-
Notifications
You must be signed in to change notification settings - Fork 0
/
anz.py
354 lines (309 loc) · 12.2 KB
/
anz.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
import itertools
from typing import List
import numpy as np
import pandas as pd
from matplotlib import pyplot as plt
import base
import tools
from scipy.signal import savgol_filter
import scipy.interpolate as interp
def _get_markers(marker_dir, bodyparts, cams):
'''
Returns data in the form of [trial][camera][body parts, frames, xys]
:param marker_dir:
:param bodyparts:
:param cams:
:return:
'''
marker_files = [tools.get_files(marker_dir, (x + '.h5')) for x in cams]
marker_files = [x for x in zip(*marker_files)]
marker_xys_per_video = [] #[video][camera][data]
for marker_file_per_camera in marker_files:
marker_xys = []
for fn in marker_file_per_camera:
data = tools.load_pose_to_np(fn, bodyparts)
for bpix in range(data.shape[0]):
x, y, score = _repair_bodyparts(
x=data[bpix, :, 0],
y=data[bpix, :, 1],
score=data[bpix, :, 2])
data[bpix, :, 0] = x
data[bpix, :, 1] = y
data[bpix, :, 2] = score
marker_xys.append(data)
marker_xys_per_video.append(marker_xys)
return marker_xys_per_video
def _get_labels(label_dir):
'''
Returns data in the form of a list of dictionaries with
key: bp
value: list of regions of form [start, end].
Also returns label names.
:param label_dir:
:return:
'''
list_of_labels = tools.get_files(label_dir, ('labels.csv'))
label_names = pd.read_csv(list_of_labels[0], index_col=0).columns.to_numpy()
# label preprocessing
label_regions_per_video = []
for fn in list_of_labels:
df = pd.read_csv(fn, index_col=0)
df = df.to_numpy().T
label_regions = {}
for label_vec, label_name in zip(df, label_names):
failed_threshold = label_vec < 1
_, repair_ixs = _chunk(
failed_threshold,
max_gap=base.LABEL_JOIN_FRAME_GAP)
label_vec[repair_ixs] = 1
chunks, _ = _chunk(
label_vec > 0,
min_gap=base.LABEL_CRITERIA_CONTIGUOUS_FRAMES)
label_regions[label_name] = chunks
label_regions_per_video.append(label_regions)
return label_regions_per_video, label_names
def _chunk(bool_array, min_gap=0, max_gap=0):
idx = np.diff(np.r_[0, bool_array, 0]).nonzero()[0]
regions = np.reshape(idx, (-1, 2))
if min_gap > 0 and max_gap > 0:
regions = np.array([r for r in regions if (
(r[1] - r[0]) <= max_gap and (r[1] - r[0] >= min_gap))])
elif min_gap > 0 and max_gap == 0:
regions = np.array([r for r in regions if (r[1] - r[0]) >= min_gap])
elif min_gap == 0 and max_gap > 0:
regions = np.array([r for r in regions if (r[1] - r[0]) <= max_gap])
else:
raise ValueError('both min and max gaps are 0')
region_ixs = np.array(
list(itertools.chain(*[range(r[0], r[1]) for r in regions]))).astype(
int)
return regions, region_ixs
def _repair_bodyparts(x, y, score):
failed_threshold = score < base.SCORE_THRESHOLD
regions, interp_ixs = _chunk(failed_threshold, max_gap=base.REPAIR_FRAME_GAP)
good_ix = np.arange(len(x))
good_ix = np.delete(good_ix, interp_ixs)
good_x = np.delete(x, interp_ixs)
new_x = np.interp(interp_ixs, good_ix, good_x)
x[interp_ixs] = new_x
good_ix = np.arange(len(y))
good_ix = np.delete(good_ix, interp_ixs)
good_y = np.delete(y, interp_ixs)
new_y = np.interp(interp_ixs, good_ix, good_y)
y[interp_ixs] = new_y
score[interp_ixs] = base.SCORE_INTERP
return x, y, score
def _chunk_bodyparts(scores,
score_threshold,
min_frames):
'''
:param scores: shape [bps, frames]
:param score_threshold:
:param min_frames:
:return:
'''
bool_arr = scores > score_threshold
bool_vec = np.all(bool_arr, axis=0)
