Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Remove skcuda from motion_correction. Implements #1416 #1433

Merged
merged 9 commits into from
Dec 11, 2024
Merged

Remove skcuda from motion_correction. Implements #1416 #1433

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
9 commits
Select commit Hold shift + click to select a range
8254fe2
Remove skcuda from motion_correction. Implements #1416
pgunn Dec 6, 2024
4d94b6e
use_cuda deprecation - revise phrasing as per kkolar
pgunn Dec 10, 2024
d31d8c8
Fix docs for arguments to tile_and_correct() and tile_and_correct_3d()
pgunn Dec 10, 2024
8898469
motion_correction: eliminate the global boolean "opencv".
pgunn Dec 10, 2024
95b8d27
motion_correction: Fix confusing errors from some functions
pgunn Dec 10, 2024
6aa2309
motion_correction: Fix more mistakes in function documentation, other…
pgunn Dec 10, 2024
54e1952
motion_correction: Remove some dead code, more docs cleanups
pgunn Dec 10, 2024
dbb0018
Update caiman/motion_correction.py
pgunn Dec 10, 2024
5a8a4dc
Update caiman/motion_correction.py
pgunn Dec 10, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
122 changes: 18 additions & 104 deletions caiman/motion_correction.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -69,15 +69,7 @@
except:
def profile(a): return a

opencv = True

try:
import pycuda.gpuarray as gpuarray
import pycuda.driver as cudadrv
import atexit
HAS_CUDA = True
except ImportError:
HAS_CUDA = False
opencv = True # FIXME How is the user meant to be able to interact with this?!
pgunn marked this conversation as resolved.
Show resolved Hide resolved

class MotionCorrect(object):
"""
Expand Down Expand Up @@ -143,8 +135,8 @@ def __init__(self, fname, min_mov=None, dview=None, max_shifts=(6, 6), niter_rig
nonneg_movie: boolean
make the SAVED movie and template mostly nonnegative by removing min_mov from movie

use_cuda : bool, optional
Use skcuda.fft (if available). Default: False
use_cuda : boolean (DEPRECATED)
No longer used, may be removed in a future version
pgunn marked this conversation as resolved.
Show resolved Hide resolved

border_nan : bool or string, optional
Specifies how to deal with borders. (True, False, 'copy', 'min')
Expand Down Expand Up @@ -204,8 +196,8 @@ def __init__(self, fname, min_mov=None, dview=None, max_shifts=(6, 6), niter_rig
self.var_name_hdf5 = var_name_hdf5
self.is3D = bool(is3D)
self.indices = indices
if self.use_cuda and not HAS_CUDA:
logger.debug("pycuda is unavailable. Falling back to default FFT.")
if self.use_cuda:
logger.warn("skcuda is no longer supported; using non-cuda algorithms")
pgunn marked this conversation as resolved.
Show resolved Hide resolved

def motion_correct(self, template=None, save_movie=False):
"""general function for performing all types of motion correction. The
Expand Down Expand Up @@ -1346,32 +1338,6 @@ def _compute_error(cross_correlation_max, src_amp, target_amp):
(src_amp * target_amp)
return np.sqrt(np.abs(error))

def init_cuda_process():
"""
Initialize a PyCUDA context at global scope so that it can be accessed
from processes when using multithreading
"""
global cudactx

cudadrv.init()
dev = cudadrv.Device(0)
cudactx = dev.make_context() # type: ignore
atexit.register(cudactx.pop) # type: ignore


def close_cuda_process(n):
"""
Cleanup cuda process
"""

global cudactx

import skcuda.misc as cudamisc
try:
cudamisc.done_context(cudactx) # type: ignore
except:
pass

def register_translation_3d(src_image, target_image, upsample_factor = 1,
space = "real", shifts_lb = None, shifts_ub = None,
max_shifts = [10,10,10]):
Expand Down Expand Up @@ -1586,8 +1552,8 @@ def register_translation(src_image, target_image, upsample_factor=1,
will be FFT'd to compute the correlation, while "fourier" data will
bypass FFT of input data. Case insensitive.

use_cuda : bool, optional
Use skcuda.fft (if available). Default: False
use_cuda : boolean (DEPRECATED)
No longer used, may be removed in a future version

Returns:
shifts : ndarray
Expand Down Expand Up @@ -1627,50 +1593,13 @@ def register_translation(src_image, target_image, upsample_factor=1,
raise NotImplementedError("Error: register_translation only supports "
"subpixel registration for 2D images")

if HAS_CUDA and use_cuda:
from skcuda.fft import Plan
from skcuda.fft import fft as cudafft
from skcuda.fft import ifft as cudaifft
try:
cudactx # type: ignore
except NameError:
init_cuda_process()

# assume complex data is already in Fourier space
if space.lower() == 'fourier':
src_freq = src_image
target_freq = target_image
# real data needs to be fft'd.
elif space.lower() == 'real':
if HAS_CUDA and use_cuda:
# src_image_cpx = np.array(src_image, dtype=np.complex128, copy=False)
# target_image_cpx = np.array(target_image, dtype=np.complex128, copy=False)

image_gpu = gpuarray.to_gpu(np.stack((src_image, target_image)).astype(np.complex128))
freq_gpu = gpuarray.empty((2, src_image.shape[0], src_image.shape[1]), dtype=np.complex128)
# src_image_gpu = gpuarray.to_gpu(src_image_cpx)
# src_freq_gpu = gpuarray.empty(src_image_cpx.shape, np.complex128)

