From b44de0e11281c723ec426f8bec8ca0897ecfe3ff Mon Sep 17 00:00:00 2001 From: Gustaf Rydholm Date: Sun, 21 Nov 2021 21:34:53 +0100 Subject: Remove VQVAE stuff, did not work... --- text_recognizer/networks/quantizer/kmeans.py | 32 ---------------------------- 1 file changed, 32 deletions(-) delete mode 100644 text_recognizer/networks/quantizer/kmeans.py (limited to 'text_recognizer/networks/quantizer/kmeans.py') diff --git a/text_recognizer/networks/quantizer/kmeans.py b/text_recognizer/networks/quantizer/kmeans.py deleted file mode 100644 index a34c381..0000000 --- a/text_recognizer/networks/quantizer/kmeans.py +++ /dev/null @@ -1,32 +0,0 @@ -"""K-means clustering for embeddings.""" -from typing import Tuple - -from einops import repeat -import torch -from torch import Tensor - -from text_recognizer.networks.quantizer.utils import norm, sample_vectors - - -def kmeans( - samples: Tensor, num_clusters: int, num_iters: int = 10 -) -> Tuple[Tensor, Tensor]: - """Compute k-means clusters.""" - D = samples.shape[-1] - - means = sample_vectors(samples, num_clusters) - - for _ in range(num_iters): - dists = samples @ means.t() - buckets = dists.max(dim=-1).indices - bins = torch.bincount(buckets, minlength=num_clusters) - zero_mask = bins == 0 - bins_min_clamped = bins.masked_fill(zero_mask, 1) - - new_means = buckets.new_zeros(num_clusters, D).type_as(samples) - new_means.scatter_add_(0, repeat(buckets, "n -> n d", d=D), samples) - new_means /= bins_min_clamped[..., None] - new_means = norm(new_means) - means = torch.where(zero_mask[..., None], means, new_means) - - return means, bins -- cgit v1.2.3-70-g09d2