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
|
- import os
- import sys
- from copy import deepcopy
- from os.path import join as oj
- import mat4py
- import numpy as np
- import pandas as pd
- from matplotlib import pyplot as plt
- try:
- from skimage.external.tifffile import imread
- except:
- from skimage.io import imread
- from sklearn.model_selection import train_test_split
- pd.options.mode.chained_assignment = None # default='warn' - caution: this turns off setting with copy warning
- import pickle as pkl
- from viz import *
- import math
- import config
- import features
- import outcomes
- import load_tracking
- from tqdm import tqdm
- import train_reg
- def load_dfs_for_lstm(dsets=['clath_aux+gak_new'],
- splits=['test'],
- meta=None,
- length=40,
- normalize=True,
- lifetime_threshold=15,
- hotspots_threshold=25,
- filter_hotspots=True,
- filter_short=True,
- padding='end'):
- '''Loads dataframes preprocessed ready for LSTM
- '''
- dfs = {}
- for dset in tqdm(dsets):
- for split in splits:
- df = get_data(dset=dset)
- if filter_short:
- df = df[df.lifetime > lifetime_threshold] # only keep hard tracks
-
- if filter_hotspots:
- if hotspots_threshold is None:
- df = df[~df.hotspots]
- else:
- df = df[df.lifetime <= hotspots_threshold]
- df = df[df.cell_num.isin(config.DSETS[dset][split])] # select train/test etc.
- feat_names = ['X_same_length_normalized'] + select_final_feats(get_feature_names(df))
- # print('mid shape', df.shape)
- # downsample tracks
- df['X_same_length'] = [features.downsample(df.iloc[i]['X'],
- length, padding=padding)
- for i in range(len(df))] # downsampling
- df['X_same_length_extended'] = [features.downsample(df.iloc[i]['X_extended'],
- length, padding=padding)
- for i in range(len(df))] # downsampling
- # normalize tracks
- df = features.normalize_track(df, track='X_same_length', by_time_point=False)
- df = features.normalize_track(df, track='X_same_length_extended', by_time_point=False)
- # regression response
- df = outcomes.add_sig_mean(df, resp_tracks=['Y'])
- df = outcomes.add_aux_dyn_outcome(df)
- df['X_max_orig'] = deepcopy(df['X_max'].values)
- # remove extraneous feats
- # df = df[feat_names + meta]
- # df = df.dropna()
- # normalize features
- if normalize:
- for feat in feat_names:
- if 'X_same_length' not in feat:
- df = features.normalize_feature(df, feat)
- dfs[(dset, split)] = deepcopy(df)
- return dfs, feat_names
-
- def get_data(dset='clath_aux+gak_a7d2', use_processed=True, save_processed=True,
- processed_file=oj(config.DIR_PROCESSED, 'df.pkl'),
- metadata_file=oj(config.DIR_PROCESSED, 'metadata.pkl'),
- use_processed_dicts=True,
- compute_dictionary_learning=False,
- outcome_def='y_consec_thresh',
- pixel_data: bool=False,
- video_data: bool=False,
- acc_thresh=0.95,
- previous_meta_file: str=None):
- '''
- Params
- ------
- use_processed: bool, optional
- determines whether to load df from cached pkl
- save_processed: bool, optional
- if not using processed, determines whether to save the df
- use_processed_dicts: bool, optional
- if False, recalculate the dictionary features
- previous_meta_file: str, optional
- filename for metadata.pkl file saved by previous preprocessing
- the thresholds for lifetime are taken from this file
- '''
- # get things based onn dset
- DSET = config.DSETS[dset]
- LABELS = config.LABELS[dset]
- processed_file = processed_file[:-4] + '_' + dset + '.pkl'
- metadata_file = metadata_file[:-4] + '_' + dset + '.pkl'
- if use_processed and os.path.exists(processed_file):
- return pd.read_pickle(processed_file)
- else:
- print('loading + preprocessing data...')
- metadata = {}
-
-
- # load tracks
- print('\tloading tracks...')
- df = load_tracking.get_tracks(data_dir=DSET['data_dir'],
- split=DSET,
- pixel_data=pixel_data,
- video_data=video_data,
- dset=dset) # note: different Xs can be different shapes
- # df = df.fillna(df.median()) # this only does anything for the dynamin tracks, where x_pos is sometimes NaN
- # print('num nans', df.isna().sum())
- df['pid'] = np.arange(df.shape[0]) # assign each track a unique id
- df['valid'] = True # all tracks start as valid
-
- # set testing tracks to not valid
- if DSET['test'] is not None:
- df['valid'][df.cell_num.isin(DSET['test'])] = False
- metadata['num_tracks'] = df.valid.sum()
- # print('training', df.valid.sum())
-
- # preprocess data
- print('\tpreprocessing data...')
