evoldoers / machineboss Goto Github PK

e.g. RNN outputs

logspace

• Implement log_sum_exp directly in generated code
• Convert probability profiles to sequences on-the-fly

• Change TransWeight -> WeightFunc
• typedef map<string,json> FuncDefs
• Change Params in WeightAlgebra to FuncDefs
• use change of variables instead of Lagrange multipliers

Use Poreover example data

To reproduce:

cd src/nano/test
biomake -B seq1-test1.decode |& less -r

equivalent to calculating sequence probability using CTC (Graves et al)

All cases (conditioned on input, output, both, or neither) can be reduced to the null machine (empty input & output) by appropriate compositions with generators/acceptors.
Then, that null machine can be solved by matrix inversion. In fact, the machine is acyclic & topologically sorted at this point, so the inversion can be done efficiently by calculating the sum-over-paths likelihood from each state to the end state.

function (subclass/compatible with #8) capable of

• eval
• deriv
• constraint (Lagrange multiplier)

Start with bitnoise test

As it currently does; but keep the names

As with #34, Machine-creating operations should propagate/merge their ParamDefs

this will make it easier to do 1D^2 (intersection of generator w/two parallel machines)

As #33, but using the compiler.

• Uses GSL maximizer with constrained parameter prior
• requires #11 #12 #13 #15

It should be proportional to the number of backward silent transitions that have been eliminated, not the number of states visited

• feasibility calculations (CPU, memory)
• test data (c.f. #32)

• feasibility calculations (CPU, memory)
• training
• search

and/or skip estimation of TraceParams

• logspace_eval = log(f(x1,x2...))
• logspace_deriv = dlog(f(x1,x2...))/dlog(xn)

Subclass of MachineCounts initialized from ForwardMatrix + BackwardMatrix. Needs #13

Only those into cpt_end states. (And specifically not those into end.)
This is causing discrepancies between Viterbi matrix score & path score.
Need to explicitly flag some states as softplus-able.

All unary & binary Machine operations should propagate/merge Constraints

Needs #14, #17

e.g.
nanomachine -R seq1-test1.raw -M test1.json -v8 -C -d TraceViterbiMatrix -f .00001
should give same results as
nanomachine -R seq1-test1.raw -M test1.json -v8 -C -d TraceViterbiMatrix -g 1
but doesn't

With evaluation (when loaded)
Special Machine constructor

Like #43, but compiled.

Compiled & uncompiled. To test gap handling.

• container for MachineTransitionScore
• requires #10

Interpreted, compiled

This means porting EvaluatedMachine, DPMatrix, ViterbiMatrix

Input and/or output can be a sequence instead of a profile

e.g. for protein to RNN, input is a sequence (protein) and output is a profile

Serializing final model to JSON is rate-limiting step.

Traceback yields alignments - needs #14

• Envelope data structure
• Initialize from MachinePath
• connected method (tests all start/end intervals overlap & touch corners)
• Adapt containers to use packed storage
• Adapt iterators (this is really #18)
• SeqPair path property (can be empty)

Banded initializer is now #66