
if level < self.level - 1:
while not os.path.exists(
F'{self.groups_path}-Level-{level}.npy'):
time.sleep(30)
groups = np.load(F'{self.groups_path}-Level-{level}.npy')
train_y, valid_y = self.get_mapping_y(groups, self.labels_num,
train_y, valid_y)
labels_num, last_groups = len(groups), self.get_inter_groups(
len(groups))
else:
groups, labels_num = None, train_y.shape[1]
last_groups = np.load(
F'{self.groups_path}-Level-{level-1}.npy')
train_loader = DataLoader(XMLDataset(
train_x,
train_y,
labels_num=labels_num,
groups=last_groups,
group_labels=group_candidates),
model_cnf['train'][level]['batch_size'],
num_workers=4,
shuffle=True)
group_scores, group_labels = valid_group
valid_loader = DataLoader(XMLDataset(valid_x,
valid_y,
training=False,
labels_num=labels_num,
groups=last_groups,
group_labels=group_labels,
group_scores=group_scores),
model_cnf['valid']['batch_size'],
num_workers=4)
model = XMLModel(network=FastAttentionRNN,
labels_num=labels_num,
emb_init=self.emb_init,
model_path=F'{self.model_path}-Level-{level}',
**data_cnf['model'],
**model_cnf['model'])
if not os.path.exists(model.model_path):
logger.info(
F'Loading parameters of Level-{level} from Level-{level-1}'
)
last_model = self.get_last_models(level - 1)
model.network.module.emb.load_state_dict(
last_model.module.emb.state_dict())
model.network.module.lstm.load_state_dict(
last_model.module.lstm.state_dict())
model.network.module.linear.load_state_dict(
last_model.module.linear.state_dict())
logger.info(
F'Training Level-{level}, '
F'Number of Labels: {labels_num}, '
F'Candidates Number: {train_loader.dataset.candidates_num}'
)
model.train(train_loader, valid_loader,
**model_cnf['train'][level])
model.optimizer = model.state = None
logger.info(F'Finish Training Level-{level}')
self.models[level] = model
if level == self.level - 1:
return
logger.info(F'Generating Candidates for Level-{level+1}, '
F'Number of Labels: {labels_num}, Top: {self.top}')
group_scores, group_labels = train_group
train_loader = DataLoader(XMLDataset(train_x,
labels_num=labels_num,
groups=last_groups,
group_labels=group_labels,
group_scores=group_scores),
model_cnf['valid']['batch_size'],
num_workers=4)
return train_y, model.predict(
train_loader, k=self.top), model.predict(valid_loader,
k=self.top)
def main(data_cnf, model_cnf, mode, tree_id):
tree_id = F'-Tree-{tree_id}' if tree_id is not None else ''
yaml = YAML(typ='safe')
data_cnf, model_cnf = yaml.load(Path(data_cnf)), yaml.load(Path(model_cnf))
model, model_name, data_name = None, model_cnf['name'], data_cnf['name']
model_path = os.path.join(model_cnf['path'], F'{model_name}-{data_name}{tree_id}')
emb_init = get_word_emb(data_cnf['embedding']['emb_init'])
logger.info(F'Model Name: {model_name}')
if mode is None or mode == 'train':
logger.info('Loading Training and Validation Set')
train_x, train_labels = get_data(data_cnf['train']['texts'], data_cnf['train']['labels'])
if 'size' in data_cnf['valid']:
random_state = data_cnf['valid'].get('random_state', 1240)
train_x, valid_x, train_labels, valid_labels = train_test_split(train_x, train_labels,
test_size=data_cnf['valid']['size'],
random_state=random_state)
else:
valid_x, valid_labels = get_data(data_cnf['valid']['texts'], data_cnf['valid']['labels'])
mlb = get_mlb(data_cnf['labels_binarizer'], np.hstack((train_labels, valid_labels)))
train_y, valid_y = mlb.transform(train_labels), mlb.transform(valid_labels)
labels_num = len(mlb.classes_)
logger.info(F'Number of Labels: {labels_num}')
logger.info(F'Size of Training Set: {len(train_x)}')
logger.info(F'Size of Validation Set: {len(valid_x)}')
logger.info('Training')
if 'cluster' not in model_cnf:
train_loader = DataLoader(MultiLabelDataset(train_x, train_y),
model_cnf['train']['batch_size'], shuffle=True, num_workers=4)
valid_loader = DataLoader(MultiLabelDataset(valid_x, valid_y, training=False),
model_cnf['valid']['batch_size'], num_workers=4)
model = Model(network=AttentionRNN, labels_num=labels_num, model_path=model_path, emb_init=emb_init,
**data_cnf['model'], **model_cnf['model'])
model.train(train_loader, valid_loader, **model_cnf['train'])
else:
model = FastAttentionXML(labels_num, data_cnf, model_cnf, tree_id)
model.train(train_x, train_y, valid_x, valid_y, mlb)
logger.info('Finish Training')
if mode is None or mode == 'eval':
logger.info('Loading Test Set')
mlb = get_mlb(data_cnf['labels_binarizer'])
labels_num = len(mlb.classes_)
test_x, _ = get_data(data_cnf['test']['texts'], None)
logger.info(F'Size of Test Set: {len(test_x)}')
logger.info('Predicting')
if 'cluster' not in model_cnf:
test_loader = DataLoader(MultiLabelDataset(test_x), model_cnf['predict']['batch_size'],
num_workers=4)
if model is None:
model = Model(network=AttentionRNN, labels_num=labels_num, model_path=model_path, emb_init=emb_init,
**data_cnf['model'], **model_cnf['model'])
scores, labels = model.predict(test_loader, k=model_cnf['predict'].get('k', 100))
else:
if model is None:
model = FastAttentionXML(labels_num, data_cnf, model_cnf, tree_id)
scores, labels = model.predict(test_x)
logger.info('Finish Predicting')
labels = mlb.classes_[labels]
output_res(data_cnf['output']['res'], F'{model_name}-{data_name}{tree_id}', scores, labels)
def spectral_clustering_train(
data_cnf, data_cnf_path, model_cnf, model_cnf_path,
emb_init, model_path, tree_id, output_suffix, dry_run,
):
train_xs = []
valid_xs = []
train_labels_list = []
valid_labels_list = []
train_ys = []
valid_ys = []
mlb_list = []
indices_list = []
n_clusters = model_cnf['spectral_clustering']['num_clusters']
n_components = model_cnf['spectral_clustering']['n_components']
alg = model_cnf['spectral_clustering']['alg']
size_min = model_cnf['spectral_clustering']['size_min']
size_max = model_cnf['spectral_clustering']['size_max']
train_x, train_labels = load_dataset(data_cnf)
if 'cluster' not in model_cnf:
mlb = get_mlb(data_cnf['labels_binarizer'], train_labels)
train_y = mlb.transform(train_labels)
logger.info('Build label adjacency matrix')
adj = train_y.T @ train_y
adj.setdiag(0)
adj.eliminate_zeros()
logger.info(f"Sparsity: {adj.count_nonzero() / adj.shape[0] ** 2}")
clustering = MySpectralClustering(n_clusters=n_clusters, affinity='precomputed',
n_components=n_components, n_init=1,
size_min=size_min,
size_max=size_max,
assign_labels=alg, n_jobs=-1)
logger.info('Start Spectral Clustering')
clustering.fit(adj)
logger.info('Finish Spectral Clustering')
groups = [[] for _ in range(n_clusters)]
for i, group in enumerate(clustering.labels_):
groups[group].append(i)
splitted_labels = []
for indices in groups:
splitted_labels.append(mlb.classes_[indices])
for labels in splitted_labels:
indices = get_splitted_samples(labels, train_labels)
indices_list.append(indices)
train_xs.append(train_x[indices])
train_labels_list.append(train_labels[indices])
if 'size' in data_cnf['valid']:
for i, (train_x, train_labels) in enumerate(zip(train_xs, train_labels_list)):
valid_size = data_cnf['valid']['size']
if len(train_x) * 0.8 > len(train_x) - valid_size:
valid_size = 0.2
train_x, valid_x, train_labels, valid_labels = train_test_split(
train_x, train_labels, test_size=valid_size,
)
train_xs[i] = train_x
train_labels_list[i] = train_labels
valid_xs.append(valid_x)
valid_labels_list.append(valid_labels)
else:
raise Exception("Setting valid set explicitly is not "
"supported spectral clustering mode.")
