to_dict() and from_dict() functionality for Coregionalize Kernel and MixedNoise Likelihood class, appveyor CI resurrected (SheffieldML#951)

This PR adds two main things to GPy:
- to- and from-dict functions for the kernels listed belop
- a fix for the appveyor CI
Please see the squashed commit messages listed below.
Authors: @gehbiszumeis @ppk42 respectively
Reviewer: @ekalosak 

---
* new: added to_dict() method to Coregionalize kernel class

* new: added to_dict() method to MixedNoise likelihood class

* fix: made Y_metadata dict content serializable

* fix: typo

* added additional needed parameters to to_dict() method for Coregionalize kernel + added _build_from_input dict method

* new: added possibility to build MixedNoise likelihood from input_dict

* Y_metadata conversion from serializable to np.array when loading from dict

* fix: rework Y_metadata part for compatibility with unittests !minor

* conda cleanup in appveyors pipeline

* conda clean up after conda update

* conda clean before conda update

* try pinning packages for conda

* revert all conda changes

* conda clean all (not only packages)

* use conda update anaconda

* pin conda package

* pin conda package

* try installing charset-normalizer beforehand

* try to get from conda-forge

* revert all conda changes

* Try to fix the conda update challange.

See: https://community.intel.com/t5/Intel-Distribution-for-Python/Conda-update-Conda-fails/td-p/1126174

It is just a try for a different context/(conda version).

* Still fixing build error on appveyor

I also use a newer miniconda version for greater python versions.

* Update appveyor.yml

Thinking it over it decided to use miniconda38 for all python versions unless python 3.5.

* revert miniconda versioning changes

* adjust GPy version in appveyor.yml

* 1st attempt bring the appveyor build to life again

* SheffieldML#955 fixing ci build on appveyor

After bringing the miniconda env to work again, the wrong matplotlib version was used. This commit should fix that.

* SheffieldML#955 Fix CI build

Freezing numpy and scipy was a bad idea.
I freeze matplotlib  dependend  on the python version only.

* add: built_from_dict method for White Kernel

Co-authored-by: Peter Paul Kiefer <[email protected]>
Co-authored-by: Peter Paul Kiefer <[email protected]>

Author: 3 people

GPy/core/gp.py

@@ -134,9 +134,10 @@ def to_dict(self, save_data=True):
         if self.mean_function is not None:
             input_dict["mean_function"] = self.mean_function.to_dict()
         input_dict["inference_method"] = self.inference_method.to_dict()
-        #FIXME: Assumes the Y_metadata is serializable. We should create a Metadata class
+        # TODO: We should create a Metadata class
         if self.Y_metadata is not None:
-            input_dict["Y_metadata"] = self.Y_metadata
+            # make Y_metadata serializable
+            input_dict["Y_metadata"] = {k: self.Y_metadata[k].tolist() for k in self.Y_metadata.keys()}
         if self.normalizer is not None:
             input_dict["normalizer"] = self.normalizer.to_dict()
         return input_dict
@@ -162,9 +163,12 @@ def _format_input_dict(input_dict, data=None):
             input_dict["mean_function"] = mean_function
         input_dict["inference_method"] = GPy.inference.latent_function_inference.LatentFunctionInference.from_dict(input_dict["inference_method"])
 
-        #FIXME: Assumes the Y_metadata is serializable. We should create a Metadata class
+        # converts Y_metadata from serializable to array. We should create a Metadata class
         Y_metadata = input_dict.get("Y_metadata")
-        input_dict["Y_metadata"] = Y_metadata
+        if isinstance(Y_metadata, dict):
+            input_dict["Y_metadata"] = {k: np.array(Y_metadata[k]) for k in Y_metadata.keys()}
+        else:
+            input_dict["Y_metadata"] = Y_metadata
 
         normalizer = input_dict.get("normalizer")
         if normalizer is not None:

GPy/core/parameterization/variational.py

@@ -106,7 +106,7 @@ def __init__(self, means=None, variances=None, name='latent space', *a, **kw):
         self.link_parameters(self.mean, self.variance)
         self.num_data, self.input_dim = self.mean.shape
         if self.has_uncertain_inputs():
-            assert self.variance.shape == self.mean.shape, "need one variance per sample and dimenion"
+            assert self.variance.shape == self.mean.shape, "need one variance per sample and dimension"
 
     def set_gradients(self, grad):
         self.mean.gradient, self.variance.gradient = grad

GPy/kern/src/coregionalize.py

@@ -134,3 +134,28 @@ def gradients_X(self, dL_dK, X, X2=None):
 
     def gradients_X_diag(self, dL_dKdiag, X):
         return np.zeros(X.shape)
+
+    def to_dict(self):
+        """
+        Convert the object into a json serializable dictionary.
+
+        Note: It uses the private method _save_to_input_dict of the parent.
+
+        :return dict: json serializable dictionary containing the needed information to instantiate the object
+        """
+
+        input_dict = super(Coregionalize, self)._save_to_input_dict()
+        input_dict["class"] = "GPy.kern.Coregionalize"
+        # W and kappa must be serializable
+        input_dict["W"] = self.W.values.tolist()
+        input_dict["kappa"] = self.kappa.values.tolist()
+        input_dict["output_dim"] = self.output_dim
+        return input_dict
+
+    @staticmethod
+    def _build_from_input_dict(kernel_class, input_dict):
+        useGPU = input_dict.pop('useGPU', None)
+        # W and kappa must be converted back to numpy arrays
+        input_dict['W'] = np.array(input_dict['W'])
+        input_dict['kappa'] = np.array(input_dict['kappa'])
+        return Coregionalize(**input_dict)

GPy/kern/src/static.py

@@ -68,6 +68,11 @@ def to_dict(self):
         input_dict = super(White, self)._save_to_input_dict()
         input_dict["class"] = "GPy.kern.White"
         return input_dict
+
+    @staticmethod
+    def _build_from_input_dict(kernel_class, input_dict):
+        useGPU = input_dict.pop('useGPU', None)
+        return White(**input_dict)
 
     def K(self, X, X2=None):
         if X2 is None:

GPy/likelihoods/mixed_noise.py

@@ -80,3 +80,32 @@ def samples(self, gp, Y_metadata):
             _ysim = np.array([np.random.normal(lik.gp_link.transf(gpj), scale=np.sqrt(lik.variance), size=1) for gpj in gp_filtered.flatten()])
             Ysim[flt,:] = _ysim.reshape(n1,N2)
         return Ysim
+
+    def to_dict(self):
+        """
+        Convert the object into a json serializable dictionary.
+
+        Note: It uses the private method _save_to_input_dict of the parent.
+
+        :return dict: json serializable dictionary containing the needed information to instantiate the object
+        """
+
+        # input_dict = super(MixedNoise, self)._save_to_input_dict()
+        input_dict = {"name": self.name,
+                      "class": "GPy.likelihoods.MixedNoise",
+                      "likelihoods_list": []}
+        for ii in range(len(self.likelihoods_list)):
+            input_dict["likelihoods_list"].append(self.likelihoods_list[ii].to_dict())
+
+        return input_dict
+
+    @staticmethod
+    def _build_from_input_dict(likelihood_class, input_dict):
+        import copy
+        input_dict = copy.deepcopy(input_dict)
+        # gp_link_dict = input_dict.pop('gp_link_dict')
+        # import GPy
+        # gp_link = GPy.likelihoods.link_functions.GPTransformation.from_dict(gp_link_dict)
+        # input_dict["gp_link"] = gp_link
+        input_dict['likelihoods_list'] = [Likelihood.from_dict(l) for l in input_dict['likelihoods_list']]
+        return likelihood_class(**input_dict)

appveyor.yml

@@ -3,18 +3,23 @@ environment:
     secure: 8/ZjXFwtd1S7ixd7PJOpptupKKEDhm2da/q3unabJ00=
   COVERALLS_REPO_TOKEN:
     secure: d3Luic/ESkGaWnZrvWZTKrzO+xaVwJWaRCEP0F+K/9DQGPSRZsJ/Du5g3s4XF+tS
-  gpy_version: 1.9.9
+  gpy_version: 1.10.0
   matrix:
     - PYTHON_VERSION: 3.5
       MINICONDA: C:\Miniconda35-x64
+      MPL_VERSION: 3.0.0
     - PYTHON_VERSION: 3.6
-      MINICONDA: C:\Miniconda36-x64
+      MINICONDA: C:\Miniconda3-x64
+      MPL_VERSION: 3.3.4
     - PYTHON_VERSION: 3.7
-      MINICONDA: C:\Miniconda36-x64
+      MINICONDA: C:\Miniconda3-x64
+      MPL_VERSION: 3.3.4
     - PYTHON_VERSION: 3.8
-      MINICONDA: C:\Miniconda36-x64
+      MINICONDA: C:\Miniconda3-x64
+      MPL_VERSION: 3.3.4
     - PYTHON_VERSION: 3.9
-      MINICONDA: C:\Miniconda36-x64
+      MINICONDA: C:\Miniconda3-x64
+      MPL_VERSION: 3.3.4
 
 #configuration:
 #  - Debug
@@ -25,7 +30,8 @@ install:
  - conda config --set always_yes yes --set changeps1 no
  - conda update -q conda
  - conda info -a
- - "conda create -q -n build-environment python=%PYTHON_VERSION% numpy scipy matplotlib"
+# github issue #955: freeze build version of matplotlib
+ - "conda create -q -n build-environment python=%PYTHON_VERSION% numpy scipy matplotlib=%MPL_VERSION%"
  - activate build-environment
  # We need wheel installed to build wheels
  - python -m pip install wheel

setup.py

@@ -118,8 +118,10 @@ def ismac():
     ext_mods = []
 
 install_requirements = ['numpy>=1.7', 'six', 'paramz>=0.9.0', 'cython>=0.29']
+matplotlib_version = 'matplotlib==3.3.4'
 if sys.version_info < (3, 6):
     install_requirements += ['scipy>=1.3.0,<1.5.0']
+    matplotlib_version = 'matplotlib==3.0.0'
 else:
     install_requirements += ['scipy>=1.3.0']
 
@@ -174,7 +176,8 @@ def ismac():
                         'optional':['mpi4py',
                                     'ipython>=4.0.0',
                                     ],
-                        'plotting':['matplotlib >= 3.0',
+                        #matplotlib Version see github issue #955
+                        'plotting':[matplotlib_version,
                                     'plotly >= 1.8.6'],
                         'notebook':['jupyter_client >= 4.0.6',
                                     'ipywidgets >= 4.0.3',