utils.py 41.7 KB
Newer Older
lucas_miranda's avatar
lucas_miranda committed
1
# @author lucasmiranda42
2
3
4
5

import cv2
import matplotlib.pyplot as plt
import multiprocessing
6
import networkx as nx
7
import numpy as np
lucas_miranda's avatar
lucas_miranda committed
8
import os
9
import pandas as pd
10
import regex as re
11
import seaborn as sns
12
from copy import deepcopy
13
from itertools import combinations, product
14
15
from joblib import Parallel, delayed
from scipy import spatial
16
from scipy import stats
17
from sklearn import mixture
18
from tqdm import tqdm
19
20
21
22
23
from typing import Tuple, Any, List, Union, Dict, NewType

# DEFINE CUSTOM ANNOTATED TYPES #


24
Coordinates = NewType("Coordinates", Any)
25
26


27
# QUALITY CONTROL AND PREPROCESSING #
28

29

lucas_miranda's avatar
lucas_miranda committed
30
31
32
33
34
35
36
37
38
39
def likelihood_qc(dframe: pd.DataFrame, threshold: float = 0.9) -> np.array:
    """Returns a DataFrame filtered dataframe, keeping only the rows entirely above the threshold.

        Parameters:
            - dframe (pandas.DataFrame): DeepLabCut output, with positions over time and associated likelihhod
            - threshold (float): minimum acceptable confidence

        Returns:
            - filt_mask (np.array): mask on the rows of dframe"""

40
41
    Likes = np.array([dframe[i]["likelihood"] for i in list(dframe.columns.levels[0])])
    Likes = np.nan_to_num(Likes, nan=1.0)
lucas_miranda's avatar
lucas_miranda committed
42
43
44
    filt_mask = np.all(Likes > threshold, axis=0)

    return filt_mask
45
46


47
48
49
50
51
52
53
54
55
def bp2polar(tab: pd.DataFrame) -> pd.DataFrame:
    """Returns the DataFrame in polar coordinates.

        Parameters:
            - tab (pandas.DataFrame):Table with cartesian coordinates

        Returns:
            - polar (pandas.DataFrame): Equivalent to input, but with values in polar coordinates"""

56
57
58
59
60
61
62
    tab_ = np.array(tab)
    complex_ = tab_[:, 0] + 1j * tab_[:, 1]
    polar = pd.DataFrame(np.array([abs(complex_), np.angle(complex_)]).T)
    polar.rename(columns={0: "rho", 1: "phi"}, inplace=True)
    return polar


63
64
65
66
67
68
69
70
71
def tab2polar(cartesian_df: pd.DataFrame) -> pd.DataFrame:
    """Returns a pandas.DataFrame in which all the coordinates are polar.

        Parameters:
            - cartesian_df (pandas.DataFrame):DataFrame containing tables with cartesian coordinates

        Returns:
            - result (pandas.DataFrame): Equivalent to input, but with values in polar coordinates"""

72
    result = []
73
74
    for df in list(cartesian_df.columns.levels[0]):
        result.append(bp2polar(cartesian_df[df]))
75
76
    result = pd.concat(result, axis=1)
    idx = pd.MultiIndex.from_product(
77
78
        [list(cartesian_df.columns.levels[0]), ["rho", "phi"]],
        names=["bodyparts", "coords"],
79
80
81
82
83
    )
    result.columns = idx
    return result


84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
def compute_dist(
    pair_array: np.array, arena_abs: int = 1, arena_rel: int = 1
) -> pd.DataFrame:
    """Returns a pandas.DataFrame with the scaled distances between a pair of body parts.

        Parameters:
            - pair_array (numpy.array): np.array of shape N * 4 containing X,y positions
            over time for a given pair of body parts
            - arena_abs (int): diameter of the real arena in cm
            - arena_rel (int): diameter of the captured arena in pixels

        Returns:
            - result (pd.DataFrame): pandas.DataFrame with the
            absolute distances between a pair of body parts"""

lucas_miranda's avatar
lucas_miranda committed
99
100
    lim = 2 if pair_array.shape[1] == 4 else 1
    a, b = pair_array[:, :lim], pair_array[:, lim:]
101
    ab = a - b
lucas_miranda's avatar
lucas_miranda committed
102

103
    dist = np.sqrt(np.einsum("...i,...i", ab, ab))
104
105
106
    return pd.DataFrame(dist * arena_abs / arena_rel)


107
108
109
110
111
112
113
114
115
116
def bpart_distance(
    dataframe: pd.DataFrame, arena_abs: int = 1, arena_rel: int = 1
) -> pd.DataFrame:
    """Returns a pandas.DataFrame with the scaled distances between all pairs of body parts.

        Parameters:
            - dataframe (pandas.DataFrame): pd.DataFrame of shape N*(2*bp) containing X,y positions
        over time for a given set of bp body parts
            - arena_abs (int): diameter of the real arena in cm
            - arena_rel (int): diameter of the captured arena in pixels
117

118
119
120
121
122
        Returns:
            - result (pd.DataFrame): pandas.DataFrame with the
            absolute distances between all pairs of body parts"""

    indexes = combinations(dataframe.columns.levels[0], 2)
123
124
125
126
127
128
129
    dists = []
    for idx in indexes:
        dist = compute_dist(np.array(dataframe.loc[:, list(idx)]), arena_abs, arena_rel)
        dist.columns = [idx]
        dists.append(dist)

    return pd.concat(dists, axis=1)
130
131


132
133
134
135
136
137
138
def angle(a: np.array, b: np.array, c: np.array) -> np.array:
    """Returns a numpy.array with the angles between the provided instances.

