[Misc] Consolidate pooler config overrides (#10351)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
This commit is contained in:
Cyrus Leung
2024-11-15 06:59:00 +00:00
committed by GitHub
parent 2ec8827288
commit 2ac6d0e75b
7 changed files with 141 additions and 190 deletions
+30 -24
View File
@@ -63,14 +63,14 @@ class Pooler(nn.Module):
return cls(
pooling_type=PoolingType[pooler_config.pooling_type]
if pooler_config.pooling_type is not None else pooling_type,
normalize=pooler_config.pooling_norm
if pooler_config.pooling_norm is not None else normalize,
softmax=pooler_config.pooling_softmax
if pooler_config.pooling_softmax is not None else softmax,
step_tag_id=pooler_config.pooling_step_tag_id
if pooler_config.pooling_step_tag_id is not None else step_tag_id,
returned_token_ids=pooler_config.pooling_returned_token_ids
if pooler_config.pooling_returned_token_ids is not None else
normalize=pooler_config.normalize
if pooler_config.normalize is not None else normalize,
softmax=pooler_config.softmax
if pooler_config.softmax is not None else softmax,
step_tag_id=pooler_config.step_tag_id
if pooler_config.step_tag_id is not None else step_tag_id,
returned_token_ids=pooler_config.returned_token_ids
if pooler_config.returned_token_ids is not None else
returned_token_ids,
)
@@ -94,10 +94,14 @@ class Pooler(nn.Module):
pooled_data = hidden_states[last_token_flat_indices]
elif self.pooling_type == PoolingType.ALL:
offset = 0
pooled_data = []
pooled_data_lst = []
for prompt_len in prompt_lens:
pooled_data.append(hidden_states[offset:offset + prompt_len])
pooled_data_i = hidden_states[offset:offset + prompt_len]
pooled_data_lst.append(pooled_data_i)
offset += prompt_len
pooled_data = torch.stack(pooled_data_lst)
elif self.pooling_type == PoolingType.MEAN:
# Calculate mean pooling
cumsum = torch.cumsum(hidden_states, dim=0)
@@ -110,24 +114,26 @@ class Pooler(nn.Module):
cumsum[end_indices - 1] - cumsum[start_indices] +
hidden_states[start_indices]) / prompt_lens.unsqueeze(1)
elif self.pooling_type == PoolingType.STEP:
if self.returned_token_ids is not None and len(
self.returned_token_ids) > 0:
logits = hidden_states[:,
self.returned_token_ids].softmax(dim=-1)
else:
logits = hidden_states.softmax(dim=-1)
returned_token_ids = self.returned_token_ids
if returned_token_ids is not None and len(returned_token_ids) > 0:
hidden_states = hidden_states[:, returned_token_ids]
logits = hidden_states.softmax(dim=-1)
step_tag_id = self.step_tag_id
offset = 0
pooled_data = []
pooled_data_lst = []
for prompt_len, seq_data_i in zip(
prompt_lens, pooling_metadata.seq_data.values()):
if self.step_tag_id is None:
pooled_data.append(logits[offset:offset + prompt_len])
else:
step_idxs = torch.tensor(
seq_data_i.prompt_token_ids) == self.step_tag_id
pooled_data.append(logits[offset:offset +
prompt_len][step_idxs])
pooled_data_i = logits[offset:offset + prompt_len]
if step_tag_id is not None:
token_ids = torch.tensor(seq_data_i.prompt_token_ids)
pooled_data_i = pooled_data_i[token_ids == step_tag_id]
offset += prompt_len
pooled_data_lst.append(pooled_data_i)
pooled_data = torch.stack(pooled_data_lst)
else:
raise ValueError(f"Invalid pooling type: {self.pooling_type}")