[UPDATE]. Clean Tr.

combopt/README.md

@@ -16,7 +16,8 @@ Download all submodules
 
 > git submodule update --init --recursive
 
-[Optional] If you plan to use the Antibody design task, install AbsolutNoLib by following the instructions from https://github.com/csi-greifflab/Absolut
+[Optional] If you plan to use the Antibody design task, install AbsolutNoLib by following the instructions
+from https://github.com/csi-greifflab/Absolut
 
 Create a virtual environment and activate it
 
@@ -42,7 +43,7 @@ Install the package itself
 - Random TSP
 
 ### Real-world
-- Antibody Design (Needs to download `Absolut!` ~ 39GB)  #TODO: indicate how to install absolute
+- Antibody Design (Needs to download `Absolut!`)
 - RNA Inverse Folding
 - EDA Sequence Optimisation (AIG sequence optimisation)
 - EDA Sequence and Parameter Optimisation (AIG sequence and parameter optimisation)

combopt/comb_opt/acq_optimizers/acq_optimizer_base.py

@@ -43,7 +43,7 @@ def optimize(self,
                  **kwargs
                  ) -> torch.Tensor:
         """
-        Function used to optimise the acquisition function. Should return a 2D tensor with shape
+        Function used to optimize the acquisition function. Should return a 2D tensor with shape
         (n_suggestions, n_dims), where n_dims is the dimensionality of x.
 
         If an optimizer does not support return batches of data, this can be handled by imposing with "assert

combopt/comb_opt/acq_optimizers/genetic_algorithm_acq_optimizer.py

@@ -142,7 +142,7 @@ def __init__(self,
         self.ga_allow_repeating_x = ga_allow_repeating_x
 
         assert self.search_space.num_nominal + self.search_space.num_ordinal == self.search_space.num_dims, \
-            'The Categorical GA acq optimiser currently only supports nominal and ordinal variables'
+            'The Categorical GA acq optimizer currently only supports nominal and ordinal variables'
 
     def optimize(self, x: torch.Tensor,
                  n_suggestions: int,

combopt/comb_opt/acq_optimizers/interleaved_search_acq_optimizer.py

@@ -59,7 +59,7 @@ def __init__(self,
         self.inverse_mapping = [(self.numeric_dims + self.search_space.nominal_dims).index(i) for i in
                                 range(self.search_space.num_dims)]
 
-        # Determine the learning rate used to optimise numeric variables if needed
+        # Determine the learning rate used to optimize numeric variables if needed
         if len(self.numeric_dims) > 0:
             if num_lr is None:
                 if self.search_space.num_disc > 0:
@@ -194,6 +194,8 @@ def _optimize(self,
                             is_valid = True
                         else:
                             tol_ -= 1
+                        if tol_ < 0:
+                            break
                     if tol_ < 0:
                         break
 

combopt/comb_opt/acq_optimizers/local_search_acq_optimizer.py

@@ -117,8 +117,6 @@ def _optimize(self,
                                                                                     search_space=self.search_space,
                                                                                     tr_manager=tr_manager,
                                                                                     n_points=self.n_random_vertices,
-                                                                                    is_numeric=self.is_numeric,
-                                                                                    is_mixed=self.is_mixed,
                                                                                     numeric_dims=self.numeric_dims,
                                                                                     discrete_choices=self.discrete_choices,
                                                                                     max_n_perturb_num=self.max_n_perturb_num,
@@ -184,8 +182,6 @@ def _optimize(self,
                                                               search_space=self.search_space,
                                                               tr_manager=tr_manager,
                                                               n_points=1,
-                                                              is_numeric=self.is_numeric,
-                                                              is_mixed=self.is_mixed,
                                                               numeric_dims=self.numeric_dims,
                                                               discrete_choices=self.discrete_choices,
                                                               max_n_perturb_num=self.max_n_perturb_num,

combopt/comb_opt/factory.py

@@ -208,9 +208,14 @@ def task_factory(task_name: str, dtype: torch.dtype = torch.float32, **kwargs) -
 
     elif task_name == 'antibody_design':
         if 'antigen' not in kwargs:
-            print('Target antigen not specified. Using antigen 1ADQ_A.')
-        task = CDRH3Design(antigen=kwargs.get('antigen', '1ADQ_A'), cdrh3_length=kwargs.get('cdrh3_length', 11),
-                           num_cpus=kwargs.get('num_cpus', 1), first_cpu=kwargs.get('first_cpu', 0))
+            print('Target antigen not specified. Using antigen 2DD8_S.')
+        task = CDRH3Design(
+            antigen=kwargs.get('antigen', '2DD8_S'),
+            cdrh3_length=kwargs.get('cdrh3_length', 11),
+            num_cpus=kwargs.get('num_cpus', 1),
+            first_cpu=kwargs.get('first_cpu', 0),
+            absolut_dir=kwargs.get('absolut_dir', None)
+        )
         search_space = search_space_factory('antibody_design', dtype, cdrh3_length=kwargs.get('cdrh3_length', 11))
 
