-
Notifications
You must be signed in to change notification settings - Fork 0
/
utils.py
268 lines (227 loc) · 8.64 KB
/
utils.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
import torch, torchvision
import pytorch_lightning as pl
from pytorch_lightning.loggers import TensorBoardLogger
from lightning.pytorch.callbacks import ProgressBar, StochasticWeightAveraging
from torchvision import datasets, transforms
from torch import nn, optim
from torch.nn import functional as F
from torch.utils.data import IterableDataset, DataLoader
import numpy as np
from scipy.special import gammaln
import matplotlib.pyplot as plt
from typing import Callable, List, Optional, Tuple, Generator, Dict
import os
import sys
import random
import io
import warnings
import time
import math
import functools
import collections
import traceback
from einops import rearrange, repeat
from dotmap import DotMap
from glob import glob
# The following code handles some bug from pytorch lightning when training on Yanke's cluster.
from pytorch_lightning.plugins.environments import SLURMEnvironment
class DisabledSLURMEnvironment(SLURMEnvironment):
def detect() -> bool:
return False
@staticmethod
def _validate_srun_used() -> None:
return
@staticmethod
def _validate_srun_variables() -> None:
return
def todevice(x, device):
if isinstance(x,dict):
for k, v in x.items():
x[k] = todevice(x[k], device)
elif isinstance(x,list):
x = [todevice(i, device) for i in x]
elif not isinstance(x,str):
x = x.to(device)
return x
def merge_with_mask(event_t_list, T_mask, mesh_t_list):
'''
Merge two t lists
Input:
event_t_list: (B, n, 1)
mesh_t_list: (B, m, 1)
T_mask: (B, n)
'''
b, n, _ = event_t_list.shape
_, m, _ = mesh_t_list.shape
total_t_list = torch.zeros(b, m+n, 1).to(event_t_list.device)
total_mask = torch.zeros(b, m+n).to(event_t_list.device)
for i in range(b):
nn = torch.sum(T_mask[i,:])
total_t_list[i,:(nn+m),0] = torch.sort(torch.cat((event_t_list[i,:nn,0], mesh_t_list[i,:,0])))[0]
total_mask[i,:(nn+m)] = 1
return total_t_list, total_mask.bool()
def interpolate(mesh_rate_list, event_t_list):
'''
Input:
mesh_rate_list: (B, resolution+1, E_bins)
event_t_list: (B, n)
Output:
event_rate_list: (B, n, E_bins)
'''
B, resolutionp1, E_bins = mesh_rate_list.shape
resolution = resolutionp1 - 1
B_scaled = event_t_list * resolution
B_floor = torch.floor(B_scaled).long()
B_ceil = torch.ceil(B_scaled).long()
B_remainder = B_scaled - B_floor.float() # (B, n)
# Gather the values from A
A_floor = mesh_rate_list.gather(1, B_floor.unsqueeze(2).expand(-1, -1, E_bins))
A_ceil = mesh_rate_list.gather(1, B_ceil.unsqueeze(2).expand(-1, -1, E_bins))
# Interpolate
event_rate_list = A_floor + (A_ceil - A_floor) * B_remainder.unsqueeze(2)
return event_rate_list
def load_from_less_latents(model, small_dset, big_dset, ckpt_path):
state = torch.load(ckpt_path)
state_dict = state['state_dict']
del state
small_latents = state_dict['latent']
del state_dict['latent']
model.load_state_dict(state_dict, strict=False)
del state_dict
# Load latents
i = 0
j = 0
m = len(small_dset)
assert small_latents.shape[0] == m
small_latents = small_latents.detach()
n = len(big_dset)
with torch.no_grad():
while True:
if i == m:
print('loading complete')
break
if j == n and i < m:
raise ValueError("Not everything is loaded")
if small_dset[i]['id'] == big_dset[j]['id']:
model.latent[j].copy_(small_latents[i])
i += 1
j += 1
else:
j += 1
return model
def loglikelihood(log_event_rate_list, T_mask, E_mask, log_mesh_rate_list, T):
'''
log likelihood of a batch of event list with the same length.
