forked from EPFL-VILAB/TaskDiscovery
-
Notifications
You must be signed in to change notification settings - Fork 0
/
utils.py
219 lines (180 loc) · 7.13 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
import os
import numpy as np
import torch
import pytorch_lightning as pl
import random
import functools
from torch.distributions import Categorical
from itertools import combinations
from sklearn.metrics import pairwise_distances
from models.tasks import CIFAR_REAL_BIN_TASKS
def hamming_sym(a, b=None, binary=True):
s = 1 - pairwise_distances(a, b, metric='hamming', n_jobs=-1)
if binary:
s[s < 0.5] = 1 - s[s < 0.5]
return s
def tonp(x):
if isinstance(x, (np.ndarray, float, int)):
return np.array(x)
return x.detach().cpu().numpy()
def viz_array_grid(array, rows, cols, padding=0, channels_last=False, normalize=False, **kwargs):
# normalization
'''
Args:
array: (N_images, N_channels, H, W) or (N_images, H, W, N_channels)
rows, cols: rows and columns of the plot. rows * cols == array.shape[0]
padding: padding between cells of plot
channels_last: for Tensorflow = True, for PyTorch = False
normalize: `False`, `mean_std`, or `min_max`
Kwargs:
if normalize == 'mean_std':
mean: mean of the distribution. Default 0.5
std: std of the distribution. Default 0.5
if normalize == 'min_max':
min: min of the distribution. Default array.min()
max: max if the distribution. Default array.max()
'''
array = tonp(array)
if not channels_last:
array = np.transpose(array, (0, 2, 3, 1))
array = array.astype('float32')
if normalize:
if normalize == 'mean_std':
mean = kwargs.get('mean', 0.5)
mean = np.array(mean).reshape((1, 1, 1, -1))
std = kwargs.get('std', 0.5)
std = np.array(std).reshape((1, 1, 1, -1))
array = array * std + mean
elif normalize == 'min_max':
min_ = kwargs.get('min', array.min())
min_ = np.array(min_).reshape((1, 1, 1, -1))
max_ = kwargs.get('max', array.max())
max_ = np.array(max_).reshape((1, 1, 1, -1))
array -= min_
array /= max_ + 1e-9
batch_size, H, W, channels = array.shape
assert rows * cols == batch_size
if channels == 1:
canvas = np.ones((H * rows + padding * (rows - 1),
W * cols + padding * (cols - 1)))
array = array[:, :, :, 0]
elif channels == 3:
canvas = np.ones((H * rows + padding * (rows - 1),
W * cols + padding * (cols - 1),
3))
else:
raise TypeError('number of channels is either 1 of 3')
for i in range(rows):
for j in range(cols):
img = array[i * cols + j]
start_h = i * padding + i * H
start_w = j * padding + j * W
canvas[start_h: start_h + H, start_w: start_w + W] = img
canvas = np.clip(canvas, 0, 1)
canvas *= 255.0
canvas = canvas.astype('uint8')
return canvas
def _entropy_with_logits(logits):
return Categorical(logits=logits).entropy()
def _mean_categorical_with_logits(logits):
logp = torch.log_softmax(logits, dim=1)
logp_mean = torch.logsumexp(logp - np.log(logits.shape[0]*1.), dim=0)
return logp_mean
def _l2_reg(hparams, params):
# l2 biased regularization
return sum([((b - p) ** 2).sum() for b, p in zip(hparams, params)])
def set_seeds(seed):
torch.manual_seed(seed)
torch.cuda.manual_seed(seed)
np.random.seed(seed)
random.seed(seed)
class CheckpointEveryNSteps(pl.Callback):
"""
Save a checkpoint every N steps, instead of Lightning's default that checkpoints
based on validation loss.
from https://github.com/PyTorchLightning/pytorch-lightning/issues/2534#issuecomment-674582085
"""
def __init__(
self,
save_step_frequency,
prefix="checkpoint",
use_modelcheckpoint_filename=False,
):
"""
Args:
save_step_frequency: how often to save in steps
prefix: add a prefix to the name, only used if
use_modelcheckpoint_filename=False
use_modelcheckpoint_filename: just use the ModelCheckpoint callback's
default filename, don't use ours.
"""
self.save_step_frequency = save_step_frequency
self.prefix = prefix
self.use_modelcheckpoint_filename = use_modelcheckpoint_filename
def on_train_batch_end(self, trainer: pl.Trainer, *_, force_save=False):
""" Check if we should save a checkpoint after every train batch """
epoch = trainer.current_epoch
global_step = trainer.global_step
if self.save_step_frequency == -1: return
if force_save or global_step % self.save_step_frequency == 0 or global_step == 1:
if self.use_modelcheckpoint_filename:
filename = trainer.checkpoint_callback.filename
else:
filename = f"{self.prefix}.ckpt"
ckpt_path = os.path.join(trainer.checkpoint_callback.dirpath, filename)
trainer.save_checkpoint(ckpt_path)
print(f'[Checkpoint] ===> Saved to {ckpt_path}')
def on_exception(self, trainer: "pl.Trainer", pl_module: "pl.LightningModule", exception: BaseException) -> None:
print('[Checkpoint] ===> Saving on interruption...')
self.on_train_batch_end(trainer, force_save=True)
def rvs(dim=3, seed=None):
"""
Return dim random perpendicular vectors in R^dim
"""
if seed is None:
random_state = np.random
else:
random_state = np.random.default_rng(seed)
H = np.eye(dim)
D = np.ones((dim,))
for n in range(1, dim):
x = random_state.normal(size=(dim-n+1,))
D[n-1] = np.sign(x[0])
x[0] -= D[n-1]*np.sqrt((x*x).sum())
# Householder transformation
Hx = (np.eye(dim-n+1) - 2.*np.outer(x, x)/(x*x).sum())
mat = np.eye(dim)
mat[n-1:, n-1:] = Hx
H = np.dot(H, mat)
# Fix the last sign such that the determinant is 1
D[-1] = (-1)**(1-(dim % 2))*D.prod()
# Equivalent to np.dot(np.diag(D), H) but faster, apparently
H = (D*H.T).T
return H
def rot_by_alpha_deg(v, deg):
theta = np.deg2rad(deg)
rot = np.array([[np.cos(theta), -np.sin(theta)], [np.sin(theta), np.cos(theta)]])
return v @ rot
def random_k_way_linear_task(K, d, seed):
basis = rvs(d, seed)[:2]
w = np.array([0, 1.])
ws = np.stack([rot_by_alpha_deg(w, 360/K*i) for i in range(K)]).T
return np.sum(basis[:, None] * ws[..., None], 0).T
def get_all_binary_tasks(classes):
tasks = list(combinations(classes, len(classes)//2))
return tasks
def get_main_tasks_idxs_from_included_classes(classes):
classes = set(classes)
reduced_tasks = list(map(set, get_all_binary_tasks(classes)))
main_task_idxs = []
_taken_tasks = []
for i, cls1 in enumerate(CIFAR_REAL_BIN_TASKS):
upd_cls1 = set(cls1).intersection(classes)
if upd_cls1 in reduced_tasks and upd_cls1 not in _taken_tasks:
main_task_idxs.append(i)
_taken_tasks.append(upd_cls1)
return main_task_idxs
def partialclass(cls, *args, **kwds):
class NewCls(cls):
__init__ = functools.partialmethod(cls.__init__, *args, **kwds)
return NewCls