labels_binarizer_path = data_cnf['labels_binarizer']
suffix = output_suffix.upper().replace('-', '_')
for i, labels in enumerate(splitted_labels):
filename = f"{labels_binarizer_path}_{suffix}_{i}"
mlb_tree = get_mlb(filename, labels[None, ...], force=True)
mlb_list.append(mlb_tree)
logger.info(f"Number of labels of cluster {i}: {len(labels):,}")
logger.info(f"Number of Training Set of cluster {i}: {len(train_xs[i]):,}")
logger.info(f"Number of Validation Set of cluster {i}: {len(valid_xs[i]):,}")
with redirect_stderr(None):
train_y = mlb_tree.transform(train_labels_list[i])
valid_y = mlb_tree.transform(valid_labels_list[i])
train_ys.append(train_y)
valid_ys.append(valid_y)
else:
if 'size' in data_cnf['valid']:
train_x, valid_x, train_labels, valid_labels = train_test_split(
train_x, train_labels, test_size=data_cnf['valid']['size'],
)
else:
valid_x, valid_labels = get_data(data_cnf['valid']['texts'], data_cnf['valid']['labels'])
mlb = get_mlb(data_cnf['labels_binarizer'], np.hstack((
train_labels, valid_labels,
)))
train_y, valid_y = mlb.transform(train_labels), mlb.transform(valid_labels)
logger.info('Training')
if 'cluster' not in model_cnf:
for i, (train_x, train_y, valid_x, valid_y) in enumerate(zip(
train_xs, train_ys, valid_xs, valid_ys,
)):
train_loader = DataLoader(
MultiLabelDataset(train_x, train_y),
model_cnf['train']['batch_size'], shuffle=True, num_workers=4)
valid_loader = DataLoader(
MultiLabelDataset(valid_x, valid_y, training=False),
model_cnf['valid']['batch_size'], num_workers=4)
model = Model(
network=AttentionRNN, labels_num=len(mlb_list[i].classes_),
model_path=f'{model_path}-{i}', emb_init=emb_init,
**data_cnf['model'], **model_cnf['model'])
if not dry_run:
logger.info(f"Start Training Cluster {i}")
model.train(train_loader, valid_loader, **model_cnf['train'])
logger.info(f"Finish Training Cluster {i}")
else:
model.save_model()
else:
model = FastAttentionXML(
len(mlb.classes_), data_cnf, model_cnf, tree_id, output_suffix,
)
if not dry_run:
model.train(train_x, train_y, valid_x, valid_y, mlb)
log_config(data_cnf_path, model_cnf_path, dry_run)
def spectral_clustering_eval(
data_cnf, model_cnf, data_name, model_name, model_path, emb_init,
tree_id, output_suffix, dry_run,
):
mlb_list = []
n_clusters = model_cnf['spectral_clustering']['num_clusters']
labels_binarizer_path = data_cnf['labels_binarizer']
scores_list = []
labels_list = []
logger.info('Loading Test Set')
test_x, _ = get_data(data_cnf['test']['texts'], None)
logger.info(F'Size of Test Set: {len(test_x):,}')
logger.info('Predicting')
if 'cluster' not in model_cnf:
suffix = output_suffix.upper().replace('-', '_')
for i in range(n_clusters):
filename = f"{labels_binarizer_path}_{suffix}_{i}"
mlb_tree = get_mlb(filename)
mlb_list.append(mlb_tree)
test_loader = DataLoader(
MultiLabelDataset(test_x),
model_cnf['predict']['batch_size'],
num_workers=4)
for i, mlb in enumerate(mlb_list):
logger.info(f"Predicting Cluster {i}")
labels_num = len(mlb.classes_)
k = model_cnf['predict'].get('k', 100) // n_clusters
model = Model(
network=AttentionRNN, labels_num=labels_num,
model_path=f'{model_path}-{i}', emb_init=emb_init,
load_model=True,