        Parameters:
            - a (2D np.array): positions over time for a bodypart
            - b (2D np.array): positions over time for a bodypart
            - c (2D np.array): positions over time for a bodypart
lucas_miranda's avatar
lucas_miranda committed
139

140
141
142
        Returns:
            - ang (1D np.array): angles between the three-point-instances"""

lucas_miranda's avatar
lucas_miranda committed
143
144
145
    ba = a - b
    bc = c - b

146
    cosine_angle = np.einsum("...i,...i", ba, bc) / (
lucas_miranda's avatar
lucas_miranda committed
147
148
        np.linalg.norm(ba, axis=1) * np.linalg.norm(bc, axis=1)
    )
149
150
151
152
153
154
155
    ang = np.arccos(cosine_angle)

    return ang


def angle_trio(bpart_array: np.array) -> np.array:
    """Returns a numpy.array with all three possible angles between the provided instances.
lucas_miranda's avatar
lucas_miranda committed
156

157
158
        Parameters:
            - bpart_array (2D numpy.array): positions over time for a bodypart
lucas_miranda's avatar
lucas_miranda committed
159

160
161
        Returns:
            - ang_trio (2D numpy.array): all-three angles between the three-point-instances"""
lucas_miranda's avatar
lucas_miranda committed
162

163
164
    a, b, c = bpart_array
    ang_trio = np.array([angle(a, b, c), angle(a, c, b), angle(b, a, c)])
lucas_miranda's avatar
lucas_miranda committed
165

166
    return ang_trio
lucas_miranda's avatar
lucas_miranda committed
167
168


169
170
171
172
def rotate(
    p: np.array, angles: np.array, origin: np.array = np.array([0, 0])
) -> np.array:
    """Returns a numpy.array with the initial values rotated by angles radians
lucas_miranda's avatar
lucas_miranda committed
173

174
175
176
177
178
179
180
        Parameters:
            - p (2D numpy.array): array containing positions of bodyparts over time
            - angles (2D numpy.array): set of angles (in radians) to rotate p with
            - origin (2D numpy.array): rotation axis (zero vector by default)

        Returns:
            - rotated (2D numpy.array): rotated positions over time"""
lucas_miranda's avatar
lucas_miranda committed
181

182
183
184
185
186
    R = np.array([[np.cos(angles), -np.sin(angles)], [np.sin(angles), np.cos(angles)]])

    o = np.atleast_2d(origin)
    p = np.atleast_2d(p)

187
188
189
190
    rotated = np.squeeze((R @ (p.T - o.T) + o.T).T)

    return rotated

191

192
193
194
def align_trajectories(data: np.array, mode: str = "all") -> np.array:
    """Returns a numpy.array with the positions rotated in a way that the center (0 vector)
    and the body part in the first column of data are aligned with the y axis.
195

196
197
198
199
200
        Parameters:
            - data (3D numpy.array): array containing positions of body parts over time, where
            shape is N (sliding window instances) * m (sliding window size) * l (features)
            - mode (string): specifies if *all* instances of each sliding window get
            aligned, or only the *center*
201

202
203
        Returns:
            - aligned_trajs (2D np.array): aligned positions over time"""
204

205
    angles = np.zeros(data.shape[0])
206
    data = deepcopy(data)
207
    dshape = data.shape
208

209
210
211
212
    if mode == "center":
        center_time = (data.shape[1] - 1) // 2
        angles = np.arctan2(data[:, center_time, 0], data[:, center_time, 1])
    elif mode == "all":
lucas_miranda's avatar
lucas_miranda committed
213
        data = data.reshape(-1, dshape[-1], order="C")
214
        angles = np.arctan2(data[:, 0], data[:, 1])
lucas_miranda's avatar
lucas_miranda committed
215
216
217
    elif mode == "none":
        data = data.reshape(-1, dshape[-1], order="C")
        angles = np.zeros(data.shape[0])
218
219
220
221
222

    aligned_trajs = np.zeros(data.shape)

    for frame in range(data.shape[0]):
        aligned_trajs[frame] = rotate(
lucas_miranda's avatar
lucas_miranda committed
223
224
            data[frame].reshape([-1, 2], order="C"), angles[frame],
        ).reshape(data.shape[1:], order="C")
225

lucas_miranda's avatar
lucas_miranda committed
226
227
    if mode == "all" or mode == "none":
        aligned_trajs = aligned_trajs.reshape(dshape, order="C")
228

229
230
231
    return aligned_trajs


232
233
234
235
236
237
238
239
240
def smooth_boolean_array(a: np.array) -> np.array:
    """Returns a boolean array in which isolated appearances of a feature are smoothened

        Parameters:
            - a (1D numpy.array): boolean instances

        Returns:
            - a (1D numpy.array): smoothened boolean instances"""

241
242
243
244
245
246
    for i in range(1, len(a) - 1):
        if a[i - 1] == a[i + 1]:
            a[i] = a[i - 1]
    return a == 1


247
248
249
def rolling_window(a: np.array, window_size: int, window_step: int) -> np.array:
    """Returns a 3D numpy.array with a sliding-window extra dimension

250
251
        Parameters:
            - a (2D np.array): N (instances) * m (features) shape
252

253
254
255
        Returns:
            - rolled_a (3D np.array):
            N (sliding window instances) * l (sliding window size) * m (features)"""
256

257
258
    shape = (a.shape[0] - window_size + 1, window_size) + a.shape[1:]
    strides = (a.strides[0],) + a.strides
259
260
    rolled_a = np.lib.stride_tricks.as_strided(
        a, shape=shape, strides=strides, writeable=True
261
    )[::window_step]
262
    return rolled_a
263

264

265
266
267
def smooth_mult_trajectory(series: np.array, alpha: float = 0.15) -> np.array:
    """Returns a smooths a trajectory using exponentially weighted averages

268
269
        Parameters:
            - series (numpy.array): 1D trajectory array with N (instances) - alpha (float): 0 <= alpha <= 1;
270
271
            indicates the inverse weight assigned to previous observations. Higher (alpha~1) indicates less smoothing;
            lower indicates more (alpha~0)
272
273
274

        Returns:
            - smoothed_series (np.array): smoothed version of the input, with equal shape"""
275
276
277
278
279

    result = [series[0]]
    for n in range(len(series)):
        result.append(alpha * series[n] + (1 - alpha) * result[n - 1])