     elif task_name == "rna_inverse_fold":

combopt/comb_opt/models/gp/combo_gp.py

@@ -461,7 +461,7 @@ def cholesky_update(self, constmean: float, log_amp: float, log_beta: torch.Floa
         self.gram_mat_update(constmean, log_amp, log_beta, log_likelihood_noise)
 
         eye_mat = torch.diag(self.gram_mat.new_ones(self.gram_mat.size(0)))
-        for jitter_const in [0, 1e-8, 1e-7, 1e-6, 1e-5, 1e-4, 1e-3]:
+        for jitter_const in [0, 1e-8, 1e-7, 1e-6, 1e-5, 1e-4, 1e-3, 1e-2]:
             chol_jitter = torch.trace(self.gram_mat).item() * jitter_const
             try:
                 # cholesky is lower triangular matrix

combopt/comb_opt/optimizers/__init__.py

@@ -14,7 +14,7 @@
 from comb_opt.optimizers.casmopolitan import Casmopolitan
 from comb_opt.optimizers.cocabo import CoCaBO
 from comb_opt.optimizers.combo import COMBO
-from comb_opt.optimizers.genetic_algorithm import PymooGeneticAlgorithm
+from comb_opt.optimizers.genetic_algorithm import PymooGeneticAlgorithm, GeneticAlgorithm
 from comb_opt.optimizers.local_search import LocalSearch
 from comb_opt.optimizers.multi_armed_bandit import MultiArmedBandit
 from comb_opt.optimizers.optimizer_base import OptimizerBase

combopt/comb_opt/optimizers/bocs.py

@@ -27,7 +27,9 @@ class BOCS(BoBase):
     @property
     def name(self) -> str:
         if self.use_tr:
-            name = f'LR ({self.model_estimator}) - Tr-Based LS acq optim'
+            name = f'LR ({self.model_estimator}) - Tr-Based SA acq optim'
+        elif self.model_estimator == "sparse_horseshoe":  # Standard BOCS
+            name = "BOCS"
         else:
             name = f'BOCS ({self.model_estimator})'
         return name
@@ -70,7 +72,7 @@ def __init__(self,
             warning_message = 'This is the general form implementation of BOCS (see Appendix A of ' + \
                               'https://arxiv.org/abs/1806.08838), which differs from the standard implementation ' + \
                               'for purely binary problems. The differences are: (1) binary variables are ' + \
-                              'represented by their one hot encoding, and (2) SA is used to optimise the acquisition' + \
+                              'represented by their one hot encoding, and (2) SA is used to optimize the acquisition' + \
                               'in place of SDP.'
             warnings.warn(warning_message, category=UserWarning)
 

combopt/comb_opt/optimizers/cocabo.py

@@ -45,6 +45,8 @@ def name(self) -> str:
                 name = f'GP ({self.model_numeric_kernel_name}) - MAB acq optim'
             elif self.is_nominal:
                 name = f'GP ({self.model_cat_kernel_name}) - MAB acq optim'
+            else:
+                raise ValueError()
 
         return name
 
@@ -54,7 +56,7 @@ def __init__(self,
                  model_numeric_kernel_name: str = 'mat52',
                  model_num_kernel_ard: bool = True,
                  model_num_kernel_lengthscale_constr: Optional[Interval] = None,
-                 model_cat_kernel_name='transformed_overlap',
+                 model_cat_kernel_name='overlap',
                  model_cat_kernel_ard: bool = True,
                  model_cat_kernel_lengthscale_constr: Optional[Interval] = None,
                  model_noise_prior: Optional[Prior] = None,

combopt/comb_opt/optimizers/combo.py

@@ -46,7 +46,7 @@ def __init__(self,
                  acq_optim_n_greedy_ascent_init: int = 20,
                  acq_optim_n_spray: int = 10,
                  acq_optim_max_n_ascent: float = float('inf'),
-                 use_tr: bool = True,
+                 use_tr: bool = False,
                  tr_restart_acq_name: str = 'lcb',
                  tr_restart_n_cand: Optional[int] = None,
                  tr_min_nominal_radius: Optional[Union[int, float]] = None,

