Sergey O

bugfix for latest jax version (jnp.take() in version jax>=0.3.8 return nans)

1

alphafold+multimer+templates returns NAN, starting with jax version 0.3.8 @YoshitakaMo traced it down to def batched_gather() in alphafold/model/utils.py you need to change: ```jnp.take(p, i, axis=axis)``` to ```jnp.take(p, i, axis=axis, mode="clip")``` https://jax.readthedocs.io/en/latest/changelog.html#jax-0-3-8-april-29-2022...

jax.lax.cond() signficantly slower when called inside jax.lax.scan()

3

I've been using cond() inside my scan() loop, and noticed it was much slower compared to just replacing it with `cond * true + (1-cond) * false`. Outside of scan(),...

convert_unifold_to_alphafold.py ?

8

Is there a convert_unifold_to_alphafold.py script? 😄

`frame_aligned_point_error()` function does not have a `loss_unit_distance` argument. Should `loss_unit_distance` be `length_scale` ?

hk.Conv2DTranspose takes FOREVER to initialize and compile

1

Not sure if this is a jax thing or dm-haiku... but recently I've been trying to use Conv2DTranspose in my model, and even for very simple case... it takes forever...

Any chance you guys might have the opposite script that converts unifold weights to alphafold? We would like to explore options to include weights in ColabFold.

the ccmpred.parameter_handling.structured_to_linear() function linearizes x_single and x_pair. by default it is assumed that x_single is a vector of size (L,20), if the flag nogapstate=False, it's assumed x_single is (L,21). So...