Change test in test_unigram_03().

src/icosagon/convlayer.py

@@ -68,33 +68,6 @@ class DecagonLayer(torch.nn.Module):
         self.next_layer_repr[fam.node_type_row].append(
             Convolutions(fam.node_type_column, convolutions))
 
-    # def build_fam_two_node_types_sym(self, fam) -> None:
-    #     convolutions_row = torch.nn.ModuleList()
-    #     convolutions_column = torch.nn.ModuleList()
-    #
-    #     if self.input_dim[fam.node_type_column] != \
-    #         self.input_dim[fam.node_type_row]:
-    #         raise ValueError('input_dim for row and column must be equal for a symmetric family')
-    #
-    #     if self.output_dim[fam.node_type_column] != \
-    #         self.output_dim[fam.node_type_row]:
-    #         raise ValueError('output_dim for row and column must be equal for a symmetric family')
-    #
-    #     for r in fam.relation_types:
-    #         assert r.adjacency_matrix is not None and \
-    #             r.adjacency_matrix_backward is not None
-    #         conv = DropoutGraphConvActivation(self.input_dim[fam.node_type_column],
-    #             self.output_dim[fam.node_type_row], r.adjacency_matrix,
-    #             self.keep_prob, self.rel_activation)
-    #         convolutions_row.append(conv)
-    #         convolutions_column.append(conv.clone(r.adjacency_matrix_backward))
-    #
-    #     self.next_layer_repr[fam.node_type_row].append(
-    #         Convolutions(fam.node_type_column, convolutions_row))
-    #
-    #     self.next_layer_repr[fam.node_type_column].append(
-    #         Convolutions(fam.node_type_row, convolutions_column))
-
     def build_fam_two_node_types(self, fam) -> None:
         convolutions_row = torch.nn.ModuleList()
         convolutions_column = torch.nn.ModuleList()
@@ -118,12 +91,6 @@ class DecagonLayer(torch.nn.Module):
         self.next_layer_repr[fam.node_type_column].append(
             Convolutions(fam.node_type_row, convolutions_column))
 
-    # def build_fam_two_node_types(self, fam) -> None:
-    #     if fam.is_symmetric:
-    #         self.build_fam_two_node_types_sym(fam)
-    #     else:
-    #         self.build_fam_two_node_types_asym(fam)
-
     def build_family(self, fam) -> None:
         if fam.node_type_row == fam.node_type_column:
             self.build_fam_one_node_type(fam)

tests/icosagon/test_sampling.py

@@ -118,7 +118,7 @@ def test_unigram_02():
 def test_unigram_03():
     range_max = 7
     distortion = 0.75
-    batch_size = 25
+    batch_size = 2500
     unigrams = [ 1, 3, 2, 1, 2, 1, 3]
     num_true = 1
 
@@ -129,8 +129,8 @@ def test_unigram_03():
     true_classes_tf = tf.convert_to_tensor(true_classes)
     true_classes_torch = torch.tensor(true_classes)
 
-    counts_tf = defaultdict(list)
-    counts_torch = defaultdict(list)
+    counts_tf = torch.zeros(range_max)
+    counts_torch = torch.zeros(range_max)
 
     for i in range(10):
         neg_samples, _, _ = tf.nn.fixed_unigram_candidate_sampler(
@@ -142,29 +142,25 @@ def test_unigram_03():
             distortion=distortion,
             unigrams=unigrams)
 
-        counts = defaultdict(int)
+        counts = torch.zeros(range_max)
         with tf.Session() as sess:
             neg_samples = neg_samples.eval()
         for x in neg_samples:
             counts[x.item()] += 1
-        for k, v in counts.items():
-            counts_tf[k].append(v)
+        counts_tf += counts
 
         neg_samples = icosagon.sampling.fixed_unigram_candidate_sampler(
             true_classes=true_classes,
             distortion=distortion,
             unigrams=unigrams)
 
-        counts = defaultdict(int)
+        counts = torch.zeros(range_max)
         for x in neg_samples:
             counts[x.item()] += 1
-        for k, v in counts.items():
-            counts_torch[k].append(v)
+        counts_torch += counts
 
-    for i in range(range_max):
-        print('counts_tf[%d]:' % i, counts_tf[i])
-        print('counts_torch[%d]:' % i, counts_torch[i])
+    print('counts_tf:', counts_tf)
+    print('counts_torch:', counts_torch)
 
-    for i in range(range_max):
-        statistic, pvalue = scipy.stats.ttest_ind(counts_tf[i], counts_torch[i])
-        assert pvalue * range_max > .05
+    distance = scipy.stats.wasserstein_distance(counts_tf, counts_torch)
+    assert distance < 2000