combopt/comb_opt/optimizers/genetic_algorithm.py

@@ -12,6 +12,10 @@
 import numpy as np
 import pandas as pd
 import torch
+
+from pymoo.config import Config
+Config.warnings['not_compiled'] = False
+
 from pymoo.algorithms.base.genetic import GeneticAlgorithm
 from pymoo.algorithms.soo.nonconvex.ga import FitnessSurvival
 from pymoo.algorithms.soo.nonconvex.ga import comp_by_cv_and_fitness
@@ -222,9 +226,9 @@ class CategoricalGeneticAlgorithm(OptimizerBase):
     @property
     def name(self) -> str:
         if self.tr_manager is not None:
-            name = 'Tr-based Categorical Genetic Algorithm'
+            name = 'Tr-based Genetic Algorithm'
         else:
-            name = 'Categorical Genetic Algorithm'
+            name = 'Genetic Algorithm'
         return name
 
     def __init__(self,
@@ -603,7 +607,7 @@ def _mutate(self, x: torch.Tensor) -> torch.Tensor:
 
 class GeneticAlgorithm(OptimizerBase):
     """
-    A Genetic Algorithm (GA) optimiser that determines which exact GA algorithm to use based on the variable types in
+    A Genetic Algorithm (GA) optimizer that determines which exact GA algorithm to use based on the variable types in
     the search space. If the search space contains only nominal variables, an elitist GA algorithm will be used. If the
     search space contains any other variable type combinations, the Mixed Variable GA from pymoo will be used (see
     https://pymoo.org/customization/mixed.html). On purely combinatorial problems, the elitist GA algorithm can

combopt/comb_opt/optimizers/mix_and_match/__init__.py

@@ -8,13 +8,17 @@
 # PARTICULAR PURPOSE. See the MIT License for more details.
 
 from comb_opt.optimizers.mix_and_match.gp_diff_ker_ga_acq_optim import GpDiffusionGaAcqOptim
+from comb_opt.optimizers.mix_and_match.gp_diff_ker_is_acq_optim import GpDiffusionIsAcqOptim
 from comb_opt.optimizers.mix_and_match.gp_diff_ker_sa_acq_optim import GpDiffusionSaAcqOptim
-from comb_opt.optimizers.mix_and_match.gp_diff_ker_is_acq_optim import GpDiffusionTrLsAcqOptim
+from comb_opt.optimizers.mix_and_match.gp_o_ker_ga_acq_optim import GpOGaAcqOptim
+from comb_opt.optimizers.mix_and_match.gp_o_ker_is_acq_optim import GpOIsAcqOptim
+from comb_opt.optimizers.mix_and_match.gp_o_ker_ls_acq_optim import GpOLsAcqOptim
+from comb_opt.optimizers.mix_and_match.gp_o_ker_sa_acq_optim import GpOSaAcqOptim
 from comb_opt.optimizers.mix_and_match.gp_ssk_ker_ls_acq_optim import GpSskLsAcqOptim
 from comb_opt.optimizers.mix_and_match.gp_ssk_ker_sa_acq_optim import GpSskSaAcqOptim
 from comb_opt.optimizers.mix_and_match.gp_to_ker_ls_acq_optim import GpToLsAcqOptim
 from comb_opt.optimizers.mix_and_match.gp_to_ker_ga_acq_optim import GpToGaAcqOptim
 from comb_opt.optimizers.mix_and_match.gp_to_ker_sa_acq_optim import GpToSaAcqOptim
 from comb_opt.optimizers.mix_and_match.lr_ga_acq_optim import LrGaAcqOptim
-from comb_opt.optimizers.mix_and_match.lr_ls_acq_optim import LrLsAcqOptim
 from comb_opt.optimizers.mix_and_match.lr_is_acq_optim import LrIsAcqOptim
+from comb_opt.optimizers.mix_and_match.lr_ls_acq_optim import LrLsAcqOptim

combopt/comb_opt/optimizers/mix_and_match/gp_diff_ker_ga_acq_optim.py

@@ -48,7 +48,7 @@ def __init__(self,
                  acq_optim_ga_num_elite: int = 10,
                  acq_optim_ga_store_x: bool = False,
                  acq_optim_ga_allow_repeating_x: bool = True,
-                 use_tr: bool = True,
+                 use_tr: bool = False,
                  tr_restart_acq_name: str = 'lcb',
                  tr_restart_n_cand: Optional[int] = None,
                  tr_min_nominal_radius: Optional[Union[int, float]] = None,

combopt/comb_opt/optimizers/mix_and_match/gp_diff_ker_is_acq_optim.py

@@ -22,7 +22,7 @@
 from comb_opt.utils.graph_utils import laplacian_eigen_decomposition
 
