Some more cleanup before 0.1.3 (#353)

* Fixed some test paths.

* Removed dependency on numdifftools.

* Fixed code causing warnings in unit tests.

* Replaced deprecated torch.triangular_solve with torch.linalg.solve_triangular.

examples/README.md

@@ -37,11 +37,6 @@ The motion planning and tactile estimation examples require `hydra` installation
 ```bash
 pip install hydra-core
 ```
-The backward modes example requires `numdifftools`, which you can install with
-
-```bash
-pip install numdifftools
-```
 
 The homography example requires `hydra`, `kornia` and `OpenCV`, which you can install with
 

examples/backward_modes.py

@@ -11,7 +11,6 @@
 import time
 from collections import defaultdict
 
-import numdifftools as nd
 import numpy as np
 import torch
 
@@ -146,20 +145,21 @@ def quad_error_fn(optim_vars, aux_vars):
 
 
 # Next we numerically check the derivative
-def fit_x(data_x_np):
-    theseus_inputs["x"] = (
-        torch.from_numpy(data_x_np).float().clone().requires_grad_().unsqueeze(0)
+with torch.no_grad():
+
+    def fn(data_x_torch):
+        theseus_inputs["x"] = data_x_torch
+        updated_inputs, _ = theseus_optim.forward(
+            theseus_inputs,
+            optimizer_kwargs={"track_best_solution": True, "verbose": False},
+        )
+        return updated_inputs["a"]
+
+    g = (
+        torch.autograd.functional.jacobian(fn, data_x.detach())[0, 0, 0]
+        .double()
+        .numpy()
     )
-    updated_inputs, _ = theseus_optim.forward(
-        theseus_inputs, optimizer_kwargs={"track_best_solution": True, "verbose": False}
-    )
-    return updated_inputs["a"].item()
-
-
-data_x_np = data_x.detach().clone().numpy()
-dfit_x = nd.Gradient(fit_x)
-g = dfit_x(data_x_np)
-
 print("\n--- Numeric derivative")
 print(g)
 

requirements/dev.txt

@@ -6,7 +6,6 @@ pre-commit>=2.9.2
 isort>=5.6.4
 differentiable-robot-model>=0.2.3
 types-PyYAML==5.4.3
-numdifftools>=0.9.40
 mock>=4.0.3
 types-mock>=4.0.8
 Sphinx==5.0.2

requirements/docs.txt

@@ -4,7 +4,6 @@ scipy>=1.5.3
 scikit-sparse>=0.4.5
 torch>=1.11
 pytest>=6.2.1
-numdifftools>=0.9.40
 pybind11>=2.7.1
 mock>=4.0.3
 Sphinx==5.0.2

tests/core/test_cost_weight.py

@@ -26,7 +26,7 @@ def test_copy_scale_cost_weight():
 
 
 def test_copy_diagonal_cost_weight():
-    diagonal = th.Variable(torch.ones(3))
+    diagonal = th.Variable(torch.ones(1, 3))
     p1 = th.DiagonalCostWeight(diagonal, name="diagonal_cost_weight")
     for the_copy in [p1.copy(), copy.deepcopy(p1)]:
         check_another_theseus_function_is_copy(p1, the_copy, new_name=f"{p1.name}_copy")
@@ -89,9 +89,9 @@ def _check(cw):
                 assert error.allclose(expected_err)
                 assert jacobians[0].allclose(expected_jac)
 
+            diagonal = diagonal.unsqueeze(0)  # add batch dimension
             _check(th.DiagonalCostWeight(diagonal.tolist()))
             _check(th.DiagonalCostWeight(diagonal))
-            _check(th.DiagonalCostWeight(diagonal.view(1, dim)))
             _check(th.DiagonalCostWeight(th.Variable(diagonal)))
 
             batched_diagonal = torch.randn(batch_size, dim)

tests/core/test_vectorizer.py

@@ -112,7 +112,7 @@ def test_correct_schemas_and_shared_vars():
     objective.add(cf6)
 
     # Not grouped with anything cf1 and cf2 because weight type is different
-    w7 = th.DiagonalCostWeight([1.0])
+    w7 = th.DiagonalCostWeight([[1.0]])
     cf7 = th.Difference(v1, tv, w7)
     objective.add(cf7)
 
@@ -178,7 +178,7 @@ def test_vectorized_error():
             for i in range(rng.choice([1, 10]))
         ]
         s_target = th.SE3.rand(1, generator=generator)
-        ws = th.DiagonalCostWeight(torch.randn(6, generator=generator))
+        ws = th.DiagonalCostWeight(torch.randn(1, 6, generator=generator))
         for s in se3s:
             objective.add(th.Difference(s, s_target, ws))
 

tests/embodied/collision/test_eff_obj_contact.py

@@ -86,7 +86,7 @@ def _create_sdf_data(sdf_idx=0):
     cell_size = 0.01
     origin_x, origin_y = 0.0, 0.0
     sdf_data = np.loadtxt(
-        open("theseus/embodied/collision/tests/sdf_data.csv", "rb"),
+        open("tests/embodied/collision/sdf_data.csv", "rb"),
         delimiter=",",
         skiprows=0,
     )

