def train(training_data,
valid_data,
vocabulary,
embedding_dim,
hidden_dim,
device,
batch_size,
lr,
embedding,
all_relations,
model=None,
epoch=100,
memory_data=[],
loss_margin=0.5,
alignment_model=None):
if model is None:
torch.manual_seed(100)
model = SimilarityModel(embedding_dim, hidden_dim, len(vocabulary),
np.array(embedding), 1, device)
loss_function = nn.MarginRankingLoss(loss_margin)
model = model.to(device)
optimizer = optim.Adam(model.parameters(), lr=lr)
best_acc = 0
memory_index = 0
for epoch_i in range(epoch):
for i in range((len(training_data) - 1) // batch_size + 1):
samples = training_data[i * batch_size:(i + 1) * batch_size]
seed_rels = []
for item in samples:
if item[0] not in seed_rels:
seed_rels.append(item[0])
if len(memory_data) > 0:
all_seen_data = []
for this_memory in memory_data:
all_seen_data += this_memory
memory_batch = memory_data[memory_index]
scores, loss = feed_samples(model, memory_batch, loss_function,
all_relations, device,
alignment_model)
optimizer.step()
memory_index = (memory_index + 1) % len(memory_data)
scores, loss = feed_samples(model, samples, loss_function,
all_relations, device, alignment_model)
optimizer.step()
del scores
del loss
return model
def train(training_data, valid_data, vocabulary, embedding_dim, hidden_dim,
device, batch_size, lr, model_path, embedding, all_relations,
model=None, epoch=100, all_seen_samples=[],
task_memory_size=100, loss_margin=0.5, all_seen_rels=[]):
if model is None:
torch.manual_seed(100)
model = SimilarityModel(embedding_dim, hidden_dim, len(vocabulary),
np.array(embedding), 1, device)
loss_function = nn.MarginRankingLoss(loss_margin)
model = model.to(device)
optimizer = optim.Adam(model.parameters(), lr=lr)
best_acc = 0
for epoch_i in range(epoch):
#print('epoch', epoch_i)
#training_data = training_data[0:100]
for i in range((len(training_data)-1)//batch_size+1):
memory_data = sample_memory_data(all_seen_samples, task_memory_size)
for this_sample in memory_data:
rel_cands = [rel for rel in all_seen_rels if rel!=this_sample[0]]
this_sample[1] = random.sample(rel_cands,
min(len(rel_cands),num_cands))
memory_data_grads = get_grads_memory_data(model, memory_data,
loss_function,
all_relations,
device)
#print(memory_data_grads)
samples = training_data[i*batch_size:(i+1)*batch_size]
feed_samples(model, samples, loss_function, all_relations, device)
sample_grad = copy_grad_data(model)
if len(memory_data_grads) > 0:
if torch.matmul(memory_data_grads,
torch.t(sample_grad.view(1,-1))) < 0:
project2cone2(sample_grad, memory_data_grads)
grad_params = get_grad_params(model)
grad_dims = [param.data.numel() for param in grad_params]
overwrite_grad(grad_params, sample_grad, grad_dims)
optimizer.step()
'''
acc=evaluate_model(model, valid_data, batch_size, all_relations, device)
if acc > best_acc:
torch.save(model, model_path)
best_model = torch.load(model_path)
return best_model
'''
return model
def main(opt):
print(opt)
random.seed(opt.random_seed)
torch.manual_seed(opt.random_seed)
np.random.seed(opt.random_seed)
np.random.RandomState(opt.random_seed)
start_time = time.time()
checkpoint_dir = os.path.join(opt.checkpoint_dir, '%.f' % start_time)
device = torch.device(('cuda:%d' % opt.cuda_id) if torch.cuda.is_available() and opt.cuda_id >= 0 else 'cpu')
# do following process
split_train_data, train_data_dict, split_test_data, split_valid_data, relation_numbers, rel_features, \
split_train_relations, vocabulary, embedding = \
load_data(opt.train_file, opt.valid_file, opt.test_file, opt.relation_file, opt.glove_file,
opt.embedding_dim, opt.task_arrange, opt.rel_encode, opt.task_num,
opt.train_instance_num, opt.dataset)
print(split_train_relations)
# ------------------------------------------------------------------------
# save cluster results
our_tasks = split_train_relations
count = 0
for i in our_tasks:
count += len(i)
portion = np.zeros(10000)
portion = portion-1
for i in range(len(our_tasks)):
for j in our_tasks[i]:
portion[j - 1] = int(i)
np.save("dataset/tacred/CML_tacred_random.npy", np.array(portion).astype(int))
# -------------------------------------------------------------------------
print('\n'.join(
['Task %d\t%s' % (index, ', '.join(['%d' % rel for rel in split_train_relations[index]])) for index in
range(len(split_train_relations))]))
task_sequence = list(range(opt.task_num))
if opt.random_idx:
for i in range(opt.random_times):
random.shuffle(task_sequence)
offset_seq = task_sequence[-opt.sequence_index:] + task_sequence[:-opt.sequence_index]
split_train_data = resort_list(split_train_data, offset_seq)
split_test_data = resort_list(split_test_data, offset_seq)
split_valid_data = resort_list(split_valid_data, offset_seq)
split_train_relations = resort_list(split_train_relations, offset_seq)
print('[%s]' % ', '.join(['Task %d' % idx for idx in offset_seq]))
relid2embedidx = {}
embedidx2relid = {}
if opt.similarity == 'kl_similarity':
kl_dist_ht = read_json(opt.kl_dist_file)
sorted_similarity_index = np.argsort(np.asarray(kl_dist_ht), axis=1) + 1
elif opt.similarity == 'glove_similarity':
glove_embedding = []
embed_id = 0
for rel_id in rel_features:
glove_embedding.append(rel_features[rel_id])
relid2embedidx[rel_id] = embed_id
embedidx2relid[embed_id] = rel_id
embed_id += 1
glove_similarity = cosine_similarity(np.asarray(glove_embedding))
glove_dist = np.sqrt(1 - np.power(np.where(glove_similarity > 1.0, 1.0, glove_similarity), 2))
sorted_embed_index = np.argsort(np.asarray(glove_dist), axis=1)
sorted_similarity_index = np.zeros(sorted_embed_index.shape)
for i in range(sorted_embed_index.shape[0]):
for j in range(sorted_embed_index.shape[1]):
sorted_similarity_index[i][j] = embedidx2relid[sorted_embed_index[i][j]]
else:
raise Exception('similarity method not implemented')
# prepare model
inner_model = SimilarityModel(opt.embedding_dim, opt.hidden_dim, len(vocabulary),
np.array(embedding), 1, device)
memory_data = []
memory_pool = []
memory_question_embed = []
memory_relation_embed = []
sequence_results = []
result_whole_test = []
seen_relations = []
all_seen_relations = []
rel2instance_memory = {}
memory_index = 0
seen_task_relations = []
rel_embeddings = []
for task_ix in range(opt.task_num): # outside loop
# reptile start model parameters pi
weights_before = deepcopy(inner_model.state_dict())
train_task = split_train_data[task_ix]
test_task = split_test_data[task_ix]
valid_task = split_valid_data[task_ix]
train_relations = split_train_relations[task_ix]
seen_task_relations.append(train_relations)
# collect seen relations
for data_item in train_task:
if data_item[0] not in seen_relations:
seen_relations.append(data_item[0])
# remove unseen relations
current_train_data = remove_unseen_relation(train_task, seen_relations, dataset=opt.dataset)
current_valid_data = remove_unseen_relation(valid_task, seen_relations, dataset=opt.dataset)
current_test_data = []
for previous_task_id in range(task_ix + 1):
current_test_data.append(
remove_unseen_relation(split_test_data[previous_task_id], seen_relations, dataset=opt.dataset))
for this_sample in current_train_data:
if this_sample[0] not in all_seen_relations:
all_seen_relations.append(this_sample[0])
update_rel_cands(memory_data, all_seen_relations, opt.num_cands)
# train inner_model
loss_function = nn.MarginRankingLoss(opt.loss_margin)
inner_model = inner_model.to(device)
optimizer = optim.Adam(inner_model.parameters(), lr=opt.learning_rate)
t = tqdm(range(opt.outside_epoch))
best_valid_acc = 0.0
early_stop = 0
best_checkpoint = ''
#
resorted_memory_pool = []
for epoch in t:
batch_num = (len(current_train_data) - 1) // opt.batch_size + 1
total_loss = 0.0
target_rel = -1
for batch in range(batch_num):
batch_train_data = current_train_data[batch * opt.batch_size: (batch + 1) * opt.batch_size]
if len(memory_data) > 0:
# CML
if target_rel == -1 or len(resorted_memory_pool) == 0:
target_rel = batch_train_data[0][0]
if opt.similarity == 'kl_similarity':
target_rel_sorted_index = sorted_similarity_index[target_rel - 1]
else:
target_rel_sorted_index = sorted_similarity_index[relid2embedidx[target_rel]]
resorted_memory_pool = resort_memory(memory_pool, target_rel_sorted_index)
if len(resorted_memory_pool) >= opt.task_memory_size:
current_memory = resorted_memory_pool[:opt.task_memory_size]
resorted_memory_pool = resorted_memory_pool[opt.task_memory_size + 1:] # update rest memory
batch_train_data.extend(current_memory)
else:
current_memory = resorted_memory_pool[:]
resorted_memory_pool = []
batch_train_data.extend(current_memory)
# MLLRE
# all_seen_data = []
# for one_batch_memory in memory_data:
# all_seen_data += one_batch_memory
#
# memory_batch = memory_data[memory_index]
# batch_train_data.extend(memory_batch)
# scores, loss = feed_samples(inner_model, memory_batch, loss_function, relation_numbers, device)
# optimizer.step()
# memory_index = (memory_index+1) % len(memory_data)
if len(rel2instance_memory) > 0: # from the second task, this will not be empty
if opt.is_curriculum_train == 'Y':
current_train_rel = batch_train_data[0][0]
current_rel_similarity_sorted_index = sorted_similarity_index[current_train_rel + 1]
seen_relation_sorted_index = []
for rel in current_rel_similarity_sorted_index:
if rel in rel2instance_memory.keys():
seen_relation_sorted_index.append(rel)
curriculum_rel_list = []
if opt.sampled_rel_num >= len(seen_relation_sorted_index):
curriculum_rel_list = seen_relation_sorted_index[:]
else:
step = len(seen_relation_sorted_index) // opt.sampled_rel_num
for i in range(0, len(seen_relation_sorted_index), step):
curriculum_rel_list.append(seen_relation_sorted_index[i])
# curriculum select relation
instance_list = []
for sampled_relation in curriculum_rel_list:
if opt.mini_batch_split == 'Y':
instance_list.append(rel2instance_memory[sampled_relation])
else:
instance_list.extend(rel2instance_memory[sampled_relation])
else:
# randomly select relation
instance_list = []
random_relation_list = random.sample(list(rel2instance_memory.keys()),
min(opt.sampled_rel_num, len(rel2instance_memory)))
for sampled_relation in random_relation_list:
if opt.mini_batch_split == 'Y':
instance_list.append(rel2instance_memory[sampled_relation])
else:
instance_list.extend(rel2instance_memory[sampled_relation])
if opt.mini_batch_split == 'Y':
for one_batch_instance in instance_list:
scores, loss = feed_samples(inner_model, one_batch_instance, loss_function,
relation_numbers, device, all_seen_relations)
optimizer.step()
else:
scores, loss = feed_samples(inner_model, instance_list, loss_function, relation_numbers, device,
all_seen_relations)
optimizer.step()
scores, loss = feed_samples(inner_model, batch_train_data, loss_function, relation_numbers, device,
all_seen_relations)
optimizer.step()
total_loss += loss
# valid test
valid_acc = evaluate_model(inner_model, current_valid_data, opt.batch_size, relation_numbers, device)
# checkpoint
checkpoint = {'net_state': inner_model.state_dict(), 'optimizer': optimizer.state_dict()}
if valid_acc > best_valid_acc:
best_checkpoint = '%s/checkpoint_task%d_epoch%d.pth.tar' % (checkpoint_dir, task_ix + 1, epoch)
if not os.path.exists(checkpoint_dir):
os.makedirs(checkpoint_dir)
torch.save(checkpoint, best_checkpoint)
best_valid_acc = valid_acc
early_stop = 0
else:
early_stop += 1
# print()
t.set_description('Task %i Epoch %i' % (task_ix + 1, epoch + 1))
t.set_postfix(loss=total_loss.item(), valid_acc=valid_acc, early_stop=early_stop,
best_checkpoint=best_checkpoint)
t.update(1)
if early_stop >= opt.early_stop and task_ix != 0:
# convergence
break
if task_ix == 0 and early_stop >= 40:
break
t.close()
print('Load best check point from %s' % best_checkpoint)
checkpoint = torch.load(best_checkpoint)
weights_after = checkpoint['net_state']
if opt.outer_step_formula == 'fixed':
outer_step_size = opt.step_size
elif opt.outer_step_formula == 'linear':
outer_step_size = opt.step_size * (1 - task_ix / opt.task_num)
elif opt.outer_step_formula == 'square_root':
outer_step_size = math.sqrt(opt.step_size * (1 - task_ix / opt.task_num))
# outer_step_size = 0.4
inner_model.load_state_dict(
{name: weights_before[name] + (weights_after[name] - weights_before[name]) * outer_step_size
for name in weights_before})
results = [evaluate_model(inner_model, test_data, opt.batch_size, relation_numbers, device)
for test_data in current_test_data]
# sample memory from current_train_data
if opt.memory_select_method == 'select_for_relation':
# sample instance for one relation
for rel in train_relations:
rel_items = remove_unseen_relation(train_data_dict[rel], seen_relations, dataset=opt.dataset)
rel_memo = select_data(inner_model, rel_items, int(opt.sampled_instance_num),
relation_numbers, opt.batch_size, device)
rel2instance_memory[rel] = rel_memo
if opt.memory_select_method == 'select_for_task':
# sample instance for one Task
rel_instance_num = math.ceil(opt.sampled_instance_num_total / len(train_relations))
for rel in train_relations:
rel_items = remove_unseen_relation(train_data_dict[rel], seen_relations, dataset=opt.dataset)
rel_memo = select_data(inner_model, rel_items, rel_instance_num,
relation_numbers, opt.batch_size, device)
rel2instance_memory[rel] = rel_memo
if opt.task_memory_size > 0:
# sample memory from current_train_data
if opt.memory_select_method == 'random':
memory_data.append(random_select_data(current_train_data, int(opt.task_memory_size)))
elif opt.memory_select_method == 'vec_cluster':
selected_memo = select_data(inner_model, current_train_data, int(opt.task_memory_size),
relation_numbers, opt.batch_size, device)
memory_data.append(selected_memo) # memorydata-list
memory_pool.extend(selected_memo)
elif opt.memory_select_method == 'difficulty':
memory_data.append()
print_list(results)
avg_result = sum(results) / len(results)
test_set_size = [len(testdata) for testdata in current_test_data]
whole_result = sum([results[i] * test_set_size[i] for i in range(len(current_test_data))]) / sum(test_set_size)
print('test_set_size: [%s]' % ', '.join([str(size) for size in test_set_size]))
print('avg_acc: %.3f, whole_acc: %.3f' % (avg_result, whole_result))
print('test_all:')
result_total_for_avg = []
result_total_for_whole = []
for epoch in range(10):
current_test_data = []
for previous_task_id in range(opt.task_num):
current_test_data.append(
remove_unseen_relation(split_test_data[previous_task_id], seen_relations, dataset=opt.dataset))
loss_function = nn.MarginRankingLoss(opt.loss_margin)
optimizer = optim.Adam(inner_model.parameters(), lr=opt.learning_rate)
optimizer.zero_grad()
for one_batch_memory in memory_data:
