DARTsort 0.3.5__py3-none-any.whl
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- dartsort/__init__.py +68 -0
- dartsort/cli.py +105 -0
- dartsort/clustering/__init__.py +3 -0
- dartsort/clustering/cluster_util.py +392 -0
- dartsort/clustering/clustering.py +765 -0
- dartsort/clustering/clustering_features.py +202 -0
- dartsort/clustering/density.py +1299 -0
- dartsort/clustering/forward_backward.py +219 -0
- dartsort/clustering/gmm/mixture.py +5790 -0
- dartsort/clustering/gmm/stable_features.py +993 -0
- dartsort/clustering/kmeans.py +448 -0
- dartsort/clustering/merge.py +501 -0
- dartsort/clustering/modes.py +214 -0
- dartsort/clustering/refine_util.py +191 -0
- dartsort/clustering/relocate.py +109 -0
- dartsort/clustering/split.py +1285 -0
- dartsort/config.py +288 -0
- dartsort/detect/__init__.py +2 -0
- dartsort/detect/detect.py +449 -0
- dartsort/detect/detect_filters.py +37 -0
- dartsort/evaluate/__init__.py +1 -0
- dartsort/evaluate/analysis.py +594 -0
- dartsort/evaluate/comparison.py +534 -0
- dartsort/evaluate/hybrid_util.py +599 -0
- dartsort/evaluate/noise_recording_tools.py +184 -0
- dartsort/evaluate/sim_template_tools.py +842 -0
- dartsort/evaluate/simkit.py +840 -0
- dartsort/evaluate/simlib.py +270 -0
- dartsort/localize/__init__.py +0 -0
- dartsort/localize/localize_torch.py +291 -0
- dartsort/localize/localize_util.py +313 -0
- dartsort/main.py +657 -0
- dartsort/peel/__init__.py +6 -0
- dartsort/peel/grab.py +208 -0
- dartsort/peel/matching.py +628 -0
- dartsort/peel/matching_util/__init__.py +3 -0
- dartsort/peel/matching_util/compressed_upsampled.py +718 -0
- dartsort/peel/matching_util/drifty.py +791 -0
- dartsort/peel/matching_util/matching_base.py +616 -0
- dartsort/peel/matching_util/pairwise.py +359 -0
- dartsort/peel/matching_util/pairwise_util.py +1317 -0
- dartsort/peel/peel_base.py +1000 -0
- dartsort/peel/running_template.py +1498 -0
- dartsort/peel/subtract.py +1123 -0
- dartsort/peel/threshold.py +471 -0
- dartsort/peel/universal.py +191 -0
- dartsort/pretrained/default_temporal_kernel.npy +0 -0
- dartsort/pretrained/detect_np1.pt +0 -0
- dartsort/pretrained/detect_np2.pt +0 -0
- dartsort/pretrained/glasbey1024.npz +0 -0
- dartsort/pretrained/offset_trained_denoiser.pt +0 -0
- dartsort/pretrained/offset_trained_denoiser_added_temp.pt +0 -0
- dartsort/pretrained/offset_trained_denoiser_added_temp_boost.pt +0 -0
- dartsort/pretrained/original_single_chan_denoiser.pt +0 -0
- dartsort/pretrained/single_chan_denoiser.pt +0 -0
- dartsort/templates/__init__.py +11 -0
- dartsort/templates/get_templates.py +771 -0
- dartsort/templates/postprocess_util.py +555 -0
- dartsort/templates/realignment.py +466 -0
- dartsort/templates/superres_util.py +204 -0
- dartsort/templates/template_util.py +518 -0
- dartsort/templates/templates.py +528 -0
- dartsort/transform/__init__.py +2 -0
- dartsort/transform/_multichan_denoiser_kit.py +650 -0
- dartsort/transform/all_transformers.py +58 -0
- dartsort/transform/amortized_localization.py +503 -0
- dartsort/transform/amplitudes.py +197 -0
- dartsort/transform/decollider.py +635 -0
- dartsort/transform/enforce_decrease.py +158 -0
- dartsort/transform/localize.py +86 -0
- dartsort/transform/matching_denoiser.py +44 -0
- dartsort/transform/pipeline.py +391 -0
- dartsort/transform/single_channel_denoiser.py +153 -0
- dartsort/transform/single_channel_templates.py +98 -0
- dartsort/transform/supervised_denoiser.py +486 -0
- dartsort/transform/temporal_pca.py +413 -0
- dartsort/transform/transform_base.py +255 -0
- dartsort/util/__init__.py +0 -0
- dartsort/util/cli_util.py +179 -0
- dartsort/util/data_util.py +1128 -0
- dartsort/util/drift_util.py +1054 -0
- dartsort/util/internal_config.py +1074 -0
- dartsort/util/interpolation_util.py +1012 -0
- dartsort/util/job_util.py +19 -0
- dartsort/util/logging_util.py +91 -0
- dartsort/util/main_util.py +235 -0
- dartsort/util/more_operators.py +269 -0
- dartsort/util/multiprocessing_util.py +211 -0
- dartsort/util/nn_util.py +461 -0
- dartsort/util/noise_util.py +1461 -0
- dartsort/util/peel_util.py +223 -0
- dartsort/util/py_util.py +213 -0
- dartsort/util/registration_util.py +128 -0
- dartsort/util/residual_util.py +30 -0
- dartsort/util/sparse_util.py +81 -0
- dartsort/util/spikeio.py +227 -0
- dartsort/util/spiketorch.py +1198 -0
- dartsort/util/testing_util/__init__.py +0 -0
- dartsort/util/testing_util/matching_debug_util.py +517 -0
- dartsort/util/testing_util/mixture_testing_util.py +291 -0
- dartsort/util/torch_optimization_util.py +156 -0
- dartsort/util/torch_util.py +75 -0
- dartsort/util/universal_util.py +269 -0
- dartsort/util/waveform_util.py +809 -0
- dartsort/vis/__init__.py +19 -0
- dartsort/vis/analysis_plots.py +392 -0
- dartsort/vis/colors.py +36 -0
- dartsort/vis/gt.py +392 -0
- dartsort/vis/layout.py +149 -0
- dartsort/vis/metrics_util.py +23 -0
- dartsort/vis/mixture.py +1489 -0
- dartsort/vis/over_time.py +514 -0
- dartsort/vis/recanim.py +429 -0
- dartsort/vis/scatterplots.py +779 -0
- dartsort/vis/sorting.py +163 -0
- dartsort/vis/unit.py +917 -0
- dartsort/vis/unit_comparison.py +1017 -0
- dartsort/vis/versus.py +300 -0
- dartsort/vis/vismain.py +537 -0
- dartsort/vis/waveforms.py +357 -0
- dartsort-0.3.5.dist-info/METADATA +93 -0
- dartsort-0.3.5.dist-info/RECORD +125 -0
- dartsort-0.3.5.dist-info/WHEEL +5 -0
- dartsort-0.3.5.dist-info/entry_points.txt +2 -0
- dartsort-0.3.5.dist-info/top_level.txt +1 -0
|
@@ -0,0 +1,635 @@
|
|
|
1
|
+
import dataclasses
|
|
2
|
+
from typing import cast
|
|
3
|
+
|
|
4
|
+
import numpy as np
|
|
5
|
+
import pandas as pd
|
|
6
|
+
import torch
|
|
7
|
+
import torch.nn.functional as F
|
|
8
|
+
from torch.utils.data import (
|
|
9
|
+
DataLoader,
|
|
10
|
+
Dataset,
|
|
11
|
+
Sampler,
|
|
12
|
+
StackDataset,
|
|
13
|
+
TensorDataset,
|
|
14
|
+
)
|
|
15
|
+
from tqdm.auto import trange
|
|
16
|
+
|
|
17
|
+
from ..util.spiketorch import reindex, spawn_torch_rg
|
|
18
|
+
from ..util.logging_util import get_logger
|
|
19
|
+
from ._multichan_denoiser_kit import (
|
|
20
|
+
BaseMultichannelDenoiser,
|
|
21
|
+
get_noise,
|
|
22
|
+
AsyncSameChannelRecordingNoiseDataset,
|
|
23
|
+
AOTIndicesWeightedRandomBatchSampler,
|
|
24
|
+
NoneDataset,
|
|
25
|
+
)
|
|
26
|
+
|
|
27
|
+
|
|
28
|
+
logger = get_logger(__name__)
|
|
29
|
+
|
|
30
|
+
|
|
31
|
+
class Decollider(BaseMultichannelDenoiser):
|
|
32
|
+
default_name = "decollider"
