Feature/init plan util #27

alexisthual · 2023-05-27T12:04:54Z

Add util function to initialise ot plan.

… early

…op early

… into feature/allow_not_setting_convergence_criteria

…ub.com:alexisthual/fugw into feature/init_plan_util

…ub.com:alexisthual/fugw into feature/allow_not_setting_convergence_criteria

…ub.com:alexisthual/fugw into feature/init_plan_util

pbarbarant

Excellent overall 👏, I just found a typo and I'm concerned about the use of torch.float32 that might lead to casting errors later on.

pbarbarant · 2023-05-28T17:33:43Z

src/fugw/utils.py

+        xb = torch.rand(n_target)
+        plan = emd_1d(xa, xb).to(dtype=torch.float32)
+    else:
+        raise Exception(f"Unknown initialisation method {method}")


Typo initialisation -> initialization

This is my attempt to defend UK English over US English ;)

pbarbarant · 2023-05-28T17:39:05Z

src/fugw/utils.py

+            "Number of source and target points must be equal "
+            "when using identity initialization."
+        )
+        plan = torch.eye(n_source, dtype=torch.float32)


Maybe you should replace dtype=torch.float32 with dtype=torch.float64 to avoid casting issues.

The default type we use in _make_tensor() is torch.float32 (or torch.int32), which is why I chose this type here 🙂

fugw/src/fugw/utils.py

Lines 38 to 55 in ffb6fc0

def _make_tensor(x, device=None, dtype=None):

"""Turn x into a torch.Tensor with suited type and device."""

if isinstance(x, np.ndarray):

tensor = torch.tensor(x)

elif isinstance(x, torch.Tensor):

tensor = x

else:

raise Exception(f"Expected np.ndarray or torch.Tensor, got {type(x)}")

# By default, cast x to float32 or int32

# depending on its original type

if dtype is None:

if tensor.is_floating_point():

dtype = torch.float32

else:

dtype = torch.int32

return tensor.to(device=device, dtype=dtype)

Okay let's keep float32 for coherence then ! 👍

pbarbarant · 2023-05-28T17:39:16Z

src/fugw/utils.py

+        plan = plan / plan.sum()
+    elif method == "entropic":
+        if weights_source is None:
+            weights_source = torch.ones(n_source, dtype=torch.float32)


pbarbarant · 2023-05-28T17:39:26Z

src/fugw/utils.py

+        if weights_source is None:
+            weights_source = torch.ones(n_source, dtype=torch.float32)
+        if weights_target is None:
+            weights_target = torch.ones(n_target, dtype=torch.float32)


pbarbarant · 2023-05-28T17:39:40Z

src/fugw/utils.py

+    elif method == "permutation":
+        xa = torch.rand(n_source)
+        xb = torch.rand(n_target)
+        plan = emd_1d(xa, xb).to(dtype=torch.float32)


…plan_util

alexisthual added 20 commits May 27, 2023 12:33

Bugfix init plan not being on device

51ee308

Allow not setting convergence criteria and add test

017d492

Rename early_stopping_threshold and add tests

c7e3d31

Bugfix test

95265db

Make stopping criteria more explicit in solvers and allow to not stop…

ed9133b

… early

Make stopping criteria more explicit in bcd iters and allow to not st…

d152cc9

…op early

Convert sparse init plan to CSR matrix

f941c81

Add POT to dependencies

80703d4

Add util function for initializing dense ot plan

478376f

Add tests

03459bc

Fix test name and torch eye

237853f

Fix marginal constraints test

c98ca0a

Add assertion failed error message

97639d5

Use init plan util in dense mapping

e407081

Bugfix missing var name in util function

e640617

Merge branch 'bugfix/init_plan_device' of github.com:alexisthual/fugw…

92ba6c9

… into feature/allow_not_setting_convergence_criteria

Merge branch 'feature/allow_not_setting_convergence_criteria' of gith…

19c6b31

…ub.com:alexisthual/fugw into feature/init_plan_util

Fix incorrect stoping criterium

6c4d2ef

Merge branch 'feature/allow_not_setting_convergence_criteria' of gith…

ddfd780

…ub.com:alexisthual/fugw into feature/allow_not_setting_convergence_criteria

Merge branch 'feature/allow_not_setting_convergence_criteria' of gith…

d8a6e58

…ub.com:alexisthual/fugw into feature/init_plan_util

pbarbarant reviewed May 28, 2023

View reviewed changes

pbarbarant approved these changes Jun 5, 2023

View reviewed changes

Base automatically changed from feature/allow_not_setting_convergence_criteria to main June 5, 2023 10:59

Merge branch 'main' of github.com:alexisthual/fugw into feature/init_…

b1803ca

…plan_util

alexisthual merged commit 8775aac into main Jun 5, 2023

alexisthual deleted the feature/init_plan_util branch June 5, 2023 11:39

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Feature/init plan util #27

Feature/init plan util #27

alexisthual commented May 27, 2023

pbarbarant left a comment

pbarbarant May 28, 2023

alexisthual Jun 5, 2023

pbarbarant May 28, 2023

alexisthual Jun 5, 2023

pbarbarant Jun 5, 2023

pbarbarant May 28, 2023

pbarbarant May 28, 2023

pbarbarant May 28, 2023

	def _make_tensor(x, device=None, dtype=None):
	"""Turn x into a torch.Tensor with suited type and device."""
	if isinstance(x, np.ndarray):
	tensor = torch.tensor(x)
	elif isinstance(x, torch.Tensor):
	tensor = x
	else:
	raise Exception(f"Expected np.ndarray or torch.Tensor, got {type(x)}")

	# By default, cast x to float32 or int32
	# depending on its original type
	if dtype is None:
	if tensor.is_floating_point():
	dtype = torch.float32
	else:
	dtype = torch.int32

	return tensor.to(device=device, dtype=dtype)

Feature/init plan util #27

Feature/init plan util #27

Conversation

alexisthual commented May 27, 2023

pbarbarant left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment