From ae78ac7480a0d90a5efee4f649c46cff20d7f4fa Mon Sep 17 00:00:00 2001 From: Jeremy Fix Date: Wed, 27 Dec 2023 16:47:35 +0100 Subject: [PATCH] inclure the overfit curve on the minibatch --- Webpage/02-pytorch-asr.md | 2 ++ 1 file changed, 2 insertions(+) diff --git a/Webpage/02-pytorch-asr.md b/Webpage/02-pytorch-asr.md index cceb717..21169c9 100644 --- a/Webpage/02-pytorch-asr.md +++ b/Webpage/02-pytorch-asr.md @@ -348,6 +348,8 @@ mymachine:~:mylogin$ python3 main_ctc.py train --debug --batch_size 16 You should pretty "quickly" see a null loss on the training set with a perfect decoding. Note that it still takes 2s. per minibatch and a hundred epochs for overfitting. +![CTC loss on the test/train/valid sets with an overfitting architecture, considering only one minibatch of size 16.](./data/02-pytorch-asr/overfit_minibatch.png){.bordered} + ### Overfitting the training set The next step is to design a sufficiently rich architecture to overfit the training set when any sort of regularization is disabled (L2/L1, dropout, data augmentation, ...) :