# target_image_gpu = gpuarray.to_gpu(target_image_cpx)
# target_freq_gpu = gpuarray.empty(target_image_cpx.shape, np.complex128)

plan = Plan(src_image.shape, np.complex128, np.complex128, batch=2)
# cudafft(src_image_gpu, src_freq_gpu, plan, scale=True)
# cudafft(target_image_gpu, target_freq_gpu, plan, scale=True)
cudafft(image_gpu, freq_gpu, plan, scale=True)
# src_freq = src_freq_gpu.get()
# target_freq = target_freq_gpu.get()
freq = freq_gpu.get()
src_freq = freq[0, :, :]
target_freq = freq[1, :, :]

# del(src_image_gpu)
# del(src_freq_gpu)
# del(target_image_gpu)
# del(target_freq_gpu)
del(image_gpu)
del(freq_gpu)
elif opencv:
if opencv:
src_freq_1 = cv2.dft(
src_image, flags=cv2.DFT_COMPLEX_OUTPUT + cv2.DFT_SCALE)
src_freq = src_freq_1[:, :, 0] + 1j * src_freq_1[:, :, 1]
Expand All @@ -1695,14 +1624,7 @@ def register_translation(src_image, target_image, upsample_factor=1,
# Whole-pixel shift - Compute cross-correlation by an IFFT
shape = src_freq.shape
image_product = src_freq * target_freq.conj()
if HAS_CUDA and use_cuda:
image_product_gpu = gpuarray.to_gpu(image_product)
cross_correlation_gpu = gpuarray.empty(
image_product.shape, np.complex128)
iplan = Plan(image_product.shape, np.complex128, np.complex128)
cudaifft(image_product_gpu, cross_correlation_gpu, iplan, scale=True)
cross_correlation = cross_correlation_gpu.get()
elif opencv:
if opencv:

image_product_cv = np.dstack(
[np.real(image_product), np.imag(image_product)])
Expand Down Expand Up @@ -2112,8 +2034,8 @@ def tile_and_correct(img, template, strides, overlaps, max_shifts, newoverlaps=N
filt_sig_size: tuple
standard deviation and size of gaussian filter to center filter data in case of one photon imaging data

use_cuda : bool, optional
Use skcuda.fft (if available). Default: False
use_cuda : boolean (DEPRECATED)
No longer used, may be removed in a future version

border_nan : bool or string, optional
specifies how to deal with borders. (True, False, 'copy', 'min')
Expand Down Expand Up @@ -2360,8 +2282,8 @@ def tile_and_correct_3d(img:np.ndarray, template:np.ndarray, strides:tuple, over
filt_sig_size: tuple
standard deviation and size of gaussian filter to center filter data in case of one photon imaging data

use_cuda : bool, optional
Use skcuda.fft (if available). Default: False
use_cuda : boolean (DEPRECATED)
No longer used, may be removed in a future version

border_nan : bool or string, optional
specifies how to deal with borders. (True, False, 'copy', 'min')
Expand Down Expand Up @@ -2389,11 +2311,6 @@ def tile_and_correct_3d(img:np.ndarray, template:np.ndarray, strides:tuple, over
img, template, upsample_factor=upsample_factor_fft, max_shifts=max_shifts)

if max_deviation_rigid == 0: # if rigid shifts only

# if shifts_opencv:
# NOTE: opencv does not support 3D operations - skimage is used instead
# else:

if gSig_filt is not None:
raise Exception(
'The use of FFT and filtering options have not been tested. Set opencv=True')
Expand Down Expand Up @@ -2760,8 +2677,8 @@ def motion_correct_batch_rigid(fname, max_shifts, dview=None, splits=56, num_spl
subidx: slice
Indices to slice

use_cuda : bool, optional
Use skcuda.fft (if available). Default: False
use_cuda : boolean (DEPRECATED)
No longer used, may be removed in a future version

indices: tuple(slice), default: (slice(None), slice(None))
Use that to apply motion correction only on a part of the FOV
Expand Down Expand Up @@ -2914,8 +2831,8 @@ def motion_correct_batch_pwrigid(fname, max_shifts, strides, overlaps, add_to_mo
save_movie_rigid: boolean
toggle save movie

use_cuda : bool, optional
Use skcuda.fft (if available). Default: False
use_cuda : boolean (DEPRECATED)
No longer used, may be removed in a future version

indices: tuple(slice), default: (slice(None), slice(None))
Use that to apply motion correction only on a part of the FOV
Expand Down Expand Up @@ -3148,10 +3065,7 @@ def motion_correction_piecewise(fname, splits, strides, overlaps, add_to_movie=0

if dview is not None:
logger.info('** Starting parallel motion correction **')
if HAS_CUDA and use_cuda:
res = dview.map(tile_and_correct_wrapper,pars)
dview.map(close_cuda_process, range(len(pars)))
elif 'multiprocessing' in str(type(dview)):
if 'multiprocessing' in str(type(dview)):
logger.debug("entering multiprocessing tile_and_correct_wrapper")
res = dview.map_async(tile_and_correct_wrapper, pars).get(4294967)
else:
Expand Down
2 changes: 0 additions & 2 deletions docs/source/CaImAn_features_and_references.rst
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -12,8 +12,6 @@ calcium (and voltage) imaging data:
- Can be run also in online mode (i.e. one frame at a time)
- Corrects for non-rigid artifacts due to raster scanning or
non-uniform brain motion
- FFTs can be computed on GPUs (experimental). Requires pycuda and
skcuda to be installed.
|
- **Source extraction**

Expand Down
Loading