- df = remove_invalid_tracks(df) # use catIdx
- # print('valid', df.valid.sum())
- df = features.add_basic_features(df)
- df = outcomes.add_outcomes(df,
- LABELS=LABELS,
- vps_data='vps' in dset)
- metadata['num_tracks_valid'] = df.valid.sum()
- metadata['num_aux_pos_valid'] = df[df.valid][outcome_def].sum()
- metadata['num_hotspots_valid'] = df[df.valid]['hotspots'].sum()
- df['valid'][df.hotspots] = False
- df, meta_lifetime = process_tracks_by_lifetime(df, outcome_def=outcome_def,
- plot=False, acc_thresh=acc_thresh,
- previous_meta_file=previous_meta_file)
- df['valid'][df.short] = False
- df['valid'][df.long] = False
- metadata.update(meta_lifetime)
- metadata['num_tracks_hard'] = df['valid'].sum()
- metadata['num_aux_pos_hard'] = int(df[df.valid == 1][outcome_def].sum())
-
- # add features
- print('\tadding features...')
- df = features.add_dasc_features(df)
- if compute_dictionary_learning:
- df = features.add_dict_features(df, use_processed=use_processed_dicts)
- # df = features.add_smoothed_tracks(df)
- # df = features.add_pcs(df)
- # df = features.add_trend_filtering(df)
- # df = features.add_binary_features(df, outcome_def=outcome_def)
- if save_processed:
- print('\tsaving...')
- pkl.dump(metadata, open(metadata_file, 'wb'))
- df.to_pickle(processed_file)
- return df
- def get_snf_mt_vs_wt():
- """Need to merge the train and test data and split
- This is because they were split discretely in a way that using X_normalized gives an unfair hint to the algorithm
- Set outcome mt to whether or not the
- """
- df_fulls = []
- for i, dsets in enumerate([['vps4_snf7'], ['vps4_snf7___key=mt']]):
- dfs, feat_names = data.load_dfs_for_lstm(dsets=dsets,
- splits=['train', 'test'],
- filter_hotspots=True,
- filter_short=False,
- lifetime_threshold=None,
- hotspots_threshold=25,
- meta=['cell_num', 'Y_sig_mean', 'Y_sig_mean_normalized'],
- normalize=False)
- print('feat_names', feat_names)
- df_full = pd.concat(list(dfs.values()))
- df_full['mt'] = i
- df_fulls.append(df_full)
- df_full = pd.concat(df_fulls)
- df_train, df_test = train_test_split(df_full, random_state=13, test_size=0.33)
- return df_train, df_test, feat_names
- def remove_invalid_tracks(df, keep=[1, 2]):
- '''Remove certain types of tracks based on cat_idx.
- Only keep cat_idx = 1 and 2
- 1-4 (non-complex trajectory - no merges and splits)
- 1 - valid
- 2 - signal occasionally drops out
- 3 - cut - starts / ends
- 4 - multiple - at the same place (continues throughout)
- 5-8 (there is merging or splitting)
- '''
- return df[df.catIdx.isin(keep)]
- def process_tracks_by_lifetime(df: pd.DataFrame, outcome_def: str,
- plot=False, acc_thresh=0.95, previous_meta_file=None):
- '''Calculate accuracy you can get by just predicting max class
- as a func of lifetime and return points within proper lifetime (only looks at training cells)
- '''
- vals = df[df.valid == 1][['lifetime', outcome_def]]
- R, C = 1, 3
- lifetimes = np.unique(vals['lifetime'])
- # cumulative accuracy for different thresholds
- accs_cum_lower = np.array([1 - np.mean(vals[outcome_def][vals['lifetime'] <= l]) for l in lifetimes])
- accs_cum_higher = np.array([np.mean(vals[outcome_def][vals['lifetime'] >= l]) for l in lifetimes]).flatten()
- if previous_meta_file is None:
- try:
- idx_thresh = np.nonzero(accs_cum_lower >= acc_thresh)[0][-1] # last nonzero index
- thresh_lower = lifetimes[idx_thresh]
- except:
- idx_thresh = 0
- thresh_lower = lifetimes[idx_thresh] - 1
- try:
- idx_thresh_2 = np.nonzero(accs_cum_higher >= acc_thresh)[0][0]
- thresh_higher = lifetimes[idx_thresh_2]
- except:
- idx_thresh_2 = lifetimes.size - 1
- thresh_higher = lifetimes[idx_thresh_2] + 1
- else:
- previous_meta = pkl.load(open(previous_meta_file, 'rb'))
- thresh_lower = previous_meta['thresh_short']
- thresh_higher = previous_meta['thresh_long']
- # only df with lifetimes in proper range
- df['short'] = df['lifetime'] <= thresh_lower