scores, loss = feed_samples(inner_model, one_batch_memory, loss_function, relation_numbers, device,
all_seen_relations)
optimizer.step()
results = [evaluate_model(inner_model, test_data, opt.batch_size, relation_numbers, device)
for test_data in current_test_data]
print(results)
avg_result = sum(results) / len(results)
test_set_size = [len(testdata) for testdata in current_test_data]
whole_result = sum([results[i] * test_set_size[i] for i in range(len(current_test_data))]) / sum(test_set_size)
print('test_set_size: [%s]' % ', '.join([str(size) for size in test_set_size]))
print('avg_acc: %.3f, whole_acc: %.3f' % (avg_result, whole_result))
result_total_for_avg.append(results)
result_total_for_whole.append(whole_result)
# clean saved parameters
files = os.listdir(checkpoint_dir)
for weigths_file in files:
os.remove(os.path.join(checkpoint_dir, weigths_file))
os.removedirs(checkpoint_dir)
# -----------------------------------------------------------
# 输出结果
avg_total = np.mean(np.array(result_total_for_avg), 1)
avg_mean, avg_interval = interval(avg_total)
whole_mean, whole_interval = interval(np.array(result_total_for_whole))
result_total = {"avg_acc": result_total_for_avg, "whole_acc": result_total_for_whole,
"avg_mean": avg_mean, "avg_interval": avg_interval.tolist(),
"whole_mean": whole_mean, "whole_interval": whole_interval.tolist()}
print(result_total)
with open(opt.result_file, "w") as file_out:
json.dump(result_total, file_out)
def main(opt):
print(opt)
# print('线性outer step formula，0.6 step size， 每task聚类取50个memo')
random.seed(opt.random_seed)
torch.manual_seed(opt.random_seed)
np.random.seed(opt.random_seed)
np.random.RandomState(opt.random_seed)
start_time = time.time()
checkpoint_dir = os.path.join(opt.checkpoint_dir, '%.f' % start_time)
device = torch.device((
'cuda:%d' % opt.cuda_id
) if torch.cuda.is_available() and opt.cuda_id >= 0 else 'cpu')
# do following process
split_train_data, train_data_dict, split_test_data, split_valid_data, relation_numbers, rel_features, \
split_train_relations, vocabulary, embedding = \
load_data(opt.train_file, opt.valid_file, opt.test_file, opt.relation_file, opt.glove_file,
opt.embedding_dim, opt.task_arrange, opt.rel_encode, opt.task_num,
opt.train_instance_num, opt.dataset)
print('\n'.join([
'Task %d\t%s' %
(index, ', '.join(['%d' % rel
for rel in split_train_relations[index]]))
for index in range(len(split_train_relations))
]))
# offset tasks
# split_train_data = offset_list(split_train_data, opt.task_offset)
# split_test_data = offset_list(split_test_data, opt.task_offset)
# split_valid_data = offset_list(split_valid_data, opt.task_offset)
# task_sq = [None] * len(split_train_relations)
# for i in range(len(split_train_relations)):
# task_sq[(i + opt.task_offset) % len(split_train_relations)] = i
# print('[%s]' % ', '.join(['Task %d' % i for i in task_sq]))
# insert 6th-task
# task_index = [[6, 0, 1, 2, 3, 4, 5, 7, 8, 9],
# [0, 6, 1, 2, 3, 4, 5, 7, 8, 9],
# [0, 1, 6, 2, 3, 4, 5, 7, 8, 9],
# [0, 1, 2, 6, 3, 4, 5, 7, 8, 9],
# [0, 1, 2, 3, 6, 4, 5, 7, 8, 9],
# [0, 1, 2, 3, 4, 6, 5, 7, 8, 9],
# [0, 1, 2, 3, 4, 5, 6, 7, 8, 9],
# [0, 1, 2, 3, 4, 5, 7, 6, 8, 9],
# [0, 1, 2, 3, 4, 5, 7, 8, 6, 9],
# [0, 1, 2, 3, 4, 5, 7, 8, 9, 6]]
# task_sequence = [[0, 1, 2, 3, 4, 5, 6, 7, 8, 9],
# [9, 0, 1, 2, 3, 4, 5, 6, 7, 8],
# [8, 9, 0, 1, 2, 3, 4, 5, 6, 7],
# [7, 8, 9, 0, 1, 2, 3, 4, 5, 6],
# [6, 7, 8, 9, 0, 1, 2, 3, 4, 5],
# [5, 6, 7, 8, 9, 0, 1, 2, 3, 4],
# [4, 5, 6, 7, 8, 9, 0, 1, 2, 3],
# [3, 4, 5, 6, 7, 8, 9, 0, 1, 2],
# [2, 3, 4, 5, 6, 7, 8, 9, 0, 1],
# [1, 2, 3, 4, 5, 6, 7, 8, 9, 0]]
task_sequence = list(range(opt.task_num))
if opt.random_idx:
for i in range(opt.random_times):
random.shuffle(task_sequence)
offset_seq = task_sequence[
-opt.sequence_index:] + task_sequence[:-opt.sequence_index]
# task_sequence = [[0, 1, 2, 3, 4, 5, 6, 7, 8, 9],
# [9, 0, 1, 2, 3, 4, 5, 6, 7, 8],
# [8, 9, 0, 1, 2, 3, 4, 5, 6, 7],
# [7, 8, 9, 0, 1, 2, 3, 4, 5, 6],
# [6, 7, 8, 9, 0, 1, 2, 3, 4, 5],
# [5, 6, 7, 8, 9, 0, 1, 2, 3, 4],
# [4, 5, 6, 7, 8, 9, 0, 1, 2, 3],
# [3, 4, 5, 6, 7, 8, 9, 0, 1, 2],
# [2, 3, 4, 5, 6, 7, 8, 9, 0, 1],
# [1, 2, 3, 4, 5, 6, 7, 8, 9, 0]]
#
# if opt.random_idx:
# random_idx = [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]
# for i in range(opt.random_times):
# random.shuffle(random_idx)
# print(str(random_idx))
#
# for i in range(len(task_sequence)):
# new_sq = [0] * len(task_sequence[i])
# for j in range(len(task_sequence[i])):
# new_sq[j] = random_idx[task_sequence[i][j]]
# task_sequence[i] = new_sq
split_train_data = resort_list(split_train_data, offset_seq)
split_test_data = resort_list(split_test_data, offset_seq)
split_valid_data = resort_list(split_valid_data, offset_seq)
split_train_relations = resort_list(split_train_relations, offset_seq)
print('[%s]' % ', '.join(['Task %d' % idx for idx in offset_seq]))
relid2embedidx = {}
embedidx2relid = {}
if opt.similarity == 'kl_similarity':
kl_dist_ht = read_json(opt.kl_dist_file)
# tmp = [[0, 1, 2, 3], [1, 0, 4, 6], [2, 4, 0, 5], [3, 6, 5, 0]]
sorted_similarity_index = np.argsort(np.asarray(kl_dist_ht),
axis=1) + 1
elif opt.similarity == 'glove_similarity':
glove_embedding = []
embed_id = 0
for rel_id in rel_features:
glove_embedding.append(rel_features[rel_id])
relid2embedidx[rel_id] = embed_id
embedidx2relid[embed_id] = rel_id
embed_id += 1
glove_similarity = cosine_similarity(np.asarray(glove_embedding))
glove_dist = np.sqrt(1 - np.power(
np.where(glove_similarity > 1.0, 1.0, glove_similarity), 2))
sorted_embed_index = np.argsort(np.asarray(glove_dist), axis=1)
sorted_similarity_index = np.zeros(sorted_embed_index.shape)
for i in range(sorted_embed_index.shape[0]):
for j in range(sorted_embed_index.shape[1]):
sorted_similarity_index[i][j] = embedidx2relid[
sorted_embed_index[i][j]]
# print()
# for i in range(1, len(rel_features) + 1):
# rel_embed = rel_features[i] # 1-80
# glove_embedding.append(rel_embed) # index 0-79, total 80
#
# glove_similarity = cosine_similarity(np.asarray(glove_embedding))
# glove_dist = np.sqrt(1 - np.power(np.where(glove_similarity > 1.0, 1.0, glove_similarity), 2))
# sorted_similarity_index = np.argsort(np.asarray(glove_dist), axis=1) + 1
else:
raise Exception('similarity method not implemented')
# prepare model
inner_model = SimilarityModel(opt.embedding_dim, opt.hidden_dim,
len(vocabulary), np.array(embedding), 1,
device)
memory_data = []
memory_pool = []
memory_question_embed = []
memory_relation_embed = []
sequence_results = []
result_whole_test = []
seen_relations = []
all_seen_relations = []
rel2instance_memory = {}
memory_index = 0
seen_task_relations = []
rel_embeddings = []
for task_ix in range(opt.task_num): # outside loop
# reptile start model parameters pi
weights_before = deepcopy(inner_model.state_dict())
train_task = split_train_data[task_ix]
test_task = split_test_data[task_ix]
valid_task = split_valid_data[task_ix]
train_relations = split_train_relations[task_ix]
seen_task_relations.append(train_relations)
# collect seen relations
for data_item in train_task:
if data_item[0] not in seen_relations:
seen_relations.append(data_item[0])
# remove unseen relations
current_train_data = remove_unseen_relation(train_task,
seen_relations,
dataset=opt.dataset)
current_valid_data = remove_unseen_relation(valid_task,
seen_relations,
