Distance (#4)

* deleted redundant graph test

* edit_distance implemented

autokeras/bayesian.py

@@ -1,26 +1,49 @@
+import math
 import numpy as np
 from scipy.linalg import cholesky, cho_solve
+from scipy.optimize import linear_sum_assignment
 
 
-def edit_distance(x, y):
-    ret = 0
-    ret += abs(x.n_conv - y.n_conv)
-    ret += abs(x.n_dense - y.n_dense)
+def layer_distance(a, b):
+    return abs(a - b) * 1.0 / max(a, b)
+
+
+def layers_distance(list_a, list_b):
+    len_a = len(list_a)
+    len_b = len(list_b)
+    f = np.zeros((len_a + 1, len_b + 1))
+    f[-1][-1] = 0
+    for i in range(-1, len_a):
+        f[i][-1] = i + 1
+    for j in range(-1, len_b):
+        f[-1][j] = j + 1
+    for i in range(len_a):
+        for j in range(len_b):
+            f[i][j] = min(f[i][j - 1] + 1, f[i - 1][j] + 1, f[i - 1][j - 1] + layer_distance(list_a[i], list_b[j]))
+    return f[len_a][len_b]
 
-    for i in range(min(x.n_conv, y.n_conv)):
-        a = x.conv_widths[i]
-        b = y.conv_widths[i]
-        ret += abs(a - b) / max(a, b)
 
-    for i in range(min(x.n_dense, y.n_dense)):
-        a = x.dense_widths[i]
-        b = y.dense_widths[i]
-        ret += abs(a - b) / max(a, b)
+def skip_connection_distance(a, b):
+    if a[2] != b[2]:
+        return 1.0
+    len_a = abs(a[1] - a[0])
+    len_b = abs(b[1] - b[0])
+    return abs(a[0] - b[0]) + abs(len_a - len_b)
 
-    for connection in x.skip_connections:
-        if connection not in y.skip_connections:
-            ret += 1
 
+def skip_connections_distance(list_a, list_b):
+    distance_matrix = np.zeros((len(list_a), len(list_b)))
+    for i, a in enumerate(list_a):
+        for j, b in enumerate(list_b):
+            distance_matrix[i][j] = skip_connection_distance(a, b)
+    return distance_matrix[linear_sum_assignment(distance_matrix)].sum() + abs(len(list_a) - len(list_b))
+
+
+def edit_distance(x, y):
+    ret = 0
+    ret += layers_distance(x.conv_widths, y.conv_widths)
+    ret += layers_distance(x.dense_widths, y.dense_widths)
+    ret += skip_connections_distance(x.skip_connections, y.skip_connections)
     return ret
 
 

tests/test_bayesian.py

@@ -6,7 +6,7 @@
 def test_edit_distance():
     descriptor1 = Graph(get_add_skip_model()).extract_descriptor()
     descriptor2 = Graph(get_concat_skip_model()).extract_descriptor()
-    assert edit_distance(descriptor1, descriptor2) == 3
+    assert edit_distance(descriptor1, descriptor2) == 2.0
 
 
 def test_gpr():

tests/test_graph.py

@@ -98,19 +98,6 @@ def test_dense_wider():
     assert np.sum(np.abs(output1 - output2)) < 1e-4
 
 
-def test_skip_add():
-    model = get_conv_model()
-    graph = NetworkMorphismGraph(model)
-    graph.to_add_skip_model(graph.layer_to_id[model.layers[1]], graph.layer_to_id[model.layers[4]])
-    new_model = graph.produce_model()
-    input_data = get_conv_data()
-
-    output1 = model.predict_on_batch(input_data).flatten()
-    output2 = new_model.predict_on_batch(input_data).flatten()
-
-    assert np.array_equal(output1, output2)
-
-
 def test_skip_add_over_pooling_stub():
     model = to_stub_model(get_pooling_model())
     graph = Graph(model)
@@ -133,19 +120,6 @@ def test_skip_add_over_pooling():
     assert np.array_equal(output1, output2)
 
 
-def test_skip_concatenate():
-    model = get_add_skip_model()
-    graph = NetworkMorphismGraph(model)
-    graph.to_concat_skip_model(graph.layer_to_id[model.layers[4]], graph.layer_to_id[model.layers[4]])
-    new_model = graph.produce_model()
-    input_data = get_conv_data()
-
-    output1 = model.predict_on_batch(input_data).flatten()
-    output2 = new_model.predict_on_batch(input_data).flatten()
-
-    assert np.sum(np.abs(output1 - output2)) < 3e-1
-
-
 def test_skip_concat_over_pooling_stub():
     model = to_stub_model(get_pooling_model())
     graph = Graph(model)
@@ -170,17 +144,6 @@ def test_skip_concat_over_pooling():
     assert np.sum(np.abs(output1 - output2)) < 4e-1
 
 
-def test_copy_model():
-    model = get_add_skip_model()
-    new_model = NetworkMorphismGraph(model).produce_model()
-    input_data = get_conv_data()
-
-    output1 = model.predict_on_batch(input_data).flatten()
-    output2 = new_model.predict_on_batch(input_data).flatten()
-
-    assert np.sum(output1 - output2) == 0
-
-
 def test_extract_descriptor_add():
     model = get_add_skip_model()
     descriptor = Graph(model).extract_descriptor()