Merge 6146b2c into 40a9466

docs/source/advancedcommands.rst

@@ -1,6 +1,65 @@
 Advanced Commands
 *****************
 
+.. _migp:
+
+Choice Variables
+================
+If MOSEK 9 is installed, GPkit supports discretized free variables with the ``mosek_conif`` solver.
+Choice variables are free in the sense of having multiple possible choices, but discretized in the
+sense of having a limited set of possible solutions.
+
+.. literalinclude:: examples/migp.py
+
+If solved with the mosek_conif solver, the script above will print::
+
+    Optimal Cost
+    ------------
+     [ 1.5       2.15      2.8       3.22     ... ]
+
+    ~~~~~~~~
+    WARNINGS
+    ~~~~~~~~
+    No Dual Solution
+    ----------------
+    This model has the discretized choice variables [x] and hence no dual
+    solution. You can fix those variables to their optimal value and get
+    sensitivities to the resulting continuous problem by updating your model's
+    substitions with `sol["choicevariables"]`.
+    ~~~~~~~~
+
+    Swept Variables
+    ---------------
+    numerator : [ 0.5
+                  1.15
+                  1.8
+                  2.45
+                  3.1
+                  3.75
+                  4.4
+                  5.05
+                  5.7
+                  6.35
+                  7         ]
+
+    Free Variables
+    --------------
+    x : [ 1
+          1
+          1
+          2
+          2
+          2
+          2
+          2
+          2
+          3
+          3         ]
+
+Note that the optimal values for ``x`` are discretized, clicking from 1
+to 2 to 3 during the sweep, and that the solution has no dual variables.
+
+
 Derived Variables
 =================
 

docs/source/examples/migp.py

@@ -0,0 +1,10 @@
+import numpy as np
+from gpkit import *
+
+x = Variable("x", choices=range(1,4))
+num = Variable("numerator", np.linspace(0.5, 7, 11))
+
+m = Model(x + num/x)
+sol = m.solve(verbosity=0)
+
+print(sol.table())

docs/source/examples/migp_output.txt

@@ -0,0 +1,59 @@
+
+Optimal Cost
+------------
+ [ 1.41      2.14      2.68      3.13     ... ]
+
+~~~~~~~~
+WARNINGS
+~~~~~~~~
+Freed Choice Variables
+----------------------
+This model has the discretized choice variables [x], but since the 'cvxopt' solver doesn't support discretization they were treated as continuous variables.
+~~~~~~~~
+
+Swept Variables
+---------------
+numerator : [ 0.5
+              1.15
+              1.8
+              2.45
+              3.1
+              3.75
+              4.4
+              5.05
+              5.7
+              6.35
+              7         ]
+
+Free Variables
+--------------
+x : [ 0.707
+      1.07
+      1.34
+      1.57
+      1.76
+      1.94
+      2.1
+      2.25
+      2.39
+      2.52
+      2.65      ]
+
+Variable Sensitivities
+----------------------
+numerator : [ +0.5
+              +0.5
+              +0.5
+              +0.5
+              +0.5
+              +0.5
+              +0.5
+              +0.5
+              +0.5
+              +0.5
+              +0.5      ]
+
+Most Sensitive Constraints (in last sweep)
+------------------------------------------
+(none)
+

docs/source/installation.rst

@@ -4,42 +4,44 @@ Installation
 ************
 
 1. If you are on Mac or Windows, we recommend installing `Anaconda <http://www.continuum.io/downloads>`_. Alternatively, `install pip and create a virtual environment <https://packaging.python.org/guides/installing-using-pip-and-virtualenv/>`_.
-2. (optional) Install the MOSEK solver as directed below
-3. Run ``pip install gpkit`` in the appropriate terminal or command prompt.
-4. Open a Python prompt and run ``import gpkit`` to finish installation and run unit tests.
+2. (optional) Install the MOSEK 9 solver with ``pip install Mosek``, then a license as described below
+3. (optional) Install the MOSEK 8 solver as described below
+4. Run ``pip install gpkit`` in the appropriate terminal or command prompt.
+5. Open a Python prompt and run ``import gpkit`` to finish installation and run unit tests.
 
 If you encounter any bugs please email ``gpkit@mit.edu``
 or `raise a GitHub issue <http://github.com/convexengineering/gpkit/issues/new>`_.
 
 
-Installing MOSEK
-================
-GPkit interfaces with two off the shelf solvers: cvxopt, and MOSEK.
+Installing MOSEK 8
+==================
+GPkit interfaces with two off the shelf solvers: cvxopt, and MOSEK (versions 8 and 9).
 Cvxopt is open source and installed by default; MOSEK requires a commercial licence or (free)
-academic license.