dataset=opt.dataset)
current_test_data = []
for previous_task_id in range(task_ix + 1):
current_test_data.append(
remove_unseen_relation(split_test_data[previous_task_id],
seen_relations,
dataset=opt.dataset))
for this_sample in current_train_data:
if this_sample[0] not in all_seen_relations:
all_seen_relations.append(this_sample[0])
update_rel_cands(memory_data, all_seen_relations, opt.num_cands)
# train inner_model
loss_function = nn.MarginRankingLoss(opt.loss_margin)
inner_model = inner_model.to(device)
optimizer = optim.Adam(inner_model.parameters(), lr=opt.learning_rate)
t = tqdm(range(opt.outside_epoch))
best_valid_acc = 0.0
early_stop = 0
best_checkpoint = ''
#
resorted_memory_pool = []
for epoch in t:
batch_num = (len(current_train_data) - 1) // opt.batch_size + 1
total_loss = 0.0
target_rel = -1
for batch in range(batch_num):
batch_train_data = current_train_data[batch *
opt.batch_size:(batch +
1) *
opt.batch_size]
if len(memory_data) > 0:
# curriculum select and organize memory
if target_rel == -1 or len(resorted_memory_pool) == 0:
target_rel = batch_train_data[0][0]
if opt.similarity == 'kl_similarity':
target_rel_sorted_index = sorted_similarity_index[
target_rel - 1]
else:
target_rel_sorted_index = sorted_similarity_index[
relid2embedidx[target_rel]]
resorted_memory_pool = resort_memory(
memory_pool, target_rel_sorted_index)
if len(resorted_memory_pool) >= opt.task_memory_size:
current_memory = resorted_memory_pool[:opt.
task_memory_size]
resorted_memory_pool = resorted_memory_pool[
opt.task_memory_size + 1:] # 更新剩余的memory
batch_train_data.extend(current_memory)
else:
current_memory = resorted_memory_pool[:]
resorted_memory_pool = [] # 更新剩余的memory
batch_train_data.extend(current_memory)
# 淘汰的做法
# if len(resorted_memory_pool) != 0:
# current_memory = resorted_memory_pool[:opt.task_memory_size]
# resorted_memory_pool = resorted_memory_pool[opt.task_memory_size + 1:] # 更新剩余的memory
# batch_train_data.extend(current_memory)
# else:
# target_rel = batch_train_data[0][0]
# target_rel_sorted_index = sorted_similarity_index[target_rel - 1]
# resorted_memory_pool = resort_memory(memory_pool, target_rel_sorted_index)
# MLLRE的做法
# all_seen_data = []
# for one_batch_memory in memory_data:
# all_seen_data += one_batch_memory
#
# memory_batch = memory_data[memory_index]
# batch_train_data.extend(memory_batch)
# scores, loss = feed_samples(inner_model, memory_batch, loss_function, relation_numbers, device)
# optimizer.step()
# memory_index = (memory_index+1) % len(memory_data)
# random.shuffle(batch_train_data)
if len(rel2instance_memory
) > 0: # from the second task, this will not be empty
if opt.is_curriculum_train == 'Y':
current_train_rel = batch_train_data[0][0]
current_rel_similarity_sorted_index = sorted_similarity_index[
current_train_rel + 1]
seen_relation_sorted_index = []
for rel in current_rel_similarity_sorted_index:
if rel in rel2instance_memory.keys():
seen_relation_sorted_index.append(rel)
curriculum_rel_list = []
if opt.sampled_rel_num >= len(
seen_relation_sorted_index):
curriculum_rel_list = seen_relation_sorted_index[:]
else:
step = len(seen_relation_sorted_index
) // opt.sampled_rel_num
for i in range(0, len(seen_relation_sorted_index),
step):
curriculum_rel_list.append(
seen_relation_sorted_index[i])
# curriculum select relation
instance_list = []
for sampled_relation in curriculum_rel_list:
if opt.mini_batch_split == 'Y':
instance_list.append(
rel2instance_memory[sampled_relation])
else:
instance_list.extend(
rel2instance_memory[sampled_relation])
else:
# randomly select relation
instance_list = []
random_relation_list = random.sample(
list(rel2instance_memory.keys()),
min(opt.sampled_rel_num, len(rel2instance_memory)))
for sampled_relation in random_relation_list:
if opt.mini_batch_split == 'Y':
instance_list.append(
rel2instance_memory[sampled_relation])
else:
instance_list.extend(
rel2instance_memory[sampled_relation])
if opt.mini_batch_split == 'Y':
for one_batch_instance in instance_list:
# curriculum_instance_list = remove_unseen_relation(curriculum_instance_list, seen_relations)
scores, loss = feed_samples(
inner_model, one_batch_instance, loss_function,
relation_numbers, device)
optimizer.step()
else:
# curriculum_instance_list = remove_unseen_relation(curriculum_instance_list, seen_relations)
scores, loss = feed_samples(inner_model, instance_list,
loss_function,
relation_numbers, device)
optimizer.step()
scores, loss = feed_samples(inner_model, batch_train_data,
loss_function, relation_numbers,
device)
optimizer.step()
total_loss += loss
# valid test
valid_acc = evaluate_model(inner_model, current_valid_data,
opt.batch_size, relation_numbers,
device)
# checkpoint
checkpoint = {
'net_state': inner_model.state_dict(),
'optimizer': optimizer.state_dict()
}
if valid_acc > best_valid_acc:
best_checkpoint = '%s/checkpoint_task%d_epoch%d.pth.tar' % (
checkpoint_dir, task_ix + 1, epoch)
if not os.path.exists(checkpoint_dir):
os.makedirs(checkpoint_dir)
torch.save(checkpoint, best_checkpoint)
best_valid_acc = valid_acc
early_stop = 0
else:
early_stop += 1
# print()
t.set_description('Task %i Epoch %i' % (task_ix + 1, epoch + 1))
t.set_postfix(loss=total_loss.item(),
valid_acc=valid_acc,
early_stop=early_stop,
best_checkpoint=best_checkpoint)
t.update(1)
if early_stop >= opt.early_stop and task_ix != 0:
# 已经充分训练了
break
if task_ix == 0 and early_stop >= 40: # 防止大数据量的task在第一轮得不到充分训练
break
t.close()
print('Load best check point from %s' % best_checkpoint)
checkpoint = torch.load(best_checkpoint)
weights_after = checkpoint['net_state']
# weights_after = inner_model.state_dict() # 经过inner_epoch轮次的梯度更新后weights
# outer_step_size = opt.step_size * (1 - task_index / opt.task_num) # linear schedule
# if outer_step_size < opt.step_size * 0.5:
# outer_step_size = opt.step_size * 0.5
if opt.outer_step_formula == 'fixed':
outer_step_size = opt.step_size
elif opt.outer_step_formula == 'linear':
outer_step_size = opt.step_size * (1 - task_ix / opt.task_num)
elif opt.outer_step_formula == 'square_root':
outer_step_size = math.sqrt(opt.step_size *
(1 - task_ix / opt.task_num))
# outer_step_size = 0.4
inner_model.load_state_dict({
name: weights_before[name] +
(weights_after[name] - weights_before[name]) * outer_step_size
for name in weights_before
})
# 用memory进行训练：
# for i in range(5):
# for one_batch_memory in memory_data:
# scores, loss = feed_samples(inner_model, one_batch_memory, loss_function, relation_numbers, device)
# optimizer.step()
results = [
evaluate_model(inner_model, test_data, opt.batch_size,
relation_numbers, device)
for test_data in current_test_data
] # 使用current model和alignment model对test data进行一个预测
# sample memory from current_train_data
if opt.memory_select_method == 'select_for_relation':
# 每个关系sample k个
for rel in train_relations:
rel_items = remove_unseen_relation(train_data_dict[rel],
seen_relations,
dataset=opt.dataset)
rel_memo = select_data(inner_model, rel_items,
int(opt.sampled_instance_num),
relation_numbers, opt.batch_size,
device)
rel2instance_memory[rel] = rel_memo
if opt.memory_select_method == 'select_for_task':
# 为每个task sample k个
rel_instance_num = math.ceil(opt.sampled_instance_num_total /
len(train_relations))
for rel in train_relations:
rel_items = remove_unseen_relation(train_data_dict[rel],
