forked from jdmanton/rlgc
-
Notifications
You must be signed in to change notification settings - Fork 0
/
rlgc.py
390 lines (298 loc) · 14.7 KB
/
rlgc.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
#!/usr/bin/env python
# Richardson-Lucy deconvolution code, using gradient consensus to stop iterations locally
# James Manton, 2023
# jmanton@mrc-lmb.cam.ac.uk
#
# Developed in collaboration with Andy York (Calico), Jan Becker (Oxford) and Craig Russell (EMBL EBI)
# Minor edits by Scott Brooks (Warwick) s.brooks.2@warwick.ac.uk
import numpy as np
import cupy as cp
import ast
import timeit
import tifffile
import argparse
from joblib import Parallel, delayed
import cProfile
import pstats
# python rlgc.py --input "/mnt/Raid_partition_1/internal_tmp/brooks/dockerdir/MantonData/2024-05-13_montage7Capture10crop448_t1.tif" --psf "/mnt/Raid_partition_1/internal_tmp/brooks/dockerdir/MantonData/448SamplePSF2_cropped.tif" --output "/mnt/Raid_partition_1/internal_tmp/brooks/dockerdir/MantonData/2024-05-13_montage7Capture10crop448_t1_rlgc_iter20.tif" --rl_output "/mnt/Raid_partition_1/internal_tmp/brooks/dockerdir/MantonData/debug_iterations20.tif" --max_iters 20
ntype = np.float16
ctype = cp.float16
rng = np.random.default_rng()
optimal_pool_size = 60 # This appears to be max windows can handle
# TODO: additional options add redundant code where
def main():
# Get input arguments
parser = argparse.ArgumentParser(formatter_class = argparse.ArgumentDefaultsHelpFormatter)
parser.add_argument('--input', type = str, required = True)
# should be list of strings
parser.add_argument('--psf', type = str, required = True)
parser.add_argument('--output', type = str, required = True)
parser.add_argument('--max_iters', type = int, default = 10)
parser.add_argument('--reblurred', type = str, required = False)
parser.add_argument('--process_psf', type = int, default = 1)
parser.add_argument('--rl_output', type = str, required = True) #originally False, return this to False later
parser.add_argument('--limit', type = float, default = 0.01)
parser.add_argument('--iters_output', type = str, required = False)
parser.add_argument('--rl_iters_output', type = str, required = False)
parser.add_argument('--updates_output', type = str, required = False)
parser.add_argument('--blur_consensus', type = int, default = 1)
args = parser.parse_args()
# # Log which files we're working with and the number of iterations
# print('')
# print('Input file: %s' % args.input)
# print('Input shape: %s' % (image.shape, ))
# print('PSF files: %s' % args.psf)
# print('PSF shape: %s' % (psf_temp.shape, ))
# print('Output file: %s' % args.output)
# print('Maximum number of iterations: %d' % args.max_iters)
# print('PSF processing: %s' % args.process_psf)
# print('')
# Print raw input arguments for debugging
print(f"Raw PSF input: {args.psf}")
# Split the comma-separated string into a list
psf_list = args.psf.split(',')
# Ensure it's a list
if not isinstance(psf_list, list):
raise ValueError("The argument for --psf should be a comma-separated list of strings.")
