Following up on #86 , development of a 1.0 compatible GPflowOpt version. This is far from done, the biggest difficulty (Acquisition) is still ahead unfortunately.

At the same time lots of tests are affected: I'm reworking them to use some of the cool pytest features. When this work is over, I hope to do another PR to improve testing further (split out computationally demanding tests to system, use mock to test things which now trigger BO runs and model optimizations)

As mentioned #4 , tests often failed (mostly on python 2.7) due to cholesky decomposition errors. First I thought this was mostly caused by updating the data and calling optimize() again, but resetting the hyperparameters wasn't working all the time. Increasing the likelihood variance sometimes helps slightly but isn't very robust either.

Right now the tests specify lengthscales for the initial model, and apply a hyperprior on the kernel variance. Each BO iteration, the hyperparameters supplied with the initial model are applied as a starting point. In addition, restarts are applied by randomizing the Params. This approach made it a lot more stable but isn't perfect yet. Especially in longer runs of BO, reverting to the supplied lengthscales each time ultimately causes crashes.

Some things we may consider:

• Normalizing the input/output data. Tested this a bit, didn't solve the issue. Additionally, the model hyperparameters loose some interpretability. Note that I think we will ultimately need this for PES anyway.
• Add a callback for re-configuring hyperparameters. Instead of reverting to the initially supplied hypers each iteration, this function is called and configures the initial state. I think for more complex modeling approaches this ultimately required, but for simple scenario's with GPR this has to work automatically.
• Applying hyperpriors is going to be important.

I'd love to hear thoughts on how to improve this.

Good morning guys,

I am currently trying to get gpflowopt up and running for some optimization problem. Naturally, I first tried the example you provided in the documentation Now, while I do get the same result as in the documentation, I am a bit puzzled about the function evaluations the optimizer is choosing. To be more specific, the optimizer always chooses to evaluate the function at the point [0.0, 0.5] in all 15 iterations. I am probably overlooking something, as this does not seem to be the desired behavior, right? The optimizer seems to be not really optimizing. Can anyone point out the mistake I made during the setup of the problem? I am pretty sure I follow the instructions of the example in the documentation to the letter.

This is the code, that I am running:

import numpy as np
from gpflowopt.domain import ContinuousParameter
import gpflow
from gpflowopt.bo import BayesianOptimizer
from gpflowopt.design import LatinHyperCube
from gpflowopt.acquisition import ExpectedImprovement
#from gpflowopt.optim import SciPyOptimizer

def fx(X):
    X = np.atleast_2d(X)
    result = np.sum(np.square(X), axis=1)[:, None]
    print("X: {}".format(X))
    print("fx: {}".format(result))
    return result


domain = ContinuousParameter('x1', -2, 2) + ContinuousParameter('x2', -1, 2)

# Use standard Gaussian process Regression
lhd = LatinHyperCube(21, domain)
X = lhd.generate()
Y = fx(X)
model = gpflow.gpr.GPR(X, Y, gpflow.kernels.Matern52(2, ARD=True))
model.kern.lengthscales.transform = gpflow.transforms.Log1pe(1e-3)

# Now create the Bayesian Optimizer
alpha = ExpectedImprovement(model)
optimizer = BayesianOptimizer(domain, alpha)

# Run the Bayesian optimization
#with optimizer.silent():
r = optimizer.optimize(fx, n_iter=15)
print(r)

And this is the output I am seeing:

python try_out_gpflowopt.py 
2017-12-11 09:28:09.790044: I tensorflow/core/platform/cpu_feature_guard.cc:137] Your CPU supports instructions that this TensorFlow binary was not compiled to use: SSE4.
1 SSE4.2 AVX AVX2 FMA
X: [[ 0.2   0.65]
 [ 0.   -0.1 ]
 [-0.8   0.5 ]
 [ 0.4   1.4 ]
 [-0.6   1.25]
 [ 1.    0.05]
 [ 1.2   0.8 ]
 [-1.   -0.25]
 [ 0.8  -0.7 ]
 [ 1.4   1.55]
 [-1.6   1.1 ]
 [-1.8   0.35]
 [-0.2  -0.85]
 [ 1.8   0.2 ]
 [-0.4   1.85]
 [ 0.6   2.  ]
 [ 2.    0.95]
 [ 1.6  -0.55]
 [-1.4   1.7 ]
 [-1.2  -1.  ]
 [-2.   -0.4 ]]
fx: [[ 0.4625]
 [ 0.01  ]
 [ 0.89  ]
 [ 2.12  ]
 [ 1.9225]
 [ 1.0025]
 [ 2.08  ]
 [ 1.0625]
 [ 1.13  ]
 [ 4.3625]
 [ 3.77  ]
 [ 3.3625]
 [ 0.7625]
 [ 3.28  ]
 [ 3.5825]
 [ 4.36  ]
 [ 4.9025]
 [ 2.8625]
 [ 4.85  ]
 [ 2.44  ]
 [ 4.16  ]]
Warning: optimization restart 4/5 failed
X: [[ 0.   0.5]]
fx: [[ 0.25]]
X: [[ 0.   0.5]]
fx: [[ 0.25]]
X: [[ 0.   0.5]]
fx: [[ 0.25]]
X: [[ 0.   0.5]]
fx: [[ 0.25]]
X: [[ -3.76012797e-10   4.99988509e-01]]
fx: [[ 0.24998851]]
Warning: optimization restart 1/5 failed
Warning: inf or nan in gradient: replacing with zeros
X: [[ 0.   0.5]]
fx: [[ 0.25]]
Warning: optimization restart 4/5 failed
Warning: inf or nan in gradient: replacing with zeros
X: [[ 0.   0.5]]
fx: [[ 0.25]]
X: [[ 0.   0.5]]
fx: [[ 0.25]]
X: [[ 0.   0.5]]
fx: [[ 0.25]]
Warning: optimization restart 3/5 failed
X: [[ 0.   0.5]]
fx: [[ 0.25]]
Warning: inf or nan in gradient: replacing with zeros
Warning: inf or nan in gradient: replacing with zeros
Warning: inf or nan in gradient: replacing with zeros
X: [[ 0.   0.5]]
fx: [[ 0.25]]
Warning: inf or nan in gradient: replacing with zeros
Warning: inf or nan in gradient: replacing with zeros
Warning: inf or nan in gradient: replacing with zeros
X: [[ 0.   0.5]]
fx: [[ 0.25]]
Warning: inf or nan in gradient: replacing with zeros
Warning: inf or nan in gradient: replacing with zeros
Warning: optimization restart 3/5 failed
Warning: inf or nan in gradient: replacing with zeros
Warning: inf or nan in gradient: replacing with zeros
Warning: inf or nan in gradient: replacing with zeros
X: [[ 0.   0.5]]
fx: [[ 0.25]]
Warning: inf or nan in gradient: replacing with zeros
Warning: inf or nan in gradient: replacing with zeros
Warning: inf or nan in gradient: replacing with zeros
X: [[ 0.   0.5]]
fx: [[ 0.25]]
Warning: inf or nan in gradient: replacing with zeros
Warning: inf or nan in gradient: replacing with zeros
Warning: inf or nan in gradient: replacing with zeros
Warning: optimization restart 3/5 failed
Warning: inf or nan in gradient: replacing with zeros
Warning: inf or nan in gradient: replacing with zeros
X: [[ 0.   0.5]]
fx: [[ 0.25]]
Warning: inf or nan in gradient: replacing with zeros
Warning: inf or nan in gradient: replacing with zeros
Warning: inf or nan in gradient: replacing with zeros
X: [[ 0.   0.5]]
fx: [[ 0.25]]
     fun: array([ 0.01])
 message: 'OK'
    nfev: 15
 success: True
       x: array([[ 0. , -0.1]])

For the sake of completeness, these are the steps I took to setup GPFlowOpt:

1. Created new conda environment based on Python 3.5
2. Cloned the gpflowopt repo
3. Ran pip install . --process-dependency-links

BTW, this is the first issue I've ever created on GiHub, so please forgive me if I am violating any conventions and please let me know if I left out crucial information.

During calls to optimize(), sometimes UserWarnings pop up:

/home/javdrher/.virtualenvs/gpflowopt/lib/python3.5/site-packages/GPflow/transforms.py:129: RuntimeWarning: overflow encountered in exp result = np.log(1. + np.exp(x)) + self._lower

Typically this is quite harmless, if its really causing troubles its usually followed by a cholesky decomposition exception. However, those warnings mess up output. Specifically in documentation notebooks. I was thinking of silencing the warnings, any reason not to?

The algorithm in Pareto::divide_conquer_nd fails when two points in the Pareto set have the same value in a certain dimension. An example is included in the modified pareto.py: The Pareto set d21 contains three points, two of which have a value of 2.0 in the first dimension. Ordering the three points results in different ways results in different values of the hypervolume (28 and 32), which are all wrong (should be 29).

The issue is in the dominance test associated with _is_test_required method and pseudo_pf. The array pseudo_pf assigns different ranks to the same values. Therefore, by reordering the Pareto set in the test case, different pseudo Pareto sets are generated for the same Pareto set.