|
|
33
|
+
|
|
34
|
+
def __init__(
|
|
35
|
+
self,
|
|
36
|
+
channel_index,
|
|
37
|
+
geom,
|
|
38
|
+
hidden_dims=(1024, 1024),
|
|
39
|
+
norm_kind="none",
|
|
40
|
+
name=None,
|
|
41
|
+
name_prefix="",
|
|
42
|
+
batch_size=256,
|
|
43
|
+
learning_rate=3e-4,
|
|
44
|
+
weight_decay=0.0,
|
|
45
|
+
n_epochs=75,
|
|
46
|
+
pad_depth_only=True,
|
|
47
|
+
channelwise_dropout_p=0.0,
|
|
48
|
+
with_conv_fullheight=False,
|
|
49
|
+
val_split_p=0.0,
|
|
50
|
+
min_epochs=10,
|
|
51
|
+
earlystop_eps=None,
|
|
52
|
+
random_seed=0,
|
|
53
|
+
res_type="none",
|
|
54
|
+
lr_schedule="CosineAnnealingLR",
|
|
55
|
+
lr_schedule_kwargs=None,
|
|
56
|
+
inference_batch_size=1024,
|
|
57
|
+
optimizer="Adam",
|
|
58
|
+
optimizer_kwargs=None,
|
|
59
|
+
nonlinearity="PReLU",
|
|
60
|
+
scaling="max",
|
|
61
|
+
signal_gates=True,
|
|
62
|
+
step_callback=None,
|
|
63
|
+
# my args. todo: port over common ones.
|
|
64
|
+
epoch_size=200 * 256,
|
|
65
|
+
inference_z_samples=10,
|
|
66
|
+
detach_amortizer=True,
|
|
67
|
+
exz_estimator="n3n",
|
|
68
|
+
inference_kind="amortized",
|
|
69
|
+
eyz_res_type="none",
|
|
70
|
+
e_exz_y_res_type="none",
|
|
71
|
+
emz_res_type="none",
|
|
72
|
+
l4_alpha=0,
|
|
73
|
+
l1_alpha=0,
|
|
74
|
+
output_l1_alpha=1e-5,
|
|
75
|
+
cycle_loss_alpha=1.0,
|
|
76
|
+
separate_cycle_net=False,
|
|
77
|
+
detach_cycle_loss=False,
|
|
78
|
+
val_noise_random_seed=0,
|
|
79
|
+
inf_net_hidden_dims=None,
|
|
80
|
+
eyz_net_hidden_dims=None,
|
|
81
|
+
queue_chunks=32,
|
|
82
|
+
):
|
|
83
|
+
assert exz_estimator in ("n2n", "2n2", "n3n", "3n3")
|
|
84
|
+
assert inference_kind in ("raw", "exz", "exz_fromz", "amortized", "exy_fake")
|
|
85
|
+
|
|
86
|
+
super().__init__(
|
|
87
|
+
geom=geom,
|
|
88
|
+
channel_index=channel_index,
|
|
89
|
+
name=name,
|
|
90
|
+
name_prefix=name_prefix,
|
|
91
|
+
hidden_dims=hidden_dims,
|
|
92
|
+
norm_kind=norm_kind,
|
|
93
|
+
batch_size=batch_size,
|
|
94
|
+
learning_rate=learning_rate,
|
|
95
|
+
weight_decay=weight_decay,
|
|
96
|
+
n_epochs=n_epochs,
|
|
97
|
+
pad_depth_only=pad_depth_only,
|
|
98
|
+
channelwise_dropout_p=channelwise_dropout_p,
|
|
99
|
+
with_conv_fullheight=with_conv_fullheight,
|
|
100
|
+
val_split_p=val_split_p,
|
|
101
|
+
min_epochs=min_epochs,
|
|
102
|
+
earlystop_eps=earlystop_eps,
|
|
103
|
+
random_seed=random_seed,
|
|
104
|
+
res_type=res_type,
|
|
105
|
+
lr_schedule=lr_schedule,
|
|
106
|
+
lr_schedule_kwargs=lr_schedule_kwargs,
|
|
107
|
+
inference_batch_size=inference_batch_size,
|
|
108
|
+
optimizer=optimizer,
|
|
109
|
+
optimizer_kwargs=optimizer_kwargs,
|
|
110
|
+
nonlinearity=nonlinearity,
|
|
111
|
+
scaling=scaling,
|
|
112
|
+
signal_gates=signal_gates,
|
|
113
|
+
step_callback=step_callback,
|
|
114
|
+
)
|
|
115
|
+
|
|
116
|
+
self.epoch_size = epoch_size
|
|
117
|
+
self.queue_chunks = queue_chunks
|
|
118
|
+
|
|
119
|
+
self.inference_z_samples = inference_z_samples
|
|
120
|
+
self.detach_amortizer = detach_amortizer
|
|
121
|
+
self.exz_estimator = exz_estimator
|
|
122
|
+
self.inference_kind = inference_kind
|
|
123
|
+
self.eyz_res_type = eyz_res_type
|
|
124
|
+
self.e_exz_y_res_type = e_exz_y_res_type
|
|
125
|
+
self.emz_res_type = emz_res_type
|
|
126
|
+
self.val_noise_random_seed = val_noise_random_seed
|
|
127
|
+
self.inf_net_hidden_dims = inf_net_hidden_dims
|
|
128
|
+
self.eyz_net_hidden_dims = eyz_net_hidden_dims
|
|
129
|
+
self.l1_alpha = l1_alpha
|
|
130
|
+
self.l4_alpha = l4_alpha
|
|
131
|
+
self.output_l1_alpha = output_l1_alpha
|
|
132
|
+
self.cycle_loss_alpha = cycle_loss_alpha
|
|
133
|
+
self.separate_cycle_net = separate_cycle_net
|
|
134
|
+
self.detach_cycle_loss = detach_cycle_loss
|
|
135
|
+
|
|
136
|
+
if separate_cycle_net:
|
|
137
|
+
assert cycle_loss_alpha > 0
|
|
138
|
+
|
|
139
|
+
def initialize_spike_length_dependent_params(self):
|
|
140
|
+
if hasattr(self, "inf_net"):
|
|
141
|
+
logger.dartsortdebug("Already initialized.")
|
|
142
|
+
return
|
|
143
|
+
self.initialize_shapes()
|
|
144
|
+
if self.exz_estimator in ("n2n", "n3n"):
|
|
145
|
+
self.eyz = self.get_mlp(
|
|
146
|
+
res_type=self.eyz_res_type, hidden_dims=self.eyz_net_hidden_dims
|
|
147
|
+
)
|
|
148
|
+
if self.exz_estimator in ("n3n", "2n2", "3n3"):
|
|
149
|
+
self.emz = self.get_mlp(res_type=self.emz_res_type, output_layer="linear")
|
|
150
|
+
if self.inference_kind == "amortized":
|
|
151
|
+
self.inf_net = self.get_mlp(
|
|
152
|
+
res_type=self.e_exz_y_res_type, hidden_dims=self.inf_net_hidden_dims
|
|
153
|
+
)
|
|
154
|
+
if self.separate_cycle_net:
|
|
155
|
+
self.den_net = self.get_mlp(
|
|
156
|
+
res_type=self.e_exz_y_res_type, hidden_dims=self.inf_net_hidden_dims
|
|
157
|
+
)
|
|
158
|
+
else:
|
|
159
|
+
self.den_net = self.inf_net
|
|
160
|
+
self.to(self.device)
|
|
161
|
+
|
|
162
|
+
def fit(self, recording, waveforms, *, channels, weights=None, **unused):
|
|
163
|
+
super().fit(recording, waveforms, channels=channels, weights=weights)
|
|
164
|
+
train_data, val_data = self._construct_datasets_from_waveforms(
|
|
165
|
+
waveforms, channels, recording, weights
|
|
166
|
+
)
|
|
167
|
+
with torch.enable_grad():
|
|
168
|
+
res = self._fit(train_data, val_data)
|
|
169
|
+
self._needs_fit = False
|
|
170
|
+
return res
|
|
171
|
+
|
|
172
|
+
def forward_unbatched(self, waveforms, channels):
|
|
173
|
+
"""Called only at inference time."""