seen_relations,
dataset=opt.dataset)
rel_memo = select_data(inner_model, rel_items,
rel_instance_num, relation_numbers,
opt.batch_size, device)
rel2instance_memory[rel] = rel_memo
if opt.task_memory_size > 0:
# sample memory from current_train_data
if opt.memory_select_method == 'random':
memory_data.append(
random_select_data(current_train_data,
int(opt.task_memory_size)))
elif opt.memory_select_method == 'vec_cluster':
selected_memo = select_data(inner_model, current_train_data,
int(opt.task_memory_size),
relation_numbers, opt.batch_size,
device)
memory_data.append(
selected_memo
) # memorydata是一个list，list中的每个元素都是一个包含selected_num个sample的list
memory_pool.extend(selected_memo)
elif opt.memory_select_method == 'difficulty':
memory_data.append()
print_list(results)
avg_result = sum(results) / len(results)
test_set_size = [len(testdata) for testdata in current_test_data]
whole_result = sum([
results[i] * test_set_size[i]
for i in range(len(current_test_data))
]) / sum(test_set_size)
print('test_set_size: [%s]' %
', '.join([str(size) for size in test_set_size]))
print('avg_acc: %.3f, whole_acc: %.3f' % (avg_result, whole_result))
# end of each task, get embeddings of all
# if len(all_seen_relations) > 1:
# rel_embed = tsne_relations(inner_model, seen_task_relations, relation_numbers, device, task_sequence[opt.sequence_index])
# rel_embeddings.append(rel_embed)
print('test_all:')
for epoch in range(10):
current_test_data = []
for previous_task_id in range(opt.task_num):
current_test_data.append(
remove_unseen_relation(split_test_data[previous_task_id],
seen_relations,
dataset=opt.dataset))
loss_function = nn.MarginRankingLoss(opt.loss_margin)
optimizer = optim.Adam(inner_model.parameters(), lr=opt.learning_rate)
optimizer.zero_grad()
for one_batch_memory in memory_data:
scores, loss = feed_samples(inner_model, one_batch_memory,
loss_function, relation_numbers,
device)
optimizer.step()
results = [
evaluate_model(inner_model, test_data, opt.batch_size,
relation_numbers, device)
for test_data in current_test_data
]
print(results)
avg_result = sum(results) / len(results)
test_set_size = [len(testdata) for testdata in current_test_data]
whole_result = sum([
results[i] * test_set_size[i]
for i in range(len(current_test_data))
]) / sum(test_set_size)
print('test_set_size: [%s]' %
', '.join([str(size) for size in test_set_size]))
print('avg_acc: %.3f, whole_acc: %.3f' % (avg_result, whole_result))
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--cuda_id',
default=0,
type=int,
help='cuda device index, -1 means use cpu')
parser.add_argument('--train_file',
default='dataset/training_data.txt',
help='train file')
parser.add_argument('--valid_file',
default='dataset/val_data.txt',
help='valid file')
parser.add_argument('--test_file',
default='dataset/val_data.txt',
help='test file')
parser.add_argument('--relation_file',
default='dataset/relation_name.txt',
help='relation name file')
parser.add_argument('--glove_file',
default='dataset/glove.6B.50d.txt',
help='glove embedding file')
parser.add_argument('--embedding_dim',
default=50,
type=int,
help='word embeddings dimensional')
parser.add_argument('--hidden_dim',
default=200,
type=int,
help='BiLSTM hidden dimensional')
parser.add_argument(
'--task_arrange',
default='random',
help='task arrangement method, e.g. cluster_by_glove_embedding, random'
)
parser.add_argument('--rel_encode',
default='glove',
help='relation encode method')
parser.add_argument(
'--meta_method',
default='reptile',
help='meta learning method, maml and reptile can be choose')
parser.add_argument('--batch_size',
default=50,
type=float,
help='Reptile inner loop batch size')
parser.add_argument('--task_num',
default=10,
type=int,
help='number of tasks')
parser.add_argument(
'--train_instance_num',
default=200,
type=int,
help='number of instances for one relation, -1 means all.')
parser.add_argument('--loss_margin',
default=0.5,
type=float,
help='loss margin setting')
parser.add_argument('--outside_epoch',
default=200,
type=float,
help='task level epoch')
parser.add_argument('--early_stop',
default=20,
type=float,
help='task level epoch')
parser.add_argument('--step_size',
default=0.4,
type=float,
help='step size Epsilon')
parser.add_argument('--learning_rate',
default=2e-2,
type=float,
help='learning rate')
parser.add_argument('--random_seed',
default=317,
type=int,
help='random seed')
parser.add_argument('--task_memory_size',
default=50,
type=int,
help='number of samples for each task')
parser.add_argument(
'--memory_select_method',
default='vec_cluster',
help=
'the method of sample memory data, e.g. vec_cluster, random, difficulty'
)
opt = parser.parse_args()
print(opt)
random.seed(opt.random_seed)
torch.manual_seed(opt.random_seed)
np.random.seed(opt.random_seed)
np.random.RandomState(opt.random_seed)
device = torch.device((
'cuda:%d' % opt.cuda_id
) if torch.cuda.is_available() and opt.cuda_id >= 0 else 'cpu')
# do following process
split_train_data, train_data_dict, split_test_data, test_data_dict, split_valid_data, valid_data_dict, \
relation_numbers, rel_features, vocabulary, embedding = \
load_data(opt.train_file, opt.valid_file, opt.test_file, opt.relation_file, opt.glove_file,
opt.embedding_dim, opt.task_arrange, opt.rel_encode, opt.task_num,
opt.train_instance_num)
# prepare model
inner_model = SimilarityModel(opt.embedding_dim, opt.hidden_dim,
len(vocabulary), np.array(embedding), 1,
device)
memory_data = [] # B
seen_relations = []
for task_index in range(opt.task_num): # outside loop
# reptile start model parameters pi
weights_before = deepcopy(inner_model.state_dict())
train_task = split_train_data[task_index]
# test_task = split_test_data[task_index]
valid_task = split_valid_data[task_index]
# collect seen relations
for data_item in train_task:
if data_item[0] not in seen_relations:
seen_relations.append(data_item[0])
# remove unseen relations
current_train_data = remove_unseen_relation(train_task, seen_relations)
current_valid_data = remove_unseen_relation(valid_task, seen_relations)
current_test_data = []
for previous_task_id in range(task_index + 1):
current_test_data.append(
remove_unseen_relation(split_test_data[previous_task_id],
seen_relations))
# train inner_model
loss_function = nn.MarginRankingLoss(opt.loss_margin)
inner_model = inner_model.to(device)
t = tqdm(range(opt.outside_epoch))
best_valid_acc = 0.0
early_stop = 0
best_checkpoint = ''
for epoch in t:
weights_task = deepcopy(weights_before)
# optimizer.zero_grad()
# inner_model.load_state_dict({name: weights_before[name] for name in weights_before})
batch_num = (len(current_train_data) -
1) // opt.train_instance_num + 1
total_loss = 0.0
weights_list = [None] * (batch_num + len(memory_data))
for batch in range(batch_num):
# one relation's train data
batch_train_data = current_train_data[batch *
opt.train_instance_num:
(batch + 1) *
opt.train_instance_num]
inner_model.load_state_dict(weights_task)
optimizer = optim.SGD(inner_model.parameters(),
lr=opt.learning_rate)
optimizer.zero_grad()
scores, loss = feed_samples(inner_model, batch_train_data,
loss_function, relation_numbers,
device)
loss.backward() # 计算反向传播梯度
# 更新参数
# for f in inner_model.parameters():
# f.data.sub_(f.grad.data * opt.learning_rate)
optimizer.step() # 更新参数
total_loss += loss
