From cbbc707f41b8b26dcec5e9bc741447a68de4dad5 Mon Sep 17 00:00:00 2001
From: many-hats <52046775+many-hats@users.noreply.github.com>
Date: Tue, 12 Apr 2022 14:45:10 -0400
Subject: [PATCH] update hubert_feature_reader to skip small chunks

Very rarely, if the chunk is smaller than the kernel size (e.g. (x.size(1) % max_chunk) < 10), the feature reader will produce a runtime error:
RuntimeError: Calculated padded input size per channel: (1). Kernel size: (10). Kernel size can't be greater than actual input size
---
 textless/data/hubert_feature_reader.py | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/textless/data/hubert_feature_reader.py b/textless/data/hubert_feature_reader.py
index 4c774b5..45af901 100644
--- a/textless/data/hubert_feature_reader.py
+++ b/textless/data/hubert_feature_reader.py
@@ -66,6 +66,8 @@ def get_features(self, x):
         feat = []
         for start in range(0, x.size(1), self.max_chunk):
             x_chunk = x[:, start : start + self.max_chunk]
+            if x_chunk.size(1) < 10:
+                continue
             feat_chunk, _ = self.model.extract_features(
                 source=x_chunk,
                 padding_mask=None,