280
281
282
    smoothed_series = np.array(result)

    return smoothed_series
283

lucas_miranda's avatar
lucas_miranda committed
284
285

# BEHAVIOUR RECOGNITION FUNCTIONS #
286
287


288
def close_single_contact(
289
290
291
292
293
294
    pos_dframe: pd.DataFrame,
    left: str,
    right: str,
    tol: float,
    arena_abs: int,
    arena_rel: int,
295
296
) -> np.array:
    """Returns a boolean array that's True if the specified body parts are closer than tol.
297

298
299
300
301
302
        Parameters:
            - pos_dframe (pandas.DataFrame): DLC output as pandas.DataFrame; only applicable
            to two-animal experiments.
            - left (string): First member of the potential contact
            - right (string): Second member of the potential contact
303
304
305
            - tol (float): maximum distance for which a contact is reported
            - arena_abs (int): length in mm of the diameter of the real arena
            - arena_rel (int): length in pixels of the diameter of the arena in the video
306

307
308
309
        Returns:
            - contact_array (np.array): True if the distance between the two specified points
            is less than tol, False otherwise"""
310

311
312
313
    close_contact = (
        np.linalg.norm(pos_dframe[left] - pos_dframe[right], axis=1) * arena_abs
    ) / arena_rel < tol
314

315
    return close_contact
316
317


318
319
320
321
322
323
324
def close_double_contact(
    pos_dframe: pd.DataFrame,
    left1: str,
    left2: str,
    right1: str,
    right2: str,
    tol: float,
325
326
    arena_abs: int,
    arena_rel: int,
327
328
329
330
331
332
333
334
335
336
337
    rev: bool = False,
) -> np.array:
    """Returns a boolean array that's True if the specified body parts are closer than tol.

        Parameters:
            - pos_dframe (pandas.DataFrame): DLC output as pandas.DataFrame; only applicable
            to two-animal experiments.
            - left1 (string): First contact point of animal 1
            - left2 (string): Second contact point of animal 1
            - right1 (string): First contact point of animal 2
            - right2 (string): Second contact point of animal 2
338
339
340
341
            - tol (float): maximum distance for which a contact is reported
            - arena_abs (int): length in mm of the diameter of the real arena
            - arena_rel (int): length in pixels of the diameter of the arena in the video
            - rev (bool): reverses the default behaviour (nose2tail contact for both mice)
342
343
344
345
346
347
348

        Returns:
            - double_contact (np.array): True if the distance between the two specified points
            is less than tol, False otherwise"""

    if rev:
        double_contact = (
349
350
351
352
353
354
355
356
            (np.linalg.norm(pos_dframe[right1] - pos_dframe[left2], axis=1) * arena_abs)
            / arena_rel
            < tol
        ) & (
            (np.linalg.norm(pos_dframe[right2] - pos_dframe[left1], axis=1) * arena_abs)
            / arena_rel
            < tol
        )
357
358
359

    else:
        double_contact = (
360
361
362
363
364
365
366
367
            (np.linalg.norm(pos_dframe[right1] - pos_dframe[left1], axis=1) * arena_abs)
            / arena_rel
            < tol
        ) & (
            (np.linalg.norm(pos_dframe[right2] - pos_dframe[left2], axis=1) * arena_abs)
            / arena_rel
            < tol
        )
368
369

    return double_contact
370
371
372


def recognize_arena(
lucas_miranda's avatar
lucas_miranda committed
373
374
375
376
377
    videos: list,
    vid_index: int,
    path: str = ".",
    recoglimit: int = 1,
    arena_type: str = "circular",
378
) -> Tuple[np.array, int, int]:
lucas_miranda's avatar
lucas_miranda committed
379
380
381
    """Returns numpy.array with information about the arena recognised from the first frames
    of the video. WARNING: estimates won't be reliable if the camera moves along the video.

382
383
384
385
386
387
        Parameters:
            - videos (list): relative paths of the videos to analise
            - vid_index (int): element of videos to use
            - path (string): full path of the directory where the videos are
            - recoglimit (int): number of frames to use for position estimates
            - arena_type (string): arena type; must be one of ['circular']
lucas_miranda's avatar
lucas_miranda committed
388

389
390
        Returns:
            - arena (np.array): 1D-array containing information about the arena.
391
392
393
            "circular" (3-element-array) -> x-y position of the center and the radius
            - h (int): height of the video in pixels
            - w (int): width of the video in pixels"""
lucas_miranda's avatar
lucas_miranda committed
394
395

    cap = cv2.VideoCapture(os.path.join(path, videos[vid_index]))
396
397

    # Loop over the first frames in the video to get resolution and center of the arena
lucas_miranda's avatar
lucas_miranda committed
398
    arena, fnum, h, w = False, 0, None, None
399
400
401
402
403
404
405
406
407
408
409
410

    while cap.isOpened() and fnum < recoglimit:
        ret, frame = cap.read()
        # if frame is read correctly ret is True
        if not ret:
            print("Can't receive frame (stream end?). Exiting ...")
            break

        if arena_type == "circular":

            # Detect arena and extract positions
            arena = circular_arena_recognition(frame)[0]
411
            if h is None and w is None:
412
413
414
415
                h, w = frame.shape[0], frame.shape[1]

        fnum += 1

416
    return arena, h, w
417
418


419
420
def circular_arena_recognition(frame: np.array) -> np.array:
    """Returns x,y position of the center and the radius of the recognised arena
lucas_miranda's avatar
lucas_miranda committed
421

422
        Parameters:
423
            - frame (np.array): numpy.array representing an individual frame of a video
424

425
426
427
        Returns:
            - circles (np.array): 3-element-array containing x,y positions of the center
            of the arena, and a third value indicating the radius"""
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453