weights_list[batch] = deepcopy(
inner_model.state_dict()) # 保存theta_t^i
if len(memory_data) > 0:
for i in range(len(memory_data)):
one_batch_memory = memory_data[i]
inner_model.load_state_dict(weights_task)
optimizer = optim.SGD(inner_model.parameters(),
lr=opt.learning_rate)
optimizer.zero_grad()
scores, loss = feed_samples(inner_model, one_batch_memory,
loss_function,
relation_numbers, device)
loss.backward()
# 更新参数
# for f in inner_model.parameters():
# f.data.sub_(f.grad.data * opt.learning_rate)
optimizer.step()
total_loss += loss
weights_list[batch_num + i] = deepcopy(
inner_model.state_dict())
outer_step_size = opt.step_size * (1 / len(weights_list))
for name in weights_before:
weights_task[
name] = weights_before[name] - outer_step_size * sum([
weights[name] - weights_before[name]
for weights in weights_list
])
# load state dict of weights_after
inner_model.load_state_dict(weights_task)
# weights_before = deepcopy(inner_model.state_dict())
del weights_list
valid_acc = evaluate_model(inner_model, current_valid_data,
opt.batch_size, relation_numbers,
device)
checkpoint = {'net_state': inner_model.state_dict()}
if valid_acc > best_valid_acc:
best_checkpoint = './checkpoint/checkpoint_task%d_epoch%d.pth.tar' % (
task_index, epoch)
torch.save(checkpoint, best_checkpoint)
best_valid_acc = valid_acc
early_stop = 0
else:
early_stop += 1
# print()
t.set_description('Task %i Epoch %i' % (task_index + 1, epoch + 1))
t.set_postfix(loss=total_loss.item(),
valid_acc=valid_acc,
early_stop=early_stop,
best_checkpoint=best_checkpoint)
t.update(1)
if early_stop >= opt.early_stop:
# 已经充分训练了
break
t.close()
# sample memory from current_train_data
if opt.memory_select_method == 'random':
memory_data.append(
random_select_data(current_train_data, opt.task_memory_size))
elif opt.memory_select_method == 'vec_cluster':
memory_data.append(
select_data(inner_model, current_train_data,
opt.task_memory_size, relation_numbers,
opt.batch_size, device)
) # memorydata是一个list，list中的每个元素都是一个包含selected_num个sample的list
elif opt.memory_select_method == 'difficulty':
memory_data.append()
results = [
evaluate_model(inner_model, test_data, opt.batch_size,
relation_numbers, device)
for test_data in current_test_data
] # 使用current model和alignment model对test data进行一个预测
print(results)
def train_memory(training_data,
valid_data,
vocabulary,
embedding_dim,
hidden_dim,
device,
batch_size,
lr,
model_path,
embedding,
all_relations,
model=None,
epoch=100,
memory_data=[],
loss_margin=0.5,
past_fisher=None,
rel_samples=[],
relation_frequences=[],
rel_embeds=None,
rel_ques_cand=None,
rel_acc_diff=None,
all_seen_rels=None,
update_rel_embed=None,
reverse_model=None,
memory_que_embed=[],
memory_rel_embed=[],
to_update_reverse=False):
if model is None:
torch.manual_seed(100)
model = SimilarityModel(embedding_dim, hidden_dim, len(vocabulary),
np.array(embedding), 1, device)
loss_function = nn.MarginRankingLoss(loss_margin)
model = model.to(device)
optimizer = optim.Adam(model.parameters(), lr=lr)
reverse_optimiser = torch.optim.Adam(reverse_model.parameters(), lr=lr)
best_acc = 0
#acc_pre=evaluate_model(model, valid_data, batch_size, all_relations, device)
given_pro = None
memory_index = 0
memory_data_grads = []
for epoch_i in range(epoch):
#print('epoch', epoch_i)
#training_data = training_data[0:100]
#for i in range((len(training_data)-1)//batch_size+1):
for i, samples in enumerate(memory_data):
#samples = training_data[i*batch_size:(i+1)*batch_size]
seed_rels = []
for item in samples:
if item[0] not in seed_rels:
seed_rels.append(item[0])
#start_time = time.time()
'''
if len(memory_data) > 0:
all_seen_data = []
for this_memory in memory_data:
all_seen_data+=this_memory
memory_batch = memory_data[memory_index]
#memory_batch = random.sample(all_seen_data,
# min(batch_size, len(all_seen_data)))
#print(memory_data)
scores, loss = feed_samples(model, memory_batch,
loss_function,
all_relations, device, reverse_model)
if to_update_reverse:
reverse_optimiser.step()
else:
optimizer.step()
memory_index = (memory_index+1)%len(memory_data)
'''
scores, loss = feed_samples(model, samples, loss_function,
all_relations, device, reverse_model)
#memory_que_embed[i], memory_rel_embed[i])
#end_time = time.time()
#print('forward time:', end_time - start_time)
sample_grad = copy_grad_data(model)
if to_update_reverse:
reverse_optimiser.step()
else:
optimizer.step()
del scores
del loss
'''
acc=evaluate_model(model, valid_data, batch_size, all_relations, device)
if acc > best_acc:
torch.save(model, model_path)
best_model = torch.load(model_path)
return best_model
'''
#acc_aft=evaluate_model(model, valid_data, batch_size, all_relations, device)
#return model, max(0, acc_aft-acc_pre)
if to_update_reverse:
return reverse_model, 0
else:
return model, 0
def train(training_data,
valid_data,
vocabulary,
embedding_dim,
hidden_dim,
device,
batch_size,
lr,
model_path,
embedding,
all_relations,
model=None,
epoch=100,
memory_data=[],
loss_margin=0.5,
past_fisher=None,
rel_samples=[],
relation_frequences=[],
rel_embeds=None,
rel_ques_cand=None,
rel_acc_diff=None,
all_seen_rels=None,
update_rel_embed=None,
reverse_model=None,
memory_que_embed=[],
memory_rel_embed=[],
to_update_reverse=False):
if model is None:
torch.manual_seed(100)
model = SimilarityModel(embedding_dim, hidden_dim, len(vocabulary),
np.array(embedding), 1, device)
loss_function = nn.MarginRankingLoss(loss_margin)
model = model.to(device)
optimizer = optim.Adam(model.parameters(), lr=lr)
reverse_optimiser = torch.optim.Adam(reverse_model.parameters(), lr=lr)
best_acc = 0
#acc_pre=evaluate_model(model, valid_data, batch_size, all_relations, device)
given_pro = None
memory_index = 0
memory_data_grads = []
for epoch_i in range(epoch):
#print('epoch', epoch_i)
#training_data = training_data[0:100]
for i in range((len(training_data) - 1) // batch_size + 1):
samples = training_data[i * batch_size:(i + 1) * batch_size]
seed_rels = []
for item in samples:
if item[0] not in seed_rels:
seed_rels.append(item[0])
'''
for item_cand in item[1]:
if item_cand not in seed_rels:
seed_rels.append(item_cand)
'''
if len(rel_samples) > 0:
memory_data = sample_constrains(rel_samples,
relation_frequences,
rel_embeds, seed_rels,
rel_ques_cand, rel_acc_diff,
given_pro, all_seen_rels)
'''
to_train_mem = memory_data
if len(memory_data) > num_constrain:
to_train_mem = random.sample(memory_data, num_constrain)
memory_data_grads = get_grads_memory_data(model, to_train_mem,
loss_function,
all_relations,
device, reverse_model,
memory_que_embed,
memory_rel_embed)
'''
#print(memory_data_grads)
#start_time = time.time()
if len(memory_data) > 0:
all_seen_data = []
for this_memory in memory_data:
all_seen_data += this_memory
memory_batch = memory_data[memory_index]
#memory_batch = random.sample(all_seen_data,
# min(batch_size, len(all_seen_data)))
#print(memory_data)
scores, loss = feed_samples(model, memory_batch, loss_function,
all_relations, device,
reverse_model)
if to_update_reverse:
reverse_optimiser.step()
else:
optimizer.step()
memory_index = (memory_index + 1) % len(memory_data)
scores, loss = feed_samples(model, samples, loss_function,
all_relations, device, reverse_model)
#end_time = time.time()
#print('forward time:', end_time - start_time)
sample_grad = copy_grad_data(model)
if len(memory_data_grads) > 0:
#if not check_constrain(memory_data_grads, sample_grad):
if True:
project2cone2(sample_grad, memory_data_grads)
if past_fisher is None:
grad_params = get_grad_params(model)
grad_dims = [
param.data.numel() for param in grad_params
]
overwrite_grad(grad_params, sample_grad, grad_dims)
if past_fisher is not None:
sample_grad = rescale_grad(sample_grad, past_fisher)
grad_params = get_grad_params(model)
grad_dims = [param.data.numel() for param in grad_params]
overwrite_grad(grad_params, sample_grad, grad_dims)
if to_update_reverse:
reverse_optimiser.step()
else:
optimizer.step()
#optimizer.step()
if (epoch_i % 5 == 0) and len(relation_frequences) > 0 and False:
update_rel_embed(model, all_seen_rels, all_relations,
rel_embeds)
samples = list(relation_frequences.keys())
#return random.sample(samples, min(len(samples), num_samples))
sample_embeds = torch.from_numpy(
np.asarray([rel_embeds[i] for i in samples]))
#seed_rel_embeds = torch.from_numpy(np.asarray(
# [rel_embeds[i] for i in seed_rels])).to(device)
sample_embeds_np = sample_embeds.cpu().double().numpy()
#given_pro = kmeans_pro(sample_embeds_np, samples, num_constrain)
if epoch_i % 5 == 0 and False:
update_rel_embed(model, all_seen_rels, all_relations,
rel_embeds)
#update_rel_cands(memory_data, all_seen_rels, rel_embeds)
del scores
del loss
'''
acc=evaluate_model(model, valid_data, batch_size, all_relations, device)
if acc > best_acc:
torch.save(model, model_path)
best_model = torch.load(model_path)
return best_model
'''
#acc_aft=evaluate_model(model, valid_data, batch_size, all_relations, device)
#return model, max(0, acc_aft-acc_pre)
if to_update_reverse:
return reverse_model, 0
else:
return model, 0
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--cuda_id',
default=0,
type=int,
help='cuda device index, -1 means use cpu')
parser.add_argument('--train_file',
default='dataset/training_data.txt',
help='train file')
parser.add_argument('--valid_file',
default='dataset/val_data.txt',
help='valid file')
parser.add_argument('--test_file',
default='dataset/val_data.txt',
help='test file')
parser.add_argument('--relation_file',
default='dataset/relation_name.txt',
help='relation name file')
parser.add_argument('--glove_file',
default='dataset/glove.6B.300d.txt',
help='glove embedding file')
parser.add_argument('--kl_dist_file',
default='dataset/kl_dist_ht.json',
help='glove embedding file')
parser.add_argument('--embedding_dim',
default=300,
type=int,
help='word embeddings dimensional')
parser.add_argument('--hidden_dim',
default=200,
type=int,
help='BiLSTM hidden dimensional')
parser.add_argument(
'--task_arrange',
default='cluster_by_glove_embedding',
help='task arrangement method, e.g. cluster_by_glove_embedding, random'
)
parser.add_argument('--rel_encode',
default='glove',
help='relation encode method')
parser.add_argument(
'--meta_method',
default='reptile',
help='meta learning method, maml and reptile can be choose')
parser.add_argument('--batch_size',
default=50,
type=float,
help='Reptile inner loop batch size')
parser.add_argument('--task_num',
default=10,
type=int,
help='number of tasks')
parser.add_argument(
'--train_instance_num',
default=200,
type=int,
help='number of instances for one relation, -1 means all.')
parser.add_argument('--loss_margin',
default=0.5,
type=float,
help='loss margin setting')
parser.add_argument('--outside_epoch',
default=200,
type=float,
help='task level epoch')
parser.add_argument('--early_stop',
default=20,
type=float,
help='task level epoch')
parser.add_argument('--step_size',
default=0.6,
type=float,
help='step size Epsilon')
parser.add_argument('--learning_rate',
default=2e-3,
type=float,
help='learning rate')
parser.add_argument('--random_seed',
default=226,
type=int,
help='random seed')
parser.add_argument('--task_memory_size',
default=50,
type=int,
help='number of samples for each task')
parser.add_argument(
'--memory_select_method',
default='vec_cluster',
help=
'the method of sample memory data, e.g. vec_cluster, random, difficulty'
)
parser.add_argument(
'--curriculum_rel_num',
default=3,
help=
'curriculum learning relation sampled number for current training relation'
)
parser.add_argument(
'--curriculum_instance_num',
default=5,
help=
'curriculum learning instance sampled number for a sampled relation')
opt = parser.parse_args()
print(opt)
random.seed(opt.random_seed)
torch.manual_seed(opt.random_seed)
np.random.seed(opt.random_seed)
np.random.RandomState(opt.random_seed)
device = torch.device((
'cuda:%d' % opt.cuda_id
) if torch.cuda.is_available() and opt.cuda_id >= 0 else 'cpu')
# do following process
split_train_data, train_data_dict, split_test_data, test_data_dict, split_valid_data, valid_data_dict, \
relation_numbers, rel_features, vocabulary, embedding = \
load_data(opt.train_file, opt.valid_file, opt.test_file, opt.relation_file, opt.glove_file,
opt.embedding_dim, opt.task_arrange, opt.rel_encode, opt.task_num,
opt.train_instance_num)
# kl similarity of the joint distribution of head and tail
kl_dist_ht = read_json(opt.kl_dist_file)
# tmp = [[0, 1, 2, 3], [1, 0, 4, 6], [2, 4, 0, 5], [3, 6, 5, 0]]
sorted_sililarity_index = np.argsort(-np.asarray(kl_dist_ht), axis=1) + 1
# prepare model
inner_model = SimilarityModel(opt.embedding_dim, opt.hidden_dim,
len(vocabulary), np.array(embedding), 1,
device)
memory_data = []
memory_question_embed = []
memory_relation_embed = []
sequence_results = []
result_whole_test = []
seen_relations = []
all_seen_relations = []
memory_index = 0
for task_index in range(opt.task_num): # outside loop
# reptile start model parameters pi
weights_before = deepcopy(inner_model.state_dict())
train_task = split_train_data[task_index]
test_task = split_test_data[task_index]
valid_task = split_valid_data[task_index]
# collect seen relations
for data_item in train_task:
if data_item[0] not in seen_relations:
seen_relations.append(data_item[0])
# remove unseen relations
current_train_data = remove_unseen_relation(train_task, seen_relations)
current_valid_data = remove_unseen_relation(valid_task, seen_relations)
current_test_data = []
for previous_task_id in range(task_index + 1):
current_test_data.append(
remove_unseen_relation(split_test_data[previous_task_id],
seen_relations))
# train inner_model
loss_function = nn.MarginRankingLoss(opt.loss_margin)
inner_model = inner_model.to(device)
optimizer = optim.Adam(inner_model.parameters(), lr=opt.learning_rate)
t = tqdm(range(opt.outside_epoch))
best_valid_acc = 0.0
early_stop = 0
best_checkpoint = ''
for epoch in t:
batch_num = (len(current_train_data) - 1) // opt.batch_size + 1
total_loss = 0.0
for batch in range(batch_num):
batch_train_data = current_train_data[batch *
opt.batch_size:(batch +
1) *
opt.batch_size]
if len(memory_data) > 0:
all_seen_data = []
for one_batch_memory in memory_data:
all_seen_data += one_batch_memory
memory_batch = memory_data[memory_index]
batch_train_data.extend(memory_batch)
# scores, loss = feed_samples(inner_model, memory_batch, loss_function, relation_numbers, device)
# optimizer.step()
memory_index = (memory_index + 1) % len(memory_data)
# random.shuffle(batch_train_data)
# curriculum before batch_train
if task_index > 0:
current_train_rel = batch_train_data[0][0]
current_rel_similarity_sorted_index = sorted_sililarity_index[
