def worker(self, done_queue, sampled_classes, source_classes):
'''
Generate one task (support and query).
Store into self.support[self.cur] and self.query[self.cur]
'''
while True:
if done_queue.qsize() > 100:
time.sleep(1)
continue
# sample examples
support_idx, query_idx = [], []
for y in sampled_classes:
tmp = np.random.permutation(len(self.idx_list[y]))
support_idx.append(self.idx_list[y][tmp[:self.args.shot]])
query_idx.append(
self.idx_list[y][tmp[self.args.shot:self.args.shot +
self.args.query]])
support_idx = np.concatenate(support_idx)
query_idx = np.concatenate(query_idx)
# aggregate examples
max_support_len = np.max(self.data['text_len'][support_idx])
max_query_len = np.max(self.data['text_len'][query_idx])
support = utils.select_subset(self.data, {},
['text', 'text_len', 'label'],
support_idx, max_support_len)
query = utils.select_subset(self.data, {},
['text', 'text_len', 'label'],
query_idx, max_query_len)
done_queue.put((support, query))
def worker(self, done_queue, sampled_classes, source_classes):
'''
Generate one task (support and query).
Store into self.support[self.cur] and self.query[self.cur]
'''
example_prob_metrix = self.example_prob_metrix
while True:
if done_queue.qsize() > self.num_episodes:
time.sleep(1)
return
# continue
# sample examples
support_idx, query_idx = [], []
if example_prob_metrix is None:
for y in sampled_classes:
tmp = np.random.permutation(len(self.idx_list[y]))
if len(tmp) < self.args.shot + self.args.query:
tmp = np.random.choice(len(self.idx_list[y]),
self.args.shot +
self.args.query,
replace=True)
support_idx.append(self.idx_list[y][tmp[:self.args.shot]])
query_idx.append(
self.idx_list[y][tmp[self.args.shot:self.args.shot +
self.args.query]])
else:
for y in sampled_classes:
tmp = np.random.choice(len(self.idx_list[y]),
self.args.shot + self.args.query,
p=example_prob_metrix[y][0],
replace=False)
support_idx.append(self.idx_list[y][tmp[:self.args.shot]])
query_idx.append(self.idx_list[y][tmp[self.args.shot:]])
support_idx = np.concatenate(support_idx)
query_idx = np.concatenate(query_idx)
# aggregate examples
max_support_len = np.max(self.data['text_len'][support_idx])
max_query_len = np.max(self.data['text_len'][query_idx])
support = utils.select_subset(self.data, {},
['text', 'text_len', 'label'],
support_idx, max_support_len)
query = utils.select_subset(self.data, {},
['text', 'text_len', 'label'],
query_idx, max_query_len)
done_queue.put((support, query))
def worker(self, done_queue):
'''
Generate one task (support and query).
Store into self.support[self.cur] and self.query[self.cur]
'''
while True:
if done_queue.qsize() > 100:
time.sleep(1)
continue
# sample ways
sampled_classes = np.random.permutation(
self.num_classes)[:self.args.way]
source_classes = []
for j in range(self.num_classes):
if j not in sampled_classes:
source_classes.append(self.all_classes[j])
source_classes = sorted(source_classes)
# sample examples
support_idx, query_idx = [], []
for y in sampled_classes:
tmp = np.random.permutation(len(self.idx_list[y]))
support_idx.append(self.idx_list[y][tmp[:self.args.shot]])
query_idx.append(
self.idx_list[y][tmp[self.args.shot:self.args.shot +
self.args.query]])
support_idx = np.concatenate(support_idx)
query_idx = np.concatenate(query_idx)
if self.args.mode == 'finetune' and len(query_idx) == 0:
query_idx = support_idx
# aggregate examples
max_support_len = np.max(self.data['text_len'][support_idx])
max_query_len = np.max(self.data['text_len'][query_idx])
support = utils.select_subset(self.data, {},
['text', 'text_len', 'label'],
support_idx, max_support_len)
query = utils.select_subset(self.data, {},
['text', 'text_len', 'label'],
query_idx, max_query_len)
done_queue.put((support, query))
def worker(self, done_queue):
'''
Generate one task (support and query).
Store into self.support[self.cur] and self.query[self.cur]
'''
while True:
if done_queue.qsize() > 100:
time.sleep(1)
continue
# sample ways
sampled_classes = np.random.permutation(
self.num_classes)[:self.args.way]
source_classes = []
for j in range(self.num_classes):
if j not in sampled_classes:
source_classes.append(self.all_classes[j])
source_classes = sorted(source_classes)
# sample examples
support_idx, query_idx = [], []
for y in sampled_classes:
tmp = np.random.permutation(len(self.idx_list[y]))
support_idx.append(
self.idx_list[y][tmp[:self.args.shot]])
query_idx.append(
self.idx_list[y][
tmp[self.args.shot:self.args.shot+self.args.query]])
support_idx = np.concatenate(support_idx)
query_idx = np.concatenate(query_idx)
if self.args.mode == 'finetune' and len(query_idx) == 0:
query_idx = support_idx
# aggregate examples
max_support_len = np.max(self.data['text_len'][support_idx])
max_query_len = np.max(self.data['text_len'][query_idx])
support = utils.select_subset(self.data, {}, ['text', 'text_len', 'label'],
support_idx, max_support_len)
query = utils.select_subset(self.data, {}, ['text', 'text_len', 'label'],
query_idx, max_query_len)
if self.args.embedding in ['idf', 'meta', 'meta_mlp']:
# compute inverse document frequency over the meta-train set
idf = stats.get_idf(self.data, source_classes)
support['idf'] = idf
query['idf'] = idf
if self.args.embedding in ['iwf', 'meta', 'meta_mlp']:
# compute SIF over the meta-train set
iwf = stats.get_iwf(self.data, source_classes)
support['iwf'] = iwf
query['iwf'] = iwf
if self.args.bert:
# prepare bert token id
# +2 becuase bert_id includes [CLS] and [SEP]
support = utils.select_subset(self.data, support, ['bert_id'],
support_idx, max_support_len+2)
query = utils.select_subset(self.data, query, ['bert_id'],
query_idx, max_query_len+2)
if 'pos' in self.args.auxiliary:
support = utils.select_subset(
self.data, support, ['head', 'tail'], support_idx)
query = utils.select_subset(
self.data, query, ['head', 'tail'], query_idx)
done_queue.put((support, query))
