rev #1

cacgan/gans/tuning.py

@@ -41,12 +41,6 @@ def tune(dataset, chk=None, n_calls=50, nepochs=10002):
     """
     hparam tune aug cycgan using gaussian process
     the objective is the mean of "mindist" from "prior" mode of cv runs
-
-    :param dataset:
-    :param chk:
-    :param n_calls:
-    :param nepochs:
-    :return:
     """
 
     def objective(hparams):

scripts/imbalanced_dataset.py

@@ -0,0 +1,13 @@
+from cacgan.data import AtmoStructureGroup
+
+if __name__ == '__main__':
+    mg = AtmoStructureGroup.from_csv()
+    amine_gs = mg.group_by("amine")
+    amine_gs = sorted(amine_gs, key=lambda x: len(x), reverse=True)
+    namines = len(amine_gs)
+    namines_unpopular = 0
+    for g in amine_gs:
+        print(g.first_amine, len(g))
+        if len(g) < 5:
+            namines_unpopular += 1
+    print(namines, namines_unpopular)

scripts/test_emd.py

@@ -0,0 +1,28 @@
+import unittest
+
+
+class EmdTest(unittest.TestCase):
+
+    def test_aginst_pyemd(self):
+        from cacgan.sinkhornOT import dummy_distance_matrix, np, torch, DEVICE, emdloss_detach
+        from pyemd import emd
+        from sklearn.preprocessing import normalize
+        a = np.random.random((4, 6))
+        b = np.random.random((4, 6))
+        a = normalize(a, axis=1, norm="l1")
+        b = normalize(b, axis=1, norm="l1")
+        d = dummy_distance_matrix(a.shape[1])
+        pyemd_results = []
+        for i in range(a.shape[0]):
+            pyemd_results.append(emd(a[i], b[i], d))
+
+        ta = torch.tensor(a).to(DEVICE, dtype=torch.float32)
+        tb = torch.tensor(b).to(DEVICE, dtype=torch.float32)
+        td = torch.tensor(d).to(DEVICE, dtype=torch.float32)
+        sinkhorn_results = emdloss_detach(ta, tb, td).cpu().detach().numpy()
+        same = np.allclose(sinkhorn_results, np.array(pyemd_results))
+        self.assertTrue(same)
+
+
+if __name__ == '__main__':
+    unittest.main()

scripts/test_layer.py

@@ -0,0 +1,24 @@
+import unittest
+
+
+class LayerTest(unittest.TestCase):
+
+    def test_filter(self):
+        from cacgan.gans.networks import ConditionalResnetGenerator, torch
+        from cacgan.utils import DEVICE
+        input_size = 6
+        ninput = 3
+        zero_columns = [2, 3]
+        generator_input = torch.rand((ninput, input_size), device=DEVICE)
+        for icol in zero_columns:
+            generator_input[:, icol] = 0
+        crg = ConditionalResnetGenerator(6, input_size, input_size)
+        crg.model.to(device=DEVICE)
+        noise = torch.rand((ninput, input_size), device=DEVICE)
+        out = crg.forward(generator_input, noise)
+        for icol in zero_columns:
+            self.assertTrue(all(out[:, icol] == 0))
+
+
+if __name__ == '__main__':
+    unittest.main()