2021-10-05 02:59:21 +00:00
|
|
|
import random
|
|
|
|
from math import prod
|
|
|
|
|
|
|
|
import torch
|
|
|
|
import torch.nn as nn
|
|
|
|
import torch.nn.functional as F
|
|
|
|
|
|
|
|
|
|
|
|
# Fits a soft-discretized input to a normal-PDF across the specified dimension.
|
|
|
|
# In other words, attempts to force the discretization function to have a mean equal utilization across all discrete
|
|
|
|
# values with the specified expected variance.
|
|
|
|
class DiscretizationLoss(nn.Module):
|
2021-10-06 23:10:50 +00:00
|
|
|
def __init__(self, discrete_bins, dim, expected_variance, store_past=0):
|
2021-10-05 02:59:21 +00:00
|
|
|
super().__init__()
|
2021-10-06 23:10:50 +00:00
|
|
|
self.discrete_bins = discrete_bins
|
2021-10-05 02:59:21 +00:00
|
|
|
self.dim = dim
|
|
|
|
self.dist = torch.distributions.Normal(0, scale=expected_variance)
|
2021-10-06 23:10:50 +00:00
|
|
|
if store_past > 0:
|
|
|
|
self.record_past = True
|
|
|
|
self.register_buffer("accumulator_index", torch.zeros(1, dtype=torch.long, device='cpu'))
|
|
|
|
self.register_buffer("accumulator_filled", torch.zeros(1, dtype=torch.long, device='cpu'))
|
|
|
|
self.register_buffer("accumulator", torch.zeros(store_past, discrete_bins))
|
|
|
|
else:
|
|
|
|
self.record_past = False
|
2021-10-05 02:59:21 +00:00
|
|
|
|
|
|
|
def forward(self, x):
|
|
|
|
other_dims = set(range(len(x.shape)))-set([self.dim])
|
|
|
|
averaged = x.sum(dim=tuple(other_dims)) / x.sum()
|
|
|
|
averaged = averaged - averaged.mean()
|
2021-10-06 23:10:50 +00:00
|
|
|
|
|
|
|
if self.record_past:
|
|
|
|
acc_count = self.accumulator.shape[0]
|
|
|
|
avg = averaged.detach().clone()
|
|
|
|
if self.accumulator_filled > 0:
|
|
|
|
averaged = torch.mean(self.accumulator, dim=0) * (acc_count-1) / acc_count + \
|
|
|
|
averaged / acc_count
|
|
|
|
|
|
|
|
# Also push averaged into the accumulator.
|
|
|
|
self.accumulator[self.accumulator_index] = avg
|
|
|
|
self.accumulator_index += 1
|
|
|
|
if self.accumulator_index >= acc_count:
|
|
|
|
self.accumulator_index *= 0
|
|
|
|
if self.accumulator_filled <= 0:
|
|
|
|
self.accumulator_filled += 1
|
|
|
|
|
2021-10-05 02:59:21 +00:00
|
|
|
return torch.sum(-self.dist.log_prob(averaged))
|
|
|
|
|
|
|
|
|
|
|
|
if __name__ == '__main__':
|
2021-10-06 23:10:50 +00:00
|
|
|
d = DiscretizationLoss(1024, 1, 1e-6, store_past=20)
|
|
|
|
for _ in range(500):
|
|
|
|
v = torch.randn(16, 1024, 500)
|
|
|
|
#for k in range(5):
|
|
|
|
# v[:, random.randint(0,8192), :] += random.random()*100
|
|
|
|
v = F.softmax(v, 1)
|
|
|
|
print(d(v))
|