Lazy-load heavy deps to halve import time (~1.5s → ~0.7s)

alexarje · claude · alexarje · commit 41432b6fe07e · 2026-06-27T22:34:46.000+02:00
Move scipy.signal (medfilt2d / signal) and scipy.stats.entropy out of module-top
imports in _filter, _impacts, _ssm and _flow, and defer IPython.display in _show,
to function-local imports. These were pulled in eagerly during
`import musicalgestures` (scipy.signal ~1s, IPython ~0.2s) but are only needed when
the relevant methods run. librosa is already lazily loaded upstream.

Co-Authored-By: Claude Opus 4.8 (1M context) &lt;noreply@anthropic.com&gt;
diff --git a/musicalgestures/_filter.py b/musicalgestures/_filter.py
@@ -1,4 +1,3 @@
-from scipy.signal import medfilt2d
 import cv2
 import numpy as np
 import matplotlib
@@ -19,6 +18,7 @@ def filter_frame(motion_frame, filtertype, threshold, kernel_size):
         np.array(uint8): The filtered frame.
     """
 
+    from scipy.signal import medfilt2d   # lazy import: keeps scipy.signal out of startup
     if filtertype.lower() == 'regular':
         motion_frame = (motion_frame > threshold*255)*motion_frame
         motion_frame = medfilt2d(motion_frame, kernel_size)
diff --git a/musicalgestures/_flow.py b/musicalgestures/_flow.py
@@ -4,7 +4,6 @@
 import math
 import weakref
 import matplotlib.pyplot as plt
-from scipy.stats import entropy
 
 import musicalgestures
 from musicalgestures._utils import MgFigure, extract_wav, embed_audio_in_video, MgProgressbar, convert_to_avi, generate_outfilename, ffmpeg_cmd, get_cuda_device_count, cuda_unavailable_reason
@@ -231,6 +230,7 @@ def dense(
 
             time = np.linspace(0, len(xvel)/fps, len(xvel))
             acceleration = self.get_acceleration(xvel, fps)
+            from scipy.stats import entropy   # lazy import: keeps scipy.stats out of startup
             acceleration_entropy = entropy(acceleration)
 
             ax.plot(time, xvel, label=f'Average acceleration: {round(np.mean(acceleration),3)} m/s\nEntropy of acceleration: {round(acceleration_entropy,3)}')
diff --git a/musicalgestures/_impacts.py b/musicalgestures/_impacts.py
@@ -2,7 +2,6 @@
 import os
 import numpy as np
 from numba import jit
-from scipy.signal import medfilt2d
 import matplotlib.pyplot as plt
 import musicalgestures
 from musicalgestures._directograms import directogram
@@ -11,6 +10,7 @@
 
 def impact_envelope(directogram, kernel_size=5):
 
+    from scipy.signal import medfilt2d   # lazy import: keeps scipy.signal out of startup
     # Apply a median filter to the directogram using a local window-size given by kernel_size
     filtered_directogram = medfilt2d(directogram, kernel_size)
     flux = np.zeros(directogram.shape)
diff --git a/musicalgestures/_show.py b/musicalgestures/_show.py
@@ -1,11 +1,6 @@
 # import cv2
 # import numpy as np
 import os
-from IPython.display import Image, display, HTML
-# try:
-from IPython.display import Video
-# except:
-#     from IPython.core.display import Video
 from base64 import b64encode
 import musicalgestures
 
@@ -24,6 +19,8 @@ def mg_show(self, filename=None, key=None, mode='windowed', window_width=640, wi
         window_title (str, optional): The title of the window. If None, the title of the window will be the file name. Defaults to None.
         ipython_kwargs (dict, optional): Additional arguments for IPython.display.Image or IPython.display.Video. Defaults to None.
     """
+    # Lazy import: keeps IPython out of `import musicalgestures` startup.
+    from IPython.display import Image, display, HTML, Video
 
     def show(file, width=640, height=480, mode='windowed', title='Untitled', parent=None, **ipython_kwargs):
         """
diff --git a/musicalgestures/_ssm.py b/musicalgestures/_ssm.py
@@ -1,7 +1,6 @@
 import os
 import cv2
 import numpy as np
-from scipy import signal
 import matplotlib as mpl
 import matplotlib.pyplot as plt
 import matplotlib.gridspec as gridspec
@@ -38,6 +37,7 @@ def inside(x, pos):
         
     formatter = FuncFormatter(inside)
 
+    from scipy import signal   # lazy import: keeps scipy.signal out of startup
     filt_kernel = np.expand_dims(signal.get_window(w_type, filt_len), axis=0)
     X_smooth = signal.convolve(X, filt_kernel, mode='same') / filt_len
     X_smooth = X_smooth[:, ::down_sampling]