    # Convert image to greyscale, threshold it, blur it and detect the biggest best fitting circle
    # using the Hough algorithm
    gray_image = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)
    ret, thresh = cv2.threshold(gray_image, 50, 255, 0)
    frame = cv2.medianBlur(thresh, 9)
    circle = cv2.HoughCircles(
        frame,
        cv2.HOUGH_GRADIENT,
        1,
        300,
        param1=50,
        param2=10,
        minRadius=0,
        maxRadius=0,
    )

    circles = []

    if circle is not None:
        circle = np.uint16(np.around(circle[0]))
        circles.append(circle)

    return circles[0]


454
455
456
457
def climb_wall(
    arena_type: str, arena: np.array, pos_dict: pd.DataFrame, tol: float, nose: str
) -> np.array:
    """Returns True if the specified mouse is climbing the wall
lucas_miranda's avatar
lucas_miranda committed
458

459
460
461
462
463
464
465
466
467
468
469
470
471
        Parameters:
            - arena_type (str): arena type; must be one of ['circular']
            - arena (np.array): contains arena location and shape details
            - pos_dict (table_dict): position over time for all videos in a project
            - tol (float): minimum tolerance to report a hit
            - nose (str): indicates the name of the body part representing the nose of
            the selected animal

        Returns:
            - climbing (np.array): boolean array. True if selected animal
            is climbing the walls of the arena"""

    nose = pos_dict[nose]
472

473
474
475
476
477
478
    if arena_type == "circular":
        center = np.array(arena[:2])
        climbing = np.linalg.norm(nose - center, axis=1) > (arena[2] + tol)

    else:
        raise NotImplementedError("Supported values for arena_type are ['circular']")
479

480
    return climbing
481
482


lucas_miranda's avatar
lucas_miranda committed
483
484
485
486
def rolling_speed(
    dframe: pd.DatetimeIndex, window: int = 10, rounds: int = 10, deriv: int = 1
) -> pd.DataFrame:
    """Returns the average speed over n frames in pixels per frame
lucas_miranda's avatar
lucas_miranda committed
487

lucas_miranda's avatar
lucas_miranda committed
488
489
490
491
492
493
        Parameters:
            - dframe (pandas.DataFrame): position over time dataframe
            - pause (int):  frame-length of the averaging window
            - rounds (int): float rounding decimals
            - deriv (int): position derivative order; 1 for speed,
            2 for acceleration, 3 for jerk, etc
lucas_miranda's avatar
lucas_miranda committed
494

lucas_miranda's avatar
lucas_miranda committed
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
        Returns:
            - speeds (pd.DataFrame): containing 2D speeds for each body part
            in the original data or their consequent derivatives"""

    original_shape = dframe.shape
    body_parts = dframe.columns.levels[0]
    speeds = pd.DataFrame

    for der in range(deriv):
        distances = np.concatenate(
            [
                np.array(dframe).reshape([-1, (2 if der == 0 else 1)], order="F"),
                np.array(dframe.shift()).reshape(
                    [-1, (2 if der == 0 else 1)], order="F"
                ),
            ],
            axis=1,
        )
lucas_miranda's avatar
lucas_miranda committed
513

lucas_miranda's avatar
lucas_miranda committed
514
515
516
517
518
519
520
        distances = np.array(compute_dist(distances))
        distances = distances.reshape(
            [original_shape[0], original_shape[1] // 2], order="F"
        )
        distances = pd.DataFrame(distances, index=dframe.index)
        speeds = np.round(distances.rolling(window).mean(), rounds)
        speeds[np.isnan(speeds)] = 0.0
lucas_miranda's avatar
lucas_miranda committed
521

lucas_miranda's avatar
lucas_miranda committed
522
        dframe = speeds
lucas_miranda's avatar
lucas_miranda committed
523

lucas_miranda's avatar
lucas_miranda committed
524
    speeds.columns = body_parts
525
526
527
528

    return speeds


529
530
531
def huddle(
    pos_dframe: pd.DataFrame, tol_forward: float, tol_spine: float, tol_speed: float
) -> np.array:
lucas_miranda's avatar
lucas_miranda committed
532
533
    """Returns true when the mouse is huddling using simple rules. (!!!) Designed to
    work with deepof's default DLC mice models; not guaranteed to work otherwise.
534

lucas_miranda's avatar
lucas_miranda committed
535
536
537
538
539
540
        Parameters:
            - pos_dframe (pandas.DataFrame):
            - tol_forward (float): Maximum tolerated distance between ears and
            forward limbs
            - tol_rear (float): Maximum tolerated average distance between spine
            body parts
lucas_miranda's avatar
lucas_miranda committed
541

lucas_miranda's avatar
lucas_miranda committed
542
543
544
545
546
547
548
549
550
551
        Returns:
            hudd (np.array): True if the animal is huddling, False otherwise
        """

    forward = (
        np.linalg.norm(pos_dframe["Left_ear"] - pos_dframe["Left_fhip"], axis=1)
        < tol_forward
    ) & (
        np.linalg.norm(pos_dframe["Right_ear"] - pos_dframe["Right_fhip"], axis=1)
        < tol_forward
552
553
    )

554
    spine = ["Spine_1", "Center", "Spine_2", "Tail_base"]
lucas_miranda's avatar
lucas_miranda committed
555
556
557
558
559
560
561
562
563
564
565
566
567
    spine_dists = []
    for comb in range(2):
        spine_dists.append(
            np.linalg.norm(
                pos_dframe[spine[comb]] - pos_dframe[spine[comb + 1]], axis=1
            )
        )
    spine = np.mean(spine_dists) < tol_spine

    hudd = forward & spine

    return hudd

568

lucas_miranda's avatar
lucas_miranda committed
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
def following_path(
    distance_dframe: pd.DataFrame,
    position_dframe: pd.DataFrame,
    follower: str,
    followed: str,
    frames: int = 20,
    tol: float = 0,
) -> np.array:
    """For multi animal videos only. Returns True if 'follower' is closer than tol to the path that
    followed has walked over the last specified number of frames