**data_cnf['model'], **model_cnf['model'])
scores, labels = model.predict(test_loader, k=k)
scores_list.append(scores)
labels_list.append(mlb.classes_[labels])
logger.info(f"Finish Prediting Cluster {i}")
scores = np.hstack(scores_list)
labels = np.hstack(labels_list)
i = np.arange(len(scores))[:, None]
j = np.argsort(scores)[:, ::-1]
scores = scores[i, j]
labels = labels[i, j]
else:
mlb = get_mlb(data_cnf['labels_binarizer'])
model = FastAttentionXML(len(mlb.classes_), data_cnf, model_cnf,
tree_id, output_suffix)
scores, labels = model.predict(test_x, model_cnf['predict'].get('k', 100))
labels = mlb.classes_[labels]
logger.info('Finish Predicting')
score_path, label_path = output_res(data_cnf['output']['res'],
f'{model_name}-{data_name}{tree_id}',
scores, labels, output_suffix)
log_results(score_path, label_path, dry_run)
def default_train(
data_cnf, data_cnf_path, model_cnf, model_cnf_path,
emb_init, model_path, tree_id, output_suffix, dry_run,
):
train_x, train_labels = load_dataset(data_cnf)
if 'size' in data_cnf['valid']:
train_x, valid_x, train_labels, valid_labels = train_test_split(
train_x, train_labels, test_size=data_cnf['valid']['size'],
)
else:
valid_x, valid_labels = get_data(data_cnf['valid']['texts'], data_cnf['valid']['labels'])
mlb = get_mlb(data_cnf['labels_binarizer'], np.hstack((
train_labels, valid_labels,
)))
freq = mlb.transform(np.hstack([train_labels, valid_labels])).sum(axis=0).A1
train_y, valid_y = mlb.transform(train_labels), mlb.transform(valid_labels)
labels_num = len(mlb.classes_)
logger.info(F'Number of Labels: {labels_num}')
logger.info(F'Size of Training Set: {len(train_x):,}')
logger.info(F'Size of Validation Set: {len(valid_x):,}')
logger.info('Training')
if 'cluster' not in model_cnf:
if 'propensity' in data_cnf:
a = data_cnf['propensity']['a']
b = data_cnf['propensity']['b']
pos_weight = get_inv_propensity(train_y, a, b)
else:
pos_weight = None
train_loader = DataLoader(
MultiLabelDataset(train_x, train_y),
model_cnf['train']['batch_size'], shuffle=True, num_workers=4)
valid_loader = DataLoader(
MultiLabelDataset(valid_x, valid_y, training=False),
model_cnf['valid']['batch_size'], num_workers=4)
if 'loss' in model_cnf:
gamma = model_cnf['loss'].get('gamma', 2.0)
loss_name = model_cnf['loss']['name']
else:
gamma = None
loss_name = 'bce'
model = Model(
network=AttentionRNN, labels_num=labels_num, model_path=model_path,
emb_init=emb_init, pos_weight=pos_weight, loss_name=loss_name, gamma=gamma,
freq=freq, **data_cnf['model'], **model_cnf['model'])
if not dry_run:
model.train(train_loader, valid_loader, mlb=mlb, **model_cnf['train'])
else:
model.save_model()
else:
model = FastAttentionXML(labels_num, data_cnf, model_cnf, tree_id, output_suffix)
if not dry_run:
model.train(train_x, train_y, valid_x, valid_y, mlb)
log_config(data_cnf_path, model_cnf_path, dry_run)
def main(data_cnf, model_cnf, mode):
model_name = os.path.split(model_cnf)[1].split(".")[0]
yaml = YAML(typ='safe')
data_cnf, model_cnf = yaml.load(Path(data_cnf)), yaml.load(Path(model_cnf))
# 設定log檔案位置
logfile("./logs/logfile_{0}_cornet_{1}_cornet_dim_{2}.log".format(
model_name, model_cnf['model']['n_cornet_blocks'],