+academic license. In MOSEK version 8 GPkit uses the command-line interface ``mskexpopt`` solver, while
+in MOSEK 9 it uses the more active exponential-cone interface (and hence supports :ref:`migp`).
 
 Mac OS X
   - If ``which gcc`` does not return anything, install the `Apple Command Line Tools <https://developer.apple.com/downloads/index.action?=command%20line%20tools>`_.
   - Download `MOSEK 8 <https://www.mosek.com/downloads/>`_, then:
       - Move the ``mosek`` folder to your home directory
       - Follow `these steps for Mac <http://docs.mosek.com/7.0/toolsinstall/Mac_OS_X_installation.html>`_.
-      - Request an `academic license file <http://license.mosek.com/academic>`_ and put it in ``~/mosek/``
+  - Request an `academic license file <http://license.mosek.com/academic>`_ and put it in ``~/mosek/``
 
 Linux
   - Download `MOSEK 8 <https://www.mosek.com/downloads/>`_, then:
       - Move the ``mosek`` folder to your home directory
       - Follow `these steps for Linux <http://docs.mosek.com/7.0/toolsinstall/Linux_UNIX_installation_instructions.html>`_.
-      - Request an `academic license file <http://license.mosek.com/academic>`_ and put it in ``~/mosek/``
+  - Request an `academic license file <http://license.mosek.com/academic>`_ and put it in ``~/mosek/``
 
 Windows
+    - Make sure ``gcc`` is on your system path.
+        - To do this, type ``gcc`` into a command prompt.
+        - If you get ``executable not found``, then install the 64-bit version (x86_64 installer architecture dropdown option) with GCC version 6.4.0 or older of `mingw <http://sourceforge.net/projects/mingw-w64/>`_.
+        - In an Anaconda command prompt (or equivalent), run ``cd C:\Program Files\mingw-w64\x86_64-6.4.0-posix-seh-rt_v5-rev0\`` (or whatever corresponds to the correct installation directory; note that if mingw is in ``Program Files (x86)`` instead of ``Program Files`` you've installed the 32-bit version by mistake)
+        - Run ``mingw-w64`` to add it to your executable path. For step 3 of the install process you'll need to run ``pip install gpkit`` from this prompt.
     - Download `MOSEK 8 <https://www.mosek.com/downloads/>`_, then:
         - Follow `these steps for Windows <http://docs.mosek.com/7.0/toolsinstall/Windows_installation.html>`_.
-        - Request an `academic license file <http://license.mosek.com/academic>`_ and put it in ``C:\Users\(your_username)\mosek\``
-        - Make sure ``gcc`` is on your system path.
-            - To do this, type ``gcc`` into a command prompt.
-            - If you get ``executable not found``, then install the 64-bit version (x86_64 installer architecture dropdown option) with GCC version 6.4.0 or older of `mingw <http://sourceforge.net/projects/mingw-w64/>`_.
-            - In an Anaconda command prompt (or equivalent), run ``cd C:\Program Files\mingw-w64\x86_64-6.4.0-posix-seh-rt_v5-rev0\`` (or whatever corresponds to the correct installation directory; note that if mingw is in ``Program Files (x86)`` instead of ``Program Files`` you've installed the 32-bit version by mistake)
-            - Run ``mingw-w64`` to add it to your executable path. For step 3 of the install process you'll need to run ``pip install gpkit`` from this prompt.
+    - Request an `academic license file <http://license.mosek.com/academic>`_ and put it in ``C:\Users\(your_username)\mosek\``
 
 Debugging your installation
 ===========================

gpkit/constraints/gp.py

@@ -69,6 +69,7 @@ class GeometricProgram:
     >>> gp.solve()
     """
     _result = solve_log = solver_out = model = v_ss = nu_by_posy = None
+    choicevaridxs = integersolve = None
 
     def __init__(self, cost, constraints, substitutions, *, checkbounds=True):
         self.cost, self.substitutions = cost, substitutions
@@ -155,6 +156,8 @@ def gen(self):
                 m_idx += 1
         self.p_idxs = np.array(self.p_idxs, "int32")  # to use array equalities
         self.varidxs = {vk: i for i, vk in enumerate(self.varlocs)}
+        self.choicevaridxs = {vk: i for i, vk in enumerate(self.varlocs)
+                              if vk.choices}
         for j, (var, locs) in enumerate(self.varlocs.items()):
             row.extend(locs)
             col.extend([j]*len(locs))
@@ -191,6 +194,9 @@ def solve(self, solver=None, *, verbosity=1, gen_result=True, **kwargs):
 
         solverargs = DEFAULT_SOLVER_KWARGS.get(solvername, {})
         solverargs.update(kwargs)
+        if self.choicevaridxs and solvername == "mosek_conif":
+            solverargs["choicevaridxs"] = self.choicevaridxs
+            self.integersolve = True
         starttime = time()
         solver_out, infeasibility, original_stdout = {}, None, sys.stdout
         try:
@@ -304,6 +310,27 @@ def _compile_result(self, solver_out):
         result["constants"] = KeyDict(self.substitutions)
         result["variables"] = KeyDict(result["freevariables"])
         result["variables"].update(result["constants"])
+        result["soltime"] = solver_out["soltime"]
+        if self.integersolve:
+            result["choicevariables"] = KeyDict( \
+                {k: v for k, v in result["freevariables"].items()
+                 if k in self.choicevaridxs})
+            result["warnings"] = {"No Dual Solution": [(\
+                "This model has the discretized choice variables"
+                " %s and hence no dual solution. You can fix those variables"
+                " to their optimal value and get sensitivities to the resulting"
+                " continuous problem by updating your model's substitions with"
+                " `sol[\"choicevariables\"]`."
