지속적인 학습을 위한 6가지 일반적인 방법 요약: 이전 데이터의 성능을 유지하면서 ML 모델을 새로운 데이터에 적용

class PromptPool(nn.Module):
def __init__(self, M = 100, hidden_size = 768, length = 20, N=5):
super().__init__()
self.pool = nn.Parameter(torch.rand(M, length, hidden_size), requires_grad=True).float()
self.keys = nn.Parameter(torch.rand(M, hidden_size), requires_grad=True).float()
 
self.length = length
self.hidden = hidden_size
self.n = N
 
nn.init.xavier_normal_(self.pool)
nn.init.xavier_normal_(self.keys)
 
def init_weights(self, embedding):
pass
 
# function to select from pool based on index
def concat(self, indices, input_embeds):
subset = self.pool[indices, :] # 2, 2, 20, 768
 
subset = subset.to("cuda:0").reshape(indices.size(0),
self.n*self.length,
self.hidden) # 2, 40, 768
 
return torch.cat((subset, input_embeds), 1)
 
# x is cls output
def query_fn(self, x):
 
# encode input x to same dim as key using cosine
x = x / x.norm(dim=1)[:, None]
k = self.keys / self.keys.norm(dim=1)[:, None]
 
scores = torch.mm(x, k.transpose(0,1).to("cuda:0"))
 
# get argmin
subsets = torch.topk(scores, self.n, 1, False).indices # k smallest
 
return subsets
 
 pool = PromptPool()

def train():
count = 0
print("*********** Started Training *************")
 
start = time.time()
for epoch in range(40):
model.eval()
pool.train()
 
optimizer.zero_grad(set_to_none=True)
lap = time.time()
 
for batch in iter(train_dataloader):
count += 1
q, p, train_labels = batch
 
queries_emb = model(input_ids=q['input_ids'].to("cuda:0"),
attention_mask=q['attention_mask'].to("cuda:0"))
passage_emb = model(input_ids=p['input_ids'].to("cuda:0"),
attention_mask=p['attention_mask'].to("cuda:0"))
 
# pool
q_idx = pool.query_fn(queries_emb)
raw_qembedding = model.model.embeddings(input_ids=q['input_ids'].to("cuda:0"))
q = pool.concat(indices=q_idx, input_embeds=raw_qembedding)
 
p_idx = pool.query_fn(passage_emb)
raw_pembedding = model.model.embeddings(input_ids=p['input_ids'].to("cuda:0"))
p = pool.concat(indices=p_idx, input_embeds=raw_pembedding)
 
qattention_mask = torch.ones(batch_size, q.size(1))
pattention_mask = torch.ones(batch_size, p.size(1))
 
queries_emb = model.model(inputs_embeds=q,
attention_mask=qattention_mask.to("cuda:0")).last_hidden_state
passage_emb = model.model(inputs_embeds=p,
attention_mask=pattention_mask.to("cuda:0")).last_hidden_state
 
q_cls = queries_emb[:, pool.n*pool.length+1, :]
p_cls = passage_emb[:, pool.n*pool.length+1, :]
 
loss, ql, pl = calc_loss(q_cls, p_cls)
loss.backward()
 
optimizer.step()
optimizer.zero_grad(set_to_none=True)
 
if count % 10 == 0:
print("Model Loss:", round(loss.item(),4), 
"| QL:", round(ql.item(),4), "| PL:", round(pl.item(),4), 
"| Took:", round(time.time() - lap), "secondsn")
 
lap = time.time()
 
if count % 40 == 0 and count > 0:
print("model saved")
torch.save(model.state_dict(), model_PATH)
torch.save(pool.state_dict(), pool_PATH)
 
if count == 4600: return
 
print("Training Took:", round(time.time() - start), "seconds")
print("n*********** Training Complete *************")

class DistilledData(nn.Module):
def __init__(self, num_labels, M, q_len=64, hidden_size=768):
super().__init__()
self.num_samples = M
self.q_len = q_len
self.num_labels = num_labels
self.data = nn.Parameter(torch.rand(num_labels, M, q_len, hidden_size), requires_grad=True) # i.e. shape: 1000, 4, 64, 768
 
# init using model embedding, xavier, or load from state dict
def init_weights(self, model, path=None):
if model:
self.data.requires_grad = False
print("Init weights using model embedding")
raw_embedding = model.model.get_input_embeddings()
soft_embeds = raw_embedding.weight[:, :].clone().detach()
nums = soft_embeds.size(0)
for i1 in range(self.num_labels):
for i2 in range(self.num_samples):
for i3 in range(self.q_len):
random_idx = random.randint(0, nums-1)
self.data[i1, i2, i3, :] = soft_embeds[random_idx, :]
print(self.data.shape)
self.data.requires_grad = True
 
if not path:
nn.init.xavier_normal_(self.data)
else:
distilled_data.load_state_dict(torch.load(path), strict=False)
 
