Feature/refactor for reduced complexity #11

tutrie · 2021-08-09T12:09:23Z

No description provided.

Refactor

2D working example gpytorch

adding MES

update

Refactor

up to date

Refactor

CUDA device agnostic implementation

irinaespejo · 2021-08-09T19:44:41Z

excursion/_diagnosis/__init__.py

+import numpy as np
+import itertools
+from .. import utils
+import torch
+
+torch.cuda.set_device(0)
+
+
+def classlabels(values, thresholds):
+    labels = np.zeros(values.shape)
+    for j in range(len(thresholds) - 1):
+        print("[{}, {}]".format(thresholds[j], thresholds[j + 1]))
+        within = np.logical_and(thresholds[j] < values, values < thresholds[j + 1])
+        labels[within] = j + 1
+    return labels
+
+
+def confusion_matrix(gps, scandetails):
+    thresholds = np.concatenate([[-np.inf], scandetails.thresholds, [np.inf]])
+    diagX = utils.mesh2points(
+        utils.mgrid(scandetails.rangedef), scandetails.rangedef[:, 2]
+    )
+
+    diagX = diagX[~scandetails.invalid_region(diagX)]
+
+    labels = list(range(1, len(thresholds)))
+
+    confusion_list, predlabels_list, truelabels_list = [], [], []
+    for i, gp in enumerate(gps):
+        predy = gps[i].predict(diagX)
+        truthy = scandetails.truth_functions[i](diagX)
+        predlabels = classlabels(predy, thresholds)
+        truelabels = classlabels(truthy, thresholds)
+
+        confusion_matrix = np.zeros((len(labels), len(labels)))
+        for pred, true in itertools.product(labels, labels):
+            print("pred {}/true {}".format(pred, true))
+            predlabels_when_true = predlabels[truelabels == true]
+            numerator = np.sum(np.where(predlabels_when_true == pred, 1, 0))
+            denominator = len(predlabels_when_true)
+            print("{}/{}".format(numerator, denominator))
+            confusion_matrix[true - 1, pred - 1] = numerator / denominator
+
+        predlabels_list.append(predlabels)
+        truelabels_list.append(truelabels)
+        confusion_list.append(confusion_matrix.tolist())
+    return confusion_list, predlabels_list, truelabels_list, diagX, labels


Unsure why the folder is called _diagnosis/ what does the underscore mean here?

irinaespejo · 2021-08-09T19:45:06Z

excursion/_diagnosis/__init__.py

+from .. import utils
+import torch
+
+torch.cuda.set_device(0)


sorry this line is legacy from my code, suggestion: delete

irinaespejo · 2021-08-09T19:45:55Z

excursion/_diagnosis/old_init_.py

+import numpy as np
+import logging
+
+log = logging.getLogger(__name__)
+
+def get_class_labels(y,class_labels,minmaxes):
+    assigned_labels = -np.ones_like(y)
+    for class_idx in class_labels:
+        class_selector = np.logical_and(y>=minmaxes[class_idx][0], y<minmaxes[class_idx][1])
+        assigned_labels[class_selector] = class_idx
+    assert np.min(assigned_labels) >= 0
+    return assigned_labels
+
+def get_status(testy,esti_y,class_labels,minmaxes):
+    true_labels = get_class_labels(testy,class_labels,minmaxes)
+    esti_labels = get_class_labels(esti_y,class_labels,minmaxes)
+    status = np.concatenate([esti_labels.reshape(-1,1),true_labels.reshape(-1,1)], axis=-1)
+    return status
+
+def get_confusion_matrices(statuses,class_labels):
+    confusion_matrix = -np.ones((len(statuses),2,2))
+    for sidx,sta in enumerate(statuses):
+        for i in class_labels:
+            for j in class_labels:
+                test_val  = i
+                truth_val = j
+                condition_on_truth = sta[sta[:,1]==truth_val]
+                is_test_condition_truth = condition_on_truth[condition_on_truth[:,0]==test_val]
+                prob = len(is_test_condition_truth)/len(condition_on_truth)
+                confusion_matrix[sidx][i][j] = prob #p(i|j)
+    return confusion_matrix
+
+
+def confusion_callback(X,y_list,gps,scandetails):
+    testX, testy_list = scandetails.testdata()
+    log.debug('computing confusion matrix based on {} testing points'.format(len(testX)))
+    estimatey_list = [gps[i].predict(testX) for i in range(len(scandetails.functions))]
+
+    class_labels = np.arange(len(scandetails.thresholds)+1)
+
+    boundaries = [-np.inf] + scandetails.thresholds + [np.inf]
+    minmaxes = [[boundaries[i],boundaries[i+1]] for i in class_labels]
+    statuses = [
+        get_status(yl,estimatey_list[i],class_labels,minmaxes)
+        for i,yl in enumerate(testy_list)
+    ]
+    matrices = get_confusion_matrices(statuses,class_labels)
+    topline = np.mean([np.mean(np.diag(cm)) for cm in matrices])
+    return {'t': topline, 'matrices': np.asarray(matrices).tolist()}
+
+def diagnose(X,y_list,gps, scandetails, callbacks = None):
+    callbacks = callbacks or {
+        'confusion': confusion_callback,
+        'npoints': lambda X,*args: len(X),
+    }
+    return {name: c(X,y_list,gps,scandetails) for name,c in callbacks.items()}


I assume this is legacy code from my repo. Let's remove it so that the PR is cleaner

irinaespejo · 2021-08-09T19:47:55Z

excursion/_diagnosis/old_samplers.py

+import numpy as np
+import itertools
+
+from .. import utils
+
+
+def regular_grid_generator(
+    scandetails, central_range=[5, 20], nsamples_per_grid=15, min_points_per_dim=2
+):
+    ndim = len(scandetails.rangedef[:, 2])
+    grids = set(
+        x
+        for y in range(*central_range)
+        for x in itertools.combinations_with_replacement([y, y - 1, y - 2], ndim)
+    )
+    grids = [g for g in grids if np.all(np.array(g) >= min_points_per_dim)]
+    grids = sorted(grids, key=lambda k: np.product(k))
+
+    def makegrids(sizes):
+        for s in range(nsamples_per_grid):
+            los = np.random.uniform(0, 0.1, size=ndim)
+            his = np.random.uniform(0.9, 1.0, size=ndim)
+            yield np.array(
+                np.meshgrid(
+                    *[np.linspace(los[i], his[i], sizes[i]) for i in range(ndim)]
+                )
+            )
+
+    for g in grids:
+        for mesh in makegrids(g):
+            X = utils.mesh2points(mesh, mesh.shape[1:])
+            for i in range(ndim):
+                vmin, vmax = (
+                    scandetails.rangedef[i][0],
+                    scandetails.rangedef[i][1],
+                )
+                X[:, i] = X[:, i] * (vmax - vmin) + vmin
+            X = X[~scandetails.invalid_region(X)]
+            yield X, g
+
+
+def latin_hypercube_generator(
+    scandetails, nsamples_per_npoints=50, point_range=[4, 100]
+):
+    ndim = len(scandetails.rangedef[:, 2])
+    import pyDOE
+
+    for npoints in range(*point_range):
+        for s in range(nsamples_per_npoints):
+            sample_n = npoints
+            while True:
+                print("sampling", sample_n)
+                X = pyDOE.lhs(ndim, samples=sample_n)
+
+                for i in range(ndim):
+                    vmin, vmax = (
+                        scandetails.rangedef[i][0],


This file is handy to preserve. I would rename it to samplers.py instead of old_samplers.py so that we know it is not legacy but actual useful code. Not sure if it should be under _diagnosis but fine for now.