I figured out two ways for the fix. One is to fix pseudo_pf by sorting the Pareto set such that the same values are assigned the same rank (e.g. using scipy.stats.rankdata). The other is to fix the dominance test by checking the actual Pareto set. The first one leads to more iterations in the algorithm. Therefore, I implemented the second approach in pareto.py with a minimum modification - although some simplifications of the code are possible.

i have installed latest gpflow version (1.2) using pip Later when i tried to install gpflowopt using the following code pip install git+https://github.com/GPflow/GPflowOpt.git

it is down grading GPflow from 1.2 to 0.4

I just downloaded GPflowOpt, yet nothing can run due to slight changes made in GPflow. For example, you have now made a 'core' subfolder and seemed to changed the AutoFlow function. I have tried to apply come changes on my own (it is not hard to change 'from gpflow.param import DataHolder, Autoflow' to two separate imports in their correct folders, params and core), but note @Autoflow calls now need to be @gpflow.autoflow. I spent several hours changing this for pretty much every function/class.

Yet now it seems certain classes are also significantly changed. 'Parameterized' no longer has the attribute 'highest_parent', something needed for SciPyOptimizer.

At this point, not a single thing can be called from GPflowOpt without error.

I implemented the recent acquisition function Max-Value Entropy Search from:

Wang, Z. & Jegelka, S.. (2017). Max-value Entropy Search for Efficient Bayesian Optimization. Proceedings of the 34th International Conference on Machine Learning, in PMLR 70:3627-3635

We named it Min-Value Entropy Search because the GPflow framework seeks the minimum if the function. There is a notebook evaluating the method on the Shekel function and it seems to perform well.

In this pull request I implemented the Approximatively Marginalised GP as described as indicated in issue #39 . It currently supports multi-output GPs. A notebook and some tests are included.

Do improve the speed of optimisation, I replaced GPR with VGP as follows:

domain = np.sum([GPflowOpt.domain.ContinuousParameter(f'mux{i}', mm[i], mx[i]) for i in range(7)]) domain += np.sum([GPflowOpt.domain.ContinuousParameter(f'muy{i}', mm[i+7], mx[i+7]) for i in range(7)]) domain += np.sum([GPflowOpt.domain.ContinuousParameter(f'sigmax{i}', 1e-7, 1.) for i in range(7)]) domain += np.sum([GPflowOpt.domain.ContinuousParameter(f'sigmay{i}', 1e-7, 1.) for i in range(7)]) domain += GPflowOpt.domain.ContinuousParameter('offset', endo * 0.7, endo * 1.3) design = GPflowOpt.design.RandomDesign(500, domain) X = design.generate() Y = np.vstack([obj(x.reshape(1, -1)) for x in X]) model = GPflow.vgp.VGP(X, Y, GPflow.kernels.RBF(29, lengthscales=X.std(axis=0)), likelihood=GPflow.likelihoods.Gaussian()) acquisition = GPflowOpt.acquisition.ExpectedImprovement(model) opt = GPflowOpt.optim.StagedOptimizer([GPflowOpt.optim.MCOptimizer(domain, 500), GPflowOpt.optim.SciPyOptimizer(domain)]) optimizer = GPflowOpt.BayesianOptimizer(domain, acquisition, optimizer=opt) optimizer.optimize(obj, n_iter=500)

GPR works, but with VGP I receive the following error:

[[Node: gradients/unnamed._models.model_datascaler.model.build_likelihood/unnamed._models.model_datascaler.model.likelihood.variational_expectations/sub_1_grad/BroadcastGradientArgs = BroadcastGradientArgs[T=DT_INT32, _device="/j ob:localhost/replica:0/task:0/gpu:0"](gradients/unnamed._models.model_datascaler.model.build_likelihood/unnamed._models.model_datascaler.model.likelihood.variational_expectations/sub_1_grad/Shape, gradients/unnamed._models.model_datascale r.model.build_likelihood/unnamed._models.model_datascaler.model.likelihood.variational_expectations/sub_1_grad/Shape_1)]] 2017-07-18 23:03:28.798171: W tensorflow/core/framework/op_kernel.cc:1158] Invalid argument: Incompatible shapes: [501,1] vs. [500,1] [[Node: gradients/unnamed._models.model_datascaler.model.build_likelihood/unnamed._models.model_datascaler.model.likelihood.variational_expectations/sub_1_grad/BroadcastGradientArgs = BroadcastGradientArgs[T=DT_INT32, _device="/j ob:localhost/replica:0/task:0/gpu:0"](gradients/unnamed._models.model_datascaler.model.build_likelihood/unnamed._models.model_datascaler.model.likelihood.variational_expectations/sub_1_grad/Shape, gradients/unnamed._models.model_datascale r.model.build_likelihood/unnamed._models.model_datascaler.model.likelihood.variational_expectations/sub_1_grad/Shape_1)]] Warning: optimization restart 1/5 failed 2017-07-18 23:03:28.898935: W tensorflow/core/framework/op_kernel.cc:1158] Invalid argument: Incompatible shapes: [500,1] vs. [501,1] [[Node: gradients/unnamed._models.model_datascaler.model.build_likelihood/add_1_grad/BroadcastGradientArgs = BroadcastGradientArgs[T=DT_INT32, _device="/job:localhost/replica:0/task:0/gpu:0"](gradients/unnamed._models.model_datas caler.model.build_likelihood/add_1_grad/Shape, gradients/unnamed._models.model_datascaler.model.build_likelihood/add_1_grad/Shape_1)]] 2017-07-18 23:03:28.898992: W tensorflow/core/framework/op_kernel.cc:1158] Invalid argument: Incompatible shapes: [500,1] vs. [501,1] [[Node: gradients/unnamed._models.model_datascaler.model.build_likelihood/add_1_grad/BroadcastGradientArgs = BroadcastGradientArgs[T=DT_INT32, _device="/job:localhost/replica:0/task:0/gpu:0"](gradients/unnamed._models.model_datas caler.model.build_likelihood/add_1_grad/Shape, gradients/unnamed._models.model_datascaler.model.build_likelihood/add_1_grad/Shape_1)]] 2017-07-18 23:03:28.899066: W tensorflow/core/framework/op_kernel.cc:1158] Invalid argument: Incompatible shapes: [500,1] vs. [501,1] [[Node: gradients/unnamed._models.model_datascaler.model.build_likelihood/add_1_grad/BroadcastGradientArgs = BroadcastGradientArgs[T=DT_INT32, _device="/job:localhost/replica:0/task:0/gpu:0"](gradients/unnamed._models.model_datas caler.model.build_likelihood/add_1_grad/Shape, gradients/unnamed._models.model_datascaler.model.build_likelihood/add_1_grad/Shape_1)]] 2017-07-18 23:03:28.899289: W tensorflow/core/framework/op_kernel.cc:1158] Invalid argument: Incompatible shapes: [500,1] vs. [501,1] [[Node: gradients/unnamed._models.model_datascaler.model.build_likelihood/add_1_grad/BroadcastGradientArgs = BroadcastGradientArgs[T=DT_INT32, _device="/job:localhost/replica:0/task:0/gpu:0"](gradients/unnamed._models.model_datas caler.model.build_likelihood/add_1_grad/Shape, gradients/unnamed._models.model_datascaler.model.build_likelihood/add_1_grad/Shape_1)]] Warning: optimization restart 2/5 failed

I'm using master GPflow and GPflowOpt on TensorFlow 1.2 and Python 3.6.

Thanks.

Following #52 here is some code which gets rid of 1 optimize call (in case of no initial design specified). The PR also includes a context which can be used to suspend all optimizes. This is mostly useful for the lower-level api.

note: In the following release the data mechanism will undergo some changes and enabling the scaling should get moved to avoid another scaling

Hello can you help me to solve this issue ERROR: Could not find a version that satisfies the requirement GPflow==0.5.0 (from gpflowopt) (from versions: 1.4.1.linux-x86_64, 1.5.0.linux-x86_64, 1.0.0, 1.1.0, 1.1.1, 1.2.0, 1.3.0, 1.4.1, 1.5.0, 1.5.1, 2.0.0rc1, 2.0.0, 2.0.1, 2.0.2, 2.0.3, 2.0.4, 2.0.5, 2.1.0, 2.1.1, 2.1.2, 2.1.3, 2.1.4, 2.1.5, 2.2.0, 2.2.1, 2.3.0, 2.3.1, 2.4.0, 2.5.1, 2.5.2) ERROR: No matching distribution found for GPflow==0.5.0

I tried to install in google colab and spyder but not working

I am very new to this BO domain and working on constraint BO and found this wonderful tool that has constrained BO application. As I can understand from the code that the acquisition function for objective and constraint are multiplied. I have a question about that. Is this constrained BO considered as coupled then?

Is there a ask/tell interface? If not, is there a workaround?

I'd like to initialize the optimizer with already known function evaluations. Likewise, I'd like to query the next data point that should be evaluated, according to the acquisition function.

Hey there, I found your project which seems promising for a problem, I am currently working on. However, I as far as I can see, the option to include discrete variables in the optimization is not yet implemented? Is this correct? and if so, are there any developement in this direction currently going on?

Hi, Before I start writing my coding with GPflowOpt, I need to check with you whether I can use a PyTorch model within the objective function? My objective function needs the decoder of a VAE defined using PyTorch. In addition to other essential parameters, I want to pass this model as one parameter to the objective function. I understand that GPflowOpt is based on TF, but I am not sure whether the objective can be any function independent of TF. Thanks in advance...