Turned Gumbel off as default (mostly for pytest right now)

config/dICP_config.yaml

@@ -8,7 +8,7 @@ dICP:
 
 
   functionality:
-    gumbel: True        # If true, use Gumbel-Softmax trick for nearest neighour
+    gumbel: False        # If true, use Gumbel-Softmax trick for nearest neighour
   # Not yet implemented
   #  svd: False           # If true, use SVD to solve pt2pt problem, no effect for pt2pl
 

tests/test_ICP.py

@@ -16,10 +16,6 @@
 def max_iterations():
     return 100
 
-@pytest.fixture
-def diff_tolerance():
-    return 0.3
-
 @pytest.fixture
 def tolerance():
     return 1e-10
@@ -36,7 +32,7 @@ def target():
     data_file_path = os.path.join(current_dir, 'data', 'points_map.npy')
     return np.load(data_file_path)
 
-def test_pt2pt_dICP(source, target, max_iterations, diff_tolerance, tolerance):
+def test_pt2pt_dICP(source, target, max_iterations, tolerance):
     """
     Test differentiable point-to-point ICP algorithm.
     """
@@ -67,10 +63,10 @@ def test_pt2pt_dICP(source, target, max_iterations, diff_tolerance, tolerance):
 
     # Check that the transformation is correct
     err_T = se3op.tran2vec(T_ts_true @ np.linalg.inv(T_ts_pred.detach().numpy()))
-    assert(np.linalg.norm(err_T) < diff_tolerance)
+    assert(np.linalg.norm(err_T) < tolerance)
 
     # Check that the transformed source is close to target
-    assert np.allclose(source_transformed.detach().numpy(), target.detach().numpy(), atol=diff_tolerance)
+    assert np.allclose(source_transformed.detach().numpy(), target.detach().numpy(), atol=1e-5)
 
     # Check that the gradient is not none
     T_ts_pred.sum().backward()
@@ -81,7 +77,7 @@ def test_pt2pt_dICP(source, target, max_iterations, diff_tolerance, tolerance):
     # Confirm gradient is not nan
     assert torch.isnan(source.grad).any() == False and torch.isnan(target.grad).any() == False
 
-def test_pt2pl_dICP(source, target, max_iterations, diff_tolerance, tolerance):
+def test_pt2pl_dICP(source, target, max_iterations, tolerance):
     """
     Test differentiable point-to-plane ICP algorithm.
     """
@@ -110,10 +106,10 @@ def test_pt2pl_dICP(source, target, max_iterations, diff_tolerance, tolerance):
 
     # Check that the transformation is correct
     err_T = se3op.tran2vec(T_ts_true @ np.linalg.inv(T_ts_pred.detach().numpy()))
-    assert(np.linalg.norm(err_T) < diff_tolerance)
+    assert(np.linalg.norm(err_T) < tolerance)
 
     # Check that the transformed source is close to target
-    assert np.allclose(source_transformed.detach().numpy(), target[:,:3].detach().numpy(), atol=diff_tolerance)
+    assert np.allclose(source_transformed.detach().numpy(), target[:,:3].detach().numpy(), atol=1e-5)
 
     # Check that the gradient is not none
     T_ts_pred.sum().backward()