|
|
174
|
+
# TODO: batch all of this.
|
|
175
|
+
waveforms, masks = self.to_nn_channels(waveforms, channels)
|
|
176
|
+
net_input = waveforms, masks.unsqueeze(1)
|
|
177
|
+
|
|
178
|
+
if self.inference_kind == "amortized":
|
|
179
|
+
pred = self.den_net(net_input)
|
|
180
|
+
elif self.inference_kind == "raw":
|
|
181
|
+
if hasattr(self, "emz"):
|
|
182
|
+
emz = self.emz(net_input)
|
|
183
|
+
pred = waveforms - emz
|
|
184
|
+
elif hasattr(self, "eyz"):
|
|
185
|
+
pred = self.eyz(net_input)
|
|
186
|
+
else:
|
|
187
|
+
assert False
|
|
188
|
+
elif self.inference_kind == "exz_fromz":
|
|
189
|
+
pred = torch.zeros_like(waveforms)
|
|
190
|
+
for j in range(self.inference_z_samples):
|
|
191
|
+
m = get_noise(
|
|
192
|
+
self.recording,
|
|
193
|
+
channels.numpy(force=True),
|
|
194
|
+
self.b.model_channel_index.numpy(force=True),
|
|
195
|
+
spike_length_samples=cast(int, self.spike_length_samples),
|
|
196
|
+
rg=None,
|
|
197
|
+
)
|
|
198
|
+
m = m.to(waveforms)
|
|
199
|
+
z = waveforms + m
|
|
200
|
+
net_input = z, masks.unsqueeze(1)
|
|
201
|
+
if self.exz_estimator == "n2n":
|
|
202
|
+
eyz = self.eyz(net_input)
|
|
203
|
+
pred += 2 * eyz - z
|
|
204
|
+
elif self.exz_estimator == "2n2":
|
|
205
|
+
emz = self.emz(net_input)
|
|
206
|
+
pred += z - 2 * emz
|
|
207
|
+
elif self.exz_estimator == "n3n":
|
|
208
|
+
eyz = self.eyz(net_input)
|
|
209
|
+
emz = self.emz(net_input)
|
|
210
|
+
pred += eyz - emz
|
|
211
|
+
elif self.exz_estimator == "3n3":
|
|
212
|
+
emz = self.emz(net_input)
|
|
213
|
+
pred += z - emz
|
|
214
|
+
else:
|
|
215
|
+
assert False
|
|
216
|
+
pred /= self.inference_z_samples
|
|
217
|
+
elif self.inference_kind == "exz":
|
|
218
|
+
if self.exz_estimator == "n2n":
|
|
219
|
+
eyz = self.eyz(net_input)
|
|
220
|
+
pred = 2 * eyz - waveforms
|
|
221
|
+
elif self.exz_estimator == "2n2":
|
|
222
|
+
emz = self.emz(net_input)
|
|
223
|
+
pred = waveforms - 2 * emz
|
|
224
|
+
elif self.exz_estimator == "n3n":
|
|
225
|
+
eyz = self.eyz(net_input)
|
|
226
|
+
emz = self.emz(net_input)
|
|
227
|
+
pred = eyz - emz
|
|
228
|
+
elif self.exz_estimator == "3n3":
|
|
229
|
+
emz = self.emz(net_input)
|
|
230
|
+
pred = waveforms - emz
|
|
231
|
+
else:
|
|
232
|
+
assert False
|
|
233
|
+
else:
|
|
234
|
+
assert False
|
|
235
|
+
|
|
236
|
+
pred = self.to_orig_channels(pred, channels)
|
|
237
|
+
|
|
238
|
+
return pred
|
|
239
|
+
|
|
240
|
+
def train_forward(self, y, m, ell, mask):
|
|
241
|
+
z = y + m
|
|
242
|
+
|
|
243
|
+
# predictions given z
|
|
244
|
+
# TODO: variance given z and put it in the loss
|
|
245
|
+
exz = eyz = emz = e_exz_y = cycle_output = None
|
|
246
|
+
net_input = z, mask.unsqueeze(1)
|
|
247
|
+
if self.exz_estimator == "n2n":
|
|
248
|
+
eyz = self.eyz(net_input)
|
|
249
|
+
exz = 2 * eyz - z
|
|
250
|
+
elif self.exz_estimator == "2n2":
|
|
251
|
+
emz = self.emz(net_input)
|
|
252
|
+
exz = z - 2 * emz
|
|
253
|
+
elif self.exz_estimator == "n3n":
|
|
254
|
+
eyz = self.eyz(net_input)
|
|
255
|
+
emz = self.emz(net_input)
|
|
256
|
+
exz = eyz - emz
|
|
257
|
+
elif self.exz_estimator == "3n3":
|
|
258
|
+
emz = self.emz(net_input)
|
|
259
|
+
exz = y - emz
|
|
260
|
+
else:
|
|
261
|
+
assert False
|
|
262
|
+
|
|
263
|
+
# predictions given y, if relevant
|
|
264
|
+
if self.inference_kind == "amortized":
|
|
265
|
+
e_exz_y = self.inf_net((y, mask.unsqueeze(1)))
|
|
266
|
+
|
|
267
|
+
if self.cycle_loss_alpha:
|
|
268
|
+
