np.float -> float

conda/das-osxarm/build.sh

@@ -6,7 +6,7 @@ export PIP_IGNORE_INSTALLED=False
 
 $PYTHON -m pip install keras-tuner==1.1.2 kt-legacy -vv --no-dependencies
 # $PYTHON -m pip install kt-legacy -vv --no-dependencies
-# $PYTHON -m pip install tensorflow-macos==2.8.0 tensorflow-metal=0.5.0
-$PYTHON -m pip install tensorflow-macos==2.9 tensorflow-metal==0.5.0
+# $PYTHON -m pip install tensorflow-macos==2.9 tensorflow-metal==0.5.0
+$PYTHON -m pip install tensorflow-macos==2.13 tensorflow-metal==1.0.0
 $PYTHON -m pip install das -vv --no-dependencies
 

conda/das-osxarm/meta.yaml

@@ -18,24 +18,24 @@ requirements:
   host:
     - python {{ python }}
     - pip
-    - h5py
-    - numpy  # <1.24
+    # - h5py
+    # - numpy
   run:
     - python {{ python }}
     - rich
     - flammkuchen
     - matplotlib
     - pandas
-    - numpy  # <1.24
+    - numpy
     - peakutils
     - pyyaml
     - defopt=6.3
     - pyside2
     - scikit-learn
     - scikit-image
     - tqdm
-    - xarray-behave
-    - apple::tensorflow-deps=2.9
+    - xarray-behave>=0.34.0
+    - apple::tensorflow-deps=2.13
     - keras
     - keras-tuner
 

src/das/__init__.py

@@ -1,2 +1,2 @@
 """DAS"""
-__version__ = "0.31.0"
+__version__ = "0.31.1"

src/das/evaluate.py

@@ -76,8 +76,8 @@ def evaluate_segment_timing(segment_labels_true, segment_labels_pred, samplerate
 # TODO: move to das.segment_utils
 def segment_timing(labels, samplerate: float):
     """Get onset and offset time (in seconds) for each segment."""
-    segment_onset_times = np.where(np.diff(labels) == 1)[0].astype(np.float) / samplerate  # explicit cast required?
-    segment_offset_times = np.where(np.diff(labels) == -1)[0].astype(np.float) / samplerate
+    segment_onset_times = np.where(np.diff(labels) == 1)[0].astype(float) / samplerate  # explicit cast required?
+    segment_offset_times = np.where(np.diff(labels) == -1)[0].astype(float) / samplerate
     return segment_onset_times, segment_offset_times
 
 

src/das/event_utils.py

@@ -71,7 +71,7 @@ def match_events(eventindices_true, eventindices_pred, tol=100):
     nearest_event = np.zeros_like(eventindices_pred)
     # find nearest true event for each predicted event
     _, nearest_event, nearest_dist = find_nearest(eventindices_true, eventindices_pred)
-    nearest_event = nearest_event.astype(np.float)
+    nearest_event = nearest_event.astype(float)
 
     # flag those that have no nearby event
     nearest_event = np.ma.masked_array(nearest_event, mask=nearest_dist > tol)

src/das/make_dataset.py

@@ -51,7 +51,7 @@ def init_store(
     for target in ["train", "val", "test"]:
         root.empty(name=f"{target}/x", shape=(0, nb_channels), chunks=(chunk_len, nb_channels), dtype=np.float16)
         root.empty(name=f"{target}/y", shape=(0, nb_classes), chunks=(chunk_len, nb_classes), dtype=np.float16)
-        # root.empty(name=f'{target}/eventtimes', shape=(0, nb_classes), chunks=(1_000,), dtype=np.float)
+        # root.empty(name=f'{target}/eventtimes', shape=(0, nb_classes), chunks=(1_000,), dtype=float)
         if make_single_class_datasets:
             for class_name in class_names[1:]:
                 root.empty(name=f"{target}/y_{class_name}", shape=(0, 2), chunks=(chunk_len, nb_classes), dtype=np.float16)
@@ -209,7 +209,7 @@ def make_gaps(
     gap_halfwidth = int(np.floor(gap_seconds * samplerate) / 2)
 
     # widen gaps between adjacent syllables of different types
-    a = y.copy().astype(np.float)
+    a = y.copy().astype(float)
     label_change = np.where(np.diff(a, axis=0) != 0)[0]
     # remove on and offsets (0->label or label->0)
     onset = a[label_change] == 0

src/das/models.py

@@ -4,7 +4,7 @@
 import tensorflow.keras.layers as kl
 from tensorflow.keras import regularizers
 from tensorflow.keras.applications.resnet_v2 import ResNet50V2
-from typing import List, Optional
+from typing import List, Optional, Tuple
 from . import tcn as tcn_layer
 from .kapre.time_frequency import Spectrogram
 from . import spec_utils
@@ -329,11 +329,30 @@ def tcn_stft_morph(
 
     x = kl.Dense(nb_classes, activation="softmax")(x)
 
-    morph_kernel_duration = 16
+    # from .morpholayers.constraints import Disk
+    # import skimage.morphology as skm
+    # def ser(length):
+    #     return skm.rectangle(length, 1)
+    # skimage.morphology.remove_small_holes
+    # skimage.morphology.remove_small_objects
+    morph_kernel_duration = 33
     if morph_kernel_duration is not None:
         x = x[:, :, tf.newaxis, :]
-        x = Closing2D(1, padding="same", kernel_size=(morph_kernel_duration, 1), kernel_regularization=l1lattice(0.002))(x)
-        x = Opening2D(1, padding="same", kernel_size=(morph_kernel_duration, 1), kernel_regularization=l1lattice(0.002))(x)
+        x = Closing2D(
+            num_filters=1,
+            padding="same",
+            kernel_size=(morph_kernel_duration, 1),
+            kernel_regularization=l1lattice(0.002),
+            # kernel_constraint=ser,
+        )(x)
+        x = Opening2D(
+            num_filters=1,
+            padding="same",
+            kernel_size=(morph_kernel_duration, 1),
+            kernel_regularization=l1lattice(0.002),
+            # kernel_constraint=ser,
+        )(x)
+
         x = x[..., 0, :]
 
     if nb_pre_conv > 0 and upsample:

src/das/morpholayers/constraints.py

@@ -5,17 +5,21 @@
 import skimage.morphology as skm
 import scipy.ndimage.morphology as snm
 