regions, region_ixs = _chunk(bool_vec, min_gap=min_frames)
return regions, region_ixs
def get_pellet_location(loc_arr, score_arr, plot):
'''
1. Assumes that xs and xss are of shape [TRIALS][FRAMES]
2. Assumes that the wheel rotation speed did not change within a session
:param loc_arr:
:param score_arr:
:return:
'''
import scipy.ndimage
num_start_frames = 150
burn_in_frames = 15
stationary_win = 10
score_threshold = base.SCORE_THRESHOLD
# find the best first stationary index
loc_arr = np.array([x[:num_start_frames] for x in loc_arr])
score_arr = np.array([x[:num_start_frames] for x in score_arr])
xmas = np.ma.masked_where(score_arr < score_threshold, loc_arr)
dxs = np.diff(xmas, axis=1)
mdxs = np.mean(dxs, axis=0)
mdxs = savgol_filter(mdxs, window_length=7, polyorder=0)
residual = np.abs(mdxs[burn_in_frames:] - 0)
ix = np.argwhere(residual < 0.1)[0][0] + burn_in_frames
# find all coordinates in which the pellet at the stationary index is
# visible and has small velocity
mask = np.logical_and(score_arr[:, ix:ix + stationary_win] > score_threshold,
np.abs(dxs[:, ix:ix + stationary_win]) < 0.4)
coords = np.ma.masked_where(np.invert(mask), loc_arr[:, ix:ix + stationary_win])
mcoords = np.min(coords, axis=1)
# interpolate and filter these coordinates across trials
mcoords_interp = np.copy(mcoords)
mask = np.ma.getmask(mcoords)
interp_ix = np.where(mask)[0]
known_ix = np.where(np.invert(mask))[0]
interp_vals = np.interp(interp_ix, known_ix, mcoords[known_ix])
mcoords_interp[interp_ix] = interp_vals
mcoords_interp = scipy.ndimage.median_filter(mcoords_interp, size=5)
if plot:
plt.figure()
plt.plot(mdxs)
plt.title('Velocity, ix at {}'.format(ix))
plt.figure()
plt.plot(mcoords)
plt.title('Unrefined coordinates')
plt.figure()
plt.plot(mcoords_interp)
plt.title('Interpolated + Median filtered coordinates')
plt.show()
return mcoords_interp
def anneal_labels(extends, grabs, window):
'''
Anneals labels together within a window gap. Assumes that a comes before b.
:param a: list of tuples denoting start and end of behavioral epoch
:param b: list of tuples denoting start and end of behavioral epoch
:return:
'''
before = []
after = []
both = []
for extend in extends:
for grab in grabs:
if np.abs(grab[0] - extend[1]) < window:
before.append(grab)
after.append(extend)
both.append([extend[0], grab[1]])
return before, after, both
def interpolate_polyline(polyline, num_points, s=10):
duplicates = []
for i in range(1, len(polyline)):
if np.allclose(polyline[i], polyline[i-1], 1e-8, 1e-8):
duplicates.append(i)
if duplicates:
polyline = np.delete(polyline, duplicates, axis=0)
tck, u = interp.splprep(polyline.T, s=s)
u = np.linspace(0.0, 1.0, num_points)
return interp.splev(u, tck)
def contiguous_interp(ixs, data):
if len(ixs) != ixs[-1] - ixs[0] + 1:
all_indices = np.arange(ixs[0], ixs[-1]+1)
interp_ixs = np.setdiff1d(all_indices, ixs)
interp_val = np.interp(interp_ixs, ixs, data[ixs])
data[interp_ixs] = interp_val
return data
def _marker_label_intersects(marker_scores,
grab_region,
score_threshold,
min_contiguous_frames):
_, loc_ixs = _chunk_bodyparts(marker_scores,
score_threshold=score_threshold,
min_frames=min_contiguous_frames)
grab_ixs = np.arange(grab_region[0], grab_region[1])
ixs = np.intersect1d(grab_ixs, loc_ixs)
return ixs
def _get_grab_locations_atomic(marker_positions,
marker_scores,
grab_region):
'''
Find grab endpoints with viable paw locations. Assumes that the first
3 frames of grab_region contains the grab endpoint.
