Pool_factor not changing embedding number

[John42506176Linux]

Discussed in #363

^{Originally posted by John42506176Linux September 7, 2024}
Hi, Colbert community :),

I'm currently testing out Colbert, and I was curious how I could test multiple degrees of token pooling similar to this https://www.answer.ai/posts/colbert-pooling.html. Currently, when changing pool_factor I haven't seen any changes in the number of embeddings.

Here is the code being used.

Version: 0.2.20

from colbert.modeling.checkpoint import Checkpoint
from colbert.infra import ColBERTConfig

answer_ai = Checkpoint("answerdotai/answerai-colbert-small-v1", colbert_config=ColBERTConfig())

vectors = answer_ai.docFromText(documents,bsize=2, pool_factor=3, showprogress=True)[0]

[jbellis]

pooling only executes on the keep_dims="flatten" path of docFromText. Related: #366

[jessiejuachon]

pooling only executes on the keep_dims="flatten" path of docFromText. Related: #366

I tried setting pool_factor=2, keep_dims='flatten', and bsize=1 just so it would enter path, but tensor.shape is the same.