-"""
-===============
-GLOBAL VARIABLE
-===============
-"""
 MIN_LATT = -1
 MAX_LATT = 0
 
 
 @tf.custom_gradient
 def rounding_op1(x):
+    """Round the input tensor to the nearest tenth.
+
+    Args:
+        x (tf.Tensor): Input tensor.
+
+    Returns:
+        tf.Tensor: Rounded tensor.
+    """
+
     def grad(dy):
         return dy
 
@@ -24,6 +28,15 @@ def grad(dy):
 
 @tf.custom_gradient
 def rounding_op2(x):
+    """Round the input tensor to the nearest hundredth.
+
+    Args:
+        x (tf.Tensor): Input tensor.
+
+    Returns:
+        tf.Tensor: Rounded tensor.
+    """
+
     def grad(dy):
         return dy
 
@@ -32,6 +45,15 @@ def grad(dy):
 
 @tf.custom_gradient
 def rounding_op3(x):
+    """Round the input tensor to the nearest thousandth.
+
+    Args:
+        x (tf.Tensor): Input tensor.
+
+    Returns:
+        tf.Tensor: Rounded tensor.
+    """
+
     def grad(dy):
         return dy
 
@@ -40,14 +62,28 @@ def grad(dy):
 
 @tf.custom_gradient
 def rounding_op4(x):
+    """Round the input tensor to the nearest ten-thousandth.
+
+    Args:
+        x (tf.Tensor): Input tensor.
+
+    Returns:
+        tf.Tensor: Rounded tensor.
+    """
+
     def grad(dy):
         return dy
 
     return tf.round(x * 10000) / 10000, grad
 
 
 class Rounding(Constraint):
-    # Using Constraint to Round values
+    """Constrains weights by rounding them to specified decimal places.
+
+    Args:
+        c (int): Number of decimal places for rounding.
+    """
+
     def __init__(self, c=4):
         self.c = c
 
@@ -67,9 +103,7 @@ def get_config(self):
 
 
 class NonPositive(Constraint):
-    """
-    Constraint to NonPositive Values
-    """
+    """Constrains weights to be non-positive values."""
 
     def __init__(self):
         self.min_value = MIN_LATT
@@ -83,9 +117,7 @@ def get_config(self):
 
 
 class NonPositiveExtensive(Constraint):
-    """
-    Constraint to NonPositive and Center equal to zero
-    """
+    """Constrains weights to be non-positive and centers equal to zero."""
 
     def __init__(self):
         self.min_value = MIN_LATT
@@ -95,7 +127,6 @@ def __call__(self, w):
         w = K.clip(w, self.min_value, 0)
         data = np.ones(w.shape)
         data[int(w.shape[0] / 2), int(w.shape[1] / 2), :, :] = 0
-        # data_tf = tf.convert_to_tensor(data, np.float32)
         w = tf.multiply(w, tf.convert_to_tensor(data, np.float32))
         return w
 
@@ -104,7 +135,8 @@ def get_config(self):
 
 
 class ZeroToOne(Constraint):
-    # Constraint between 0 to 1 Values
+    """Constrains weights to be between 0 and 1."""
+
     def __init__(self):
         self.min_value = 0.0
         self.max_value = 1.0
@@ -117,9 +149,7 @@ def get_config(self):
 
 
 class Lattice(Constraint):
-    """
-    Contraint to Value Lattice Value
-    """
+    """Constrains weights to be within a lattice range."""
 
     def __init__(self):
         self.min_value = MIN_LATT
@@ -134,9 +164,7 @@ def get_config(self):
 
 
 class SEconstraint(Constraint):
-    """
-    Constraint any SE Shape
-    """
+    """Constrains weights to any structured element (SE) shape."""
 
     def __init__(self, SE=skm.disk(1)):
         self.min_value = MIN_LATT
@@ -156,17 +184,13 @@ def get_config(self):
 
 
 class Disk(Constraint):
-    """
-    Constraint to Disk Shape
-    Only for square filters.
-    """
+    """Constrains weights to a disk shape (only for square filters)."""
 
     def __init__(self):
         self.min_value = MIN_LATT
         self.max_value = -MIN_LATT
 
     def __call__(self, w):
-        # print('DISK CONSTRAINT',w.shape)
         data = skm.disk(int(w.shape[0] / 2))
         data = np.repeat(data[:, :, np.newaxis], w.shape[2], axis=2)
         data = np.repeat(data[:, :, :, np.newaxis], w.shape[3], axis=3)
@@ -179,17 +203,13 @@ def get_config(self):
 
 
 class Diamond(Constraint):
-    """
-    Constraint to Diamond Shape
-    Only for square filters.
-    """
+    """Constrains weights to a diamond shape (only for square filters)."""
 
     def __init__(self):
         self.min_value = MIN_LATT
         self.max_value = -MIN_LATT
 
     def __call__(self, w):
-        # print('DIAMOND CONSTRAINT',w.shape)
         data = skm.diamond(int(w.shape[0] / 2))
         data = np.repeat(data[:, :, np.newaxis], w.shape[2], axis=2)
         data = np.repeat(data[:, :, :, np.newaxis], w.shape[3], axis=3)