irinaespejo

More comments on core files of the refactoring

irinaespejo · 2021-08-09T19:51:14Z

excursion/_optimize/__init__.py

+import numpy as np
+import datetime
+import torch
+import gpytorch
+
+from . import core
+
+torch.cuda.set_device(0)
+
+
+def gridsearch(gps, X, scandetails):
+    thresholds = [-np.inf] + scandetails.thresholds + [np.inf]
+    print("info_gain")
+    acqval = np.array(
+        [
+            core.info_gain(xtest, gps, thresholds, scandetails.meanX)
+            for xtest in scandetails.acqX
+        ]
+    )
+    print(acqval)
+    newx = None
+    for i, cacq in enumerate(scandetails.acqX[np.argsort(acqval)]):
+        if cacq.tolist() not in X.tolist():
+            print("taking new x. best non-existent index {} {}".format(i, cacq))
+            newx = cacq
+            return newx, acqval
+        else:
+            print("{} is good but already there".format(cacq))
+    print("returning None.. something must be wrong")
+    return None, None
+
+
+def gridsearch_gpytorch(gps, X, scandetails):
+    thresholds = [-np.inf] + scandetails.thresholds + [np.inf]
+    print("info_gain")
+    acqval = np.array(
+        [
+            core.info_gain_gpytorch(xtest, gps, thresholds, scandetails.meanX)
+            for xtest in scandetails.acqX
+        ]
+    )
+
+    newx = None
+    for i, cacq in enumerate(scandetails.acqX[np.argsort(acqval)]):
+        if cacq.tolist() not in X.tolist():
+            print("taking new x. best non-existent index {} {}".format(i, cacq))
+            newx = cacq
+            return newx, acqval
+        else:
+            print("{} is good but already there".format(cacq))
+    print("returning None.. something must be wrong")
+    return None, None
+
+
+def batched_gridsearch(gps, X, scandetails, gp_maker=None, batchsize=1):
+    newX = np.empty((0, X.shape[-1]))
+    my_gps = gps
+    my_y_list = [gp.y_train_ for gp in my_gps]
+    myX = X
+
+    acqinfos = []
+
+    while True:
+        newx, acqinfo = gridsearch(my_gps, myX, scandetails)
+        newX = np.concatenate([newX, np.asarray([newx])])
+        acqinfos.append(acqinfo)
+        if (len(newX)) == batchsize:
+            print("we got our batch")
+            return newX, acqinfos
+
+        print("do fake update")
+        myX = np.concatenate([myX, newX])
+
+        newy_list = [gp.predict(newX) for gp in my_gps]
+        for i, newy in enumerate(newy_list):
+            print("new y i: {} {}".format(i, newy))
+            my_y_list[i] = np.concatenate([my_y_list[i], newy])
+        print("build fake gps")
+        my_gps = [gp_maker(myX, my_y) for my_y in my_y_list]


I don't understand the underscore in the folder. Does this mean there is an optimize/ folder somehwere else. If o, what is the difference?

irinaespejo · 2021-08-09T19:51:50Z

excursion/_optimize/core.py

@@ -0,0 +1,268 @@
+import numpy as np


Same comment as with _optimize/__init__.py, underscore?

irinaespejo · 2021-08-09T19:53:59Z

excursion/acquisition/MES.py

+            entropy_grid[p_j > 0] -= torch.log(torch.exp(p_j[p_j > 0])) * torch.log(torch.exp(torch.log(p_j[p_j > 0])))
+
+        acq_cand_vals = entropy_grid
+        self.acq_vals = torch.clone(entropy_grid)


Why is torch.clone necessary? Can we do self.acq_vals = acq_cand_vals or am I missing something?

irinaespejo · 2021-08-09T19:55:11Z

excursion/acquisition/PES.py

+        self.batch = batch
+        self.acq_vals = None
+
+    def _acquire(self, gp, thresholds, X_pointsgrid, x_candidate):


Why does class MES() have a method called acquire but PES has _acquire ? I think without underscore is fine

irinaespejo · 2021-08-09T19:57:56Z

excursion/acquisition/PPES.py

+def PPES(gp, testcase, thresholds, x_candidate):
+    """
+    Calculates information gain of choosing x_candidadate as next point to evaluate.
+    Performs this calculation with the Predictive Entropy Search approximation weighted by the posterior.
+    Roughly,
+    PES(x_candidate) = int Y(x)dx { H[Y(x_candidate)] - E_{S(x=j)} H[Y(x_candidate)|S(x=j)] }
+    Notation: PES(x_candidate) = int dx H0 - E_Sj H1
+
+    """
+
+    # compute predictive posterior of Y(x) | train data
+    raise NotImplmentedError(
+        "Should be same strcture as PES but the cumulative info gain is weighted"
+    )
+


Remove file altogether, sorry it's legacy

irinaespejo · 2021-08-09T20:01:07Z

excursion/acquisition/PES.py

+    def acquire(self, gp, thresholds, X_pointsgrid):
+        """
+        Calculates information gain of choosing x_candidate as next point to evaluate.
+        Performs this calculation with the Predictive Entropy Search approximation. Roughly,
+        PES(x_candidate) = int dx { H[Y(x_candidate)] - E_{S(x=j)} H[Y(x_candidate)|S(x=j)] }
+        Notation: PES(x_candidate) = int dx H0 - E_Sj H1
+
+        """
+
+        # compute predictive posterior of Y(x) | train data
+        # likelihood = gp.likelihood
+        # gp.eval()
+        # likelihood.eval()
+        acquisition_values = []
+
+        # for x_candidate in X_grid:
+        for x_candidate in X_pointsgrid:
+            x_candidate = x_candidate.view(1, -1).to(device=self.device, dtype=self.dtype)
+            acquisition_values.append(self._acquire(gp, thresholds, X_pointsgrid, x_candidate))
+
+        acq_cand_vals = torch.Tensor(acquisition_values).to(device=self.device, dtype=self.dtype)
+        self.acq_vals = torch.clone(acq_cand_vals)
+
+        return X_pointsgrid[self.get_first_max_index(gp, X_pointsgrid, acq_cand_vals)]


ok I see the utility of _acquire here. Same comment with torch.clone. The code for PES looks good to me.

irinaespejo · 2021-08-09T20:04:15Z

excursion/acquisition/base.py

+    def set_params(self, **params):
+        """
+        Set the parameters of this acquisition function.
+        Parameters
+        ----------
+        **params : dict
+            Function parameters.
+        Returns
+        -------
+        self : object
+            Function instance.
+        """
+        if not params:
+            # Simple optimization to gain speed (inspect is slow)
+            return self
+        for key, value in params.items():
+            setattr(self, key, value)
+
+        return self


We don't use acquisition function with parameters, that would be a bit meta. Let's erase the method if you can't give a use case.

irinaespejo · 2021-08-09T20:04:20Z

excursion/acquisition/base.py

+from collections import defaultdict
+
+
+class AcquisitionFunction(object):


Also, why is there no init method?

No init since it is an abstract interface and not an abstract object. this ensures all the future acq classes which implement it will be able to use the ask-and-tell API

irinaespejo · 2021-08-09T20:07:07Z

excursion/excursion.py

+        self.init_n_points = init_n_points # future place to store init points that were a snapshot of past training
+        self.dtype = torch.float64
+        self._invalid_region = None
+


self.device = device ?? with device coming from

if cuda available:
device = torch.device('cuda')
else:
device = torch.device('cpu')
If you have another way of incorporating device lmk

device comes from the algorithm options. bit convoluted at the moment i think... also this was an oversight bc it should not always be torch.float64, unless it should?

If it's too much hassle let's leave the device and dtype like that but please remmeber to add it to the ongoing TO DO list in this issue #12

end ✔️

irinaespejo · 2021-08-09T20:10:08Z

excursion/learner.py

+class _Learner(object):
+    def __init__(self, problem_details: ExcursionProblem = None, algorithm_options: dict = None):
+        self.problem_details = problem_details
+        self.options = algorithm_options
+
+    def ask(self, npoints: int = None):
+        """
+        Suggest a new point to evaluate.
+        Parameters
+        ----------
+        npoints : int, default: None
+            Determine how many points to ask for batch acquisition.
+        Returns
+        -------
+        points : object
+            Some kind of array object e.g. torch.Tensor or numpy.ndarray
+        """
+        raise NotImplementedError()
+
+    def tell(self, x, y):
+        raise NotImplementedError()
+
+    def evaluate(self, x):
+        raise NotImplementedError()
+
+    def evaluate_and_tell(self, x):
+        raise NotImplementedError()
+
+    def run(self, n_iterations, plot_result=False):
+        raise NotImplementedError()
+
+    def evaluate_metrics(self):
+        raise NotImplementedError()


We discussed _Learner in a meeting and agreed to remove? @lukasheinrich @tutrie

irinaespejo · 2021-08-09T20:17:54Z

excursion/unfactored/madgraph5atlasval/.gitmodules

+[submodule "DileptonReinterpretationProj"]
+	path = DileptonReinterpretationProj
+	url = https://gitlab.cern.ch/prieck/DileptonReinterpretationProj.git
+[submodule "excursion"]
+	path = excursion
+	url = https://github.com/diana-hep/excursion.git


Everything under the folder 'excursion/madgraph5atlasval/' should get removed, legacy 🔥 📄

irinaespejo · 2021-08-09T20:29:02Z

testing/algorithm_specs.yaml

+#
+# ALGORITHM OPTS
+#
+
+example: '1Dtoyanalysis'
+
+#number of initial points
+ninit: 1
+
+#number of iterations
+nupdates: 1
+
+##Define the init training data type
+init_type: 'random'
+
+#if init_type == 'custom' then fill in the lines below
+#X_train:
+#  [0.,0.]
+
+#likelihood
+likelihood:
+  type: 'GaussianLikelihood'
+  epsilon: 0.
+
+#model
+model:
+  type: 'GridGP'
+  kernel: 'RBF'
+  prior: 'Constant'
+
+
+#acquisition function
+acq:
+  acq_type: 'MES'
+  batch: False
+  batchtype: None
+  batchsize: None


ok the testing/ folder is legacy but should stay to make unit tests with pytest. I'm gonna put the refactoring of the testing/ folder to adapt it to your code in a TO DO list. No action to take for now.

irinaespejo · 2021-08-09T20:32:34Z

excursion/unfactored/checkmate.py

+import json
+import numpy as np
+import pkg_resources
+import pickle
+from excursion.utils import mgrid, mesh2points
+import torch
+import sklearn.preprocessing
+
+
+datafile = pkg_resources.resource_filename(
+    "excursion", "unfactored/data/checkmate_dense.json"
+)
+
+def modify(zv):
+    return np.log(zv) - np.log(0.05)
+
+
+truthX, truthy_obs, truthy_exp = [], [], []
+for p, _, result in json.load(open(datafile))["precomputed"]:
+    if p[0] < p[1] + 200:
+        continue
+    truthX.append(p)
+    truthy_obs.append(
+        max(float(result[1]["observed_CLs"]), 0.001) if result[1] else 0.001
+    )
+    truthy_exp.append(
+        max(float(result[1]["expected_CLs"]), 0.001) if result[1] else 0.001
+    )
+
+truthX = np.array(truthX)
+truthy_obs = np.array(truthy_obs)
+truthy_obs = modify(truthy_obs)
+
+truthy_exp = np.array(truthy_exp)
+truthy_exp = modify(truthy_exp)
+
+
+scaler = sklearn.preprocessing.MinMaxScaler()
+scaler.fit(truthX)
+truthX = scaler.transform(truthX)
+
+picklefile = pkg_resources.resource_filename(
+    "excursion", "unfactored/data/checkmate.pkl"
+)
+d = pickle.load(open(picklefile, "rb"))
+
+print('d', type(d))
+print(d)
+
+
+def truth_obs(X):
+    from scipy.interpolate import griddata
+    grid_data = torch.from_numpy(griddata(truthX,truthy_obs,X))
+    return grid_data
+
+
+
+
+def invalid_region(x):
+    oX = scaler.inverse_transform(x)
+    return oX[:, 0] < oX[:, 1] + 202
+
+
+thresholds = torch.Tensor([0.05])
+true_functions = [truth_obs]
+n_dims = 2
+
+rangedef = np.array([[0.0, 1.0, 101], [0.0, 1.0, 101]])
+plot_meshgrid = mgrid(rangedef)
+X_plot = mesh2points(plot_meshgrid, rangedef[:, 2])
+X = torch.from_numpy(X_plot)


Should the unfactored/ folder get 🔥 ? I think so, we have completely changed the testcase import module in favor of ExcrusionProblem class

irinaespejo · 2021-08-09T20:42:26Z

excursion/models/base.py

+        Set the parameters of this initial point generator.
+        Parameters


A Gaussian Process has hyperparameters and GPyTorch takes control of that. I'm unsure what parameters set_params refers to.

this is responsible for setting the initial parameters in the builders. it simplifies constructing objects for use with the ask and tell api. they dont have to change the learner or the optimizer. just make a new acquisition function and model and can do custom stuff with very little work. I can create an example, with botorch maybe?

If you could create a very simple use case for this, that'd be helpful. Thank you!
Before creating an example, maybe the famous diagram connecting the classes and objects is enough for me to understand the utility.

irinaespejo · 2021-08-09T20:51:25Z

excursion/models/gp.py

+            targets = targets + self.epsilon * MultivariateNormal(torch.zeros(len(targets)), torch.eye(len(targets)))\
+                .sample(torch.Size([])).to(device=self.device, dtype=self.dtype)


It's better to put NotImplementedError here.
The issue is subtle, when there is a noisy observation it means that querying the black box function ans receiving target_observed = target_true + epsilon. Epsilon is a random variable and everytime it is queried like in MultivariateNormal(torch.zeros(len(targets)), torch.eye(len(targets))) it will give a different error ⚠️
This can introduce bugs along the way because if we are querying target_observed anywhere else in the code they are going to be different quantities ❗

discussed on Slack, y(x) is only called once. no action to take

end ✔️

irinaespejo · 2021-08-09T20:53:09Z

excursion/optimize/_optimizer.py

+class _Optimizer(object):
+    def __init__(self, details: ExcursionProblem, device: str, n_funcs: int = None,
+                 base_estimator: str or list or ExcursionModel = "ExactGP", n_initial_points=None,
+                 initial_point_generator="random", acq_func: str = "MES", fit_optimizer=None,
+                 base_estimator_kwargs=None, fit_optimizer_kwargs=None, acq_func_kwargs=None, jump_start: bool = True):
+        raise NotImplementedError()
+
+    def ask(self, n_points=None, batch_kwarg={}):
+        raise NotImplementedError()
+
+    def _ask(self):
+        raise NotImplementedError()
+
+    def tell(self, x, y, fit=True):
+        raise NotImplementedError()
+
+    def _tell(self, x, y, fit=True):
+        """Perform the actual work of incorporating one or more new points.
+        See `tell()` for the full description.
+        This method exists to give access to the internals of adding points
+        by side stepping all input validation and transformation."""
+        raise NotImplementedError()
+
+    def update_next(self):
+        """Updates the value returned by opt.ask(). Does so by calling the acquisition func. Useful if a parameter
+        was updated after ask was called."""
+        raise NotImplementedError()
+
+    def fit(self):
+        """Calls the model's fit method. Give user access to this command as well.
+        """
+        raise NotImplementedError()
+
+    def get_result(self):
+        """Returns the most recent result as a new object. self.result stores the log if log = true.
+        """
+        raise NotImplementedError()


Same as with _Learner, do we need a private class? @lukasheinrich @tutrie

irinaespejo · 2021-08-09T20:55:03Z

excursion/test_suite/algorithm_options/algorithm_specs_2d_bench.yaml

@@ -0,0 +1,49 @@
+#


either we use a folder called test_suite/ or testing/ the name testing/ has the advantage that pytest automatically detects this is the folder to use for unit tests.

irinaespejo · 2021-08-09T20:57:27Z

excursion/optimize/estimator_backup.py

@@ -0,0 +1,518 @@
+from .builders import build_sampler, build_model, build_acquisition_func


Why have two folders _optimize/ and optimize/ ?

irinaespejo · 2021-08-09T21:01:17Z

excursion/optimize/builders.py

+        if acq_function == "mes":
+            acq_function = MES(device=kwargs['device'], dtype=kwargs['dtype'])
+        if acq_function == "pes":
+            acq_function = PES()


PES() should, at least, get the same arguments as MES: device=kwargs['device'], dtype=kwargs['dtype']

it does, set_params takes care of it. I will change it was an oversight

cool

end ✔️

irinaespejo

Last comments

made acquisition builder signature for MES and PES the same by removing dtype and device from MES letting it be set by set_params instead

…hub.com/tutrie/excursion into FEATURE/refactor_for_reduced_complexity

This reverts commit 33db5cd.

discovered bug with shared memory in the result object for the list objects. self.result.update updates lists of result objects for dif problems

we use the excursionresult to store data don't need two copies

…d test case and new example

irinaespejo added 30 commits March 24, 2020 18:26

code refactoring, goytorch integration, 1D example

be9c2f6

Merge pull request #1 from irinaespejo/refactor

fb62778

Refactor

2D working example gpytorch

d42a48b

2D working example gpytorch

2fd84f8

Merge pull request #2 from irinaespejo/refactor

d3bdb52

Refactor

2D working example gpytorch

10a7d04

Merge pull request #3 from irinaespejo/refactor

6991a5d

2D working example gpytorch

adding MES

7a97030

Merge branch 'master' of https://github.com/irinaespejo/excursion

7ad2ec5

Merge pull request #4 from irinaespejo/refactor

5ef5c07

adding MES

add pic

9845c3b

Merge branch 'master' of https://github.com/irinaespejo/excursion

ab2f201

update

add pic

214aada

add pic

b6b0123

add pic

f8bd688

add pic

9ab8468

add gifs

64db51e

new features

53a84bc

starting to make changes

7a657cd

reorganization in a ExcursionSetEstimation class

b1937e0

working refactory all dimensions toy examples

9a5ddf9

Merge pull request #5 from irinaespejo/refactor

80b2918

Refactor

up to date

9bde66a

Merge pull request #6 from irinaespejo/refactor

372da4a

up to date

introduced GridKernelRegression

a9c21a3

introduced GridKernelRegression

b4516cf

Merge pull request #7 from irinaespejo/refactor

e14b5a7

Refactor

CUDA device agnostic implementation

62c0be2

Merge pull request #8 from irinaespejo/refactor

9071e89

CUDA device agnostic implementation

fix minor argparse bug

e568c04

irinaespejo mentioned this pull request Aug 9, 2021

TO DOs #12

Open

2 tasks

irinaespejo reviewed Aug 9, 2021

View reviewed changes

irinaespejo suggested changes Aug 9, 2021

View reviewed changes

irinaespejo reviewed Aug 9, 2021

View reviewed changes

irinaespejo suggested changes Aug 9, 2021

View reviewed changes

tutrie added 13 commits August 9, 2021 17:17

deleting some things, fixed notebooks

0799947

deleted legacy files, fixed MES builder, new .gitignore

76697e3

made acquisition builder signature for MES and PES the same by removing dtype and device from MES letting it be set by set_params instead

Delete .idea directory

33db5cd

removed old results

68773bd

Merge branch 'FEATURE/refactor_for_reduced_complexity' of https://git…

ed72ada

…hub.com/tutrie/excursion into FEATURE/refactor_for_reduced_complexity

Revert "Delete .idea directory"

a51ff74

This reverts commit 33db5cd.

tried adding .idea to local exclude

25d4c61

found bug and update dtype implementation

e2eae77

discovered bug with shared memory in the result object for the list objects. self.result.update updates lists of result objects for dif problems

fixed bug with excursion result instance variables

c296c8b

removed _tell and data object

2b0cc96

we use the excursionresult to store data don't need two copies

getting ready for last legacy deletion

3d18c4e

updated import structure, deleted legacy, removed unsued code, create…

24e87f4

…d test case and new example

final deletes

5cf5a91

tutrie closed this Aug 16, 2021

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Feature/refactor for reduced complexity #11

Feature/refactor for reduced complexity #11

tutrie commented Aug 9, 2021

irinaespejo Aug 9, 2021

irinaespejo Aug 9, 2021

irinaespejo Aug 9, 2021

irinaespejo Aug 9, 2021

irinaespejo left a comment

irinaespejo Aug 9, 2021

irinaespejo Aug 9, 2021

irinaespejo Aug 9, 2021

irinaespejo Aug 9, 2021

irinaespejo Aug 9, 2021

irinaespejo Aug 9, 2021

irinaespejo Aug 9, 2021

irinaespejo Aug 9, 2021

tutrie Aug 9, 2021

irinaespejo Aug 10, 2021 •

edited

Loading

irinaespejo Aug 9, 2021

tutrie Aug 9, 2021

irinaespejo Aug 10, 2021

irinaespejo Aug 9, 2021

irinaespejo Aug 9, 2021

irinaespejo Aug 9, 2021

irinaespejo Aug 9, 2021

irinaespejo Aug 9, 2021

tutrie Aug 9, 2021

irinaespejo Aug 10, 2021

irinaespejo Aug 9, 2021

irinaespejo Aug 12, 2021

irinaespejo Aug 9, 2021

irinaespejo Aug 9, 2021

irinaespejo Aug 9, 2021

irinaespejo Aug 9, 2021

tutrie Aug 9, 2021

irinaespejo Aug 10, 2021

irinaespejo left a comment

		from collections import defaultdict


		class AcquisitionFunction(object):

		Set the parameters of this initial point generator.
		Parameters

		targets = targets + self.epsilon * MultivariateNormal(torch.zeros(len(targets)), torch.eye(len(targets)))\
		.sample(torch.Size([])).to(device=self.device, dtype=self.dtype)

		@@ -0,0 +1,518 @@
		from .builders import build_sampler, build_model, build_acquisition_func

Feature/refactor for reduced complexity #11

Feature/refactor for reduced complexity #11

Conversation

tutrie commented Aug 9, 2021

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

irinaespejo left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

irinaespejo Aug 10, 2021 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

irinaespejo left a comment

Choose a reason for hiding this comment

irinaespejo Aug 10, 2021 •

edited

Loading