r(t1) * ... * r(tn) * exp(-integral(r(t)))
We take the log likelihood for better computational performance
Input:
log_event_rate_list: (B, n_event, E_bins)
T_mask: (B, n_event), if mask == 0 then it's a padding
E_mask: (B, n_event, E_bins)
log_mesh_rate_list: (B, n_mesh, E_bins)
T: (B,)
'''
B, n_mesh, E_bins = log_mesh_rate_list.shape
integral = 0.5 * (
torch.sum(log_mesh_rate_list[:,1:,:].exp(), dim=(1,2))
+ torch.sum(log_mesh_rate_list[:,:-1,:].exp(), dim=(1,2))
) * T / (n_mesh-1) # (B,)
return ((log_event_rate_list * T_mask.unsqueeze(-1) * E_mask).sum(dim=(1,2)) - integral).mean()
def total_variation(rate_list, T_mask=None):
'''
Calculate total variation for a log rate list. The absolute value of the first entry is calculated twice
Input:
rate_list: (B, n, E_bins)
T_mask: (B, n)
'''
if T_mask is not None:
rate_list = rate_list * T_mask.unsqueeze(-1)
b, n, e = rate_list.shape
s = 0
for i in range(b):
nn = torch.sum(T_mask[i,:])
temp = rate_list[i,:nn,:]
s += torch.sum(torch.abs(torch.diff(temp,dim=0)))
return s / b
else:
return (rate_list[:,1:,:] - rate_list[:,:-1,:]).abs().sum(dim=(1,2)).mean()
def total_variation_normalized(rate_list, T_mask=None):
'''
Calculate total variation for a log rate list. The absolute value of the first entry is calculated twice
Input:
rate_list: (B, n, E_bins)
T_mask: (B, n)
'''
if T_mask is not None:
rate_list = rate_list * T_mask.unsqueeze(-1)
b, n, e = rate_list.shape
s = 0
for i in range(b):
nn = torch.sum(T_mask[i,:])
if nn == 1:
continue
temp = rate_list[i,:nn,:]
s += torch.diff(temp,dim=0).abs().mean()
return s / b
else:
return (rate_list[:,1:,:] - rate_list[:,:-1,:]).abs().mean()
def visualize_hist(times, t_scale):
times = times / t_scale
plt.hist(times, bins = torch.arange(torch.ceil(torch.max(times))))
def plot_recon_grid(collated_outputs,
indices,
title_size=26,
label_size=22,
tick_size=20,
figsize=(12,12),
nbins=96,
t_scale=28800,
B=64,
):
k = len(indices)
kk = np.sqrt(k)
assert kk == int(kk)
kk = int(kk)
fig, axes = plt.subplots(kk,kk, figsize=figsize)
for i, total_index in enumerate(indices):
batch_index = total_index // B
index = total_index % B
mask = collated_outputs['mask'][batch_index][index]
times = collated_outputs['event_list'][batch_index][index,mask,0] * t_scale / 3600
total_mask = collated_outputs['total_mask'][batch_index][index]
total_times = collated_outputs['total_list'][batch_index][index,total_mask,0] * t_scale / 3600
rates = collated_outputs['total_rates'][batch_index][index,total_mask] / nbins
T = collated_outputs['T'][total_index] * t_scale
axes[i//kk,i%kk].hist(times, bins = nbins,label='Actual counts')
axes[i//kk,i%kk].plot(total_times, torch.sum(rates,dim=-1),label='Fitted rate',linewidth=2)
axes[i//kk,i%kk].tick_params(axis='both', which='major', labelsize=tick_size)
plt.tight_layout()
# Need modification, only at 0
def plot_recon_grid_energy(collated_outputs,
indices,
title_size=26,
label_size=22,
tick_size=20,
figsize=(12,12),
nbins=96,
t_scale=28800,
B=64,
):
k = len(indices)
kk = np.sqrt(k)
assert kk == int(kk)
kk = int(kk)
fig, axes = plt.subplots(kk,kk, figsize=figsize)
for i, total_index in enumerate(indices):
batch_index = total_index // B
index = total_index % B
mask = collated_outputs['mask'][batch_index][index]
times = collated_outputs['event_list'][batch_index][index,mask,0] * t_scale / 3600
total_mask = collated_outputs['total_mask'][batch_index][index]
total_times = collated_outputs['total_list'][batch_index][index,total_mask,0] * t_scale / 3600
rates = collated_outputs['total_rates'][batch_index][index,total_mask] / nbins
T = collated_outputs['T'][total_index] * t_scale
axes[i//kk,i%kk].hist(times, bins = nbins,label='Actual counts')
axes[i//kk,i%kk].plot(total_times, torch.sum(rates,dim=-1),label='Fitted rate',linewidth=2)
axes[i//kk,i%kk].tick_params(axis='both', which='major', labelsize=tick_size)
plt.tight_layout()