Bert-as-service: Classifier Predictions

[loretoparisi]

Thanks for this service, it works like a charm and better than Google's!
How it difficult to support classifier predictions for the classification task here from a pre-trained model?

[hanxiao]

Thanks! Supporting classification isn't hard. I will implement it as a feature in the future.

For now, I'd like to implement more pooling strategies.

[loretoparisi]

@hanxiao thank you that would be awesome! I was looking at the bert repo classification task here https://github.com/google-research/bert/blob/60454702590a6c69bd45c5d4258c7e17b8a3e1da/run_classifier.py
By the way this is far more complex how it could actually be.

[mortimer125]

hi, could I put this sentence vector with pooling_strategy being CLS_TOKEN to a two-layers NN to do classification tasks?

[hanxiao]

@loretoparisi I write one based on tf.estimator.DNNClassifier, checkout example5 is this something you have in mind?

[loretoparisi]

@hanxiao amazing 💯 🥇 This is exactly what I was looking for! The reference CAIL dataset is also a very good example for both the text embedding and labels encoding.
Closing! Thank you!

[hanxiao]

FYI, I made a visualization of different BERT layers, may help you understand the pool_layer and pool_strategy. See: https://github.com/hanxiao/bert-as-service#q-so-which-layer-and-which-pooling-strategy-is-the-best