with tifffile.TiffFile(args.input) as tif:
full_image = tif.asarray()
# metadata = tif.shaped_metadata
imagej_metadata = tif.imagej_metadata if tif.is_imagej else {}
print(args.input)
image_shape = get_image_shape(full_image, psf_list)
# Load PSFs
psfs = load_psfs(args, psf_list, image_shape)
tifffile.imwrite("processedpsfs.tif", psfs.get(), bigtiff=True)
recon, recon_rl = triage(args, full_image, psfs)
with tifffile.TiffWriter(args.output, imagej=True) as tif_writer:
combined_metadata = {'axes': 'TZCYX'}
combined_metadata.update(imagej_metadata)
tif_writer.write(recon, metadata=combined_metadata)
print(recon.shape)
# Save RL output if argument given
# if args.rl_output is not None:
# tifffile.imwrite(args.rl_output, recon_rl, bigtiff=True, metadata={'axes': 'TZCYX'}, imagej_metadata=metadata)
# not implemeted for 5D
# # Save full iterations if argument given
# if (args.iters_output is not None):
# tifffile.imwrite(args.iters_output, iters[0:num_iters, :, :, :], bigtiff=True)
# not implemeted for 5D
# # Save full RL iterations if argument given
# if (args.rl_iters_output is not None):
# tifffile.imwrite(args.rl_iters_output, rl_iters[0:num_iters, :, :, :], bigtiff=True)
# not implemeted for 5D
# # Save full updates if argument given
# if (args.updates_output is not None):
# tifffile.imwrite(args.updates_output, updates, bigtiff=True)
def triage(args, full_image, psfs):
# if there is only 1 timepoint just run original code, else assume ztcyx
# TODO: come up with something better than enumerating possibilities + channel lazy 2 GPU processing
if full_image.ndim <= 3:
return rlgc(args, full_image, psfs[0], 0, 0)
# TODO: MULTIchannel we need to make the shape
elif full_image.ndim == 4:
if len(psfs) == 1:
return rlgc_4D(args, full_image, psfs[0], 0)
else:
return rlgc_4D_multichannel(args, full_image, psfs)
elif full_image.ndim == 5:
# TODO: Should this also have some alignment if psfs are very different?
return rlgc_5D(args, full_image, psfs)
def get_image_shape(full_image, psf_list):
# assume tzcyx or zcyx
shape = full_image.shape
if len(shape) == 4 and len(psf_list) == 2:
first_element = shape[0]
else:
first_element = shape[1]
last_two_elements = shape[-2:]
# Combine into a new tuple
return (first_element, *last_two_elements)
def load_psfs(args, psf_list, image_shape):
# this is a little inefficient, but avoiding lists, should we store the first read?
# psf_temp = tifffile.imread(psf_list[0])e
new_shape = (len(psf_list),) + image_shape
psfs = cp.zeros(new_shape, dtype=ctype)
for i, psf_path in enumerate(psf_list):
# Load and pad PSF if necessary
print(psf_path)
psf_temp = tifffile.imread(psf_path)
# Add new z-axis if we have 2D data
if psf_temp.ndim == 2:
psf_temp = np.expand_dims(psf_temp, axis=0)
if (args.process_psf):
print("Processing PSF...")
# Take upper left 16x16 pixels to estimate noise level and create appropriate fake noise
noisy_region = psf_temp[0:16, 0:16, 0:16]
background = np.random.normal(np.mean(noisy_region), np.std(noisy_region), image_shape)
psf = np.zeros(image_shape)
# Place PSF is the top left corner of an array with the same shape of the original image
psf[:psf_temp.shape[0], :psf_temp.shape[1], :psf_temp.shape[2]] = psf_temp
# This looks like it centers the PSF, but for larger PSFs in comparison to image size will not centre well, without this it doesnt work, what does it do?
