-
Notifications
You must be signed in to change notification settings - Fork 0
/
bcdataset.py
261 lines (234 loc) · 10.4 KB
/
bcdataset.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
import glob
import numpy as np
import os
import random
import torch
from expelliarmus import Wizard
from torch.utils.data import Dataset
def save_dataset_to_raw_files(data_folder, chars, fidxs, dst_folder, expelliarmus=False):
"""Save dataset to raw files."""
os.makedirs(dst_folder, exist_ok=True)
if expelliarmus:
wizard = Wizard(encoding='evt2')
for ch in chars:
for fi in fidxs:
print(ch, fi, end='...\r')
path = f'{data_folder}/{ch}{fi}.npy'
d = np.load(path, allow_pickle=True)
# filter out bins with <1k events
d = d[[e.shape[0] >= 1_000 for e in d]]
for evarridx, evarr in enumerate(d):
if evarr['t'].max() - evarr['t'].min() > 1_000:
print(f'time too long: {evarr["t"].max() - evarr["t"].min()}, '
f'{evarr["t"].max()}, {evarr["t"].min()}')
if evarr is None:
print(f'evarr empty')
if expelliarmus:
new_path = os.path.join(dst_folder, f'{ch}{fi}_{evarridx}.raw')
wizard.save(fpath=new_path, arr=evarr)
else:
np.save(os.path.join(dst_folder, f'{ch}{fi}_{evarridx}.npy'), evarr)
print('done!')
def convert_ev_arr_to_tensor(ev_arr, timestep_us=1):
"""Convert event array (xypt compressed numpy) to tensor of shape (2*32*24, 1000)."""
# timestep is in us
n_timesteps = np.ceil(1_000 / timestep_us).astype(int)
tnsr = np.zeros((2, 32, 24, n_timesteps))
t_start = ev_arr['t'].min() // 1_000 * 1_000
assert ev_arr['t'].max() - t_start <= 1_000, 'event array too long'
tnsr[ev_arr['p'], ev_arr['x'], ev_arr['y'], (ev_arr['t'] - t_start) // timestep_us] = 1
tnsr = torch.from_numpy(tnsr.reshape(-1, n_timesteps)).float()
return tnsr
class BinCytometryDataset(Dataset):
TRAIN_SPLIT = 0.8
CHARS = "AB"
FIDXS = list(range(1, 5))
def __init__(self, data_folder, train=True, test_fi=None, temp_split=False, seed=42, ext='npy',
timestep_us=1):
"""Dataset for binary cytometry data. Each sample is a 2D tensor of shape (2, 32, 24, 1000).
This dataset has two modes:
1) if test_fi is None, all files are used for training and testing. If temp_split is True,
the dataset is split into train/test respecting temporal order. If not, randomly.
2) if test_fi is not None, all files except test_fi are used for training and test_fi is
used for testing. temp_split has no effect here.
Parameters
----------
data_folder : str
path of dataset root
train : bool, optional
train/test flag, by default True
test_fi: int, optional
index (1-4) of file to use for testing data (use all others for training), default: None
if None, load data from all files and then split into train/test
temp_split : bool, optional
split dataset into train/test respecting temporal order, by default False
if False, split dataset into train/test randomly
!!! only works if test_fi is None !!! # CODEX WROTE THIS
seed: int, optional
random seed, by default 42 (used for splitting dataset into train/test)
ext: str, optional
file extension in data_folder ('raw' for expelliarmus, 'npy' for numpy), default: 'npy'
"""
self.data_folder = data_folder
self.EXT = ext
self.n_total = len(glob.glob(f'{data_folder}/*.{self.EXT}'))
self.timestep_us = timestep_us
if 1000 // self.timestep_us != 1000 / self.timestep_us:
print('[WARNING] 1000/timestep_us not an integer - last time bin contains fewer events')
if self.EXT == 'raw':
print('[WARNING] using expelliarmus is unstable [WARNING]')
self.wizard = Wizard(encoding='evt2')
raise NotImplementedError('using expelliarmus is unstable')
# temporary list of all possible combinations of char and fidx
chfi_list = [f'{ch}{fi}' for ch in self.CHARS for fi in self.FIDXS]
############################
# if test_fi is given
if test_fi is not None:
chfi_to_n_total = {
chfi: len(glob.glob(f'{data_folder}/{chfi}_*.{self.EXT}'))
for chfi in chfi_list
}
if train:
self.sample_filenames = [
f'{data_folder}/{ch}{fi}_{idx}.{self.EXT}'
for ch in self.CHARS for fi in self.FIDXS
for idx in range(chfi_to_n_total[f'{ch}{fi}'])
if fi != test_fi
]
else:
# test_filenames = glob.glob(f'{data_folder}/?{test_fi}_*.{self.EXT}')
self.sample_filenames = [
f'{data_folder}/{ch}{fi}_{idx}.{self.EXT}'
for ch in self.CHARS for fi in self.FIDXS
for idx in range(chfi_to_n_total[f'{ch}{fi}'])
if fi == test_fi
]
return
############################
# if test_fi is *not* given
print('no test_fi given...')