model_cnf['model']['cornet_dim']))
model, model_name, data_name = None, model_cnf['name'], data_cnf['name']
model_path = os.path.join(
model_cnf['path'],
F'{model_name}-{data_name}-{model_cnf["model"]["n_cornet_blocks"]}-{model_cnf["model"]["cornet_dim"]}'
)
emb_init = get_word_emb(data_cnf['embedding']['emb_init'])
logger.info(F'Model Name: {model_name}')
# summary(model_dict[model_name])
if mode is None or mode == 'train':
logger.info('Loading Training and Validation Set')
train_x, train_labels = get_data(data_cnf['train']['texts'],
data_cnf['train']['labels'])
if 'size' in data_cnf['valid']:
random_state = data_cnf['valid'].get('random_state', 1240)
train_x, valid_x, train_labels, valid_labels = train_test_split(
train_x,
train_labels,
test_size=data_cnf['valid']['size'],
random_state=random_state)
else:
valid_x, valid_labels = get_data(data_cnf['valid']['texts'],
data_cnf['valid']['labels'])
mlb = get_mlb(data_cnf['labels_binarizer'],
np.hstack((train_labels, valid_labels)))
train_y, valid_y = mlb.transform(train_labels), mlb.transform(
valid_labels)
labels_num = len(mlb.classes_)
logger.info(F'Number of Labels: {labels_num}')
logger.info(F'Size of Training Set: {len(train_x)}')
logger.info(F'Size of Validation Set: {len(valid_x)}')
logger.info('Training')
train_loader = DataLoader(MultiLabelDataset(train_x, train_y),
model_cnf['train']['batch_size'],
shuffle=True,
num_workers=4)
valid_loader = DataLoader(MultiLabelDataset(valid_x,
valid_y,
training=True),
model_cnf['valid']['batch_size'],
num_workers=4)
if 'gpipe' not in model_cnf:
model = Model(network=model_dict[model_name],
labels_num=labels_num,
model_path=model_path,
emb_init=emb_init,
**data_cnf['model'],
**model_cnf['model'])
else:
model = GPipeModel(model_name,
labels_num=labels_num,
model_path=model_path,
emb_init=emb_init,
**data_cnf['model'],
**model_cnf['model'])
loss, p1, p5 = model.train(train_loader, valid_loader,
**model_cnf['train'])
np.save(
model_cnf['np_loss'] + "{0}_cornet_{1}_cornet_dim_{2}.npy".format(
model_name, model_cnf['model']['n_cornet_blocks'],
model_cnf['model']['cornet_dim']), loss)
np.save(
model_cnf['np_p1'] + "{0}_cornet_{1}_cornet_dim_{2}.npy".format(
model_name, model_cnf['model']['n_cornet_blocks'],
model_cnf['model']['cornet_dim']), p1)
np.save(
model_cnf['np_p5'] + "{0}_cornet_{1}_cornet_dim_{2}.npy".format(
model_name, model_cnf['model']['n_cornet_blocks'],
model_cnf['model']['cornet_dim']), p5)
logger.info('Finish Training')
if mode is None or mode == 'eval':
logger.info('Loading Test Set')
logger.info('model path: ', model_path)
mlb = get_mlb(data_cnf['labels_binarizer'])
labels_num = len(mlb.classes_)
test_x, _ = get_data(data_cnf['test']['texts'], None)
logger.info(F'Size of Test Set: {len(test_x)}')
logger.info('Predicting')
test_loader = DataLoader(MultiLabelDataset(test_x),
model_cnf['predict']['batch_size'],
num_workers=4)
if 'gpipe' not in model_cnf:
if model is None:
model = Model(network=model_dict[model_name],
labels_num=labels_num,
model_path=model_path,
emb_init=emb_init,
**data_cnf['model'],
**model_cnf['model'])
else:
if model is None:
model = GPipeModel(model_name,
labels_num=labels_num,
model_path=model_path,
