heavyedge-dataset 1.0.0.post0__py3-none-any.whl → 1.1.0__py3-none-any.whl

heavyedge_dataset/__init__.py

@@ -9,10 +9,13 @@ import numbers
 from collections.abc import Sequence
 
 import numpy as np
+from heavyedge.api import landmarks_type3
 from torch.utils.data import Dataset
 
 __all__ = [
     "ProfileDataset",
+    "PseudoLandmarkDataset",
+    "MathematicalLandmarkDataset",
 ]
 
 
@@ -47,8 +50,10 @@ class ProfileDataset(Dataset):
     (22, 2, 3200)
     >>> lengths.shape
     (22,)
+    >>> import matplotlib.pyplot as plt  # doctest: +SKIP
+    ... plt.plot(*profiles.transpose(1, 2, 0))
 
-    Should this dataset be used for :class:`torch.utils.data.DataLoader`,
+    Should the dataset be used for :class:`torch.utils.data.DataLoader`,
     ``collate_fn`` argument should be passed to the data loader.
 
     >>> from torch.utils.data import DataLoader
@@ -117,3 +122,151 @@ class ProfileDataset(Dataset):
     def __getitems__(self, idxs):
         # PyTorch API
         return self.__getitem__(idxs)
+
+
+class PseudoLandmarkDataset(Dataset):
+    """Dataset for pseudo-landmarks of edge profiles
+
+    Parameters
+    ----------
+    file : heavyedge.ProfileData
+        Open hdf5 file.
+    m : {1, 2}
+        Dimension of landmark coordinates.
+    k : int
+        Number of landmarks to sample.
+    transform : callable, optional
+        Optional transformation to be applied on samples.
+
+    Examples
+    --------
+    >>> from heavyedge import ProfileData, get_sample_path
+    >>> from heavyedge_dataset import PseudoLandmarkDataset
+    >>> with ProfileData(get_sample_path("Prep-Type1.h5")) as file:
+    ...     dataset = PseudoLandmarkDataset(file, 1, 10)
+    ...     data = dataset[:]
+    >>> data.shape
+    (18, 1, 10)
+    >>> import matplotlib.pyplot as plt  # doctest: +SKIP
+    ... plt.plot(*data.transpose(1, 2, 0))
+
+    Because sampling pseudo-landmark requires loading full profile data,
+    loading large dataset can cause memory failure even if the output data is managable.
+    This can be avoided by batched loading with :class:`torch.utils.data.DataLoader`.
+
+    >>> import numpy as np
+    >>> from torch.utils.data import DataLoader
+    >>> with ProfileData(get_sample_path("Prep-Type1.h5")) as file:
+    ...     dataset = PseudoLandmarkDataset(file, 1, 10)
+    ...     loader = DataLoader(dataset, batch_size=10)
+    ...     data = np.concatenate(list(loader))
+    >>> data.shape
+    (18, 1, 10)
+    """
+
+    def __init__(self, file, m, k, transform=None):
+        self.profiles = ProfileDataset(file, m=m)
+        self.k = k
+        self.transform = transform
+
+    def __len__(self):
+        return len(self.profiles)
+
+    def __getitem__(self, idx):
+        if isinstance(idx, numbers.Integral):
+            Y, L = self.profiles[idx]
+            Ys, Ls = [Y], [L]
+        else:
+            Ys, Ls = self.profiles[idx]
+
+        X = []
+        for Y, L in zip(Ys, Ls):
+            idxs = np.linspace(0, L - 1, self.k, dtype=int)
+            X.append(Y[:, idxs])
+        ret = np.array(X)
+        if self.transform is not None:
+            ret = self.transform(ret)
+        return ret
+
+    def __getitems__(self, idxs):
+        # PyTorch API
+        return self.__getitem__(idxs)
+
+
+class MathematicalLandmarkDataset(Dataset):
+    """Dataset for mathematical landmarks of edge profiles.
+
+    Loads data as a tuple of two numpy arrays:
+
+    1. Landmark coordinates, shape: (N, m, 4).
+    2. Average plateau height, shape: (N,).
+
+    N is the number of loaded data and m is dimension of coordinates.
+
+    Parameters
+    ----------
+    file : heavyedge.ProfileData
+        Open hdf5 file.
+    m : {1, 2}
+        Dimension of landmark coordinates.
+    sigma : scalar
+        Standard deviation of Gaussian kernel for landmark detection.
+    transform : callable, optional
+        Optional transformation to be applied on samples.
+
+    Examples
+    --------
+    >>> from heavyedge import ProfileData, get_sample_path
+    >>> from heavyedge_dataset import MathematicalLandmarkDataset
+    >>> with ProfileData(get_sample_path("Prep-Type3.h5")) as file:
+    ...     dataset = MathematicalLandmarkDataset(file, 2, 32)
+    ...     landmarks, height = dataset[:]
+    >>> landmarks.shape
+    (35, 2, 4)
+    >>> height.shape
+    (35,)
+
+    Because sampling mathematical landmark requires loading full profile data,
+    loading large dataset can cause memory failure even if the output data is managable.
+    This can be avoided by batched loading with :class:`torch.utils.data.DataLoader`.
+    Note that ``collate_fn`` argument should be passed to the data loader.
+
+    >>> import numpy as np
+    >>> from torch.utils.data import DataLoader
+    >>> with ProfileData(get_sample_path("Prep-Type3.h5")) as file:
+    ...     dataset = MathematicalLandmarkDataset(file, 2, 32)
+    ...     loader = DataLoader(dataset, batch_size=10, collate_fn=lambda x: x)
+    ...     landmarks = np.concatenate([lm for lm, _ in loader])
+    >>> landmarks.shape
+    (35, 2, 4)
+    """
+
+    def __init__(self, file, m, sigma, transform=None):
+        self.profiles = ProfileDataset(file, m=m)
+        self.sigma = sigma
+        self.transform = transform
+
+    def __len__(self):
+        return len(self.profiles)
+
+    def __getitem__(self, idx):
+        if isinstance(idx, numbers.Integral):
+            Y, L = self.profiles[idx]
+            Ys, Ls = [Y], [L]
+        else:
+            Ys, Ls = self.profiles[idx]
+
+        X, H = [], []
+        for Y, L in zip(Ys, Ls):
+            idxs = np.flip(landmarks_type3(Y[-1, :L], self.sigma))
+            X.append(Y[:, idxs])
+            H.append(np.mean(Y[-1, : idxs[0]]))
+
+        ret = np.array(X), np.array(H)
+        if self.transform is not None:
+            ret = self.transform(ret)
+        return ret
+
+    def __getitems__(self, idxs):
+        # PyTorch API
+        return self.__getitem__(idxs)

