-
Notifications
You must be signed in to change notification settings - Fork 0
/
sample_channel_fitnesses.py
executable file
·77 lines (63 loc) · 2.22 KB
/
sample_channel_fitnesses.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
import data
import models
import jax
from jax import jit, vmap
import jax.numpy as jnp
import util as u
import argparse
import sys
import simple_ga
import numpy as np
parser = argparse.ArgumentParser(
formatter_class=argparse.ArgumentDefaultsHelpFormatter)
parser.add_argument('--params', type=str, required=True)
parser.add_argument('--split', type=str, required=True)
parser.add_argument('--num-examples', type=int, required=True)
parser.add_argument('--model-type', type=str, default='single',
help="model type; 'single' or 'multi-res'")
opts = parser.parse_args()
print(opts, file=sys.stderr)
assert opts.model_type in ['single', 'multi-res']
dataset = data.dataset(split=opts.split,
batch_size=opts.num_examples)
for x, y_true in dataset:
break
params = u.load_params(opts.params)
@jit
def mean_loss(member):
if opts.model_type == 'single':
# member denotes a channel mask we want to apply to
# entire x batch
model = models.construct_single_trunk_model()
mask_tile_shape = list(x.shape)
mask_tile_shape[-1] = 1
mask = jnp.tile(member, mask_tile_shape)
logits = model.apply(params, x * mask)
else: # multi-res
# member denotes channel selection handled in model
model = models.construct_multires_model()
logits = model.apply(params, x, member)
return u.softmax_cross_entropy(logits, y_true).mean()
@jit
def channel_penalty(member):
if opts.model_type == 'single':
return 0
penalty = jnp.sum(jnp.equal(member, 0)) * 0.8 # x64
penalty += jnp.sum(jnp.equal(member, 1)) * 0.4 # x32
penalty += jnp.sum(jnp.equal(member, 2)) * 0.2 # x16
penalty += jnp.sum(jnp.equal(member, 3)) * 0.1 # x8
# channel 4, x0, is free to use
return penalty
def new_member():
if opts.model_type == 'single':
return np.random.randint(0, 2, size=(13,))
else: # multi-res
return np.random.randint(0, 5, size=(13,))
print("member\tloss\tpenalty")
for _ in range(1000):
random_member = new_member()
m = list(random_member)
l = float(mean_loss(random_member))
p = float(channel_penalty(random_member))
print("\t".join(map(str, [m, l, p])))
sys.stdout.flush()