current_train_rel + 1]
seen_relation_sorted_index = []
for rel in current_rel_similarity_sorted_index:
if rel in seen_relations:
seen_relation_sorted_index.append(rel)
curriculum_rel_list = []
if opt.curriculum_rel_num >= len(
seen_relation_sorted_index):
curriculum_rel_list = seen_relation_sorted_index[:]
else:
step = len(seen_relation_sorted_index
) // opt.curriculum_rel_num
for i in range(0, len(seen_relation_sorted_index),
step):
curriculum_rel_list.append(
seen_relation_sorted_index[i])
curriculum_instance_list = []
for curriculum_rel in curriculum_rel_list:
curriculum_instance_list.extend(
random.sample(train_data_dict[curriculum_rel],
opt.curriculum_instance_num))
curriculum_instance_list = remove_unseen_relation(
curriculum_instance_list, seen_relations)
# optimizer.zero_grad()
scores, loss = feed_samples(inner_model,
curriculum_instance_list,
loss_function,
relation_numbers, device)
# loss.backward()
optimizer.step()
scores, loss = feed_samples(inner_model, batch_train_data,
loss_function, relation_numbers,
device)
optimizer.step()
total_loss += loss
# valid test
valid_acc = evaluate_model(inner_model, current_valid_data,
opt.batch_size, relation_numbers,
device)
# checkpoint
checkpoint = {
'net_state': inner_model.state_dict(),
'optimizer': optimizer.state_dict()
}
if valid_acc > best_valid_acc:
best_checkpoint = './checkpoint/checkpoint_task%d_epoch%d.pth.tar' % (
task_index, epoch)
torch.save(checkpoint, best_checkpoint)
best_valid_acc = valid_acc
early_stop = 0
else:
early_stop += 1
# print()
t.set_description('Task %i Epoch %i' % (task_index + 1, epoch + 1))
t.set_postfix(loss=total_loss.item(),
valid_acc=valid_acc,
early_stop=early_stop,
best_checkpoint=best_checkpoint)
t.update(1)
if early_stop >= opt.early_stop:
# 已经充分训练了
break
t.close()
print('Load best check point from %s' % best_checkpoint)
checkpoint = torch.load(best_checkpoint)
weights_after = checkpoint['net_state']
# weights_after = inner_model.state_dict() # 经过inner_epoch轮次的梯度更新后weights
# if task_index == opt.task_num - 1:
# outer_step_size = opt.step_size * (1 - 5 / opt.task_num)
# else:
outer_step_size = math.sqrt(opt.step_size *
(1 - task_index / opt.task_num))
# outer_step_size = opt.step_size / opt.task_num
# outer_step_size = 0.4
inner_model.load_state_dict({
name: weights_before[name] +
(weights_after[name] - weights_before[name]) * outer_step_size
for name in weights_before
})
# 用memory进行训练：
# for i in range(5):
# for one_batch_memory in memory_data:
# scores, loss = feed_samples(inner_model, one_batch_memory, loss_function, relation_numbers, device)
# optimizer.step()
results = [
evaluate_model(inner_model, test_data, opt.batch_size,
relation_numbers, device)
for test_data in current_test_data
] # 使用current model和alignment model对test data进行一个预测
# sample memory from current_train_data
if opt.memory_select_method == 'random':
memory_data.append(
random_select_data(current_train_data,
int(opt.task_memory_size / results[-1])))
elif opt.memory_select_method == 'vec_cluster':
memory_data.append(
select_data(inner_model, current_train_data,
int(opt.task_memory_size / results[-1]),
relation_numbers, opt.batch_size, device)
) # memorydata是一个list，list中的每个元素都是一个包含selected_num个sample的list
elif opt.memory_select_method == 'difficulty':
memory_data.append()
print_list(results)
avg_result = sum(results) / len(results)
test_set_size = [len(testdata) for testdata in current_test_data]
whole_result = sum([
results[i] * test_set_size[i]
for i in range(len(current_test_data))
]) / sum(test_set_size)
print('test_set_size: [%s]' %
', '.join([str(size) for size in test_set_size]))
print('avg_acc: %.3f, whole_acc: %.3f' % (avg_result, whole_result))
print('test_all:')
for epoch in range(10):
current_test_data = []
for previous_task_id in range(opt.task_num):
current_test_data.append(
remove_unseen_relation(split_test_data[previous_task_id],
seen_relations))
loss_function = nn.MarginRankingLoss(opt.loss_margin)
optimizer = optim.Adam(inner_model.parameters(), lr=opt.learning_rate)
optimizer.zero_grad()
for one_batch_memory in memory_data:
scores, loss = feed_samples(inner_model, one_batch_memory,
loss_function, relation_numbers,
device)
optimizer.step()
results = [
evaluate_model(inner_model, test_data, opt.batch_size,
relation_numbers, device)
for test_data in current_test_data
]
print(results)
avg_result = sum(results) / len(results)
test_set_size = [len(testdata) for testdata in current_test_data]
whole_result = sum([
results[i] * test_set_size[i]
for i in range(len(current_test_data))
]) / sum(test_set_size)
print('test_set_size: [%s]' %
', '.join([str(size) for size in test_set_size]))
print('avg_acc: %.3f, whole_acc: %.3f' % (avg_result, whole_result))
def train(training_data,
valid_data,
vocabulary,
embedding_dim,
hidden_dim,
device,
batch_size,
lr,
model_path,
embedding,
all_relations,
model=None,
epoch=100,
grad_means=[],
grad_fishers=[],
loss_margin=2.0):
if model is None:
torch.manual_seed(100)
model = SimilarityModel(embedding_dim, hidden_dim, len(vocabulary),
np.array(embedding), 1, device)
loss_function = nn.MarginRankingLoss(loss_margin)
model = model.to(device)
optimizer = optim.Adam(model.parameters(), lr=lr)
best_acc = 0
for epoch_i in range(epoch):
#print('epoch', epoch_i)
#training_data = training_data[0:100]
for i in range((len(training_data) - 1) // batch_size + 1):
samples = training_data[i * batch_size:(i + 1) * batch_size]
questions, relations, relation_set_lengths = process_samples(
samples, all_relations, device)
#print('got data')
ranked_questions, reverse_question_indexs = \
ranking_sequence(questions)
ranked_relations, reverse_relation_indexs =\
ranking_sequence(relations)
question_lengths = [len(question) for question in ranked_questions]
relation_lengths = [len(relation) for relation in ranked_relations]
#print(ranked_questions)
pad_questions = torch.nn.utils.rnn.pad_sequence(ranked_questions)
pad_relations = torch.nn.utils.rnn.pad_sequence(ranked_relations)
#print(pad_questions)
pad_questions = pad_questions.to(device)
pad_relations = pad_relations.to(device)
#print(pad_questions)
model.zero_grad()
model.init_hidden(device, sum(relation_set_lengths))
all_scores = model(pad_questions, pad_relations, device,
reverse_question_indexs,
reverse_relation_indexs, question_lengths,
relation_lengths)
all_scores = all_scores.to('cpu')
pos_scores = []
neg_scores = []
start_index = 0
for length in relation_set_lengths:
pos_scores.append(all_scores[start_index].expand(length - 1))
neg_scores.append(all_scores[start_index + 1:start_index +
length])
start_index += length
pos_scores = torch.cat(pos_scores)
neg_scores = torch.cat(neg_scores)
loss = loss_function(
pos_scores, neg_scores,
torch.ones(
sum(relation_set_lengths) - len(relation_set_lengths)))
loss = loss.sum()
#loss.to(device)
#print(loss)
for i in range(len(grad_means)):
grad_mean = grad_means[i]
grad_fisher = grad_fishers[i]
#print(param_loss(model, grad_mean, grad_fisher, p_lambda))
loss += param_loss(model, grad_mean, grad_fisher,
p_lambda).to('cpu')
loss.backward()
optimizer.step()
'''
acc=evaluate_model(model, valid_data, batch_size, all_relations, device)
if acc > best_acc:
torch.save(model, model_path)
best_model = torch.load(model_path)
return best_model
'''
return model