{heavyedge_dataset-1.0.0.post0.dist-info → heavyedge_dataset-1.1.0.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: heavyedge-dataset
-Version: 1.0.0.post0
+Version: 1.1.0
 Summary: PyTorch-compatible edge profile dataset API
 Author-email: Jisoo Song <jeesoo9595@snu.ac.kr>
 License-Expression: MIT
@@ -28,6 +28,7 @@ Provides-Extra: doc
 Requires-Dist: sphinx; extra == "doc"
 Requires-Dist: numpydoc; extra == "doc"
 Requires-Dist: pydata_sphinx_theme; extra == "doc"
+Requires-Dist: matplotlib; extra == "doc"
 Provides-Extra: dev
 Requires-Dist: flake8; extra == "dev"
 Requires-Dist: black; extra == "dev"
@@ -48,7 +49,7 @@ Package to load edge profile data as PyTorch dataset.
 
 ## Usage
 
-HeavyEdge-Dataset provides `ProfileDataset` which wraps profile data file.
+HeavyEdge-Dataset provides dataset classes profile data file.
 
 A simple use case to load two-dimensional coordinates of profiles and their lengths:
 

heavyedge_dataset-1.1.0.dist-info/RECORD

@@ -0,0 +1,6 @@
+heavyedge_dataset/__init__.py,sha256=V7EonirGRfibsJAQbrEmpOpz53v4zpylWHR2zrMaNPE,8772
+heavyedge_dataset-1.1.0.dist-info/licenses/LICENSE,sha256=pBq2E7YJkUcEINdYeERL4RVFOQICd_MwJq6OusuAPGc,1066
+heavyedge_dataset-1.1.0.dist-info/METADATA,sha256=38Jf4MZANfXyZNQS2GH_JZZ4MzrSzAhN2UeqWTFMzK8,3457
+heavyedge_dataset-1.1.0.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
+heavyedge_dataset-1.1.0.dist-info/top_level.txt,sha256=wpRFI8TlkYFGetc17appkyybauBvzhGGvyueunsdJTc,18
+heavyedge_dataset-1.1.0.dist-info/RECORD,,

heavyedge_dataset-1.0.0.post0.dist-info/RECORD

@@ -1,6 +0,0 @@
-heavyedge_dataset/__init__.py,sha256=CmquaxuwXJYTCGInmzI8Tp0Z9hJBVtEK9eH3NLyoREY,3737
-heavyedge_dataset-1.0.0.post0.dist-info/licenses/LICENSE,sha256=pBq2E7YJkUcEINdYeERL4RVFOQICd_MwJq6OusuAPGc,1066
-heavyedge_dataset-1.0.0.post0.dist-info/METADATA,sha256=34mMpniE5Nn4E6mwdVKhT4m6mnEUKRoCce9Tv6aGNHI,3434
-heavyedge_dataset-1.0.0.post0.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
-heavyedge_dataset-1.0.0.post0.dist-info/top_level.txt,sha256=wpRFI8TlkYFGetc17appkyybauBvzhGGvyueunsdJTc,18
-heavyedge_dataset-1.0.0.post0.dist-info/RECORD,,