 
-class GpDiffusionTrLsAcqOptim(BoBase):
+class GpDiffusionIsAcqOptim(BoBase):
 
     @property
     def name(self) -> str:
@@ -48,7 +48,7 @@ def __init__(self,
                  acq_optim_num_optimizer: str = 'sgd',
                  acq_optim_num_lr: Optional[float] = 1e-3,
                  acq_optim_nominal_tol: int = 100,
-                 use_tr: bool = True,
+                 use_tr: bool = False,
                  tr_restart_acq_name: str = 'lcb',
                  tr_restart_n_cand: Optional[int] = None,
                  tr_min_nominal_radius: Optional[Union[int, float]] = None,
@@ -168,6 +168,6 @@ def __init__(self,
 
         self.use_tr = use_tr
 
-        super(GpDiffusionTrLsAcqOptim, self).__init__(search_space, n_init, model, acq_func, acq_optim, tr_manager,
+        super(GpDiffusionIsAcqOptim, self).__init__(search_space, n_init, model, acq_func, acq_optim, tr_manager,
                                                       dtype,
                                                       device)

combopt/comb_opt/optimizers/mix_and_match/gp_o_ker_ga_acq_optim.py

@@ -57,7 +57,7 @@ def __init__(self,
                  acq_optim_ga_num_elite: int = 10,
                  acq_optim_ga_store_x: bool = False,
                  acq_optim_ga_allow_repeating_x: bool = True,
-                 use_tr: bool = True,
+                 use_tr: bool = False,
                  tr_restart_acq_name: str = 'lcb',
                  tr_restart_n_cand: Optional[int] = None,
                  tr_min_num_radius: Optional[Union[int, float]] = None,

combopt/comb_opt/optimizers/mix_and_match/gp_o_ker_is_acq_optim.py

@@ -58,7 +58,7 @@ def __init__(self,
                  acq_optim_num_optimizer: str = 'sgd',
                  acq_optim_num_lr: Optional[float] = 1e-3,
                  acq_optim_nominal_tol: int = 100,
-                 use_tr: bool = True,
+                 use_tr: bool = False,
                  tr_restart_acq_name: str = 'lcb',
                  tr_restart_n_cand: Optional[int] = None,
                  tr_min_num_radius: Optional[Union[int, float]] = None,
@@ -142,7 +142,7 @@ def __init__(self,
                                         numeric_kernel_name='mat52',
                                         numeric_kernel_use_ard=model_num_kernel_ard,
                                         numeric_lengthscale_constraint=model_num_kernel_lengthscale_constr,
-                                        nominal_kernel_name='transformed_overlap',
+                                        nominal_kernel_name='overlap',
                                         nominal_kernel_use_ard=model_cat_kernel_ard,
                                         nominal_lengthscale_constraint=model_cat_kernel_lengthscale_constr)
 

combopt/comb_opt/optimizers/mix_and_match/gp_o_ker_ls_acq_optim.py

@@ -56,7 +56,7 @@ def __init__(self,
                  acq_optim_n_greedy_ascent_init: int = 20,
                  acq_optim_n_spray: int = 10,
                  acq_optim_max_n_ascent: float = float('inf'),
-                 use_tr: bool = True,
+                 use_tr: bool = False,
                  tr_restart_acq_name: str = 'lcb',
                  tr_restart_n_cand: Optional[int] = None,
                  tr_min_num_radius: Optional[Union[int, float]] = None,

combopt/comb_opt/optimizers/mix_and_match/gp_o_ker_sa_acq_optim.py

@@ -54,7 +54,7 @@ def __init__(self,
                  acq_optim_num_iter: int = 200,
                  acq_optim_init_temp: int = 1,
                  acq_optim_tolerance: int = 100,
-                 use_tr: bool = True,
+                 use_tr: bool = False,
                  tr_restart_acq_name: str = 'lcb',
                  tr_restart_n_cand: Optional[int] = None,
                  tr_min_num_radius: Optional[Union[int, float]] = None,