# function to sample a passage and positive sample as in the article, i am doing dense retrieval
def get_sample(self, label):
q_idx = random.randint(0, self.num_samples-1)
sampled_dist_q = self.data[label, q_idx, :, :]
 
p_idx = random.randint(0, self.num_samples-1)
while q_idx == p_idx:
p_idx = random.randint(0, self.num_samples-1)
sampled_dist_p = self.data[label, p_idx, :, :]
 
return sampled_dist_q, sampled_dist_p, q_idx, p_idx

def distll_train(chunk_size=32):
count, times = 0, 0
print("*********** Started Training *************")
start = time.time()
lap = time.time()
 
for epoch in range(40):
distilled_data.train()
 
for batch in iter(train_dataloader):
count += 1
# get real query, pos, label, distilled data query, distilled data pos, ... from batch
q, p, train_labels, dq, dp, q_indexes, p_indexes = batch
 
for idx in range(0, dq['input_ids'].size(0), chunk_size):
model.train()
 
with torch.enable_grad():
# train on distiled data first
x1 = dq['input_ids'][idx:idx+chunk_size].clone().detach().requires_grad_(True)
x2 = dp['input_ids'][idx:idx+chunk_size].clone().detach().requires_grad_(True)
q_emb = model(inputs_embeds=x1.to("cuda:0"),
attention_mask=dq['attention_mask'][idx:idx+chunk_size].to("cuda:0")).cpu()
p_emb = model(inputs_embeds=x2.to("cuda:0"),
attention_mask=dp['attention_mask'][idx:idx+chunk_size].to("cuda:0"))
loss = default_loss(q_emb.to("cuda:0"), p_emb)
del q_emb, p_emb
 
loss.backward(retain_graph=True, create_graph=False)
state_dict = model.state_dict()
 
# update model weights
with torch.no_grad():
for idx, param in enumerate(model.parameters()):
if param.requires_grad and not param.grad is None:
param.data -= (param.grad*3e-5)
 
# real data
model.eval()
q_embs = []
p_embs = []
for k in range(0, len(q['input_ids']), chunk_size):
with torch.no_grad():
q_emb = model(input_ids=q['input_ids'][k:k+chunk_size].to("cuda:0"),).cpu()
p_emb = model(input_ids=p['input_ids'][k:k+chunk_size].to("cuda:0"),).cpu()
q_embs.append(q_emb)
p_embs.append(p_emb)
q_embs = torch.cat(q_embs, 0)
p_embs = torch.cat(p_embs, 0)
r_loss = default_loss(q_embs.to("cuda:0"), p_embs.to("cuda:0"))
del q_embs, p_embs
 
# distill backward
if count % 2 == 0:
d_grad = torch.autograd.grad(inputs=[x1.to("cuda:0")],#, x2.to("cuda:0")],
outputs=loss,
grad_outputs=r_loss)
indexes = q_indexes
else:
d_grad = torch.autograd.grad(inputs=[x2.to("cuda:0")],
outputs=loss,
grad_outputs=r_loss)
indexes = p_indexes
loss.detach()
r_loss.detach()
 
grads = torch.zeros(distilled_data.data.shape) # lbl, 10, 100, 768
for i, k in enumerate(indexes):
grads[train_labels[i], k, :, :] = grads[train_labels[i], k, :, :].to("cuda:0") 
+ d_grad[0][i, :, :]
distilled_data.data.grad = grads
data_optimizer.step()
data_optimizer.zero_grad(set_to_none=True)
 
model.load_state_dict(state_dict)
model_optimizer.step()
model_optimizer.zero_grad(set_to_none=True)
 
if count % 10 == 0:
print("Count:", count ,"| Data:", round(loss.item(), 4), "| Model:", 
round(r_loss.item(),4), "| Time:", round(time.time() - lap, 4))
# print()
lap = time.time()
 
if count % 100 == 0:
torch.save(model.state_dict(), model_PATH)
torch.save(distilled_data.state_dict(), distill_PATH)
 
if loss < 0.1 and r_loss < 1:
times += 1
 
if times > 100:
print("Training Took:", round(time.time() - start), "seconds")
print("n*********** Training Complete *************")
return
del loss, r_loss, grads, q, p, train_labels, dq, dp, x1, x2, state_dict
 
print("Training Took:", round(time.time() - start), "seconds")
print("n*********** Training Complete *************")