1. Assumes marker_positions has shape [body parts, frames]
2. Assumes marker_scores has shape [body parts, frames]
3. Assumes grab_region is of shape [start_ix, end_ix]
:param marker_positions:
:param label_regions:
:return:
'''
score_threshold = base.SCORE_THRESHOLD
min_contiguous_frames = base.CRITERIA_CONTIGUOUS_FRAMES
window = 5
ixs = _marker_label_intersects(marker_scores,
grab_region,
score_threshold,
min_contiguous_frames)
mean_marker_position = np.mean(marker_positions, axis=0)
pos = np.mean(mean_marker_position[ixs][:window])
return pos
def outcome_diagnostics(annotated_outcomes: List[base.OUTCOME],
computed_outcomes: List[base.OUTCOME],
mask):
'''
Compare ground_truth labels with computed outcomes. Print mistakes if
there are discrepancies that are not masked already
:param mask: bool array of same length as each outcome
'''
assert len(annotated_outcomes) == len(computed_outcomes)
assert len(annotated_outcomes) == len(mask)
for i, (x, y) in enumerate(zip(annotated_outcomes, computed_outcomes)):
if x!=y and mask[i]:
print(i, x, y, 'MASKED' if not mask[i] else '')
def outcome_truth_table(dropped_regions, chew_regions, grabbed_regions):
'''
All inputs are list of regions. Can be empty.
:param dropped_regions:
:param chew_regions:
:param grabbed_regions:
:return:
'''
any_drop = len(dropped_regions) > 0
any_chew = len(chew_regions) > 0
any_grab = len(grabbed_regions) > 0
if not any_drop and any_chew:
outcome = base.OUTCOME.SUCCESS
elif any_drop and not any_chew:
outcome = base.OUTCOME.FAIL
elif any_drop and any_chew:
if chew_regions[0][0] <= dropped_regions[0][0]:
outcome = base.OUTCOME.DROP_AFTER_GRAB
else:
outcome = base.OUTCOME.DROP_FP
elif len(dropped_regions) == 0 and len(chew_regions) == 0:
if any_grab:
outcome = base.OUTCOME.FAIL
else:
outcome = base.OUTCOME.NO_ATTEMPTS
else:
raise ValueError('outcome is not recognized')
return outcome
def grab_truth_table(outcome, grabbed_regions, chew_regions, dropped_regions):
if len(dropped_regions) > 1:
print('more than 1 dropped region')
done = False
pregrab = base.GRABTYPES.FAIL_WITH_PELLET
grab_outcomes = []
if outcome == base.OUTCOME.FAIL and len(dropped_regions) == 0:
grab_outcomes = [pregrab for x in grabbed_regions]
done = True
elif outcome == base.OUTCOME.FAIL and len(dropped_regions) > 0:
during = base.GRABTYPES.DROPPED
post = base.GRABTYPES.FAIL_POST_DROP
key_region = dropped_regions[0]
elif outcome == base.OUTCOME.DROP_AFTER_GRAB:
during = base.GRABTYPES.DROPPED
post = base.GRABTYPES.FAIL_POST_DROP
key_region = chew_regions[0]
elif outcome == base.OUTCOME.SUCCESS or outcome == base.OUTCOME.DROP_FP:
during = base.GRABTYPES.SNATCHED
post = base.GRABTYPES.FAIL_POST_SNATCH
key_region = chew_regions[0]
elif outcome == base.OUTCOME.NO_ATTEMPTS:
grab_outcomes = []
done = True
else:
raise ValueError(f'outcome: {outcome} not recognized')
if not done:
n_grabs = len(grabbed_regions)
for g in range(n_grabs):
current_grab = grabbed_regions[g]
if current_grab[0] <= key_region[0]:
if g + 1 < n_grabs:
next_grab = grabbed_regions[g+1]
else:
next_grab = None
if next_grab is None or next_grab[0] >= key_region[0]:
grab_outcomes.append(during)
assert current_grab[0] - key_region[0] < 100
elif next_grab[0] < key_region[0]:
grab_outcomes.append(base.GRABTYPES.FAIL_WITH_PELLET)
else:
raise ValueError('possibility not considered')
else:
grab_outcomes.append(post)
return grab_outcomes