combopt/comb_opt/optimizers/mix_and_match/gp_to_ker_ga_acq_optim.py

@@ -57,7 +57,7 @@ def __init__(self,
                  acq_optim_ga_num_elite: int = 10,
                  acq_optim_ga_store_x: bool = False,
                  acq_optim_ga_allow_repeating_x: bool = True,
-                 use_tr: bool = True,
+                 use_tr: bool = False,
                  tr_restart_acq_name: str = 'lcb',
                  tr_restart_n_cand: Optional[int] = None,
                  tr_min_num_radius: Optional[Union[int, float]] = None,

combopt/comb_opt/optimizers/mix_and_match/gp_to_ker_ls_acq_optim.py

@@ -56,7 +56,7 @@ def __init__(self,
                  acq_optim_n_greedy_ascent_init: int = 20,
                  acq_optim_n_spray: int = 10,
                  acq_optim_max_n_ascent: float = float('inf'),
-                 use_tr: bool = True,
+                 use_tr: bool = False,
                  tr_restart_acq_name: str = 'lcb',
                  tr_restart_n_cand: Optional[int] = None,
                  tr_min_num_radius: Optional[Union[int, float]] = None,

combopt/comb_opt/optimizers/mix_and_match/gp_to_ker_sa_acq_optim.py

@@ -54,7 +54,7 @@ def __init__(self,
                  acq_optim_num_iter: int = 200,
                  acq_optim_init_temp: int = 1,
                  acq_optim_tolerance: int = 100,
-                 use_tr: bool = True,
+                 use_tr: bool = False,
                  tr_restart_acq_name: str = 'lcb',
                  tr_restart_n_cand: Optional[int] = None,
                  tr_min_num_radius: Optional[Union[int, float]] = None,

combopt/comb_opt/optimizers/simulated_annealing.py

@@ -286,3 +286,14 @@ def sample_unseen_nominal_neighbour(self, x_nominal: torch.Tensor):
             x_nominal_neighbour = x_nominal_neighbour.view(-1)
 
         return x_nominal_neighbour
+
+    def fill_field_after_pkl_load(self, search_space: SearchSpace, **kwargs):
+        """ As some elements are not pickled, need to reinstantiate them """
+        self.search_space = search_space
+
+    def __getstate__(self):
+        d = dict(self.__dict__)
+        to_remove = ["search_space"]  # fields to remove when pickling this object
+        for attr in to_remove:
+            del d[attr]
+        return d

combopt/comb_opt/tasks/antibody_design/cdrh3_design.py

@@ -9,6 +9,7 @@
 
 import os
 import subprocess
+from typing import Optional
 
 import numpy as np
 import pandas as pd
@@ -24,7 +25,8 @@ class CDRH3Design(TaskBase):
     def name(self) -> str:
         return f'{self.antigen} Antibody Design'
 
-    def __init__(self, antigen: str = '1ADQ_A', cdrh3_length: int = 11, num_cpus: int = 10, first_cpu: int = 0):
+    def __init__(self, antigen: str = '1ADQ_A', cdrh3_length: int = 11, num_cpus: int = 10, first_cpu: int = 0,
+                 absolut_dir: Optional[str] = None):
         super(CDRH3Design, self).__init__()
         self.num_cpus = num_cpus
         self.first_cpu = first_cpu
@@ -35,15 +37,19 @@ def __init__(self, antigen: str = '1ADQ_A', cdrh3_length: int = 11, num_cpus: in
         self.amino_acid_to_idx = {aa: i for i, aa in enumerate(self.amino_acids)}
         self.idx_to_amino_acid = {value: key for key, value in self.amino_acid_to_idx.items()}
 
-        self.AbsolutNoLib_dir = get_AbsolutNoLib_dir()
+        self.AbsolutNoLib_dir = get_AbsolutNoLib_dir(absolut_dir)
         self.valid_antigens = get_valid_antigens(self.AbsolutNoLib_dir)
+        self.need_to_check_precomputed_antigen_structure = True
         assert antigen in self.valid_antigens, f'Specified antigen is not valid. Please choose of from: \n\n {self.valid_antigens}'
-        download_precomputed_antigen_structure(self.AbsolutNoLib_dir, self.antigen)
 
     def evaluate(self, x: pd.DataFrame) -> np.ndarray:
 
         assert os.path.exists(os.path.join(self.AbsolutNoLib_dir, 'antigen_data', f'{self.antigen}'))
 
+        if self.need_to_check_precomputed_antigen_structure:
+            download_precomputed_antigen_structure(self.AbsolutNoLib_dir, self.antigen)
+            self.need_to_check_precomputed_antigen_structure = False
+
         # Change working directory
         current_dir = os.getcwd()
         os.chdir(os.path.join(self.AbsolutNoLib_dir, 'antigen_data', f'{self.antigen}'))