        Parameters:
            - distance_dframe (pandas.DataFrame): distances between bodyparts; generated by the preprocess module
            - position_dframe (pandas.DataFrame): position of bodyparts; generated by the preprocess module
            - follower (str) identifier for the animal who's following
            - followed (str) identifier for the animal who's followed
            - frames (int) frames in which to track whether the process consistently occurs,
            - tol (float) Maximum distance for which True is returned

        Returns:
            - follow (np.array): boolean sequence, True if conditions are fulfilled, False otherwise"""
590
591

    # Check that follower is close enough to the path that followed has passed though in the last frames
lucas_miranda's avatar
lucas_miranda committed
592
593
594
    shift_dict = {
        i: position_dframe[followed + "_Tail_base"].shift(i) for i in range(frames)
    }
595
596
    dist_df = pd.DataFrame(
        {
lucas_miranda's avatar
lucas_miranda committed
597
598
599
            i: np.linalg.norm(
                position_dframe[follower + "_Nose"] - shift_dict[i], axis=1
            )
600
601
602
603
604
605
            for i in range(frames)
        }
    )

    # Check that the animals are oriented follower's nose -> followed's tail
    right_orient1 = (
lucas_miranda's avatar
lucas_miranda committed
606
607
608
609
        distance_dframe[tuple(sorted([follower + "_Nose", followed + "_Tail_base"]))]
        < distance_dframe[
            tuple(sorted([follower + "_Tail_base", followed + "_Tail_base"]))
        ]
610
611
612
    )

    right_orient2 = (
lucas_miranda's avatar
lucas_miranda committed
613
614
        distance_dframe[tuple(sorted([follower + "_Nose", followed + "_Tail_base"]))]
        < distance_dframe[tuple(sorted([follower + "_Nose", followed + "_Nose"]))]
615
616
    )

lucas_miranda's avatar
lucas_miranda committed
617
618
    follow = np.all(
        np.array([(dist_df.min(axis=1) < tol), right_orient1, right_orient2]), axis=0,
619
620
    )

lucas_miranda's avatar
lucas_miranda committed
621
622
    return follow

623

lucas_miranda's avatar
lucas_miranda committed
624
def single_behaviour_analysis(
625
626
627
628
629
630
631
632
    behaviour_name: str,
    treatment_dict: dict,
    behavioural_dict: dict,
    plot: int = 0,
    stat_tests: bool = True,
    save: str = None,
    ylim: float = None,
) -> list:
633
    """Given the name of the behaviour, a dictionary with the names of the groups to compare, and a dictionary
634
635
636
637
638
639
640
641
642
643
644
645
646
647
       with the actual tags, outputs a box plot and a series of significance tests amongst the groups

        Parameters:
            - behaviour_name (str): name of the behavioural trait to analize
            - treatment_dict (dict): dictionary containing video names as keys and experimental conditions as values
            - behavioural_dict (dict): tagged dictionary containing video names as keys and annotations as values
            - plot (int): Silent if 0; otherwise, indicates the dpi of the figure to plot
            - stat_tests (bool): performs FDR corrected Mann-U non-parametric tests among all groups if True
            - save (str): Saves the produced figure to the specified file
            - ylim (float): y-limit for the boxplot. Ignored if plot == False

        Returns:
            - beh_dict (dict): dictionary containing experimental conditions as keys and video names as values
            - stat_dict (dict): dictionary containing condition pairs as keys and stat results as values"""
648
649
650
651
652
653
654
655
656
657

    beh_dict = {condition: [] for condition in treatment_dict.keys()}

    for condition in beh_dict.keys():
        for ind in treatment_dict[condition]:
            beh_dict[condition].append(
                np.sum(behavioural_dict[ind][behaviour_name])
                / len(behavioural_dict[ind][behaviour_name])
            )

658
    return_list = [beh_dict]
659

660
    if plot > 0:
661

662
        fig, ax = plt.subplots(dpi=plot)
663

664
665
666
667
668
669
670
671
672
        sns.boxplot(
            list(beh_dict.keys()), list(beh_dict.values()), orient="vertical", ax=ax
        )

        ax.set_title("{} across groups".format(behaviour_name))
        ax.set_ylabel("Proportion of frames")

        if ylim is not None:
            ax.set_ylim(ylim)
673

674
        if save is not None:
675
676
            plt.savefig(save)

677
        return_list.append(ax)
678

679
680
    if stat_tests:
        stat_dict = {}
681
        for i in combinations(treatment_dict.keys(), 2):
682
683
684
685
686
687
688
689
690
691
692
            # Solves issue with automatically generated examples
            if (
                beh_dict[i[0]] == beh_dict[i[1]]
                or np.var(beh_dict[i[0]]) == 0
                or np.var(beh_dict[i[1]]) == 0
            ):
                stat_dict[i] = "Identical sources. Couldn't run"
            else:
                stat_dict[i] = stats.mannwhitneyu(
                    beh_dict[i[0]], beh_dict[i[1]], alternative="two-sided"
                )
693
        return_list.append(stat_dict)
694

695
    return return_list
696
697


698
699
700
701
702
703
704
705
706
707
708
709
def max_behaviour(
    behaviour_dframe: pd.DataFrame, window_size: int = 10, stepped: bool = False
) -> np.array:
    """Returns the most frequent behaviour in a window of window_size frames

        Parameters:
                - behaviour_dframe (pd.DataFrame): boolean matrix containing occurrence
                of tagged behaviours per frame in the video
                - window_size (int): size of the window to use when computing
                the maximum behaviour per time slot
                - stepped (bool): sliding windows don't overlap if True. False by default

710
711
712
        Returns:
            - max_array (np.array): string array with the most common behaviour per instance
            of the sliding window"""
713
714
715
716
717
718
719

    speeds = [col for col in behaviour_dframe.columns if "speed" in col.lower()]

    behaviour_dframe = behaviour_dframe.drop(speeds, axis=1).astype("float")
    win_array = behaviour_dframe.rolling(window_size, center=True).sum()
    if stepped:
        win_array = win_array[::window_size]
720
721
    max_array = win_array[1:].idxmax(axis=1)

722
723
724
725
    return np.array(max_array)


# MACHINE LEARNING FUNCTIONS #
726
727


728
729
730
731
732
733
734
735
736
737
738
739
740
def gmm_compute(x: np.array, n_components: int, cv_type: str) -> list:
    """Fits a Gaussian Mixture Model to the provided data and returns evaluation metrics.

