Adds method to read the pooling types from model's files (#9506)

Signed-off-by: Flavia Beo <flavia.beo@ibm.com>
Signed-off-by: Max de Bayser <mbayser@br.ibm.com>
Co-authored-by: Max de Bayser <mbayser@br.ibm.com>
This commit is contained in:
Flávia Béo
2024-11-07 08:42:40 +00:00
committed by GitHub
co-authored by Max de Bayser
parent e036e527a0
commit aa9078fa03
10 changed files with 342 additions and 25 deletions
+13 -1
View File
@@ -16,6 +16,7 @@ class PoolingType(IntEnum):
ALL = 1
CLS = 2
STEP = 3
MEAN = 4
class Pooler(nn.Module):
@@ -27,7 +28,7 @@ class Pooler(nn.Module):
3. Returns structured results as `PoolerOutput`.
Attributes:
pooling_type: The type of pooling to use (LAST, ALL, CLS).
pooling_type: The type of pooling to use.
normalize: Whether to normalize the pooled data.
"""
@@ -97,6 +98,17 @@ class Pooler(nn.Module):
for prompt_len in prompt_lens:
pooled_data.append(hidden_states[offset:offset + prompt_len])
offset += prompt_len
elif self.pooling_type == PoolingType.MEAN:
# Calculate mean pooling
cumsum = torch.cumsum(hidden_states, dim=0)
start_indices = torch.cat([
torch.tensor([0], device=hidden_states.device),
torch.cumsum(prompt_lens[:-1], dim=0)
])
end_indices = torch.cumsum(prompt_lens, dim=0)
pooled_data = (
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: