matplotlib-torch 0.1.0__tar.gz

matplotlib_torch-0.1.0/PKG-INFO

@@ -0,0 +1,29 @@
+Metadata-Version: 2.4
+Name: matplotlib-torch
+Version: 0.1.0
+Summary: Autocast PyTorch tensors to numpy/matplotlib-CPU arrays in matplotlib.
+Requires-Python: >=3.9
+Description-Content-Type: text/markdown
+Requires-Dist: matplotlib
+Requires-Dist: torch
+Provides-Extra: test
+Requires-Dist: pytest; extra == "test"
+
+# matplotlib-torch
+
+Transparently pass PyTorch tensors (CPU, CUDA, MPS, with/without grad) to
+matplotlib. Avoid `.detach().cpu().numpy()` boilerplate.
+
+```python
+import matplotlib_torch
+matplotlib_torch.activate()
+
+import matplotlib.pyplot as plt
+import torch
+
+plt.plot(torch.tensor([1., 2., 3.], device='cuda'))   # just works
+```
+
+## Install
+
+    pip install matplotlib-torch

matplotlib_torch-0.1.0/README.md

@@ -0,0 +1,18 @@
+# matplotlib-torch
+
+Transparently pass PyTorch tensors (CPU, CUDA, MPS, with/without grad) to
+matplotlib. Avoid `.detach().cpu().numpy()` boilerplate.
+
+```python
+import matplotlib_torch
+matplotlib_torch.activate()
+
+import matplotlib.pyplot as plt
+import torch
+
+plt.plot(torch.tensor([1., 2., 3.], device='cuda'))   # just works
+```
+
+## Install
+
+    pip install matplotlib-torch

matplotlib_torch-0.1.0/matplotlib_torch/__init__.py

@@ -0,0 +1,47 @@
+"""Autocast PyTorch tensors to numpy arrays for matplotlib.
+
+Strategy: patch ``torch.Tensor.__array__`` so that any numpy conversion of a
+tensor (on any device, with any grad state) detaches, moves to CPU, and returns
+an ndarray. matplotlib's plotting paths all funnel through ``np.array(x)`` /
+``x.__array__()``, so this single hook covers ``plot``, ``scatter``, ``imshow``,
+``bar``, ``hist``, etc.
+
+Notes:
+- This is intentionally a global patch on torch, scoped to when the package is
+  activated. It also silences numpy 2.0's ``copy`` kwarg deprecation warning
+  that affects torch's stock ``__array__``.
+- matplotlib's ``units`` registry is not used: matplotlib's
+  ``cbook._unpack_to_numpy`` runs before unit conversion and erases the torch
+  type, so a registered converter would never be reached.
+"""
+from __future__ import annotations
+
+import torch
+
+__all__ = ["activate", "deactivate", "is_active"]
+
+_ORIGINAL_array = torch.Tensor.__array__
+
+
+def __array__(self, dtype=None, copy=None):
+    # Unconditional detach + cpu: handles CUDA/MPS/CPU, leaf/non-leaf,
+    # requires_grad True/False, in one path. Matches the package contract that
+    # tensors on any device "just work" with matplotlib.
+    arr = self.detach().cpu().numpy()
+    if dtype is not None:
+        arr = arr.astype(dtype)
+    return arr
+
+
+def activate() -> None:
+    """Install the torch-aware ``__array__`` patch."""
+    torch.Tensor.__array__ = __array__
+
+
+def deactivate() -> None:
+    """Restore torch's stock ``__array__``."""
+    torch.Tensor.__array__ = _ORIGINAL_array
+
+
+def is_active() -> bool:
+    return torch.Tensor.__array__ is __array__

matplotlib_torch-0.1.0/matplotlib_torch.egg-info/PKG-INFO

@@ -0,0 +1,29 @@
+Metadata-Version: 2.4
+Name: matplotlib-torch
+Version: 0.1.0
+Summary: Autocast PyTorch tensors to numpy/matplotlib-CPU arrays in matplotlib.
+Requires-Python: >=3.9
+Description-Content-Type: text/markdown
+Requires-Dist: matplotlib
+Requires-Dist: torch
+Provides-Extra: test
+Requires-Dist: pytest; extra == "test"
+
+# matplotlib-torch
+
+Transparently pass PyTorch tensors (CPU, CUDA, MPS, with/without grad) to
+matplotlib. Avoid `.detach().cpu().numpy()` boilerplate.
+
+```python
+import matplotlib_torch
+matplotlib_torch.activate()
+
+import matplotlib.pyplot as plt
+import torch
+
+plt.plot(torch.tensor([1., 2., 3.], device='cuda'))   # just works
+```
+
+## Install
+
+    pip install matplotlib-torch

matplotlib_torch-0.1.0/matplotlib_torch.egg-info/SOURCES.txt

@@ -0,0 +1,9 @@
+README.md
+pyproject.toml
+matplotlib_torch/__init__.py
+matplotlib_torch.egg-info/PKG-INFO
+matplotlib_torch.egg-info/SOURCES.txt
+matplotlib_torch.egg-info/dependency_links.txt
+matplotlib_torch.egg-info/requires.txt
+matplotlib_torch.egg-info/top_level.txt
+tests/test_basic.py

matplotlib_torch-0.1.0/matplotlib_torch.egg-info/dependency_links.txt

@@ -0,0 +1 @@
+

matplotlib_torch-0.1.0/matplotlib_torch.egg-info/requires.txt

@@ -0,0 +1,5 @@
+matplotlib
+torch
+
+[test]
+pytest

matplotlib_torch-0.1.0/matplotlib_torch.egg-info/top_level.txt

@@ -0,0 +1 @@
+matplotlib_torch

matplotlib_torch-0.1.0/pyproject.toml

@@ -0,0 +1,17 @@
+[build-system]
+requires = ["setuptools>=61"]
+build-backend = "setuptools.build_meta"
+
+[project]
+name = "matplotlib-torch"
+version = "0.1.0"
+description = "Autocast PyTorch tensors to numpy/matplotlib-CPU arrays in matplotlib."
+readme = "README.md"
+requires-python = ">=3.9"
+dependencies = ["matplotlib", "torch"]
+
+[project.optional-dependencies]
+test = ["pytest"]
+
+[tool.setuptools.packages.find]
+include = ["matplotlib_torch*"]

matplotlib_torch-0.1.0/setup.cfg

@@ -0,0 +1,4 @@
+[egg_info]
+tag_build = 
+tag_date = 0
+

matplotlib_torch-0.1.0/tests/test_basic.py

@@ -0,0 +1,99 @@
+import matplotlib
+
+matplotlib.use("Agg")  # headless for tests
+import matplotlib.pyplot as plt
+import numpy as np
+import pytest
+import torch
+
+import matplotlib_torch
+
+
+@pytest.fixture(autouse=True)
+def _activate():
+    matplotlib_torch.activate()
+    yield
+    matplotlib_torch.deactivate()
+
+
+def test_is_active_toggle():
+    assert matplotlib_torch.is_active()
+    matplotlib_torch.deactivate()
+    assert not matplotlib_torch.is_active()
+    matplotlib_torch.activate()
+    assert matplotlib_torch.is_active()
+
+
+def test_cpu_tensor():
+    x = torch.arange(5, dtype=torch.float32)
+    (line,) = plt.plot(x)
+    np.testing.assert_array_equal(line.get_xdata(), np.arange(5))
+    np.testing.assert_array_equal(line.get_ydata(), np.arange(5))
+
+
+def test_cuda_tensor():
+    if not torch.cuda.is_available():
+        pytest.skip("no cuda")
+    x = torch.arange(5, dtype=torch.float32, device="cuda")
+    (line,) = plt.plot(x)
+    assert isinstance(line.get_ydata(), np.ndarray)
+    np.testing.assert_array_equal(line.get_ydata(), np.arange(5))
+
+
+def test_grad_tensor():
+    x = torch.arange(5, dtype=torch.float32, requires_grad=True)
+    y = x * 2
+    (line,) = plt.plot(x, y)
+    np.testing.assert_array_equal(line.get_ydata(), np.arange(5) * 2)
+
+
+def test_xy_both_tensors():
+    x = torch.linspace(0, 1, 10)
+    y = torch.linspace(1, 2, 10)
+    (line,) = plt.plot(x, y)
+    np.testing.assert_allclose(line.get_xdata(), np.linspace(0, 1, 10))
+    np.testing.assert_allclose(line.get_ydata(), np.linspace(1, 2, 10))
+
+
+def test_scatter():
+    x = torch.randn(20)
+    y = torch.randn(20)
+    c = torch.rand(20)
+    sc = plt.scatter(x, y, c=c)
+    assert sc.get_offsets().shape == (20, 2)
+
+
+def test_scatter_cuda():
+    if not torch.cuda.is_available():
+        pytest.skip("no cuda")
+    x = torch.randn(20, device="cuda")
+    y = torch.randn(20, device="cuda")
+    sc = plt.scatter(x, y)
+    assert sc.get_offsets().shape == (20, 2)
+
+
+def test_imshow():
+    img = torch.rand(8, 8, 3)
+    im = plt.imshow(img)
+    arr = im.get_array()
+    assert arr.shape == (8, 8, 3)
+
+
+def test_imshow_cuda():
+    if not torch.cuda.is_available():
+        pytest.skip("no cuda")
+    img = torch.rand(8, 8, 3, device="cuda")
+    im = plt.imshow(img)
+    assert im.get_array().shape == (8, 8, 3)
+
+
+def test_non_tensor_passes_through():
+    # plain numpy / lists should still work alongside tensors.
+    (line,) = plt.plot([1, 2, 3], np.array([4, 5, 6]))
+    np.testing.assert_array_equal(line.get_ydata(), [4, 5, 6])
+
+
+def test_scalar_tensor():
+    # A 0-d tensor as a y value should not crash the converter.
+    (line,) = plt.plot([0], torch.tensor(2.0))
+    assert line.get_ydata()[0] == 2.0