        Parameters:
            - x (numpy.array): data matrix to train the model
            - n_components (int): number of Gaussian components to use
            - cv_type (str): covariance matrix type to use.
            Must be one of "spherical", "tied", "diag", "full"

        Returns:
            - gmm_eval (list): model and associated BIC for downstream selection
    """

741
742
743
744
745
746
747
    gmm = mixture.GaussianMixture(
        n_components=n_components,
        covariance_type=cv_type,
        max_iter=100000,
        init_params="kmeans",
    )
    gmm.fit(x)
748
749
750
751
752
    gmm_eval = [gmm, gmm.bic(x)]
    return gmm_eval


def gmm_model_selection(
753
    x: pd.DataFrame,
754
755
756
757
758
759
760
761
    n_components_range: range,
    part_size: int,
    n_runs: int = 100,
    n_cores: int = False,
    cv_types: Tuple = ("spherical", "tied", "diag", "full"),
) -> Tuple[List[list], List[np.ndarray], Union[int, Any]]:
    """Runs GMM clustering model selection on the specified X dataframe, outputs the bic distribution per model,
       a vector with the median BICs and an object with the overall best model
762

763
        Parameters:
764
            - x (pandas.DataFrame): data matrix to train the models
765
766
767
768
769
            - n_components_range (range): generator with numbers of components to evaluate
            - n_runs (int): number of bootstraps for each model
            - part_size (int): size of bootstrap samples for each model
            - n_cores (int): number of cores to use for computation
            - cv_types (tuple): Covariance Matrices to try. All four available by default
770

771
772
773
774
775
776
777
        Returns:
            - bic (list): All recorded BIC values for all attempted parameter combinations
            (useful for plotting)
            - m_bic(list): All minimum BIC values recorded throughout the process
            (useful for plottinh)
            - best_bic_gmm (sklearn.GMM): unfitted version of the best found model
    """
778
779
780
781
782
783
784
785

    # Set the default of n_cores to the most efficient value
    if not n_cores:
        n_cores = min(multiprocessing.cpu_count(), n_runs)

    bic = []
    m_bic = []
    lowest_bic = np.inf
786
    best_bic_gmm = 0
787
788
789
790
791
792
793
794

    pbar = tqdm(total=len(cv_types) * len(n_components_range))

    for cv_type in cv_types:

        for n_components in n_components_range:

            res = Parallel(n_jobs=n_cores, prefer="threads")(
795
796
797
798
                delayed(gmm_compute)(
                    x.sample(part_size, replace=True), n_components, cv_type
                )
                for _ in range(n_runs)
799
800
801
802
803
804
805
806
807
808
809
            )
            bic.append([i[1] for i in res])

            pbar.update(1)
            m_bic.append(np.median([i[1] for i in res]))
            if m_bic[-1] < lowest_bic:
                lowest_bic = m_bic[-1]
                best_bic_gmm = res[0][0]

    return bic, m_bic, best_bic_gmm

810
811

# RESULT ANALYSIS FUNCTIONS #
812
813
814


def cluster_transition_matrix(
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
    cluster_sequence: np.array,
    nclusts: int,
    autocorrelation: bool = True,
    return_graph: bool = False,
) -> Tuple[Union[nx.Graph, Any], np.ndarray]:
    """Computes the transition matrix between clusters and the autocorrelation in the sequence.

        Parameters:
            - cluster_sequence (numpy.array):
            - nclusts (int):
            - autocorrelation (bool):
            - return_graph (bool):

        Returns:
            - trans_normed (numpy.array / networkx.Graph:
            - autocorr (numpy.array):
831
832
833
    """

    # Stores all possible transitions between clusters
834
835
836
    clusters = [str(i) for i in range(nclusts)]
    cluster_sequence = cluster_sequence.astype(str)

837
838
839
840
841
842
843
    trans = {t: 0 for t in product(clusters, clusters)}
    k = len(clusters)

    # Stores the cluster sequence as a string
    transtr = "".join(list(cluster_sequence))

    # Assigns to each transition the number of times it occurs in the sequence
844
    for t in trans.keys():
845
846
847
        trans[t] = len(re.findall("".join(t), transtr, overlapped=True))

    # Normalizes the counts to add up to 1 for each departing cluster
848
849
    trans_normed = np.zeros([k, k]) + 1e-5
    for t in trans.keys():
850
        trans_normed[int(t[0]), int(t[1])] = np.round(
851
852
853
            trans[t]
            / (sum({i: j for i, j in trans.items() if i[0] == t[0]}.values()) + 1e-5),
            3,
854
855
856
857
858
859
860
861
        )

    # If specified, returns the transition matrix as an nx.Graph object
    if return_graph:
        trans_normed = nx.Graph(trans_normed)

    if autocorrelation:
        cluster_sequence = list(map(int, cluster_sequence))
862
863
        autocorr = np.corrcoef(cluster_sequence[:-1], cluster_sequence[1:])
        return trans_normed, autocorr
864
865
866

    return trans_normed

867

868
869
870
871
872
873
# MAIN BEHAVIOUR TAGGING FUNCTION #


def rule_based_tagging(
    tracks: List,
    videos: List,
874
    coordinates: Coordinates,
875
876
877
878
879
880
881
882
883
884
885
886
    vid_index: int,
    arena_abs: int,
    animal_ids: List = None,
    show: bool = False,
    save: bool = False,
    fps: float = 25.0,
    speed_pause: int = 50,
    frame_limit: float = np.inf,
    recog_limit: int = 1,
    path: str = os.path.join("./"),
    arena_type: str = "circular",
    classifiers: Dict = None,
887
888
) -> pd.DataFrame:
    """Outputs a dataframe with the motives registered per frame."""
889
890
891

    vid_name = re.findall("(.*?)_", tracks[vid_index])[0]

