aistats algorithm added

README.md

@@ -1,6 +1,6 @@
 # Complexity
 *Authors*: Ricardo Vilalta, Dainis Boumber, Michael Meskhi.
 
-**Description **
+
 
 TODO

complexity.py

@@ -46,7 +46,9 @@ def plot_ds(grid_size, loc, X, y, xx, yy, title, seeds=None, colspan=1, rowspan=
 '''
     Perform Active Learning
     QueryStrategy (Random Sampling or Uncertainty Sampling)
+
 '''
+
 def active(classifiers, datasets, experiments, query_strat, quota=25, plot_every_n=5):
     for dataset_index, ((X_src, y_src), (X_tgt, y_tgt)) in enumerate(datasets):
         u_tgt = [None] * len(X_tgt)

modules/active_da.py

@@ -0,0 +1,105 @@
+# Necessities
+import numpy as np
+import random
+import scipy.spatial
+import sklearn.metrics as metr
+from numpy.core.multiarray import ndarray
+
+import modules.util as u
+from modules.oracle import Oracle
+
+#untested but should work
+# example usage (pseudocodish):
+# cda = CADA(MNIST.X, MNIST.y)
+# ixs = query(MNIST_M.X, 100)
+# X_train, y_train = MNIST_M.X[ixs], MNIST_M.y[ixs]
+# classifier.fit(X_train, y_train)
+# classifier.predict(MNIST_M.X[-ixs])
+
+class CADA(object):
+    '''
+    1.  Compute complexity measure on source domain at different levels of locality (size of neighborhood).
+    2. Choose the single neighborhood size K* that keeps entropy below a predefined threshold.
+    3. Sample N examples from target domain randomly.
+    4. Grow a window around each example of size K*.
+    5. All examples within each window are banned from sampling.
+    6. Go to step 3 until no more examples are left for sampling.
+    7. Create a model with the queried examples on target. (outside the scope of this class, can use any model you want)
+    '''
+    def __init__(self, source_X, source_y, max_entropy=0.1, f_samples=0.01, window_growth_rate=0.1, mink=1, maxk=-1, n_samples=-1):
+        assert(len(source_X)==len(source_y))
+        assert(f_samples <= 1.0 and max_entropy <= 1.0)
+        assert (n_samples < len(source_y))
+        assert(mink > 0 and maxk < len(source_y))
+        #all sane
+
+        #how many do we actually sample in Step 3
+        self.source_X = source_X
+        self.source_y = source_y
+        #build the tree of the distribution that we do nn on
+        self.tree = scipy.spatial.cKDTree(self.source_X, leafsize=32, compact_nodes=False, balanced_tree=False)
+        print('build done')
+        if n_samples == -1:
+            self.n_samples = int(len(source_y) * f_samples)
+            if self.n_samples == 0:
+                self.n_samples = 1
+        else:
+            self.n_samples = n_samples
+        #init sampling
+        self.seeds = np.random.random_integers(0, len(source_X) - 1, self.n_samples)
+        self.classes = set(source_y)
+        stepsize = int(maxk * window_growth_rate)
+        if stepsize == 0:
+            stepsize = 1
+        self.Ks = np.arange(mink, maxk, step=stepsize)  # ckdTree starts counting from 1
+        self.Hs = np.zeros(len(self.Ks))
+        self.ws = np.zeros((len(self.seeds), len(self.Ks)))
+        self.banned = np.zeros((self.seeds, len(self.Ks)))
+        self.K = 0
+
+        for i, k in enumerate(self.Ks):
+            for j, seed in enumerate(self.seeds):
+                self.ws[j, i] = self._H(k=k, seed=seed)
+
+            # add up entropy for each window as they grow
+            self.Hs[i] = np.sum(self.ws[:, i]) / len(self.seeds)
+
+            if self.Hs[i] > max_entropy:
+                if i > 0:
+                    self.K = self.Ks[i-1]
+                break # done with step 2
+
+    # returns indices into target_X
+
+    def query(self, target_X, N):
+        if N > len(target_X):
+            raise AttributeError
+
+        target_banned = np.zeros(len(target_X))
+        queried_example_indices = []
+        while 0 in target_banned: # step 6 check
+            not_banned_ix = [i for i, banned in enumerate(target_banned) if banned != 1]
+            example_indices = random.sample(not_banned_ix, min(len(not_banned_ix), N)) # step 4
+            queried_example_indices.extend(example_indices)
+
+            for example_ix in example_indices:
+                _, ii = self._nearest_neighbors(self.K, example_ix) # step 4
+                target_banned[ii.flatten().squeeze()] = 1 # step 5
+
+        return queried_example_indices
+
+
+    def _nearest_neighbors(self, k, seed):
+        return self.tree.query(self.source_X[seed, :], k=k, n_jobs=-1)
+
+    def _H(self, k, seed):
+        H = 0
+        _, ii = self._nearest_neighbors(k, seed)
+        neighbors = self.source_y[ii]
+        for c in self.classes:
+            same_c = np.where(neighbors == c)[0]
+            r = len(same_c)/float(k)
+            if r > 0:
+                H += (r * np.log2(r))
+        return -H
+