+                % sorted(self.choicevaridxs.keys()), self.choicevaridxs)]}
+            return SolutionArray(result)
+        elif self.choicevaridxs:
+            result["warnings"] = {"Freed Choice Variables": [(\
+                "This model has the discretized choice variables"
+                " %s, but since the '%s' solver doesn't support discretization"
+                " they were treated as continuous variables."
+                % (sorted(self.choicevaridxs.keys()), solver_out["solver"]),
+                self.choicevaridxs)]}
+
         result["sensitivities"] = {"constraints": {}}
         la, self.nu_by_posy = self._generate_nula(solver_out)
         cost_senss = sum(nu_i*exp for (nu_i, exp) in zip(self.nu_by_posy[0],
@@ -342,7 +369,6 @@ def _compile_result(self, solver_out):
         result["sensitivities"]["variables"] = KeyDict(gpv_ss)
         result["sensitivities"]["constants"] = \
             result["sensitivities"]["variables"]  # NOTE: backwards compat.
-        result["soltime"] = solver_out["soltime"]
         return SolutionArray(result)
 
     def check_solution(self, cost, primal, nu, la, tol, abstol=1e-20):
@@ -379,6 +405,8 @@ def almost_equal(num1, num2):
                 raise Infeasible("Primal solution violates constraint: %s is "
                                  "greater than 1" % primal_exp_vals[mi].sum())
         # check dual sol #
+        if self.integersolve:
+            return
         # note: follows dual formulation in section 3.1 of
         # http://web.mit.edu/~whoburg/www/papers/hoburg_phd_thesis.pdf
         if not almost_equal(self.nu_by_posy[0].sum(), 1):

gpkit/small_classes.py

@@ -74,9 +74,10 @@ def __init__(self, output=None, *, verbosity=0):
 
     def write(self, writ):
         "Append and potentially write the new line."
+        if writ[:2] == "b'":
+            writ = writ[2:-1]
         if writ != "\n":
-            writ = writ.rstrip("\n")
-            self.append(str(writ))
+            self.append(writ.rstrip("\n"))
         if self.verbosity > 0:  # pragma: no cover
             self.output.write(writ)
 

gpkit/solution_array.py

@@ -243,7 +243,9 @@ def warnings_table(self, _, **kwargs):
         if len(data_vec) == 0:
             continue
         if not hasattr(data_vec, "shape"):
-            data_vec = [data_vec]
+            data_vec = [data_vec]  # not a sweep
+        if all((data == data_vec[0]).all() for data in data_vec[1:]):
+            data_vec = [data_vec[0]]  # warnings identical across all sweeps
         for i, data in enumerate(data_vec):
             if len(data) == 0:
                 continue
@@ -340,7 +342,8 @@ class SolutionArray(DictOfLists):
     """
     modelstr = ""
     _name_collision_varkeys = None
-    table_titles = {"sweepvariables": "Swept Variables",
+    table_titles = {"choicevariables": "Choice Variables",
+                    "sweepvariables": "Swept Variables",
                     "freevariables": "Free Variables",
                     "constants": "Fixed Variables",  # TODO: change everywhere
                     "variables": "Variables"}
@@ -683,7 +686,7 @@ def table(self, showvars=(),
         -------
         str
         """
-        if sortmodelsbysenss:
+        if sortmodelsbysenss and "sensitivities" in self:
             kwargs["sortmodelsbysenss"] = self["sensitivities"]["models"]
         else:
             kwargs["sortmodelsbysenss"] = False
@@ -700,7 +703,10 @@ def table(self, showvars=(),
         showvars = self._parse_showvars(showvars)
         strs = []
         for table in tables:
-            if table == "cost":
+            if "sensitivities" not in self and ("sensitivities" in table or
+                                                "constraints" in table):
+                continue
+            elif table == "cost":
                 cost = self["cost"]  # pylint: disable=unsubscriptable-object
                 if kwargs.get("latex", None):  # cost is not printed for latex
                     continue