assert e_exz_y is not None
|
|
269
|
+
cycle_targ = e_exz_y.detach() if self.detach_cycle_loss else e_exz_y
|
|
270
|
+
cycle_input = cycle_targ + ell
|
|
271
|
+
cycle_output = self.den_net((cycle_input, mask.unsqueeze(1)))
|
|
272
|
+
|
|
273
|
+
return dict(
|
|
274
|
+
exz=exz,
|
|
275
|
+
eyz=eyz,
|
|
276
|
+
emz=emz,
|
|
277
|
+
e_exz_y=e_exz_y,
|
|
278
|
+
cycle_output=cycle_output,
|
|
279
|
+
)
|
|
280
|
+
|
|
281
|
+
def loss(
|
|
282
|
+
self,
|
|
283
|
+
mask,
|
|
284
|
+
waveforms,
|
|
285
|
+
m,
|
|
286
|
+
net_outputs,
|
|
287
|
+
l1_alpha=None,
|
|
288
|
+
l4_alpha=None,
|
|
289
|
+
output_l1_alpha=None,
|
|
290
|
+
):
|
|
291
|
+
loss_dict = {}
|
|
292
|
+
mask = mask.unsqueeze(1)
|
|
293
|
+
|
|
294
|
+
exz = net_outputs["exz"]
|
|
295
|
+
eyz = net_outputs["eyz"]
|
|
296
|
+
emz = net_outputs["emz"]
|
|
297
|
+
e_exz_y = net_outputs["e_exz_y"]
|
|
298
|
+
cycle_output = net_outputs["cycle_output"]
|
|
299
|
+
|
|
300
|
+
if eyz is not None:
|
|
301
|
+
eyz_mask = mask * eyz
|
|
302
|
+
loss_dict["eyz"] = F.mse_loss(eyz_mask, mask * waveforms)
|
|
303
|
+
if l1_alpha:
|
|
304
|
+
loss_dict["eyz_l1"] = (
|
|
305
|
+
l4_alpha * (eyz - waveforms).mul_(mask).abs_().mean()
|
|
306
|
+
)
|
|
307
|
+
if l4_alpha:
|
|
308
|
+
loss_dict["eyz_l4"] = (
|
|
309
|
+
l4_alpha * ((eyz - waveforms).mul_(mask) ** 4).mean()
|
|
310
|
+
)
|
|
311
|
+
if output_l1_alpha:
|
|
312
|
+
loss_dict["eyz_ol1"] = output_l1_alpha * eyz_mask.abs().mean()
|
|
313
|
+
if emz is not None:
|
|
314
|
+
loss_dict["emz"] = F.mse_loss(mask * emz, mask * m)
|
|
315
|
+
if l1_alpha:
|
|
316
|
+
loss_dict["emz_l1"] = l4_alpha * (emz - m).mul_(mask).abs_().mean()
|
|
317
|
+
if l4_alpha:
|
|
318
|
+
loss_dict["emz_l4"] = l4_alpha * ((emz - m).mul_(mask) ** 4).mean()
|
|
319
|
+
if e_exz_y is not None:
|
|
320
|
+
to_amortize = exz
|
|
321
|
+
if self.detach_amortizer:
|
|
322
|
+
# should amortize-ability affect the learning of eyz, emz?
|
|
323
|
+
to_amortize = to_amortize.detach()
|
|
324
|
+
am_mask = mask * to_amortize
|
|
325
|
+
loss_dict["e_exz_y"] = F.mse_loss(am_mask, mask * e_exz_y)
|
|
326
|
+
if l1_alpha:
|
|
327
|
+
loss_dict["e_exz_y_l1"] = (
|
|
328
|
+
l4_alpha * (to_amortize - e_exz_y).mul_(mask).abs_().mean()
|
|
329
|
+
)
|
|
330
|
+
if l4_alpha:
|
|
331
|
+
loss_dict["e_exz_y_l4"] = (
|
|
332
|
+
l4_alpha * ((to_amortize - e_exz_y).mul_(mask) ** 4).mean()
|
|
333
|
+
)
|
|
334
|
+
if output_l1_alpha:
|
|
335
|
+
loss_dict["e_exz_y_ol1"] = output_l1_alpha * am_mask.abs().mean()
|
|
336
|
+
if cycle_output is not None:
|
|
337
|
+
coef = 1 if self.separate_cycle_net else self.cycle_loss_alpha
|
|
338
|
+
cycle_targ = e_exz_y.detach() if self.detach_cycle_loss else e_exz_y
|
|
339
|
+
loss_dict["cycle"] = coef * F.mse_loss(
|
|
340
|
+
mask * cycle_targ, mask * cycle_output
|
|
341
|
+
)
|
|
342
|
+
if l1_alpha:
|
|
343
|
+
loss_dict["cycle_l1"] = (coef * l1_alpha) * (
|
|
344
|
+
(cycle_targ - cycle_output).mul_(mask).abs_().mean()
|
|
345
|
+
)
|
|
346
|
+
return loss_dict
|
|
347
|
+
|
|
348
|
+
def _fit(
|
|
349
|
+
self,
|
|
350
|
+
train_data: "DecolliderDataLoader",
|
|
351
|
+
val_data: "DecolliderDataLoader | None",
|
|
352
|
+
):
|
|
353
|
+
optimizer = self.get_optimizer()
|
|
354
|
+
scheduler = self.get_scheduler(optimizer)
|
|
355
|
+
|
|
356
|
+
last_val_loss = None
|
|
357
|
+
train_records = []
|
|
358
|
+
|
|
359
|
+
with trange(self.n_epochs, desc="Epochs", unit="epoch") as pbar:
|
|
360
|
+
for epoch in pbar:
|
|
361
|
+
# deal with random indices...
|
|
362
|
+
train_data.refresh()
|
|
363
|
+
|
|
364
|
+
# Training phase
|
|
365
|
+
self.train()
|
|
366
|
+
train_losses = {}
|
|
367
|
+
for (
|
|
368
|
+
waveform_batch,
|
|
369
|
+
channels_batch,
|
|
370
|
+
noise_batch,
|
|
371
|
+
cnoise_batch,
|
|
372
|
+
) in train_data:
|
|
373
|
+
waveform_batch = waveform_batch.to(self.device)
|
|
374
|
+
channels_batch = channels_batch.to(self.device)
|
|
375
|
+
waveform_batch = reindex(
|
|
376
|
+
channels_batch,
|
|
377
|
+
waveform_batch,
|
|
378
|
+
self.relative_index,
|
|
379
|
+
pad_value=0.0,
|
|
380
|
+
)
|
|
381
|
+
|
|
382
|
+
optimizer.zero_grad()
|
|
383
|
+
m = noise_batch.to(waveform_batch)
|
|
384
|
+
ell = None
|
|
385
|
+
if cnoise_batch is not None:
|
|
386
|
+
ell = cnoise_batch.to(waveform_batch)
|
|
387
|
+
mask = self.get_masks(channels_batch).to(waveform_batch)
|
|
388
|
+
fres = self.train_forward(waveform_batch, m, ell, mask)
|
|
389
|
+
loss_dict = self.loss(
|
|
390
|
+
mask,
|
|
391
|
+
waveform_batch,
|
|
392
|
+
m,
|
|
393
|
+
fres,
|
|
394
|
+
l1_alpha=self.l1_alpha,
|
|
395
|
+
l4_alpha=self.l4_alpha,
|
|
396
|
+
output_l1_alpha=self.output_l1_alpha,
|
|
397
|
+
)
|
|
398
|
+
loss = sum(loss_dict.values())
|
|
399
|
+
loss.backward() # type: ignore
|
|
400
|
+
optimizer.step()
|
|
401
|
+
|
|
402
|
+
for k, v in loss_dict.items():
|
|
403
|
+
train_losses[k] = v + train_losses.get(k, 0.0)
|
|
404
|
+
# // epoch loop
|
|
405
|
+
train_data.cleanup()
|
|
406
|
+
train_losses = {
|
|
407
|
+
k: v.item() / len(train_data) for k, v in train_losses.items()
|
|
408
|
+
}
|
|
409
|
+
train_records.append({**train_losses})
|
|
410
|
+
|
|
411
|
+
# Validation phase (only if val_loader is not None)
|
|
412
|
+
val_losses = {}
|
|
413
|
+
val_loss = None
|
|
414
|
+
if val_data is not None:
|
|
415
|
+
self.eval()
|
|
416
|
+
val_losses = {}
|
|
417
|
+
with torch.no_grad():
|
|
418
|
+
for (
|
|
419
|
+
waveform_batch,
|
|
420
|
+
channels_batch,
|
|
421
|
+
noise_batch,
|
|
422
|
+
ell_batch,
|
|
423
|
+
) in val_data:
|
|
424
|
+
waveform_batch = waveform_batch.to(self.device)
|
|
425
|
+
channels_batch = channels_batch.to(self.device)
|
|
426
|
+
noise_batch = noise_batch.to(self.device)
|
|
427
|
+
if ell_batch is not None:
|
|
428
|
+
ell_batch = ell_batch.to(self.device)
|
|
429
|
+
|
|
430
|
+
waveform_batch, mask = self.to_nn_channels(
|
|
431
|
+
waveform_batch, channels_batch
|
|
432
|
+
)
|
|
433
|
+
m = noise_batch.to(waveform_batch)
|
|
434
|
+
fres = self.train_forward(
|
|
435
|
+
waveform_batch, m, ell_batch, mask
|
|
436
|
+
)
|
|
437
|
+
loss_dict = self.loss(
|
|
438
|
+
mask,
|
|
439
|
+
waveform_batch,
|
|
440
|
+
m,
|
|
441
|
+
fres,
|
|
442
|
+
l1_alpha=self.l1_alpha,
|
|
443
|
+
l4_alpha=self.l4_alpha,
|
|
444
|
+
output_l1_alpha=self.output_l1_alpha,
|
|
445
|
+
)
|
|
446
|
+
for k, v in loss_dict.items():
|
|
447
|
+
val_losses[k] = v + val_losses.get(k, 0.0)
|
|
448
|
+
|
|
449
|
+
val_losses = {
|
|
450
|
+
k: v.item() / len(val_data) for k, v in val_losses.items()
|
|
451
|
+
}
|
|
452
|
+
val_loss = sum(val_losses.values())
|
|
453
|
+
train_records[-1]["val_loss"] = val_loss
|
|
454
|
+
|
|
455
|
+
if (
|
|
456
|
+
self.earlystop_eps is not None
|
|
457
|
+
and last_val_loss is not None
|
|
458
|
+
and val_loss - last_val_loss > self.earlystop_eps
|
|
459
|
+
):
|
|
460
|
+
if epoch >= self.min_epochs:
|
|
461
|
+
print(f"Early stopping after {epoch} epochs.")
|
|
462
|
+
break
|
|
463
|
+
last_val_loss = val_loss
|
|
464
|
+
|
|
465
|
+
if self.step_callback is not None:
|
|
466
|
+
self.step_callback(self, epoch, val_loss)
|
|
467
|
+
|
|
468
|
+
# Print loss summary
|
|
469
|
+
loss_str = f"Train {loss:.4f} " + "|".join( # type: ignore
|
|
470
|
+
f"{k}: {v:.3f}" for k, v in train_losses.items()
|
|
471
|
+
)
|
|
472
|
+
if val_data is not None:
|
|
473
|
+
loss_str += f" Val {val_loss:.4f}" + "|".join(
|
|
474
|
+
f"{k}: {v:.3f}" for k, v in val_losses.items()
|
|
475
|
+
)
|
|
476
|
+
pbar.set_description(f"Epochs [{loss_str}]")
|
|
477
|
+
|
|
478
|
+
self.step_scheduler(scheduler, loss, val_loss) # type: ignore
|
|
479
|
+
|
|
480
|
+
train_df = pd.DataFrame.from_records(train_records)
|
|
481
|
+
return train_df
|
|
482
|
+
|
|
483
|
+
def _construct_datasets_from_waveforms(
|
|
484
|
+
self, waveforms, channels, recording, weights=None
|
|
485
|
+
):
|
|
486
|
+
rg = np.random.default_rng(self.random_seed)
|
|
487
|
+
|
|
488
|
+
val_size = 0
|
|
489
|
+
train_indices = slice(None)
|
|
490
|
+
val_indices = None
|
|
491
|
+
if self.val_split_p:
|
|
492
|
+
num_samples = len(waveforms)
|
|
493
|
+
val_size = int(self.val_split_p * num_samples)
|
|
494
|
+
train_size = num_samples - val_size
|
|
495
|
+
train_indices = rg.choice(num_samples, size=train_size, replace=False)
|
|
496
|
+
val_indices = np.setdiff1d(np.arange(num_samples), train_indices)
|
|
497
|
+
|
|
498
|
+
spike_length_samples = waveforms.shape[1]
|
|
499
|
+
|
|
500
|
+
# training dataset
|
|
501
|
+
train_waveforms = waveforms[train_indices]
|
|
502
|
+
train_channels = channels[train_indices]
|
|
503
|
+
train_dataset = TensorDataset(train_waveforms, train_channels)
|
|
504
|
+
train_noise_dataset = AsyncSameChannelRecordingNoiseDataset(
|
|
505
|
+
recording,
|
|
506
|
+
train_channels.numpy(force=True),
|
|
507
|
+
self.b.model_channel_index.numpy(force=True),
|
|
508
|
+
spike_length_samples=spike_length_samples,
|
|
509
|
+
generator=spawn_torch_rg(rg),
|
|
510
|
+
queue_chunks=self.queue_chunks,
|
|
511
|
+
)
|
|
512
|
+
if self.cycle_loss_alpha:
|
|
513
|
+
train_cycle_noise_dataset = AsyncSameChannelRecordingNoiseDataset(
|
|
514
|
+
recording,
|
|
515
|
+
train_channels.numpy(force=True),
|
|
516
|
+
self.b.model_channel_index.numpy(force=True),
|
|
517
|
+
spike_length_samples=spike_length_samples,
|
|
518
|
+
generator=spawn_torch_rg(rg),
|
|
519
|
+
queue_chunks=self.queue_chunks,
|
|
520
|
+
)
|
|
521
|
+
else:
|
|
522
|
+
train_cycle_noise_dataset = NoneDataset(len(train_channels))
|
|
523
|
+
|
|
524
|
+
train_stack_dataset = StackDataset(
|
|
525
|
+
train_dataset, train_noise_dataset, train_cycle_noise_dataset
|
|
526
|
+
)
|
|
527
|
+
train_weights = None if weights is None else weights[train_indices]
|
|
528
|
+
train_sampler = AOTIndicesWeightedRandomBatchSampler(
|
|
529
|
+
n_examples=len(train_channels),
|
|
530
|
+
weights=train_weights,
|
|
531
|
+
replacement=train_weights is not None,
|
|
532
|
+
batch_size=self.batch_size,
|
|
533
|
+
generator=spawn_torch_rg(rg),
|
|
534
|
+
epoch_size=self.epoch_size,
|
|
535
|
+
)
|
|
536
|
+
train_loader = DataLoader(
|
|
537
|
+
train_stack_dataset,
|
|
538
|
+
sampler=train_sampler,
|
|
539
|
+
num_workers=0,
|
|
540
|
+
batch_size=None,
|
|
541
|
+
)
|
|
542
|
+
train_data = DecolliderDataLoader(
|
|
543
|
+
loader=train_loader,
|
|
544
|
+
sampler=train_sampler,
|