892
893
    distances = coordinates.get_coords()[vid_name]
    speeds = coordinates.get_coords(speed=1)[vid_name]
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
    arena, h, w = recognize_arena(videos, vid_index, path, recog_limit, arena_type)

    # Dictionary with motives per frame
    behavioural_tags = []
    if animal_ids:
        behavioural_tags.append(["nose2nose", "sidebyside", "sidereside"])
        for _id in animal_ids:
            for behaviour in [
                "_nose2tail",
                "_climbing",
                "_huddle",
                "_following",
                "_speed",
            ]:
                behavioural_tags.append(_id + behaviour)

    else:
911
        behavioural_tags += ["huddle", "climbing", "speed"]
912

913
    tag_dict = {tag: np.zeros(distances.shape[0]) for tag in behavioural_tags}
914
915
916
917
918

    if animal_ids:
        # Define behaviours that can be computed on the fly from the distance matrix
        tag_dict["nose2nose"] = smooth_boolean_array(
            close_single_contact(
919
                distances,
920
921
922
923
924
925
926
927
928
                animal_ids[0] + "_Nose",
                animal_ids[1] + "_Nose",
                15.0,
                arena_abs,
                arena[2],
            )
        )
        tag_dict[animal_ids[0] + "_nose2tail"] = smooth_boolean_array(
            close_single_contact(
929
                distances,
930
931
932
933
934
935
936
937
938
                animal_ids[0] + "_Nose",
                animal_ids[1] + "_Tail_base",
                15.0,
                arena_abs,
                arena[2],
            )
        )
        tag_dict[animal_ids[1] + "_nose2tail"] = smooth_boolean_array(
            close_single_contact(
939
                distances,
940
941
942
943
944
945
946
947
948
                animal_ids[1] + "_Nose",
                animal_ids[0] + "_Tail_base",
                15.0,
                arena_abs,
                arena[2],
            )
        )
        tag_dict["sidebyside"] = smooth_boolean_array(
            close_double_contact(
949
                distances,
950
951
952
953
954
955
956
957
958
959
960
961
                animal_ids[0] + "_Nose",
                animal_ids[0] + "_Tail_base",
                animal_ids[1] + "_Nose",
                animal_ids[1] + "_Tail_base",
                15.0,
                rev=False,
                arena_abs=arena_abs,
                arena_rel=arena[2],
            )
        )
        tag_dict["sidereside"] = smooth_boolean_array(
            close_double_contact(
962
                distances,
963
964
965
966
967
968
969
970
971
972
973
974
975
                animal_ids[0] + "_Nose",
                animal_ids[0] + "_Tail_base",
                animal_ids[1] + "_Nose",
                animal_ids[1] + "_Tail_base",
                15.0,
                rev=True,
                arena_abs=arena_abs,
                arena_rel=arena[2],
            )
        )
        for _id in animal_ids:
            tag_dict[_id + "_following"] = smooth_boolean_array(
                following_path(
976
977
                    distances[vid_name],
                    distances,
978
979
980
981
982
983
984
985
986
987
                    follower=_id,
                    followed=[i for i in animal_ids if i != _id][0],
                    frames=20,
                    tol=20,
                )
            )
            tag_dict[_id + "_climbwall"] = smooth_boolean_array(
                pd.Series(
                    (
                        spatial.distance.cdist(
988
                            np.array(distances[_id + "_Nose"]), np.array([arena[:2]])
989
990
                        )
                        > (w / 200 + arena[2])
991
992
                    ).reshape(distances.shape[0]),
                    index=distances.index,
993
994
                )
            )
995
            tag_dict[_id + "_speed"] = speeds[_id + "_speed"]
996
997

    else:
998
        print(w)
999
1000
1001
1002
        tag_dict["climbwall"] = smooth_boolean_array(
            pd.Series(
                (
                    spatial.distance.cdist(
1003
                        np.array(distances["Nose"]), np.array([arena[:2]])
1004
1005
                    )
                    > (w / 200 + arena[2])
1006
1007
                ).reshape(distances.shape[0]),
                index=distances.index,
1008
1009
            )
        )
1010
        tag_dict["speed"] = speeds["Center"]
1011

1012
    if classifiers and "huddle" in classifiers:
1013
1014
        mouse_X = {
            _id: np.array(
1015
                distances[vid_name][
1016
1017
                    [
                        j
1018
                        for j in distances[vid_name].keys()
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
                        if (len(j) == 2 and _id in j[0] and _id in j[1])
                    ]
                ]
            )
            for _id in animal_ids
        }
        for _id in animal_ids:
            tag_dict[_id + "_huddle"] = smooth_boolean_array(
                classifiers["huddle"].predict(mouse_X[_id])
            )
    else:
        try:
            for _id in animal_ids:
                tag_dict[_id + "_huddle"] = smooth_boolean_array(
1033
                    huddle(distances, 25, 25, 5)
1034
1035
                )
        except TypeError:
1036
            tag_dict["huddle"] = smooth_boolean_array(huddle(distances, 25, 25, 5))
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197

    # if any([show, save]):
    #     cap = cv2.VideoCapture(path + videos[vid_index])

    # # Keep track of the frame number, to align with the tracking data
    # fnum = 0
    # if save:
    #     writer = None

