masking from child sequences; DNSM model bugfix; Yun branch lengths

[matsen]

• take the mask from the child sequence
• fix bug in DNSM models which was interpreting "N" as an ambiguous AA
• implementing Yun-style branch lengths
• eliminating the pre-training before branch length optimization

[matsen]

Yun's derivation:

[mmjohn]

Looks great!

Running make test in my epam conda env gives the following errors on both ermine and local installs:

===================================== short test summary info ======================================
ERROR tests/test_dnsm.py::test_crepe_roundtrip - NameError: name 'training_method' is not defined
ERROR tests/test_netam.py::test_write_output - NameError: name 'training_method' is not defined
ERROR tests/test_netam.py::test_standardize_model_rates - NameError: name 'training_method' is not defined

[matsen]

Thanks for catching the test failure!! That was a spur of the moment last decision. Fixed now.

[mmjohn]

Everything's working for me now

[matsen]

I decided to eliminate the short training before the first round of branch length optimization, which appeared here:

diff --git a/netam/framework.py b/netam/framework.py
index 42b348c..ec13b9d 100644
--- a/netam/framework.py
+++ b/netam/framework.py
@@ -656,9 +656,6 @@ class Burrito(ABC):
         else:
             raise ValueError(f"Unknown training method {training_method}")
         loss_history_l = []
-        self.mark_branch_lengths_optimized(0)
-        loss_history_l.append(self.train(3))
         optimize_branch_lengths()
         self.mark_branch_lengths_optimized(0)
         for cycle in range(cycle_count):

It was making the non-fixed SHM models harder to train, and actually an untrained model isn't so crazy scale-wise for the DNSM:

(Erick, this is here)