|
545
|
+
noise_dataset=train_noise_dataset,
|
|
546
|
+
cycle_noise_dataset=train_cycle_noise_dataset,
|
|
547
|
+
spike_length_samples=spike_length_samples,
|
|
548
|
+
)
|
|
549
|
+
|
|
550
|
+
# initialize validation datasets only if val_split_p > 0
|
|
551
|
+
if val_size > 0:
|
|
552
|
+
val_waveforms = waveforms[val_indices]
|
|
553
|
+
val_channels = channels[val_indices]
|
|
554
|
+
val_noise = get_noise(
|
|
555
|
+
recording,
|
|
556
|
+
val_channels.numpy(force=True),
|
|
557
|
+
self.b.model_channel_index.numpy(force=True),
|
|
558
|
+
spike_length_samples=spike_length_samples,
|
|
559
|
+
rg=rg,
|
|
560
|
+
)
|
|
561
|
+
if self.cycle_loss_alpha:
|
|
562
|
+
cycle_val_noise = get_noise(
|
|
563
|
+
recording,
|
|
564
|
+
val_channels.numpy(force=True),
|
|
565
|
+
self.b.model_channel_index.numpy(force=True),
|
|
566
|
+
spike_length_samples=spike_length_samples,
|
|
567
|
+
rg=rg,
|
|
568
|
+
)
|
|
569
|
+
cycle_val_noise = TensorDataset(cycle_val_noise)
|
|
570
|
+
else:
|
|
571
|
+
cycle_val_noise = NoneDataset(len(train_channels))
|
|
572
|
+
val_dataset = TensorDataset(val_waveforms, val_channels)
|
|
573
|
+
val_noise_dataset = TensorDataset(val_noise)
|
|
574
|
+
|
|
575
|
+
# val set does not need shuffling
|
|
576
|
+
val_loader = DataLoader(
|
|
577
|
+
val_dataset,
|
|
578
|
+
num_workers=self.n_data_workers, # type: ignore
|
|
579
|
+
persistent_workers=bool(self.n_data_workers),
|
|
580
|
+
batch_size=self.batch_size,
|
|
581
|
+
)
|
|
582
|
+
val_data = DecolliderDataLoader(
|
|
583
|
+
loader=val_loader,
|
|
584
|
+
sampler=None,
|
|
585
|
+
noise_dataset=val_noise_dataset,
|
|
586
|
+
cycle_noise_dataset=cycle_val_noise,
|
|
587
|
+
spike_length_samples=spike_length_samples,
|
|
588
|
+
)
|
|
589
|
+
else:
|
|
590
|
+
val_data = None
|
|
591
|
+
|
|
592
|
+
return train_data, val_data
|
|
593
|
+
|
|
594
|
+
|
|
595
|
+
@dataclasses.dataclass(kw_only=True, frozen=True)
|
|
596
|
+
class DecolliderDataLoader:
|
|
597
|
+
loader: DataLoader
|
|
598
|
+
sampler: Sampler | None
|
|
599
|
+
noise_dataset: Dataset
|
|
600
|
+
cycle_noise_dataset: Dataset | NoneDataset
|
|
601
|
+
spike_length_samples: int
|
|
602
|
+
|
|
603
|
+
def __len__(self):
|
|
604
|
+
return len(self.loader)
|
|
605
|
+
|
|
606
|
+
def __iter__(self):
|
|
607
|
+
"""Flatten..."""
|
|
608
|
+
for batch in self.loader:
|
|
609
|
+
batch_flat = []
|
|
610
|
+
for item in batch:
|
|
611
|
+
if isinstance(item, (list, tuple)):
|
|
612
|
+
batch_flat.extend(item)
|
|
613
|
+
else:
|
|
614
|
+
batch_flat.append(item)
|
|
615
|
+
yield batch_flat
|
|
616
|
+
|
|
617
|
+
def refresh(self):
|
|
618
|
+
if hasattr(self.sampler, "refresh"):
|
|
619
|
+
self.sampler.refresh()
|
|
620
|
+
if hasattr(self.noise_dataset, "refresh"):
|
|
621
|
+
self.noise_dataset.refresh(self.sampler.indices)
|
|
622
|
+
if self.cycle_noise_dataset is not None and hasattr(
|
|
623
|
+
self.cycle_noise_dataset, "refresh"
|
|
624
|
+
):
|
|
625
|
+
self.cycle_noise_dataset.refresh(self.sampler.indices)
|
|
626
|
+
else:
|
|
627
|
+
assert not hasattr(self.noise_dataset, "refresh")
|
|
628
|
+
|
|
629
|
+
def cleanup(self):
|
|
630
|
+
if hasattr(self.noise_dataset, "cleanup"):
|
|
631
|
+
self.noise_dataset.cleanup()
|
|
632
|
+
if self.cycle_noise_dataset is not None and hasattr(
|
|
633
|
+
self.cycle_noise_dataset, "cleanup"
|
|
634
|
+
):
|
|
635
|
+
self.cycle_noise_dataset.cleanup()
|