    #     # Loop over the frames in the video
    #     pbar = tqdm(total=min(dframe.shape[0] - recog_limit, frame_limit))
    #     while cap.isOpened() and fnum < frame_limit:
    #
    #         ret, frame = cap.read()
    #         # if frame is read correctly ret is True
    #         if not ret:
    #             print("Can't receive frame (stream end?). Exiting ...")
    #             break
    #
    #         font = cv2.FONT_HERSHEY_COMPLEX_SMALL
    #
    #         if like_qc_dict[vid_name][fnum]:
    #
    #             # Extract positions
    #             pos_dict = {
    #                 i: np.array([dframe[i]["x"][fnum], dframe[i]["y"][fnum]])
    #                 for i in dframe.columns.levels[0]
    #                 if i != "Like_QC"
    #             }
    #
    #             if h is None and w is None:
    #                 h, w = frame.shape[0], frame.shape[1]
    #
    #             # Label positions
    #             downleft = (int(w * 0.3 / 10), int(h / 1.05))
    #             downright = (int(w * 6.5 / 10), int(h / 1.05))
    #             upleft = (int(w * 0.3 / 10), int(h / 20))
    #             upright = (int(w * 6.3 / 10), int(h / 20))
    #
    #             # Display all annotations in the output video
    #             if tag_dict["nose2nose"][fnum] and not tag_dict["sidebyside"][fnum]:
    #                 cv2.putText(
    #                     frame,
    #                     "Nose-Nose",
    #                     (downleft if bspeed > wspeed else downright),
    #                     font,
    #                     1,
    #                     (255, 255, 255),
    #                     2,
    #                 )
    #             if tag_dict["bnose2tail"][fnum] and not tag_dict["sidereside"][fnum]:
    #                 cv2.putText(
    #                     frame, "Nose-Tail", downleft, font, 1, (255, 255, 255), 2
    #                 )
    #             if tag_dict["wnose2tail"][fnum] and not tag_dict["sidereside"][fnum]:
    #                 cv2.putText(
    #                     frame, "Nose-Tail", downright, font, 1, (255, 255, 255), 2
    #                 )
    #             if tag_dict["sidebyside"][fnum]:
    #                 cv2.putText(
    #                     frame,
    #                     "Side-side",
    #                     (downleft if bspeed > wspeed else downright),
    #                     font,
    #                     1,
    #                     (255, 255, 255),
    #                     2,
    #                 )
    #             if tag_dict["sidereside"][fnum]:
    #                 cv2.putText(
    #                     frame,
    #                     "Side-Rside",
    #                     (downleft if bspeed > wspeed else downright),
    #                     font,
    #                     1,
    #                     (255, 255, 255),
    #                     2,
    #                 )
    #             if tag_dict["bclimbwall"][fnum]:
    #                 cv2.putText(
    #                     frame, "Climbing", downleft, font, 1, (255, 255, 255), 2
    #                 )
    #             if tag_dict["wclimbwall"][fnum]:
    #                 cv2.putText(
    #                     frame, "Climbing", downright, font, 1, (255, 255, 255), 2
    #                 )
    #             if tag_dict["bhuddle"][fnum] and not tag_dict["bclimbwall"][fnum]:
    #                 cv2.putText(frame, "huddle", downleft, font, 1, (255, 255, 255), 2)
    #             if tag_dict["whuddle"][fnum] and not tag_dict["wclimbwall"][fnum]:
    #                 cv2.putText(frame, "huddle", downright, font, 1, (255, 255, 255), 2)
    #             if tag_dict["bfollowing"][fnum] and not tag_dict["bclimbwall"][fnum]:
    #                 cv2.putText(
    #                     frame,
    #                     "*f",
    #                     (int(w * 0.3 / 10), int(h / 10)),
    #                     font,
    #                     1,
    #                     ((150, 150, 255) if wspeed > bspeed else (150, 255, 150)),
    #                     2,
    #                 )
    #             if tag_dict["wfollowing"][fnum] and not tag_dict["wclimbwall"][fnum]:
    #                 cv2.putText(
    #                     frame,
    #                     "*f",
    #                     (int(w * 6.3 / 10), int(h / 10)),
    #                     font,
    #                     1,
    #                     ((150, 150, 255) if wspeed < bspeed else (150, 255, 150)),
    #                     2,
    #                 )
    #
    #             if (bspeed == None and wspeed == None) or fnum % speed_pause == 0:
    #                 bspeed = tag_dict["bspeed"][fnum]
    #                 wspeed = tag_dict["wspeed"][fnum]
    #
    #             cv2.putText(
    #                 frame,
    #                 "W: " + str(np.round(wspeed, 2)) + " mmpf",
    #                 (upright[0] - 20, upright[1]),
    #                 font,
    #                 1,
    #                 ((150, 150, 255) if wspeed < bspeed else (150, 255, 150)),
    #                 2,
    #             )
    #             cv2.putText(
    #                 frame,
    #                 "B: " + str(np.round(bspeed, 2)) + " mmpf",
    #                 upleft,
    #                 font,
    #                 1,
    #                 ((150, 150, 255) if bspeed < wspeed else (150, 255, 150)),
    #                 2,
    #             )
    #
    #             if show:
    #                 cv2.imshow("frame", frame)
    #
    #             if save:
    #
    #                 if writer is None:
    #                     # Define the codec and create VideoWriter object.The output is stored in 'outpy.avi' file.
    #                     # Define the FPS. Also frame size is passed.
    #                     writer = cv2.VideoWriter()
    #                     writer.open(
    #                         re.findall("(.*?)_", tracks[vid_index])[0] + "_tagged.avi",
    #                         cv2.VideoWriter_fourcc(*"MJPG"),
    #                         fps,
    #                         (frame.shape[1], frame.shape[0]),
    #                         True,
    #                     )
    #                 writer.write(frame)
    #
    #         if cv2.waitKey(1) == ord("q"):
    #             break
    #
    #         pbar.update(1)
    #         fnum += 1

    # cap.release()
    # cv2.destroyAllWindows()

1198
    tag_df = pd.DataFrame(tag_dict)
1199

1200
    return tag_df
1201
1202


1203
1204
# TODO:
#    - Add sequence plot to single_behaviour_analysis (show how the condition varies across a specified time window)
1205
1206
1207
#    - Add digging to rule_based_tagging
#    - Add center to rule_based_tagging
#    - Check for features requested by Joeri
1208
1209
1210
1211

#    - Check speed. Avoid recomputing unnecessarily
#    - Pass thresholds as parameters of the function. Provide defaults (we should tune them in the future)
#    - Check if attributes I'm asking for (eg arena) are already stored in Table_dict metadata