for axis, axis_size in enumerate(psf_temp.shape):
psf = np.roll(psf, -int(axis_size / 2), axis=axis)
# Calculate the shift values to center psf_temp in psf
shifts = [-(psf_temp.shape[i] // 2) + (psf.shape[i] // 2) for i in range(len(psf_temp.shape))]
# # Roll psf to center psf_temp
# for axis, shift in enumerate(shifts):
# psf = np.roll(psf, shift, axis=axis)
# remove any dirt on the coverslip then replace with background
# if (args.process_psf):
# psf = np.where(psf < np.mean(noisy_region), background, psf)
# psf = psf_temp
psf = psf / np.sum(psf)
# Load data PSF onto GPU
psfs[i] = cp.array(psf, dtype=ctype)
return psfs
# NOT PARAELLEL
# def rlgc_4D(args, full_image, psf, chstr):
# recon_4D = np.zeros(full_image.shape, dtype=ntype)
# recon_rl_4D = np.zeros(full_image.shape, dtype=ntype)
# for timepoint in range(full_image.shape[0]):
# recon, recon_rl = rlgc(args, full_image[timepoint], psf, timepoint, chstr)
# recon_4D[timepoint]= recon
# recon_rl_4D[timepoint]= recon_rl
# return recon_4D, recon_rl_4D
# def rlgc_4D_multichannel(args, full_image, psfs):
# recon_4D = np.zeros(full_image.shape, dtype=ntype)
# recon_rl_4D = np.zeros(full_image.shape, dtype=ntype)
# for channel in range(len(psfs)):
# recon, recon_rl = rlgc(args, full_image[:,channel], psfs[channel], 0, channel)
# recon_4D[:,channel]= recon
# recon_rl_4D[:,channel]= recon_rl
# return recon_4D, recon_rl_4D
# def rlgc_5D(args, full_image,psfs):
# recon_5D = np.zeros(full_image.shape, dtype=ntype)
# recon_rl_5D = np.zeros(full_image.shape, dtype=ntype)
# print(full_image.shape)
# for channel in range(len(psfs)):
# recon, recon_rl = rlgc_4D(args, full_image[:,:,channel], psfs[channel], channel)
# recon_5D[:,:,channel]= recon
# recon_rl_5D[:,:,channel]= recon_rl
# return recon_5D, recon_rl_5D
def rlgc_4D(args, full_image, psf, chstr):
results = Parallel(n_jobs=optimal_pool_size)(
delayed(rlgc)(args, full_image[timepoint], psf, timepoint, chstr) for timepoint in range(full_image.shape[0])
)
recon_4D, recon_rl_4D = zip(*results)
return np.array(recon_4D), np.array(recon_rl_4D)
def rlgc_4D_multichannel(args, full_image, psfs):
results = Parallel(n_jobs=optimal_pool_size)(
delayed(rlgc)(args, full_image[:, channel], psfs[channel], 0, channel) for channel in range(len(psfs))
)
recon_4D, recon_rl_4D = zip(*results)
return np.array(recon_4D), np.array(recon_rl_4D)
def rlgc_5D(args, full_image, psfs):
# We will handle parallelism at this level only
results = Parallel(n_jobs=optimal_pool_size)(
delayed(rlgc_5D_single_channel)(args, full_image, psfs, channel) for channel in range(len(psfs))
)
recon_5D, recon_rl_5D = zip(*results)
return np.array(recon_5D), np.array(recon_rl_5D)
def rlgc_5D_single_channel(args, full_image, psfs, channel):
recon_channel = []
recon_rl_channel = []
for timepoint in range(full_image.shape[0]):
recon, recon_rl = rlgc(args, full_image[timepoint, :, channel], psfs[channel], timepoint, channel)
recon_channel.append(recon)
recon_rl_channel.append(recon_rl)
return np.array(recon_channel), np.array(recon_rl_channel)
def rlgc(args, image, psf, tp, ch):
print(f"Processing timepoint {str(tp)}, channel {str(ch)}")
# Load data and PSF onto GPU
# TODO: Specify GPU
image = cp.array(image, dtype=ctype)
# Calculate OTF and transpose
# TODO put this in load psf
otf = cp.fft.rfftn(psf)
psfT = cp.flip(psf, (0, 1, 2))
otfT = cp.fft.rfftn(psfT)
# Get dimensions of data
num_z = image.shape[0]
num_y = image.shape[1]
num_x = image.shape[2]
