mythrex / deep_dream_textual_data Goto Github PK

Hey, Great push.

I was doing boosting on activations for Bert-pre-trained embeddings.
Results::::

Final nodes output: tensor([[ 1.3311, -0.3895]], grad_fn=)
actual: negative Predictions: negative
Dreaming!!
Predictions after boosting:
Final nodes output: tensor([[62.9412, -0.3895]], grad_fn=)
actual: negative Predictions: negative

---

Final nodes output: tensor([[-0.5708, -0.1172]], grad_fn=)
actual: positive Predictions: positive
Dreaming!!
Predictions after boosting:
Final nodes output: tensor([[ -0.5708, 199.3414]], grad_fn=)
actual: positive Predictions: positive

---

If you see the results after embeddings boosting, its certainly increasing, but only the actual label part.
for eg: for 1st example its going from 1.3111 -> 62.9412 but positive part is staying the same.
why this is the case?

Note: 1-> positive and 0 -> negative.

Let me know why this is the case, or am I doing something wrong? Thankyou!