emb_init=emb_init,
**data_cnf['model'],
**model_cnf['model'])
scores, labels = model.predict(test_loader,
k=model_cnf['predict'].get('k', 3801))
logger.info('Finish Predicting')
labels = mlb.classes_[labels]
output_res(data_cnf['output']['res'], F'{model_name}-{data_name}',
scores, labels)
def splitting_head_tail_train(
data_cnf,
data_cnf_path,
model_cnf,
model_cnf_path,
emb_init,
model_path,
tree_id,
output_suffix,
dry_run,
split_ratio,
):
train_x, train_labels = load_dataset(data_cnf)
logger.info(f'Split head and tail labels: {split_ratio}')
head_labels, head_labels_i, tail_labels, tail_labels_i = get_head_tail_labels(
train_labels,
split_ratio,
)
train_h_x = train_x[head_labels_i]
train_h_labels = train_labels[head_labels_i]
train_t_x = train_x[tail_labels_i]
train_t_labels = train_labels[tail_labels_i]
if 'size' in data_cnf['valid']:
valid_size = data_cnf['valid']['size']
train_h_x, valid_h_x, train_h_labels, valid_h_labels = train_test_split(
train_h_x,
train_h_labels,
test_size=valid_size if len(train_h_x) > 2 * valid_size else 0.1,
)
train_t_x, valid_t_x, train_t_labels, valid_t_labels = train_test_split(
train_t_x,
train_t_labels,
test_size=valid_size if len(train_t_x) > 2 * valid_size else 0.1,
)
else:
valid_x, valid_labels = get_data(data_cnf['valid']['texts'],
data_cnf['valid']['labels'])
valid_h_labels_i, valid_t_labels_i = get_head_tail_samples(
head_labels,
tail_labels,
valid_labels,
)
valid_t_x = valid_x[valid_h_labels_i]
valid_h_x = valid_x[valid_t_labels_i]
valid_h_labels = valid_x[valid_h_labels_i]
valid_t_labels = valid_x[valid_t_labels_i]
labels_binarizer_path = data_cnf['labels_binarizer']
mlb_h = get_mlb(f"{labels_binarizer_path}_h_{split_ratio}",
head_labels[None, ...])
mlb_t = get_mlb(f"{labels_binarizer_path}_t_{split_ratio}",
tail_labels[None, ...])
with redirect_stderr(None):
train_h_y = mlb_h.transform(train_h_labels)
valid_h_y = mlb_h.transform(valid_h_labels)
train_t_y = mlb_t.transform(train_t_labels)
valid_t_y = mlb_t.transform(valid_t_labels)
logger.info(f'Number of Head Labels: {len(head_labels):,}')
logger.info(f'Number of Tail Labels: {len(tail_labels):,}')
logger.info(f'Size of Head Training Set: {len(train_h_x):,}')
logger.info(f'Size of Head Validation Set: {len(valid_h_x):,}')
logger.info(f'Size of Tail Training Set: {len(train_t_x):,}')
logger.info(f'Size of Tail Validation Set: {len(valid_t_x):,}')
logger.info('Training')
if 'cluster' not in model_cnf:
train_h_loader = DataLoader(MultiLabelDataset(train_h_x, train_h_y),
model_cnf['train']['batch_size'],
shuffle=True,
num_workers=4)
valid_h_loader = DataLoader(MultiLabelDataset(valid_h_x,
valid_h_y,
training=False),
model_cnf['valid']['batch_size'],
num_workers=4)
head_model = Model(network=AttentionRNN,
labels_num=len(head_labels),
model_path=f'{model_path}-head',
emb_init=emb_init,
**data_cnf['model'],
**model_cnf['model'])
if not dry_run:
logger.info('Training Head Model')
head_model.train(train_h_loader, valid_h_loader,
**model_cnf['train'])
logger.info('Finish Traning Head Model')
else:
head_model.save_model()
train_t_loader = DataLoader(MultiLabelDataset(train_t_x, train_t_y),
model_cnf['train']['batch_size'],
shuffle=True,
num_workers=4)
valid_t_loader = DataLoader(MultiLabelDataset(valid_t_x,
valid_t_y,
training=False),
model_cnf['valid']['batch_size'],
num_workers=4)
tail_model = Model(network=AttentionRNN,
labels_num=len(tail_labels),
model_path=f'{model_path}-tail',
emb_init=emb_init,
**data_cnf['model'],
**model_cnf['model'])
if not dry_run:
logger.info('Training Tail Model')
tail_model.train(train_t_loader, valid_t_loader,
**model_cnf['train'])
logger.info('Finish Traning Tail Model')
else:
tail_model.save_model()
else:
raise Exception("FastAttention is not currently supported for "
"splited head and tail dataset")
log_config(data_cnf_path, model_cnf_path, dry_run)
return head_model, tail_model, head_labels, tail_labels
def splitting_head_tail_eval(
data_cnf,
model_cnf,
data_name,
model_name,
model_path,
emb_init,
tree_id,
output_suffix,
dry_run,
split_ratio,
head_labels,
tail_labels,
head_model,
tail_model,
):
logger.info('Loading Test Set')
mlb = get_mlb(data_cnf['labels_binarizer'])
labels_num = len(mlb.classes_)
test_x, _ = get_data(data_cnf['test']['texts'], None)
logger.info(F'Size of Test Set: {len(test_x):,}')
labels_binarizer_path = data_cnf['labels_binarizer']
mlb_h = get_mlb(f"{labels_binarizer_path}_h_{split_ratio}")
mlb_t = get_mlb(f"{labels_binarizer_path}_t_{split_ratio}")
if head_labels is None:
train_x, train_labels = get_data(data_cnf['train']['texts'],
data_cnf['train']['labels'])
head_labels, _, tail_labels, _ = get_head_tail_labels(
train_labels,
split_ratio,
)
h_labels_i = np.nonzero(mlb.transform(head_labels[None, ...]).toarray())[0]
t_labels_i = np.nonzero(mlb.transform(tail_labels[None, ...]).toarray())[0]
logger.info('Predicting')
if 'cluster' not in model_cnf:
test_loader = DataLoader(MultiLabelDataset(test_x),
model_cnf['predict']['batch_size'],
num_workers=4)
if head_model is None:
head_model = Model(network=AttentionRNN,
labels_num=len(head_labels),
model_path=f'{model_path}-head',
emb_init=emb_init,
load_model=True,
**data_cnf['model'],
**model_cnf['model'])
logger.info('Predicting Head Model')
h_k = model_cnf['predict'].get('top_head_k', 30)
scores_h, labels_h = head_model.predict(test_loader, k=h_k)
labels_h = mlb_h.classes_[labels_h]
logger.info('Finish Predicting Head Model')
if tail_model is None:
tail_model = Model(network=AttentionRNN,
labels_num=len(tail_labels),
model_path=f'{model_path}-tail',
emb_init=emb_init,
load_model=True,
**data_cnf['model'],
**model_cnf['model'])
logger.info('Predicting Tail Model')
t_k = model_cnf['predict'].get('top_tail_k', 70)
scores_t, labels_t = tail_model.predict(test_loader, k=t_k)
labels_t = mlb_t.classes_[labels_t]
logger.info('Finish Predicting Tail Model')
scores = np.c_[scores_h, scores_t]
labels = np.c_[labels_h, labels_t]
i = np.arange(len(scores))[:, None]
j = np.argsort(scores)[:, ::-1]
scores = scores[i, j]
labels = labels[i, j]
else:
raise Exception("FastAttention is not currently supported for "
"splited head and tail dataset")
logger.info('Finish Predicting')
score_path, label_path = output_res(data_cnf['output']['res'],
f'{model_name}-{data_name}{tree_id}',
scores, labels, output_suffix)
log_results(score_path, label_path, dry_run)