# map from id (e.g. A1) to index list of dataset
idx_map = {
chfi: list(range(len(glob.glob(f'{data_folder}/{chfi}_*.{self.EXT}'))))
for chfi in chfi_list
}
if not temp_split:
# shuffle indices
for chfi in chfi_list:
random.seed(seed)
random.shuffle(idx_map[chfi])
# map from id (e.g. A1) to number of samples in train/test
chfi_to_n_train = {
chfi: int(self.TRAIN_SPLIT * len(glob.glob(f'{data_folder}/{chfi}_*.{self.EXT}')))
for chfi in chfi_list
}
chfi_to_n_test = {
chfi: len(glob.glob(f'{data_folder}/{chfi}_*.{self.EXT}')) - chfi_to_n_train[chfi]
for chfi in chfi_list
}
all_filenames = {
chfi: [f'{data_folder}/{chfi}_{idx}.{self.EXT}' for idx in idx_map[chfi]]
for chfi in chfi_list
}
if train:
self.sample_filenames = [
all_filenames[chfi][idx]
for chfi in chfi_list
for idx in range(chfi_to_n_train[chfi])
]
else:
self.sample_filenames = [
all_filenames[chfi][idx]
for chfi in chfi_list
for idx in range(
chfi_to_n_train[chfi],
chfi_to_n_train[chfi] + chfi_to_n_test[chfi]
)
]
def __len__(self):
"""Returns the number of samples in the dataset."""
return len(self.sample_filenames)
def __getitem__(self, idx, return_trial=False):
"""Returns a sample from the dataset.
data: tensor of shape (2, 32, 24, 1000)
label: int with value 0 or 1
"""
sample_filename = self.sample_filenames[idx]
label = int(os.path.basename(sample_filename)[0] == 'B')
trial = int(os.path.basename(sample_filename)[1])
if self.EXT == 'raw':
ev_arr = self.wizard.read(sample_filename)
else:
ev_arr = np.load(sample_filename)
assert ev_arr is not None, f'event array is None: {sample_filename}'
data = convert_ev_arr_to_tensor(ev_arr, self.timestep_us)
assert data is not None, f'event array is longer than 1000 time steps? {sample_filename}'
if return_trial:
return data, label, trial
else:
return data, label
def get_datasets(data_folder, seed=42, temp_split=False, run_checks=True):
"""Get train and test datasets."""
print('getting datasets...', end='\r')
train_dataset = BinCytometryDataset(
data_folder=data_folder,
train=True,
temp_split=temp_split,
seed=seed
)
test_dataset = BinCytometryDataset(
data_folder=data_folder,
train=False,
temp_split=temp_split,
seed=seed
)
print('running checks...', end='\r')
if run_checks:
# check that train/test split is correct
for e in train_dataset.sample_filenames:
assert e not in test_dataset.sample_filenames, f'train/test split overlap: {e}'
assert train_dataset.n_total == test_dataset.n_total
ntotal = train_dataset.n_total
assert len(train_dataset) + len(test_dataset) == ntotal
assert len(train_dataset.sample_filenames) + len(test_dataset.sample_filenames) == ntotal
assert abs(len(train_dataset) - int(ntotal * train_dataset.TRAIN_SPLIT)) < 10
assert abs(len(test_dataset) - int(ntotal * (1 - train_dataset.TRAIN_SPLIT))) < 10
# return datasets
print('successfully loaded train and test datasets')
return train_dataset, test_dataset
def get_datasets_fold(data_folder, test_file_idx, run_checks=True, seed=42, timestep_us=1):
print('getting datasets...', end='\r')
ds_tr = BinCytometryDataset(
data_folder=data_folder,
train=True,
test_fi=test_file_idx,
seed=seed,
timestep_us=timestep_us
)
ds_te = BinCytometryDataset(
data_folder=data_folder,
train=False,
test_fi=test_file_idx,
seed=seed,
timestep_us=timestep_us
)
# check that train/test split is correct
if run_checks:
print('running checks...', end='\r')
for e in ds_tr.sample_filenames:
assert e not in ds_te.sample_filenames, f'train/test split overlap: {e}'
test_filenames = glob.glob(f'{data_folder}/?{test_file_idx}_*.npy')
train_filenames = [e for idx in range(1, 5) if idx != test_file_idx for e in glob.glob(f'{data_folder}/?{idx}_*.npy')]
assert len(train_filenames) == len(ds_tr.sample_filenames)
assert len(test_filenames) == len(ds_te.sample_filenames)
assert len(train_filenames) == len(set(train_filenames) & set(ds_tr.sample_filenames))
assert len(test_filenames) == len(set(test_filenames) & set(ds_te.sample_filenames))
assert ds_tr.n_total == ds_te.n_total
ntotal = ds_tr.n_total
assert len(ds_tr) + len(ds_te) == ntotal
assert len(ds_tr.sample_filenames) + len(ds_te.sample_filenames) == ntotal
# assert right number of timestep
assert ds_tr[0][0].shape[-1] == np.ceil(1000 / timestep_us).astype(int)
# return datasets
print('successfully loaded train and test datasets')
return ds_tr, ds_te