ameerazam08's picture
Upload folder using huggingface_hub (#1)
9e7a39a
raw
history blame
1.68 kB
from abc import abstractmethod
from typing import Any, Tuple
import torch
import torch.nn as nn
import torch.nn.functional as F
from ....modules.distributions.distributions import DiagonalGaussianDistribution
class AbstractRegularizer(nn.Module):
def __init__(self):
super().__init__()
def forward(self, z: torch.Tensor) -> Tuple[torch.Tensor, dict]:
raise NotImplementedError()
@abstractmethod
def get_trainable_parameters(self) -> Any:
raise NotImplementedError()
class DiagonalGaussianRegularizer(AbstractRegularizer):
def __init__(self, sample: bool = True):
super().__init__()
self.sample = sample
def get_trainable_parameters(self) -> Any:
yield from ()
def forward(self, z: torch.Tensor) -> Tuple[torch.Tensor, dict]:
log = dict()
posterior = DiagonalGaussianDistribution(z)
if self.sample:
z = posterior.sample()
else:
z = posterior.mode()
kl_loss = posterior.kl()
kl_loss = torch.sum(kl_loss) / kl_loss.shape[0]
log["kl_loss"] = kl_loss
return z, log
def measure_perplexity(predicted_indices, num_centroids):
# src: https://github.com/karpathy/deep-vector-quantization/blob/main/model.py
# eval cluster perplexity. when perplexity == num_embeddings then all clusters are used exactly equally
encodings = (
F.one_hot(predicted_indices, num_centroids).float().reshape(-1, num_centroids)
)
avg_probs = encodings.mean(0)
perplexity = (-(avg_probs * torch.log(avg_probs + 1e-10)).sum()).exp()
cluster_use = torch.sum(avg_probs > 0)
return perplexity, cluster_use