Training and evaluation loss goes down however, WER score stays the same

ahabahab8 · May 23, 2022, 12:07pm

Hello.
I have a *.ipynb file in which i am trying to fine-tune wav2vec2.0 XLSR large model on common voice Farsi dataset.
I don’t know why but as steps go further, loss decreases but WER score stays the same.
This is the link to my *.ipynb file.

github.com

rsd16/xlrsr-53-for-common-voice-persian-dataset/blob/main/nouvoversion2.ipynb

{
  "nbformat": 4,
  "nbformat_minor": 0,
  "metadata": {
    "colab": {
      "name": "nouvoversion2.ipynb",
      "provenance": [],
      "collapsed_sections": []
    },
    "kernelspec": {
      "name": "python3",
      "display_name": "Python 3"
    },
    "language_info": {
      "name": "python"
    },
    "accelerator": "GPU",
    "widgets": {
      "application/vnd.jupyter.widget-state+json": {
        "37bb09bf9dbd41e0bb670a6daa9d8928": {

This file has been truncated. show original

Can anybody please help me with this?
Thank you

Topic		Replies	Views
Wav2vec2 xlsr nan train loss Models	1	854	June 14, 2021
`nan` training loss but eval loss does improve over time Research	5	3433	October 10, 2022
Fine-tuning Wav2v2.0: Loss increasing, WER decreasing Models	2	478	June 30, 2023
Finetuning Wav2Vec2 loss constant Beginners	1	206	August 14, 2023
Inconsistent evaluation result (WER) when finetuning wav2vev2 pretrained model Beginners	2	388	November 7, 2023

Training and evaluation loss goes down however, WER score stays the same

Related Topics