- df['long'] = df['lifetime'] >= thresh_higher
- n = vals.shape[0]
- n_short = np.sum(df['short'])
- n_long = np.sum(df['long'])
- acc_short = 1 - np.mean(vals[outcome_def][vals['lifetime'] <= thresh_lower])
- acc_long = np.mean(vals[outcome_def][vals['lifetime'] >= thresh_higher])
- metadata = {'num_short': n_short, 'num_long': n_long, 'acc_short': acc_short,
- 'acc_long': acc_long, 'thresh_short': thresh_lower, 'thresh_long': thresh_higher}
- if plot:
- plt.figure(figsize=(12, 4), dpi=200)
- plt.subplot(R, C, 1)
- outcome = df[outcome_def]
- plt.hist(df['lifetime'][outcome == 1], label='aux+', alpha=1, color=cb, bins=25)
- plt.hist(df['lifetime'][outcome == 0], label='aux-', alpha=0.7, color=cr, bins=25)
- plt.xlabel('lifetime')
- plt.ylabel('count')
- plt.legend()
- plt.subplot(R, C, 2)
- plt.plot(lifetimes, accs_cum_lower, color=cr)
- # plt.axvline(thresh_lower)
- plt.axvspan(0, thresh_lower, alpha=0.2, color=cr)
- plt.ylabel('fraction of negative events')
- plt.xlabel(f'lifetime <= value\nshaded includes {n_short / n * 100:0.0f}% of pts')
- plt.subplot(R, C, 3)
- plt.plot(lifetimes, accs_cum_higher, cb)
- plt.axvspan(thresh_higher, max(lifetimes), alpha=0.2, color=cb)
- plt.ylabel('fraction of positive events')
- plt.xlabel(f'lifetime >= value\nshaded includes {n_long / n * 100:0.0f}% of pts')
- plt.tight_layout()
- return df, metadata
- def get_feature_names(df):
- '''Returns features (all of which are scalar)
- Removes metadata + time-series columns + outcomes
- '''
- ks = list(df.keys())
- feat_names = [
- k for k in ks
- if not k.startswith('y')
- and not k.startswith('Y')
- and not k.startswith('Z')
- and not k.startswith('pixel')
- # and not k.startswith('pc_')
- and not k in ['catIdx', 'cell_num', 'pid', 'valid', # metadata
- 'X', 'X_pvals', 'X_starts', 'X_ends', 'X_extended', # curves
- 'short', 'long', 'hotspots', 'sig_idxs', # should be weeded out
- 'X_max_around_Y_peak', 'X_max_after_Y_peak', # redudant with X_max / fall
- 'X_max_diff', 'X_peak_idx', # unlikely to be useful
- 'x_pos', 'z_pos', # curves
- 't', 'x_pos_seq', 'y_pos_seq', 'z_pos_seq', # curves
- 'X_smooth_spl', 'X_smooth_spl_dx', 'X_smooth_spl_d2x', # curves
- 'X_quantiles',
- ]
- ]
- return feat_names
- def select_final_feats(feat_names, binarize=False):
- feat_names = [x for x in feat_names
- if not x.startswith('sc_') # sparse coding
- and not x.startswith('nmf_') # nmf
- and not x in ['center_max', 'left_max', 'right_max', 'up_max', 'down_max',
- 'X_max_around_Y_peak', 'X_max_after_Y_peak', 'X_max_diff_after_Y_peak']
- and not x.startswith('pc_')
- and not 'extended' in x
- and not x == 'slope_end'
- and not '_tf_smooth' in x
- and not 'local' in x
- and not 'last' in x
- and not 'video' in x
- and not x == 'X_quantiles'
- # and not 'X_peak' in x
- # and not 'slope' in x
- # and not x in ['fall_final', 'fall_slope', 'fall_imp', 'fall']
- ]
- if binarize:
- feat_names = [x for x in feat_names if 'binary' in x]
- else:
- feat_names = [x for x in feat_names if not 'binary' in x]
- return feat_names
- if __name__ == '__main__':
-
- # process original data (and save out lifetime thresholds)
- # dset_orig = 'clath_aux+gak_a7d2'
- # df = get_data(dset=dset_orig) # save out orig
-
- # process new data (using lifetime thresholds from original data)
- outcome_def = 'y_consec_sig'
- for dset in ['vps4_snf7___key=mt', 'vps4_snf7']: # two keys from the same file
- # for dset in config.DSETS.keys():
- df = get_data(dset=dset, previous_meta_file=None, save_processed=True)
- # df = get_data(dset=dset, previous_meta_file=f'{config.DIR_PROCESSED}/metadata_{dset_orig}.pkl')
- print(dset, 'num cells', len(df['cell_num'].unique()), 'num tracks', df.shape[0], 'num aux+',
- df[outcome_def].sum(), 'aux+ fraction', (df[outcome_def].sum() / df.shape[0]).round(3),
- 'valid', df.valid.sum(), 'valid aux+', df[df.valid][outcome_def].sum(), 'valid aux+ fraction',
- (df[df.valid][outcome_def].sum() / df.valid.sum()).round(3))
|