num_pixels = num_z * num_y * num_x
# Calculate Richardson-Lucy iterations
HTones = fftconv(cp.ones_like(image), otfT)
recon = cp.ones((num_z, num_y, num_x))
recon_rl = cp.ones((num_z, num_y, num_x))
if (args.iters_output is not None):
iters = np.zeros((args.max_iters, num_z, num_y, num_x))
if (args.rl_iters_output is not None):
rl_iters = np.zeros((args.max_iters, num_z, num_y, num_x))
if (args.updates_output is not None):
updates = np.zeros((args.max_iters, num_z, num_y, num_x))
num_iters = 0
for iter in range(args.max_iters):
start_time = timeit.default_timer()
print(image.shape)
# Split recorded image into 50:50 images
# TODO: make this work on the GPU (for some reason, we get repeating blocks with a naive conversion to cupy)
split1 = rng.binomial(image.get().astype('int64'), p=0.5)
split1 = cp.array(split1)
split2 = image - split1
# Calculate prediction
Hu = fftconv(recon, otf)
# Calculate updates for split images and full images (H^T (d / Hu))
ratio1 = split1 / (0.5 * (Hu + 1E-12))
ratio2 = split2 / (0.5 * (Hu + 1E-12))
HTratio1 = fftconv(ratio1, otfT)
HTratio2 = fftconv(ratio2, otfT)
ratio = image / (Hu + 1E-12)
HTratio = fftconv(ratio, otfT)
HTratio = HTratio / HTones
# Normalise update steps by H^T(1) and only update pixels in full estimate where split updates agree in 'sign'
update1 = HTratio1 / HTones
update2 = HTratio2 / HTones
if (args.blur_consensus != 0):
shouldNotUpdate = fftconv(fftconv((update1 - 1) * (update2 - 1), otf), otfT) < 0
else:
shouldNotUpdate = (update1 - 1) * (update2 - 1) < 0
HTratio[shouldNotUpdate] = 1
# Save previous estimate to check we're not wasting our time updating small values
previous_recon = recon
# Update estimate
recon = recon * HTratio
# Add to full iterations output if asked to by user
if (args.iters_output is not None):
iters[iter, :, :, :] = recon.get()
if (args.updates_output is not None):
updates[iter, :, :, :] = HTratio.get()
# Also calculate normal RL update if asked to by user
if args.rl_output is not None:
Hu_rl = fftconv(recon_rl, otf)
ratio_rl = image / (Hu_rl + 1E-12)
HTratio_rl = fftconv(ratio_rl, otfT)
recon_rl = recon_rl * HTratio_rl / HTones
if (args.rl_iters_output is not None):
rl_iters[iter, :, :, :] = recon_rl.get()
calc_time = timeit.default_timer() - start_time
num_updated = num_pixels - cp.sum(shouldNotUpdate)
max_relative_delta = cp.max((recon - previous_recon) / cp.max(recon))
print(f"Timepoint {str(tp)}, channel {str(ch)}" + "Iteration %03d completed in %1.3f s. %1.2f %% of image updated. Update range: %1.2f to %1.2f. Largest relative delta = %1.3f" % (iter + 1, calc_time, 100 * num_updated / num_pixels, cp.min(HTratio), cp.max(HTratio), max_relative_delta))
num_iters = num_iters + 1
if (num_updated / num_pixels < args.limit):
break
if (max_relative_delta < 0.01):
break
# not implemented for 5D
# Reblur, collect from GPU and save if argument given
# if args.reblurred is not None:
# reblurred = fftconv(recon, otf)
# reblurred = reblurred.get()
# tifffile.imwrite(args.reblurred, reblurred, bigtiff=True)
# Collect reconstruction from GPU and save, for now leave on GPU
# recon = recon.get()
if args.rl_output is not None:
# recon_rl = recon_rl.get()
return recon.get(), recon_rl.get()
return recon.get()
def fftconv(x, H):
return cp.fft.irfftn(cp.fft.rfftn(x) * H, x.shape)
if __name__ == '__main__':
profiler = cProfile.Profile()
profiler.enable()
main()
profiler.disable()
stats = pstats.Stats(profiler).sort_stats('cumtime')
stats.print_stats(10)