tests/embodied/measurements/test_between.py

@@ -86,9 +86,7 @@ def test_error_between_so2():
     ).unsqueeze(1)
 
     num_val = len(so2_data)
-    between_errors = np.load(
-        "theseus/embodied/measurements/tests/between_errors_so2.npy"
-    )
+    between_errors = np.load("tests/embodied/measurements/between_errors_so2.npy")
     z = 0
     # between_errors was generated using all 3-combinations of the 5 SO2 above
     for i in range(num_val):
@@ -113,9 +111,7 @@ def test_error_between_se2():
         ]
     )
     num_val = len(se2_data)
-    between_errors = np.load(
-        "theseus/embodied/measurements/tests/between_errors_se2.npy"
-    )
+    between_errors = np.load("tests/embodied/measurements/between_errors_se2.npy")
     z = 0
     # between_errors was generated using all 3-combinations of the 5 SE2 above
     for i in range(num_val):

tests/embodied/misc/test_variable_difference.py

@@ -79,7 +79,7 @@ def test_error_local_cost_fn_so2():
     ).unsqueeze(1)
 
     num_val = len(so2_data)
-    sq_dist_errors = np.load("theseus/embodied/misc/tests/sq_dist_errors_so2.npy")
+    sq_dist_errors = np.load("tests/embodied/misc/sq_dist_errors_so2.npy")
     k = 0
     # sq_dist_errors was generated using all 2-combinations of the 5 SO2 above
     for i in range(num_val):
@@ -102,7 +102,7 @@ def test_error_local_cost_fn_se2():
         ]
     )
     num_val = len(se2_data)
-    sq_dist_errors = np.load("theseus/embodied/misc/tests/sq_dist_errors_se2.npy")
+    sq_dist_errors = np.load("tests/embodied/misc/sq_dist_errors_se2.npy")
     k = 0
     # sq_dist_errors was generated using all 2-combinations of the 5 SE2 above
     for i in range(num_val):

tests/optimizer/nonlinear/test_backwards.py

@@ -3,7 +3,6 @@
 # This source code is licensed under the MIT license found in the
 # LICENSE file in the root directory of this source tree.
 
-import numdifftools as nd
 import pytest  # noqa: F401
 import torch
 
@@ -63,21 +62,19 @@ def quad_error_fn(optim_vars, aux_vars):
 
 
 def test_backwards():
-    # First we use numdifftools to numerically compute the gradient
+    # First we use torch.autograd.functional to numerically compute the gradient
     # the optimal a w.r.t. the x part of the data
-    def fit_x(data_x_np):
-        theseus_inputs["x"] = (
-            torch.from_numpy(data_x_np).float().clone().requires_grad_().unsqueeze(0)
-        )
-        updated_inputs, _ = theseus_optim.forward(
-            theseus_inputs,
-            optimizer_kwargs={"track_best_solution": True, "verbose": False},
-        )
-        return updated_inputs["a"].item()
-
-    data_x_np = data_x.detach().clone().numpy()
-    dfit_x = nd.Gradient(fit_x)
-    da_dx_numeric = torch.from_numpy(dfit_x(data_x_np)).float()
+    with torch.no_grad():
+
+        def fn(data_x_torch):
+            theseus_inputs["x"] = data_x_torch
+            updated_inputs, _ = theseus_optim.forward(
+                theseus_inputs,
+                optimizer_kwargs={"track_best_solution": True, "verbose": False},
+            )
+            return updated_inputs["a"]
+
+        da_dx_numeric = torch.autograd.functional.jacobian(fn, data_x.detach())[0, 0, 0]
 
     theseus_inputs["x"] = data_x
     updated_inputs, _ = theseus_optim.forward(

tests/test_dlm_perturbation.py

@@ -93,7 +93,7 @@ def test_backward_pass_se3_runs():
     loss0 = None
     for _ in range(5):
         adam.zero_grad()
-        with th.enable_lie_tangent():
+        with th.enable_lie_tangent(), th.no_lie_group_check(silent=True):
             out, _ = layer.forward(
                 {"target": target_data},
                 optimizer_kwargs={

tests/test_theseus_layer.py

@@ -248,7 +248,7 @@ def _run_optimizer_test(
     # need to pass updated values through "objective.update". I'm doing it this way
     # to check that update works properly
     cost_weight = th.DiagonalCostWeight(
-        th.Variable(torch.empty(num_points), name="cost_weight_values")
+        th.Variable(torch.empty(1, num_points), name="cost_weight_values")
     )
 
     # Here we create the outer loop models and optimizers for the cost weight

theseus/theseus_layer.py

@@ -117,7 +117,7 @@ def compute_samples(
             mean=torch.zeros((n_vars, n_samples), device=delta.device),
             std=torch.ones((n_vars, n_samples), device=delta.device),
         )
-        delta_samples = (torch.triangular_solve(y, sqrt_AtA).solution) + (
+        delta_samples = (torch.linalg.solve_triangular(sqrt_AtA, y, upper=True)) + (
             delta.unsqueeze(-1)
         ).repeat(1, 1, n_samples)